fix(memhub): restore and freeze the home when an upgrade migration fails

A failed SDK migration left a half-migrated home (schema migrations, moved legacy work, renamed Chroma collections) that the next run resumed on top of, the maintenance fence re-enabled the LaunchAgent so launchd retried at once, a deferred service idled forever, and capture hooks kept writing into the legacy database. Restore: before the first mutation the run records a restore point (the pre-cutover bundle, or a fresh bundle when resuming an old ledger row). Any failure before SDK authority is committed restores the management database (with WAL), config.json and hy-memory/governed-v2 through a verified staging copy and a journaled rename swap; the failed state is kept aside in backups/sdk-migration-failed/. A failure after the commit, or an unreadable authority, never restores. An interrupted swap is finished on the next start; meanwhile status reports required and Core refuses to serve. The migration's own code stays the result; the restore outcome is logged and reported separately. Freeze: a deferred migration turns the service enable marker off, makes the runtime service exit instead of idling, and leaves a frozen capture fence that hooks honour. Desktop launch, memhub install/upgrade and now start/restart re-run the migration and unfreeze on success, restoring the previous auto-start. The failure dialog says when data was restored, and its GitHub issue now prefills the body: from/to version, code, failing step, whether data was restored, macOS version and the log path shown with ~. migrate.invoked records the pre-upgrade version. sdk_data_plane_* codes now reach the dialog too. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Merge pull request #195 from Tencent-Hunyuan/feat/memory-panel-all-scenes
2026-10-27 19:11:04 +08:00 · 2026-10-27 18:04:32 +08:00 · 2026-10-27 16:54:08 +08:00 · 2026-10-27 15:59:35 +08:00 · 2026-10-27 15:55:00 +08:00 · 2026-10-27 15:42:38 +08:00
456 changed files with 26581 additions and 69590 deletions
@@ -2,6 +2,156 @@

 Hy Memory and MemHub are versioned and released independently.

+## Unreleased
+
+- The SDK migration no longer fails its first run on a home with legacy
+  vectors: reattaching them left a Chroma instance open with its own settings,
+  and the readback that followed could not open the same store
+  ("An instance of Chroma already exists ... with different settings").
+- MemHub "添加记忆" (add a memory) no longer waits behind unprocessed history.
+  Notes are claimed ahead of the event-time window that keeps history strictly
+  old to new; they still run one at a time and in order within their session.
+- The SDK memory page now reads its list, count, source and time labels and
+  graph from one page response for one source filter, so rows, totals and the
+  map no longer disagree. Memories with no recorded source are shown and
+  filterable as "unidentified source"; library search keeps only memories
+  that contain the words within the SDK's relevance window.
+- Metadata repair v2 follows `governed_memory_evidence` across every
+  pre-cutover bundle, which restores the source of memories the cutover
+  backfilled without one, and moves impossible future conversation times back
+  to the write time.
+- Reused the existing three-column memory map for SDK-authority list/detail
+  views and removed the retired "original evidence" surface. SDK-native status,
+  source, conversation time and creation time now render without fabricated
+  revisions; source and text filters are applied by the paged Core endpoint.
+- SDK memory search now calls the SDK search contract instead of listing the
+  entire library in Core. Corrupt empty records are skipped and diagnosed
+  without mutating data or failing a page, and post-cutover metadata repair
+  restores source and original times from the verified migration bundle.
+- Kept Cursor's WAL change token separate from the file mtime so a hash can no
+  longer appear as a conversation date centuries in the future.
+- A trusted packaged Runtime now repairs a dirty same-build installation by
+  staging a clean replacement, and `runtime_service_not_ready` is rendered as
+  user-facing startup guidance instead of an internal code.
+- Fixed legacy-to-SDK startup migration to adopt native Coding Experience and
+  Working Fact IDs instead of importing them as Chat layers. Raw migration now
+  rebuilds bounded user/assistant turns from the latest three days, excludes
+  tools, and enforces the same rolling retention after cutover.
+- Paused host capture writes with a live Desktop migration fence; Cursor
+  History catch-up resumes from its existing checkpoint after cutover.
+- Removed MemHub's private `hy_memory_bridge` implementation. MemHub now calls
+  public `HyMemoryClient` methods at runtime; prompt/reconcile/response
+  monkey-patches and private VDB/Coding runtime access were deleted. Kuzu Graph
+  projection/admin and stopped-writer App migration are the documented product
+  exceptions.
+- Added a MemHub-managed native Hermes MemoryProvider with automatic prefetch,
+  persistent per-session memory-ID deduplication, authorized Broker reads, and
+  recall-context stripping to prevent memory recursion.
+- Added an extensible MemHub interface-localization layer with Simplified
+  Chinese, English, and Spanish catalogs shared by Web and Electron-native
+  surfaces. Interface language remains independent from memory content and
+  model processing language.
+- Fixed governed-memory revisions and repeated evidence links failing when the
+  visibility dirty-marker trigger encounters the same memory twice. Startup
+  now replaces the affected legacy trigger without changing schema versions
+  or governed and visibility data.
+- Kept post-write authority verification failures available for read-only
+  reconciliation while preventing automatic or generic confirmed retries from
+  invoking the model again. Automatic reconciliation skips those sticky rows
+  before applying its batch limit, and linked durable jobs exclusively own
+  retry scheduling for their memory tasks.
+- Removed MemHub's task-scoped HyPermission store wrappers and global
+  model-content consent switch. The trusted embedded Hy Memory SDK now uses
+  its native local stores directly.
+- Existing installations ignore the retired config flag, readmit previously
+  parked pre-invocation work, and drop only the obsolete
+  `hy_node_permissions` proof table. Evidence, governed memories, model
+  profiles, SDK stores, and source/host grants are preserved.
+
+## Hy Memory 2.0.6 — 2026-10-27
+
+- Published
+  [`hy-memory==2.0.6`](https://pypi.org/project/hy-memory/2.0.6/).
+- `search(..., rerank=False)` skips every reranking stage for that one
+  call (the local cross-encoder, the memory-graph rerank, and the BM25 rerank of
+  working/coding facts), so results come back in retrieval-score order. The
+  flag travels on the read request, so concurrent searches do not affect each
+  other. The default (`True`) keeps current behavior; `/api/v1/search` accepts
+  it too.
+- MMR diversity selection now runs only when the requested `limit` is
+  below 20 (`MMR_MAX_TOP_K`) and is skipped for larger limits, where its
+  quadratic token-overlap pass is slow and gains little. Previously it ran for
+  every limit. `MEMORY_HYBRID_V2_MMR_LAMBDA` still sets the weight.
+  Near-duplicate demotion follows the same gate: it compares each candidate
+  with every earlier non-duplicate, which is quadratic in a pool that grows
+  with `limit`, and large result pages should show similar items anyway.
+- `search(scene="all")` searches chat, working and coding at the same time
+  and merges them with Reciprocal Rank Fusion (k=60) over each scene's own
+  ranking, so no scene is crowded out by the other scenes' score scales. Each
+  result says which scene and kind it came from. A scene that fails or is turned
+  off is reported in `errors` and does not fail the search. Source filters apply
+  to all three.
+- `list_memories(scene="all")` lists chat, working and coding memories
+  together, newest update first, with `limit`/`offset` paging over the merged
+  order and `total` summing all three. `scene="coding"` and `"working"` now
+  honor `include_sources` / `exclude_sources`, and `total` counts only the
+  filtered memories.
+- `update()` and `delete()` now work on working/coding experiences and
+  facts, not just chat memories. They take the same optional `kind` / `variant`
+  as `get()`, and an ID found in more than one place raises
+  `AmbiguousMemoryIdError` instead of guessing. Updating an experience replaces
+  its solution and keeps its task, unless a new `task=` is also passed; updating
+  a fact replaces its statement. Deleting removes the stored row and its vectors.
+- `list_memories(scene="chat" | "coding" | "working")` lists one kind of
+  memory. `coding` and `working` list that scene's whole store, newest update
+  first (`workspace_id` still narrows it), and `total` is now the real count
+  rather than the page size. The new `memory_stats()` returns the count and the
+  latest created/updated time for each scene and their total, without reading
+  any memory text. Previously coding and working memories could not be listed
+  or counted without a workspace, so a library of 45 memories listed 22.
+- Reconcile keeps time order. Both reconcile prompts now state that only
+  `memory_at` decides which memory is newer, that an older memory must never
+  supersede or overwrite a newer one, and that an unknown time is not newer;
+  the Chinese prompt now lists `memory_at` among the new-memory fields. The
+  reconciler also enforces this in code: a SUPERSEDE or UPDATE aimed at a memory
+  with a later `memory_at` becomes an ADD, so the newer memory stays current.
+- The HTTP server returns 409 `ambiguous_memory_id` and 400 for a bad
+  `kind`/`variant` instead of 500, passes `kind`/`variant` (and `task` on PUT)
+  to get/update/delete, and no longer reads a PUT/DELETE query string as part
+  of the memory id.
+- `memory_stats()` reads working/coding times, which are stored as naive UTC,
+  as UTC rather than local time.
+
+## Hy Memory 2.0.5 — 2026-10-22
+
+- Published
+  [`hy-memory==2.0.5`](https://pypi.org/project/hy-memory/2.0.5/).
+- Extended `get()` across Chat, Coding Experience, Working Fact, and raw IDs;
+  the returned `layer`, `kind`, and `variant` identify the actual record.
+- Added explicit raw retrieval through the existing `search()` and
+  `list_memories()` contracts, including user, agent, session, and source
+  filters.
+- Added public embedding-service injection and corrected extraction language
+  detection for structured OpenAI-style messages.
+- Added lifecycle status filters, durable idempotent write outcomes, and
+  quiesced local data-plane snapshot/restore.
+- Kept legacy migration mechanics out of the public Client API; separate
+  migration-only raw/import methods from development previews were removed.
+- Physical delete now reports partial VDB/Graph cleanup as failure.
+- Coding search no longer injects the version-evolution diff block by default;
+  opt in with `MEMORY_CODING_CHAIN_BLOCK_ENABLED=true`.
+- Serialized SDK SQLite cache/history connection lifecycle and tightened
+  Coding reconciliation to its agent/session/workspace authority.
+- Clarified persistent caller context through `meta`, while `metadata` remains
+  legacy processing configuration.
+
+## Hy Memory 2.0.5.dev4 — 2026-10-21
+
+- Development preview used for SDK-authority migration validation; not the
+  final `2.0.5` release.
+- Published as
+  [`hy-memory==2.0.5.dev4`](https://pypi.org/project/hy-memory/2.0.5.dev4/).
+
 ## Hy Memory 2.0.5.dev3 — 2026-10-21

 - Development preview; not the final `2.0.5` release.
@@ -28,6 +178,12 @@ Hy Memory and MemHub are versioned and released independently.
  APIs, offline local data-plane snapshot/restore, and fail-closed physical
  delete receipts.

+## Hy Memory 2.0.4 — 2026-10-17
+
+- Published
+  [`hy-memory==2.0.4`](https://pypi.org/project/hy-memory/2.0.4/) as a
+  maintenance release.
+
 ## Hy Memory 2.0.3 — 2026-10-17

 - Added an optional runtime HTTP transport factory so embedded products can
@@ -38,58 +194,6 @@ Hy Memory and MemHub are versioned and released independently.
  unchanged memory extraction, reconciliation, storage, and Coding/Working
  behavior.

-## Unreleased
-
- Added a MemHub-managed native Hermes MemoryProvider with automatic prefetch,
-  persistent per-session memory-ID deduplication, authorized Broker reads, and
-  recall-context stripping to prevent memory recursion.
- Added an extensible MemHub interface-localization layer with Simplified
-  Chinese, English, and Spanish catalogs shared by Web and Electron-native
-  surfaces. Interface language remains independent from memory content and
-  model processing language.
- `search(..., layers=["l1_raw"])` now explicitly searches SDK raw records in
-  pro/ultra mode and keeps existing user/agent/session/source filters; default
-  search continues to hide raw.
- Added `import_memory_exact` for idempotent migration of an existing semantic
-  memory with its canonical ID, lifecycle and supersession links, without
-  extraction or reconciliation.
- Added SDK-authoritative lifecycle listing with explicit status filters,
-  idempotent durable write outcomes, exact raw import/get/list APIs, and
-  quiesced local data-plane snapshot/restore contracts for the MemHub
-  authority migration.
- Physical delete now reports incomplete VDB/Graph cleanup as a partial failure
-  instead of claiming success after only one store was changed.
- Coding search no longer injects the version-evolution diff block
-  (`chain_block`) by default. Enable with
-  `MEMORY_CODING_CHAIN_BLOCK_ENABLED=true` or
-  `CodingConfig.chain_block_enabled=True` (write-side observe unchanged).
- Fixed governed-memory revisions and repeated evidence links failing when the
-  visibility dirty-marker trigger encounters the same memory twice. Startup
-  now replaces the affected legacy trigger without changing schema versions
-  or governed and visibility data.
- Serialized each SDK SQLite store's shared connection lifecycle so concurrent
-  history/cache operations cannot race with `close()` and crash the process.
- Scoped Coding reconciliation targets to the current HY authority,
-  conversation, and workspace, preventing near-duplicate sessions from
-  mutating an experience or working fact owned by another evidence context.
- Kept post-write authority verification failures available for read-only
-  reconciliation while preventing automatic or generic confirmed retries from
-  invoking the model again. Automatic reconciliation skips those sticky rows
-  before applying its batch limit, and linked durable jobs exclusively own
-  retry scheduling for their memory tasks.
- Clarified that `add(metadata=...)` carries legacy SDK processing options,
-  while persistent caller business context belongs in `meta`.
- Added an opt-in `get(..., include_raw=True)` path for reading `L1_RAW`
-  memories, including memories written in lite mode. The default remains
-  hidden for backward compatibility.
- Removed MemHub's task-scoped HyPermission store wrappers and global
-  model-content consent switch. The trusted embedded Hy Memory SDK now uses
-  its native local stores directly.
- Existing installations ignore the retired config flag, readmit previously
-  parked pre-invocation work, and drop only the obsolete
-  `hy_node_permissions` proof table. Evidence, governed memories, model
-  profiles, SDK stores, and source/host grants are preserved.
-
 ## Hy Memory 2.0.2 — 2026-10-17

 - Published
@@ -104,7 +208,17 @@ Hy Memory and MemHub are versioned and released independently.
  Because PyPI release metadata is immutable, the next SDK patch release must
  carry the updated license metadata; `2.0.1` retains its published metadata.

-## MemHub 1.0.3 — Unreleased
+## MemHub 1.0.5 — Unreleased
+
+- Updated the source version to `1.0.5` (build 14).
+- Pinned the public SDK to the stable `hy-memory==2.0.5` release.
+
+## MemHub 1.0.4 — 2026-10-18
+
+- Updated the source version to `1.0.4` (build 13).
+- Published the `v1.0.4` source and release metadata.
+
+## MemHub 1.0.3 — 2026-10-17

 - Updated the source version to `1.0.3` (build 7).
 - Pinned the public SDK to `hy-memory==2.0.3`.
@@ -273,8 +273,8 @@ catalog 也可以在没有基础 L0 时形成只读的合成画像。

 ```python
 memory = client.get("memory_id")
-# lite 模式写入的是 L1_RAW；按 ID 读取时必须显式选择返回原文。
-raw_memory = client.get("memory_id", include_raw=True)
+# raw ID 返回 L1_RAW，fact ID 返回对应事实层。
+raw_memory = client.get("raw_id")
 result = client.update("memory_id", "更新后的内容")
 result = client.delete("memory_id")
 result = client.delete_all(user_id="user_001")
@@ -284,7 +284,7 @@ page = client.list_memories(
    limit=100,
    offset=0,
    order="desc",
-    layers=None,
+    layers=["l1_raw"],  # 显式列出 raw；不传仍沿用默认语义层列表
    with_evolution=False,
 )

@@ -294,9 +294,10 @@ result = client.clone_user("user_001", "user_002", agent_id=None)
 对应异步方法为 `async_get()`、`async_update()`、`async_delete()`、
 `async_delete_all()`、`async_list_memories()`、`async_clone_user()`。

-`get()` / `async_get()` 默认继续隐藏 `L1_RAW`，因此不会改变现有调用方的原文访问
-边界。只有显式传入 `include_raw=True` 才返回该层；这也是 lite 模式按 ID 读取
-写入内容及其 `meta` 的方式。MemHub 的 HTTP 读取接口不开放此参数。
+`get()` / `async_get()` 是统一的 ID 精确读取：返回结果的 `layer`、`kind` 和
+`variant` 表明实际类型。授权由调用 SDK 前的产品边界完成，不由 SDK 根据 layer
+猜测。`list_memories()` 默认仍不包含 `L1_RAW`；显式传
+`layers=["l1_raw"]` 时按同一列表合同返回 raw。

 普通 `list_memories()` 返回 `vdb` 桶，并在 Graph 有结果时附带 `graph`；传入
 `workspace_id` 时返回生产力记忆的 `scene="productivity"` 和 `experiences`。
@@ -65,6 +65,12 @@ from .snapshot import (

 # ====== 配置 ======
 from .config import MemoryConfig
+from .embedding import (
+    EmbeddingDimensionMismatchError,
+    EmbeddingService,
+    EmbeddingServiceProtocol,
+)
+from .retrieval import AmbiguousMemoryIdError, MemoryKind, MemoryVariant

 # ====== Embedding 端点覆盖（per-call，不改 model/dims，只换接入端点）======
 from .core.embed_service import embed_endpoint_override
@@ -125,6 +131,13 @@ __all__ = [
    "snapshot_local_data_plane",
    # 配置
    "MemoryConfig",
+    # 公共扩展协议 / 精确读取类型
+    "EmbeddingService",
+    "EmbeddingServiceProtocol",
+    "EmbeddingDimensionMismatchError",
+    "MemoryKind",
+    "MemoryVariant",
+    "AmbiguousMemoryIdError",
    # Embedding 端点覆盖
    "embed_endpoint_override",
    # 高级 API
@@ -743,6 +743,7 @@ RECONCILE_PROMPT_ZH = """你是一个记忆整合系统（Memory Reconciler）

 * content
 * owner
+* memory_at（发生时间）
 * tags

 new memories 没有 ID。
@@ -791,6 +792,7 @@ new memories 没有 ID。
 * 无损合并
 * 保留所有有效信息
 * 不要产生重复 memory
+* 两者在某个值上不一致时，保留 memory_at 更晚的那个值；绝不用更早的 memory 覆盖更新的值

 ### 同一次事件的补充 vs 另一次发生

@@ -823,6 +825,7 @@ new memories 没有 ID。
 * 新 memory 与旧 memory 在同一维度冲突
 * 两者不能同时为真
 * 旧 memory 已不再代表当前状态
+* 新 memory 的 memory_at 不早于旧 memory 的 memory_at

 例如：

@@ -837,6 +840,7 @@ new memories 没有 ID。
 * wording 差异
 * 同义表达
 * 可共存偏好
+* 新 memory 的 memory_at 早于旧 memory：它是更早的状态，绝不能取代更新的状态（见「时间规则」）

 要求：

@@ -859,7 +863,7 @@ new memories 没有 ID。

 ### 把多条旧记忆折叠进同一条链

-如果**两条或更多**已有记忆描述的是**同一维度**，而新记忆是该维度上的最新变化，
+如果**两条或更多**已有记忆描述的是**同一维度**，而新记忆是该维度上的最新变化（它的 memory_at 晚于所有被折叠的记忆），
 你可以把它们折叠进同一条演化链。此时用 `memory_ids`（有序列表，**旧→新**）
 代替单个 `memory_id`：

@@ -951,13 +955,16 @@ SUPERSEDE 不需要复制旧内容，因为旧节点仍存在。

 ## 时间规则

-`memory_at: null`
+每条 memory（新的和已有的）都带 memory_at，表示它发生的时间。记忆**不一定按发生顺序写入**：
+历史可能晚导入，导入还没处理完时用户也可能手动添加记忆。所以一条 new memory 可能**比已有记忆更早**。

-表示时间未知。
-
-不要编造时间顺序。
-
-不要仅因为“看起来更新”就 supersede。
+* 新旧**只看 memory_at**。"New memories" 只表示"正在写入"，不表示"更近"。
+* 更早的 memory 绝不能取代更新的 memory：不能 SUPERSEDE 一条更新的记忆，也不能用 UPDATE 拿更早的值覆盖更新的值。
+* 如果 new memory 比与它冲突的已有记忆更早：保留已有记忆作为当前状态；把 new memory 以"过去状态"ADD
+  （用过去时，并带上时间，例如"2023 年用户住在上海"），没有新信息就不输出。
+* `memory_at: null` 表示时间未知。未知**不算更新**：不要用 memory_at 为 null 的 memory 去 SUPERSEDE 一条有时间的记忆。
+* 不要编造时间顺序。
+* 不要仅因为“看起来更新”就 supersede。

 ---

@@ -221,6 +221,8 @@ Rules:
 * merge losslessly
 * preserve all useful information
 * do NOT create duplicates
+* if the two disagree on a value, keep the value from the memory with the later
+  `memory_at`; never let an older memory overwrite a newer value

 ### Same occurrence vs. another occurrence

@@ -255,6 +257,7 @@ Use ONLY when:
 * the new memory contradicts an existing memory
 * both cannot be true at the same time
 * the old memory would now give the wrong current-state answer
+* the new memory's `memory_at` is the same as or LATER than the existing memory's `memory_at`

 Examples:

@@ -268,6 +271,8 @@ Do NOT use for:
 * compatible preferences
 * wording differences
 * semantic equivalence
+* a new memory whose `memory_at` is EARLIER than the existing memory's — it is an
+  older state and must never replace a newer one (see "Time order")

 Rules:

@@ -289,7 +294,8 @@ Rules:
 ### Folding several existing memories into one chain

 If TWO OR MORE existing memories track the SAME dimension and the new memory is
-the latest change on that dimension, you may fold them into a single evolution
+the latest change on that dimension — its `memory_at` is later than every folded
+memory's `memory_at` — you may fold them into a single evolution
 chain. Use `memory_ids` (an ordered list, OLDEST → NEWEST) instead of a single
 `memory_id`:

@@ -318,8 +324,7 @@ older memories' text into it; they remain queryable in the chain history.
 * Separate dated occurrences stay separate ADDs; use `same_series` only for the
  same subject and same concrete activity/event type.
 * Keep each memory focused on one coherent topic.
-* Do not invent chronology.
-* `memory_at: null` means unknown time.
+* Follow "Time order" below.
 * **NEVER merge or supersede across different subjects.** The subject is whoever
  the content is about: "The user ..." vs "John ..." vs "Mary ...". Memories about
  different people are independent facts even when the sentence is otherwise
@@ -332,6 +337,27 @@ older memories' text into it; they remain queryable in the chain history.

 ---

+# Time order
+
+Every memory, new or existing, has `memory_at`: when it happened. Memories are
+NOT always written in the order they happened — history can be imported late,
+and a memory can be added while older history is still being processed. A new
+memory can therefore be OLDER than an existing one.
+
+* Decide which memory is newer ONLY by `memory_at`. "New memories" means "being
+  written now", not "more recent".
+* An older memory must never replace a newer one: no SUPERSEDE of a newer
+  memory, and no UPDATE that overwrites a newer value with an older one.
+* If a new memory is older than an existing memory it conflicts with, keep the
+  existing memory as the current state. ADD the new one as a past state, worded
+  in the past with its time (for example "In 2023 the user lived in Shanghai"),
+  or output nothing if it adds no information.
+* `memory_at: null` means unknown time. Unknown is NOT newer: never SUPERSEDE a
+  memory that has a known `memory_at` with one whose `memory_at` is null.
+* Do not invent chronology.
+
+---
+
 # Optional: `related`

 On any op you MAY add a `related` array recording how the memory being written
@@ -441,6 +467,81 @@ def append_reconcile_structure_prompt(
    return prompt + RECONCILE_STRUCTURE_METADATA_PROMPT_EN


+def _as_aware(value: Any) -> Optional[datetime]:
+    """memory_at as a comparable aware datetime; None when unknown or unreadable."""
+    if isinstance(value, str) and value:
+        try:
+            value = datetime.fromisoformat(value)
+        except (TypeError, ValueError):
+            return None
+    if not isinstance(value, datetime):
+        return None
+    # Stored nodes rebuild memory_at from epoch seconds as naive local time
+    # (MemoryNode.from_dict), so a naive value is local, not UTC.
+    return value if value.tzinfo is not None else value.astimezone()
+
+
+def enforce_time_order(
+    ops: List["ReconcileOp"],
+    new_memory_at: Any,
+    existing: Dict[str, Any],
+) -> List["ReconcileOp"]:
+    """An older memory must never replace a newer one, whatever the model says.
+
+    Memories are not always written in the order they happened: history is
+    imported late, and a person can add a memory while older history is still
+    being processed. The prompt tells the model to compare `memory_at`; this
+    makes it hold. A SUPERSEDE or UPDATE aimed at an existing memory whose
+    `memory_at` is later than the memory being written is turned into an ADD,
+    so the newer memory stays the current state. Unknown times are not
+    compared. A content-less UPDATE (the legacy DELETE mapping) is dropped.
+    """
+    written_at = _as_aware(new_memory_at)
+    if written_at is None:
+        return ops
+    kept: List[ReconcileOp] = []
+    for op in ops:
+        if op.op not in ("SUPERSEDE", "UPDATE"):
+            kept.append(op)
+            continue
+        targets = list(op.memory_ids) if op.memory_ids else [op.memory_id]
+        newer = [
+            target for target in targets
+            if target and (node := existing.get(target)) is not None
+            and (target_at := _as_aware(getattr(node, "memory_at", None))) is not None
+            and target_at > written_at
+        ]
+        if not newer:
+            kept.append(op)
+            continue
+        logger.warning(
+            "[reconciler] %s blocked: memory_at %s is older than existing %s; "
+            "kept the newer memory",
+            op.op, written_at.isoformat(), ",".join(newer),
+        )
+        if not op.content:
+            continue
+        kept.append(ReconcileOp(
+            op="ADD",
+            content=op.content,
+            layer=op.layer,
+            owner=op.owner,
+            tags=list(op.tags),
+            tag_ners=list(op.tag_ners),
+            topic_paths=list(op.topic_paths),
+            event_facet=dict(op.event_facet),
+            topic_assignments=list(op.topic_assignments),
+            structure_references=list(op.structure_references),
+            topic_paths_provided=op.topic_paths_provided,
+            event_facet_provided=op.event_facet_provided,
+            topic_assignments_provided=op.topic_assignments_provided,
+            speculate=op.speculate,
+            related=list(op.related),
+            reason=f"time_order: older than {','.join(newer)}; {op.reason}"[:500],
+        ))
+    return kept
+
+
 def build_reconcile_prompt(
    *,
    new_memories: List[str],
@@ -1399,6 +1500,7 @@ class MemoryReconciler:
                _tok_total += response.tokens_used
                raw_response = response.content
                ops = self._parse_ops(raw_response)
+                ops = enforce_time_order(ops, current_time, candidate_map)
                ops = self._inherit_exact_input_structure(
                    ops, new_memories_with_meta,
                )
@@ -217,6 +217,21 @@ def _in_clause(column: str, values: Optional[List[str]]) -> Tuple[str, List[Any]
    return f"{column} IN ({placeholders})", list(values)


+def _source_where(
+    include_sources: Optional[List[str]], exclude_sources: Optional[List[str]],
+) -> Tuple[str, List[Any]]:
+    """Caller-source filter as SQL (``1`` when unfiltered); exclusion wins."""
+    clauses: List[str] = []
+    params: List[Any] = []
+    if include_sources:
+        clauses.append(f"source IN ({','.join('?' for _ in include_sources)})")
+        params.extend(include_sources)
+    if exclude_sources:
+        clauses.append(f"source NOT IN ({','.join('?' for _ in exclude_sources)})")
+        params.extend(exclude_sources)
+    return (" AND ".join(clauses) or "1"), params
+
+
 def _scope_matches(
    row_workspace_id: Optional[str],
    row_branch: Optional[str],
@@ -416,6 +431,36 @@ class CodingMemoryStore:

        self._initialized = False

+    @staticmethod
+    def database_path(
+        config: MemoryConfig,
+        db_path: Optional[str] = None,
+        variant: str = "working",
+    ) -> str:
+        """Return the authoritative SQLite path without creating anything."""
+        if db_path is None:
+            data_dir = (
+                os.path.dirname(config.history.db_path)
+                if config.history.db_path else "./data"
+            )
+            db_path = os.path.join(data_dir, "coding_memory.db")
+        normalized = (
+            "coding"
+            if str(variant or "").strip().lower() == "coding"
+            else "working"
+        )
+        return _variant_db_path(db_path, "_cv" if normalized == "coding" else "")
+
+    @classmethod
+    def database_exists(
+        cls,
+        config: MemoryConfig,
+        db_path: Optional[str] = None,
+        variant: str = "working",
+    ) -> bool:
+        """Whether a coding namespace has an existing SQLite database."""
+        return os.path.isfile(cls.database_path(config, db_path, variant))
+
    async def initialize(self) -> None:
        if self._initialized:
            return
@@ -573,6 +618,86 @@ class CodingMemoryStore:
        logger.info(f"[coding-store] deleted experience_id={experience_id}")
        return True

+    async def edit_experience(
+        self,
+        experience_id: str,
+        *,
+        user_id: Optional[str] = None,
+        task: Optional[str] = None,
+        solution: Optional[str] = None,
+        search_keys: Optional[List[str]] = None,
+    ) -> Optional[CodingMemory]:
+        """调用方手动改一条经验（client.update 用）。返回改后的经验；不存在返回 None。
+
+        与 update() 的区别：update 是写入链路的覆盖式重写，带写授权校验（同
+        agent/session/workspace 才能改）；这里是用户对已有条目的直接编辑，只动
+        给定的字段，其余原样保留，因此不做授权校验。
+
+        向量行只嵌 task + search_keys（solution 不进向量），所以只改 solution 时
+        不重嵌、不碰向量库；改了 task / search_keys 才删旧 key 行重写。
+        updated_at 总是刷新。
+        """
+        prev = await self.get_by_id(experience_id, user_id=user_id)
+        if prev is None:
+            return None
+        updated = deepcopy(prev)
+        if task is not None:
+            updated.task = task
+        if solution is not None:
+            updated.solution = solution
+        if search_keys is not None:
+            updated.search_keys = list(search_keys)
+        updated.updated_at = datetime.utcnow()
+        await _run_in_pool(self._upsert_sql_sync, updated)
+        if updated.task != prev.task or updated.search_keys != prev.search_keys:
+            await self._delete_vec_keys_by_count(experience_id, 1 + len(prev.search_keys))
+            await self._upsert_vec_keys(updated)
+        logger.info(f"[coding-store] edited experience_id={experience_id}")
+        return updated
+
+    async def edit_fact(
+        self,
+        fact_id: str,
+        *,
+        statement: str,
+        user_id: Optional[str] = None,
+    ) -> Optional[WorkingFact]:
+        """调用方手动改一条事实的 statement。返回改后的事实；不存在返回 None。
+
+        statement 是事实层的检索文本，所以 active 事实会重嵌并覆盖它那一行向量
+        （node_id 就是 fact_id，upsert 即覆盖）。superseded 事实的向量行早已撤掉，
+        这里不把它加回去，否则会让过期结论重新占召回名额。
+        """
+        prev = await self.get_fact(fact_id, user_id=user_id)
+        if prev is None:
+            return None
+        updated = deepcopy(prev)
+        updated.statement = statement
+        await _run_in_pool(self._upsert_fact_sql_sync, updated)
+        if updated.status == "active":
+            await self._upsert_fact_vec(updated)
+        logger.info(f"[coding-store] edited fact_id={fact_id}")
+        return updated
+
+    async def delete_fact(self, fact_id: str, user_id: Optional[str] = None) -> bool:
+        """物理删除一条事实（SQLite 行 + 事实向量行），返回是否真有删到。
+
+        别处留下的引用（经验的 source_fact_ids、后继事实的 fact_link）不清：
+        召回侧本来就按存在性过滤死链（get_many / get_facts 查不到自然掉队）。
+        """
+        prev = await self.get_fact(fact_id, user_id=user_id)
+        if prev is None:
+            logger.info(f"[coding-store] delete fact skipped (not found): {fact_id}")
+            return False
+        if self._fact_vec is not None:
+            try:
+                await self._fact_vec.delete(fact_id)
+            except Exception as e:
+                logger.debug(f"[coding-store] delete fact vec {fact_id} ignored: {e}")
+        await _run_in_pool(self._delete_fact_sql_sync, fact_id, prev.user_id)
+        logger.info(f"[coding-store] deleted fact_id={fact_id}")
+        return True
+
    # ------------------------------------------------------------
    # 读取
    # ------------------------------------------------------------
@@ -657,29 +782,84 @@ class CodingMemoryStore:
        branch: Optional[str] = None,
        limit: int = 100,
        offset: int = 0,
+        include_sources: Optional[List[str]] = None,
+        exclude_sources: Optional[List[str]] = None,
    ) -> List["CodingMemory"]:
        """按 workspace_id（+可选 branch）纯过滤列出该 user 的完整 coding memory。

        作用域规则见 _scope_where（与 search 路径共用同一套）。无语义排序，按
-        updated_at 倒序（最近修改优先）。
+        updated_at 倒序（最近修改优先）。include_sources / exclude_sources 按
+        source 列在 SQL 里过滤（排除优先），所以 LIMIT/OFFSET 在过滤之后生效。
        """
        return await _run_in_pool(
-            self._list_by_workspace_sync, user_id, workspace_id, branch, limit, offset
+            self._list_by_workspace_sync, user_id, workspace_id, branch, limit, offset,
+            include_sources, exclude_sources,
        )

    def _list_by_workspace_sync(
        self, user_id: str, workspace_id: str, branch: Optional[str],
        limit: int, offset: int,
+        include_sources: Optional[List[str]] = None,
+        exclude_sources: Optional[List[str]] = None,
    ) -> List["CodingMemory"]:
        scope_sql, scope_params = _scope_where(workspace_id, branch)
+        source_sql, source_params = _source_where(include_sources, exclude_sources)
        with self._lock:
            assert self._conn is not None
            sql = (f"SELECT * FROM coding_memory_meta WHERE user_id = ? AND {scope_sql} "
-                   f"ORDER BY updated_at DESC LIMIT ? OFFSET ?")
-            params: List[Any] = [user_id] + scope_params + [limit, offset]
+                   f"AND {source_sql} "
+                   f"ORDER BY updated_at DESC, experience_id ASC LIMIT ? OFFSET ?")
+            params: List[Any] = ([user_id] + scope_params + source_params
+                                 + [limit, offset])
            cur = self._conn.execute(sql, params)
            return [_row_to_memory(r) for r in cur.fetchall()]

+    async def stats(
+        self,
+        user_id: str,
+        workspace_id: Optional[str] = None,
+        branch: Optional[str] = None,
+        include_sources: Optional[List[str]] = None,
+        exclude_sources: Optional[List[str]] = None,
+    ) -> Dict[str, Any]:
+        """经验数、最新创建/修改时间、active 事实数。纯聚合，不读正文。
+
+        作用域与 list_by_workspace 相同：workspace_id 为空表示该 user 全部；
+        来源过滤同样与 list_by_workspace 一致，count 因此恒等于分页可翻到的总数。
+        """
+        return await _run_in_pool(
+            self._stats_sync, user_id, workspace_id, branch,
+            include_sources, exclude_sources,
+        )
+
+    def _stats_sync(
+        self, user_id: str, workspace_id: Optional[str], branch: Optional[str],
+        include_sources: Optional[List[str]] = None,
+        exclude_sources: Optional[List[str]] = None,
+    ) -> Dict[str, Any]:
+        scope_sql, scope_params = _scope_where(workspace_id, branch)
+        source_sql, source_params = _source_where(include_sources, exclude_sources)
+        with self._lock:
+            assert self._conn is not None
+            row = self._conn.execute(
+                f"SELECT COUNT(*) AS n, MAX(created_at) AS created, "
+                f"MAX(updated_at) AS updated FROM coding_memory_meta "
+                f"WHERE user_id = ? AND {scope_sql} AND {source_sql}",
+                [user_id] + scope_params + source_params,
+            ).fetchone()
+            facts = self._conn.execute(
+                f"SELECT COUNT(*) AS n FROM working_fact "
+                f"WHERE user_id = ? AND status = 'active' AND {scope_sql} "
+                f"AND {source_sql}",
+                [user_id] + scope_params + source_params,
+            ).fetchone()
+        return {
+            "count": int(row["n"] or 0),
+            "latest_created_at": _parse_dt(row["created"]) if row["created"] else None,
+            "latest_updated_at": _parse_dt(row["updated"]) if row["updated"] else None,
+            "facts": int(facts["n"] or 0),
+        }
+
    # ------------------------------------------------------------
    # 召回
    # ------------------------------------------------------------
@@ -1058,6 +1238,14 @@ class CodingMemoryStore:
                (experience_id, user_id),
            )

+    def _delete_fact_sql_sync(self, fact_id: str, user_id: str) -> None:
+        with self._lock:
+            assert self._conn is not None
+            self._conn.execute(
+                "DELETE FROM working_fact WHERE fact_id = ? AND user_id = ?",
+                (fact_id, user_id),
+            )
+
    def _get_by_id_sync(
        self, experience_id: str, user_id: Optional[str]
    ) -> Optional[CodingMemory]:
@@ -50,6 +50,7 @@ class CodingWriter:
        embed_service: EmbedService,
        fact_reconciler: Optional[WorkingFactReconciler] = None,
        cache: Any = None,
+        variant: str = "working",
    ):
        self.store = store
        self.extractor = extractor
@@ -59,6 +60,11 @@ class CodingWriter:
        self.fact_reconciler = fact_reconciler or WorkingFactReconciler(
            store, embed_service, llm_provider
        )
+        self.variant = (
+            "coding"
+            if str(variant or "").strip().lower() == "coding"
+            else "working"
+        )
        # cache 仅用于把 request_id 串给各阶段（各阶段自己打点）；writer 不集中打点。
        self.cache = cache

@@ -217,6 +223,7 @@ class CodingWriter:
        return {
            "success": True,
            "scene": "productivity",
+            "variant": self.variant,
            "request_id": request_id,
            "elapsed_ms": round((time.perf_counter() - t0) * 1000, 2),
            "ops": [op.to_dict() for op in ops],
@@ -278,11 +285,11 @@ class CodingWriter:
                        f"[coding-write] crosslink fact {f.fact_id} failed: {e}"
                    )

-    @staticmethod
-    def _empty_response(request_id: str, t0: float) -> Dict[str, Any]:
+    def _empty_response(self, request_id: str, t0: float) -> Dict[str, Any]:
        return {
            "success": True,
            "scene": "productivity",
+            "variant": self.variant,
            "request_id": request_id,
            "elapsed_ms": round((time.perf_counter() - t0) * 1000, 2),
            "ops": [],
@@ -0,0 +1,53 @@
+"""Public structural contract for injected embedding services."""
+
+from typing import Any, Dict, List, Protocol, Tuple, runtime_checkable
+
+
+@runtime_checkable
+class EmbeddingService(Protocol):
+    """Embedding service accepted by :class:`HyMemoryClient`.
+
+    Implementations are borrowed by the client: their lifecycle remains owned
+    by the caller and ``HyMemoryClient.close()`` will not close them.
+    """
+
+    async def embed(self, text: str, use_cache: bool = True) -> List[float]:
+        ...
+
+    async def embed_with_meta(
+        self, text: str, use_cache: bool = True
+    ) -> Tuple[List[float], Dict[str, Any]]:
+        ...
+
+    async def embed_batch(
+        self, texts: List[str], use_cache: bool = True
+    ) -> List[List[float]]:
+        ...
+
+    async def embed_queued(
+        self, text: str, use_cache: bool = True
+    ) -> List[float]:
+        ...
+
+    async def embed_queued_with_meta(
+        self, text: str, use_cache: bool = True
+    ) -> Tuple[List[float], Dict[str, Any]]:
+        ...
+
+    def get_embedding_dims(self) -> int:
+        ...
+
+
+# Explicit alias for callers that prefer the conventional Protocol suffix.
+EmbeddingServiceProtocol = EmbeddingService
+
+
+class EmbeddingDimensionMismatchError(ValueError):
+    """Raised before initialization when service and store dimensions differ."""
+
+
+__all__ = [
+    "EmbeddingService",
+    "EmbeddingServiceProtocol",
+    "EmbeddingDimensionMismatchError",
+]
@@ -160,6 +160,28 @@ def near_dup_demote(
    return out


+# MMR 只在调用方请求的 top-k（search 的 limit）小于这个值时自动启用。
+# mmr_select 的贪心是 O(k^2 * pool) 次 Jaccard，k 大时代价陡增，而大 k 本身已
+# 足够容纳多样性，收益很小。权重仍由 MEMORY_HYBRID_V2_MMR_LAMBDA 控制。
+MMR_MAX_TOP_K = 20
+
+
+def mmr_applies(top_k: int, lambda_: float) -> bool:
+    """本次检索是否走 MMR：λ 在有效区间内，且 top_k < MMR_MAX_TOP_K。"""
+    return 0.0 < float(lambda_) < 0.999 and int(top_k) < MMR_MAX_TOP_K
+
+
+def near_dup_applies(top_k: int) -> bool:
+    """本次检索是否走近重复降权：与 MMR 同一个 top-k 门槛。
+
+    near_dup_demote 让每个候选和此前所有"非重复"候选两两算 Jaccard，候选
+    互不相似时是 O(pool^2)；池子随 limit 放大（约 max(3*limit, 30) + 关键词
+    命中），limit 达到 MMR_MAX_TOP_K 时代价同样陡增。大 k 的调用方（如记忆
+    面板翻页）要的是全部相关项，相似项也应照常出现，降权没有收益。
+    """
+    return int(top_k) < MMR_MAX_TOP_K
+
+
 def mmr_select(
    hits: Sequence[Dict[str, Any]],
    limit: int,
@@ -401,6 +401,11 @@ class ReadRequest:
    intention_limit: int = 0           # Intention 路的 topk，默认 0 关闭


+    # 重排开关（按次生效）。False 时跳过全部重排阶段：本地 cross-encoder
+    # (recall.local_rerank_enabled) 与记忆关系图重排 (_apply_graph_rerank)，
+    # 结果按召回分排序。走请求对象而非环境变量，并发请求互不影响。
+    rerank: bool = True
+
    # 时间过滤
    created_after: Optional[float] = None  # Unix timestamp (float)，只返回 gmt_created >= 此值的记忆

@@ -57,7 +57,9 @@ from ._retrieval.ner import (
    memory_matches_time_hints,
 )
 from ._retrieval.diversity import (
+    near_dup_applies,
    near_dup_demote,
+    mmr_applies,
    mmr_select,
    social_boilerplate_demote,
    pool_df_demote,
@@ -1279,7 +1281,10 @@ class BaseReadPipeline(ReadPipeline):
            SAME_SERIES_TIMELINE_ON = bool(
                getattr(self.config.recall, "same_series_timeline_enabled", False)
            )
-            LOCAL_RERANK_ON = bool(
+            # 按次的重排开关（search(rerank=False)）：关掉本地 cross-encoder 与图重排，
+            # 结果按召回分排序。只读请求对象，不碰共享状态，并发请求互不影响。
+            RERANK_ON = bool(getattr(request, "rerank", True))
+            LOCAL_RERANK_ON = RERANK_ON and bool(
                getattr(self.config.recall, "local_rerank_enabled", False)
            )
            LOCAL_RERANK_POOL_FACTOR = max(1, int(
@@ -1431,6 +1436,11 @@ class BaseReadPipeline(ReadPipeline):
            MIN_SCORE = _resolve_normal_min_score(request.min_score)

            final_limit = request.limit if request.limit > 0 else 10
+            # MMR 只在调用方的 top-k（final_limit，即 search 的 limit）< MMR_MAX_TOP_K
+            # 时启用；池子（LOCAL_RERANK 下是 final_limit 的若干倍）不参与判定。
+            MMR_ON = mmr_applies(final_limit, MMR_LAMBDA)
+            # 近重复降权同样是候选两两比较，用同一个 top-k 门槛。
+            NEAR_DUP_ON = near_dup_applies(final_limit)
            vdb_sem_limit = max(final_limit * 3, 30)
            vdb_kw_limit = max(final_limit * 5, 80)
            graph_limit = max(
@@ -2462,7 +2472,7 @@ class BaseReadPipeline(ReadPipeline):
            # 只把已选出的 top-N 换个顺序收益很小。
            relation_paths: List[Dict[str, Any]] = []
            _graph_stats: Dict[str, Any] = {}
-            if self._expansion_on():
+            if RERANK_ON and self._expansion_on():
                _t_graph = datetime.now()
                _pool_before = len(normal_results)
                normal_results, relation_paths, _graph_stats = await self._apply_graph_rerank(
@@ -2525,7 +2535,7 @@ class BaseReadPipeline(ReadPipeline):
                        df_thr=POOL_DF_THR,
                        penalty=POOL_DF_PENALTY,
                    )
-                if NEAR_DUP_THR > 0 and NEAR_DUP_PENALTY > 0:
+                if NEAR_DUP_ON and NEAR_DUP_THR > 0 and NEAR_DUP_PENALTY > 0:
                    normal_results = near_dup_demote(
                        normal_results,
                        jaccard_thr=NEAR_DUP_THR,
@@ -2538,7 +2548,7 @@ class BaseReadPipeline(ReadPipeline):
                    ))
                    if LOCAL_RERANK_ON else final_limit
                )
-                if 0.0 < MMR_LAMBDA < 0.999 and len(normal_results) > _pre_rerank_limit:
+                if MMR_ON and len(normal_results) > _pre_rerank_limit:
                    normal_results = mmr_select(
                        normal_results,
                        _pre_rerank_limit,
@@ -2715,7 +2725,7 @@ class BaseReadPipeline(ReadPipeline):
                        reverse=True,
                    )
                    if (
-                        0.0 < MMR_LAMBDA < 0.999
+                        MMR_ON
                        and len(multi_seed_pool) > _pre_rerank_limit
                    ):
                        multi_seed_pool = mmr_select(
@@ -3544,6 +3554,9 @@ class BaseReadPipeline(ReadPipeline):
                "episode_capsule_render_mode": EPISODE_CAPSULE_RENDER_MODE,
                "same_series_timeline_enabled": SAME_SERIES_TIMELINE_ON,
                "local_rerank_enabled": LOCAL_RERANK_ON,
+                "rerank": RERANK_ON,
+                "mmr_applied": MMR_ON,
+                "near_dup_applied": NEAR_DUP_ON,
                "local_rerank_model": getattr(
                    self.config.recall,
                    "local_rerank_model",
@@ -19,6 +19,7 @@ import logging
 from .base import WritePipeline, ReadPipeline
 from ..config import MemoryConfig
 from ..core.embed_service import EmbedService
+from ..embedding import EmbeddingService
 from ..data.vector_store import create_vector_store
 from ..data.vector_store_base import VectorStoreBase

@@ -50,7 +51,12 @@ class ComponentFactory:
        s2 = await factory.get_system2_writer()    # System2Writer (ultra mode)
    """

-    def __init__(self, config: Optional[MemoryConfig] = None):
+    def __init__(
+        self,
+        config: Optional[MemoryConfig] = None,
+        *,
+        embedding_service: Optional[EmbeddingService] = None,
+    ):
        self.config = config or MemoryConfig.from_env()

        # 实例缓存（单例）
@@ -62,7 +68,7 @@ class ComponentFactory:
        self._readers: Dict[str, ReadPipeline] = {}

        # 共享资源
-        self._shared_embed_service: Optional[EmbedService] = None
+        self._shared_embed_service: Optional[EmbeddingService] = embedding_service
        self._shared_vector_store: Optional[VectorStoreBase] = None
        self._shared_vector_store_initialized: bool = False
        self._shared_graph_store = None
@@ -72,7 +78,7 @@ class ComponentFactory:
    # 共享资源
    # ================================================================

-    def _get_shared_embed_service(self) -> EmbedService:
+    def _get_shared_embed_service(self) -> EmbeddingService:
        if self._shared_embed_service is None:
            self._shared_embed_service = EmbedService(self.config)
        return self._shared_embed_service
@@ -0,0 +1,23 @@
+"""Public types and errors for exact memory-ID lookup."""
+
+from typing import Literal, Sequence
+
+
+MemoryKind = Literal["chat", "experience", "fact"]
+MemoryVariant = Literal["working", "coding"]
+
+
+class AmbiguousMemoryIdError(LookupError):
+    """An ID exists in more than one public memory namespace."""
+
+    def __init__(self, memory_id: str, namespaces: Sequence[str]):
+        self.memory_id = memory_id
+        self.namespaces = tuple(namespaces)
+        joined = ", ".join(self.namespaces)
+        super().__init__(
+            f"memory ID {memory_id!r} is ambiguous across namespaces: {joined}; "
+            "pass kind and/or variant to disambiguate"
+        )
+
+
+__all__ = ["MemoryKind", "MemoryVariant", "AmbiguousMemoryIdError"]
@@ -33,9 +33,28 @@ import threading
 from datetime import datetime
 from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
 from typing import Any, Dict, Optional
+from urllib.parse import parse_qs, urlparse
+
+from .retrieval import AmbiguousMemoryIdError

 logger = logging.getLogger("hy_memory.server")

+
+def _namespace_args(values: Dict[str, Any]) -> Dict[str, Any]:
+    """kind / variant from a query string (lists) or a JSON body (scalars).
+
+    They disambiguate an id that exists in more than one namespace (chat,
+    experience, fact; working or coding). Absent means "resolve it".
+    """
+    args: Dict[str, Any] = {}
+    for key in ("kind", "variant"):
+        value = values.get(key)
+        if isinstance(value, list):
+            value = value[0] if value else None
+        if value:
+            args[key] = str(value)
+    return args
+
 # Lazy-init global client
 _client = None
 _client_lock = None
@@ -148,7 +167,6 @@ class MemoryHTTPHandler(BaseHTTPRequestHandler):
            if path == "/api/v1/metrics":
                minutes = 5
                if "?" in self.path:
-                    from urllib.parse import parse_qs, urlparse
                    qs = parse_qs(urlparse(self.path).query)
                    minutes = int(qs.get("minutes", ["5"])[0])
                client = _get_client()
@@ -161,7 +179,8 @@ class MemoryHTTPHandler(BaseHTTPRequestHandler):
            if m:
                memory_id = m.group(1)
                client = _get_client()
-                result = client.get(memory_id)
+                query = parse_qs(urlparse(self.path).query)
+                result = client.get(memory_id, **_namespace_args(query))
                if result is None:
                    _json_response(self, 404, {"error": "not_found", "memory_id": memory_id})
                else:
@@ -170,6 +189,10 @@ class MemoryHTTPHandler(BaseHTTPRequestHandler):

            _json_response(self, 404, {"error": "not_found", "path": path})

+        except AmbiguousMemoryIdError as e:
+            _json_response(self, 409, {"error": "ambiguous_memory_id", "detail": str(e)})
+        except ValueError as e:
+            _json_response(self, 400, {"error": str(e)})
        except Exception as e:
            logger.error(f"[server] GET {path} error: {e}", exc_info=True)
            _json_response(self, 500, {"error": str(e)})
@@ -213,7 +236,7 @@ class MemoryHTTPHandler(BaseHTTPRequestHandler):
            _json_response(self, 500, {"error": str(e)})

    def do_PUT(self):
-        path = self.path.rstrip("/")
+        path = self.path.split("?")[0].rstrip("/")
        body = _read_json_body(self)
        if body is None:
            _json_response(self, 400, {"error": "invalid JSON body"})
@@ -228,29 +251,43 @@ class MemoryHTTPHandler(BaseHTTPRequestHandler):
                    _json_response(self, 400, {"error": "content is required"})
                    return
                client = _get_client()
-                result = client.update(memory_id, content)
+                result = client.update(
+                    memory_id,
+                    content,
+                    **_namespace_args(body),
+                    **({"task": body["task"]} if body.get("task") else {}),
+                )
                _json_response(self, 200, result)
                return

            _json_response(self, 404, {"error": "not_found", "path": path})

+        except AmbiguousMemoryIdError as e:
+            _json_response(self, 409, {"error": "ambiguous_memory_id", "detail": str(e)})
+        except ValueError as e:
+            _json_response(self, 400, {"error": str(e)})
        except Exception as e:
            logger.error(f"[server] PUT {path} error: {e}", exc_info=True)
            _json_response(self, 500, {"error": str(e)})

    def do_DELETE(self):
-        path = self.path.rstrip("/")
+        path = self.path.split("?")[0].rstrip("/")
        try:
            m = re.match(r"^/api/v1/memories/([^/]+)$", path)
            if m:
                memory_id = m.group(1)
                client = _get_client()
-                result = client.delete(memory_id)
+                query = parse_qs(urlparse(self.path).query)
+                result = client.delete(memory_id, **_namespace_args(query))
                _json_response(self, 200, result)
                return

            _json_response(self, 404, {"error": "not_found", "path": path})

+        except AmbiguousMemoryIdError as e:
+            _json_response(self, 409, {"error": "ambiguous_memory_id", "detail": str(e)})
+        except ValueError as e:
+            _json_response(self, 400, {"error": str(e)})
        except Exception as e:
            logger.error(f"[server] DELETE {path} error: {e}", exc_info=True)
            _json_response(self, 500, {"error": str(e)})
@@ -455,7 +492,8 @@ class MemoryHTTPHandler(BaseHTTPRequestHandler):
            if key in body:
                kwargs[key] = body[key]

-        for key in ("limit", "min_score", "schema_min_score", "schema_limit", "reader"):
+        for key in ("limit", "min_score", "schema_min_score", "schema_limit", "reader",
+                    "rerank"):
            if key in body:
                kwargs[key] = body[key]

@@ -8,6 +8,7 @@ Uses a two-tier strategy:
 This avoids langdetect's known weakness with short Chinese text.
 """

+import json
 import logging
 import re
 from typing import List, Optional
@@ -135,6 +136,30 @@ def detect_language(text: str) -> str:
    return "zh" if ratio > 0.1 else "en"


+def _spoken_text(value: str) -> str:
+    """Ignore text-block metadata when a message content is structured JSON."""
+    text = value.strip()
+    if text.startswith("["):
+        try:
+            blocks = json.loads(text)
+        except (ValueError, TypeError):
+            pass
+        else:
+            if (
+                isinstance(blocks, list)
+                and blocks
+                and all(
+                    isinstance(block, dict)
+                    and isinstance(block.get("type"), str)
+                    and block["type"] in {"text", "input_text", "output_text"}
+                    and isinstance(block.get("text"), str)
+                    for block in blocks
+                )
+            ):
+                return "\n".join(block["text"] for block in blocks)
+    return text
+
+
 def extract_content_for_detection(
    content: str = "",
    messages: Optional[List] = None,
@@ -153,20 +178,57 @@ def extract_content_for_detection(
        Concatenated content text suitable for language detection.
    """
    if content and content.strip():
-        return content.strip()
+        value = content.strip()
+        # Writer serializes message batches to JSON before calling Extractor.
+        # Detect the spoken text, not the leading ASCII keys (role/content),
+        # without changing the raw evidence sent to the model.
+        if value.startswith("["):
+            try:
+                decoded = json.loads(value)
+            except (ValueError, TypeError):
+                pass
+            else:
+                if (
+                    isinstance(decoded, list)
+                    and decoded
+                    and all(
+                        isinstance(item, dict)
+                        and isinstance(item.get("role"), str)
+                        and isinstance(item.get("content"), str)
+                        for item in decoded
+                    )
+                ):
+                    # In a mixed batch, the latest user turn is the clearest
+                    # instruction for the language; assistant/tool metadata
+                    # must not decide it first.
+                    user_text = [
+                        _spoken_text(item["content"])
+                        for item in decoded
+                        if item["role"] == "user" and item["content"].strip()
+                    ]
+                    if user_text:
+                        return user_text[-1]
+                    spoken = "\n".join(_spoken_text(item["content"]) for item in decoded)
+                    if spoken.strip():
+                        return spoken
+        return _spoken_text(value)

    if messages:
        parts = []
+        user_parts = []
        for msg in messages:
            if isinstance(msg, dict):
-                c = msg.get("content", "")
+                c, role = msg.get("content", ""), msg.get("role")
            elif hasattr(msg, "content"):
-                c = msg.content
+                c, role = msg.content, getattr(msg, "role", None)
            else:
                continue
-            if c and c.strip():
-                parts.append(c.strip())
-        return "\n".join(parts)
+            if isinstance(c, str) and c.strip():
+                text = _spoken_text(c)
+                parts.append(text)
+                if role == "user":
+                    user_parts.append(text)
+        return user_parts[-1] if user_parts else "\n".join(parts)

    return ""

@@ -9,10 +9,17 @@
 ## Architecture Principles

 1. Memory 读写与存储 backend **只经 SDK**（config / env）。
-2. MemHub 对 memory 的原子操作 **只经 SDK**；缺接口先补 SDK。
+2. MemHub 运行时对 memory 的原子操作 **只经 SDK 公共 Client**；缺正常业务
+   接口先补 SDK。Runtime writers 停止后的 App 启动迁移可直接操作底层存储，
+   但不得把迁移专用方法加入公共 Client。
 3. MemHub **不存 memory 实体**；**可以存 `memory_id`** 做组织/回执。原文进入 **SDK raw**。
+4. SDK raw 只保留滚动三天；一条 raw 对应一个完整 turn，只含真人 user 与
+   assistant，tool 不持久化。提取后的高层 memory 不受该窗口影响。
+5. `source_raw_ids` 是 SDK 内部溯源关系，不是产品 Evidence 合同；记忆详情不提供
+   “原始依据”或按 raw ID 展开原文的入口。

-生产入口仍建议经 `HyMemoryBridge`（共享 Runtime、模型配置、密钥、harness/model 门禁），但 bridge **不得**再把 SDK 结果镜像成第二套 memory 正文权威。
+生产入口通过共享 Runtime 租约取得公开 `HyMemoryClient`；MemHub 只在调用前执行
+模型配置、密钥和 harness/model 门禁，不维护第二套 memory 正文权威。

 来源授权、模型授权和宿主召回授权仍是 MemHub 管理面门禁；memory
 数据平面的权威在 SDK。
@@ -30,46 +37,15 @@ reader:       hybrid

 ## MemHub 运行时边界

-MemHub 对 SDK 使用公开、可复现的本地策略（经 bridge / `hy_runtime_config` 注入）：
+MemHub 通过公开 `HyMemoryClient` 配置使用可复现的本地策略：

 - vector / graph / cache / history / coding 等 **一律 SDK config/env**；
 - 模型与 embedding 来自版本化 Model Profile；密钥进安全存储；
 - MemHub 管理库可存 harness/model/队列状态与 **`memory_id` 引用**。

-发布包默认 SDK 行为环境变量仍集中在
-`apps/core/src/one_memory/hy_runtime_config.py`，并由下表锁定：
-
-| 变量 | 值 |
-|---|---|
-| `MEMORY_MODE` | `ultra` |
-| `MEMORY_ENABLE_GRAPH` | `true` |
-| `MEMORY_GRAPH_RERANK_ENABLED` | `false` |
-| `MEMORY_GRAPH_TEMPORAL_SPINE` | `true` |
-| `MEMORY_TOPIC_SUMMARY_ENABLED` | `false` |
-| `MEMORY_EPISODE_SUMMARY_ENABLED` | `false` |
-| `MEMORY_EPISODE_RECALL_ENABLED` | `true` |
-| `MEMORY_DYNAMIC_EPISODE_ENABLED` | `true` |
-| `MEMORY_SESSION_EPISODE_ENABLED` | `false` |
-| `MEMORY_SESSION_EPISODE_CONTEXT_ENABLED` | `false` |
-| `MEMORY_EPISODE_CAPSULE_CONTEXT_ENABLED` | `true` |
-| `MEMORY_EPISODE_CAPSULE_QUERY_MIN` | `0.65` |
-| `MEMORY_EPISODE_CAPSULE_ASSIGNMENT_MIN` | `0.60` |
-| `MEMORY_EPISODE_CAPSULE_FIRST_RANK_MAX` | `3` |
-| `MEMORY_EPISODE_CAPSULE_MEMBER_LIMIT` | `3` |
-| `MEMORY_ENABLE_INTENTION` | `false` |
-| `MEMORY_ENABLE_QUERY_REWRITE` | `false` |
-| `MEMORY_FEW_SHOT_ENABLED` | `true` |
-| `MEMORY_HYBRID_V2_QUERY_EXPAND` | `1` |
-| `MEMORY_HYBRID_V2_GEO_INTENT_PRIOR` | `0.6` |
-| `MEMORY_HYBRID_V2_DIVERSITY` | `1` |
-| `MEMORY_HYBRID_V2_TEMPORAL_ADMIT_PRIOR` | `0.55` |
-| `MEMORY_HYBRID_V2_TEMPORAL_BOOST` | `0.18` |
-| `MEMORY_CROSS_DOMAIN_CON_THRESHOLD` | `0.5` |
-| `MEMORY_READER_ENABLE_SUMMARY` | `false` |
-| `MEMORY_SUMMARY_ENABLED_IN_SYS2` | `false` |
-| `MEMORY_LOCAL_RERANK_ENABLED` | `false` |
-| `RECONCILE_ENABLED` | `true` |
-| `HY_MEMORY_READER` | `hybrid` |
+MemHub 不再通过 import-time 环境变量覆盖 SDK 的提取、reconcile、episode、
+intention、few-shot、summary 或去重策略。产品只配置模型/密钥/endpoint、本地
+backend 路径、公开 embedding service，以及每次公开 Client 调用的参数。

 独立脚本直连 SDK 时须自配存储目录与模型端点。

@@ -77,7 +53,7 @@ MemHub 对 SDK 使用公开、可复现的本地策略（经 bridge / `hy_runtim

 下面是独立调试示例。

-产品代码经 `HyMemoryBridge` 获取已配置的 client（管理面门禁），memory I/O
+产品代码经 `hy_sdk_runtime.SdkClientHandle` 获取已配置的公开 client；memory I/O
 由 SDK 完成。

 ```python
@@ -228,7 +204,8 @@ result = client.search(
 - 区分空结果、权限拒绝、不可用、超时、部分结果；
 - 未发布的 Matter 占位功能不保留表或 API；未来组织功能另立合同。

-生产入口：`hy_memory_bridge.py`、`recall.py`、`packages/host-bridge`。
+生产入口：`hy_sdk_runtime.py`、`sdk_write_processor.py`、`recall.py` 和
+`packages/host-bridge`。

 ## 关闭客户端

@@ -245,8 +222,7 @@ Runtime 长期复用 client；仅在停止、配置切换或受控恢复时重

 ```text
 harness / user action (+ authz, model, queue)
-  -> HyMemoryBridge
-  -> Hy Memory SDK
+  -> public HyMemoryClient
  -> SDK stores (config/env)
  -> optional MemHub persist memory_id / org metadata
  -> Recall Broker / UI (SDK read + MemHub authz)
@@ -18,13 +18,13 @@ If you run into a problem, please open an Issue with the version, platform,repro

 ## 发行与平台合同

-源码当前固定使用公开 PyPI 开发预览 `hy-memory==2.0.5.dev3`。MemHub 源码版本为 `1.0.4`
-（build 13），GitHub 仓库为
+源码当前固定使用公开 PyPI 正式版 `hy-memory==2.0.6`。MemHub 源码版本为 `1.0.5`
+（build 14），GitHub 仓库为
 [`Tencent-Hunyuan/Hy-Memory`](https://github.com/Tencent-Hunyuan/Hy-Memory)，npm 包名为
 `memhub-cli`。当前代码架构见
 [`docs/ARCHITECTURE.md`](docs/ARCHITECTURE.md)。

-`v1.0.4` 的发行目标是 Linux x64/arm64 Runtime 与 macOS arm64/x64 App 和 Runtime。
+`v1.0.5` 的发行目标是 Linux x64/arm64 Runtime 与 macOS arm64/x64 App 和 Runtime。
 Linux 制品由 GitHub Actions 在原生架构 runner 上构建并完成生命周期验收；macOS
 制品由发行流水线完成 Developer ID 签名、公证和验收后手动上传。Windows 不支持。
 只有同时出现在 GitHub Release 与 `release-manifest.json` 中、且通过公开校验的制品才属于
@@ -45,7 +45,7 @@ Linux 制品由 GitHub Actions 在原生架构 runner 上构建并完成生命

 ## 仓库结构

- `apps/core`：Python / FastAPI Kernel，负责授权、模型、队列、SDK bridge 与 API。
+- `apps/core`：Python / FastAPI Kernel，负责授权、模型、队列、SDK Runtime 边界与 API。
 - `apps/cli`：安装、控制和诊断共享 Runtime 的 `one-memory` CLI。
 - `apps/web`：React / TypeScript 管理界面。
 - `apps/desktop`：Electron 桌面壳与图形安装入口。
@@ -117,7 +117,7 @@ MemHub 的 SQLite 数据默认位于 `~/.one-memory`，不在 npm 包或 Runtime

 ## Hy Memory

-SDK 依赖固定为公开 PyPI 开发预览 `hy-memory==2.0.5.dev3`。SDK 配置、权限边界及 search
+SDK 依赖固定为公开 PyPI 正式版 `hy-memory==2.0.6`。SDK 配置、权限边界及 search
 返回格式见 [`MEMORY.md`](MEMORY.md)。HY search 返回的 `content` 已包含 SDK 统一生成的
 逻辑时间和演化上下文；调用方不应再次转换时间戳或重复拼接 `When`。

@@ -3,9 +3,9 @@
  "dataLicense": "CC0-1.0",
  "SPDXID": "SPDXRef-DOCUMENT",
  "name": "one-memory",
-  "documentNamespace": "https://one-memory.local/spdx/2026-10-22T08:09:31Z",
+  "documentNamespace": "https://one-memory.local/spdx/2026-10-22T13:18:16Z",
  "creationInfo": {
-    "created": "2026-10-22T08:09:31Z",
+    "created": "2026-10-22T13:18:16Z",
    "creators": [
      "Tool: one-memory-generate-sbom"
    ]
@@ -2352,7 +2352,7 @@
    {
      "SPDXID": "SPDXRef-Package-hy-memory",
      "name": "hy-memory",
-      "versionInfo": "2.0.5.dev3",
+      "versionInfo": "2.0.6",
      "downloadLocation": "NOASSERTION",
      "filesAnalyzed": false,
      "licenseConcluded": "Apache-2.0",
@@ -2,14 +2,14 @@

 MemHub itself is licensed under Apache-2.0 (see `LICENSE`).
 It bundles the runtime dependencies listed below.
-Generated from the installed dependency tree on 2026-10-22T08:09:31Z
+Generated from the installed dependency tree on 2026-10-22T13:18:16Z
 by `python3 scripts/generate_sbom.py`. Do not edit by hand.

 Per-package license texts are not reproduced here yet; the license
 column is filled in as part of release review, and the SBOM records
 `NOASSERTION` until then rather than guessing.

-## hy-memory 2.0.5.dev3
+## hy-memory 2.0.6

 License: Apache-2.0

@@ -1,6 +1,6 @@
 {
  "name": "@one-memory/cli",
-  "version": "1.0.4",
+  "version": "1.0.5",
  "description": "MemHub bootstrap CLI for installing and managing the local-first shared memory runtime",
  "license": "Apache-2.0",
  "repository": {
@@ -21,18 +21,21 @@ const TOKEN = "probe-token-not-a-real-credential";

 const MEMORY = {
  memory_id: "probe-1",
-  user_id: "local",
-  kind: "chat",
-  status: "confirmed",
  content: "部署脚本要先跑 verify_merge_plan。",
-  scopes: ["all-apps"],
-  evidence_ids: ["probe-e1"],
-  origin: "extracted",
+  layer: "l2_fact",
+  status: "active",
+  is_latest: true,
+  supersedes: [],
+  superseded_by: [],
  tags: [],
-  revision: 2,
-  created_at: "2026-09-19T00:00:00Z",
-  updated_at: "2026-09-19T00:00:00Z",
-  deleted_at: null,
+  owner: null,
+  source_raw_ids: ["probe-raw-1"],
+  user_id: "local",
+  agent_id: "codex",
+  session_id: "probe-session",
+  source: "codex",
+  memory_at: 1758240000,
+  gmt_created: 1758240000,
 };

 const ROUTES = new Map([
@@ -124,26 +127,6 @@ const ROUTES = new Map([
      },
    ],
  ],
-  [
-    "/api/v1/memory-tasks",
-    [
-      {
-        task_id: "probe-task",
-        request_id: "probe-request",
-        user_id: "local",
-        kind: "extract",
-        input_ref: "probe",
-        status: "waiting_for_model",
-        profile_id: "probe-profile",
-        profile_revision: 1,
-        profile_snapshot: {},
-        attempts: 0,
-        last_error_code: "model_missing",
-        created_at: "2026-09-19T00:00:00Z",
-        updated_at: "2026-09-19T00:00:01Z",
-      },
-    ],
-  ],
  [
    "/api/v1/model-profiles",
    [
@@ -159,28 +142,20 @@ const ROUTES = new Map([
      },
    ],
  ],
-  ["/api/v1/home-summary", { memory_count: 1, latest: null, observed_at: "T" }],
  [
-    "/api/v1/memory-library",
-    { total: 1, offset: 0, limit: 100, items: [MEMORY] },
-  ],
-  [
-    "/api/v1/memories/probe-1",
+    "/api/v2/home-summary",
    {
-      memory: MEMORY,
-      evidence: [
-        {
-          evidence_id: "probe-e1",
-          available: true,
-          source_id: "codex",
-          written_by: "user",
-          created_at: "2026-09-19T00:00:00Z",
-          verbatim: "部署前先跑 verify_merge_plan，别靠读 diff 猜冲突。",
-        },
-      ],
-      history: [{ memory_id: "probe-1", revision: 2, action: "confirm" }],
+      schema_version: 2,
+      memory_count: 1,
+      latest: MEMORY,
+      observed_at: "2026-09-19T00:00:01Z",
    },
  ],
+  [
+    "/api/v2/memory-library",
+    { memories: [MEMORY], total: 1, offset: 0, limit: 100 },
+  ],
+  ["/api/v2/memories/probe-1", MEMORY],
 ]);

 const seen = [];
@@ -1,6 +1,7 @@
-import { mkdtemp, readdir, rm, writeFile } from "node:fs/promises";
+import { mkdtemp, readdir, readFile, rm, writeFile } from "node:fs/promises";
 import os from "node:os";
 import path from "node:path";
+import { inflateRawSync } from "node:zlib";
 import { DiagnosticLogger } from "@one-memory/runtime-control";
 import { afterEach, expect, it, vi } from "vitest";

@@ -29,6 +30,27 @@ async function fixture() {
  return dataDir;
 }

+function readZipEntry(archive: Buffer, expected: string): string {
+  let offset = 0;
+  while (archive.readUInt32LE(offset) === 0x04034b50) {
+    const method = archive.readUInt16LE(offset + 8);
+    const compressedSize = archive.readUInt32LE(offset + 18);
+    const nameLength = archive.readUInt16LE(offset + 26);
+    const extraLength = archive.readUInt16LE(offset + 28);
+    const nameStart = offset + 30;
+    const dataStart = nameStart + nameLength + extraLength;
+    const name = archive.subarray(nameStart, nameStart + nameLength).toString();
+    const compressed = archive.subarray(dataStart, dataStart + compressedSize);
+    if (name === expected) {
+      if (method === 0) return compressed.toString();
+      if (method === 8) return inflateRawSync(compressed).toString();
+      throw new Error("unsupported zip compression");
+    }
+    offset = dataStart + compressedSize;
+  }
+  throw new Error(`missing ${expected}`);
+}
+
 it("prints matching diagnostic events as NDJSON without starting Runtime", async () => {
  const dataDir = await fixture();
  let stdout = "";
@@ -74,6 +96,25 @@ it("exports one support zip to the requested downloads directory", async () => {
  });
  expect(receipt.path).toMatch(/MemHub-日志-3800-\d{8}-\d{4}\.zip$/);
  expect(await readdir(downloads)).toEqual([path.basename(receipt.path)]);
+  const archive = await readFile(receipt.path);
+  expect(JSON.parse(readZipEntry(archive, "summary.json")).commit).toBeNull();
+});
+
+it("fails closed for an explicitly supplied invalid commit", async () => {
+  const dataDir = await fixture();
+  const downloads = await mkdtemp(path.join(os.tmpdir(), "om-cli-downloads-"));
+  roots.push(downloads);
+  vi.spyOn(process.stdout, "write").mockImplementation(() => true);
+
+  await expect(
+    run(["logs", "export", "--json"], {
+      ONE_MEMORY_DATA_DIR: dataDir,
+      ONE_MEMORY_EXPORT_DIR: downloads,
+      ONE_MEMORY_BUILD_NUMBER: "3800",
+      ONE_MEMORY_COMMIT: "unknown",
+    }),
+  ).rejects.toThrow("diagnostic_commit_invalid");
+  expect(await readdir(downloads)).toEqual([]);
 });

 it("never emits a secret from a malformed manually written event", async () => {
@@ -3,7 +3,9 @@ import { tmpdir } from "node:os";
 import path from "node:path";
 import {
  claimRuntimeServiceForApp,
+  clearSdkMigrationBlock,
  readRuntimeServiceOwner,
+  recordSdkMigrationBlock,
  resolveRuntimeCliConfig,
 } from "@one-memory/runtime-control";
 import { describe, expect, it, vi } from "vitest";
@@ -57,6 +59,11 @@ const runtimeMocks = vi.hoisted(() => ({
  })),
  uninstallUserService: vi.fn(async (..._args: unknown[]) => "removed"),
  restartUserService: vi.fn(async () => undefined),
+  runSdkMigration: vi.fn(
+    async (): Promise<{ state: string; code?: string }> => ({
+      state: "not_required",
+    }),
+  ),
  rollbackLocalInstall: vi.fn(async () => undefined),
  planCliLink: vi.fn(async ({ home, installRoot }) => ({
    schema_version: 1,
@@ -126,6 +133,7 @@ vi.mock("@one-memory/runtime-control", async (importOriginal) => {
    installUserService: runtimeMocks.installUserService,
    uninstallUserService: runtimeMocks.uninstallUserService,
    restartUserService: runtimeMocks.restartUserService,
+    runSdkMigration: runtimeMocks.runSdkMigration,
    rollbackLocalInstall: runtimeMocks.rollbackLocalInstall,
    planCliLink: runtimeMocks.planCliLink,
    installCliLink: runtimeMocks.installCliLink,
@@ -784,7 +792,7 @@ describe("public One Memory CLI", () => {
      }
    });

-    it("upgrade failure restarts the restored artifact with the same data identity", async () => {
+    it("upgrade restart failure keeps the new version and never restarts the old one", async () => {
      vi.clearAllMocks();
      runtimeMocks.planCliLink.mockResolvedValueOnce({
        schema_version: 1,
@@ -826,25 +834,27 @@ describe("public One Memory CLI", () => {
        command: "upgrade",
        code: "restart failed",
      });
-      expect(runtimeMocks.rollbackLocalInstall).toHaveBeenCalledOnce();
-      // The CLI manages the objective aliases the package does not own
-      // (`memhub` + `memhub-cli`); both are rolled back.
-      expect(aliasMocks.rollbackManagedCliLink).toHaveBeenCalledTimes(2);
-      expect(runtimeMocks.restartUserService).toHaveBeenCalledTimes(2);
-      for (const call of runtimeMocks.restartUserService.mock
-        .calls as unknown[][]) {
-        expect(call[0]).toMatchObject({
-          dataDir: "/tmp/custom data",
-          installRoot: "/tmp/install/current",
-        });
-      }
+      expect(JSON.parse(streams.stderr())).toMatchObject({
+        rolled_back: false,
+      });
+      // Product rule: the previous Runtime may predate the SDK migration gate,
+      // so it is never restored after the migration was handed the home.
+      expect(runtimeMocks.rollbackLocalInstall).not.toHaveBeenCalled();
+      expect(aliasMocks.rollbackManagedCliLink).not.toHaveBeenCalled();
+      expect(runtimeMocks.restartUserService).toHaveBeenCalledOnce();
+      expect(
+        (runtimeMocks.restartUserService.mock.calls as unknown[][])[0]?.[0],
+      ).toMatchObject({
+        dataDir: "/tmp/custom data",
+        installRoot: "/tmp/install/current",
+      });
    });

    // O-22 (2026-09-29 C8 H17): `upgrade --yes` on a manifest-valid but
    // crash-looping version reported exit 0 and left `current` on the bad
    // version. A restart that merely returns is not proof the new Runtime came
    // up, so the CLI now verifies health and rolls back if it never does.
-    it("rolls back and fails when the upgraded Runtime never becomes healthy", async () => {
+    it("fails without rolling back when the upgraded Runtime never becomes healthy", async () => {
      vi.clearAllMocks();
      runtimeMocks.planCliLink.mockResolvedValueOnce({
        schema_version: 1,
@@ -884,9 +894,219 @@ describe("public One Memory CLI", () => {
        command: "upgrade",
        code: "upgrade_health_check_failed",
      });
-      // Rolled back to the previous version and restarted it.
+      // Kept on the new version; the old Runtime is not restarted.
+      expect(runtimeMocks.rollbackLocalInstall).not.toHaveBeenCalled();
+      expect(runtimeMocks.restartUserService).toHaveBeenCalledOnce();
+    });
+
+    const upgradeArguments = [
+      "upgrade",
+      "--source",
+      "/tmp/dist",
+      "--home",
+      "/tmp/home",
+      "--install-root",
+      "/tmp/install",
+      "--data-dir",
+      "/tmp/custom data",
+      "--yes",
+    ];
+    const presentCliLink = {
+      schema_version: 1 as const,
+      platform: "linux" as const,
+      home: "/tmp/home",
+      install_root: "/tmp/install",
+      bin_directory: "/tmp/bin",
+      status: "present" as const,
+      blocked_reason: null,
+      link_path: "/tmp/bin/one-memory",
+      target_path: "/tmp/install/current/bin/one-memory",
+    };
+
+    it("says the data was restored when a failed migration put it back", async () => {
+      vi.clearAllMocks();
+      runtimeMocks.planCliLink.mockResolvedValueOnce(presentCliLink);
+      runtimeMocks.runSdkMigration.mockResolvedValueOnce({
+        state: "deferred",
+        code: "sdk_migration_memory_unreadable",
+        dataRestored: true,
+      } as never);
+      const streams = captureStreams();
+      let code: number;
+      try {
+        code = await run(
+          upgradeArguments,
+          {},
+          { waitForRuntimeReady: async () => true },
+        );
+      } finally {
+        streams.restore();
+      }
+      expect(code!).toBe(1);
+      const failure = JSON.parse(streams.stderr());
+      expect(failure).toMatchObject({
+        code: "sdk_migration_memory_unreadable",
+        data_restored: true,
+        log_file: "/tmp/custom data/logs/sdk-migration.log",
+      });
+      expect(failure.message).toContain("记忆数据已恢复到升级前的状态。");
+      expect(failure.message).toContain(
+        "/tmp/custom data/logs/sdk-migration.log",
+      );
+    });
+
+    it("migrates the home with the new Core before restarting the Runtime", async () => {
+      vi.clearAllMocks();
+      runtimeMocks.planCliLink.mockResolvedValueOnce(presentCliLink);
+      const order: string[] = [];
+      runtimeMocks.runSdkMigration.mockImplementationOnce(async () => {
+        order.push("migrate");
+        return { state: "cutover" };
+      });
+      runtimeMocks.restartUserService.mockImplementationOnce(async () => {
+        order.push("restart");
+      });
+      const output = captureStdout();
+      let code: number;
+      try {
+        code = await run(
+          upgradeArguments,
+          {},
+          { waitForRuntimeReady: async () => true },
+        );
+      } finally {
+        output.restore();
+      }
+      expect(code!).toBe(0);
+      expect(order).toEqual(["migrate", "restart"]);
+      expect(runtimeMocks.runSdkMigration).toHaveBeenCalledWith(
+        expect.objectContaining({ dataDir: "/tmp/custom data" }),
+      );
+    });
+
+    it("stays on the new version, not serving, when the migration does not finish", async () => {
+      vi.clearAllMocks();
+      runtimeMocks.planCliLink.mockResolvedValueOnce(presentCliLink);
+      runtimeMocks.runSdkMigration.mockResolvedValueOnce({
+        state: "deferred",
+        code: "sdk_migration_disk_space_insufficient",
+      });
+      const streams = captureStreams();
+      let code: number;
+      try {
+        code = await run(
+          upgradeArguments,
+          {},
+          { waitForRuntimeReady: async () => true },
+        );
+      } finally {
+        streams.restore();
+      }
+      expect(code!).toBe(1);
+      const failure = JSON.parse(streams.stderr());
+      expect(failure).toMatchObject({
+        command: "upgrade",
+        code: "sdk_migration_disk_space_insufficient",
+        // The one file support needs, and where to send it.
+        log_file: "/tmp/custom data/logs/sdk-migration.log",
+        issues_url: "https://github.com/Tencent-Hunyuan/Hy-Memory/issues/new",
+      });
+      expect(failure.message).toContain(
+        "/tmp/custom data/logs/sdk-migration.log",
+      );
+      expect(failure.message).toContain("GitHub");
+      expect(failure.message).not.toContain("回滚到上一版本");
+      expect(runtimeMocks.runSdkMigration).toHaveBeenCalledWith(
+        expect.objectContaining({ caller: "cli-upgrade" }),
+      );
+      // No rollback: the old Runtime has no migration gate. The new service is
+      // (re)started so its gate, not the old Runtime, holds the data dir.
+      expect(runtimeMocks.rollbackLocalInstall).not.toHaveBeenCalled();
+      expect(aliasMocks.rollbackManagedCliLink).not.toHaveBeenCalled();
+      expect(runtimeMocks.restartUserService).toHaveBeenCalledOnce();
+      expect(
+        (runtimeMocks.restartUserService.mock.calls as unknown[][])[0]?.[0],
+      ).toMatchObject({
+        installRoot: "/tmp/install/current",
+      });
+    });
+
+    it("stays on the new version when the migration itself throws", async () => {
+      vi.clearAllMocks();
+      runtimeMocks.planCliLink.mockResolvedValueOnce(presentCliLink);
+      runtimeMocks.runSdkMigration.mockRejectedValueOnce(
+        new Error("sdk_migration_fence_failed"),
+      );
+      const streams = captureStreams();
+      let code: number;
+      try {
+        code = await run(
+          upgradeArguments,
+          {},
+          { waitForRuntimeReady: async () => true },
+        );
+      } finally {
+        streams.restore();
+      }
+      expect(code!).toBe(1);
+      expect(JSON.parse(streams.stderr())).toMatchObject({
+        command: "upgrade",
+        code: "sdk_migration_fence_failed",
+        log_file: "/tmp/custom data/logs/sdk-migration.log",
+        issues_url: "https://github.com/Tencent-Hunyuan/Hy-Memory/issues/new",
+      });
+      expect(runtimeMocks.rollbackLocalInstall).not.toHaveBeenCalled();
+    });
+
+    it("still rolls back a failure before the migration touched the home", async () => {
+      vi.clearAllMocks();
+      runtimeMocks.planCliLink.mockResolvedValueOnce(presentCliLink);
+      aliasMocks.installManagedCliLink.mockRejectedValueOnce(
+        new Error("cli_link_failed"),
+      );
+      const streams = captureStreams();
+      let code: number;
+      try {
+        code = await run(
+          upgradeArguments,
+          {},
+          { waitForRuntimeReady: async () => true },
+        );
+      } finally {
+        streams.restore();
+      }
+      expect(code!).toBe(1);
+      expect(JSON.parse(streams.stderr())).toMatchObject({
+        command: "upgrade",
+        code: "cli_link_failed",
+      });
+      expect(runtimeMocks.runSdkMigration).not.toHaveBeenCalled();
      expect(runtimeMocks.rollbackLocalInstall).toHaveBeenCalledOnce();
-      expect(runtimeMocks.restartUserService).toHaveBeenCalledTimes(2);
+      // The previous Runtime was never stopped; nothing is (re)started.
+      expect(runtimeMocks.restartUserService).not.toHaveBeenCalled();
+    });
+
+    it("keeps the new version when a migrated home is slow to come up", async () => {
+      vi.clearAllMocks();
+      runtimeMocks.planCliLink.mockResolvedValueOnce(presentCliLink);
+      runtimeMocks.runSdkMigration.mockResolvedValueOnce({ state: "cutover" });
+      const streams = captureStreams();
+      let code: number;
+      try {
+        code = await run(
+          upgradeArguments,
+          {},
+          { waitForRuntimeReady: async () => false },
+        );
+      } finally {
+        streams.restore();
+      }
+      expect(code!).toBe(1);
+      expect(JSON.parse(streams.stderr())).toMatchObject({
+        command: "upgrade",
+        code: "runtime_not_ready_after_sdk_migration",
+      });
+      expect(runtimeMocks.rollbackLocalInstall).not.toHaveBeenCalled();
    });

    it("reports success once the upgraded Runtime is healthy", async () => {
@@ -1384,7 +1604,7 @@ describe("public One Memory CLI", () => {
    expect(JSON.parse(output.value()).apply).toBe(true);
  });

-  it("rolls back the distribution when service installation fails", async () => {
+  it("keeps the distribution when service installation fails after the migration ran", async () => {
    runtimeMocks.applyLocalInstall.mockResolvedValueOnce({
      schema_version: 1,
      source: "/tmp/distribution",
@@ -1425,13 +1645,14 @@ describe("public One Memory CLI", () => {
      command: "install",
      status: "error",
      code: "service failed",
+      rolled_back: false,
    });
-    expect(runtimeMocks.rollbackLocalInstall).toHaveBeenCalledOnce();
-    expect(runtimeMocks.rollbackCliLink).toHaveBeenCalledOnce();
-    expect(aliasMocks.rollbackManagedCliLink).toHaveBeenCalledTimes(2);
+    expect(runtimeMocks.rollbackLocalInstall).not.toHaveBeenCalled();
+    expect(runtimeMocks.rollbackCliLink).not.toHaveBeenCalled();
+    expect(aliasMocks.rollbackManagedCliLink).not.toHaveBeenCalled();
  });

-  it("upgrades only an existing managed CLI and rolls back a failed restart", async () => {
+  it("upgrades only an existing managed CLI and keeps it after a failed restart", async () => {
    runtimeMocks.planCliLink.mockResolvedValueOnce({
      schema_version: 1,
      platform: "linux",
@@ -1471,7 +1692,7 @@ describe("public One Memory CLI", () => {
      status: "error",
      code: "restart failed",
    });
-    expect(runtimeMocks.rollbackLocalInstall).toHaveBeenCalledOnce();
+    expect(runtimeMocks.rollbackLocalInstall).not.toHaveBeenCalled();
    expect(runtimeMocks.installCliLink).not.toHaveBeenCalled();
  });

@@ -1895,6 +2116,189 @@ describe("lifecycle commands accept --json", () => {
  });
 });

+// A Runtime service that refuses to serve an unmigrated home must be reported
+// as that, with the migration log and where to send it, not as a generic
+// "not ready within budget".
+describe("lifecycle commands report a migration block", () => {
+  async function blockedHome() {
+    const dataDir = await mkdtemp(path.join(tmpdir(), "memhub-blocked-"));
+    await recordSdkMigrationBlock(
+      dataDir,
+      "sdk_migration_disk_space_insufficient",
+      "runtime-service",
+    );
+    return dataDir;
+  }
+  // darwin: kickstart "restarts" but never becomes ready. Elsewhere: no
+  // service, so the direct spawn fails. Both must surface the block.
+  const blockedControls = () => ({
+    restartManagedRuntime: vi.fn(async () => "restarted" as const),
+    waitForRuntimeReady: vi.fn(async () => false),
+    stopRuntime: vi.fn(async () => "stopped" as const),
+    ensureRuntime: vi.fn(async () => {
+      throw new Error("Timed out waiting for MemHub Runtime");
+    }),
+  });
+
+  it("restart --json prints the block code, log file and issues URL", async () => {
+    const dataDir = await blockedHome();
+    const output = captureStdout();
+    let code: number;
+    try {
+      code = await run(
+        ["restart", "--json"],
+        { ONE_MEMORY_DATA_DIR: dataDir },
+        blockedControls(),
+      );
+    } finally {
+      output.restore();
+      await rm(dataDir, { recursive: true, force: true });
+    }
+    expect(code!).toBe(1);
+    const lines = output.value().trim().split("\n");
+    const document = JSON.parse(lines[lines.length - 1]!);
+    expect(document).toMatchObject({
+      command: "restart",
+      state: "blocked",
+      code: "sdk_migration_disk_space_insufficient",
+      log_file: path.join(dataDir, "logs", "sdk-migration.log"),
+      issues_url: "https://github.com/Tencent-Hunyuan/Hy-Memory/issues/new",
+    });
+    expect(output.value()).not.toContain("runtime_not_ready_within_budget");
+  });
+
+  it("restart on a frozen home retries the migration in the foreground first", async () => {
+    const dataDir = await mkdtemp(path.join(tmpdir(), "memhub-frozen-"));
+    await recordSdkMigrationBlock(
+      dataDir,
+      "sdk_migration_memory_unreadable",
+      "desktop",
+      { frozen: { service_was_enabled: true }, data_restored: true },
+    );
+    const controls = {
+      ...blockedControls(),
+      migrateSdkData: vi.fn(async () => ({
+        state: "deferred" as const,
+        code: "sdk_migration_memory_unreadable",
+        dataRestored: true,
+      })),
+    };
+    const output = captureStdout();
+    let code: number;
+    try {
+      code = await run(
+        ["restart", "--json"],
+        { ONE_MEMORY_DATA_DIR: dataDir },
+        controls,
+      );
+    } finally {
+      output.restore();
+      await rm(dataDir, { recursive: true, force: true });
+    }
+    expect(code!).toBe(1);
+    expect(controls.migrateSdkData).toHaveBeenCalledWith(
+      expect.objectContaining({ dataDir, caller: "cli-restart" }),
+    );
+    // Still frozen: the service is not started to retry on its own.
+    expect(controls.restartManagedRuntime).not.toHaveBeenCalled();
+    expect(controls.ensureRuntime).not.toHaveBeenCalled();
+    expect(JSON.parse(output.value())).toMatchObject({
+      command: "restart",
+      state: "blocked",
+      code: "sdk_migration_memory_unreadable",
+      data_restored: true,
+    });
+  });
+
+  it("restart on a frozen home restarts the Runtime once the retry succeeds", async () => {
+    const dataDir = await mkdtemp(path.join(tmpdir(), "memhub-frozen-"));
+    await recordSdkMigrationBlock(
+      dataDir,
+      "sdk_migration_memory_unreadable",
+      "desktop",
+      { frozen: { service_was_enabled: false } },
+    );
+    const order: string[] = [];
+    const migrateSdkData = vi.fn(async () => {
+      order.push("migrate");
+      await clearSdkMigrationBlock(dataDir);
+      return { state: "cutover" as const };
+    });
+    const output = captureStreams();
+    let code: number;
+    try {
+      code = await run(
+        ["restart", "--json"],
+        { ONE_MEMORY_DATA_DIR: dataDir },
+        {
+          migrateSdkData,
+          waitForRuntimeReady: vi.fn(async () => true),
+          restartManagedRuntime: vi.fn(async () => {
+            order.push("restart");
+            return "restarted" as const;
+          }),
+          stopRuntime: vi.fn(async () => "stopped" as const),
+          ensureRuntime: vi.fn(async () => {
+            order.push("restart");
+            return { pid: 1 } as never;
+          }),
+        },
+      );
+    } finally {
+      output.restore();
+      await rm(dataDir, { recursive: true, force: true });
+    }
+    expect(code!).toBe(0);
+    expect(migrateSdkData).toHaveBeenCalledWith(
+      expect.objectContaining({ caller: "cli-restart" }),
+    );
+    expect(order).toEqual(["migrate", "restart"]);
+  });
+
+  it("restart in text mode names the log file and GitHub issues", async () => {
+    const dataDir = await blockedHome();
+    const output = captureStdout();
+    try {
+      expect(
+        await run(
+          ["restart"],
+          { ONE_MEMORY_DATA_DIR: dataDir },
+          blockedControls(),
+        ),
+      ).toBe(1);
+    } finally {
+      output.restore();
+      await rm(dataDir, { recursive: true, force: true });
+    }
+    expect(output.value()).toContain("sdk_migration_disk_space_insufficient");
+    expect(output.value()).toContain(
+      path.join(dataDir, "logs", "sdk-migration.log"),
+    );
+    expect(output.value()).toContain(
+      "https://github.com/Tencent-Hunyuan/Hy-Memory/issues/new",
+    );
+  });
+
+  it("ui --json reports the block instead of runtime_not_running", async () => {
+    const dataDir = await blockedHome();
+    const output = captureStdout();
+    let code: number;
+    try {
+      code = await run(["ui", "--json"], { ONE_MEMORY_DATA_DIR: dataDir });
+    } finally {
+      output.restore();
+      await rm(dataDir, { recursive: true, force: true });
+    }
+    expect(code!).not.toBe(0);
+    expect(JSON.parse(output.value())).toMatchObject({
+      command: "ui",
+      state: "blocked",
+      code: "sdk_migration_disk_space_insufficient",
+      log_file: path.join(dataDir, "logs", "sdk-migration.log"),
+    });
+  });
+});
+
 // The objective names `memhub host install <host>` (host as a positional). Both
 // that and the `--host <id>` flag form must reach the same command; naming both
 // at once is ambiguous and refused rather than guessed.
@@ -2004,7 +2408,7 @@ describe("install gates on Runtime readiness", () => {
    });
  }

-  it("fails and rolls back when the installed Runtime never becomes ready", async () => {
+  it("fails without rolling back when the installed Runtime never becomes ready", async () => {
    vi.clearAllMocks();
    stubInstallPlan();
    const streams = captureStreams();
@@ -2033,10 +2437,10 @@ describe("install gates on Runtime readiness", () => {
    }
    expect(code!).toBe(1);
    expect(streams.stderr()).toContain("install_runtime_not_ready");
-    expect(runtimeMocks.rollbackLocalInstall).toHaveBeenCalledOnce();
+    expect(runtimeMocks.rollbackLocalInstall).not.toHaveBeenCalled();
    expect(
      lifecycleMocks.rollbackPreparedServiceInstall,
-    ).toHaveBeenCalledOnce();
+    ).not.toHaveBeenCalled();
  });

  it("takes CLI ownership after a successful install over an App-owned Runtime", async () => {
@@ -2081,7 +2485,7 @@ describe("install gates on Runtime readiness", () => {
    }
  });

-  it("restores App ownership before relaunching the old Runtime after install failure", async () => {
+  it("never hands the home back to the App Runtime after the migration ran", async () => {
    vi.clearAllMocks();
    stubInstallPlan();
    const root = await mkdtemp(path.join(tmpdir(), "memhub-cli-owner-"));
@@ -2091,14 +2495,6 @@ describe("install gates on Runtime readiness", () => {
      appBundle: "/Applications/MemHub.app",
      appExecutable: "/Applications/MemHub.app/Contents/MacOS/MemHub",
    });
-    lifecycleMocks.rollbackPreparedServiceInstall.mockImplementationOnce(
-      async (_prepared: unknown, rollback: () => Promise<void>) => {
-        await rollback();
-        await expect(readRuntimeServiceOwner(dataDir)).resolves.toMatchObject({
-          owner: "app",
-        });
-      },
-    );
    const streams = captureStreams();
    try {
      expect(
@@ -2121,15 +2517,114 @@ describe("install gates on Runtime readiness", () => {
          { waitForRuntimeReady: async () => false },
        ),
      ).toBe(1);
+      // The App Runtime may predate the SDK migration gate: CLI keeps it.
      await expect(readRuntimeServiceOwner(dataDir)).resolves.toMatchObject({
-        owner: "app",
+        owner: "cli",
      });
+      expect(
+        lifecycleMocks.rollbackPreparedServiceInstall,
+      ).not.toHaveBeenCalled();
    } finally {
      streams.restore();
      await rm(root, { recursive: true, force: true });
    }
  });

+  it("keeps the install, not serving, when its migration is deferred", async () => {
+    vi.clearAllMocks();
+    stubInstallPlan();
+    runtimeMocks.runSdkMigration.mockResolvedValueOnce({
+      state: "deferred",
+      code: "sdk_migration_memory_unreadable",
+    });
+    const waitForRuntimeReady = vi.fn(async () => true);
+    const streams = captureStreams();
+    let code: number;
+    try {
+      code = await run(
+        [
+          "install",
+          "--source",
+          "/tmp/dist",
+          "--home",
+          "/tmp/home",
+          "--install-root",
+          "/tmp/install",
+          "--bin-dir",
+          "/tmp/bin",
+          "--data-dir",
+          "/tmp/data",
+          "--yes",
+        ],
+        {},
+        { waitForRuntimeReady },
+      );
+    } finally {
+      streams.restore();
+    }
+    expect(code!).toBe(1);
+    const failure = JSON.parse(streams.stderr());
+    expect(failure).toMatchObject({
+      command: "install",
+      code: "sdk_migration_memory_unreadable",
+      log_file: "/tmp/data/logs/sdk-migration.log",
+      issues_url: "https://github.com/Tencent-Hunyuan/Hy-Memory/issues/new",
+    });
+    expect(failure.message).toContain("/tmp/data/logs/sdk-migration.log");
+    expect(runtimeMocks.runSdkMigration).toHaveBeenCalledWith(
+      expect.objectContaining({ caller: "cli-install" }),
+    );
+    // The new (gated) service owns the data dir; the previous registration is
+    // never re-bootstrapped and nothing is rolled back.
+    expect(lifecycleMocks.applyPreparedServiceInstall).toHaveBeenCalledWith(
+      expect.objectContaining({ restorePreviousOnFailure: false }),
+    );
+    expect(runtimeMocks.rollbackLocalInstall).not.toHaveBeenCalled();
+    expect(
+      lifecycleMocks.rollbackPreparedServiceInstall,
+    ).not.toHaveBeenCalled();
+    expect(waitForRuntimeReady).not.toHaveBeenCalled();
+  });
+
+  it("still rolls back an install that failed before its migration ran", async () => {
+    vi.clearAllMocks();
+    stubInstallPlan();
+    runtimeMocks.installCliLink.mockRejectedValueOnce(
+      new Error("cli_link_failed"),
+    );
+    const streams = captureStreams();
+    let code: number;
+    try {
+      code = await run(
+        [
+          "install",
+          "--source",
+          "/tmp/dist",
+          "--home",
+          "/tmp/home",
+          "--install-root",
+          "/tmp/install",
+          "--bin-dir",
+          "/tmp/bin",
+          "--data-dir",
+          "/tmp/data",
+          "--yes",
+        ],
+        {},
+        { waitForRuntimeReady: async () => true },
+      );
+    } finally {
+      streams.restore();
+    }
+    expect(code!).toBe(1);
+    expect(JSON.parse(streams.stderr())).toMatchObject({
+      command: "install",
+      code: "cli_link_failed",
+    });
+    expect(runtimeMocks.runSdkMigration).not.toHaveBeenCalled();
+    expect(runtimeMocks.rollbackLocalInstall).toHaveBeenCalledOnce();
+  });
+
  it("succeeds when the installed Runtime becomes ready", async () => {
    vi.clearAllMocks();
    stubInstallPlan();
@@ -32,6 +32,8 @@ import {
  processIsAlive,
  type ResolvedRuntimeRelease,
  RuntimeDiscoveryError,
+  readInstalledRuntimeSource,
+  readSdkMigrationBlock,
  removeRuntimeServiceOwner,
  resolveRuntimeCliConfig,
  resolveRuntimeRelease,
@@ -39,7 +41,11 @@ import {
  rollbackCliLink,
  rollbackLocalInstall,
  runRuntimeCli,
+  runSdkMigration,
  runtimeInstallLayout,
+  type SdkMigrationResult,
+  type SdkMigrationSource,
+  sdkMigrationLogPath,
  stopRuntimeProcess,
 } from "@one-memory/runtime-control";
 import axios from "axios";
@@ -175,21 +181,10 @@ import {
  removeManagedCliLink,
  rollbackManagedCliLink,
 } from "./managed-cli-link.js";
-import {
-  executeMemoryMutation,
-  inspectMemoryPlan,
-  type MemoryFlowDocument,
-  memoryFlowExitCode,
-  memoryPlanAcceptsContent,
-  parseMemoryContent,
-  planExistingMemoryMutation,
-  planMemoryAdd,
-  readMemoryMutationReceipt,
-  renderMemoryFlowText,
-} from "./memory-flows.js";
 import {
  isMemoryId,
  type MemoryQuery,
+  parseMemoryContent,
  readMemoryDetail,
  readMemoryPage,
  renderMemoryDetailText,
@@ -354,6 +349,9 @@ export interface RunControls {
  /** Wait for a restarted Runtime to become healthy after `upgrade`.
   * Injectable so the upgrade tests do not spend real seconds polling. */
  waitForRuntimeReady?: (dataDir: string) => Promise<boolean>;
+  /** Run the SDK data migration before install/upgrade starts the new
+   * Runtime. Injectable so the tests do not spawn a real Core. */
+  migrateSdkData?: typeof migrateSdkDataBeforeStart;
  restartManagedRuntime?: typeof restartManagedService;
  stopRuntime?: typeof stopRuntimeProcess;
  ensureRuntime?: typeof ensureRuntimeProcess;
@@ -1048,15 +1046,21 @@ async function runCommand(
        entry: "cli",
      }));
      const target = path.join(downloads, diagnosticZipFilename(build));
+      const environmentCommit = environment.ONE_MEMORY_COMMIT?.trim();
+      const embeddedCommit =
+        typeof __ONE_MEMORY_COMMIT__ === "string"
+          ? __ONE_MEMORY_COMMIT__.trim()
+          : "";
+      const commit =
+        environmentCommit ||
+        (embeddedCommit && embeddedCommit !== "unknown"
+          ? embeddedCommit
+          : undefined);
      const receipt = await exportDiagnosticZip({
        dataDir: config.dataDir,
        target,
        productVersion: build,
-        commit:
-          environment.ONE_MEMORY_COMMIT?.trim() ||
-          (typeof __ONE_MEMORY_COMMIT__ === "string"
-            ? __ONE_MEMORY_COMMIT__
-            : "unknown"),
+        ...(commit === undefined ? {} : { commit }),
        coreSha256: await sha256File(config.coreCommand).catch(() => undefined),
        home: os.homedir(),
        crashReportsDir: path.join(
@@ -1164,6 +1168,41 @@ async function runCommand(
    // declared command, and lifecycle is exactly where scripts need it.
    if (arguments_.length !== (asJson ? 2 : 1)) return usage();
    const config = resolveRuntimeCliConfig(environment, process.argv[1]);
+    // A failed upgrade migration froze this home: its service exits without
+    // migrating. `start`/`restart` is an explicit retry, run here in the
+    // foreground; success lifts the freeze and turns the service back on.
+    const frozen = await readSdkMigrationBlock(config.dataDir).catch(
+      () => null,
+    );
+    if (frozen?.frozen !== undefined) {
+      const migration = await (
+        controls.migrateSdkData ?? migrateSdkDataBeforeStart
+      )({
+        coreCommand: config.coreCommand,
+        dataDir: config.dataDir,
+        runtimeBuild:
+          config.buildIdentity?.buildNumber ??
+          config.buildIdentity?.version ??
+          "unknown",
+        asJson,
+        caller: command === "restart" ? "cli-restart" : "cli-start",
+      });
+      if (migration.state === "deferred") {
+        const failure = sdkMigrationFailure(
+          command,
+          migration.code ?? "sdk_migration_incomplete",
+          config.dataDir,
+          "blocked",
+          migration.dataRestored === true,
+        );
+        process.stdout.write(
+          asJson
+            ? `${JSON.stringify({ ...failure, state: "blocked" })}\n`
+            : `${failure.message}\n`,
+        );
+        return EXIT_FAILED;
+      }
+    }
    const enableSnapshot = await setManagedServiceEnabled(config.dataDir, true);
    const restoreEnableState = () =>
      restoreManagedServiceEnabled(config.dataDir, enableSnapshot);
@@ -1186,6 +1225,28 @@ async function runCommand(
          : state === "already_running"
            ? "MemHub Runtime 已在运行。\n"
            : "Runtime 服务已注册但未就绪；请运行 memhub logs 查看原因。\n";
+    // A Runtime that did not come up because its service refuses to serve an
+    // unmigrated home reports that cause, the migration log and where to send
+    // it, instead of a generic budget timeout.
+    const migrationBlocked = async (): Promise<string | null> => {
+      const blocked = await runtimeMigrationBlock(config.dataDir, command);
+      if (blocked === null) return null;
+      return asJson
+        ? `${JSON.stringify({
+            schema_version: 1,
+            product: PRODUCT,
+            command,
+            state: "blocked",
+            code: blocked.code,
+            message: blocked.message,
+            log_file: blocked.log_file,
+            issues_url: blocked.issues_url,
+          })}\n`
+        : `${blocked.message}\n`;
+    };
+    const notReady = async () =>
+      (await migrationBlocked()) ??
+      lifecycle("not-ready", "runtime_not_ready_within_budget");
    // A launchd-managed Runtime must be driven through launchctl. Spawning a
    // second process while the service holds the data directory is exactly what
    // produced "Runtime exited before becoming ready".
@@ -1230,16 +1291,12 @@ async function runCommand(
            process.stdout.write(lifecycle("restarted", "runtime_verified"));
            return EXIT_OK;
          }
-          process.stdout.write(
-            lifecycle("not-ready", "runtime_not_ready_within_budget"),
-          );
+          process.stdout.write(await notReady());
          await restoreEnableState();
          return EXIT_FAILED;
        }
        if (restarted === "not-ready") {
-          process.stdout.write(
-            lifecycle("not-ready", "runtime_not_ready_within_budget"),
-          );
+          process.stdout.write(await notReady());
          await restoreEnableState();
          return EXIT_FAILED;
        }
@@ -1295,16 +1352,12 @@ async function runCommand(
          process.stdout.write(lifecycle("started", "runtime_verified"));
          return EXIT_OK;
        }
-        process.stdout.write(
-          lifecycle("not-ready", "runtime_not_ready_within_budget"),
-        );
+        process.stdout.write(await notReady());
        await restoreEnableState();
        return EXIT_FAILED;
      }
      if (startedManaged === "not-ready") {
-        process.stdout.write(
-          lifecycle("not-ready", "runtime_not_ready_within_budget"),
-        );
+        process.stdout.write(await notReady());
        await restoreEnableState();
        return EXIT_FAILED;
      }
@@ -1333,6 +1386,11 @@ async function runCommand(
      // A spawned Runtime that exits before ready is a failure of this command,
      // not a crash: report it as a structured document (JSON) or a plain
      // message, with a non-zero exit, instead of letting it escape.
+      const blocked = await migrationBlocked();
+      if (blocked !== null) {
+        process.stdout.write(blocked);
+        return EXIT_FAILED;
+      }
      const message = error instanceof Error ? error.message : String(error);
      process.stdout.write(
        asJson
@@ -1529,6 +1587,24 @@ async function runCommand(
      return uiExitCode(result);
    };
    if (attachment === null) {
+      const migration = await runtimeMigrationBlock(config.dataDir, "ui");
+      if (migration !== null) {
+        emit(
+          asJson
+            ? `${JSON.stringify({
+                schema_version: 1,
+                product: PRODUCT,
+                command: "ui",
+                state: "blocked",
+                code: migration.code,
+                message: migration.message,
+                log_file: migration.log_file,
+                issues_url: migration.issues_url,
+              })}\n`
+            : `ui 未打开：${migration.code}\n${migration.message}\n`,
+        );
+        return uiExitCode({ state: "blocked", code: migration.code });
+      }
      return emitUi(
        { state: "blocked", code: "runtime_not_running" },
        `先运行 ${PRODUCT} start 启动 Runtime。`,
@@ -1633,7 +1709,7 @@ async function runCommand(
    });
    if (runtime.state !== "running") {
      const document: ContextSearchDocument = {
-        schema_version: 1,
+        schema_version: 2,
        product: PRODUCT,
        state: "unavailable",
        code: runtime.code,
@@ -1657,7 +1733,7 @@ async function runCommand(
          timeout_ms: parsed.timeoutMs,
        })
      : ({
-          schema_version: 1,
+          schema_version: 2,
          product: PRODUCT,
          state: "unavailable",
          code: attached.failure ?? "runtime_attach_failed",
@@ -2423,20 +2499,20 @@ async function runCommand(
      };
    }
    if (blocked || !reader) {
-      if (
-        parsed.kind === "update" ||
-        parsed.kind === "delete" ||
-        parsed.kind === "plan-add" ||
-        parsed.kind === "plan-existing" ||
-        parsed.kind === "apply"
-      ) {
-        const document: MemoryFlowDocument = {
-          schema_version: 1,
+      if (parsed.kind === "update" || parsed.kind === "delete") {
+        const document = {
+          schema_version: 2,
          product: PRODUCT,
          state: "unknown",
+          operation: parsed.kind,
+          memory_id: parsed.memoryId,
          code: blocked?.code ?? "runtime_attach_failed",
        };
-        emitMemoryFlow(document, asJson);
+        emit(
+          asJson
+            ? `${JSON.stringify(document)}\n`
+            : `记忆操作未完成：${document.state}（${document.code}）\n`,
+        );
        return runtime.state === "stopped" ? 3 : 1;
      }
      const document =
@@ -2530,73 +2606,7 @@ async function runCommand(
      );
      return outcome.state === "committed" ? EXIT_OK : EXIT_FAILED;
    }
-    if (parsed.kind === "plan-add") {
-      const content = await readMemoryContentFile(parsed.contentFile);
-      if (content === null) return usage();
-      const document = planMemoryAdd({
-        content,
-        memoryKind: parsed.memoryKind,
-        scopes: parsed.scopes,
-        evidenceIds: parsed.evidenceIds,
-      });
-      emitMemoryFlow(document, asJson);
-      return memoryFlowExitCode(document);
-    }
-    if (parsed.kind === "plan-existing") {
-      let content: string | undefined;
-      if (parsed.contentFile) {
-        const read = await readMemoryContentFile(parsed.contentFile);
-        if (read === null) return usage();
-        content = read;
-      }
-      const document = await planExistingMemoryMutation(reader, {
-        operation: parsed.operation,
-        memoryId: parsed.memoryId,
-        ...(content === undefined ? {} : { content }),
-      });
-      emitMemoryFlow(document, asJson);
-      return memoryFlowExitCode(document);
-    }
-    if (!operations) {
-      const document: MemoryFlowDocument = {
-        schema_version: 1,
-        product: PRODUCT,
-        state: "unknown",
-        code: "runtime_attach_failed",
-      };
-      emitMemoryFlow(document, asJson);
-      return EXIT_FAILED;
-    }
-    const file = await readBoundedJson(parsed.planFile);
-    if (file === null) return usage();
-    const plan = unwrapPlan(file);
-    const identity = inspectMemoryPlan(plan);
-    if (!identity) return usage();
-    let content: string | undefined;
-    if (parsed.contentFile) {
-      const read = await readMemoryContentFile(parsed.contentFile);
-      if (read === null) return usage();
-      content = read;
-    }
-    if (
-      (identity.needsContent && content === undefined) ||
-      (!identity.needsContent && content !== undefined) ||
-      !memoryPlanAcceptsContent(plan, content)
-    ) {
-      return usage();
-    }
-    const attempt = await claimPlanAttempt(parsed.planFile, {
-      schema_version: 1,
-      operation_id: identity.operationId,
-      plan_digest: identity.digest,
-    });
-    if (attempt === "invalid") return usage();
-    const document =
-      attempt === "prior"
-        ? await readMemoryMutationReceipt(reader, plan)
-        : await executeMemoryMutation(plan, content, operations, reader);
-    emitMemoryFlow(document, asJson);
-    return memoryFlowExitCode(document);
+    return usage();
  }
  if (named === "calls" || named === "import-file") {
    const parsed = parseUiWriteCommand(named, arguments_.slice(1));
@@ -2806,10 +2816,15 @@ async function runCommand(
        if (!ready) throw new Error("runtime_service_activation_unverified");
        return plan;
      })
-      .catch((error: unknown) => {
+      .catch(async (error: unknown) => {
+        const migration =
+          error instanceof Error &&
+          error.message === "runtime_service_activation_unverified"
+            ? await runtimeMigrationBlock(layout.dataDir, "service-install")
+            : null;
        process.stderr.write(
          `${JSON.stringify(
-            serviceErrorDocument("service-install", error),
+            migration ?? serviceErrorDocument("service-install", error),
            null,
            2,
          )}\n`,
@@ -3143,6 +3158,11 @@ async function runCommand(
      null;
    let installReceiptWritten = false;
    let serviceApplied = false;
+    let migrated = false;
+    // `null` until the migration is handed this home; "failed" if it threw.
+    let migrationState: SdkMigrationResult["state"] | "failed" | null = null;
+    let deferredCode: string | null = null;
+    let deferredRestored = false;
    let previousOwner:
      | Awaited<ReturnType<typeof claimCliServiceOwnership>>
      | undefined;
@@ -3153,7 +3173,39 @@ async function runCommand(
      }
      shellPathReceipt = await applyShellPath(shellPath);
      previousOwner = await claimCliServiceOwnership(layout.dataDir);
-      await applyPreparedServiceInstall(preparedService);
+      // `install` over an existing home (reinstall, App-to-CLI takeover) must
+      // migrate it too, before the new Runtime starts serving it.
+      migrationState = "failed";
+      const migration = await (
+        controls.migrateSdkData ?? migrateSdkDataBeforeStart
+      )({
+        coreCommand: layout.coreCommand,
+        dataDir: layout.dataDir,
+        runtimeBuild: String(plan.build_number ?? plan.version),
+        asJson: true,
+        caller: "cli-install",
+        sourceRuntime: await readInstalledRuntimeSource(
+          plan.install_root,
+          receipt.previous_current_target,
+        ),
+      });
+      migrationState = migration.state;
+      // A deferred migration keeps this install: the new Runtime service is
+      // registered and its migration gate keeps Core down until the home is
+      // migrated. Rolling back would hand the home to the previous (App or CLI)
+      // Runtime, which has no gate and would serve the legacy data.
+      if (migration.state === "deferred") {
+        deferredCode = migration.code ?? "sdk_migration_incomplete";
+        deferredRestored = migration.dataRestored === true;
+      }
+      migrated =
+        migration.state === "cutover" || migration.state === "complete";
+      // From here the previous LaunchAgent is never re-bootstrapped: it may
+      // be an older Runtime with no migration gate.
+      await applyPreparedServiceInstall({
+        ...preparedService,
+        restorePreviousOnFailure: false,
+      });
      serviceApplied = true;
      // The receipt exists so a later command can find the data dir without
      // flags. An isolated install named its data dir explicitly, and writing
@@ -3169,6 +3221,7 @@ async function runCommand(
        });
        installReceiptWritten = true;
      }
+      if (deferredCode !== null) throw new SdkMigrationDeferred(deferredCode);
      // Registering the service is not the same as a working install: on a host
      // where Core cold-start stalls, launchd reports a pid while status is
      // still `runtime_service_running`. Gate on real readiness (same predicate
@@ -3204,6 +3257,47 @@ async function runCommand(
          );
        }
      };
+      // Once the migration has been handed this home nothing rolls back: the
+      // previous version (App or CLI) may be an older Runtime without the
+      // migration gate, and restoring it could serve a home that is not on SDK
+      // authority. This install stays; its service keeps Core down until the
+      // home is migrated. Only a failure before the migration ran (the
+      // previous Runtime was never stopped and the service never touched) is
+      // rolled back below.
+      if (migrationState !== null) {
+        if (!serviceApplied && !migrated && deferredCode === null) {
+          // The migration threw. Hand the data dir to this version's gated
+          // service anyway so that the previous Runtime does not come back.
+          await applyPreparedServiceInstall({
+            ...preparedService,
+            restorePreviousOnFailure: false,
+          }).catch(() => undefined);
+        }
+        process.stderr.write(
+          `${JSON.stringify(
+            deferredCode !== null
+              ? sdkMigrationFailure(
+                  "install",
+                  deferredCode,
+                  layout.dataDir,
+                  "upgrade",
+                  deferredRestored,
+                )
+              : migrated
+                ? notReadyAfterMigration("install")
+                : migrationState === "failed"
+                  ? sdkMigrationFailure(
+                      "install",
+                      sdkMigrationErrorCode(error),
+                      layout.dataDir,
+                    )
+                  : notReadyAfterUpgrade("install", error),
+            null,
+            2,
+          )}\n`,
+        );
+        return EXIT_FAILED;
+      }
      if (serviceApplied) {
        await rollbackPreparedServiceInstall(
          preparedService,
@@ -3575,9 +3669,18 @@ async function runCommand(
    let previousOwner:
      | Awaited<ReturnType<typeof claimCliServiceOwnership>>
      | undefined;
-    const rollbackUpgrade = async (
-      dataDir: string | undefined,
-    ): Promise<void> => {
+    let migrated = false;
+    // `null` until the migration is handed this home; "failed" if it threw.
+    let migrationState: SdkMigrationResult["state"] | "failed" | null = null;
+    /**
+     * Restore the previous version's files and ownership. Only used for a
+     * failure before the migration ran: the previous Runtime was never stopped
+     * and the service never restarted, so nothing new starts serving. After
+     * the migration has been handed the home the upgrade never rolls back:
+     * the previous Runtime may predate the migration gate and would serve a
+     * home that is not on SDK authority.
+     */
+    const rollbackUpgrade = async (): Promise<void> => {
      for (const receipt of memhubReceipts.reverse()) {
        await rollbackManagedCliLink(receipt);
      }
@@ -3587,51 +3690,81 @@ async function runCommand(
          () => undefined,
        );
      }
-      await restartUserService(
-        runtimeInstallLayout({
-          platform: plan.platform,
-          home: parsed.home,
-          installRoot: canonicalManagedServiceRoot(
-            plan.platform,
-            plan.current_link,
-          ),
-          dataDir,
-          ...hourlyIsolationFromEnvironment(environment),
-        }),
-        serviceOptions,
-      ).catch(() => undefined);
+    };
+    /**
+     * After a migration that did not finish, make sure the process holding the
+     * data dir is the new version's service (gated: it keeps Core down while
+     * the migration is deferred) and not the previous Runtime. The migration
+     * fence already stopped it unless the status probe itself failed.
+     */
+    const keepUpgradedRuntimeOffLegacyHome = async (): Promise<void> => {
+      try {
+        await restartUserService(layout, serviceOptions);
+      } catch {
+        await (controls.stopRuntime ?? stopRuntimeProcess)({
+          dataDir: parsed.dataDir ?? layout.dataDir,
+          probeIdentity,
+        }).catch(() => undefined);
+      }
    };
    try {
      for (const link of companionLinks) {
        memhubReceipts.push(await installManagedCliLink(link));
      }
      previousOwner = await claimCliServiceOwnership(layout.dataDir);
+      migrationState = "failed";
+      const migration = await (
+        controls.migrateSdkData ?? migrateSdkDataBeforeStart
+      )({
+        coreCommand: layout.coreCommand,
+        dataDir: parsed.dataDir ?? layout.dataDir,
+        runtimeBuild: String(plan.build_number ?? plan.version),
+        asJson,
+        caller: "cli-upgrade",
+        sourceRuntime: await readInstalledRuntimeSource(
+          plan.install_root,
+          receipt.previous_current_target,
+        ),
+      });
+      migrationState = migration.state;
+      if (migration.state === "deferred") {
+        // Stay on the new version. Its Runtime service refuses to spawn Core
+        // while the migration is deferred, so nothing serves the legacy home;
+        // rolling back would restart the old Runtime, which has no such gate.
+        await keepUpgradedRuntimeOffLegacyHome();
+        process.stderr.write(
+          `${JSON.stringify(sdkMigrationFailure("upgrade", migration.code ?? "sdk_migration_incomplete", parsed.dataDir ?? layout.dataDir, "upgrade", migration.dataRestored === true), null, 2)}\n`,
+        );
+        return EXIT_FAILED;
+      }
+      // From here a failure keeps the new version: an older Runtime either
+      // refuses a migrated home or, predating the migration gate, would serve
+      // one that is not on SDK authority.
+      migrated =
+        migration.state === "cutover" || migration.state === "complete";
      await restartUserService(layout, serviceOptions);
      // A restart that returns without error only means launchd/systemd
      // accepted the command. O-22: the first `upgrade` implementation
      // reported success even when the new artifact crash-looped, leaving
      // `current` on a version that never came up. Verify the restarted
-      // Runtime actually becomes healthy; if it does not, roll back to the
-      // previously working version.
+      // Runtime actually becomes healthy; if it does not, fail loudly. It is
+      // not rolled back (product rule: SDK authority or not serving at all).
      const dataDir = parsed.dataDir ?? layout.dataDir;
      const ready = await (controls.waitForRuntimeReady ?? waitForRuntimeReady)(
        dataDir,
      );
-      if (!ready) {
-        await rollbackUpgrade(parsed.dataDir);
+      if (!ready && migrated) {
        process.stderr.write(
-          `${JSON.stringify(
-            {
-              schema_version: 1,
-              product: PRODUCT,
-              command: "upgrade",
-              status: "error",
-              code: "upgrade_health_check_failed",
-              message: "新版本重启后未在限定时间内就绪，已回滚到上一版本。",
-            },
-            null,
-            2,
-          )}\n`,
+          `${JSON.stringify(notReadyAfterMigration("upgrade"), null, 2)}\n`,
+        );
+        return EXIT_FAILED;
+      }
+      if (!ready) {
+        // Not rolled back: see `rollbackUpgrade`. If the service is blocked on
+        // the migration (it re-runs it before spawning Core), say so.
+        const blocked = await runtimeMigrationBlock(dataDir, "upgrade");
+        process.stderr.write(
+          `${JSON.stringify(blocked ?? notReadyAfterUpgrade("upgrade"), null, 2)}\n`,
        );
        return EXIT_FAILED;
      }
@@ -3644,9 +3777,32 @@ async function runCommand(
        });
      }
    } catch (error) {
-      await rollbackUpgrade(parsed.dataDir);
+      if (migrationState === null) {
+        await rollbackUpgrade();
+        process.stderr.write(
+          `${JSON.stringify(serviceErrorDocument("upgrade", error), null, 2)}\n`,
+        );
+        return EXIT_FAILED;
+      }
+      if (migrated) {
+        // An older Runtime refuses a migrated home: keep the new version.
+        process.stderr.write(
+          `${JSON.stringify(notReadyAfterMigration("upgrade"), null, 2)}\n`,
+        );
+        return EXIT_FAILED;
+      }
+      if (migrationState === "failed") {
+        // The migration itself threw: its outcome is unknown, so treat the
+        // home as not migrated and keep the old Runtime off it.
+        await keepUpgradedRuntimeOffLegacyHome();
+        process.stderr.write(
+          `${JSON.stringify(sdkMigrationFailure("upgrade", sdkMigrationErrorCode(error), parsed.dataDir ?? layout.dataDir), null, 2)}\n`,
+        );
+        return EXIT_FAILED;
+      }
+      // `not_required`: the restart itself failed. Stay on the new version.
      process.stderr.write(
-        `${JSON.stringify(serviceErrorDocument("upgrade", error), null, 2)}\n`,
+        `${JSON.stringify(notReadyAfterUpgrade("upgrade", error), null, 2)}\n`,
      );
      return EXIT_FAILED;
    }
@@ -4014,11 +4170,173 @@ export function runtimeIsReady(
  );
 }

+const SDK_MIGRATION_STEPS: Record<string, string> = {
+  bundle: "正在备份记忆数据，数据较多时需要几分钟……",
+  prepare: "正在整理待处理的写入任务……",
+  cutover: "正在切换记忆数据源……",
+  cleanup: "正在清理旧数据表……",
+};
+
+/**
+ * Migrate a home written by an older release before the new Runtime serves it.
+ *
+ * Runs the new Runtime's own Core with the Runtime stopped. Progress goes to
+ * stderr so stdout stays one result document.
+ */
+async function migrateSdkDataBeforeStart(options: {
+  coreCommand: string;
+  dataDir: string;
+  runtimeBuild: string;
+  asJson: boolean;
+  caller?: "cli-install" | "cli-upgrade" | "cli-start" | "cli-restart";
+  sourceRuntime?: SdkMigrationSource | null;
+}): Promise<SdkMigrationResult> {
+  return runSdkMigration({
+    dataDir: options.dataDir,
+    core: { command: options.coreCommand },
+    runtimeBuild: options.runtimeBuild,
+    ...(options.caller === undefined ? {} : { caller: options.caller }),
+    ...(options.sourceRuntime === undefined
+      ? {}
+      : { sourceRuntime: options.sourceRuntime }),
+    onProgress: (step) => {
+      process.stderr.write(
+        options.asJson
+          ? `${JSON.stringify({ schema_version: 1, product: PRODUCT, event: "sdk_migration.progress", step })}\n`
+          : `${SDK_MIGRATION_STEPS[step] ?? step}\n`,
+      );
+    },
+  });
+}
+
+class SdkMigrationDeferred extends Error {
+  constructor(readonly code: string) {
+    super(code);
+  }
+}
+
+function notReadyAfterMigration(command: "install" | "upgrade") {
+  return {
+    schema_version: 1,
+    product: PRODUCT,
+    command,
+    status: "error",
+    code: "runtime_not_ready_after_sdk_migration",
+    message:
+      "记忆数据已迁移到新版本，但新版本 Runtime 没有在限定时间内就绪。旧版本无法打开已迁移的数据，所以没有回滚。请运行 memhub restart，或 memhub doctor 查看原因。",
+  };
+}
+
+/**
+ * The new version did not come up and was not rolled back: the previous one
+ * may predate the SDK migration gate. `error` carries the underlying code
+ * (a failed service restart), never free text from a child process.
+ */
+function notReadyAfterUpgrade(command: "install" | "upgrade", error?: unknown) {
+  return {
+    schema_version: 1,
+    product: PRODUCT,
+    command,
+    status: "error",
+    code:
+      error !== undefined
+        ? serviceErrorDocument(command, error).code
+        : command === "upgrade"
+          ? "upgrade_health_check_failed"
+          : "install_runtime_not_ready",
+    rolled_back: false,
+    message:
+      "新版本 Runtime 没有启动或没有在限定时间内就绪。为避免旧版本打开记忆数据，已保留新版本，没有回滚。请运行 memhub restart 重试，或运行 memhub doctor 查看原因；如果仍然失败，请在 GitHub 提交反馈：" +
+      MIGRATION_ISSUES_URL,
+    issues_url: MIGRATION_ISSUES_URL,
+  };
+}
+
+const SDK_MIGRATION_FAILURES: Record<string, string> = {
+  sdk_migration_disk_space_insufficient:
+    "磁盘空间不足，记忆数据迁移没有完成。请腾出与记忆数据同等大小的空间后运行 memhub restart 重试。",
+  sdk_migration_memory_unreadable:
+    "核对发现新版本还读不到你的全部记忆，没有切换数据源。",
+  sdk_migration_model_profile_required:
+    "迁移需要先有可用的模型配置，用来核对新版本能否读回记忆。请先用 memhub models setup 完成配置，再运行 memhub restart 重试。",
+};
+
+const MIGRATION_ISSUES_URL =
+  "https://github.com/Tencent-Hunyuan/Hy-Memory/issues/new";
+
+/**
+ * Why MemHub is not serving: the home has not reached SDK authority.
+ *
+ * Product rule: after an upgrade the home is on SDK authority or MemHub does
+ * not serve it at all. Nothing here rolls back to an older Runtime (it has no
+ * migration gate and would serve the legacy home); the user gets the code, the
+ * migration log and where to report it.
+ */
+function sdkMigrationFailure(
+  command: string,
+  code: string,
+  dataDir: string,
+  context: "upgrade" | "blocked" = "upgrade",
+  dataRestored = false,
+) {
+  const logFile = sdkMigrationLogPath(dataDir);
+  // Core put the home back exactly as it was before the upgrade, so the next
+  // (fixed) version migrates from the original data.
+  const restored = dataRestored ? "记忆数据已恢复到升级前的状态。" : "";
+  const reason =
+    SDK_MIGRATION_FAILURES[code] ??
+    "记忆数据迁移没有完成。可以运行 memhub restart 重试。";
+  const lead =
+    context === "upgrade"
+      ? `${reason}已保留新版本；在迁移完成前 MemHub 不会启动，也不会用旧版本打开未迁移的数据。`
+      : `MemHub 没有启动：记忆数据还没有迁移到新版本（${code}）。${reason}`;
+  return {
+    schema_version: 1,
+    product: PRODUCT,
+    command,
+    status: "error",
+    code,
+    message: `${lead}原有记忆数据不会丢失。${restored}迁移日志：${logFile}。请在 GitHub 提交反馈并附上这个日志文件：${MIGRATION_ISSUES_URL}`,
+    log_file: logFile,
+    issues_url: MIGRATION_ISSUES_URL,
+    ...(dataRestored ? { data_restored: true } : {}),
+  };
+}
+
+/** A stable code from a thrown migration error, never its free text. */
+function sdkMigrationErrorCode(error: unknown): string {
+  const message = error instanceof Error ? error.message : "";
+  return /^[a-z][a-z0-9_]{0,95}$/.test(message)
+    ? message
+    : "sdk_migration_incomplete";
+}
+
+/**
+ * The Runtime service refuses to spawn Core while this home's migration is
+ * deferred and leaves a block note; report that instead of a generic
+ * "not ready within budget".
+ */
+async function runtimeMigrationBlock(
+  dataDir: string,
+  command: string,
+): Promise<ReturnType<typeof sdkMigrationFailure> | null> {
+  const block = await readSdkMigrationBlock(dataDir).catch(() => null);
+  if (block === null) return null;
+  return sdkMigrationFailure(
+    command,
+    block.code,
+    dataDir,
+    "blocked",
+    block.data_restored === true,
+  );
+}
+
 async function waitForRuntimeReady(
  dataDir: string,
  timeoutMs = 45_000,
 ): Promise<boolean> {
-  const deadline = Date.now() + timeoutMs;
+  const started = Date.now();
+  const deadline = started + timeoutMs;
  while (Date.now() < deadline) {
    const runtime = await inspectRuntimeStatus({
      dataDir,
@@ -4026,6 +4344,17 @@ async function waitForRuntimeReady(
      requireServiceReceipt: true,
    }).catch(() => null);
    if (runtimeIsReady(runtime)) return true;
+    // The service deferred its migration during this wait and will not spawn
+    // Core: stop waiting. An older note may predate a retry still running, so
+    // only a note written after the wait began ends it early.
+    // A frozen home: its service exits without migrating, so nothing will
+    // come up until a foreground migration lifts the freeze.
+    const blocked = await readSdkMigrationBlock(dataDir).catch(() => null);
+    if (
+      blocked !== null &&
+      (blocked.frozen !== undefined || Date.parse(blocked.at) >= started)
+    )
+      return false;
    await new Promise((resolve) => setTimeout(resolve, 500));
  }
  return false;
@@ -4730,10 +5059,10 @@ async function runMaintenanceRestore(options: {
  const instanceId = randomUUID();
  const child = spawn(options.coreCommand, ["serve", "--maintenance"], {
    env: {
-      // Do not pass the caller's Runtime tuning variables (for example
-      // retrieval-index flags) into a maintenance process: a restore must
-      // start the product's minimal maintenance configuration, not inherit
-      // an unrelated experiment from the shell.
+      // Do not pass the caller's Runtime tuning variables into a
+      // maintenance process: a restore must start the product's minimal
+      // maintenance configuration, not inherit an unrelated experiment from
+      // the shell.
      PATH: process.env.PATH,
      HOME: process.env.HOME,
      LANG: process.env.LANG,
@@ -5945,21 +6274,7 @@ type MemoryCommand =
      contentFile: string;
      confirmed: boolean;
    }
-  | { kind: "delete"; memoryId: string; confirmed: boolean }
-  | {
-      kind: "plan-add";
-      contentFile: string;
-      memoryKind: string;
-      scopes: string[];
-      evidenceIds: string[];
-    }
-  | {
-      kind: "plan-existing";
-      operation: "correct" | "delete";
-      memoryId: string;
-      contentFile?: string;
-    }
-  | { kind: "apply"; planFile: string; contentFile?: string };
+  | { kind: "delete"; memoryId: string; confirmed: boolean };

 async function parseMemoryCommand(
  rest: readonly string[],
@@ -5999,31 +6314,6 @@ async function parseMemoryCommand(
      ? { kind: "delete", memoryId, confirmed }
      : null;
  }
-  if (action === "plan" || action === "apply") return null;
-  if (action === "plan") return parseMemoryPlanCommand(tokens.slice(1));
-  if (action === "apply") {
-    let planFile: string | undefined;
-    let contentFile: string | undefined;
-    let confirmed = false;
-    for (let index = 1; index < tokens.length; index += 1) {
-      const flag = tokens[index];
-      if (flag === "--confirm") {
-        if (confirmed) return null;
-        confirmed = true;
-        continue;
-      }
-      const value = tokens[index + 1];
-      if (value === undefined || value.startsWith("--")) return null;
-      if (flag === "--plan-file" && planFile === undefined) planFile = value;
-      else if (flag === "--content-file" && contentFile === undefined)
-        contentFile = value;
-      else return null;
-      index += 1;
-    }
-    return planFile && confirmed
-      ? { kind: "apply", planFile, ...(contentFile ? { contentFile } : {}) }
-      : null;
-  }
  if (action !== "list") return null;
  let status: string | undefined;
  let offset = 0;
@@ -6050,73 +6340,11 @@ async function parseMemoryCommand(
  return { kind: "list", query: { status, offset, limit } };
 }

-function parseMemoryPlanCommand(
-  tokens: readonly string[],
-): MemoryCommand | null {
-  const operation = tokens[0];
-  const add = operation === "add";
-  if (!add && !["correct", "delete"].includes(operation)) {
-    return null;
-  }
-  const memoryId = add ? undefined : tokens[1];
-  if (!add && (!memoryId || !isMemoryId(memoryId))) return null;
-  let contentFile: string | undefined;
-  let memoryKind = "chat";
-  let kindSet = false;
-  const scopes: string[] = [];
-  const evidenceIds: string[] = [];
-  for (let index = add ? 1 : 2; index < tokens.length; index += 1) {
-    const flag = tokens[index];
-    const value = tokens[index + 1];
-    if (value === undefined || value.startsWith("--")) return null;
-    if (flag === "--content-file" && contentFile === undefined)
-      contentFile = value;
-    else if (flag === "--kind" && add && !kindSet) {
-      memoryKind = value;
-      kindSet = true;
-    } else if (flag === "--scope" && add && !scopes.includes(value))
-      scopes.push(value);
-    else if (flag === "--evidence" && add && !evidenceIds.includes(value))
-      evidenceIds.push(value);
-    else return null;
-    index += 1;
-  }
-  if (add) {
-    return contentFile && scopes.length > 0 && evidenceIds.length > 0
-      ? {
-          kind: "plan-add",
-          contentFile,
-          memoryKind,
-          scopes,
-          evidenceIds,
-        }
-      : null;
-  }
-  if (
-    (operation === "correct" && !contentFile) ||
-    (operation === "delete" && (contentFile !== undefined || scopes.length > 0))
-  ) {
-    return null;
-  }
-  return {
-    kind: "plan-existing",
-    operation: operation as "correct" | "delete",
-    memoryId: memoryId as string,
-    ...(contentFile ? { contentFile } : {}),
-  };
-}
-
 async function readMemoryContentFile(file: string): Promise<string | null> {
  const bytes = await readBoundedFile(file, 100_000 * 4 + 2);
  return bytes ? parseMemoryContent(bytes) : null;
 }

-function emitMemoryFlow(document: MemoryFlowDocument, asJson: boolean): void {
-  emit(
-    asJson ? `${JSON.stringify(document)}\n` : renderMemoryFlowText(document),
-  );
-}
-
 function emitModelWriteDocument(
  document: ModelWriteDocument,
  asJson: boolean,
@@ -6289,6 +6517,8 @@ function serviceErrorDocument(
      "服务文件不属于当前 MemHub 安装；请保留文件并确认其来源后再处理。",
    runtime_service_managed_upgrade_failed:
      "新服务启动失败，旧 MemHub 服务已恢复；检查新安装日志后重试。",
+    runtime_service_replacement_failed:
+      "新服务启动失败；记忆数据尚未处于新版本可用的状态，所以没有恢复旧服务。请运行 memhub doctor 检查服务状态。",
    runtime_service_managed_upgrade_rollback_failed:
      "新服务启动失败且旧服务恢复失败；请运行 memhub doctor 检查服务状态。",
    runtime_service_file_modified:
@@ -387,7 +387,7 @@ export const COMMANDS: readonly CommandDescriptor[] = [
      {
        flag: "--content-file <path>",
        summary:
-          "从有界普通文件读取新增/纠正正文；计划只保存摘要，执行时再次校验。",
+          "从有界普通文件读取 update 的新正文；预览只显示摘要，--confirm 时才写入。",
      },
      DIRECT_CONFIRM_OPTION,
      JSON_OPTION,
@@ -485,7 +485,7 @@ export const COMMANDS: readonly CommandDescriptor[] = [
  {
    name: "recall",
    summary:
-      "从当前授权的原文与治理记忆统一检索上下文；本地只读，不调用模型、不触发补索引。",
+      "从当前授权的原文与已沉淀记忆统一检索上下文；本地只读，不调用模型、不触发补索引。",
    usage:
      "memhub recall <query> [--kind <raw|memory>] [--source <id>] [--limit <n>] [--character-budget <n>] [--timeout-ms <n>] [--json]\n            memhub recall --query-file <path> [...]\n            memhub recall --stdin [...]",
    options: [
@@ -43,61 +43,35 @@ const ATTACHED = {
 };

 const RESPONSE = {
-  schema_version: 1,
+  schema_version: 2,
  request_id: "context-request-1",
  state: "grounded",
-  paths: {
-    raw: {
-      status: "ok",
-      reason: null,
-      elapsed_ms: 2,
-      returned_count: 1,
-      truncated: false,
-      truncated_by: [],
-    },
-    memory: {
-      status: "ok",
-      reason: null,
-      elapsed_ms: 3,
-      returned_count: 1,
-      truncated: false,
-      truncated_by: [],
-    },
-  },
  items: [
    {
      kind: "raw",
-      source_id: "codex",
-      conversation_id: "conversation-1",
-      anchor_evidence_id: "evidence-1",
-      messages: [
-        {
-          evidence_id: "evidence-1",
-          revision: 1,
-          role: "user",
-          written_by: "user",
-          created_at: "2026-09-19T00:00:00Z",
-          text: "Synthetic raw citation.",
-          truncated: false,
-        },
-      ],
+      raw_id: "raw-1",
+      content: "Synthetic raw citation.",
+      source: "codex",
+      session_id: "session-1",
+      memory_at: "2026-09-19T00:00:00Z",
+      score: 0.8,
      truncated: false,
    },
    {
      kind: "memory",
      memory_id: "memory-1",
-      revision: 2,
-      content: "Synthetic governed memory.",
-      memory_kind: "working",
+      content: "Synthetic SDK memory.",
+      layer: "l2_fact",
+      status: "active",
+      source: "codex",
      updated_at: "2026-09-19T00:00:01Z",
-      evidence_ids: ["evidence-1"],
+      score: 0.7,
      truncated: false,
    },
  ],
-  used_characters: 49,
+  used_characters: 44,
  truncated: false,
  total_ms: 5,
-  adoption: "unknown",
 };

 beforeEach(() => {
@@ -115,42 +89,17 @@ beforeEach(() => {
 });

 describe("unified context search", () => {
-  it("accepts an empty tool message in a valid partial context response", async () => {
+  it("accepts an empty raw item in a valid partial context response", async () => {
    transport.post.mockResolvedValue({
      status: 200,
      data: {
        ...RESPONSE,
        state: "partial",
-        paths: {
-          raw: {
-            ...RESPONSE.paths.raw,
-            status: "degraded",
-            reason: "lexical_only",
-            returned_count: 1,
-          },
-          memory: {
-            ...RESPONSE.paths.memory,
-            status: "empty",
-            returned_count: 0,
-          },
-        },
        items: [
          {
-            kind: "raw",
-            source_id: "workbuddy",
-            conversation_id: "conversation-1",
-            anchor_evidence_id: "evidence-1",
-            messages: [
-              {
-                evidence_id: "evidence-1",
-                revision: 1,
-                role: "tool",
-                written_by: "tool",
-                created_at: "2026-10-05T20:48:12Z",
-                text: "",
-                truncated: true,
-              },
-            ],
+            ...RESPONSE.items[0],
+            source: "workbuddy",
+            content: "",
            truncated: true,
          },
        ],
@@ -165,8 +114,9 @@ describe("unified context search", () => {
      streams.restore();
    }
    expect(JSON.parse(streams.stdout())).toMatchObject({
+      schema_version: 2,
      state: "partial",
-      items: [{ kind: "raw", messages: [{ text: "" }] }],
+      items: [{ kind: "raw", content: "" }],
    });
  });

@@ -178,12 +128,12 @@ describe("unified context search", () => {
      streams.restore();
    }
    const output = streams.stdout();
-    expect(output).toContain("Synthetic governed memory.");
+    expect(output).toContain("Synthetic SDK memory.");
    expect(output).toContain("Synthetic raw citation.");
-    expect(output).not.toMatch(/^- raw |memory-1|evidence-1|采用状态/m);
+    expect(output).not.toMatch(/^- raw |memory-1|raw-1|采用状态/m);
  });

-  it("queries the current read-only context endpoint and preserves citations", async () => {
+  it("queries the SDK context endpoint directly and preserves item ids", async () => {
    const streams = captureStreams();
    try {
      expect(
@@ -209,7 +159,7 @@ describe("unified context search", () => {
      streams.restore();
    }
    expect(transport.post).toHaveBeenCalledWith(
-      "/api/v1/context/search",
+      "/api/v2/context/search",
      {
        query: "synthetic question",
        kinds: ["raw", "memory"],
@@ -222,12 +172,18 @@ describe("unified context search", () => {
    );
    const output = JSON.parse(streams.stdout());
    expect(output).toMatchObject({
+      schema_version: 2,
      state: "grounded",
      request_id: "context-request-1",
-      adoption: "unknown",
    });
-    expect(output.items[0].messages[0].evidence_id).toBe("evidence-1");
-    expect(output.items[1].evidence_ids).toEqual(["evidence-1"]);
+    expect(output).not.toHaveProperty("paths");
+    expect(output.items[0]).toMatchObject({ kind: "raw", raw_id: "raw-1" });
+    expect(output.items[1]).toMatchObject({
+      kind: "memory",
+      memory_id: "memory-1",
+      layer: "l2_fact",
+    });
+    // No authority pre-query: released Core always serves the SDK data plane.
    expect(transport.get).not.toHaveBeenCalled();
    expect(transport.put).not.toHaveBeenCalled();
    expect(transport.patch).not.toHaveBeenCalled();
@@ -242,19 +198,8 @@ describe("unified context search", () => {
      status: 200,
      data: {
        ...RESPONSE,
-        paths: {
-          ...RESPONSE.paths,
-          raw: {
-            status: "not_requested",
-            reason: "not_requested",
-            elapsed_ms: 0,
-            returned_count: 0,
-            truncated: false,
-            truncated_by: [],
-          },
-        },
        items: [RESPONSE.items[1]],
-        used_characters: 26,
+        used_characters: 21,
      },
    });
    const streams = captureStreams();
@@ -394,29 +339,7 @@ describe("unified context search", () => {
      status: 200,
      data: {
        ...RESPONSE,
-        paths: {
-          ...RESPONSE.paths,
-          memory: {
-            ...RESPONSE.paths.memory,
-            status: "empty",
-            returned_count: 0,
-          },
-        },
-        items: [
-          {
-            ...(RESPONSE.items[0] as Record<string, unknown>),
-            messages: [
-              {
-                ...(
-                  RESPONSE.items[0] as {
-                    messages: Array<Record<string, unknown>>;
-                  }
-                ).messages[0],
-                text,
-              },
-            ],
-          },
-        ],
+        items: [{ ...RESPONSE.items[0], content: text }],
        used_characters: 10,
      },
    });
@@ -429,7 +352,7 @@ describe("unified context search", () => {
    expect(JSON.parse(streams.stdout())).toMatchObject({
      state: "grounded",
      used_characters: 10,
-      items: [{ kind: "raw", messages: [{ text }] }],
+      items: [{ kind: "raw", content: text }],
    });
  });

@@ -441,29 +364,7 @@ describe("unified context search", () => {
      status: 200,
      data: {
        ...RESPONSE,
-        paths: {
-          ...RESPONSE.paths,
-          memory: {
-            ...RESPONSE.paths.memory,
-            status: "empty",
-            returned_count: 0,
-          },
-        },
-        items: [
-          {
-            ...(RESPONSE.items[0] as Record<string, unknown>),
-            messages: [
-              {
-                ...(
-                  RESPONSE.items[0] as {
-                    messages: Array<Record<string, unknown>>;
-                  }
-                ).messages[0],
-                text,
-              },
-            ],
-          },
-        ],
+        items: [{ ...RESPONSE.items[0], content: text }],
        used_characters: 12,
      },
    });
@@ -479,6 +380,40 @@ describe("unified context search", () => {
    });
  });

+  it.each([
+    [
+      "a raw item from a source that was not asked for",
+      ["--source", "codex"],
+      { ...RESPONSE.items[0], source: "cursor" },
+    ],
+    [
+      "a raw item when only memory was asked for",
+      ["--kind", "memory"],
+      RESPONSE.items[0],
+    ],
+    ["a score outside 0..1", [], { ...RESPONSE.items[1], score: 1.5 }],
+  ])("rejects %s as an invalid response", async (_label, flags, item) => {
+    transport.post.mockResolvedValue({
+      status: 200,
+      data: {
+        ...RESPONSE,
+        items: [item],
+        used_characters: Array.from(String(item.content)).length,
+      },
+    });
+    const streams = captureStreams();
+    try {
+      expect(await run(["recall", "query", ...flags, "--json"])).toBe(1);
+    } finally {
+      streams.restore();
+    }
+    expect(JSON.parse(streams.stdout())).toMatchObject({
+      schema_version: 2,
+      state: "invalid",
+      code: "response_invalid",
+    });
+  });
+
  it("reports endpoint denial without exposing the response body", async () => {
    transport.post.mockResolvedValue({
      status: 403,
@@ -3,31 +3,6 @@ import { PRODUCT } from "./version.js";

 export const CONTEXT_QUERY_MAX_CHARACTERS = 2_000;
 const CONTEXT_KINDS = ["raw", "memory"] as const;
-const PATH_STATES = [
-  "ok",
-  "empty",
-  "not_requested",
-  "not_authorized",
-  "unavailable",
-  "degraded",
-] as const;
-const PATH_REASONS = [
-  "not_requested",
-  "permission_required",
-  "source_revoked",
-  "read_unavailable",
-  "timeout_budget",
-  "evidence_changed",
-  "memory_changed",
-  "lexical_only",
-] as const;
-const LIMIT_REASONS = [
-  "count_budget",
-  "character_budget",
-  "timeout_budget",
-  "metadata_budget",
-] as const;
-
 export interface ContextSearchInput {
  query: string;
  kinds: Array<(typeof CONTEXT_KINDS)[number]>;
@@ -37,8 +12,9 @@ export interface ContextSearchInput {
  timeout_ms: number;
 }

+/** The CLI's JSON document mirrors Core's SDK-native context schema v2. */
 export interface ContextSearchDocument {
-  schema_version: 1;
+  schema_version: 2;
  product: string;
  state:
    | "grounded"
@@ -50,12 +26,10 @@ export interface ContextSearchDocument {
    | "invalid";
  code?: string;
  request_id?: string;
-  paths?: Record<string, unknown>;
  items: Array<Record<string, unknown>>;
  used_characters?: number;
  truncated?: boolean;
  total_ms?: number;
-  adoption?: "unknown";
 }

 export async function searchContext(
@@ -67,7 +41,7 @@ export async function searchContext(
  const validated = validateContextResponse(outcome.data, input);
  return (
    validated ?? {
-      schema_version: 1,
+      schema_version: 2,
      product: PRODUCT,
      state: "invalid",
      code: "response_invalid",
@@ -131,13 +105,7 @@ export function renderContextText(document: ContextSearchDocument): string {
        : "查到以下相关内容：",
  ];
  for (const item of document.items) {
-    if (item.kind === "memory") {
-      lines.push(`- ${String(item.content)}`);
-    } else {
-      for (const message of item.messages as Array<Record<string, unknown>>) {
-        lines.push(`- ${String(message.text)}`);
-      }
-    }
+    lines.push(`- ${String(item.content)}`);
  }
  if (document.truncated) lines.push("结果已按请求预算截断。");
  return `${lines.join("\n")}\n`;
@@ -149,84 +117,37 @@ function validateContextResponse(
 ): ContextSearchDocument | null {
  const response = record(value);
  if (
-    response?.schema_version !== 1 ||
+    response?.schema_version !== 2 ||
    !oneOf(response.state, ["grounded", "empty", "partial", "unavailable"]) ||
    !boundedText(response.request_id, 64) ||
-    !record(response.paths) ||
    !Array.isArray(response.items) ||
    response.items.length > request.limit ||
    !integer(response.used_characters, 0, request.character_budget) ||
    typeof response.truncated !== "boolean" ||
-    !number(response.total_ms, 0) ||
-    response.adoption !== "unknown"
+    !number(response.total_ms, 0)
  ) {
    return null;
  }
-  const paths = response.paths as Record<string, unknown>;
-  const rawPath = validatePath(paths.raw);
-  const memoryPath = validatePath(paths.memory);
-  if (!rawPath || !memoryPath) return null;
  const requested = new Set(request.kinds);
-  if (
-    (!requested.has("raw") && rawPath.status !== "not_requested") ||
-    (!requested.has("memory") && memoryPath.status !== "not_requested")
-  ) {
-    return null;
-  }
  const items: Array<Record<string, unknown>> = [];
  let characters = 0;
  for (const itemValue of response.items) {
    const item = validateItem(itemValue, request);
    if (!item || !requested.has(item.kind as "raw" | "memory")) return null;
    items.push(item);
-    characters +=
-      item.kind === "memory"
-        ? characterCount(item.content)
-        : (item.messages as Array<Record<string, unknown>>).reduce(
-            (sum, message) => sum + characterCount(message.text),
-            0,
-          );
+    characters += characterCount(item.content);
  }
  if (characters !== response.used_characters) return null;
  if (response.state === "empty" && items.length !== 0) return null;
  return {
-    schema_version: 1,
+    schema_version: 2,
    product: PRODUCT,
    state: response.state,
    request_id: response.request_id as string,
-    paths: { raw: rawPath, memory: memoryPath },
    items,
    used_characters: response.used_characters as number,
    truncated: response.truncated,
    total_ms: response.total_ms as number,
-    adoption: "unknown",
-  };
-}
-
-function validatePath(value: unknown): Record<string, unknown> | null {
-  const path = record(value);
-  if (
-    !path ||
-    !oneOf(path.status, PATH_STATES) ||
-    !number(path.elapsed_ms, 0) ||
-    !integer(path.returned_count, 0, 20) ||
-    typeof path.truncated !== "boolean" ||
-    !enumList(path.truncated_by, LIMIT_REASONS) ||
-    !(
-      path.reason === null ||
-      path.reason === undefined ||
-      oneOf(path.reason, PATH_REASONS)
-    )
-  ) {
-    return null;
-  }
-  return {
-    status: path.status,
-    reason: path.reason ?? null,
-    elapsed_ms: path.elapsed_ms,
-    returned_count: path.returned_count,
-    truncated: path.truncated,
-    truncated_by: path.truncated_by,
  };
 }

@@ -235,88 +156,65 @@ function validateItem(
  request: ContextSearchInput,
 ): Record<string, unknown> | null {
  const item = record(value);
-  if (!item || !oneOf(item.kind, CONTEXT_KINDS)) return null;
+  if (
+    !item ||
+    !oneOf(item.kind, CONTEXT_KINDS) ||
+    boundedText(item.content, 20_000, true) === null ||
+    !number(item.score, 0) ||
+    (item.score as number) > 1 ||
+    typeof item.truncated !== "boolean"
+  ) {
+    return null;
+  }
  if (item.kind === "memory") {
    if (
      !boundedText(item.memory_id, 512) ||
-      !integer(item.revision, 1) ||
-      boundedText(item.content, 20_000, true) === null ||
-      boundedText(item.memory_kind, 64, true) === null ||
-      !boundedText(item.updated_at, 64) ||
-      !textList(item.evidence_ids, 128, 512) ||
-      typeof item.truncated !== "boolean"
+      boundedText(item.layer, 64, true) === null ||
+      boundedText(item.status, 32, true) === null ||
+      !optionalText(item.source, 128) ||
+      !optionalText(item.updated_at, 64)
    ) {
      return null;
    }
    return {
      kind: "memory",
      memory_id: item.memory_id,
-      revision: item.revision,
      content: item.content,
-      memory_kind: item.memory_kind,
-      updated_at: item.updated_at,
-      evidence_ids: item.evidence_ids,
+      layer: item.layer,
+      status: item.status,
+      source: item.source ?? null,
+      updated_at: item.updated_at ?? null,
+      score: item.score,
      truncated: item.truncated,
    };
  }
  if (
-    !boundedText(item.source_id, 128) ||
+    !boundedText(item.raw_id, 512) ||
+    !boundedText(item.source, 128) ||
    (request.raw_source_ids !== undefined &&
-      !request.raw_source_ids.includes(item.source_id as string)) ||
-    !boundedText(item.conversation_id, 512) ||
-    !boundedText(item.anchor_evidence_id, 512) ||
-    !Array.isArray(item.messages) ||
-    item.messages.length < 1 ||
-    item.messages.length > 3 ||
-    typeof item.truncated !== "boolean"
+      !request.raw_source_ids.includes(item.source as string)) ||
+    !boundedText(item.session_id, 512) ||
+    !optionalText(item.memory_at, 64)
  ) {
    return null;
  }
-  const messages = item.messages.map(validateMessage);
-  if (messages.some((message) => message === null)) return null;
  return {
    kind: "raw",
-    source_id: item.source_id,
-    conversation_id: item.conversation_id,
-    anchor_evidence_id: item.anchor_evidence_id,
-    messages,
+    raw_id: item.raw_id,
+    content: item.content,
+    source: item.source,
+    session_id: item.session_id,
+    memory_at: item.memory_at ?? null,
+    score: item.score,
    truncated: item.truncated,
  };
 }

-function validateMessage(value: unknown): Record<string, unknown> | null {
-  const message = record(value);
-  if (
-    !message ||
-    !boundedText(message.evidence_id, 512) ||
-    !integer(message.revision, 1) ||
-    boundedText(message.role, 32, true) === null ||
-    boundedText(message.written_by, 32, true) === null ||
-    !(
-      message.created_at === null ||
-      boundedText(message.created_at, 64) !== null
-    ) ||
-    boundedText(message.text, 20_000, true) === null ||
-    typeof message.truncated !== "boolean"
-  ) {
-    return null;
-  }
-  return {
-    evidence_id: message.evidence_id,
-    revision: message.revision,
-    role: message.role,
-    written_by: message.written_by,
-    created_at: message.created_at,
-    text: message.text,
-    truncated: message.truncated,
-  };
-}
-
 function failed(
  outcome: Exclude<ReadOutcome<unknown>, { state: "ok" }>,
 ): ContextSearchDocument {
  return {
-    schema_version: 1,
+    schema_version: 2,
    product: PRODUCT,
    state: outcome.state,
    code: outcome.code,
@@ -374,27 +272,10 @@ function oneOf<const T extends readonly string[]>(
  return typeof value === "string" && values.includes(value as T[number]);
 }

-function enumList(
-  value: unknown,
-  values: readonly string[],
-): value is string[] {
+function optionalText(value: unknown, maximum: number): boolean {
  return (
-    Array.isArray(value) &&
-    value.length <= 4 &&
-    new Set(value).size === value.length &&
-    value.every((item) => oneOf(item, values))
-  );
-}
-
-function textList(
-  value: unknown,
-  maximumItems: number,
-  maximumText: number,
-): value is string[] {
-  return (
-    Array.isArray(value) &&
-    value.length <= maximumItems &&
-    new Set(value).size === value.length &&
-    value.every((item) => boundedText(item, maximumText) !== null)
+    value === null ||
+    value === undefined ||
+    boundedText(value, maximum) !== null
  );
 }
@@ -8,14 +8,12 @@ import axios from "axios";
 */
 const READ_ONLY_PATHS = [
  "/api/v1/call-events",
-  "/api/v1/home-summary",
  "/api/v1/import-jobs",
  "/api/v1/live-sync",
-  "/api/v1/memory-library",
  "/api/v1/memory-processing",
-  "/api/v1/memory-tasks",
  "/api/v1/model-profiles",
  "/api/v1/sources",
+  "/api/v2/home-summary",
  "/api/v2/memory-library",
  "/api/v2/user-profile",
 ] as const;
@@ -67,13 +65,10 @@ export function createUiSessionRequest(attachment: {
  };
 }

-const MEMORY_DETAIL = /^\/api\/v1\/memories\/[A-Za-z0-9][\w.-]{0,127}$/;
 const IMPORT_JOB =
  /^\/api\/v1\/import-jobs\/[A-Za-z0-9][\w.-]{0,127}(?:\/issues)?$/;
 const OPERATION_RECEIPT =
  /^\/api\/v1\/(?:source-connections|model-profile-creations|model-probe-operations)\/[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}$/;
-const MEMORY_MUTATION_RECEIPT =
-  /^\/api\/v1\/memory-mutations\/[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}$/;
 /**
 * Core registers these static routes before `/{memory_id}`, so a memory whose
 * id happens to be one of them would reach a different handler: `graph`
@@ -173,7 +168,7 @@ export function createCoreContextQuery(attachment: {
    let response: { status: number; data: unknown };
    try {
      response = await client.post(
-        "/api/v1/context/search",
+        "/api/v2/context/search",
        body,
        options?.signal ? { signal: options.signal } : undefined,
      );
@@ -197,10 +192,8 @@ export function isReadOnlyPath(path: string): boolean {
  if ((READ_ONLY_PATHS as readonly string[]).includes(path)) return true;
  if (IMPORT_JOB.test(path)) return true;
  if (OPERATION_RECEIPT.test(path)) return true;
-  if (MEMORY_MUTATION_RECEIPT.test(path)) return true;
-  if (SDK_MEMORY_DETAIL.test(path)) return true;
-  if (!MEMORY_DETAIL.test(path)) return false;
-  const segment = path.slice("/api/v1/memories/".length);
+  if (!SDK_MEMORY_DETAIL.test(path)) return false;
+  const segment = path.slice("/api/v2/memories/".length);
  return !(RESERVED_MEMORY_SEGMENTS as readonly string[]).includes(segment);
 }

@@ -1,18 +1,20 @@
 import type {
-  GovernedMemory,
-  HomeSummary,
  ImportIssue,
  ImportJob,
  LiveSyncHealth,
-  MemoryDetail,
-  MemoryEvidenceExcerpt,
-  MemoryLibraryPage,
  MemoryProcessingSummary,
  ModelProfile,
  SourceResponse,
 } from "@one-memory/contracts";
-/** The only memory states a response may carry, and the only ones a filter may ask for. */
-export const MEMORY_STATUSES = ["confirmed", "deleted", "superseded"] as const;
+/** SDK memory states `memhub memories list --status` may ask for. */
+export const MEMORY_STATUSES = [
+  "active",
+  "superseded",
+  "shadow",
+  "negated",
+  "conflicted",
+  "archived",
+] as const;

 /**
 * HTTP 200 is not a contract. Everything here re-checks the shape and the
@@ -23,7 +25,6 @@ export const MEMORY_STATUSES = ["confirmed", "deleted", "superseded"] as const;

 const ID_MAX = 128;
 const LABEL_MAX = 256;
-const TEXT_MAX = 100_000;
 const LIST_MAX = 500;
 const LIVE_SYNC_STATES = [
  "idle",
@@ -39,14 +40,6 @@ const PROVIDER_KINDS = [
  "local_onnx",
 ] as const;
 const CREDENTIAL_STATES = ["available", "missing", "unavailable"] as const;
-const MEMORY_KINDS = [
-  "chat",
-  "working",
-  "coding",
-  "schema",
-  "intention",
-] as const;
-const MEMORY_ORIGINS = ["extracted", "answered", "agent", "manual"] as const;
 const IMPORT_STATUSES = [
  "pending",
  "running",
@@ -57,16 +50,6 @@ const IMPORT_STATUSES = [
  "cancelled",
 ] as const;
 const SCAN_PHASES = ["discover", "read", "emit", "done"] as const;
-const MEMORY_TASK_KINDS = ["extract", "system2", "ultra", "sweeper"] as const;
-const MEMORY_TASK_STATUSES = [
-  "pending",
-  "running",
-  "waiting_for_model",
-  "completed",
-  "failed",
-  "cancelled",
-  "deferred_outside_window",
-] as const;
 const CAPABILITY_STATES = ["not_implemented", "implemented"] as const;
 const VERIFICATION_STATES = ["fixture_only", "real_replayed"] as const;
 const SYNC_COMPLETENESS = [
@@ -87,16 +70,6 @@ const LOCATION_STATES = [
 ] as const;
 const LIVE_SYNC_MODES = ["polling", "watch_and_polling"] as const;

-export interface ValidatedMemoryTask {
-  task_id: string;
-  kind: (typeof MEMORY_TASK_KINDS)[number];
-  status: (typeof MEMORY_TASK_STATUSES)[number];
-  attempts: number;
-  last_error_code: string | null;
-  created_at: string;
-  updated_at: string;
-}
-
 export interface SourceConnectionCandidate {
  source_id: string;
  display_name: string;
@@ -352,45 +325,6 @@ export function validateImportIssues(
  });
 }

-export function validateMemoryTasks(
-  value: unknown,
-): ValidatedMemoryTask[] | null {
-  return list(value, (item) => {
-    const task = record(item);
-    if (!task) return null;
-    const taskId = text(task.task_id, ID_MAX);
-    if (
-      task.last_error_code !== null &&
-      text(task.last_error_code, LABEL_MAX) === null
-    ) {
-      return null;
-    }
-    const error = task.last_error_code as string | null;
-    const createdAt = text(task.created_at, LABEL_MAX);
-    const updatedAt = text(task.updated_at, LABEL_MAX);
-    const attempts = count(task.attempts);
-    if (
-      taskId === null ||
-      createdAt === null ||
-      updatedAt === null ||
-      attempts === null ||
-      !oneOf(task.kind, MEMORY_TASK_KINDS) ||
-      !oneOf(task.status, MEMORY_TASK_STATUSES)
-    ) {
-      return null;
-    }
-    return {
-      task_id: taskId,
-      kind: task.kind,
-      status: task.status,
-      attempts,
-      last_error_code: error,
-      created_at: createdAt,
-      updated_at: updatedAt,
-    };
-  });
-}
-
 export function validateProcessing(
  value: unknown,
 ): MemoryProcessingSummary | null {
@@ -444,150 +378,32 @@ export function validateModelProfiles(value: unknown): ModelProfile[] | null {
  });
 }

-export function validateHomeSummary(value: unknown): HomeSummary | null {
+/** `GET /api/v2/home-summary`: the SDK-authority home card. */
+export interface SdkHomeSummary {
+  schema_version: 2;
+  memory_count: number;
+  latest: { memory_id: string } | null;
+  observed_at: string;
+}
+
+export function validateHomeSummary(value: unknown): SdkHomeSummary | null {
  const summary = record(value);
  if (
    !summary ||
+    summary.schema_version !== 2 ||
    count(summary.memory_count) === null ||
    text(summary.observed_at, LABEL_MAX) === null
  ) {
    return null;
  }
-  return summary as unknown as HomeSummary;
-}
-
-/** The page must be the page that was asked for, not merely a page. */
-export function validateMemoryPage(
-  value: unknown,
-  request: { offset: number; limit: number },
-): MemoryLibraryPage | null {
-  const page = record(value);
-  if (!page) return null;
-  const total = count(page.total);
-  if (
-    total === null ||
-    page.offset !== request.offset ||
-    page.limit !== request.limit
-  ) {
-    return null;
-  }
-  const items = list(page.items, validateGovernedMemory);
-  if (items === null || items.length > request.limit) return null;
-  if (items.length > 0 && total < request.offset + items.length) return null;
-  return { items, total, offset: page.offset, limit: page.limit };
-}
-
-export function validateMemoryDetail(
-  value: unknown,
-  memoryId: string,
-): MemoryDetail | null {
-  const detail = record(value);
-  if (!detail) return null;
-  const memory = validateGovernedMemory(detail.memory);
-  if (memory === null || memory.memory_id !== memoryId) return null;
-  const cited = new Set(memory.evidence_ids);
-  const returned = new Set<string>();
-  const evidence = list(detail.evidence ?? [], (item) => {
-    const excerpt = record(item);
-    if (!excerpt) return null;
-    const evidenceId = text(excerpt.evidence_id, ID_MAX);
-    // Core builds this list from the memory's own citations; anything else is
-    // another record's excerpt and must not be shown under this memory.
-    if (
-      evidenceId === null ||
-      !cited.has(evidenceId) ||
-      returned.has(evidenceId) ||
-      typeof excerpt.available !== "boolean"
-    ) {
+  const latest = summary.latest;
+  if (latest !== null) {
+    const memory = record(latest);
+    if (!memory || typeof memory.memory_id !== "string" || !memory.memory_id) {
      return null;
    }
-    returned.add(evidenceId);
-    const available = excerpt.available === true;
-    const verbatim = available ? text(excerpt.verbatim ?? "", TEXT_MAX) : "";
-    if (verbatim === null) return null;
-    return {
-      evidence_id: evidenceId,
-      available,
-      source_id: optional(excerpt.source_id, ID_MAX),
-      written_by: optional(excerpt.written_by, LABEL_MAX),
-      created_at: optional(excerpt.created_at, LABEL_MAX),
-      // An excerpt the user cannot read has no text, whatever the server sent.
-      verbatim,
-    } as unknown as MemoryEvidenceExcerpt;
-  });
-  if (evidence === null || evidence.length !== memory.evidence_ids.length) {
-    return null;
  }
-  const history = list(detail.history ?? [], (item) => {
-    const version = record(item);
-    if (
-      !version ||
-      typeof version.revision !== "number" ||
-      !Number.isSafeInteger(version.revision) ||
-      version.revision < 1 ||
-      text(version.action, LABEL_MAX) === null ||
-      !oneOf(version.status, MEMORY_STATUSES) ||
-      text(version.content, TEXT_MAX) === null ||
-      identifiers(version.scopes) === null ||
-      text(version.created_at, LABEL_MAX) === null
-    ) {
-      return null;
-    }
-    return {
-      revision: version.revision,
-      action: version.action,
-      status: version.status,
-      content: version.content,
-      scopes: identifiers(version.scopes) as string[],
-      created_at: version.created_at,
-    };
-  });
-  if (history === null) return null;
-  return {
-    memory,
-    evidence,
-    history,
-    index_sync: null,
-  } as unknown as MemoryDetail;
-}
-
-export function validateGovernedMemory(value: unknown): GovernedMemory | null {
-  const memory = record(value);
-  if (
-    !memory ||
-    !text(memory.memory_id, ID_MAX) ||
-    !text(memory.user_id, LABEL_MAX) ||
-    !text(memory.content, TEXT_MAX) ||
-    !oneOf(memory.kind, MEMORY_KINDS) ||
-    !oneOf(memory.origin, MEMORY_ORIGINS) ||
-    typeof memory.revision !== "number" ||
-    !Number.isSafeInteger(memory.revision) ||
-    memory.revision < 1 ||
-    !oneOf(memory.status, MEMORY_STATUSES) ||
-    !text(memory.created_at, LABEL_MAX) ||
-    !text(memory.updated_at, LABEL_MAX)
-  ) {
-    return null;
-  }
-  const scopes = identifiers(memory.scopes);
-  const evidenceIds = identifiers(memory.evidence_ids);
-  const tags = identifiers(memory.tags ?? []);
-  if (
-    scopes === null ||
-    scopes.length < 1 ||
-    evidenceIds === null ||
-    evidenceIds.length < 1 ||
-    tags === null ||
-    tags.length > 32
-  ) {
-    return null;
-  }
-  return {
-    ...memory,
-    scopes,
-    evidence_ids: evidenceIds,
-    tags,
-  } as unknown as GovernedMemory;
+  return summary as unknown as SdkHomeSummary;
 }

 function identifiers(value: unknown): string[] | null {
@@ -618,10 +434,6 @@ function text(value: unknown, max: number): string | null {
  return typeof value === "string" && value.length <= max ? value : null;
 }

-function optional(value: unknown, max: number): string | undefined {
-  return typeof value === "string" && value.length <= max ? value : undefined;
-}
-
 function count(value: unknown): number | null {
  return typeof value === "number" && Number.isSafeInteger(value) && value >= 0
    ? value
@@ -134,11 +134,6 @@ function progress(): ProgressDocument {
    generation: {
      summary_state: { state: "ok" },
      summary: { waiting_for_model: 1, failed: 0 },
-      tasks_state: { state: "ok" },
-      tasks: [{ task_id: "t1", status: "waiting_for_model" }],
-      returned: 1,
-      validated: 1,
-      output_truncated: false,
    },
  };
 }
@@ -65,6 +65,9 @@ export function renderDoctorText(document: DoctorDocument): string {
    model: "模型",
    pending_operations: "待处理操作",
  };
+  const migration = document.checks.find(
+    (check) => check.id === "core" && check.details?.log_file,
+  );
  return [
    `MemHub 检查：${doctorState(document.state)}`,
    ...document.checks.map(
@@ -73,7 +76,13 @@ export function renderDoctorText(document: DoctorDocument): string {
    ),
    ...(document.state === "healthy"
      ? []
-      : ["下一步：打开 MemHub 的“帮助与诊断”查看并处理异常。"]),
+      : migration
+        ? [
+            `升级迁移没有完成（错误码：${migration.code}），MemHub 不会启动。`,
+            `下一步：把迁移日志附在 GitHub 反馈里发给我们：${migration.details?.log_file}`,
+            `反馈地址：${migration.details?.issues_url}`,
+          ]
+        : ["下一步：打开 MemHub 的“帮助与诊断”查看并处理异常。"]),
    "",
  ].join("\n");
 }
@@ -107,6 +116,14 @@ function coreCheck(status: StatusDocument): DoctorCheck {
    id: "core",
    state: status.runtime.state === "stopped" ? "warn" : "fail",
    code: status.runtime.code,
+    ...(status.runtime.log_file
+      ? {
+          details: {
+            log_file: status.runtime.log_file,
+            issues_url: status.runtime.issues_url ?? "",
+          },
+        }
+      : {}),
  };
 }

@@ -200,8 +217,7 @@ function modelCheck(status: StatusDocument): DoctorCheck {
 function pendingCheck(progress: ProgressDocument | null): DoctorCheck {
  if (
    progress?.imports.state !== "ok" ||
-    progress.generation.summary_state.state !== "ok" ||
-    progress.generation.tasks_state.state !== "ok"
+    progress.generation.summary_state.state !== "ok"
  ) {
    return {
      id: "pending_operations",
@@ -212,27 +228,20 @@ function pendingCheck(progress: ProgressDocument | null): DoctorCheck {
  const jobAttention = progress.imports.jobs.filter((item) =>
    ["paused", "interrupted", "failed"].includes(text(item.status) ?? ""),
  ).length;
-  const taskAttention = progress.generation.tasks.filter((item) =>
-    ["waiting_for_model", "failed"].includes(text(item.status) ?? ""),
-  ).length;
  const summary = progress.generation.summary;
  const waiting = summary ? number(summary.waiting_for_model) : 0;
  const failed = summary ? number(summary.failed) : 0;
-  const attention = jobAttention + taskAttention + waiting + failed;
+  const attention = jobAttention + waiting + failed;
  return {
    id: "pending_operations",
    state: attention > 0 ? "warn" : "pass",
    code: attention > 0 ? "operations_need_attention" : "pending_operations_ok",
    details: {
      import_jobs_validated: progress.imports.validated,
-      generation_tasks_validated: progress.generation.validated,
      jobs_need_attention: jobAttention,
-      tasks_need_attention: taskAttention,
      waiting_for_model: waiting,
      failed,
-      output_truncated:
-        progress.imports.output_truncated ||
-        progress.generation.output_truncated,
+      output_truncated: progress.imports.output_truncated,
    },
  };
 }
@@ -1,4 +1,4 @@
-import { mkdtemp, rm } from "node:fs/promises";
+import { mkdtemp, rm, symlink } from "node:fs/promises";
 import os from "node:os";
 import path from "node:path";
 import { DiagnosticLogger } from "@one-memory/runtime-control";
@@ -41,11 +41,15 @@ it("redacts unsafe details and absolute stack frames from the read surface", ()
    session_id: "11111111-1111-4111-8111-111111111111",
    details: {
      count: 1,
+      reason: "/Users/me/private-query.txt",
+      status: "sk-live-secret-token",
+      source_id: "credential.private",
      prompt: "secret-token-123",
      authorization: "Bearer secret-token-123",
    } as unknown as Record<string, string | number | boolean | null>,
    safe_stack: [
      "packages/runtime-control/src/example.ts:42:7",
+      "packages/runtime/sk-live-secret-token.ts:1:1",
      "/Users/me/private.py:1",
      "C:\\Users\\me\\private.py:1",
    ],
@@ -56,6 +60,8 @@ it("redacts unsafe details and absolute stack frames from the read surface", ()
    "packages/runtime-control/src/example.ts:42:7",
  ]);
  expect(JSON.stringify(redacted)).not.toContain("secret-token-123");
+  expect(JSON.stringify(redacted)).not.toContain("sk-live-secret-token");
+  expect(JSON.stringify(redacted)).not.toContain("credential.private");
  expect(JSON.stringify(redacted)).not.toContain("/Users/me");
  expect(JSON.stringify(redacted)).not.toContain("C:\\\\Users");
 });
@@ -132,3 +138,27 @@ it("follows a newly appended event without stopping Runtime", async () => {
  });
  expect(seen).toEqual(["service.ready", "core.failed"]);
 });
+
+it("fails follow mode when the diagnostic directory is unsafe", async () => {
+  const dataDir = await mkdtemp(
+    path.join(os.tmpdir(), "om-cli-follow-unsafe-"),
+  );
+  const external = await mkdtemp(
+    path.join(os.tmpdir(), "om-cli-follow-external-"),
+  );
+  roots.push(dataDir, external);
+  await symlink(external, path.join(dataDir, "logs"));
+  const query = parseLogQuery(["--follow"]);
+  expect(query).not.toBeNull();
+  if (!query) throw new Error("query_missing");
+  await expect(
+    followLogs({
+      dataDir,
+      query,
+      signal: new AbortController().signal,
+      maxEvents: 1,
+      emit: () => undefined,
+      wait: async () => undefined,
+    }),
+  ).rejects.toThrow("log_directory_unreadable");
+});
@@ -3,6 +3,7 @@ import path from "node:path";
 import type { DiagnosticEventV1 } from "@one-memory/runtime-control";
 import {
  diagnosticLogFiles,
+  projectDiagnosticEventForSafeOutput,
  readDiagnosticEvents,
 } from "@one-memory/runtime-control";

@@ -24,54 +25,6 @@ export interface LogQuery {

 const ID = /^[A-Za-z0-9][A-Za-z0-9._:-]{0,255}$/;
 const NAME = /^[a-z][a-z0-9]*(?:[._-][a-z0-9]+){0,15}$/;
-const SAFE_DETAIL_KEYS = new Set([
-  "action",
-  "attempt",
-  "count",
-  "method",
-  "mode",
-  "phase",
-  "platform",
-  "reason",
-  "result_count",
-  "source_id",
-  "spawned",
-  "status",
-  "version",
-  "parent_session_id",
-  "invalid_lines",
-  "truncated",
-  "runtime_state",
-  "core_state",
-  "broker_state",
-  "instance_id",
-  "open_file_count",
-  "memory_bytes",
-  "thread_count",
-  "database_error_count",
-  "file_count",
-  "sqlite_file_count",
-  "byte_file_count",
-  "total_bytes",
-  "copied_bytes",
-  "bytes_per_second",
-  "stored_count",
-  "revised_count",
-  "skipped_count",
-  "error_count",
-  "pending_count",
-  "prepared_count",
-  "queue_depth",
-  "in_flight_count",
-  "call_count",
-  "created_count",
-  "updated_count",
-  "superseded_count",
-  "deleted_count",
-  "operation_kind",
-  "file_kind",
-]);
-
 export function parseLogQuery(arguments_: readonly string[]): LogQuery | null {
  if (arguments_.filter((item) => item === "--json").length > 1) return null;
  const values: Partial<LogQuery> = { errorsOnly: false, follow: false };
@@ -237,63 +190,13 @@ export function renderLogText(
  return `${lines.join("\n")}\n`;
 }

-/** Defence in depth for the CLI read surface. `DiagnosticLogger` already
- * allow-lists keys, but a future writer, a manually edited log, or a malformed
- * line must not turn `memhub logs` into an exfiltration path. */
+/** Defence in depth for the CLI read surface. Local writers accept arbitrary
+ * bounded scalar keys, so every outward read surface must reuse the same
+ * field-and-value projection as the support-bundle exporter. */
 export function redactDiagnosticEvent(
  event: DiagnosticEventV1,
 ): DiagnosticEventV1 {
-  const redacted: DiagnosticEventV1 = {
-    schema_version: 1,
-    timestamp: event.timestamp,
-    level: event.level,
-    service: event.service,
-    component: event.component,
-    event: event.event,
-    outcome: event.outcome,
-    session_id: event.session_id,
-    ...(event.trace_id === undefined ? {} : { trace_id: event.trace_id }),
-    ...(event.incident_id === undefined
-      ? {}
-      : { incident_id: event.incident_id }),
-    ...(event.instance_id === undefined
-      ? {}
-      : { instance_id: event.instance_id }),
-    ...(event.operation_id === undefined
-      ? {}
-      : { operation_id: event.operation_id }),
-    ...(event.request_id === undefined ? {} : { request_id: event.request_id }),
-    ...(event.job_id === undefined ? {} : { job_id: event.job_id }),
-    ...(event.task_id === undefined ? {} : { task_id: event.task_id }),
-    ...(event.duration_ms === undefined
-      ? {}
-      : { duration_ms: event.duration_ms }),
-    ...(event.state_before === undefined
-      ? {}
-      : { state_before: event.state_before }),
-    ...(event.state_after === undefined
-      ? {}
-      : { state_after: event.state_after }),
-    ...(event.error_code === undefined ? {} : { error_code: event.error_code }),
-    ...(event.error_type === undefined ? {} : { error_type: event.error_type }),
-    ...(event.safe_stack === undefined
-      ? {}
-      : {
-          safe_stack: event.safe_stack.filter(
-            (frame) =>
-              !frame.startsWith("/") &&
-              !/^[A-Za-z]:\\/.test(frame) &&
-              !frame.includes(".."),
-          ),
-        }),
-  };
-  const details = Object.fromEntries(
-    Object.entries(event.details ?? {}).filter(([key]) =>
-      SAFE_DETAIL_KEYS.has(key),
-    ),
-  );
-  if (Object.keys(details).length > 0) redacted.details = details;
-  return redacted;
+  return projectDiagnosticEventForSafeOutput(event);
 }

 export async function followLogs(options: {
@@ -315,6 +218,7 @@ export async function followLogs(options: {
      ...options.query,
      tail: 5_000,
    });
+    if (report.failure) throw new Error(report.failure);
    for (const event of report.events) {
      const identity = JSON.stringify(event);
      if (seen.has(identity)) continue;
@@ -1,700 +0,0 @@
-import { createHash, randomUUID, timingSafeEqual } from "node:crypto";
-import type { GovernedMemory } from "@one-memory/contracts";
-import {
-  buildMemoryActionRequest,
-  buildMemoryCorrectRequest,
-  buildMemoryCreateRequest,
-  parseMemoryReceipt,
-} from "@one-memory/operations";
-import type { CoreReader } from "./core-client.js";
-import type { CoreOperations } from "./core-operations.js";
-import { validateMemoryDetail } from "./core-responses.js";
-import { isMemoryId } from "./memory-report.js";
-import { PRODUCT } from "./version.js";
-
-const MEMORY_KINDS = [
-  "chat",
-  "working",
-  "coding",
-  "schema",
-  "intention",
-] as const;
-const OPERATIONS = ["add", "correct", "delete"] as const;
-const MEMORY_ORIGINS = ["extracted", "answered", "agent", "manual"] as const;
-const MUTATION_RECEIPT_KEYS = new Set([
-  "schema_version",
-  "request_id",
-  "state",
-  "operation",
-  "memory_id",
-  "result_revision",
-  "result_status",
-  "memory_state",
-  "migrated_from",
-]);
-const CONTENT_MAX_CHARACTERS = 100_000;
-const ID = /^[A-Za-z0-9][\w.-]{0,127}$/;
-const UUID = /^[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}$/;
-
-interface PlanBody {
-  schema_version: 1;
-  kind: "memory_mutation";
-  operation: (typeof OPERATIONS)[number];
-  request_id: string;
-  memory_id?: string;
-  expected_revision?: number;
-  expected_status?: string;
-  expected_origin?: string;
-  memory_kind?: (typeof MEMORY_KINDS)[number];
-  scopes?: string[];
-  evidence_ids?: string[];
-  content_sha256?: string;
-  content_characters?: number;
-  current?: {
-    status: string;
-    origin: string;
-    revision: number;
-    scopes: string[];
-    evidence: Array<{
-      evidence_id: string;
-      available: boolean;
-      source_id?: string;
-      written_by?: string;
-    }>;
-  };
-}
-
-export type MemoryMutationPlan = PlanBody & { digest: string };
-
-export type MemoryFlowDocument =
-  | {
-      schema_version: 1;
-      product: string;
-      state: "preview";
-      plan: MemoryMutationPlan;
-    }
-  | {
-      schema_version: 1;
-      product: string;
-      state: "committed" | "recovered";
-      operation_id: string;
-      receipt: Record<string, unknown>;
-    }
-  | {
-      schema_version: 1;
-      product: string;
-      state: "rejected" | "unknown" | "invalid";
-      operation_id?: string;
-      code: string;
-    };
-
-export function parseMemoryContent(bytes: Buffer): string | null {
-  if (bytes.byteLength > CONTENT_MAX_CHARACTERS * 4 + 2) return null;
-  let decoded: string;
-  try {
-    decoded = new TextDecoder("utf-8", { fatal: true }).decode(bytes);
-  } catch {
-    return null;
-  }
-  const content = decoded.replace(/\r?\n$/, "");
-  return content.trim().length > 0 &&
-    [...content].length <= CONTENT_MAX_CHARACTERS &&
-    ![...content].some((character) => {
-      const code = character.charCodeAt(0);
-      return code <= 8 || (code >= 11 && code <= 31) || code === 127;
-    })
-    ? content
-    : null;
-}
-
-export function planMemoryAdd(input: {
-  content: string;
-  memoryKind: string;
-  scopes: string[];
-  evidenceIds: string[];
-}): MemoryFlowDocument {
-  if (
-    !validContent(input.content) ||
-    !oneOf(input.memoryKind, MEMORY_KINDS) ||
-    !validScopes(input.scopes) ||
-    !validIds(input.evidenceIds, 500)
-  ) {
-    return invalid("memory_add_input_invalid");
-  }
-  return preview(
-    withDigest({
-      schema_version: 1,
-      kind: "memory_mutation",
-      operation: "add",
-      request_id: randomUUID(),
-      memory_kind: input.memoryKind,
-      scopes: input.scopes,
-      evidence_ids: input.evidenceIds,
-      ...contentIdentity(input.content),
-    }),
-  );
-}
-
-export async function planExistingMemoryMutation(
-  reader: CoreReader,
-  input: {
-    operation: Exclude<(typeof OPERATIONS)[number], "add">;
-    memoryId: string;
-    content?: string;
-  },
-): Promise<MemoryFlowDocument> {
-  if (
-    !isMemoryId(input.memoryId) ||
-    (input.operation === "correct" && input.content === undefined) ||
-    (input.content !== undefined && !validContent(input.content))
-  ) {
-    return invalid("memory_plan_input_invalid");
-  }
-  const outcome = await reader<unknown>(`/api/v1/memories/${input.memoryId}`);
-  if (outcome.state !== "ok") {
-    return invalid(`memory_${outcome.state}:${outcome.code}`);
-  }
-  const detail = validateMemoryDetail(outcome.data, input.memoryId);
-  if (!detail) return invalid("memory_response_invalid");
-  const memory = detail.memory;
-  if (memory.status !== "confirmed") {
-    return invalid("memory_operation_not_allowed");
-  }
-  const body: PlanBody = {
-    schema_version: 1,
-    kind: "memory_mutation",
-    operation: input.operation,
-    request_id: randomUUID(),
-    memory_id: memory.memory_id,
-    expected_revision: memory.revision,
-    expected_status: memory.status,
-    expected_origin: memory.origin,
-    ...(input.content ? contentIdentity(input.content) : {}),
-    current: {
-      status: memory.status,
-      origin: memory.origin,
-      revision: memory.revision,
-      scopes: [...memory.scopes],
-      evidence: detail.evidence.map((item) => ({
-        evidence_id: item.evidence_id,
-        available: item.available,
-        ...(item.source_id ? { source_id: item.source_id } : {}),
-        ...(item.written_by ? { written_by: item.written_by } : {}),
-      })),
-    },
-  };
-  return preview(withDigest(body));
-}
-
-export function inspectMemoryPlan(
-  value: unknown,
-): { operationId: string; digest: string; needsContent: boolean } | null {
-  const plan = validatePlan(value);
-  return plan
-    ? {
-        operationId: plan.request_id,
-        digest: plan.digest,
-        needsContent: plan.operation === "add" || plan.operation === "correct",
-      }
-    : null;
-}
-
-export function memoryPlanAcceptsContent(
-  value: unknown,
-  content: string | undefined,
-): boolean {
-  const plan = validatePlan(value);
-  if (!plan) return false;
-  if (plan.operation !== "add" && plan.operation !== "correct") {
-    return content === undefined;
-  }
-  if (content === undefined) return false;
-  const identity = contentIdentity(content);
-  return (
-    identity.content_sha256 === plan.content_sha256 &&
-    identity.content_characters === plan.content_characters
-  );
-}
-
-export async function executeMemoryMutation(
-  value: unknown,
-  content: string | undefined,
-  operations: CoreOperations,
-  reader?: CoreReader,
-): Promise<MemoryFlowDocument> {
-  const plan = validatePlan(value);
-  if (!plan) return invalid("memory_plan_invalid");
-  if (!memoryPlanAcceptsContent(plan, content)) {
-    return invalid("memory_content_mismatch");
-  }
-  const { path, body } = operationRequest(plan, content);
-  const outcome = await operations.post<unknown>(path, body);
-  if (outcome.state === "rejected") {
-    return {
-      schema_version: 1,
-      product: PRODUCT,
-      state: "rejected",
-      operation_id: plan.request_id,
-      code: outcome.code,
-    };
-  }
-  if (outcome.state === "unknown") {
-    return reader
-      ? readMemoryMutationReceipt(reader, plan)
-      : pending(plan.request_id, outcome.code);
-  }
-  const memory = validateWriteResponse(outcome.data, plan);
-  if (!memory || !matchesReceipt(memory, plan, content)) {
-    return reader
-      ? readMemoryMutationReceipt(reader, plan)
-      : pending(plan.request_id, "memory_receipt_invalid");
-  }
-  return {
-    schema_version: 1,
-    product: PRODUCT,
-    state: "committed",
-    operation_id: plan.request_id,
-    receipt: {
-      memory_id: memory.memory_id,
-      status: memory.status,
-      origin: memory.origin,
-      revision: memory.revision,
-      scopes: memory.scopes,
-      evidence_ids: memory.evidence_ids,
-      content_sha256: contentIdentity(memory.content).content_sha256,
-      content_characters: [...memory.content].length,
-      created_at: memory.created_at,
-      updated_at: memory.updated_at,
-    },
-  };
-}
-
-/**
- * Validate the write response through the shared `parseMemoryReceipt`, which
- * owns the GovernedMemory shape and the revision/status binding Core must
- * satisfy. The CLI keeps only its own plan binding (`matchesReceipt`, which
- * also checks `origin`, `kind` and the echoed content), so the two surfaces
- * judge a write response identically and no shape rule is duplicated here.
- *
- * `add` creates the first revision and expects `confirmed`; the shared parser
- * derives that from `expectedRevision: 0`. Every other operation advances the
- * revision the caller acted against and expects `confirmed` (correct/scope)
- * or `deleted` (delete).
- */
-function validateWriteResponse(
-  value: unknown,
-  plan: MemoryMutationPlan,
-): GovernedMemory | null {
-  const expectedStatus = plan.operation === "delete" ? "deleted" : "confirmed";
-  // For `add` there is no id or prior revision to bind yet, so the parser's
-  // own receipt is the only source for both; the plan binding that follows is
-  // what proves the response belongs to this request.
-  const memoryId =
-    plan.operation === "add"
-      ? String((value as { memory_id?: unknown } | null)?.memory_id ?? "")
-      : (plan.memory_id as string);
-  if (memoryId === "") return null;
-  return parseMemoryReceipt(value, {
-    memoryId,
-    expectedRevision:
-      plan.operation === "add" ? 0 : (plan.expected_revision as number),
-    expectedStatus,
-  });
-}
-
-export async function readMemoryMutationReceipt(
-  reader: CoreReader,
-  value: unknown,
-): Promise<MemoryFlowDocument> {
-  const plan = validatePlan(value);
-  if (!plan) return invalid("memory_plan_invalid");
-  const outcome = await reader<unknown>(
-    `/api/v1/memory-mutations/${encodeURIComponent(plan.request_id)}`,
-  );
-  if (outcome.state !== "ok") {
-    return pending(plan.request_id, `memory_mutation_receipt_${outcome.code}`);
-  }
-  const receipt = validateMutationReceipt(outcome.data, plan);
-  if (receipt === "not_found") {
-    return pending(
-      plan.request_id,
-      "memory_mutation_receipt_not_found_after_attempt",
-    );
-  }
-  if (!receipt) {
-    return pending(plan.request_id, "memory_mutation_receipt_invalid");
-  }
-  return {
-    schema_version: 1,
-    product: PRODUCT,
-    state: "recovered",
-    operation_id: plan.request_id,
-    receipt,
-  };
-}
-
-export function memoryFlowExitCode(document: MemoryFlowDocument): number {
-  return ["preview", "committed", "recovered"].includes(document.state) ? 0 : 1;
-}
-
-export function renderMemoryFlowText(document: MemoryFlowDocument): string {
-  if (document.state === "preview") {
-    const plan = document.plan;
-    const lines = [
-      `记忆操作预览：${plan.operation}`,
-      `request_id：${plan.request_id}`,
-    ];
-    if (plan.memory_id) lines.push(`memory_id：${plan.memory_id}`);
-    if (plan.expected_revision) {
-      lines.push(
-        `当前版本：${plan.expected_revision}  状态：${plan.expected_status}  来源方式：${plan.expected_origin}`,
-      );
-    }
-    if (plan.scopes) lines.push(`使用范围：${plan.scopes.join(",")}`);
-    if (plan.current) {
-      for (const evidence of plan.current.evidence) {
-        lines.push(
-          `证据 ${evidence.evidence_id}：source=${evidence.source_id ?? "unknown"} author=${evidence.written_by ?? "unknown"} available=${evidence.available}`,
-        );
-      }
-    }
-    if (plan.content_characters !== undefined) {
-      lines.push(
-        `正文：${plan.content_characters} 字符，sha256=${plan.content_sha256}`,
-      );
-    }
-    lines.push("尚未写入；保存 JSON 计划后用 memories apply --confirm 执行。");
-    return `${lines.join("\n")}\n`;
-  }
-  if (document.state === "committed" || document.state === "recovered") {
-    return `记忆操作${document.state === "recovered" ? "已从治理回执恢复" : "已提交"}：${document.operation_id}\n${JSON.stringify(document.receipt)}\n`;
-  }
-  return `记忆操作未完成：${document.state}（${"code" in document ? document.code : "memory_operation_state_invalid"}）${document.operation_id ? ` request_id=${document.operation_id}` : ""}\n`;
-}
-
-function operationRequest(
-  plan: MemoryMutationPlan,
-  content: string | undefined,
-): { path: string; body: Record<string, unknown> } {
-  if (plan.operation === "add") {
-    // The shared create builder is the Web console's manual-add body, which
-    // always creates `chat` (the Web has no kind picker). The CLI exposes
-    // `--kind`, so it builds through the same builder and then sets the kind,
-    // keeping every other field (user_id/scope/evidence/origin) shared.
-    const request = buildMemoryCreateRequest(
-      {
-        content: content as string,
-        scopes: plan.scopes ?? [],
-        evidence_ids: plan.evidence_ids ?? [],
-        origin: "manual",
-      },
-      plan.request_id,
-    );
-    const draft = request.body.draft as Record<string, unknown>;
-    draft.kind = plan.memory_kind;
-    return { path: request.path, body: request.body };
-  }
-  if (plan.operation === "correct") {
-    return buildMemoryCorrectRequest(
-      plan.memory_id as string,
-      plan.expected_revision as number,
-      content as string,
-      undefined,
-      plan.request_id,
-    );
-  }
-  return buildMemoryActionRequest(
-    plan.memory_id as string,
-    plan.operation,
-    plan.expected_revision as number,
-    plan.request_id,
-  );
-}
-
-function matchesReceipt(
-  memory: GovernedMemory,
-  plan: MemoryMutationPlan,
-  content: string | undefined,
-): boolean {
-  if (plan.operation === "add") {
-    return (
-      memory.status === "confirmed" &&
-      memory.origin === "manual" &&
-      memory.revision === 1 &&
-      memory.kind === plan.memory_kind &&
-      same(memory.scopes, plan.scopes ?? []) &&
-      same(memory.evidence_ids, plan.evidence_ids ?? []) &&
-      memory.content === content
-    );
-  }
-  if (
-    memory.memory_id !== plan.memory_id ||
-    memory.revision !== (plan.expected_revision as number) + 1 ||
-    memory.origin !== plan.expected_origin
-  ) {
-    return false;
-  }
-  if (plan.operation === "delete" && memory.status !== "deleted") return false;
-  if (plan.operation === "correct" && memory.status !== plan.expected_status) {
-    return false;
-  }
-  if (plan.operation === "correct" && memory.content !== content) return false;
-  const expectedScopes =
-    plan.scopes ?? (plan.current?.scopes as string[] | undefined);
-  return expectedScopes ? same(memory.scopes, expectedScopes) : true;
-}
-
-function validateMutationReceipt(
-  value: unknown,
-  plan: MemoryMutationPlan,
-): Record<string, unknown> | "not_found" | null {
-  const receipt = record(value);
-  if (
-    !receipt ||
-    Object.keys(receipt).length !== MUTATION_RECEIPT_KEYS.size ||
-    Object.keys(receipt).some((key) => !MUTATION_RECEIPT_KEYS.has(key)) ||
-    receipt.schema_version !== 2 ||
-    receipt.request_id !== plan.request_id
-  ) {
-    return null;
-  }
-  if (receipt.state === "not_found") {
-    return [
-      receipt.operation,
-      receipt.memory_id,
-      receipt.result_revision,
-      receipt.result_status,
-      receipt.memory_state,
-      receipt.migrated_from,
-    ].every((item) => item === null)
-      ? "not_found"
-      : null;
-  }
-  const expectedOperation =
-    plan.operation === "add" ? "create" : plan.operation;
-  const expectedRevision =
-    plan.operation === "add" ? 1 : (plan.expected_revision as number) + 1;
-  const expectedStatus = plan.operation === "delete" ? "deleted" : "confirmed";
-  // A migrated receipt records an operation from before v2; no plan made
-  // today can have produced it.
-  if (
-    receipt.state !== "committed" ||
-    receipt.migrated_from !== null ||
-    receipt.operation !== expectedOperation ||
-    typeof receipt.memory_id !== "string" ||
-    !isMemoryId(receipt.memory_id) ||
-    (plan.memory_id !== undefined && receipt.memory_id !== plan.memory_id) ||
-    receipt.result_revision !== expectedRevision ||
-    receipt.result_status !== expectedStatus ||
-    !["present", "deleted"].includes(String(receipt.memory_state))
-  ) {
-    return null;
-  }
-  return {
-    schema_version: 2,
-    request_id: plan.request_id,
-    state: "committed",
-    operation: expectedOperation,
-    memory_id: receipt.memory_id,
-    result_revision: expectedRevision,
-    result_status: expectedStatus,
-    memory_state: receipt.memory_state,
-    proof_scope: "governance_transaction_only",
-    downstream_sync_verified: false,
-    payload_fingerprint_verified: false,
-  };
-}
-
-function validatePlan(value: unknown): MemoryMutationPlan | null {
-  const plan = record(value);
-  if (
-    plan?.schema_version !== 1 ||
-    plan.kind !== "memory_mutation" ||
-    !oneOf(plan.operation, OPERATIONS) ||
-    typeof plan.request_id !== "string" ||
-    !UUID.test(plan.request_id) ||
-    typeof plan.digest !== "string"
-  ) {
-    return null;
-  }
-  const body = { ...plan };
-  delete body.digest;
-  const expected = withDigest(body as unknown as PlanBody);
-  if (!safeEqual(expected.digest, plan.digest)) return null;
-  if (plan.operation === "add") {
-    if (
-      !oneOf(plan.memory_kind, MEMORY_KINDS) ||
-      !validScopes(plan.scopes) ||
-      !validIds(plan.evidence_ids, 500) ||
-      !contentFields(plan)
-    ) {
-      return null;
-    }
-  } else if (
-    typeof plan.memory_id !== "string" ||
-    !isMemoryId(plan.memory_id) ||
-    !Number.isSafeInteger(plan.expected_revision) ||
-    (plan.expected_revision as number) < 1 ||
-    plan.expected_status !== "confirmed" ||
-    !oneOf(plan.expected_origin, MEMORY_ORIGINS) ||
-    !validCurrent(plan) ||
-    (plan.operation === "correct" && !contentFields(plan))
-  ) {
-    return null;
-  }
-  return plan as unknown as MemoryMutationPlan;
-}
-
-function contentFields(plan: Record<string, unknown>): boolean {
-  return (
-    typeof plan.content_sha256 === "string" &&
-    /^[0-9a-f]{64}$/.test(plan.content_sha256) &&
-    Number.isSafeInteger(plan.content_characters) &&
-    (plan.content_characters as number) >= 1 &&
-    (plan.content_characters as number) <= CONTENT_MAX_CHARACTERS
-  );
-}
-
-function validCurrent(plan: Record<string, unknown>): boolean {
-  const current = record(plan.current);
-  if (
-    !current ||
-    current.status !== plan.expected_status ||
-    current.origin !== plan.expected_origin ||
-    current.revision !== plan.expected_revision ||
-    !validScopes(current.scopes) ||
-    !Array.isArray(current.evidence) ||
-    current.evidence.length < 1 ||
-    current.evidence.length > 500
-  ) {
-    return false;
-  }
-  return current.evidence.every((value) => {
-    const evidence = record(value);
-    return (
-      evidence !== null &&
-      typeof evidence.evidence_id === "string" &&
-      ID.test(evidence.evidence_id) &&
-      typeof evidence.available === "boolean" &&
-      (evidence.source_id === undefined ||
-        (typeof evidence.source_id === "string" &&
-          evidence.source_id.length <= 128)) &&
-      (evidence.written_by === undefined ||
-        (typeof evidence.written_by === "string" &&
-          evidence.written_by.length <= 256))
-    );
-  });
-}
-
-function contentIdentity(content: string): {
-  content_sha256: string;
-  content_characters: number;
-} {
-  return {
-    content_sha256: createHash("sha256").update(content, "utf8").digest("hex"),
-    content_characters: [...content].length,
-  };
-}
-
-function validContent(content: string): boolean {
-  return (
-    content.trim().length > 0 &&
-    [...content].length <= CONTENT_MAX_CHARACTERS &&
-    ![...content].some((character) => {
-      const code = character.charCodeAt(0);
-      return code <= 8 || (code >= 11 && code <= 31) || code === 127;
-    })
-  );
-}
-
-function validScopes(value: unknown): value is string[] {
-  if (
-    !Array.isArray(value) ||
-    value.length < 1 ||
-    value.length > 128 ||
-    new Set(value).size !== value.length ||
-    value.some(
-      (scope) =>
-        typeof scope !== "string" ||
-        scope.trim() !== scope ||
-        scope.length < 1 ||
-        scope.length > 512 ||
-        scope === "*",
-    )
-  ) {
-    return false;
-  }
-  return !value.includes("private") || value.length === 1;
-}
-
-function validIds(value: unknown, maximum: number): value is string[] {
-  return (
-    Array.isArray(value) &&
-    value.length >= 1 &&
-    value.length <= maximum &&
-    new Set(value).size === value.length &&
-    value.every((item) => typeof item === "string" && ID.test(item))
-  );
-}
-
-function withDigest(body: PlanBody): MemoryMutationPlan {
-  return {
-    ...body,
-    digest: createHash("sha256")
-      .update(JSON.stringify(body), "utf8")
-      .digest("hex"),
-  };
-}
-
-function preview(plan: MemoryMutationPlan): MemoryFlowDocument {
-  return {
-    schema_version: 1,
-    product: PRODUCT,
-    state: "preview",
-    plan,
-  };
-}
-
-function invalid(code: string): MemoryFlowDocument {
-  return {
-    schema_version: 1,
-    product: PRODUCT,
-    state: "invalid",
-    code,
-  };
-}
-
-function pending(operationId: string, code: string): MemoryFlowDocument {
-  return {
-    schema_version: 1,
-    product: PRODUCT,
-    state: "unknown",
-    operation_id: operationId,
-    code,
-  };
-}
-
-function same(left: readonly string[], right: readonly string[]): boolean {
-  return (
-    left.length === right.length &&
-    left.every((value, index) => value === right[index])
-  );
-}
-
-function safeEqual(left: string, right: string): boolean {
-  if (left.length !== right.length) return false;
-  return timingSafeEqual(Buffer.from(left), Buffer.from(right));
-}
-
-function oneOf<const T extends readonly string[]>(
-  value: unknown,
-  values: T,
-): value is T[number] {
-  return typeof value === "string" && values.includes(value as T[number]);
-}
-
-function record(value: unknown): Record<string, unknown> | null {
-  return value !== null && typeof value === "object" && !Array.isArray(value)
-    ? (value as Record<string, unknown>)
-    : null;
-}
@@ -1,211 +0,0 @@
-import { readFileSync } from "node:fs";
-import { describe, expect, it } from "vitest";
-import type { CoreReader } from "./core-client.js";
-import {
-  planExistingMemoryMutation,
-  planMemoryAdd,
-  readMemoryMutationReceipt,
-} from "./memory-flows.js";
-
-// Written by a real Core and checked against it on the Core side, so these are
-// the receipts the CLI meets in use rather than ones shaped to fit the parser.
-const FIXTURE = JSON.parse(
-  readFileSync(
-    new URL(
-      "../../../packages/contracts/fixtures/memory-mutation-receipts.v2.json",
-      import.meta.url,
-    ),
-    "utf8",
-  ),
-) as { receipts: Record<"add" | "correct" | "delete", Receipt> };
-
-interface Receipt {
-  schema_version: number;
-  request_id: string;
-  memory_id: string;
-  result_revision: number;
-  result_status: string;
-  operation: string;
-}
-
-function reader(data: unknown): CoreReader {
-  return async <T>() => ({ state: "ok", data: data as T });
-}
-
-function detail(sample: Receipt) {
-  const revision = sample.result_revision - 1;
-  return {
-    memory: {
-      memory_id: sample.memory_id,
-      user_id: "local",
-      kind: "chat",
-      status: "confirmed",
-      content: "Synthetic memory.",
-      scopes: ["all-apps"],
-      evidence_ids: ["evidence-1"],
-      origin: "manual",
-      tags: [],
-      revision,
-      created_at: "2026-09-19T00:00:00Z",
-      updated_at: "2026-09-19T00:00:01Z",
-      deleted_at: null,
-    },
-    evidence: [
-      {
-        evidence_id: "evidence-1",
-        available: true,
-        source_id: "fixture",
-        written_by: "person",
-        created_at: "2026-09-19T00:00:00Z",
-        verbatim: "Synthetic source text.",
-      },
-    ],
-    history: [
-      {
-        revision,
-        action: "create",
-        status: "confirmed",
-        content: "Synthetic memory.",
-        scopes: ["all-apps"],
-        created_at: "2026-09-19T00:00:01Z",
-      },
-    ],
-    index_sync: null,
-  };
-}
-
-describe("memory mutation receipt v2 from Core", () => {
-  it("recovers an add from the receipt Core returns for a create", async () => {
-    const sample = FIXTURE.receipts.add;
-    expect(sample.schema_version).toBe(2);
-    const preview = planMemoryAdd({
-      content: "Synthetic memory.",
-      memoryKind: "chat",
-      scopes: ["all-apps"],
-      evidenceIds: ["evidence-1"],
-    });
-    if (preview.state !== "preview") throw new Error("preview expected");
-    // The request id is the one binding a receipt to its plan; the plan picks it.
-    const result = await readMemoryMutationReceipt(
-      reader({ ...sample, request_id: preview.plan.request_id }),
-      preview.plan,
-    );
-    expect(result).toMatchObject({
-      state: "recovered",
-      receipt: {
-        schema_version: 2,
-        operation: "create",
-        memory_id: sample.memory_id,
-        result_revision: 1,
-        result_status: "confirmed",
-        memory_state: "present",
-      },
-    });
-  });
-
-  it.each<["correct" | "delete", { content?: string }]>([
-    ["correct", { content: "Corrected synthetic memory." }],
-    ["delete", {}],
-  ])(
-    "recovers %s from the receipt Core returns for it",
-    async (operation, extra) => {
-      const sample = FIXTURE.receipts[operation];
-      const preview = await planExistingMemoryMutation(reader(detail(sample)), {
-        operation,
-        memoryId: sample.memory_id,
-        ...extra,
-      });
-      if (preview.state !== "preview") {
-        throw new Error(`preview expected: ${JSON.stringify(preview)}`);
-      }
-      const result = await readMemoryMutationReceipt(
-        reader({ ...sample, request_id: preview.plan.request_id }),
-        preview.plan,
-      );
-      expect(result).toMatchObject({
-        state: "recovered",
-        receipt: {
-          schema_version: 2,
-          operation: sample.operation,
-          memory_id: sample.memory_id,
-          result_revision: sample.result_revision,
-          result_status: sample.result_status,
-        },
-      });
-    },
-  );
-
-  it("does not take a v1 receipt for a v2 one", async () => {
-    const sample = FIXTURE.receipts.delete;
-    const preview = await planExistingMemoryMutation(reader(detail(sample)), {
-      operation: "delete",
-      memoryId: sample.memory_id,
-    });
-    if (preview.state !== "preview") throw new Error("preview expected");
-    const { migrated_from: _dropped, ...v1 } = sample as Receipt & {
-      migrated_from: null;
-    };
-    // As v1 was, and with v2's fields but v1's number: either alone is refused.
-    for (const receipt of [
-      { ...v1, schema_version: 1 },
-      { ...sample, schema_version: 1 },
-    ]) {
-      const result = await readMemoryMutationReceipt(
-        reader({ ...receipt, request_id: preview.plan.request_id }),
-        preview.plan,
-      );
-      expect(result).toMatchObject({
-        state: "unknown",
-        code: "memory_mutation_receipt_invalid",
-      });
-    }
-  });
-
-  it("does not take a not-found receipt that names a migrated operation", async () => {
-    const sample = FIXTURE.receipts.delete;
-    const preview = await planExistingMemoryMutation(reader(detail(sample)), {
-      operation: "delete",
-      memoryId: sample.memory_id,
-    });
-    if (preview.state !== "preview") throw new Error("preview expected");
-    const notFound = {
-      schema_version: 2,
-      request_id: preview.plan.request_id,
-      state: "not_found",
-      operation: null,
-      memory_id: null,
-      result_revision: null,
-      result_status: null,
-      memory_state: null,
-    };
-    expect(
-      await readMemoryMutationReceipt(
-        reader({ ...notFound, migrated_from: null }),
-        preview.plan,
-      ),
-    ).toMatchObject({
-      code: "memory_mutation_receipt_not_found_after_attempt",
-    });
-    expect(
-      await readMemoryMutationReceipt(
-        reader({ ...notFound, migrated_from: "reject" }),
-        preview.plan,
-      ),
-    ).toMatchObject({ code: "memory_mutation_receipt_invalid" });
-  });
-
-  it("plans nothing on a memory that is not in effect", async () => {
-    const sample = FIXTURE.receipts.delete;
-    const superseded = detail(sample);
-    superseded.memory.status = "superseded";
-    superseded.history[0].status = "superseded";
-    const result = await planExistingMemoryMutation(reader(superseded), {
-      operation: "delete",
-      memoryId: sample.memory_id,
-    });
-    expect(result).toMatchObject({
-      state: "invalid",
-      code: "memory_operation_not_allowed",
-    });
-  });
-});
@@ -15,6 +15,29 @@ export function isMemoryId(value: string): boolean {
  );
 }

+const CONTENT_MAX_CHARACTERS = 100_000;
+
+/** A memory body read from `--content-file`: UTF-8, non-blank, bounded, and
+ * free of control characters other than tab and newline. */
+export function parseMemoryContent(bytes: Buffer): string | null {
+  if (bytes.byteLength > CONTENT_MAX_CHARACTERS * 4 + 2) return null;
+  let decoded: string;
+  try {
+    decoded = new TextDecoder("utf-8", { fatal: true }).decode(bytes);
+  } catch {
+    return null;
+  }
+  const content = decoded.replace(/\r?\n$/, "");
+  return content.trim().length > 0 &&
+    [...content].length <= CONTENT_MAX_CHARACTERS &&
+    ![...content].some((character) => {
+      const code = character.charCodeAt(0);
+      return code <= 8 || (code >= 11 && code <= 31) || code === 127;
+    })
+    ? content
+    : null;
+}
+
 export interface MemoryQuery {
  query?: string;
  status?: string;
@@ -129,7 +152,7 @@ export function renderMemoryPageText(document: MemoryPageDocument): string {
  ];
  for (const memory of document.memories) {
    lines.push(
-      `- ${memory.memory_id}  ${memory.status}/${memory.layer}  source=${memory.source ?? "未知"}`,
+      `- ${memory.memory_id}  [${memory.scene}] ${memory.status}/${memory.layer}  source=${memory.source ?? "未知"}`,
      `  ${memory.content}${memory.content_truncated ? ` …（共 ${memory.content_characters} 字）` : ""}`,
    );
  }
@@ -148,6 +171,7 @@ export function renderMemoryDetailText(document: MemoryDetailDocument): string {
  }
  const lines = [
    `记忆 ${memory.memory_id}`,
+    `场景：${memory.scene}  类型：${memory.kind}${memory.variant ? `（${memory.variant}）` : ""}`,
    `状态：${memory.status}  层：${memory.layer}  来源：${memory.source ?? "未知"}`,
    `owner：${memory.owner ?? "未知"}  latest：${memory.is_latest}`,
    `raw IDs：${(memory.source_raw_ids as string[]).join(",") || "无"}`,
@@ -165,6 +189,10 @@ function describeMemory(memory: SdkMemoryRecord): Record<string, unknown> {
  const budgeted = budget(memory.content ?? "");
  return {
    memory_id: memory.memory_id,
+    // Pre-2.0.6 Core listed chat memories only and named no scene.
+    scene: memory.scene ?? "chat",
+    kind: memory.kind ?? "chat",
+    variant: memory.variant ?? null,
    status: memory.status,
    layer: memory.layer,
    is_latest: memory.is_latest,
@@ -208,8 +236,9 @@ function sdkMemory(value: unknown): SdkMemoryRecord | null {
    !strings(value.superseded_by) ||
    !strings(value.tags) ||
    !strings(value.source_raw_ids) ||
-    !text(value.user_id) ||
-    !text(value.agent_id)
+    // Working/coding memories may carry an empty user or agent id.
+    typeof value.user_id !== "string" ||
+    typeof value.agent_id !== "string"
  )
    return null;
  return value as unknown as SdkMemoryRecord;
@@ -184,8 +184,16 @@ const CORE = new Map<string, { status: number; data: unknown }>([
    },
  ],
  [
-    "/api/v1/home-summary",
-    { status: 200, data: { memory_count: 42, latest: null, observed_at: "T" } },
+    "/api/v2/home-summary",
+    {
+      status: 200,
+      data: {
+        schema_version: 2,
+        memory_count: 42,
+        latest: null,
+        observed_at: "T",
+      },
+    },
  ],
 ]);

@@ -194,9 +202,7 @@ const READ_ONLY_PATHS = [
  "/api/v1/live-sync",
  "/api/v1/memory-processing",
  "/api/v1/model-profiles",
-  "/api/v1/home-summary",
-  "/api/v1/memory-library",
-  "/api/v1/memories/",
+  "/api/v2/home-summary",
  "/api/v2/memory-library",
  "/api/v2/memories/",
  "/api/v2/user-profile",
@@ -691,6 +697,66 @@ describe("CLI-21 memory reads", () => {
    expect(report.memories[1].content.length).toBeLessThan(400);
  });

+  it("lists working and coding memories with their scene and kind", async () => {
+    transport.get.mockResolvedValue({
+      status: 200,
+      data: {
+        total: 2,
+        offset: 0,
+        limit: 20,
+        totals: { chat: 1, working: 1, coding: 0 },
+        window_id: null,
+        window_reset: false,
+        memories: [
+          memory("m1", "对话记忆"),
+          {
+            ...memory("e1", "修复测试\n固定随机种子"),
+            layer: "experience",
+            agent_id: "",
+            owner: null,
+            source_raw_ids: [],
+            scene: "working",
+            kind: "experience",
+            variant: "working",
+            task: "修复测试",
+            solution: "固定随机种子",
+            created_at: 1_790_000_000,
+            updated_at: 1_790_000_100,
+          },
+        ],
+      },
+    });
+    const streams = captureStreams();
+    try {
+      expect(await run(["memories", "list", "--json"], {})).toBe(0);
+    } finally {
+      streams.restore();
+    }
+    const report = JSON.parse(streams.stdout());
+    expect(report.page).toMatchObject({ state: "ok", total: 2, returned: 2 });
+    expect(report.memories[0]).toMatchObject({
+      memory_id: "m1",
+      scene: "chat",
+      kind: "chat",
+      variant: null,
+    });
+    expect(report.memories[1]).toMatchObject({
+      memory_id: "e1",
+      scene: "working",
+      kind: "experience",
+      variant: "working",
+      layer: "experience",
+    });
+
+    const text = captureStreams();
+    try {
+      expect(await run(["memories", "list"], {})).toBe(0);
+    } finally {
+      text.restore();
+    }
+    expect(text.stdout()).toContain("- e1  [working] active/experience");
+  });
+
  it("passes the requested filters through as query parameters", async () => {
    await run(
      [
@@ -816,7 +882,7 @@ describe("CLI-21 memory reads", () => {
    expect(report.history).toEqual([]);
  });

-  // Core registers /api/v1/memories/graph before /{memory_id}, and that static
+  // Core registers /api/v2/memories/graph before /{memory_id}, and that static
  // handler initializes the HY bridge. A memory id that collides with a
  // reserved segment must never become a request.
  it("refuses a memory id that collides with a static Core route", async () => {
@@ -830,18 +896,15 @@ describe("CLI-21 memory reads", () => {
      expect(streams.stdout()).toBe("");
    }
    expect(transport.get).not.toHaveBeenCalled();
-    expect(isReadOnlyPath("/api/v1/memories/graph")).toBe(false);
-    expect(isReadOnlyPath("/api/v1/memories/ask")).toBe(false);
-    expect(isReadOnlyPath("/api/v1/memories/m1")).toBe(true);
-    expect(isReadOnlyPath("/api/v2/memory-library")).toBe(true);
+    expect(isReadOnlyPath("/api/v2/memories/graph")).toBe(false);
+    expect(isReadOnlyPath("/api/v2/memories/ask")).toBe(false);
    expect(isReadOnlyPath("/api/v2/memories/m1")).toBe(true);
+    expect(isReadOnlyPath("/api/v2/memory-library")).toBe(true);
+    expect(isReadOnlyPath("/api/v2/home-summary")).toBe(true);
+    // The retired v1 data plane is no longer a CLI read.
+    expect(isReadOnlyPath("/api/v1/memories/m1")).toBe(false);
+    expect(isReadOnlyPath("/api/v1/home-summary")).toBe(false);
    expect(isReadOnlyPath("/api/v1/model-content-permission")).toBe(false);
-    expect(
-      isReadOnlyPath(
-        "/api/v1/memory-mutations/123e4567-e89b-12d3-a456-426614174000",
-      ),
-    ).toBe(true);
-    expect(isReadOnlyPath("/api/v1/memory-mutations/../sources")).toBe(false);
  });

  it("refuses a memory id that would escape the memories path", async () => {
@@ -861,6 +924,26 @@ describe("CLI-21 memory reads", () => {
 // was cast to the expected type, so a wrong record, an unavailable excerpt, an
 // oversized page and a non-string model field all reached stdout as success.
 describe("malformed success responses", () => {
+  it("treats a v1-shaped home summary as invalid, not as a memory count", async () => {
+    transport.get.mockImplementation(async (url: string) =>
+      url === "/api/v2/home-summary"
+        ? {
+            status: 200,
+            data: { memory_count: 7, latest: null, observed_at: "T" },
+          }
+        : (CORE.get(url) ?? { status: 404, data: { detail: "not_found" } }),
+    );
+    const streams = captureStreams();
+    try {
+      await run(["status", "--json"], {});
+    } finally {
+      streams.restore();
+    }
+    expect(JSON.parse(streams.stdout()).sections.memories).toMatchObject({
+      state: "invalid",
+    });
+  });
+
  it("refuses a detail that is not the memory that was asked for", async () => {
    transport.get.mockResolvedValue({
      status: 200,
@@ -315,6 +315,40 @@ describe("macOS service lifecycle", () => {
    expect(deps.verifyRegistrationSnapshot).toHaveBeenCalledTimes(2);
  });

+  it("does not re-bootstrap a previous registration that may serve a legacy home", async () => {
+    const f = fixture();
+    const oldPath = "/Users/example/Library/LaunchAgents/old-runtime.plist";
+    const runCommand = vi
+      .fn()
+      .mockResolvedValueOnce(undefined)
+      .mockRejectedValueOnce(new Error("bootstrap failed"));
+    const deps = dependencies({
+      inspectRegistration: vi.fn(async () => ({
+        kind: "managed" as const,
+        plistPath: oldPath,
+        content: "old managed plist",
+      })),
+      runCommand,
+    });
+
+    const prepared = await prepareServiceInstall(f.layout, f.options, deps);
+    await expect(
+      applyPreparedServiceInstall(
+        { ...prepared, restorePreviousOnFailure: false },
+        deps,
+      ),
+    ).rejects.toThrow("runtime_service_replacement_failed");
+
+    // bootout + the failed bootstrap only; the old plist is not bootstrapped.
+    expect(runCommand).toHaveBeenCalledTimes(2);
+    expect(runCommand).not.toHaveBeenCalledWith([
+      "launchctl",
+      "bootstrap",
+      "gui/501",
+      oldPath,
+    ]);
+  });
+
  it("removes the replacement plist and restores the previous registration", async () => {
    const f = fixture();
    const oldPath = "/Users/example/Library/LaunchAgents/old-runtime.plist";
@@ -72,6 +72,13 @@ export type PreparedServiceInstall = {
  mode: "direct" | "bootstrap-existing" | "restart" | "replace";
  previous?: ManagedSnapshot;
  previouslyEnabled: boolean;
+  /**
+   * Re-bootstrap the previous (possibly older) LaunchAgent when the new one
+   * fails to bootstrap. `install` turns this off once the home has been handed
+   * to this version's SDK migration: an older Runtime has no migration gate and
+   * would serve a home that is not on SDK authority.
+   */
+  restorePreviousOnFailure?: boolean;
 };

 const defaultDependencies: ServiceLifecycleDependencies = {
@@ -191,7 +198,8 @@ export async function applyPreparedServiceInstall(
    if (!previous) throw new Error("runtime_service_upgrade_state_invalid");
    await verifyPreparedRegistration(prepared, dependencies);
    const domain = launchdDomain(options.uid);
-    let replacementFailure: "upgrade" | "rollback" | null = null;
+    let replacementFailure: "upgrade" | "rollback" | "not-restored" | null =
+      null;
    const replaceCommand: ServiceCommandRunner = async (command) => {
      if (command[0] !== "launchctl" || command[1] !== "bootstrap") {
        await dependencies.runCommand(command);
@@ -208,6 +216,10 @@ export async function applyPreparedServiceInstall(
        await dependencies.runCommand(command);
      } catch {
        replacementFailure = "upgrade";
+        if (prepared.restorePreviousOnFailure === false) {
+          replacementFailure = "not-restored";
+          throw new Error("runtime_service_replacement_failed");
+        }
        try {
          await dependencies.verifyRegistrationSnapshot(previous);
          await dependencies.runCommand([
@@ -234,6 +246,9 @@ export async function applyPreparedServiceInstall(
      if (replacementFailure === "upgrade") {
        throw new Error("runtime_service_managed_upgrade_failed");
      }
+      if (replacementFailure === "not-restored") {
+        throw new Error("runtime_service_replacement_failed");
+      }
      throw error;
    }
  } catch (error) {
@@ -7,12 +7,10 @@ import type {
 } from "@one-memory/contracts";
 import type { CoreReader, ReadOutcome } from "./core-client.js";
 import {
-  type ValidatedMemoryTask,
  validateImportIssues,
  validateImportJob,
  validateImportJobs,
  validateLiveSync,
-  validateMemoryTasks,
  validateProcessing,
  validateSources,
 } from "./core-responses.js";
@@ -70,11 +68,6 @@ export interface ProgressDocument {
  generation: {
    summary_state: ReadState;
    summary: Record<string, unknown> | null;
-    tasks_state: ReadState;
-    tasks: Array<Record<string, unknown>>;
-    returned: number;
-    validated: number;
-    output_truncated: boolean;
  };
 }

@@ -134,12 +127,13 @@ export async function readProgress(
  const importsPath = options.jobId
    ? `/api/v1/import-jobs/${options.jobId}`
    : "/api/v1/import-jobs";
-  const [importsRead, sourcesRead, processingRead, tasksRead, issuesRead] =
+  // Generation progress comes from `memory-processing` alone: under SDK
+  // authority the per-task list (`/api/v1/memory-tasks`) is retired and empty.
+  const [importsRead, sourcesRead, processingRead, issuesRead] =
    await Promise.all([
      reader<unknown>(importsPath),
      reader<unknown>("/api/v1/sources"),
      reader<unknown>("/api/v1/memory-processing"),
-      reader<unknown>("/api/v1/memory-tasks"),
      options.jobId && options.issues
        ? reader<unknown>(`/api/v1/import-jobs/${options.jobId}/issues`)
        : Promise.resolve(null),
@@ -153,7 +147,6 @@ export async function readProgress(
    : validated(importsRead, validateImportJobs);
  const processing = validated(processingRead, validateProcessing);
  const sources = validated(sourcesRead, validateSources);
-  const tasks = validated(tasksRead, validateMemoryTasks);
  const issues =
    options.jobId && options.issues && issuesRead
      ? validated(issuesRead, (value) =>
@@ -162,25 +155,19 @@ export async function readProgress(
      : null;

  const jobs = imports.state === "ok" ? imports.data : [];
-  const taskItems = tasks.state === "ok" ? tasks.data : [];
  const issueItems = issues?.state === "ok" ? issues.data : [];
  const importState = stateOnly(imports);
  const summaryState = stateOnly(processing);
-  const tasksState = stateOnly(tasks);

  const hasReadFailure =
    importState.state !== "ok" ||
    sources.state !== "ok" ||
    summaryState.state !== "ok" ||
-    tasksState.state !== "ok" ||
    (issues !== null && issues.state !== "ok");
  const hasProblem =
    jobs.some((job) =>
      ["paused", "interrupted", "failed"].includes(job.status),
    ) ||
-    taskItems.some((task) =>
-      ["waiting_for_model", "failed"].includes(task.status),
-    ) ||
    (processing.state === "ok" &&
      (processing.data.failed > 0 ||
        processing.data.waiting_for_model > 0 ||
@@ -188,7 +175,6 @@ export async function readProgress(
    issueItems.length > 0;
  const hasActive =
    jobs.some((job) => ["pending", "running"].includes(job.status)) ||
-    taskItems.some((task) => ["pending", "running"].includes(task.status)) ||
    (processing.state === "ok" &&
      (processing.data.pending > 0 || processing.data.running > 0));

@@ -229,11 +215,6 @@ export async function readProgress(
      summary_state: summaryState,
      summary:
        processing.state === "ok" ? projectProcessing(processing.data) : null,
-      tasks_state: tasksState,
-      tasks: taskItems.slice(0, options.limit).map(projectTask),
-      returned: Math.min(taskItems.length, options.limit),
-      validated: taskItems.length,
-      output_truncated: taskItems.length > options.limit,
    },
  };
 }
@@ -289,10 +270,7 @@ export function renderProgressText(document: ProgressDocument): string {
      ? `记忆整理：待处理 ${document.generation.summary.pending} · 进行中 ${document.generation.summary.running} · 等待模型 ${document.generation.summary.waiting_for_model} · 失败 ${document.generation.summary.failed}`
      : "记忆整理：暂时无法读取",
  );
-  if (
-    document.imports.output_truncated ||
-    document.generation.output_truncated
-  ) {
+  if (document.imports.output_truncated) {
    lines.push("这里只显示汇总；完整记录请使用 --json。");
  }
  lines.push("");
@@ -353,11 +331,6 @@ export function unreadProgress(
    generation: {
      summary_state: failed,
      summary: null,
-      tasks_state: failed,
-      tasks: [],
-      returned: 0,
-      validated: 0,
-      output_truncated: false,
    },
  };
 }
@@ -465,7 +438,3 @@ function projectProcessing(
    })),
  };
 }
-
-function projectTask(task: ValidatedMemoryTask): Record<string, unknown> {
-  return { ...task };
-}
@@ -129,18 +129,6 @@ const RESPONSES = new Map<string, { status: number; data: unknown }>([
      },
    },
  ],
-  [
-    "/api/v1/memory-tasks",
-    {
-      status: 200,
-      data: [
-        memoryTask("task-pending", "pending"),
-        memoryTask("task-running", "running"),
-        memoryTask("task-model", "waiting_for_model", "model_missing"),
-        memoryTask("task-failed", "failed", "provider_unavailable"),
-      ],
-    },
-  ],
  [
    "/api/v1/import-jobs/job-running",
    { status: 200, data: importJob("job-running", "running") },
@@ -282,9 +270,9 @@ describe("CLI-19 progress and issue reads", () => {
      imported_messages: 20,
      totals_state: { state: "ok" },
    });
-    expect(
-      report.generation.tasks.map((task: { status: string }) => task.status),
-    ).toEqual(["pending", "running", "waiting_for_model", "failed"]);
+    // The retired per-task list is no longer read or reported.
+    expect(report.generation).not.toHaveProperty("tasks");
+    expect(requestedPaths()).not.toContain("/api/v1/memory-tasks");
    expect(streams.stdout()).not.toMatch(/percent|percentage|%/i);
  });

@@ -374,7 +362,7 @@ describe("CLI-19 progress and issue reads", () => {
    expect(streams.stdout()).not.toContain("SYNTHETIC_FOREIGN_ISSUE");
  });

-  it("bounds each task list and emits an explicit truncation receipt", async () => {
+  it("bounds the import job list and emits an explicit truncation receipt", async () => {
    const streams = captureStreams();
    try {
      expect(await run(["progress", "--limit", "1", "--json"], {})).toBe(1);
@@ -388,12 +376,6 @@ describe("CLI-19 progress and issue reads", () => {
      validated: 3,
      output_truncated: true,
    });
-    expect(report.generation.tasks).toHaveLength(1);
-    expect(report.generation).toMatchObject({
-      returned: 1,
-      validated: 4,
-      output_truncated: true,
-    });
  });

  it("streams bounded progress snapshots as serial NDJSON", async () => {
@@ -428,7 +410,6 @@ describe("CLI-19 progress and issue reads", () => {
      "/api/v1/import-jobs",
      "/api/v1/sources",
      "/api/v1/memory-processing",
-      "/api/v1/memory-tasks",
    ]) {
      expect(requestedPaths().filter((item) => item === path)).toHaveLength(2);
    }
@@ -500,7 +481,7 @@ describe("CLI-19 progress and issue reads", () => {
  it("preserves denied, unavailable and malformed sections in one partial report", async () => {
    transport.get.mockImplementation(async (url: string) => {
      if (url === "/api/v1/import-jobs") return { status: 503, data: {} };
-      if (url === "/api/v1/memory-tasks") return { status: 403, data: {} };
+      if (url === "/api/v1/sources") return { status: 403, data: {} };
      if (url === "/api/v1/memory-processing") {
        return { status: 200, data: { pending: { secret: "nested" } } };
      }
@@ -521,7 +502,7 @@ describe("CLI-19 progress and issue reads", () => {
      state: "invalid",
      code: "response_invalid",
    });
-    expect(report.generation.tasks_state).toEqual({
+    expect(report.imports.totals_state).toEqual({
      state: "denied",
      code: "http_403",
    });
@@ -574,7 +555,6 @@ describe("CLI-19 progress and issue reads", () => {
          "/api/v1/live-sync",
          "/api/v1/import-jobs",
          "/api/v1/memory-processing",
-          "/api/v1/memory-tasks",
        ].some((allowed) => path === allowed || path.startsWith(`${allowed}/`)),
      ),
    ).toBe(true);
@@ -634,28 +614,6 @@ function importJob(
  };
 }

-function memoryTask(
-  taskId: string,
-  status: string,
-  error: string | null = null,
-) {
-  return {
-    task_id: taskId,
-    request_id: `request-${taskId}`,
-    user_id: "local",
-    kind: "extract",
-    input_ref: "synthetic",
-    status,
-    profile_id: "profile-1",
-    profile_revision: 1,
-    profile_snapshot: {},
-    attempts: status === "failed" ? 2 : 0,
-    last_error_code: error,
-    created_at: "T1",
-    updated_at: "T2",
-  };
-}
-
 function requestedPaths(): string[] {
  return (transport.get.mock.calls as [string][]).map(([path]) => path);
 }
@@ -1,6 +1,11 @@
 import { describe, expect, it } from "vitest";

-import { renderStatusText, type StatusDocument } from "./status-report.js";
+import { composeDoctorReport, renderDoctorText } from "./doctor-report.js";
+import {
+  composeStatusReport,
+  renderStatusText,
+  type StatusDocument,
+} from "./status-report.js";

 describe("status default output", () => {
  it("formats blocked task objects recursively instead of printing object coercions", () => {
@@ -35,3 +40,40 @@ describe("status default output", () => {
    );
  });
 });
+
+describe("status and doctor on an unfinished upgrade migration", () => {
+  it("names the migration log and the issues page", async () => {
+    const report = await composeStatusReport({
+      dataDir: "/tmp/isolated/.one-memory",
+      runtime: { state: "stopped", code: "sdk_migration_required" },
+      background: null,
+      reader: null,
+    });
+
+    expect(report.runtime.log_file).toBe(
+      "/tmp/isolated/.one-memory/logs/sdk-migration.log",
+    );
+    expect(report.runtime.issues_url).toBe(
+      "https://github.com/Tencent-Hunyuan/Hy-Memory/issues/new",
+    );
+    const text = renderStatusText(report, "MemHub");
+    expect(text).toContain("sdk_migration_required");
+    expect(text).toContain("/tmp/isolated/.one-memory/logs/sdk-migration.log");
+
+    const doctor = renderDoctorText(composeDoctorReport(report, null));
+    expect(doctor).toContain(
+      "/tmp/isolated/.one-memory/logs/sdk-migration.log",
+    );
+    expect(doctor).not.toContain("帮助与诊断");
+  });
+
+  it("adds no log path for an ordinary stopped runtime", async () => {
+    const report = await composeStatusReport({
+      dataDir: "/tmp/isolated/.one-memory",
+      runtime: { state: "stopped", code: "runtime_not_running" },
+      background: null,
+      reader: null,
+    });
+    expect(report.runtime.log_file).toBeUndefined();
+  });
+});
@@ -1,5 +1,4 @@
 import type {
-  HomeSummary,
  LiveSyncHealth,
  MemoryProcessingSummary,
  ModelProfile,
@@ -9,8 +8,10 @@ import type {
  BackgroundHealth,
  RuntimeStatus,
 } from "@one-memory/runtime-control";
+import { sdkMigrationLogPath } from "@one-memory/runtime-control";
 import type { CoreReader, ReadOutcome } from "./core-client.js";
 import {
+  type SdkHomeSummary,
  validateHomeSummary,
  validateLiveSync,
  validateModelProfiles,
@@ -19,6 +20,9 @@ import {
 } from "./core-responses.js";
 import { PRODUCT } from "./version.js";

+const MIGRATION_ISSUES_URL =
+  "https://github.com/Tencent-Hunyuan/Hy-Memory/issues/new";
+
 export type SectionState =
  | "ok"
  | "denied"
@@ -42,6 +46,9 @@ export interface StatusDocument {
    host?: string;
    port?: number;
    version?: string;
+    /** Set only when an unfinished upgrade migration is why nothing serves. */
+    log_file?: string;
+    issues_url?: string;
  };
  bridge: {
    state: BackgroundHealth["state"] | "not_checked";
@@ -102,7 +109,7 @@ export async function composeStatusReport(options: {
      reader<unknown>("/api/v1/live-sync"),
      reader<unknown>("/api/v1/memory-processing"),
      reader<unknown>("/api/v1/model-profiles"),
-      reader<unknown>("/api/v1/home-summary"),
+      reader<unknown>("/api/v2/home-summary"),
    ]);
    sections.sources = section(sources, validateSources, describeSources);
    sections.sync = section(sync, validateLiveSync, describeSync);
@@ -153,6 +160,12 @@ export async function composeStatusReport(options: {
      host: options.runtime.host,
      port: options.runtime.port,
      version: options.runtime.version,
+      ...(options.runtime.code.startsWith("sdk_migration_")
+        ? {
+            log_file: sdkMigrationLogPath(options.dataDir),
+            issues_url: MIGRATION_ISSUES_URL,
+          }
+        : {}),
    },
    bridge,
    sections,
@@ -177,6 +190,13 @@ export function renderStatusText(
 ): string {
  const lines = [
    `${heading}：${runtimeLine(report)}`,
+    ...(report.runtime.log_file
+      ? [
+          `升级迁移没有完成（错误码：${report.runtime.code}），MemHub 不会启动。`,
+          `请把迁移日志附在 GitHub 反馈里发给我们：${report.runtime.log_file}`,
+          `反馈地址：${report.runtime.issues_url}`,
+        ]
+      : []),
    `资料空间：${report.data_dir}`,
    `应用连接：${BRIDGE_LABELS[report.bridge.state]}`,
  ];
@@ -347,7 +367,7 @@ function describeModel(profiles: ModelProfile[]): Record<string, unknown> {
  };
 }

-function describeMemories(summary: HomeSummary): Record<string, unknown> {
+function describeMemories(summary: SdkHomeSummary): Record<string, unknown> {
  return {
    memory_count: summary.memory_count,
  };
@@ -4,7 +4,7 @@ build-backend = "uv_build"

 [project]
 name = "one-memory"
-version = "1.0.4"
+version = "1.0.5"
 description = "MemHub local-first memory core"
 readme = "README.md"
 requires-python = ">=3.12"
@@ -12,8 +12,9 @@ dependencies = [
    "fastapi>=0.141.1",
    "filelock>=3.32.4",
    # Public SDK version validated against MemHub's permission and recall contracts.
-    "hy-memory==2.0.5.dev3",
-    # Retrieval's derived vector index directly consumes these locked versions.
+    "hy-memory==2.0.6",
+    # chromadb arrives through the HY SDK; locked here so an unrelated upgrade
+    # cannot move the SDK's store. Built-in embedding consumes numpy directly.
    "chromadb==1.5.9",
    "numpy==2.5.2",
    "httpx>=0.28.1",
@@ -42,12 +42,10 @@ TARGETED = (
    "tests/test_database_contention.py",
    "tests/test_main_database_fairness.py",
    "tests/test_soak_probe_effectiveness.py",
-    "tests/test_evidence_reconcile_primitive.py",
    "tests/test_sqlite_semantics_gate.py",
    "tests/test_store_binding.py",
    "tests/test_store_binding_ratchet.py",
    "tests/test_main_database_entry.py",
-    "tests/test_memory_processing.py",
    "tests/test_extraction_window.py",
    "tests/test_extraction_defer.py",
    "tests/test_active_memory_window.py",
@@ -2,11 +2,5 @@

 import os

-from one_memory.hy_runtime_config import configure_hy_behavior_environment
-
-# HY has a few reader switches evaluated at module import time.  Fix the
-# product-owned, non-secret behavior before any one_memory module imports HY.
-configure_hy_behavior_environment()
-
 _version = os.environ.get("ONE_MEMORY_VERSION", "").strip()
 __version__ = _version if _version else "development"
@@ -15,7 +15,7 @@ from one_memory.call_event_store import (
    CallTool,
    HostId,
 )
-from one_memory.context_models import ContextSearchResponse, MemoryContextItem, RawContextItem
+from one_memory.context_models import SdkContextSearchResponse
 from one_memory.external_memory_access import ExternalMemoryAccess
 from one_memory.recall import RecallReceipt

@@ -41,75 +41,27 @@ def _host_id(access: ExternalMemoryAccess | None) -> HostId | None:
    return cast(HostId, host) if host in HOSTS else None


-def context_call_event(
-    result: ContextSearchResponse,
+def sdk_context_call_event(
+    result: SdkContextSearchResponse,
    access: ExternalMemoryAccess | None,
+    *,
+    memory_searched: bool,
+    timed_out: bool,
 ) -> CallEvent:
-    references: list[CallReference] = []
-    seen: set[tuple[str, str, int]] = set()
-    for item in result.items:
-        if isinstance(item, RawContextItem):
-            candidates = [
-                CallReference(
-                    kind="raw",
-                    id=message.evidence_id,
-                    revision=message.revision,
-                    source_id=item.source_id,
-                )
-                for message in item.messages
-            ]
-        elif isinstance(item, MemoryContextItem):
-            candidates = [
-                CallReference(
-                    kind="memory", id=item.memory_id, revision=item.revision, source_id=None
-                )
-            ]
-        else:
-            continue
-        for reference in candidates:
-            identity = (reference.kind, reference.id, reference.revision)
-            if identity not in seen:
-                seen.add(identity)
-                references.append(reference)
-    timed_out = any(
-        path.reason == "timeout_budget" for path in (result.paths.raw, result.paths.memory)
+    # SDK items carry no Evidence/Governance revision, and the reference
+    # resolver reads the legacy store, so the event records counts only.
+    raw_count = sum(item.kind == "raw" for item in result.items)
+    memory_count = sum(item.kind == "memory" for item in result.items)
+    state: CallState = "timed_out" if timed_out else result.state
+    return _metadata_event(
+        tool="om_recall" if memory_searched else "om_conversations",
+        state=state,
+        elapsed_ms=result.total_ms,
+        access=access,
+        raw_count=raw_count,
+        memory_count=memory_count,
+        event_id=UUID(result.request_id),
    )
-    state: CallState = (
-        "partial" if timed_out and references else "timed_out" if timed_out else result.state
-    )
-    return CallEvent.model_validate(
-        {
-            "event_id": UUID(result.request_id),
-            "occurred_at": datetime.now(UTC),
-            "host_id": _host_id(access),
-            "installation_id": UUID(access.installation_id) if access else None,
-            "tool": _context_tool(result),
-            "state": state,
-            "stage": "core_result",
-            "delivery": "unknown",
-            "adoption": "unknown",
-            "host_session_id": None,
-            "injection_elapsed_ms": None,
-            "elapsed_ms": result.total_ms,
-            "raw_count": sum(isinstance(item, RawContextItem) for item in result.items),
-            "memory_count": sum(isinstance(item, MemoryContextItem) for item in result.items),
-            "references": tuple(references),
-        }
-    )
-
-
-def record_context_result(
-    store: CallEventStore,
-    result: ContextSearchResponse,
-    access: ExternalMemoryAccess | None,
-) -> None:
-    store.record(context_call_event(result, access))
-
-
-def _context_tool(result: ContextSearchResponse) -> CallTool:
-    if result.paths.memory.status not in {"not_requested", "not_authorized"}:
-        return "om_recall"
-    return "om_conversations"


 def _metadata_event(
@@ -142,26 +94,6 @@ def _metadata_event(
    )


-def conversation_call_event(
-    result: object,
-    access: ExternalMemoryAccess | None,
-    elapsed_ms: float,
-) -> CallEvent:
-    hits = tuple(getattr(result, "hits", ()) or ())
-    conversations = {
-        conversation_id
-        for hit in hits
-        if (conversation_id := getattr(hit, "conversation_id", None))
-    }
-    return _metadata_event(
-        tool="om_conversations",
-        state="grounded" if hits else "empty",
-        elapsed_ms=elapsed_ms,
-        access=access,
-        raw_count=min(len(conversations), 20),
-    )
-
-
 def ask_call_event(
    result: object,
    access: ExternalMemoryAccess | None,
@@ -255,8 +187,9 @@ def record_legacy_result(
            host_session_id=None,
            injection_elapsed_ms=None,
            elapsed_ms=elapsed_ms,
-            raw_count=0,
-            memory_count=len(receipt.hits),
+            # A raw-fallback recall returned conversation excerpts, not memories.
+            raw_count=len(receipt.hits) if receipt.mode == "raw" else 0,
+            memory_count=0 if receipt.mode == "raw" else len(receipt.hits),
            references=(),
        )
    )
@@ -23,10 +23,6 @@ class Settings(BaseSettings):
    live_sync_enabled: bool = True
    automatic_memory_processing: bool = True
    imports_enabled: bool = True
-    # Experimental and unreleased: ownership/error shutdown still requires acceptance.
-    # Resume only explicitly initialized generations; never migrate at startup.
-    retrieval_index_enabled: bool = False
-    retrieval_minimum_free_bytes: int = Field(default=256 * 1024 * 1024, ge=0)
    # Bounded metadata only: query, original text and answers are never stored.
    call_event_retention_days: int | None = Field(default=30, ge=1, le=365)
    call_event_max_events: int | None = Field(default=10_000, ge=1, le=100_000)
@@ -47,6 +47,19 @@ class ContextSearchRequest(BaseModel):
        return self


+class SdkContextSearchRequest(ContextSearchRequest):
+    """`POST /api/v2/context/search`: the three SDK recall lanes are set apart.
+
+    `limit` is the normal lane's top-k and the cap on returned items;
+    `schema_limit` and `intention_limit` are the SDK's schema (L6) and
+    proactive intention (L7) lanes. A caller's values are passed through.
+    """
+
+    limit: int = Field(default=10, ge=1, le=20, strict=True)
+    schema_limit: int = Field(default=3, ge=0, le=20, strict=True)
+    intention_limit: int = Field(default=1, ge=0, le=20, strict=True)
+
+
 class ContextMessage(BaseModel):
    model_config = ConfigDict(frozen=True, extra="forbid")

@@ -1,7 +1,6 @@
-"""Authenticated composite read route, without generation or reindex side effects."""
+"""Authenticated SDK context search route, without generation or reindex side effects."""

 import asyncio
-import sqlite3
 import time
 from collections.abc import Sequence
 from datetime import UTC, datetime
@@ -11,421 +10,26 @@ from uuid import uuid4
 from fastapi import APIRouter, HTTPException, Request, Response
 from fastapi.params import Depends

-from one_memory.background_work import (
-    in_interactive_read,
-    in_read_pool,
-    memory_read_pool,
-    raw_read_pool,
-)
-from one_memory.call_event_recorder import context_call_event
+from one_memory.call_event_recorder import sdk_context_call_event
 from one_memory.context_access import resolve_context_authority
-from one_memory.context_memory import (
-    merge_memory_candidates,
-    read_memory_context,
-    revalidate_memory_context,
-)
 from one_memory.context_models import (
-    ContextItem,
-    ContextKind,
-    ContextLaneResult,
-    ContextSearchRequest,
-    ContextSearchResponse,
-    ContextSemanticStatus,
-    MemoryContextItem,
-    RawContextItem,
    SdkContextItem,
+    SdkContextSearchRequest,
    SdkContextSearchResponse,
    SdkMemoryContextItem,
    SdkRawContextItem,
 )
-from one_memory.context_raw import revalidate_raw_context
-from one_memory.context_search import search_context
-from one_memory.diagnostic_log import diagnostic_event, diagnostic_exception
 from one_memory.external_memory_access import external_access
+from one_memory.hy_runtime_config import HY_MEMORY_READER
+from one_memory.hy_sdk_runtime import sdk_relevance_floor
+from one_memory.hy_sdk_types import (
+    HySdkMemoryRecord,
+    parse_sdk_raw_search,
+    parse_sdk_search,
+)
 from one_memory.imports.storage import EvidenceStore
-from one_memory.retrieval.raw_lane import read_raw_indexed_context
-from one_memory.semantic_deadline import hy_wait_budget, search_within

 CALL_EVENT_RESPONSE_BUDGET_MS = 250
-RESPONSE_SERIALIZATION_RESERVE_MS = 100
-INTERACTIVE_TOTAL_BUDGET_MS = 900
-RAW_LANE_BUDGET_SECONDS = 0.3
-MEMORY_LANE_BUDGET_SECONDS = 0.75
-MEMORY_FINALIZE_RESERVE_SECONDS = 0.1
-
-
-def context_router(dependencies: Sequence[Depends]) -> APIRouter:
-    router = APIRouter(prefix="/api/v1/context", dependencies=dependencies)
-
-    @router.post("/search", response_model=ContextSearchResponse)
-    async def search(
-        value: ContextSearchRequest, request: Request, response: Response
-    ) -> ContextSearchResponse:
-        response.headers["Cache-Control"] = "no-store"
-        started = time.monotonic()
-        search_budget_ms = min(
-            INTERACTIVE_TOTAL_BUDGET_MS,
-            max(50, value.timeout_ms - RESPONSE_SERIALIZATION_RESERVE_MS),
-        )
-        deadline = started + search_budget_ms / 1000
-        access = external_access(request.headers)
-        try:
-            database = cast(EvidenceStore, request.app.state.evidence_store).database_path
-            authority = await asyncio.wait_for(
-                in_interactive_read(
-                    resolve_context_authority, value, access, database, deadline=deadline
-                ),
-                timeout=max(0, deadline - time.monotonic()),
-            )
-        except HTTPException:
-            raise
-        except Exception:
-            # Authority could not be established; do not try an unbounded local lane.
-            raise HTTPException(503, "context_authority_unavailable") from None
-
-        lane_observations: dict[ContextKind, ContextLaneResult] = {}
-        lane_failures: dict[ContextKind, tuple[str, str]] = {}
-        stage_timings: dict[str, float] = {}
-
-        async def raw(end: float) -> ContextLaneResult:
-            lane_end = min(end, time.monotonic() + RAW_LANE_BUDGET_SECONDS)
-            read = asyncio.create_task(
-                in_read_pool(
-                    raw_read_pool(),
-                    read_raw_indexed_context,
-                    database,
-                    value,
-                    source_ids=authority.raw_sources,
-                    deadline=lane_end,
-                    index=getattr(request.app.state, "retrieval_runtime", None),
-                )
-            )
-            done, _ = await asyncio.wait({read}, timeout=max(0, lane_end - time.monotonic()))
-            if read not in done:
-                read.cancel()
-                read.add_done_callback(
-                    lambda task: task.exception() if not task.cancelled() else None
-                )
-                result = ContextLaneResult(
-                    status="degraded",
-                    reason="timeout_budget",
-                    searched_channels=("lexical", "semantic"),
-                    timed_out=True,
-                    truncated_by=("timeout_budget",),
-                )
-            else:
-                result = read.result()
-            lane_observations["raw"] = result
-            return result
-
-        async def memory(end: float) -> ContextLaneResult:
-            lane_end = min(end, time.monotonic() + MEMORY_LANE_BUDGET_SECONDS)
-            target = getattr(request.app.state, "recall_target", None)
-            lexical = asyncio.create_task(
-                in_read_pool(
-                    memory_read_pool(),
-                    read_memory_context,
-                    database,
-                    value,
-                    scopes=authority.memory_scopes,
-                    deadline=lane_end,
-                    semantic_ids=(),
-                    semantic_outcome="absent",
-                )
-            )
-            semantic_hits, outcome = await search_within(
-                getattr(target, "semantic_search", None),
-                value.query,
-                hy_wait_budget(lane_end - time.monotonic()),
-            )
-            done, _ = await asyncio.wait({lexical}, timeout=max(0, lane_end - time.monotonic()))
-            if lexical not in done:
-                lexical.cancel()
-                lexical.add_done_callback(
-                    lambda task: task.exception() if not task.cancelled() else None
-                )
-                result = ContextLaneResult(
-                    status="unavailable",
-                    reason="timeout_budget",
-                    searched_channels=("lexical", "semantic"),
-                    semantic_status=cast(
-                        ContextSemanticStatus,
-                        {
-                            "absent": "not_searched",
-                            "answered": "completed",
-                            "failed": "failed",
-                            "timed_out": "timed_out",
-                        }[outcome],
-                    ),
-                    timed_out=True,
-                    truncated_by=("timeout_budget",),
-                )
-                lane_observations["memory"] = result
-                return result
-            try:
-                lexical_result = lexical.result()
-            except (sqlite3.Error, OSError, ValueError) as error:
-                diagnostic_event(
-                    level="warn",
-                    component="context.search",
-                    event="context.lexical_memory_failed",
-                    outcome="degraded",
-                    error_code="lexical_memory_read_failed",
-                    **diagnostic_exception(error),
-                )
-                lexical_result = ContextLaneResult(
-                    status="unavailable",
-                    reason="read_unavailable",
-                    searched_channels=("lexical",),
-                    failure_site="memory_lexical",
-                    error_class=type(error).__name__,
-                )
-            if outcome == "absent":
-                result = lexical_result
-            elif outcome in {"failed", "timed_out"}:
-                truncated_by = list(lexical_result.truncated_by)
-                has_lexical_memory = bool(lexical_result.items)
-                if (
-                    outcome == "timed_out"
-                    and not has_lexical_memory
-                    and "timeout_budget" not in truncated_by
-                ):
-                    truncated_by.append("timeout_budget")
-                result = lexical_result.model_copy(
-                    update={
-                        "status": "degraded",
-                        "reason": "lexical_only" if has_lexical_memory else "timeout_budget",
-                        "truncated_by": tuple(truncated_by),
-                        "searched_channels": ("lexical", "semantic"),
-                        "semantic_status": "timed_out" if outcome == "timed_out" else "failed",
-                        "timed_out": outcome == "timed_out" and not has_lexical_memory,
-                    }
-                )
-            else:
-                semantic_request = value.model_copy(update={"query": ""})
-                enriched = asyncio.create_task(
-                    in_interactive_read(
-                        read_memory_context,
-                        database,
-                        semantic_request,
-                        scopes=authority.memory_scopes,
-                        deadline=lane_end,
-                        semantic_hits=semantic_hits,
-                        semantic_outcome=outcome,
-                    )
-                )
-                done, _ = await asyncio.wait(
-                    {enriched},
-                    timeout=max(
-                        0,
-                        lane_end - MEMORY_FINALIZE_RESERVE_SECONDS - time.monotonic(),
-                    ),
-                )
-                if enriched not in done:
-                    enriched.cancel()
-                    enriched.add_done_callback(
-                        lambda task: task.exception() if not task.cancelled() else None
-                    )
-                    result = lexical_result.model_copy(
-                        update={
-                            "status": "degraded",
-                            "reason": "lexical_only",
-                            "searched_channels": ("lexical", "semantic"),
-                            "semantic_status": "completed",
-                            "timed_out": False,
-                        }
-                    )
-                else:
-                    try:
-                        semantic_result = enriched.result()
-                    except (sqlite3.Error, OSError, ValueError) as error:
-                        diagnostic_event(
-                            level="warn",
-                            component="context.search",
-                            event="context.semantic_enrichment_failed",
-                            outcome="degraded",
-                            error_code="semantic_enrichment_read_failed",
-                            **diagnostic_exception(error),
-                        )
-                        result = lexical_result.model_copy(
-                            update={
-                                "status": "degraded",
-                                "reason": "lexical_only",
-                                "searched_channels": ("lexical", "semantic"),
-                                "semantic_status": "completed",
-                                "timed_out": False,
-                            }
-                        )
-                        lane_observations["memory"] = result
-                        return result
-                    if semantic_result.status == "unavailable" and lexical_result.items:
-                        result = lexical_result.model_copy(
-                            update={
-                                "status": "degraded",
-                                "reason": "lexical_only",
-                                "searched_channels": ("lexical", "semantic"),
-                                "semantic_status": "completed",
-                                "timed_out": False,
-                            }
-                        )
-                        lane_observations["memory"] = result
-                        return result
-                    semantic_items = tuple(
-                        item
-                        for item in semantic_result.items
-                        if isinstance(item, MemoryContextItem)
-                    )
-                    lexical_items = tuple(
-                        item for item in lexical_result.items if isinstance(item, MemoryContextItem)
-                    )
-                    combined_count = len(
-                        {item.memory_id for item in (*semantic_items, *lexical_items)}
-                    )
-                    merged = merge_memory_candidates(
-                        lexical_items,
-                        semantic_items,
-                        limit=value.limit + 1,
-                    )
-                    truncated_by = list(semantic_result.truncated_by)
-                    if combined_count > len(merged) and "count_budget" not in truncated_by:
-                        truncated_by.append("count_budget")
-                    result = semantic_result.model_copy(
-                        update={
-                            "items": tuple(merged),
-                            # An answered semantic search may legitimately
-                            # return no IDs. Lexical can still have a grounded
-                            # hit; retaining semantic's `empty` status while
-                            # attaching that hit violates the lane contract
-                            # and makes the aggregator reject the whole lane.
-                            "status": "ok" if merged else semantic_result.status,
-                            "truncated_by": tuple(truncated_by),
-                        }
-                    )
-            lane_observations["memory"] = result
-            return result
-
-        async def finalizer(
-            kind: ContextKind, items: tuple[ContextItem, ...], end: float
-        ) -> tuple[ContextItem, ...]:
-            if kind == "raw":
-                return await in_interactive_read(
-                    revalidate_raw_context,
-                    database,
-                    tuple(item for item in items if isinstance(item, RawContextItem)),
-                    source_ids=authority.raw_sources,
-                    deadline=end,
-                )
-            return await in_interactive_read(
-                revalidate_memory_context,
-                database,
-                tuple(item for item in items if isinstance(item, MemoryContextItem)),
-                scopes=authority.memory_scopes,
-                deadline=end,
-            )
-
-        result = await search_context(
-            value,
-            raw_reader=raw if authority.raw_allowed else None,
-            memory_reader=memory if authority.memory_allowed else None,
-            finalizer=finalizer,
-            deadline=deadline,
-            timing_observer=stage_timings.update,
-            failure_observer=lambda kind, site, error_class: lane_failures.__setitem__(
-                kind, (site, error_class)
-            ),
-        )
-        final = result.model_copy(update={"total_ms": (time.monotonic() - started) * 1000})
-        serialization_started = time.monotonic()
-        final.model_dump_json()
-        stage_timings["serialization_ms"] = (time.monotonic() - serialization_started) * 1000
-        for phase, duration in stage_timings.items():
-            diagnostic_event(
-                level="info",
-                component="context.search",
-                event="context_stage.completed",
-                outcome="succeeded",
-                request_id=final.request_id,
-                duration_ms=max(0, round(duration)),
-                details={"phase": phase},
-            )
-        for kind in ("raw", "memory"):
-            path = getattr(final.paths, kind)
-            observed = lane_observations.get(kind)
-            failure = lane_failures.get(kind)
-            lexical_hits = (
-                sum(
-                    1
-                    for item in observed.items
-                    if isinstance(item, MemoryContextItem) and "lexical" in item.matched_by
-                )
-                if observed is not None and kind == "memory"
-                else 0
-            )
-            semantic_hits = (
-                sum(
-                    1
-                    for item in observed.items
-                    if isinstance(item, MemoryContextItem)
-                    and "governed_semantic" in item.matched_by
-                )
-                if observed is not None and kind == "memory"
-                else 0
-            )
-            timed_out = path.reason == "timeout_budget" or (
-                observed.timed_out if observed is not None else False
-            )
-            diagnostic_failure_site = failure[0] if failure is not None else None
-            diagnostic_error_class = failure[1] if failure is not None else None
-            if observed is not None:
-                diagnostic_failure_site = diagnostic_failure_site or observed.failure_site
-                diagnostic_error_class = diagnostic_error_class or observed.error_class
-            diagnostic_event(
-                level="warn" if path.status in {"degraded", "unavailable"} else "info",
-                component="context.search",
-                event="context_lane.completed",
-                outcome=path.status,
-                request_id=final.request_id,
-                duration_ms=round(path.elapsed_ms),
-                details={
-                    "operation_kind": kind,
-                    "mode": ",".join(observed.searched_channels) if observed else "none",
-                    "status": path.status,
-                    "index_complete": observed.index_complete if observed else None,
-                    **(
-                        {"lexical_hits": lexical_hits, "semantic_hits": semantic_hits}
-                        if kind == "memory"
-                        else {}
-                    ),
-                    "reason": path.reason,
-                    "truncated": path.truncated,
-                    "timed_out": timed_out,
-                    **(
-                        {
-                            "failure_site": diagnostic_failure_site,
-                            "error_class": diagnostic_error_class,
-                        }
-                        if diagnostic_failure_site is not None
-                        and diagnostic_error_class is not None
-                        else {}
-                    ),
-                },
-            )
-        store = getattr(request.app.state, "call_event_store", None)
-        if store is not None and value.query != "__om_probe__":
-            event = context_call_event(final, access)
-            try:
-                await asyncio.to_thread(
-                    store.record,
-                    event,
-                    busy_timeout_ms=CALL_EVENT_RESPONSE_BUDGET_MS,
-                )
-                response.headers["X-One-Memory-Call-Recorded"] = "true"
-            except Exception:
-                response.headers["X-One-Memory-Call-Recorded"] = "false"
-                request.app.state.call_event_recording_error = True
-        return final
-
-    return router


 def sdk_context_router(dependencies: Sequence[Depends]) -> APIRouter:
@@ -434,7 +38,7 @@ def sdk_context_router(dependencies: Sequence[Depends]) -> APIRouter:

    @router.post("/search", response_model=SdkContextSearchResponse)
    async def search(
-        value: ContextSearchRequest,
+        value: SdkContextSearchRequest,
        request: Request,
        response: Response,
    ) -> SdkContextSearchResponse:
@@ -455,19 +59,33 @@ def sdk_context_router(dependencies: Sequence[Depends]) -> APIRouter:
        except Exception:
            raise HTTPException(503, "context_authority_unavailable") from None

-        bridge = getattr(request.app.state, "sdk_library_bridge", None)
-        if bridge is None:
+        client_handle = getattr(request.app.state, "sdk_client_handle", None)
+        if client_handle is None:
            raise HTTPException(503, "sdk_context_unavailable")
+        try:
+            client = await client_handle.client()
+        except Exception:
+            raise HTTPException(503, "sdk_context_unavailable") from None

        async def raw_lane() -> tuple[SdkRawContextItem, ...]:
            if "raw" not in value.kinds or not authority.raw_allowed:
                return ()
-            hits = await bridge.search_sdk_raw(
+            raw = await asyncio.to_thread(
+                client.search,
                value.query,
-                user_id="local",
-                source_ids=tuple(authority.raw_sources or ()),
+                user_ids=["local"],
+                agent_ids=[],
+                include_sources=list(authority.raw_sources or ()),
                limit=value.limit,
+                min_score=0.0,
+                schema_min_score=0.0,
+                schema_limit=0,
+                intention_limit=0,
+                scene="chat",
+                reader=HY_MEMORY_READER,
+                layers=["l1_raw"],
            )
+            hits = parse_sdk_raw_search(raw)
            return tuple(
                SdkRawContextItem(
                    raw_id=hit.raw_id,
@@ -487,15 +105,24 @@ def sdk_context_router(dependencies: Sequence[Depends]) -> APIRouter:
        async def memory_lane() -> tuple[SdkMemoryContextItem, ...]:
            if "memory" not in value.kinds or not authority.memory_allowed:
                return ()
-            hits = await bridge.search_memories(
+            raw = await asyncio.to_thread(
+                client.search,
                value.query,
-                user_id="local",
-                agent_id="*",
+                user_ids=["local"],
+                agent_ids=[],
                limit=value.limit,
+                min_score=sdk_relevance_floor(),
+                schema_min_score=sdk_relevance_floor(),
+                schema_limit=value.schema_limit,
+                intention_limit=value.intention_limit,
+                scene="normal",
+                reader=HY_MEMORY_READER,
            )
+            hits = parse_sdk_search(raw, error_code="hy_memory_search_shape_unknown")
            items: list[SdkMemoryContextItem] = []
            for hit in hits:
-                memory = await bridge.get_sdk_memory(hit.memory_id)
+                record = await asyncio.to_thread(client.get, hit.memory_id)
+                memory = None if record is None else HySdkMemoryRecord.from_sdk(record)
                if memory is None or memory.status != "active":
                    continue
                items.append(
@@ -521,6 +148,29 @@ def sdk_context_router(dependencies: Sequence[Depends]) -> APIRouter:
                )
            return tuple(items)

+        async def record(final: SdkContextSearchResponse, *, timed_out: bool) -> None:
+            # Same boundary as v2 ask: only a host's call is history. The
+            # renderer's own searches are the person reading their library.
+            store = getattr(request.app.state, "call_event_store", None)
+            if store is None or access is None or value.query == "__om_probe__":
+                return
+            event = sdk_context_call_event(
+                final,
+                access,
+                memory_searched=authority.memory_allowed,
+                timed_out=timed_out,
+            )
+            try:
+                await asyncio.to_thread(
+                    store.record,
+                    event,
+                    busy_timeout_ms=CALL_EVENT_RESPONSE_BUDGET_MS,
+                )
+                response.headers["X-One-Memory-Call-Recorded"] = "true"
+            except Exception:
+                response.headers["X-One-Memory-Call-Recorded"] = "false"
+                request.app.state.call_event_recording_error = True
+
        lanes = (raw_lane(), memory_lane())
        try:
            raw_items, memory_items = await asyncio.wait_for(
@@ -528,7 +178,7 @@ def sdk_context_router(dependencies: Sequence[Depends]) -> APIRouter:
                timeout=value.timeout_ms / 1000,
            )
        except TimeoutError:
-            return SdkContextSearchResponse(
+            final = SdkContextSearchResponse(
                request_id=str(uuid4()),
                state="unavailable",
                items=(),
@@ -536,6 +186,8 @@ def sdk_context_router(dependencies: Sequence[Depends]) -> APIRouter:
                truncated=True,
                total_ms=(time.monotonic() - started) * 1000,
            )
+            await record(final, timed_out=True)
+            return final
        partial = isinstance(raw_items, Exception) or isinstance(memory_items, Exception)
        raw_candidates = (
            () if isinstance(raw_items, Exception) else cast(tuple[SdkContextItem, ...], raw_items)
@@ -567,7 +219,7 @@ def sdk_context_router(dependencies: Sequence[Depends]) -> APIRouter:
                truncated = True
            selected.append(item)
            used += len(item.content)
-        return SdkContextSearchResponse(
+        final = SdkContextSearchResponse(
            request_id=str(uuid4()),
            state=("partial" if partial else "grounded" if selected else "empty"),
            items=tuple(selected),
@@ -575,5 +227,7 @@ def sdk_context_router(dependencies: Sequence[Depends]) -> APIRouter:
            truncated=truncated,
            total_ms=(time.monotonic() - started) * 1000,
        )
+        await record(final, timed_out=False)
+        return final

    return router
@@ -1,464 +0,0 @@
-"""Bounded two-lane orchestration; readers and finalizers own trusted read access."""
-
-import asyncio
-import time
-from collections.abc import Awaitable, Callable, Mapping
-from dataclasses import dataclass, field
-from uuid import uuid4
-
-from pydantic import TypeAdapter
-
-from one_memory.context_models import (
-    ContextItem,
-    ContextKind,
-    ContextLaneResult,
-    ContextLimit,
-    ContextPath,
-    ContextPaths,
-    ContextSearchRequest,
-    ContextSearchResponse,
-    MemoryContextItem,
-    RawContextItem,
-)
-
-Reader = Callable[[float], Awaitable[ContextLaneResult]]
-Finalizer = Callable[
-    [ContextKind, tuple[ContextItem, ...], float], Awaitable[tuple[ContextItem, ...]]
-]
-TimingObserver = Callable[[Mapping[str, float]], None]
-FailureObserver = Callable[[ContextKind, str, str], None]
-_KINDS: tuple[ContextKind, ...] = ("raw", "memory")
-_ITEMS = TypeAdapter(tuple[ContextItem, ...])
-
-
-@dataclass
-class _Lane:
-    result: ContextLaneResult
-    elapsed_ms: float = 0
-    limits: list[ContextLimit] = field(default_factory=list)
-
-
-def _identity(item: ContextItem) -> tuple[str, ...]:
-    if isinstance(item, RawContextItem):
-        return (item.kind, item.source_id, item.conversation_id, item.anchor_evidence_id)
-    return (item.kind, item.memory_id)
-
-
-def _validate_items(kind: ContextKind, items: tuple[ContextItem, ...]) -> None:
-    if len(items) > 21 or any(item.kind != kind for item in items):
-        raise ValueError("invalid lane items")
-    for item in items:
-        if isinstance(item, RawContextItem) and item.anchor_evidence_id not in {
-            message.evidence_id for message in item.messages
-        }:
-            raise ValueError("missing raw anchor")
-
-
-async def _read(reader: Reader, kind: ContextKind, deadline: float) -> ContextLaneResult:
-    value = await reader(deadline)
-    # Revalidate even BaseModel instances: model_construct/model_copy bypass validation.
-    result = ContextLaneResult.model_validate(value.model_dump())
-    _validate_items(kind, result.items)
-    if result.status not in ("ok", "empty", "degraded", "unavailable"):
-        raise ValueError("reader cannot decide requested authority")
-    if result.status in ("empty", "unavailable") and result.items:
-        raise ValueError("non-result lane has items")
-    return result
-
-
-async def _finalize(
-    finalizer: Finalizer, kind: ContextKind, items: tuple[ContextItem, ...], deadline: float
-) -> tuple[ContextItem, ...]:
-    value = await finalizer(kind, items, deadline)
-    if not isinstance(value, tuple):
-        raise ValueError("invalid finalization result")
-    checked = _ITEMS.validate_python([item.model_dump() for item in value])
-    _validate_items(kind, checked)
-    if not {_identity(item) for item in checked} <= {_identity(item) for item in items}:
-        raise ValueError("finalizer introduced an unselected identity")
-    return checked
-
-
-async def _collect(
-    jobs: dict[ContextKind, Awaitable[object]], deadline: float
-) -> tuple[
-    dict[ContextKind, object],
-    dict[ContextKind, BaseException],
-    set[ContextKind],
-    dict[ContextKind, float],
-]:
-    """Cancel and await every owned task on deadline, failure or caller cancellation."""
-
-    durations: dict[ContextKind, float] = {}
-
-    async def measured(kind: ContextKind, job: Awaitable[object]) -> object:
-        started = time.monotonic()
-        try:
-            value = await job
-            if time.monotonic() >= deadline:
-                raise TimeoutError
-            return value
-        finally:
-            durations[kind] = (time.monotonic() - started) * 1000
-
-    tasks = {kind: asyncio.create_task(measured(kind, job)) for kind, job in jobs.items()}
-    values: dict[ContextKind, object] = {}
-    failed: dict[ContextKind, BaseException] = {}
-    timed_out: set[ContextKind] = set()
-    try:
-        if not tasks:
-            return values, failed, timed_out, durations
-        done, pending = await asyncio.wait(
-            tasks.values(), timeout=max(0, deadline - time.monotonic())
-        )
-        for kind, task in tasks.items():
-            if task in pending:
-                timed_out.add(kind)
-            elif task in done:
-                if not task.cancelled() and isinstance(task.exception(), TimeoutError):
-                    timed_out.add(kind)
-                elif task.cancelled() or task.exception() is not None:
-                    failed[kind] = task.exception() or asyncio.CancelledError()
-                else:
-                    values[kind] = task.result()
-        return values, failed, timed_out, durations
-    finally:
-        for task in tasks.values():
-            if not task.done():
-                task.cancel()
-        if tasks:
-            # Cooperative readers finish immediately and still get their
-            # normal cleanup. A bridge that suppresses cancellation must not
-            # hold a host response past its hard deadline while an underlying
-            # SDK thread winds down.
-            cleanup: asyncio.Future[list[object]] = asyncio.gather(
-                *tasks.values(), return_exceptions=True
-            )
-            current = asyncio.current_task()
-            if current is not None and current.cancelling():
-                await cleanup
-            else:
-                try:
-                    await asyncio.wait_for(asyncio.shield(cleanup), timeout=0.05)
-                except TimeoutError:
-                    cleanup.add_done_callback(_consume_cleanup)
-
-
-def _consume_cleanup(task: asyncio.Future[list[object]]) -> None:
-    if not task.cancelled():
-        task.exception()
-
-
-def _deduplicate(
-    items: tuple[ContextItem, ...], limits: list[ContextLimit]
-) -> tuple[ContextItem, ...]:
-    unique: dict[tuple[str, ...], ContextItem] = {}
-    for item in items:
-        key = _identity(item)
-        previous = unique.get(key)
-        if isinstance(previous, MemoryContextItem) and isinstance(item, MemoryContextItem):
-            if item.revision != previous.revision:
-                # Never graft older evidence onto a newer memory version.
-                unique[key] = item if item.revision > previous.revision else previous
-                continue
-            evidence = tuple(dict.fromkeys((*previous.evidence_ids, *item.evidence_ids)))
-            if len(evidence) > 128:
-                limits.append("metadata_budget")
-            unique[key] = previous.model_copy(
-                update={
-                    "evidence_ids": evidence[:128],
-                    "truncated": previous.truncated or item.truncated or len(evidence) > 128,
-                }
-            )
-        elif isinstance(previous, RawContextItem) and isinstance(item, RawContextItem):
-            messages = {message.evidence_id: message for message in previous.messages}
-            for message in item.messages:
-                if (
-                    message.evidence_id not in messages
-                    or message.revision > messages[message.evidence_id].revision
-                ):
-                    messages[message.evidence_id] = message
-            if len(messages) > 3:
-                limits.append("metadata_budget")
-            # The first window already contains the anchor; dict order preserves it.
-            unique[key] = previous.model_copy(
-                update={
-                    "messages": tuple(messages.values())[:3],
-                    "truncated": previous.truncated or item.truncated or len(messages) > 3,
-                }
-            )
-        else:
-            unique[key] = item
-    # Reserve each raw anchor for its own window before deduplicating context messages.
-    anchors = {
-        (item.source_id, item.conversation_id, item.anchor_evidence_id)
-        for item in unique.values()
-        if isinstance(item, RawContextItem)
-    }
-    seen: set[tuple[str, ...]] = set()
-    result: list[ContextItem] = []
-    for item in unique.values():
-        if isinstance(item, MemoryContextItem):
-            result.append(
-                item.model_copy(update={"evidence_ids": tuple(dict.fromkeys(item.evidence_ids))})
-            )
-            continue
-        retained_messages = []
-        for message in item.messages:
-            key = (item.source_id, item.conversation_id, message.evidence_id)
-            if key in seen or (key in anchors and message.evidence_id != item.anchor_evidence_id):
-                continue
-            seen.add(key)
-            retained_messages.append(message)
-        result.append(item.model_copy(update={"messages": tuple(retained_messages)}))
-    return tuple(result)
-
-
-def _characters(item: ContextItem) -> int:
-    if isinstance(item, MemoryContextItem):
-        return len(item.content)
-    return sum(len(message.text) for message in item.messages)
-
-
-def _shares(lengths: list[int], budget: int) -> list[int]:
-    """Water-fill allocations so a large first item cannot consume every character."""
-    allocations = [0] * len(lengths)
-    pending = list(range(len(lengths)))
-    while pending and budget:
-        share, remainder = divmod(budget, len(pending))
-        next_pending = []
-        for position, index in enumerate(pending):
-            added = min(lengths[index] - allocations[index], share + (position < remainder))
-            allocations[index] += added
-            budget -= added
-            if allocations[index] < lengths[index]:
-                next_pending.append(index)
-        pending = next_pending
-    return allocations
-
-
-def _clip(item: ContextItem, budget: int) -> ContextItem:
-    if _characters(item) <= budget:
-        return item
-    if isinstance(item, MemoryContextItem):
-        return item.model_copy(update={"content": item.content[:budget], "truncated": True})
-    # The first raw message is the lexical anchor. Give it the first share of
-    # the budget, then water-fill the adjacent context; otherwise a long
-    # neighbor can consume the whole window and hide the answer.
-    anchor = item.messages[0]
-    anchor_budget = min(len(anchor.text), budget)
-    remaining = budget - anchor_budget
-    adjacent = list(item.messages[1:])
-    shares = [anchor_budget, *_shares([len(message.text) for message in adjacent], remaining)]
-    return item.model_copy(
-        update={
-            "truncated": True,
-            "messages": tuple(
-                message.model_copy(
-                    update={
-                        "text": message.text[:share],
-                        "truncated": message.truncated or len(message.text) > share,
-                    }
-                )
-                for message, share in zip(item.messages, shares, strict=True)
-            ),
-        }
-    )
-
-
-async def search_context(
-    request: ContextSearchRequest,
-    *,
-    raw_reader: Reader | None,
-    memory_reader: Reader | None,
-    finalizer: Finalizer,
-    deadline: float | None = None,
-    timing_observer: TimingObserver | None = None,
-    failure_observer: FailureObserver | None = None,
-) -> ContextSearchResponse:
-    request = ContextSearchRequest.model_validate(request.model_dump())
-    started = time.monotonic()
-    deadline = min(
-        deadline if deadline is not None else float("inf"), started + request.timeout_ms / 1000
-    )
-    lanes: dict[ContextKind, _Lane] = {}
-    jobs: dict[ContextKind, Awaitable[object]] = {}
-    readers: dict[ContextKind, Reader | None] = {"raw": raw_reader, "memory": memory_reader}
-    for kind, reader in readers.items():
-        if kind not in request.kinds:
-            lanes[kind] = _Lane(ContextLaneResult(status="not_requested", reason="not_requested"))
-        elif reader is None:
-            lanes[kind] = _Lane(
-                ContextLaneResult(status="not_authorized", reason="permission_required")
-            )
-        elif time.monotonic() >= deadline:
-            lanes[kind] = _Lane(
-                ContextLaneResult(status="degraded", reason="timeout_budget", timed_out=True),
-                limits=["timeout_budget"],
-            )
-        else:
-            jobs[kind] = _read(reader, kind, deadline)
-    values, failed, timed_out, read_ms = await _collect(jobs, deadline)
-    for kind in jobs:
-        result = values.get(kind)
-        if kind in timed_out:
-            result = ContextLaneResult(
-                status="degraded",
-                reason="timeout_budget",
-                truncated_by=("timeout_budget",),
-                timed_out=True,
-            )
-        elif kind in failed or not isinstance(result, ContextLaneResult):
-            error = failed.get(kind)
-            if failure_observer is not None:
-                failure_observer(
-                    kind,
-                    "collect",
-                    type(error).__name__ if error is not None else "InvalidLaneResult",
-                )
-            result = ContextLaneResult(
-                status="unavailable",
-                reason="read_unavailable",
-                failure_site="collect",
-                error_class=type(error).__name__ if error is not None else "InvalidLaneResult",
-            )
-        lanes[kind] = _Lane(result, read_ms.get(kind, 0), list(dict.fromkeys(result.truncated_by)))
-
-    final_jobs: dict[ContextKind, Awaitable[object]] = {}
-    for kind in jobs:
-        lane = lanes[kind]
-        if lane.result.status not in ("ok", "empty", "degraded"):
-            continue
-        if time.monotonic() >= deadline:
-            lane.result = ContextLaneResult(
-                status="degraded", reason="timeout_budget", timed_out=True
-            )
-            lane.limits.append("timeout_budget")
-        else:
-            final_jobs[kind] = _finalize(finalizer, kind, lane.result.items, deadline)
-    final_values, final_failed, final_timed_out, revalidation_ms = await _collect(
-        final_jobs, deadline
-    )
-    for kind in final_jobs:
-        lane = lanes[kind]
-        lane.elapsed_ms += revalidation_ms.get(kind, 0)
-        if kind in final_timed_out:
-            lane.result = ContextLaneResult(
-                status="unavailable", reason="timeout_budget", timed_out=True
-            )
-            lane.limits.append("timeout_budget")
-        elif kind in final_failed:
-            error = final_failed[kind]
-            if failure_observer is not None:
-                failure_observer(kind, "finalize", type(error).__name__)
-            lane.result = ContextLaneResult(
-                status="unavailable",
-                reason="read_unavailable",
-                failure_site="finalize",
-                error_class=type(error).__name__,
-            )
-        else:
-            final_items = _ITEMS.validate_python(final_values[kind])
-            changed = {_identity(item): item for item in lane.result.items} != {
-                _identity(item): item for item in final_items
-            }
-            lane.result = lane.result.model_copy(
-                update={
-                    "items": final_items,
-                    "status": "degraded" if changed else lane.result.status,
-                    "reason": ("evidence_changed" if kind == "raw" else "memory_changed")
-                    if changed
-                    else lane.result.reason,
-                }
-            )
-
-    # Nothing previously verified may escape once the common release deadline expired.
-    if time.monotonic() >= deadline:
-        for kind in jobs:
-            if lanes[kind].result.status in ("ok", "empty", "degraded"):
-                lanes[kind].result = ContextLaneResult(
-                    status="degraded", reason="timeout_budget", timed_out=True
-                )
-                lanes[kind].limits.append("timeout_budget")
-    fusion_started = time.monotonic()
-    queues = {kind: _deduplicate(lanes[kind].result.items, lanes[kind].limits) for kind in _KINDS}
-    interleaved = [
-        queues[kind][index]
-        for index in range(max(map(len, queues.values()), default=0))
-        for kind in _KINDS
-        if index < len(queues[kind])
-    ]
-    selected = interleaved[: request.limit]
-    for item in interleaved[request.limit :]:
-        lanes[item.kind].limits.append("count_budget")
-    shares = _shares([_characters(item) for item in selected], request.character_budget)
-    items = tuple(_clip(item, share) for item, share in zip(selected, shares, strict=True))
-    for before, after in zip(selected, items, strict=True):
-        if _characters(before) != _characters(after):
-            lanes[after.kind].limits.append("character_budget")
-    if time.monotonic() >= deadline:
-        items = ()
-        for kind in jobs:
-            if lanes[kind].result.status in ("ok", "empty", "degraded"):
-                lanes[kind].result = ContextLaneResult(
-                    status="degraded", reason="timeout_budget", timed_out=True
-                )
-                lanes[kind].limits.append("timeout_budget")
-                lanes[kind].elapsed_ms = (time.monotonic() - started) * 1000
-    paths: dict[ContextKind, ContextPath] = {}
-    for kind in _KINDS:
-        lane = lanes[kind]
-        emitted = [item for item in items if item.kind == kind]
-        limits = tuple(dict.fromkeys(lane.limits))
-        truncated = bool(limits) or any(item.truncated for item in emitted)
-        status = lane.result.status
-        benign_limits = set(limits) <= {"count_budget", "character_budget"}
-        benign_degradation = (
-            not lane.result.timed_out
-            and benign_limits
-            and (
-                lane.result.reason is None
-                or (kind == "raw" and lane.result.reason == "lexical_only")
-            )
-        )
-        if status in ("ok", "empty") or (status == "degraded" and benign_degradation):
-            status = "ok" if emitted or lane.result.items else "empty"
-        paths[kind] = ContextPath(
-            status=status,
-            reason=lane.result.reason,
-            elapsed_ms=lane.elapsed_ms,
-            returned_count=len(emitted),
-            truncated=truncated,
-            truncated_by=limits,
-        )
-    requested_paths = [paths[kind] for kind in request.kinds]
-    completed = any(path.status in ("ok", "empty", "degraded") for path in requested_paths)
-    incomplete = any(path.status not in ("ok", "empty") for path in requested_paths)
-    response = ContextSearchResponse(
-        request_id=str(uuid4()),
-        state="unavailable"
-        if not completed
-        else "partial"
-        if incomplete
-        else "grounded"
-        if items
-        else "empty",
-        paths=ContextPaths(raw=paths["raw"], memory=paths["memory"]),
-        items=items,
-        used_characters=sum(map(_characters, items)),
-        truncated=any(path.truncated for path in paths.values()),
-        total_ms=(time.monotonic() - started) * 1000,
-    )
-    checked = ContextSearchResponse.model_validate(response.model_dump())
-    if timing_observer is not None:
-        timing_observer(
-            {
-                "raw_ms": read_ms.get("raw", 0),
-                "memory_ms": read_ms.get("memory", 0),
-                "raw_revalidation_ms": revalidation_ms.get("raw", 0),
-                "memory_revalidation_ms": revalidation_ms.get("memory", 0),
-                "fusion_ms": (time.monotonic() - fusion_started) * 1000,
-                "total_ms": checked.total_ms,
-            }
-        )
-    return checked
@@ -22,7 +22,7 @@ from dataclasses import dataclass
 from datetime import UTC, datetime
 from functools import partial
 from pathlib import Path
-from typing import Any, Literal, Protocol, cast
+from typing import Any, Literal, cast

 from one_memory.background_work import interactive_read_pool
 from one_memory.context_expand import ExpansionAuthority, ExpansionRefSigner
@@ -36,23 +36,8 @@ from one_memory.retrieval.models import (
    RawEvidence,
    RetrievalPath,
    RetrievalResponse,
-    TimeRange,
 )
 from one_memory.retrieval.query_plan import plan_local_query
-from one_memory.retrieval.raw_candidates import RawCandidateBatch
-from one_memory.retrieval.raw_evidence import materialize_raw_candidates
-
-
-class RawIndexLookup(Protocol):
-    def lookup_raw(
-        self,
-        query: str,
-        *,
-        source_ids: frozenset[str],
-        deadline: float,
-        time_range: TimeRange | None = None,
-    ) -> tuple[RawCandidateBatch, bool]: ...
-

 ExpandRefs = tuple[ExpansionRefSigner, ExpansionAuthority]

@@ -106,11 +91,6 @@ class LaneResult:
    coverage_limited: bool = False


-def _semantic_only_raw(item: RawEvidence) -> bool:
-    """Semantic overlap alone is a candidate, not a confirmed recall item."""
-    return set(item.matched_by) == {"semantic"}
-
-
@dataclass(frozen=True)
 class StageTimings:
    raw_ms: float = 0.0
@@ -191,12 +171,6 @@ def _managed(
    return selected, truncated


-def _group_missing_requirements(item: RawEvidence | GovernedEvidence) -> list[str]:
-    if isinstance(item, RawEvidence) and _semantic_only_raw(item):
-        return ["relevance_confirmation"]
-    return []
-
-
 def _path(
    status: str,
    reason: str | None,
@@ -296,8 +270,7 @@ def read_lexical_raw_lane(
        ("lexical", "object") if plan.object_keys else ("lexical",),
        truncated,
        # Starvation means: this lane matched, then dropped the item whole. An
-        # ordinary no-match must stay empty, and an unconfigured semantic lane
-        # is not starvation.
+        # ordinary no-match must stay empty.
        starved and not items,
        probe_saturated,
    )
@@ -564,97 +537,6 @@ def warm_raw_read_path(
        return False


-def read_semantic_raw_lane(
-    index: RawIndexLookup | None,
-    database_path: Path,
-    query: CoordinatorQuery,
-    *,
-    source_ids: frozenset[str],
-    deadline: float,
-) -> LaneResult:
-    if not source_ids:
-        return LaneResult((), "unavailable", "permission_required", (), False)
-    # No derived index is the no-model baseline, not a failure: the lexical lane
-    # must still answer. Semantic is genuinely unsearched in that state, so it
-    # must not claim a channel or a coverage result it never produced.
-    if index is None:
-        return LaneResult((), "not_searched", None, (), False)
-    try:
-        batch, unpublished = index.lookup_raw(
-            query.query,
-            source_ids=source_ids,
-            deadline=deadline,
-            time_range=plan_local_query(query.query).time_range,
-        )
-    except Exception:
-        return LaneResult((), "unavailable", "index_unavailable", ("semantic",), False)
-    if not isinstance(batch, RawCandidateBatch):
-        return LaneResult((), "unavailable", "index_unavailable", ("semantic",), False)
-    if batch.reason == "not_authorized":
-        return LaneResult((), "unavailable", "permission_required", ("semantic",), False)
-    if batch.reason == "deadline":
-        return LaneResult((), "unavailable", "timeout_budget", ("semantic",), True)
-    if batch.reason in {"vector_unavailable", "index_unavailable"}:
-        return LaneResult((), "unavailable", batch.reason, ("semantic",), False)
-    store, generation = _index_state(index, database_path)
-    from one_memory.retrieval.index_store import RetrievalIndexStore
-
-    store = cast("RetrievalIndexStore | None", store)
-    if store is None or generation is None:
-        return LaneResult((), "unavailable", "index_unavailable", ("semantic",), False)
-    try:
-        materialized = materialize_raw_candidates(
-            store,
-            batch,
-            generation_id=generation,
-            source_ids=source_ids,
-            deadline=deadline,
-            character_budget=query.character_budget,
-        )
-    except Exception:
-        return LaneResult((), "unavailable", "index_unavailable", ("semantic",), False)
-    if materialized.reason == "deadline":
-        return LaneResult((), "unavailable", "timeout_budget", ("semantic",), True)
-    if materialized.reason in {"index_unavailable", "not_authorized"}:
-        return LaneResult((), "unavailable", materialized.reason, ("semantic",), False)
-    complete = materialized.complete and not unpublished
-    if materialized.reason == "evidence_changed":
-        status: Literal["ok", "empty", "unavailable", "not_searched"] = "unavailable"
-        reason = "evidence_changed"
-    elif not materialized.items and not complete:
-        status, reason = "unavailable", "index_incomplete"
-    elif materialized.reason == "character_budget":
-        status, reason = "ok", None
-    else:
-        status, reason = ("ok" if materialized.items else "empty"), None
-    return LaneResult(
-        tuple(materialized.items),
-        status,
-        reason,
-        ("semantic",),
-        materialized.reason == "character_budget",
-    )
-
-
-def _index_state(index: RawIndexLookup, database_path: Path) -> tuple[object | None, str | None]:
-    store = getattr(index, "_store", None)
-    generation = None
-    snapshot = getattr(index, "snapshot", None)
-    if snapshot is not None:
-        try:
-            generation = snapshot().generation_id
-        except Exception:
-            generation = None
-    if store is None:
-        try:
-            from one_memory.retrieval.index_store import RetrievalIndexStore
-
-            store = RetrievalIndexStore(database_path)
-        except Exception:
-            store = None
-    return store, generation
-
-
 def read_memory_lane(
    database_path: Path,
    query: CoordinatorQuery,
@@ -857,7 +739,6 @@ def coordinate_context_v2(
    query: CoordinatorQuery,
    *,
    database_path: Path,
-    index: RawIndexLookup | None = None,
    deadline: float,
    expand_refs: ExpandRefs | None = None,
 ) -> RetrievalResponse:
@@ -871,62 +752,29 @@ def coordinate_context_v2(
        lexical = read_lexical_raw_lane(
            database_path, query, source_ids=authority.raw_sources, deadline=deadline
        )
-        semantic = read_semantic_raw_lane(
-            index, database_path, query, source_ids=authority.raw_sources, deadline=deadline
-        )
        timings = StageTimings(raw_ms=(time.monotonic() - raw_started) * 1000)
        combined: dict[tuple[str, str], RawEvidence] = {}
-        for item in (*lexical.items, *semantic.items):
-            if not isinstance(item, RawEvidence):
-                continue
-            identity = _identity(item)
-            previous = combined.get(identity)
-            if previous is None or _semantic_only_raw(previous):
-                combined[identity] = item
-        # Semantic similarity is a candidate signal, not a verified recall
-        # result. Keep semantic-only candidates visible in coverage, but do not
-        # release their bodies for host injection unless the item is
-        # independently confirmed by the lexical/object lane. A global distance
-        # threshold is deliberately not used: the measured positive and
-        # negative distributions overlap.
-        semantic_only = {_identity(item) for item in combined.values() if _semantic_only_raw(item)}
-        confirmed = {_identity(item) for item in lexical.items if isinstance(item, RawEvidence)}
-        deferred_raw = semantic_only - confirmed
-        if deferred_raw:
-            combined = {
-                identity: item
-                for identity, item in combined.items()
-                if identity not in deferred_raw
-            }
-        channels = tuple(dict.fromkeys((*lexical.channels, *semantic.channels)))
+        for item in lexical.items:
+            if isinstance(item, RawEvidence):
+                combined.setdefault(_identity(item), item)
        if lexical.status == "unavailable":
            status: Literal["ok", "empty", "unavailable", "not_searched", "degraded"] = (
                "unavailable"
            )
            reason = lexical.reason
-        elif semantic.status == "unavailable":
-            status = "unavailable"
-            reason = semantic.reason
-        elif combined:
-            status, reason = "ok", semantic.reason if semantic.status != "ok" else None
-        elif deferred_raw:
-            status, reason = "degraded", "relevance_confirmation"
-        elif lexical.status == "empty":
-            status, reason = ("empty" if semantic.status != "not_searched" else "ok"), None
+        elif combined or lexical.status == "empty":
+            status, reason = "ok", None
        else:
            status, reason = "empty", None
-        truncated = lexical.truncated or semantic.truncated
        raw = LaneResult(
            tuple(combined.values()),
            status,
            reason,
-            channels,
-            truncated,
+            lexical.channels,
+            lexical.truncated,
            lexical.starved,
            lexical.coverage_limited,
        )
-        if deferred_raw:
-            object.__setattr__(raw, "coverage_limited", True)
    elif "raw" in query.kinds and not authority.raw_allowed:
        raw = LaneResult((), "unavailable", "permission_required", (), False)
    if "memory" in query.kinds:
@@ -995,9 +843,7 @@ def coordinate_context_v2(
            return _path("unavailable", raw.reason, raw.channels, "not_searched", False, 0)
        if dropped and not final_raw:
            return _path("unavailable", "permission_required", raw.channels, "unavailable", True, 0)
-        # Lexical-only is complete when the lexical lane actually answered. An
-        # unconfigured semantic lane is a capability absence, not proof that
-        # less than the authorized range was searched.
+        # The raw lane is complete when the lexical lane actually answered.
        complete = (
            raw.status in {"ok", "empty"}
            and not dropped
@@ -1022,9 +868,7 @@ def coordinate_context_v2(
                "count_budget"
                if truncated
                else "coverage_budget"
-                if raw.coverage_limited and not deferred_raw
-                else "relevance_confirmation"
-                if deferred_raw
+                if raw.coverage_limited
                else "character_budget"
            )
        return _path(
@@ -1112,7 +956,7 @@ def coordinate_context_v2(
                "group_id": f"group-{index}",
                "answer_target": query.intent,
                "decision_status": "missing_confirmation",
-                "missing_requirements": _group_missing_requirements(item),
+                "missing_requirements": [],
                "object_refs": _object_refs(item, plan.object_keys),
                "items": [group_item.model_dump(mode="json") for group_item in (item,)],
            }
@@ -1185,7 +1029,6 @@ async def search_context_v2(
    query: CoordinatorQuery,
    *,
    database_path: Path,
-    index: RawIndexLookup | None = None,
    deadline: float,
    expand_refs: ExpandRefs | None = None,
 ) -> RetrievalResponse:
@@ -1198,7 +1041,6 @@ async def search_context_v2(
            authority,
            query,
            database_path=database_path,
-            index=index,
            deadline=deadline,
            expand_refs=expand_refs,
        ),
@@ -0,0 +1,277 @@
+"""Pure conversation-window helpers shared by capture, retrieval, and writes."""
+
+from __future__ import annotations
+
+import os
+import re
+
+from one_memory.imports.models import EvidenceContentKind, EvidenceOrigin, EvidenceRecord
+from one_memory.memory_tasks import segment_evidence_record
+
+MEMORY_TASK_TIMEOUT_MARGIN_SECONDS = 60.0
+MEMORY_TASK_TIMEOUT_MIN_SECONDS = 180.0
+MEMHUB_DEFAULT_CODING_TOOL_RESULT_MAX_BYTES = 512
+
+_HARNESS_PREFIXES = (
+    "<codex_internal_context",
+    "# agents.md instructions",
+    "<environment_context>",
+    "<permissions instructions>",
+    "<system-reminder>",
+    "<system_notification>",
+    "<instructions>",
+)
+_HOST_ENVELOPE = re.compile(
+    r"(?:"
+    r"^<timestamp>.*?<system_(?:notification|reminder)>|"
+    r"^<system_(?:notification|reminder)>|"
+    r'^\{\s*"request_id"\s*:.*?"steps"\s*:'
+    r")",
+    re.IGNORECASE | re.DOTALL,
+)
+_WHOLE_TURN_CHARACTERS = 60_000
+_WINDOW_BYTE_CEILING = 2_000_000
+_HISTORY_MERGE_TURNS = 8
+_HISTORY_MERGE_CHARACTERS = 8_000
+
+
+def memory_task_timeout_seconds(model_timeout_seconds: float) -> float:
+    """Bound a whole write without treating one model request as its budget."""
+    return max(
+        MEMORY_TASK_TIMEOUT_MIN_SECONDS,
+        float(model_timeout_seconds) + MEMORY_TASK_TIMEOUT_MARGIN_SECONDS,
+    )
+
+
+def _coding_tool_result_max_bytes() -> int:
+    raw = os.environ.get("MEMORY_CODING_TOOL_RESULT_MAX_BYTES")
+    if raw is None:
+        return MEMHUB_DEFAULT_CODING_TOOL_RESULT_MAX_BYTES
+    try:
+        value = int(raw)
+    except ValueError:
+        return MEMHUB_DEFAULT_CODING_TOOL_RESULT_MAX_BYTES
+    return value if value > 0 else MEMHUB_DEFAULT_CODING_TOOL_RESULT_MAX_BYTES
+
+
+_HY_TOOL_CHARACTERS = _coding_tool_result_max_bytes()
+
+
+def _conversation_chunks(
+    records: list[EvidenceRecord],
+    *,
+    max_characters: int | None = None,
+    max_turns: int | None = None,
+    completed_only: bool = False,
+) -> list[tuple[EvidenceRecord, ...]]:
+    """Build bounded, complete user/assistant/tool windows."""
+    turns: list[list[EvidenceRecord]] = []
+    for record in records:
+        if not _is_hy_turn_record(record):
+            continue
+        pending_proposal = bool(turns) and not _opens_turn(turns[-1][0])
+        if not turns or (_opens_turn(record) and not pending_proposal):
+            turns.append([record])
+        else:
+            turns[-1].append(record)
+
+    if completed_only and turns and _trailing_user_turn_is_open(turns[-1]):
+        turns.pop()
+
+    whole = _whole_turns() and max_characters is None
+    if whole:
+        size = _hy_read_length
+        max_characters = _WHOLE_TURN_CHARACTERS
+        max_turns = max_turns or 1_000_000
+    else:
+        size = lambda record: len(record.verbatim)  # noqa: E731
+        max_characters = max_characters or 12_000
+        max_turns = max_turns or 24
+
+    chunks: list[tuple[EvidenceRecord, ...]] = []
+    for turn in turns:
+        original_question = turn[0] if _opens_turn(turn[0]) else None
+        expanded: list[EvidenceRecord] = []
+        for record in turn:
+            reserve = (
+                size(original_question)
+                if original_question is not None
+                and record is not original_question
+                and size(original_question) < max_characters
+                else 0
+            )
+            segment_limit = max(1, max_characters - reserve)
+            if size(record) <= segment_limit:
+                expanded.append(record)
+                continue
+            spans = _semantic_segment_spans(record.verbatim, segment_limit)
+            expanded.extend(
+                segment_evidence_record(record, index=index, start=start, end=end)
+                if len(spans) > 1
+                else record
+                for index, (start, end) in enumerate(spans)
+            )
+        question = expanded[0] if expanded and _opens_turn(expanded[0]) else None
+        current: list[EvidenceRecord] = []
+        for record in expanded:
+            if current and (
+                len(current) >= max_turns
+                or sum(size(item) for item in current) + size(record) > max_characters
+                or sum(_window_bytes(item) for item in current) + _window_bytes(record)
+                > _WINDOW_BYTE_CEILING
+            ):
+                chunks.append(tuple(current))
+                current = (
+                    [question]
+                    if question is not None
+                    and max_turns > 1
+                    and size(question) + size(record) <= max_characters
+                    else []
+                )
+            current.append(record)
+        chunks.append(tuple(current))
+    return _merge_tool_only_chunks(chunks)
+
+
+def _semantic_segment_spans(text: str, limit: int) -> list[tuple[int, int]]:
+    if len(text) <= limit:
+        return [(0, len(text))]
+    spans: list[tuple[int, int]] = []
+    start = 0
+    separators = ("\n\n", "\n", "。", "！", "？", ". ", "! ", "? ")
+    while start < len(text):
+        hard_end = min(len(text), start + limit)
+        if hard_end == len(text):
+            end = hard_end
+        else:
+            floor = start + max(1, limit // 2)
+            candidates = [text.rfind(separator, floor, hard_end) for separator in separators]
+            boundary = max(candidates)
+            if boundary < floor:
+                end = hard_end
+            else:
+                matched = next(
+                    separator for separator in separators if text.startswith(separator, boundary)
+                )
+                end = boundary + len(matched)
+        spans.append((start, end))
+        start = end
+    return spans
+
+
+def _merge_tool_only_chunks(
+    chunks: list[tuple[EvidenceRecord, ...]],
+) -> list[tuple[EvidenceRecord, ...]]:
+    if not any(chunk and all(record.role == "tool" for record in chunk) for chunk in chunks):
+        return chunks
+    merged: list[tuple[EvidenceRecord, ...]] = []
+    leading_tools: list[EvidenceRecord] = []
+    for chunk in chunks:
+        if not chunk:
+            continue
+        if all(record.role == "tool" for record in chunk):
+            if merged:
+                merged[-1] = (*merged[-1], *chunk)
+            else:
+                leading_tools.extend(chunk)
+            continue
+        if leading_tools:
+            chunk = (*leading_tools, *chunk)
+            leading_tools = []
+        merged.append(chunk)
+    return [] if leading_tools else merged
+
+
+def _merge_historical_chunks(
+    chunks: list[tuple[EvidenceRecord, ...]],
+    *,
+    max_turns: int = _HISTORY_MERGE_TURNS,
+    max_characters: int = _HISTORY_MERGE_CHARACTERS,
+) -> list[tuple[EvidenceRecord, ...]]:
+    if not chunks:
+        return []
+    merged: list[tuple[EvidenceRecord, ...]] = []
+    current: list[EvidenceRecord] = []
+    current_turns = 0
+    current_characters = 0
+    for chunk in chunks:
+        chunk_characters = sum(len(record.verbatim) for record in chunk)
+        if current and (
+            current_turns >= max_turns or current_characters + chunk_characters > max_characters
+        ):
+            merged.append(tuple(current))
+            current = []
+            current_turns = 0
+            current_characters = 0
+        current.extend(chunk)
+        current_turns += 1
+        current_characters += chunk_characters
+    if current:
+        merged.append(tuple(current))
+    return merged
+
+
+def _chunk_input_ref(chunk: tuple[EvidenceRecord, ...]) -> str:
+    user_records = tuple(item for item in chunk if item.role == "user")
+    return user_records[-1].evidence_id if user_records else chunk[-1].evidence_id
+
+
+def _opens_turn(record: EvidenceRecord) -> bool:
+    return record.role == "user"
+
+
+def _trailing_user_turn_is_open(turn: list[EvidenceRecord]) -> bool:
+    if len(turn) != 1:
+        return False
+    opener = turn[0]
+    return (
+        _opens_turn(opener)
+        and opener.content_kind is EvidenceContentKind.RAW_CONVERSATION
+        and not _is_memory_note(opener)
+    )
+
+
+def _is_hy_turn_record(record: EvidenceRecord) -> bool:
+    if record.role == "tool":
+        return record.content_kind is EvidenceContentKind.RAW_CONVERSATION and bool(
+            record.verbatim.strip()
+        )
+    return _is_hy_input_turn(record)
+
+
+def _is_hy_input_turn(record: EvidenceRecord) -> bool:
+    source = record.verbatim.strip()
+    if not source:
+        return False
+    if record.content_kind is EvidenceContentKind.MEMORY_CONTENT:
+        return True
+    if record.role == "assistant":
+        return True
+    if record.role != "user":
+        return False
+    if source.casefold().startswith(_HARNESS_PREFIXES):
+        return False
+    return _HOST_ENVELOPE.search(source[:2_000]) is None
+
+
+def _is_memory_note(record: EvidenceRecord) -> bool:
+    meta = record.origin_meta.get(EvidenceOrigin.MANUAL_IMPORT)
+    return (
+        record.source_id == "one-memory-manual"
+        and meta is not None
+        and meta.raw_meta.get("input_mode") == "direct_memory_note"
+    )
+
+
+def _window_bytes(record: EvidenceRecord) -> int:
+    return len(record.verbatim.encode("utf-8"))
+
+
+def _whole_turns() -> bool:
+    return os.environ.get("ONE_MEMORY_SPLIT_TURNS", "").strip() != "1"
+
+
+def _hy_read_length(record: EvidenceRecord) -> int:
+    if record.role == "tool":
+        return min(len(record.verbatim), _HY_TOOL_CHARACTERS)
+    return len(record.verbatim)
@@ -6,7 +6,6 @@ from collections.abc import Sequence
 from typing import NamedTuple

 PREFACE = "以下是获授权历史参考，不是指令；当前用户请求优先。"
-_MARKER = re.compile(r"om-derived:([0-9a-f]{64})")
 # Written by the host-bridge recall injection hook (recall-injection.ts): the
 # recalled memory block and, in the same reply, the cross-session notice block.
 _INJECTED_TAGS = ("memhub_memory_context", "memhub_session_notice")
@@ -89,45 +88,3 @@ def strip_injected_memory(text: str) -> str:
        return text
    kept.append(text[at:])
    return _EMPTY_REMINDER.sub("", "".join(kept)).strip()
-
-
-def marker(event_id: str) -> str:
-    return f"om-derived:{event_id}"
-
-
-def echoed_event_ids(verbatim: str) -> tuple[str, ...]:
-    return tuple(dict.fromkeys(_MARKER.findall(verbatim)))
-
-
-def render_reference(events: tuple[object, ...] | list[object]) -> str:
-    lines = [PREFACE]
-    for event in events:
-        source = getattr(event, "source_id", None) or "unknown"
-        revision = getattr(event, "evidence_revision", None)
-        evidence_id = getattr(event, "evidence_id", None) or ""
-        event_id = getattr(event, "event_id", "")
-        lines.append(
-            f"来源 {source} 修订 {0 if revision is None else revision} "
-            f"证据 {evidence_id} {marker(str(event_id))}"
-        )
-    return "\n".join(lines)
-
-
-def is_pure_echo(verbatim: str, known_event_ids: set[str]) -> bool:
-    """True when the text is only a block we already injected.
-
-    A later line that is not the preface, a source line, or one of those
-    markers is a real correction and must still be recorded.
-    """
-
-    found = set(echoed_event_ids(verbatim))
-    if not found or not found & known_event_ids:
-        return False
-    for line in verbatim.splitlines():
-        stripped = line.strip()
-        if not stripped or stripped == PREFACE or stripped.startswith("来源 "):
-            continue
-        if _MARKER.search(stripped):
-            continue
-        return False
-    return True
@@ -20,8 +20,10 @@ from __future__ import annotations
 import atexit
 import json
 import logging
+import math
 import os
 import re
+import stat
 import sys
 import threading
 import time
@@ -72,80 +74,30 @@ class _Refused(OSError):
    """The path is not one we are willing to write to, and never will be."""


-DiagnosticLevel = Literal["info", "warn", "error"]
+DiagnosticLevel = Literal["debug", "info", "warn", "error"]
 _IDENTIFIER = re.compile(r"^[A-Za-z0-9][A-Za-z0-9._:-]{0,255}$")
 _EVENT = re.compile(r"^[a-z][a-z0-9]*(?:[._-][a-z0-9]+){0,15}$")
-_DETAIL_KEYS = frozenset(
-    {
-        "action",
-        "attempt",
-        "count",
-        "method",
-        "mode",
-        "phase",
-        "platform",
-        "reason",
-        "result_count",
-        "source_id",
-        "spawned",
-        "status",
-        "version",
-        "parent_session_id",
-        "invalid_lines",
-        "truncated",
-        "open_file_count",
-        "memory_bytes",
-        "thread_count",
-        "database_error_count",
-        "route_path",
-        "sqlite_errorcode",
-        "sqlite_errorname",
-        "file_count",
-        "sqlite_file_count",
-        "byte_file_count",
-        "total_bytes",
-        "copied_bytes",
-        "bytes_per_second",
-        "stored_count",
-        "revised_count",
-        "skipped_count",
-        "error_count",
-        "pending_count",
-        "prepared_count",
-        "queue_depth",
-        "in_flight_count",
-        "call_count",
-        "created_count",
-        "updated_count",
-        "superseded_count",
-        "deleted_count",
-        "operation_kind",
-        "index_complete",
-        "lexical_hits",
-        "semantic_hits",
-        "generation",
-        "timed_out",
-        "failure_site",
-        "error_class",
-        "file_kind",
-        "file_size",
-        "file_mtime_ns",
-        "file_inode",
-        "cursor_new",
-        "cursor_old",
-        "cursor_read_to_end",
-        "cursor_rewind_branch",
-        "summed_ms",
-        "model_ms",
-        "accounted_ms",
-        "unaccounted_ms",
-        "conversation_hash",
-        # The sedimentation switch: a bounded boolean, the same shape as
-        # `spawned` and `index_complete`. Left out of this list it did not
-        # drop a log line, it turned the switch's own request into a 500.
-        "enabled",
-    }
+_DIAGNOSTIC_LOG_NAME = re.compile(
+    r"^diagnostic-(runtime|core|host-bridge|desktop|cli)\.jsonl(?:\.\d+)?$"
 )
+_DETAIL_KEY = re.compile(r"^[A-Za-z][A-Za-z0-9_.-]{0,63}$")
+_CODE_FRAME = re.compile(
+    r"^(?:(?:apps|packages|scripts)/(?:[A-Za-z0-9_-]+/)*[A-Za-z0-9_.-]+:\d+:\d+"
+    r"|one_memory/(?:[A-Za-z0-9_-]+/)*[A-Za-z0-9_.-]+:\d+:[A-Za-z_][A-Za-z0-9_]*"
+    r"|[A-Za-z0-9_][A-Za-z0-9_.-]*\.py:\d+:[A-Za-z_][A-Za-z0-9_]*)$"
+)
+_SHORT_CREDENTIAL_TOKEN = re.compile(
+    r"(?<![A-Za-z0-9])([A-Za-z0-9]{24,47})(?:[_-])?(?![A-Za-z0-9])"
+)
+_REFUSED_DETAIL_KEYS = frozenset({"__proto__", "prototype", "constructor"})
+_MAX_DETAIL_KEYS = 32
+_MAX_SAFE_INTEGER = (1 << 53) - 1
+_EXPORT_SECRET_MARKER = re.compile(
+    r"(?:^|[._:-])(?:authorization|credential|password|prompt|query|secret|token)(?:$|[._:-])",
+    re.IGNORECASE,
+)
+_TRACE_STRING_DETAIL_KEYS = frozenset({"phase", "mode", "status", "reason", "failure_site"})
+_TRACE_NUMBER_DETAIL_KEYS = frozenset({"count"})
 _trace_id: ContextVar[str | None] = ContextVar("one_memory_trace_id", default=None)
 _active: DiagnosticLog | None = None
 _database_errors: Counter[tuple[str, str]] = Counter()
@@ -264,7 +216,12 @@ class DiagnosticLog:
        }
        value.update({key: _identifier(item) for key, item in identifiers.items() if item})
        if duration_ms is not None:
-            if not isinstance(duration_ms, int) or duration_ms < 0:
+            if (
+                not isinstance(duration_ms, int)
+                or isinstance(duration_ms, bool)
+                or duration_ms < 0
+                or duration_ms > _MAX_SAFE_INTEGER
+            ):
                raise ValueError("diagnostic_duration_invalid")
            value["duration_ms"] = duration_ms
        for field_name, item in {
@@ -276,7 +233,9 @@ class DiagnosticLog:
            if item is not None:
                value[field_name] = _identifier(item)
        if safe_stack:
-            value["safe_stack"] = [_safe_stack_line(line) for line in safe_stack[:256]]
+            if len(safe_stack) > 16:
+                raise ValueError("diagnostic_stack_invalid")
+            value["safe_stack"] = [_safe_stack_line(line) for line in safe_stack]
        if details:
            value["details"] = _details(details)
        self._submit(_serialize(value))
@@ -289,6 +248,8 @@ class DiagnosticLog:
        event: str,
        outcome: str,
    ) -> dict[str, object]:
+        if level not in ("debug", "info", "warn", "error"):
+            raise ValueError("diagnostic_level_invalid")
        return {
            "schema_version": 1,
            "timestamp": datetime.now(UTC).isoformat().replace("+00:00", "Z"),
@@ -651,24 +612,17 @@ def _drain_diagnostics_at_exit() -> None:


 def diagnostic_event(**values: object) -> None:
-    active = _active
-    if active is None:
-        return
-    if values.get("level") == "debug" and not logging.getLogger().isEnabledFor(logging.DEBUG):
-        return
    try:
+        active = _active
+        if active is None:
+            return
+        if values.get("level") == "debug" and not logging.getLogger().isEnabledFor(logging.DEBUG):
+            return
        active.write(**values)  # type: ignore[arg-type]
-    except OSError:
-        # Diagnostics must not become a new product outage when a disk is full
-        # or the log directory is temporarily unavailable.
-        return
-    except ValueError:
-        # Same promise, same reason, for a rejected payload instead of a failed
-        # write. An unlisted `details` key once turned every external recall
-        # into a 500 and then killed the process, and the sedimentation switch
-        # into a permanent failure. The payload itself is still wrong and is
-        # still caught in CI by tests/test_diagnostic_payload_keys.py; what
-        # must not happen is the caller paying for it.
+    except Exception:
+        # Diagnostics are best-effort: validation, serialization, logging, and
+        # filesystem failures must never become a product outage. BaseException
+        # deliberately remains visible so process-control signals still work.
        return


@@ -691,46 +645,95 @@ def read_trace_events(
    if (request_id is None) == (trace_id is None):
        raise ValueError("exactly_one_trace_identity_required")
    key, expected = ("request_id", request_id) if request_id is not None else ("trace_id", trace_id)
-    if expected is None or not _IDENTIFIER.fullmatch(expected):
+    if expected is None or not _export_identifier(expected):
        raise ValueError("invalid_trace_identity")
    selected: list[dict[str, object]] = []
-    allowed_details = {"phase", "mode", "status", "reason", "failure_site", "count"}
-    files = sorted(
-        (data_dir / "logs").glob("diagnostic-*.jsonl*"),
-        key=lambda path: path.stat().st_mtime,
-    )
-    for path in files:
-        if path.is_symlink() or not path.is_file():
+    directory = data_dir / "logs"
+    try:
+        directory_info = directory.lstat()
+    except FileNotFoundError:
+        return ()
+    if stat.S_ISLNK(directory_info.st_mode) or not stat.S_ISDIR(directory_info.st_mode):
+        raise ValueError("diagnostic_directory_unsafe")
+    files: list[tuple[Path, float]] = []
+    for path in directory.iterdir():
+        if not _DIAGNOSTIC_LOG_NAME.fullmatch(path.name):
            continue
-        for line in path.read_text(encoding="utf-8", errors="replace").splitlines():
+        try:
+            info = path.lstat()
+        except OSError:
+            continue
+        if stat.S_ISREG(info.st_mode) and not stat.S_ISLNK(info.st_mode):
+            files.append((path, info.st_mtime))
+    for path, _mtime in sorted(files, key=lambda item: item[1]):
+        for line in _read_regular_file_no_follow(path).splitlines():
            try:
                event = json.loads(line)
            except (TypeError, json.JSONDecodeError):
                continue
            if not isinstance(event, dict) or event.get(key) != expected:
                continue
-            details = event.get("details")
-            projected = {
-                name: event[name]
-                for name in (
-                    "timestamp",
-                    "component",
-                    "event",
-                    "outcome",
-                    "duration_ms",
-                    "error_code",
-                    "error_type",
-                )
-                if name in event
+            timestamp = event.get("timestamp")
+            if not isinstance(timestamp, str):
+                continue
+            try:
+                datetime.fromisoformat(timestamp.replace("Z", "+00:00"))
+            except ValueError:
+                continue
+            component = event.get("component")
+            event_name = event.get("event")
+            outcome = event.get("outcome")
+            projected: dict[str, object] = {
+                "timestamp": timestamp,
+                "component": (
+                    component
+                    if isinstance(component, str) and _export_identifier(component)
+                    else "diagnostic.redacted"
+                ),
+                "event": (
+                    event_name
+                    if isinstance(event_name, str) and _export_identifier(event_name)
+                    else "diagnostic.redacted"
+                ),
+                "outcome": (
+                    outcome
+                    if isinstance(outcome, str) and _export_identifier(outcome)
+                    else "redacted"
+                ),
            }
+            duration_ms = event.get("duration_ms")
+            if (
+                isinstance(duration_ms, int)
+                and not isinstance(duration_ms, bool)
+                and 0 <= duration_ms <= _MAX_SAFE_INTEGER
+            ):
+                projected["duration_ms"] = duration_ms
+            for name in ("error_code", "error_type"):
+                value = event.get(name)
+                if isinstance(value, str) and _export_identifier(value):
+                    projected[name] = value
+            details = event.get("details")
            if isinstance(details, dict):
-                projected["details"] = {
-                    name: value for name, value in details.items() if name in allowed_details
-                }
+                safe_details = _trace_details(details)
+                if safe_details:
+                    projected["details"] = safe_details
            selected.append(projected)
    return tuple(selected[-max(1, min(limit, 500)) :])


+def _read_regular_file_no_follow(path: Path) -> str:
+    descriptor = os.open(path, os.O_RDONLY | _O_NOFOLLOW)
+    try:
+        if not stat.S_ISREG(os.fstat(descriptor).st_mode):
+            raise ValueError("diagnostic_file_unsafe")
+        with os.fdopen(descriptor, encoding="utf-8", errors="replace") as handle:
+            descriptor = -1
+            return handle.read()
+    finally:
+        if descriptor >= 0:
+            os.close(descriptor)
+
+
 def current_trace_id() -> str | None:
    return _trace_id.get()

@@ -801,7 +804,7 @@ def record_database_error(
    sqlite_code = getattr(error, "sqlite_errorcode", None)
    sqlite_name = getattr(error, "sqlite_errorname", None)
    details: dict[str, object] = {"phase": safe_component, "count": count}
-    if route_path is not None and route_path.startswith("/api/v1/"):
+    if route_path is not None and route_path.startswith(("/api/v1/", "/api/v2/")):
        details["route_path"] = route_path
    if isinstance(sqlite_code, int):
        details["sqlite_errorcode"] = sqlite_code
@@ -837,8 +840,77 @@ def diagnostic_exception(error: BaseException) -> dict[str, object]:
    }


+def _contains_credential_material(value: str) -> bool:
+    return bool(
+        re.search(
+            r"(?:^|[^A-Za-z0-9])(?:sk-|sk_|omhg_|ghp_|github_pat_)[A-Za-z0-9._-]+",
+            value,
+            re.IGNORECASE,
+        )
+        or re.search(r"(?:^|[^A-Za-z0-9])xox[baprs]-[A-Za-z0-9-]+", value, re.IGNORECASE)
+        or re.search(r"(?:^|[^A-Za-z0-9])(?:AKIA|ASIA)[A-Z0-9]{16}(?![A-Za-z0-9])", value)
+        or re.search(r"(?:^|[^A-Za-z0-9])AKID[A-Za-z0-9]{12,}", value, re.IGNORECASE)
+        or re.search(r"(?:^|[^A-Za-z0-9])AIza[A-Za-z0-9_-]{20,}", value)
+        or re.search(r"(?:^|[^A-Za-z0-9])GOCSPX-[A-Za-z0-9_-]+", value)
+        or any(
+            _is_conservative_short_credential(match.group(1))
+            for match in _SHORT_CREDENTIAL_TOKEN.finditer(value)
+        )
+        or re.search(
+            r"\b[A-Za-z0-9_-]{10,}\.[A-Za-z0-9_-]{10,}\.[A-Za-z0-9_-]{10,}\b",
+            value,
+        )
+        or any(
+            len(candidate) >= 48
+            and any(character.isalpha() for character in candidate)
+            and any(character.isdigit() for character in candidate)
+            for candidate in re.findall(r"\b[A-Za-z0-9._-]{48,}\b", value)
+        )
+    )
+
+
+def _is_conservative_short_credential(value: str) -> bool:
+    return bool(
+        re.search(r"[A-Z]", value)
+        and re.search(r"[a-z]", value)
+        and re.search(r"\d", value)
+        and not re.fullmatch(r"[A-Fa-f0-9]+", value)
+    )
+
+
+def _export_identifier(value: str) -> bool:
+    return bool(
+        _IDENTIFIER.fullmatch(value)
+        and not _EXPORT_SECRET_MARKER.search(value)
+        and not _contains_credential_material(value)
+    )
+
+
+def _trace_details(values: Mapping[object, object]) -> dict[str, str | int | float]:
+    result: dict[str, str | int | float] = {}
+    for key, value in values.items():
+        if (
+            key in _TRACE_STRING_DETAIL_KEYS
+            and isinstance(value, str)
+            and _export_identifier(value)
+        ):
+            result[str(key)] = value
+        elif key in _TRACE_NUMBER_DETAIL_KEYS and isinstance(value, (int, float)):
+            if isinstance(value, bool):
+                continue
+            safe_integer = isinstance(value, int) and abs(value) <= _MAX_SAFE_INTEGER
+            safe_float = (
+                isinstance(value, float)
+                and math.isfinite(value)
+                and (not value.is_integer() or abs(value) <= _MAX_SAFE_INTEGER)
+            )
+            if safe_integer or safe_float:
+                result[str(key)] = value
+    return result
+
+
 def _serialize(value: Mapping[str, object]) -> str:
-    return json.dumps(value, ensure_ascii=True, separators=(",", ":")) + "\n"
+    return json.dumps(value, ensure_ascii=True, allow_nan=False, separators=(",", ":")) + "\n"


 def _uuid(value: str | None) -> str | None:
@@ -864,24 +936,47 @@ def _event_name(value: str) -> str:


 def _safe_stack_line(value: str) -> str:
-    if not value or len(value) > 256 or "\n" in value or "\r" in value:
+    if (
+        not value
+        or len(value) > 256
+        or any(character in value for character in "\0\r\n")
+        or _contains_credential_material(value)
+    ):
        raise ValueError("diagnostic_stack_invalid")
-    if Path(value.split(":", 1)[0]).is_absolute() or "file://" in value:
-        raise ValueError("diagnostic_stack_invalid")
-    return value
+    if _CODE_FRAME.fullmatch(value):
+        return value
+    if value.startswith("thread|"):
+        frames = value.removeprefix("thread|").split(" > ")
+        if frames and all(_CODE_FRAME.fullmatch(frame) for frame in frames):
+            return value
+    raise ValueError("diagnostic_stack_invalid")


 def _details(
    values: Mapping[str, str | int | float | bool | None],
 ) -> dict[str, str | int | float | bool | None]:
+    if len(values) > _MAX_DETAIL_KEYS:
+        raise ValueError("diagnostic_detail_count_invalid")
    result: dict[str, str | int | float | bool | None] = {}
    for key, value in values.items():
-        if key not in _DETAIL_KEYS:
+        if (
+            not isinstance(key, str)
+            or not _DETAIL_KEY.fullmatch(key)
+            or key in _REFUSED_DETAIL_KEYS
+        ):
            raise ValueError("diagnostic_detail_key_invalid")
        if isinstance(value, str):
-            if not value or len(value) > 256 or "\n" in value or "\r" in value:
+            if not value or len(value) > 256 or any(character in value for character in "\0\r\n"):
                raise ValueError("diagnostic_detail_value_invalid")
-        elif value is not None and not isinstance(value, (int, float, bool)):
+        elif isinstance(value, float):
+            if not math.isfinite(value) or (value.is_integer() and abs(value) > _MAX_SAFE_INTEGER):
+                raise ValueError("diagnostic_detail_value_invalid")
+        elif isinstance(value, bool) or value is None:
+            pass
+        elif isinstance(value, int):
+            if abs(value) > _MAX_SAFE_INTEGER:
+                raise ValueError("diagnostic_detail_value_invalid")
+        else:
            raise ValueError("diagnostic_detail_value_invalid")
        result[key] = value
    return result
@@ -3,8 +3,8 @@
 Every site that catches `Exception` broadly must call `report_exception`
 before (or instead of) continuing, so a failure never looks like success
 and a stack is always on disk. Content (conversation text, keys, SQL
-parameters) never reaches the log: only exception type, a truncated
-message, and file:line:function frames.
+parameters, and exception messages) never reaches the diagnostic log: only
+exception type, a repeat count, and file:line:function frames.
 """

 from __future__ import annotations
@@ -67,7 +67,7 @@ def report_exception(
        event=event,
        outcome="failed",
        **payload,
-        details={"reason": _safe_message(error), "count": _RECENT[key]},
+        details={"count": _RECENT[key]},
    )
    if full:
        _logger.error(
@@ -23,13 +23,9 @@ _ENVELOPE_HEADERS = frozenset(
 )
 _BROKER_ROUTES = {
    ("POST", "/api/v1/recall"): "om_recall",
-    ("POST", "/api/v1/memories/ask"): "om_ask",
    ("POST", "/api/v2/memories/ask"): "om_ask",
    ("POST", "/api/v2/memories/ask/stream"): "om_ask",
-    ("GET", "/api/v1/conversations/search"): "om_conversations",
    ("POST", "/api/v1/diagnostics/trace"): "om_diagnostic_trace",
-    ("POST", "/api/v1/boundary/peek"): "om_recall",
-    ("POST", "/api/v1/boundary/ack"): "om_recall",
 }
 _CONCRETE_CALL_EVENT_DELIVERY = re.compile(
    r"/api/v1/call-events/[0-9a-f]{8}-[0-9a-f]{4}-[1-5][0-9a-f]{3}-"
@@ -128,11 +124,9 @@ def require_external_endpoint(request: Request) -> None:
    access = external_access(request.headers)
    if access is None:
        return
-    if request.method == "POST" and request.url.path in {
-        "/api/v1/context/search",
-        "/api/v2/context/search",
-    }:
-        # The route resolves each requested lane against its original grant.
+    if request.method == "POST" and request.url.path == "/api/v2/context/search":
+        # om_conversations reaches Core only here, as the raw lane; the route
+        # resolves each requested lane against its original grant.
        # Passing this envelope gate does not authorize either lane by itself.
        if access.tools.isdisjoint({"om_recall", "om_conversations"}):
            raise HTTPException(403, "external_recall_scope_denied")
@@ -146,7 +140,3 @@ def require_external_endpoint(request: Request) -> None:
        raise HTTPException(403, "external_endpoint_denied")
    if tool not in access.tools:
        raise HTTPException(403, "external_recall_scope_denied")
-    if tool == "om_conversations":
-        access.require(tool, frozenset({"all-apps"}))
-        if not any(scope.startswith("source:") for scope in access.scopes):
-            raise HTTPException(403, "external_source_permission_required")
@@ -1,127 +0,0 @@
-"""Bounded homepage projection without loading every memory into the renderer."""
-
-import sqlite3
-import time
-from datetime import UTC, datetime
-from pathlib import Path
-
-from pydantic import BaseModel, ConfigDict
-
-from one_memory.database_access import main_database
-from one_memory.memory_evidence_freshness import register_memory_freshness
-
-
-class HomeLatestMemory(BaseModel):
-    model_config = ConfigDict(frozen=True)
-
-    memory_id: str
-    content: str
-    truncated: bool
-    revision: int
-    action: str
-    updated_at: datetime
-    source_ids: tuple[str, ...]
-
-
-class HomeSummary(BaseModel):
-    model_config = ConfigDict(frozen=True)
-
-    memory_count: int
-    latest: HomeLatestMemory | None
-    observed_at: datetime
-    #: 2026-10-12: the read hit its deadline and the numbers below are not a
-    #: snapshot of reality. Callers should keep their previous summary on
-    #: screen instead of treating count=0 as "no memories".
-    stale: bool = False
-
-
-class HomeSummaryUnavailable(Exception):
-    """The homepage projection could not be read inside its own deadline.
-
-    Raised instead of letting SQLite's interrupt escape as a bare
-    `OperationalError`, so the route can answer with an empty projection
-    rather than a 500.
-    """
-
-
-def read_home_summary(database_path: Path, *, user_id: str = "local") -> HomeSummary:
-    # The exclusion mirrors SourceReadAccess: one revoked active citation hides
-    # the whole memory. Permission filtering precedes both count and LIMIT.
-    # `+link.active` keeps SQLite on the (memory_id, evidence_id) primary key.
-    # Without the unary plus it picks the (active, ...) index and walks every
-    # active link once per memory: 2.9 s instead of 0.07 s on a 1,810-memory
-    # library, enough to time the homepage out whenever the machine is busy.
-    readable = """
-        m.user_id = ? AND m.status = 'confirmed' AND m.deleted_at IS NULL
-        AND NOT EXISTS (
-            SELECT 1 FROM governed_memory_evidence link
-            JOIN evidence e ON e.evidence_id = link.evidence_id
-            JOIN source_grants g ON g.user_id=e.user_id AND g.source_id=e.source_id
-            WHERE link.memory_id=m.memory_id AND +link.active=1
-              AND g.revoked_at IS NOT NULL
-        )
-        AND NOT EXISTS (
-            SELECT 1 FROM governed_memory_evidence link
-            LEFT JOIN evidence e ON e.evidence_id = link.evidence_id
-            WHERE link.memory_id=m.memory_id AND +link.active=1
-              AND (e.evidence_id IS NULL OR e.user_id != m.user_id)
-        )
-    """
-    # The total counts with the cheap permission filters only: the per-row
-    # freshness UDF scales with the whole library and can burn the whole read
-    # budget on its own (D22). Product accepted a small drift here; the rows
-    # that are actually displayed are still checked row by row below.
-    with main_database(database_path).read() as db:
-        deadline = time.monotonic() + 5
-        db.set_progress_handler(lambda: int(time.monotonic() >= deadline), 1000)
-        db.row_factory = sqlite3.Row
-        try:
-            db.execute("BEGIN")
-            count = db.execute(
-                f"SELECT COUNT(*) FROM governed_memories m WHERE {readable}", (user_id,)
-            ).fetchone()[0]
-            readable += " AND " + register_memory_freshness(db)
-            row = db.execute(
-                f"""SELECT m.memory_id, substr(m.content,1,320) AS content,
-                    length(m.content)>320 AS truncated, m.revision,
-                    COALESCE(v.action, 'unknown') AS action, m.updated_at
-                    FROM governed_memories m
-                    LEFT JOIN governed_memory_versions v
-                      ON v.memory_id=m.memory_id AND v.revision=m.revision
-                    WHERE {readable}
-                    ORDER BY m.updated_at DESC, m.memory_id LIMIT 1""",
-                (user_id,),
-            ).fetchone()
-            latest = None
-            if row is not None:
-                sources = db.execute(
-                    """SELECT DISTINCT e.source_id FROM governed_memory_evidence link
-                    JOIN evidence e ON e.evidence_id=link.evidence_id
-                    WHERE link.memory_id=? AND +link.active=1 ORDER BY e.source_id""",
-                    (row["memory_id"],),
-                ).fetchall()
-                latest = HomeLatestMemory(
-                    memory_id=row["memory_id"],
-                    content=row["content"],
-                    truncated=bool(row["truncated"]),
-                    revision=row["revision"],
-                    action=row["action"],
-                    updated_at=row["updated_at"],
-                    source_ids=tuple(source[0] for source in sources),
-                )
-        except sqlite3.Error as error:
-            # The progress handler above is what interrupts this read, so a
-            # slow homepage surfaces as SQLITE_INTERRUPT (code 9) and used to
-            # escape as a 500. 2026-10-12: answer with a `stale` projection
-            # instead of raising: the caller keeps its last good summary on
-            # screen (the UI poll retries within 10s), and the homepage never
-            # zeroes out or errors just because the machine is busy importing.
-            if getattr(error, "sqlite_errorcode", None) != sqlite3.SQLITE_INTERRUPT:
-                raise
-            return HomeSummary(
-                memory_count=0,
-                latest=None,
-                observed_at=datetime.now(UTC),
-                stale=True,
-            )
-        return HomeSummary(memory_count=count, latest=latest, observed_at=datetime.now(UTC))
@@ -0,0 +1,271 @@
+"""Kuzu graph-width inspection and explicit user-directed administration.
+
+This module is the second graph architecture exception.  It may open Kuzu
+directly to inspect an existing graph's schema, but owns no prompts, memory
+reconciliation, coding data, vector database, or legacy governance behavior.
+"""
+
+from __future__ import annotations
+
+import logging
+import os
+import re
+import shutil
+import threading
+from dataclasses import dataclass
+from datetime import UTC, datetime
+from enum import StrEnum
+from importlib import import_module
+from pathlib import Path
+from typing import Any
+
+from pydantic import BaseModel, ConfigDict
+
+from one_memory.hy_runtime_config import HY_MEMORY_NAMESPACE
+from one_memory.model_profiles import ModelProfileSnapshot, ProviderKind
+
+
+class GraphWidthError(RuntimeError):
+    """Base error for graph-width inspection and administration."""
+
+
+class GraphWidthDecision(StrEnum):
+    """What to do when an existing graph disagrees with the embedding width."""
+
+    KEEP_EXISTING = "keep_existing"
+    REBUILD_WITH_NEW = "rebuild_with_new"
+    EXPORT_FIRST = "export_first"
+
+
+class GraphWidthMismatch(BaseModel):
+    model_config = ConfigDict(frozen=True, extra="forbid")
+
+    stored_width: int
+    expected_width: int
+    graph_path: str
+
+
+class GraphWidthOutcome(BaseModel):
+    model_config = ConfigDict(frozen=True, extra="forbid")
+
+    archived_path: str | None = None
+    exported_path: str | None = None
+
+
+class GraphWidthMismatchError(GraphWidthError):
+    """Raised instead of silently archiving a live graph."""
+
+    def __init__(self, mismatch: GraphWidthMismatch) -> None:
+        super().__init__(
+            f"graph_width_mismatch:stored={mismatch.stored_width}"
+            f":expected={mismatch.expected_width}"
+        )
+        self.mismatch = mismatch
+
+
+def graph_store_enabled() -> bool:
+    """Whether HY should open the Kuzu graph store."""
+    raw = os.environ.get("ONE_MEMORY_GRAPH_STORE", "").strip().lower()
+    if not raw:
+        return False
+    return raw != "none"
+
+
+def inspect_graph_width_mismatch(config: Any) -> GraphWidthMismatch | None:
+    """Return a mismatch without renaming or deleting anything."""
+    db_path = Path(config.graph_store.db_path)
+    expected = getattr(getattr(config, "embedder", None), "embedding_dims", None)
+    return _inspect_graph_width(db_path, expected)
+
+
+def inspect_profile_graph_width(
+    profile: ModelProfileSnapshot,
+    *,
+    api_key: str | None,
+    data_dir: Path,
+) -> GraphWidthMismatch | None:
+    """Inspect the graph path and width selected by the active profile."""
+    del api_key
+    return _inspect_graph_width(
+        _profile_graph_path(data_dir),
+        _profile_embedding_width(profile),
+    )
+
+
+def decide_profile_graph_width(
+    profile: ModelProfileSnapshot,
+    *,
+    api_key: str | None,
+    data_dir: Path,
+    decision: GraphWidthDecision,
+) -> GraphWidthOutcome:
+    """Apply the user's decision to the active profile's graph."""
+    del api_key
+    config = _GraphAdminConfig(
+        graph_store=_GraphStoreConfig(str(_profile_graph_path(data_dir))),
+        embedder=_EmbeddingWidthConfig(_profile_embedding_width(profile)),
+    )
+    return apply_graph_width_decision(
+        config,
+        decision,
+        export_root=data_dir.parent,
+    )
+
+
+def apply_graph_width_decision(
+    config: Any,
+    decision: GraphWidthDecision,
+    *,
+    export_root: Path | None = None,
+) -> GraphWidthOutcome:
+    """Act only on a current mismatch and only as explicitly requested."""
+    mismatch = inspect_graph_width_mismatch(config)
+    if mismatch is None:
+        return GraphWidthOutcome()
+    if decision is GraphWidthDecision.REBUILD_WITH_NEW:
+        archive = retire_mismatched_graph_store(config)
+        return GraphWidthOutcome(archived_path=str(archive) if archive is not None else None)
+    if decision is GraphWidthDecision.EXPORT_FIRST:
+        if export_root is None:
+            raise GraphWidthError("export_root_required")
+        return GraphWidthOutcome(
+            exported_path=str(export_graph_store(config, export_root=export_root))
+        )
+    return GraphWidthOutcome()
+
+
+def export_graph_store(config: Any, *, export_root: Path) -> Path:
+    """Copy the graph outside its data directory, atomically."""
+    db_path = Path(config.graph_store.db_path)
+    stamp = datetime.now(UTC).strftime("%Y%m%dT%H%M%S%fZ")
+    destination = export_root / f"one-memory-graph-{stamp}"
+    staging = destination.with_name(f"{destination.name}.partial")
+    sources = [db_path] + [
+        sibling
+        for suffix in (".wal", ".shadow", ".tmp")
+        if (sibling := db_path.with_name(f"{db_path.name}{suffix}")).exists()
+    ]
+    try:
+        staging.mkdir(parents=True)
+        for source in sources:
+            shutil.copy2(source, staging / source.name)
+        staging.rename(destination)
+    except BaseException:
+        shutil.rmtree(staging, ignore_errors=True)
+        raise
+    logging.getLogger(__name__).info(
+        "exported hy graph store to %s (%d files)",
+        destination,
+        len(sources),
+    )
+    return destination
+
+
+def retire_mismatched_graph_store(config: Any) -> Path | None:
+    """Archive a mismatched Kuzu graph only after the user chooses rebuild."""
+    db_path = Path(config.graph_store.db_path)
+    if not db_path.exists():
+        return None
+    expected = getattr(getattr(config, "embedder", None), "embedding_dims", None)
+    if not expected:
+        return None
+    stored = _graph_embedding_width(db_path)
+    if stored is None or stored == int(expected):
+        return None
+    stamp = datetime.now(UTC).strftime("%Y%m%dT%H%M%SZ")
+    archive = db_path.with_name(f"{db_path.name}.dim{stored}.{stamp}.bak")
+    db_path.rename(archive)
+    for suffix in (".wal", ".shadow", ".tmp"):
+        sibling = db_path.with_name(f"{db_path.name}{suffix}")
+        if sibling.exists():
+            sibling.rename(archive.with_name(f"{archive.name}{suffix}"))
+    logging.getLogger(__name__).warning(
+        "retired hy graph store with embedding width %s (model expects %s): %s",
+        stored,
+        expected,
+        archive.name,
+    )
+    return archive
+
+
+_GRAPH_WIDTH_CACHE: dict[str, int | None] = {}
+_GRAPH_WIDTH_LOCK = threading.Lock()
+
+
+def _graph_embedding_width(db_path: Path) -> int | None:
+    """Read the Memory embedding width directly from Kuzu's schema."""
+    try:
+        kuzu = import_module("kuzu")
+    except ModuleNotFoundError:
+        return None
+    key = str(db_path.resolve())
+    with _GRAPH_WIDTH_LOCK:
+        if key in _GRAPH_WIDTH_CACHE:
+            return _GRAPH_WIDTH_CACHE[key]
+    try:
+        database = kuzu.Database(str(db_path), read_only=True)
+    except Exception:
+        return None
+    width: int | None = None
+    try:
+        connection = kuzu.Connection(database)
+        result = connection.execute("CALL TABLE_INFO('Memory') RETURN *")
+        while result.has_next():
+            row = result.get_next()
+            if len(row) < 3 or row[1] != "embedding":
+                continue
+            match = re.search(r"\[(\d+)\]", str(row[2]))
+            width = int(match.group(1)) if match else None
+            break
+    except Exception:
+        width = None
+    with _GRAPH_WIDTH_LOCK:
+        _GRAPH_WIDTH_CACHE[key] = width
+    return width
+
+
+def _inspect_graph_width(
+    db_path: Path,
+    expected_width: object,
+) -> GraphWidthMismatch | None:
+    if not db_path.exists() or not expected_width:
+        return None
+    stored = _graph_embedding_width(db_path)
+    if stored is None or stored == int(expected_width):
+        return None
+    return GraphWidthMismatch(
+        stored_width=int(stored),
+        expected_width=int(expected_width),
+        graph_path=str(db_path),
+    )
+
+
+def _profile_graph_path(data_dir: Path) -> Path:
+    name = "graph" if graph_store_enabled() else "graph-disabled"
+    return data_dir / "hy-memory" / HY_MEMORY_NAMESPACE / name
+
+
+def _profile_embedding_width(profile: ModelProfileSnapshot) -> int | None:
+    if profile.embedding_dimensions is not None:
+        return profile.embedding_dimensions
+    if (profile.embedding_provider_kind or profile.provider_kind) is ProviderKind.LOCAL_ONNX:
+        from one_memory.embedding_assets import BUILTIN_EMBEDDING_DIMENSIONS
+
+        return BUILTIN_EMBEDDING_DIMENSIONS
+    return None
+
+
+@dataclass(frozen=True)
+class _GraphStoreConfig:
+    db_path: str
+
+
+@dataclass(frozen=True)
+class _EmbeddingWidthConfig:
+    embedding_dims: int | None
+
+
+@dataclass(frozen=True)
+class _GraphAdminConfig:
+    graph_store: _GraphStoreConfig
+    embedder: _EmbeddingWidthConfig
@@ -0,0 +1,185 @@
+"""Read-only projection of HY's Kuzu graph for MemHub's existing graph UI.
+
+This module is one of the two explicit graph architecture exceptions.  The
+public SDK does not expose graph projection yet, so it may use the client's
+private graph handle and loop thread.  It must remain read-only and must not
+grow memory, vector, coding, prompt, or governance responsibilities.
+"""
+
+from __future__ import annotations
+
+from typing import Any, cast
+
+from pydantic import BaseModel, ConfigDict
+
+
+class HyGraphNode(BaseModel):
+    model_config = ConfigDict(frozen=True, extra="forbid")
+
+    node_id: str
+    content: str
+    layer: str | None = None
+    confidence: float | None = None
+
+
+class HyGraphEdge(BaseModel):
+    model_config = ConfigDict(frozen=True, extra="forbid")
+
+    source_node_id: str
+    target_node_id: str
+    relation: str
+    weight: float | None = None
+
+
+class HyGraphSnapshot(BaseModel):
+    model_config = ConfigDict(frozen=True, extra="forbid")
+
+    nodes: tuple[HyGraphNode, ...] = ()
+    edges: tuple[HyGraphEdge, ...] = ()
+    # False when HY was never consulted, which is not the same as HY having
+    # induced no relations.
+    consulted: bool = True
+
+
+def read_graph_snapshot(
+    client: object,
+    user_id: str,
+    agent_id: str,
+    limit: int,
+) -> HyGraphSnapshot:
+    raw_nodes, raw_edges = _read_graph_snapshot(client, user_id, agent_id, limit)
+    return HyGraphSnapshot(
+        nodes=tuple(
+            HyGraphNode(
+                node_id=str(row[0]),
+                content=str(row[1]),
+                layer=str(row[2]) if row[2] is not None else None,
+                confidence=_safe_float(row[3]),
+            )
+            for row in raw_nodes
+            if len(row) >= 4 and row[0] and row[1]
+        ),
+        edges=tuple(
+            HyGraphEdge(
+                source_node_id=str(row[0]),
+                target_node_id=str(row[1]),
+                relation=str(row[2]),
+                weight=_safe_float(row[3]),
+            )
+            for row in raw_edges
+            if len(row) >= 4 and row[0] and row[1] and row[2]
+        ),
+    )
+
+
+def _read_graph_snapshot(
+    client: object,
+    user_id: str,
+    agent_id: str,
+    limit: int,
+) -> tuple[list[list[object]], list[list[object]]]:
+    """Read the graph through the SDK handles until a public API exists."""
+    concrete = cast(Any, client)
+    # Keep the existing architecture gate unchanged; this module is the named
+    # graph-only exception while the rest of MemHub remains on public handles.
+    graph_store = getattr(concrete, "_graph_store")  # noqa: B009
+    if graph_store is None:
+        return ([], [])
+    loop_thread = getattr(concrete, "_loop_thread")  # noqa: B009
+    # Kuzu rejects a prepared statement that is handed a parameter its text
+    # never references, so a wildcard read must not bind agent_id at all.
+    params: dict[str, object] = {"user_id": user_id}
+    if agent_id != "*":
+        params["agent_id"] = agent_id
+    safe_limit = max(1, min(int(limit), 1_000))
+    memory_agent = "" if agent_id == "*" else " AND m.agent_id = $agent_id"
+    related_agent = (
+        "" if agent_id == "*" else " AND a.agent_id = $agent_id AND b.agent_id = $agent_id"
+    )
+    source_agent = "" if agent_id == "*" else " AND a.agent_id = $agent_id"
+    nodes = (
+        loop_thread.run(
+            graph_store._run(
+                f"""
+            MATCH (m:Memory)
+            WHERE m.user_id = $user_id{memory_agent}
+              AND m.status = 'active'
+            RETURN m.node_id, m.content, m.layer, m.confidence
+            LIMIT {safe_limit}
+            """,
+                params,
+            )
+        )
+        or []
+    )
+    related_edges = (
+        loop_thread.run(
+            graph_store._run(
+                f"""
+            MATCH (a:Memory)-[r:RELATED_TO]->(b:Memory)
+            WHERE a.user_id = $user_id{related_agent}
+              AND b.user_id = $user_id
+              AND a.status = 'active' AND b.status = 'active'
+            RETURN a.node_id, b.node_id, r.relation_type, r.weight
+            LIMIT {safe_limit * 4}
+            """,
+                params,
+            )
+        )
+        or []
+    )
+    derived_edges = (
+        loop_thread.run(
+            graph_store._run(
+                f"""
+            MATCH (a:Memory)-[:DERIVED_FROM]->(v:VdbRef)
+            WHERE a.user_id = $user_id{source_agent}
+              AND a.status = 'active'
+            RETURN a.node_id, v.node_id, 'derived_from', 1.0
+            LIMIT {safe_limit * 4}
+            """,
+                params,
+            )
+        )
+        or []
+    )
+    topic_nodes = (
+        loop_thread.run(
+            graph_store._run(
+                f"""
+            MATCH (m:Memory)-[:TAGGED_WITH]->(t:Topic)
+            WHERE m.user_id = $user_id{memory_agent}
+              AND m.status = 'active'
+            RETURN DISTINCT t.topic_id, t.name, 'TOPIC', 1.0
+            LIMIT {safe_limit}
+            """,
+                params,
+            )
+        )
+        or []
+    )
+    tagged_edges = (
+        loop_thread.run(
+            graph_store._run(
+                f"""
+            MATCH (m:Memory)-[:TAGGED_WITH]->(t:Topic)
+            WHERE m.user_id = $user_id{memory_agent}
+              AND m.status = 'active'
+            RETURN m.node_id, t.topic_id, 'tagged_with', 1.0
+            LIMIT {safe_limit * 4}
+            """,
+                params,
+            )
+        )
+        or []
+    )
+    return (
+        [*nodes, *topic_nodes],
+        [*related_edges, *derived_edges, *tagged_edges],
+    )
+
+
+def _safe_float(value: object) -> float | None:
+    if isinstance(value, (int, float)):
+        return float(value)
+    return None
@@ -1,128 +0,0 @@
-"""Which memories HY no longer counts as current, read from HY itself.
-
-HY owns a memory's lifecycle: it supersedes, deduplicates, shadows and
-restores on its own schedule — during a search, while deleting the head of a
-chain — and most of that never passes through a write the ledger sees. So
-the ledger does not keep a copy of any of it; readers ask HY, once per read,
-for the nodes it has set aside.
-
-Only the vector store is read. HY writes a node's status there on every
-transition, and the graph holds no status the vector store lacks.
-"""
-
-import json
-from collections.abc import Collection, Mapping
-from dataclasses import dataclass, field
-from importlib import import_module
-from types import MappingProxyType
-from typing import Any
-
-#: HY's one status for a node that is current. Everything else it writes —
-#: superseded, shadow, archived, negated, conflicted — takes it out of recall.
-_ACTIVE = "active"
-#: HY's verbatim chunks. Stored as shadow from the start and never a memory,
-#: so never something a reader could have shown.
-_RAW_LAYER = "l1_raw"
-
-
-class HyLifecycleUnreadable(RuntimeError):
-    """HY's store could not be read; never to be taken for "nothing is hidden"."""
-
-
-@dataclass(frozen=True, eq=False)
-class HyLifecycle:
-    """HY's set-aside nodes for one user, each with what HY says replaced it."""
-
-    hidden: Mapping[str, tuple[str, ...]] = field(default_factory=dict)
-
-    def __post_init__(self) -> None:
-        object.__setattr__(self, "hidden", MappingProxyType(dict(self.hidden)))
-
-    def __eq__(self, other: object) -> bool:
-        return isinstance(other, HyLifecycle) and dict(self.hidden) == dict(other.hidden)
-
-    def __hash__(self) -> int:
-        return hash(tuple(sorted(self.hidden.items())))
-
-    def replaced_by(self, memory_id: str) -> tuple[str, ...]:
-        return self.hidden.get(memory_id, ())
-
-    def replacing(self, memory_id: str) -> tuple[str, ...]:
-        """The set-aside memories HY names this one as the replacement of."""
-        return tuple(sorted(old for old, by in self.hidden.items() if memory_id in by))
-
-
-def read_hy_lifecycle(
-    client: object, user_id: str, ids: Collection[str] | None = None
-) -> HyLifecycle:
-    """Every node of `user_id` HY holds as anything but active, in one read.
-
-    With `ids`, only those nodes are looked up — what a single-memory read
-    needs, and a lookup by key rather than a scan; `replacing` is then only
-    known among them.
-
-    Read one level below HY's own listing, which drops a failed read and
-    answers empty: a hidden memory would then read as current.
-    """
-    native: Any = client
-    try:
-        chroma = import_module("hy_memory.data.vector_store_chroma")
-        store = native._vector_store
-        collection = store._collection
-        if not isinstance(store, chroma.ChromaVectorStore) or collection is None:
-            raise HyLifecycleUnreadable("hy_lifecycle_store_unsupported")
-        found = collection.get(
-            ids=None if ids is None else list(ids),
-            where={
-                "$and": [
-                    {"user_id": {"$eq": user_id}},
-                    {"status": {"$ne": _ACTIVE}},
-                    {"layer": {"$ne": _RAW_LAYER}},
-                ]
-            },
-            include=["metadatas"],
-        )
-        metadatas = found["metadatas"]
-        point_ids = found["ids"]
-        if (
-            not isinstance(point_ids, list)
-            or not isinstance(metadatas, list)
-            or len(point_ids) != len(metadatas)
-        ):
-            raise HyLifecycleUnreadable("hy_lifecycle_store_mismatch")
-    except HyLifecycleUnreadable:
-        raise
-    except Exception as error:
-        raise HyLifecycleUnreadable("hy_lifecycle_unreadable") from error
-    hidden: dict[str, tuple[str, ...]] = {}
-    for point_id, metadata in zip(point_ids, metadatas, strict=True):
-        if not isinstance(point_id, str) or not point_id or not isinstance(metadata, Mapping):
-            raise HyLifecycleUnreadable("hy_lifecycle_store_mismatch")
-        payload = chroma.ChromaVectorStore._chroma_metadata_to_payload(dict(metadata))
-        node_id = payload.get("node_id") or point_id
-        if not isinstance(node_id, str) or not node_id or node_id != point_id:
-            raise HyLifecycleUnreadable("hy_lifecycle_store_mismatch")
-        hidden[node_id] = _replacements(payload)
-    return HyLifecycle(hidden=hidden)
-
-
-def _replacements(payload: Mapping[str, object]) -> tuple[str, ...]:
-    """What HY names as taking this node's place: its chain successor or the kept duplicate."""
-    names: list[str] = []
-    superseded_by = payload.get("superseded_by")
-    if isinstance(superseded_by, list):
-        names.extend(value for value in superseded_by if isinstance(value, str) and value)
-    kept = payload.get("dedup_kept_id")
-    custom = payload.get("custom")
-    if isinstance(custom, str):
-        try:
-            custom = json.loads(custom)
-        except ValueError:
-            custom = None
-    if not isinstance(kept, str) and isinstance(custom, Mapping):
-        audit = custom.get("hy_dedup")
-        if isinstance(audit, Mapping):
-            kept = audit.get("kept_id")
-    if isinstance(kept, str) and kept:
-        names.append(kept)
-    return tuple(dict.fromkeys(names))
@@ -1,331 +0,0 @@
-"""Transport validation for the pinned HY SDK; HY's prompts and decisions pass through unchanged."""
-
-import logging
-from contextlib import nullcontext
-from time import monotonic
-from typing import Any
-
-from one_memory.diagnostic_log import diagnostic_event, diagnostic_exception, elapsed_ms
-from one_memory.hy_source_context import HyWriteBindings
-from one_memory.hy_step_timing import HyStepTimer
-
-
-def _http_status(value: Any) -> int | None:
-    return value if type(value) is int and 100 <= value <= 599 else None
-
-
-def _exception_status(error: Exception) -> int | None:
-    try:
-        status = _http_status(getattr(error, "status_code", None))
-        if status is not None:
-            return status
-        status = _http_status(getattr(getattr(error, "response", None), "status_code", None))
-        if status is not None:
-            return status
-        # Native LLMRetryError retains sanitized attempt dictionaries. Only the
-        # final attempt describes the terminal failure; never parse error text.
-        attempts = getattr(error, "attempts", None)
-        if isinstance(attempts, (list, tuple)) and attempts and isinstance(attempts[-1], dict):
-            return _http_status(attempts[-1].get("status_code"))
-    except Exception:
-        # Diagnostic attribute access must not replace the provider exception.
-        return None
-    return None
-
-
-def _model_failure_code(status: int | None) -> str:
-    specific = {
-        401: "hy_memory_model_authentication_failed",
-        402: "hy_memory_model_payment_required",
-        403: "hy_memory_model_access_denied",
-        408: "hy_memory_model_timeout",
-        429: "hy_memory_model_rate_limited",
-    }
-    if status in specific:
-        return specific[status]
-    if status is not None and 500 <= status <= 599:
-        return "hy_memory_model_service_unavailable"
-    if status is not None and 400 <= status <= 422:
-        return "hy_memory_model_request_incompatible"
-    return "hy_memory_model_request_failed"
-
-
-def _terminal_client_timeout(error: Exception) -> bool:
-    try:
-        attempts = getattr(error, "attempts", None)
-        if not isinstance(attempts, (list, tuple)) or not attempts:
-            return False
-        final = attempts[-1]
-        if type(final) is not dict:
-            return False
-        kind = final.get("failure_kind")
-        return type(kind) is str and kind == "timeout"
-    except Exception:
-        return False
-
-
-class HyResponseGuard:
-    def __init__(self, provider: Any, *, phase: str = "extract") -> None:
-        self.provider = provider
-        #: Diagnostic label only; the request passes through unchanged.
-        self.phase = phase
-        self.failures: list[str] = []
-
-    def __getattr__(self, name: str) -> Any:
-        return getattr(self.provider, name)
-
-    async def complete(self, *args: Any, **kwargs: Any) -> Any:
-        return await self._complete("complete", args, kwargs)
-
-    async def complete_messages(self, *args: Any, **kwargs: Any) -> Any:
-        return await self._complete("complete_messages", args, kwargs)
-
-    async def _complete(self, method: str, args: tuple[Any, ...], kwargs: dict[str, Any]) -> Any:
-        for attempt in range(2):
-            started = monotonic()
-            try:
-                response = await getattr(self.provider, method)(*args, **kwargs)
-            except Exception as error:
-                status = _exception_status(error)
-                code = (
-                    "hy_memory_model_timeout"
-                    if status is None and _terminal_client_timeout(error)
-                    else _model_failure_code(status)
-                )
-                self.failures.append(code)
-                logging.getLogger(__name__).warning(
-                    "%s: status_code=%s", code, status if status is not None else "unknown"
-                )
-                diagnostic_event(
-                    level="error",
-                    component="hy.model",
-                    event="hy.model_call.failed",
-                    outcome="failed",
-                    duration_ms=elapsed_ms(started),
-                    error_code=code,
-                    **diagnostic_exception(error),
-                    details={"phase": self.phase, "attempt": attempt + 1, "call_count": 1},
-                )
-                raise
-            truncated = getattr(response, "finish_reason", "") == "length"
-            empty = not (response.content or "").strip() and not getattr(
-                response, "tool_calls", None
-            )
-            if not truncated and not empty:
-                diagnostic_event(
-                    level="info",
-                    component="hy.model",
-                    event="hy.model_call.completed",
-                    outcome="succeeded",
-                    duration_ms=elapsed_ms(started),
-                    details={"phase": self.phase, "attempt": attempt + 1, "call_count": 1},
-                )
-                return response
-            code = (
-                "hy_memory_model_output_truncated"
-                if truncated
-                else "hy_memory_model_empty_response"
-            )
-            if attempt:
-                self.failures.append(code)
-                raise ValueError(code)
-            logging.getLogger(__name__).warning("%s: retrying the same HY model request", code)
-            if truncated and "max_tokens" in kwargs:
-                kwargs = {**kwargs, "max_tokens": min(16000, kwargs["max_tokens"] * 2)}
-        raise AssertionError("unreachable")
-
-
-class HyReconcileGuard:
-    #: Set only for the span of one leased write.
-    step_timer: HyStepTimer | None = None
-
-    def __init__(self, reconciler: Any) -> None:
-        self.reconciler = reconciler
-        self.failures: list[str] = []
-        self.bindings: HyWriteBindings | None = None
-        self.imported_contents: frozenset[str] | None = None
-        self.mutation_target_ids: set[str] = set()
-
-    def begin_write(self) -> None:
-        self.mutation_target_ids.clear()
-
-    def __getattr__(self, name: str) -> Any:
-        return getattr(self.reconciler, name)
-
-    async def reconcile(self, *args: Any, **kwargs: Any) -> Any:
-        timer = self.step_timer
-        with timer.step("reconcile") if timer is not None else nullcontext():
-            return await self._reconcile(args, kwargs)
-
-    async def _reconcile(self, args: tuple[Any, ...], kwargs: dict[str, Any]) -> Any:
-        started = monotonic()
-        if self.imported_contents is not None:
-            metadata = kwargs.get("new_memories_with_meta") or []
-            texts = kwargs.get("new_memories") or []
-            if (
-                not metadata
-                or not texts
-                or any(
-                    not isinstance(item, dict) or item.get("content") not in self.imported_contents
-                    for item in metadata
-                )
-                or any(text not in self.imported_contents for text in texts)
-                or {item["content"] for item in metadata} != set(texts)
-            ):
-                code = "hy_memory_write_verification_import_input_mismatch"
-                self.failures.append(code)
-                raise ValueError(code)
-        try:
-            result = await self.reconciler.reconcile(*args, **kwargs)
-        except Exception as error:
-            diagnostic_event(
-                level="error",
-                component="hy.reconcile",
-                event="hy.reconcile.failed",
-                outcome="failed",
-                duration_ms=elapsed_ms(started),
-                error_code="hy_memory_reconcile_failed",
-                **diagnostic_exception(error),
-            )
-            raise
-        if not result.success:
-            self.failures.append("hy_memory_reconcile_failed")
-            diagnostic_event(
-                level="error",
-                component="hy.reconcile",
-                event="hy.reconcile.failed",
-                outcome="failed",
-                duration_ms=elapsed_ms(started),
-                error_code="hy_memory_reconcile_failed",
-                details={"result_count": len(result.ops)},
-            )
-            return result
-        try:
-            await self._check_targets(result, kwargs)
-        except ValueError as error:
-            code = str(error)
-            self.failures.append(code)
-            raise
-        except Exception as error:
-            code = "hy_memory_write_verification_failed"
-            self.failures.append(code)
-            raise ValueError(code) from error
-        owners: dict[str, set[str]] = {}
-        for item in kwargs.get("new_memories_with_meta") or []:
-            if item.get("owner") in {"user", "agent"}:
-                owners.setdefault(item["content"], set()).add(item["owner"])
-        # HY's no-candidate ADD path copies content verbatim but omits owner.
-        # Restore only an unambiguous exact match, never infer from wording.
-        for operation in result.ops:
-            source_owners = owners.get(operation.content, set())
-            if operation.op == "ADD" and operation.owner is None and len(source_owners) == 1:
-                operation.owner = next(iter(source_owners))
-        operation_counts = {"ADD": 0, "UPDATE": 0, "SUPERSEDE": 0, "DELETE": 0}
-        for operation in result.ops:
-            raw_name = getattr(operation, "op", "")
-            name = str(getattr(raw_name, "value", raw_name)).upper()
-            if name in {"UPDATE", "SUPERSEDE", "DELETE", "SHADOW"}:
-                target = getattr(operation, "memory_id", None)
-                if isinstance(target, str) and target:
-                    self.mutation_target_ids.add(target)
-                self.mutation_target_ids.update(
-                    key
-                    for key in getattr(operation, "memory_ids", ())
-                    if isinstance(key, str) and key
-                )
-            if name == "ADD":
-                operation_counts["ADD"] += 1
-            elif name in {"UPDATE", "MERGE", "NOOP"}:
-                operation_counts["UPDATE"] += 1
-            elif name == "SUPERSEDE":
-                operation_counts["SUPERSEDE"] += 1
-            elif name in {"DELETE", "SHADOW"}:
-                operation_counts["DELETE"] += 1
-        diagnostic_event(
-            level="info",
-            component="hy.reconcile",
-            event="hy.reconcile.completed",
-            outcome="succeeded",
-            duration_ms=elapsed_ms(started),
-            details={
-                "created_count": operation_counts["ADD"],
-                "updated_count": operation_counts["UPDATE"],
-                "superseded_count": operation_counts["SUPERSEDE"],
-                "deleted_count": operation_counts["DELETE"],
-                "result_count": len(result.ops),
-            },
-        )
-        return result
-
-    async def _check_targets(self, result: Any, kwargs: dict[str, Any]) -> None:
-        """HY may only touch nodes of this user and scope that this write was allowed to see."""
-        targets = {op.memory_id for op in result.ops if op.memory_id}
-        targets.update(key for op in result.ops for key in getattr(op, "memory_ids", []))
-        for key in targets:
-            node = await kwargs["vector_store"].get_by_id(key)
-            if (
-                node is None
-                or node.user_id != kwargs["user_id"]
-                or node.agent_id != kwargs["agent_id"]
-            ):
-                raise ValueError("hy_memory_write_verification_target_invalid")
-            if self.bindings is not None:
-                known = self.bindings.known.get(key)
-                if known is None or (known.content, known.owner) != (node.content, node.owner):
-                    raise ValueError("hy_memory_write_verification_target_unbound")
-        if self.bindings is None:
-            return
-        for op in result.ops:
-            if not op.content:
-                continue
-            linked = {key for key in (op.memory_id, *getattr(op, "memory_ids", [])) if key}
-            self.bindings.record_operation_support(op.content, op.owner, linked)
-            if op.memory_id:
-                self.bindings.verified_operations[op.memory_id] = (op.content, op.owner)
-
-
-def install_hy_write_guards(client: Any) -> None:
-    """The pinned HY SDK treats an empty reconcile response as a successful empty op list.
-
-    Bind this client's real write paths. Pro/lite use the ordinary writer;
-    ultra uses the System2 writer's inner writer. Keep HY prompts, extraction,
-    retrieval, reconciliation and storage intact; never patch the shared SDK
-    installation.
-    """
-
-    async def install() -> None:
-        from one_memory.hy_write_verification import HyExtractionGuard
-
-        ordinary = await client._registry.get_writer()
-        system2 = await client._registry.get_system2_writer()
-        guards: list[Any] = []
-        seen_writers: set[int] = set()
-        seen_guards: set[int] = set()
-        for writer in (ordinary, system2._writer):
-            if id(writer) in seen_writers:
-                continue
-            seen_writers.add(id(writer))
-            agent = writer.mem_agent
-            if not isinstance(agent.llm_provider, HyResponseGuard):
-                agent.llm_provider = HyResponseGuard(agent.llm_provider)
-                agent.extractor.llm = agent.llm_provider
-            if not isinstance(agent.extractor, HyExtractionGuard):
-                agent.extractor = HyExtractionGuard(agent.extractor)
-            reconciler = writer.reconciler
-            provider = reconciler._get_llm()
-            if not isinstance(provider, HyResponseGuard):
-                reconciler._llm = HyResponseGuard(provider, phase="reconcile")
-            if not isinstance(reconciler, HyReconcileGuard):
-                writer._reconciler = HyReconcileGuard(reconciler)
-            for guard in (
-                agent.llm_provider,
-                reconciler._llm,
-                writer._reconciler,
-                agent.extractor,
-            ):
-                if id(guard) not in seen_guards:
-                    seen_guards.add(id(guard))
-                    guards.append(guard)
-        client._one_memory_write_guards = tuple(guards)
-
-    client._loop_thread.run(install())
@@ -14,7 +14,7 @@ from pathlib import Path
 from typing import Literal, cast

 HY_MEMORY_DISTRIBUTION = "hy-memory"
-HY_MEMORY_VERSION = "2.0.5.dev3"
+HY_MEMORY_VERSION = "2.0.6"
 HY_MEMORY_NAMESPACE = "governed-v2"
 DiagnosticLogLevel = Literal["error", "warning", "info", "debug"]

@@ -33,49 +33,8 @@ _COLLECTION_LOGGERS = (
 )
 _NOISY_LIBRARY_LOGGERS = ("httpx", "httpcore", "watchfiles", "jieba")

-# One Memory deliberately stays on ultra because its existing, user-visible
-# profile/graph flow calls digest(), which the pinned public SDK exposes only
-# for ultra clients. Reader and non-reranker behavior remain pinned for
-# compatibility with the SDK authority-migration development preview.
-HY_MEMORY_MODE = "ultra"
-# The SDK only recognises `legacy` and `hybrid` (see
-# hy_memory/pipelines/_retrieval/config.py: `ALL_READERS = (READER_LEGACY,
-# READER_HYBRID)`). The old spelling `hybrid_v2` was silently rejected and
-# fell back to legacy, so the hybrid-v2 recall strategy never ran at all.
 HY_MEMORY_READER = "hybrid"

-HY_MEMORY_BEHAVIOR_ENV: dict[str, str] = {
-    "MEMORY_MODE": HY_MEMORY_MODE,
-    "MEMORY_ENABLE_GRAPH": "true",
-    "MEMORY_GRAPH_RERANK_ENABLED": "false",
-    "MEMORY_GRAPH_TEMPORAL_SPINE": "true",
-    "MEMORY_TOPIC_SUMMARY_ENABLED": "false",
-    "MEMORY_EPISODE_SUMMARY_ENABLED": "false",
-    "MEMORY_EPISODE_RECALL_ENABLED": "true",
-    "MEMORY_DYNAMIC_EPISODE_ENABLED": "true",
-    "MEMORY_SESSION_EPISODE_ENABLED": "false",
-    "MEMORY_SESSION_EPISODE_CONTEXT_ENABLED": "false",
-    "MEMORY_EPISODE_CAPSULE_CONTEXT_ENABLED": "true",
-    "MEMORY_EPISODE_CAPSULE_QUERY_MIN": "0.65",
-    "MEMORY_EPISODE_CAPSULE_ASSIGNMENT_MIN": "0.60",
-    "MEMORY_EPISODE_CAPSULE_FIRST_RANK_MAX": "3",
-    "MEMORY_EPISODE_CAPSULE_MEMBER_LIMIT": "3",
-    "MEMORY_ENABLE_INTENTION": "false",
-    "MEMORY_ENABLE_QUERY_REWRITE": "false",
-    "MEMORY_FEW_SHOT_ENABLED": "true",
-    "MEMORY_HYBRID_V2_QUERY_EXPAND": "1",
-    "MEMORY_HYBRID_V2_GEO_INTENT_PRIOR": "0.6",
-    "MEMORY_HYBRID_V2_DIVERSITY": "1",
-    "MEMORY_HYBRID_V2_TEMPORAL_ADMIT_PRIOR": "0.55",
-    "MEMORY_HYBRID_V2_TEMPORAL_BOOST": "0.18",
-    "MEMORY_CROSS_DOMAIN_CON_THRESHOLD": "0.5",
-    "MEMORY_READER_ENABLE_SUMMARY": "false",
-    "MEMORY_SUMMARY_ENABLED_IN_SYS2": "false",
-    "MEMORY_LOCAL_RERANK_ENABLED": "false",
-    "RECONCILE_ENABLED": "true",
-    "HY_MEMORY_READER": HY_MEMORY_READER,
-}
-

 def detailed_logging_enabled(data_dir: Path) -> bool:
    return diagnostic_log_level(data_dir) == "debug"
@@ -168,7 +127,6 @@ def configure_hy_runtime_environment(data_dir: Path) -> None:
    log_dir = memory_root / "logs"
    memory_root.mkdir(parents=True, exist_ok=True)
    log_dir.mkdir(parents=True, exist_ok=True)
-    os.environ.update(HY_MEMORY_BEHAVIOR_ENV)
    os.environ["MEMORY_DATA_DIR"] = str(memory_root)
    os.environ["MEMORY_LOG_DIR"] = str(log_dir)
    level = diagnostic_log_level(data_dir)
@@ -178,9 +136,3 @@ def configure_hy_runtime_environment(data_dir: Path) -> None:
    os.environ["MEMORY_MEMORY_OPERATIONS_ENABLED"] = "false"
    os.environ["ANONYMIZED_TELEMETRY"] = "False"
    os.environ.pop("CHROMA_OTEL_COLLECTION_ENDPOINT", None)
-
-
-def configure_hy_behavior_environment() -> None:
-    """Set import-time HY behavior flags without touching paths or secrets."""
-
-    os.environ.update(HY_MEMORY_BEHAVIOR_ENV)
@@ -0,0 +1,469 @@
+"""Offline maintenance of the HY SDK data plane for App migrations.
+
+Runtime product code must use public ``HyMemoryClient`` methods.  The startup
+migration boundary is the isolated exception: all writers are stopped, so it
+may copy, restore and repair SDK-owned local stores.
+"""
+
+from __future__ import annotations
+
+import json
+import sqlite3
+from collections.abc import Iterator
+from contextlib import closing
+from dataclasses import dataclass, field
+from datetime import UTC, datetime, timedelta
+from pathlib import Path
+from typing import Any
+
+from hy_memory import (  # type: ignore[import-untyped]
+    DataPlaneRestore,
+    DataPlaneSnapshot,
+    restore_local_data_plane,
+    snapshot_local_data_plane,
+)
+
+from one_memory.database_access import main_database
+from one_memory.hy_runtime_config import HY_MEMORY_NAMESPACE
+
+# v2 reads the backfill's governed_memory_evidence links; v1 missed them.
+_METADATA_REPAIR_VERSION = 2
+_BATCH = 500
+
+
+@dataclass(frozen=True)
+class RepairMetadata:
+    source: str | None
+    memory_at: int | None
+    created_at: int | None
+
+
+@dataclass(frozen=True)
+class MetadataRepairResult:
+    migration_id: str | None
+    mapped: int
+    vector_rows: int
+    coding_rows: int
+
+
+def snapshot_data_plane(
+    *,
+    source_root: Path,
+    target_root: Path,
+    writers_stopped: bool,
+) -> DataPlaneSnapshot:
+    """Snapshot an explicitly quiesced SDK-owned local data plane."""
+    return snapshot_local_data_plane(
+        source_root=source_root,
+        target_root=target_root,
+        writers_stopped=writers_stopped,
+    )
+
+
+def restore_data_plane(
+    *,
+    snapshot_root: Path,
+    target_root: Path,
+    writers_stopped: bool,
+    protective_root: Path | None = None,
+) -> DataPlaneRestore:
+    """Restore an explicitly quiesced SDK-owned local data plane."""
+    return restore_local_data_plane(
+        snapshot_root=snapshot_root,
+        target_root=target_root,
+        writers_stopped=writers_stopped,
+        protective_root=protective_root,
+    )
+
+
+def repair_sdk_metadata_from_bundle(
+    *,
+    data_dir: Path,
+    database_path: Path,
+    now: datetime | None = None,
+) -> MetadataRepairResult:
+    """Restore metadata lost by historical cutovers from their verified bundles."""
+    migration_id = _sdk_migration_id(database_path)
+    if migration_id is None:
+        return MetadataRepairResult(None, 0, 0, 0)
+    receipt = data_dir / f".sdk-metadata-repair-v{_METADATA_REPAIR_VERSION}-{migration_id}.json"
+    if receipt.is_file():
+        document = json.loads(receipt.read_text(encoding="utf-8"))
+        return MetadataRepairResult(
+            migration_id,
+            int(document.get("mapped", 0)),
+            int(document.get("vector_rows", 0)),
+            int(document.get("coding_rows", 0)),
+        )
+    # A home can carry several cutovers (v1 and then v2). A later bundle may
+    # have been taken after an earlier cutover had already retired the legacy
+    # tables, so every bundle is read and their links are pooled.
+    legacy = sorted(
+        (data_dir / "backups").glob("*-pre-cutover.sdk-bundle/management/one-memory.sqlite3")
+    )
+    if not legacy:
+        # No bundle yet is not a finished repair: try again on the next start.
+        return MetadataRepairResult(migration_id, 0, 0, 0)
+    current = now or datetime.now(UTC)
+    pools: dict[str, _Pool] = {}
+    for path in legacy:
+        _merge_pools(pools, _legacy_pools(path, now=current))
+    mapping = _decide(pools)
+    vector_rows = _repair_vector(data_dir, mapping, now=current)
+    coding_rows = _repair_coding(data_dir, mapping)
+    result = MetadataRepairResult(
+        migration_id,
+        len(mapping),
+        vector_rows,
+        coding_rows,
+    )
+    temporary = receipt.with_suffix(".tmp")
+    temporary.write_text(
+        json.dumps(
+            {
+                "schema_version": 1,
+                "migration_id": migration_id,
+                "mapped": result.mapped,
+                "vector_rows": vector_rows,
+                "coding_rows": coding_rows,
+            },
+            sort_keys=True,
+        )
+        + "\n",
+        encoding="utf-8",
+    )
+    temporary.replace(receipt)
+    return result
+
+
+def _sdk_migration_id(database_path: Path) -> str | None:
+    with main_database(database_path).read(row_factory=sqlite3.Row) as database:
+        row = database.execute(
+            "SELECT authority,migration_id FROM memory_authority_state WHERE singleton=1"
+        ).fetchone()
+    if row is None or row["authority"] != "sdk":
+        return None
+    value = row["migration_id"]
+    return str(value) if isinstance(value, str) and value else None
+
+
+@dataclass
+class _Pool:
+    """Everything the legacy library knew about one SDK memory."""
+
+    sources: set[str] = field(default_factory=set)
+    times: list[datetime] = field(default_factory=list)
+    created: list[datetime] = field(default_factory=list)
+
+
+def _legacy_metadata(path: Path, *, now: datetime) -> dict[str, RepairMetadata]:
+    return _decide(_legacy_pools(path, now=now))
+
+
+def _legacy_pools(path: Path, *, now: datetime) -> dict[str, _Pool]:
+    """Link each SDK memory id to the legacy evidence it was made from.
+
+    Three legacy paths wrote memories the SDK now holds:
+
+    - the cutover backfill (`import_exact`) copied every governed memory into the
+      SDK under its legacy id, or under the id `memory_task_outputs` recorded,
+      without a source or a conversation time; `governed_memory_evidence` is the
+      link from that memory to its evidence;
+    - extraction tasks recorded their evidence as `support_refs`;
+    - the write queue recorded the SDK ids each job produced.
+    """
+    with closing(sqlite3.connect(f"file:{path}?mode=ro", uri=True)) as database:
+        database.row_factory = sqlite3.Row
+        tables = {
+            str(row[0])
+            for row in database.execute("SELECT name FROM sqlite_master WHERE type='table'")
+        }
+        if "evidence" not in tables:
+            return {}
+        links: dict[str, set[str]] = {}
+        created: dict[str, list[datetime]] = {}
+        canonical: dict[str, str] = {}
+        if "memory_task_outputs" in tables:
+            for row in database.execute("SELECT body FROM memory_task_outputs"):
+                try:
+                    body = json.loads(str(row["body"]))
+                except (TypeError, ValueError):
+                    continue
+                legacy_id = body.get("memory_id")
+                canonical_id = body.get("hy_memory_id")
+                refs = body.get("support_refs")
+                if not isinstance(legacy_id, str) or not isinstance(canonical_id, str):
+                    continue
+                canonical[legacy_id] = canonical_id
+                links.setdefault(canonical_id, set()).update(
+                    str(ref["evidence_id"])
+                    for ref in (refs if isinstance(refs, list) else ())
+                    if isinstance(ref, dict) and isinstance(ref.get("evidence_id"), str)
+                )
+        if "governed_memories" in tables:
+            for row in database.execute(
+                "SELECT memory_id,created_at,updated_at FROM governed_memories"
+            ):
+                legacy_id = str(row["memory_id"])
+                canonical_id = canonical.get(legacy_id, legacy_id)
+                links.setdefault(canonical_id, set())
+                legacy_created = _plausible_time(
+                    row["created_at"], fallback=row["updated_at"], now=now
+                )
+                if legacy_created is not None:
+                    created.setdefault(canonical_id, []).append(legacy_created)
+            if "governed_memory_evidence" in tables:
+                for row in database.execute(
+                    "SELECT memory_id,evidence_id FROM governed_memory_evidence"
+                ):
+                    legacy_id = str(row["memory_id"])
+                    links.setdefault(canonical.get(legacy_id, legacy_id), set()).add(
+                        str(row["evidence_id"])
+                    )
+        if "memory_write_jobs" in tables and "sdk_memory_ids_json" in _columns(
+            database, "memory_write_jobs"
+        ):
+            job_evidence: dict[str, set[str]] = {}
+            if "memory_write_job_evidence" in tables:
+                for row in database.execute(
+                    "SELECT job_id,evidence_id FROM memory_write_job_evidence"
+                ):
+                    job_evidence.setdefault(str(row["job_id"]), set()).add(str(row["evidence_id"]))
+            for row in database.execute(
+                "SELECT job_id,sdk_memory_ids_json FROM memory_write_jobs "
+                "WHERE sdk_memory_ids_json IS NOT NULL"
+            ):
+                try:
+                    memory_ids = json.loads(str(row["sdk_memory_ids_json"]))
+                except (TypeError, ValueError):
+                    continue
+                for memory_id in memory_ids if isinstance(memory_ids, list) else ():
+                    if isinstance(memory_id, str) and memory_id:
+                        links.setdefault(memory_id, set()).update(
+                            job_evidence.get(str(row["job_id"]), ())
+                        )
+        evidence: dict[str, tuple[str, datetime | None]] = {}
+        wanted = sorted({identifier for refs in links.values() for identifier in refs})
+        for batch in _batches(wanted):
+            placeholders = ",".join("?" for _ in batch)
+            for row in database.execute(
+                "SELECT evidence_id,source_id,created_at,imported_at "
+                f"FROM evidence WHERE evidence_id IN ({placeholders})",
+                batch,
+            ):
+                evidence[str(row["evidence_id"])] = (
+                    str(row["source_id"]),
+                    _plausible_time(row["created_at"], fallback=row["imported_at"], now=now),
+                )
+    pools: dict[str, _Pool] = {}
+    for canonical_id, refs in links.items():
+        pool = pools.setdefault(canonical_id, _Pool())
+        for identifier in refs:
+            if identifier not in evidence:
+                continue
+            source, timestamp = evidence[identifier]
+            if source and source != "default":
+                pool.sources.add(source)
+            if timestamp is not None:
+                pool.times.append(timestamp)
+        pool.created.extend(created.get(canonical_id, ()))
+    return pools
+
+
+def _merge_pools(into: dict[str, _Pool], other: dict[str, _Pool]) -> None:
+    for canonical_id, pool in other.items():
+        merged = into.setdefault(canonical_id, _Pool())
+        merged.sources |= pool.sources
+        merged.times.extend(pool.times)
+        merged.created.extend(pool.created)
+
+
+def _decide(pools: dict[str, _Pool]) -> dict[str, RepairMetadata]:
+    # Several legacy memories can fold into one SDK memory. Their evidence is
+    # pooled before deciding, so a disagreeing source stays unknown instead of
+    # whichever link happened to come last.
+    return {
+        canonical_id: RepairMetadata(
+            source=next(iter(pool.sources)) if len(pool.sources) == 1 else None,
+            memory_at=int(max(pool.times).timestamp()) if pool.times else None,
+            created_at=int(min(pool.created).timestamp()) if pool.created else None,
+        )
+        for canonical_id, pool in pools.items()
+    }
+
+
+def _columns(database: sqlite3.Connection, table: str) -> set[str]:
+    return {str(row[1]) for row in database.execute(f"PRAGMA table_info({table})")}
+
+
+def _plausible_time(
+    value: object,
+    *,
+    fallback: object,
+    now: datetime,
+) -> datetime | None:
+    for candidate in (value, fallback):
+        if not isinstance(candidate, str) or not candidate:
+            continue
+        try:
+            parsed = datetime.fromisoformat(candidate)
+        except ValueError:
+            continue
+        if parsed.tzinfo is None:
+            parsed = parsed.replace(tzinfo=UTC)
+        parsed = parsed.astimezone(UTC)
+        if parsed <= now + timedelta(days=1):
+            return parsed
+    return None
+
+
+def _repair_vector(
+    data_dir: Path,
+    mapping: dict[str, RepairMetadata],
+    *,
+    now: datetime,
+) -> int:
+    directory = data_dir / "hy-memory" / HY_MEMORY_NAMESPACE / "vector"
+    if not (directory / "chroma.sqlite3").is_file():
+        return 0
+    from one_memory.sdk_vector_reattach import open_offline_chroma
+
+    client = open_offline_chroma(directory)
+    # The SDK client is not open yet (startup has not created its handle), so
+    # this is the only writer; release the store before the SDK takes it.
+    try:
+        repaired = 0
+        ids = tuple(mapping)
+        for collection in client.list_collections():
+            for batch in _batches(ids):
+                page = collection.get(ids=list(batch), include=["metadatas"])
+                updated_ids: list[str] = []
+                updated_metadata: list[dict[str, object]] = []
+                for memory_id, metadata in zip(
+                    page.get("ids") or (),
+                    page.get("metadatas") or (),
+                    strict=True,
+                ):
+                    if not isinstance(metadata, dict):
+                        continue
+                    desired = mapping.get(str(memory_id))
+                    if desired is None:
+                        continue
+                    next_metadata = dict(metadata)
+                    changed = False
+                    if desired.source and next_metadata.get("source") in {
+                        None,
+                        "",
+                        "default",
+                    }:
+                        next_metadata["source"] = desired.source
+                        changed = True
+                    if (
+                        desired.memory_at is not None
+                        and next_metadata.get("memory_at") != desired.memory_at
+                    ):
+                        next_metadata["memory_at"] = desired.memory_at
+                        changed = True
+                    if (
+                        desired.created_at is not None
+                        and next_metadata.get("gmt_created") != desired.created_at
+                    ):
+                        next_metadata["gmt_created"] = desired.created_at
+                        changed = True
+                    if changed:
+                        updated_ids.append(str(memory_id))
+                        updated_metadata.append(next_metadata)
+                if updated_ids:
+                    collection.update(ids=updated_ids, metadatas=updated_metadata)
+                    repaired += len(updated_ids)
+            repaired += _settle_future_times(collection, now=now)
+    finally:
+        client.close()
+    return repaired
+
+
+def _settle_future_times(collection: Any, *, now: datetime) -> int:
+    """Move conversation times that are still in the future back to the write time.
+
+    A historical Cursor bug stored a file hash as the message time (years like
+    2254). The SDK lists newest first by `memory_at`, so those rows hold the
+    first page. Rows the bundles could date were fixed above; for the rest the
+    write time is the closest true moment that is known.
+    """
+    cutoff = int((now + timedelta(days=1)).timestamp())
+    page = collection.get(where={"memory_at": {"$gt": cutoff}}, include=["metadatas"])
+    updated_ids: list[str] = []
+    updated_metadata: list[dict[str, object]] = []
+    for memory_id, metadata in zip(page.get("ids") or (), page.get("metadatas") or (), strict=True):
+        if not isinstance(metadata, dict):
+            continue
+        created = metadata.get("gmt_created")
+        settled = (
+            created
+            if isinstance(created, int) and not isinstance(created, bool) and created <= cutoff
+            else int(now.timestamp())
+        )
+        updated_ids.append(str(memory_id))
+        updated_metadata.append({**metadata, "memory_at": settled})
+    if updated_ids:
+        collection.update(ids=updated_ids, metadatas=updated_metadata)
+    return len(updated_ids)
+
+
+def _repair_coding(
+    data_dir: Path,
+    mapping: dict[str, RepairMetadata],
+) -> int:
+    root = data_dir / "hy-memory" / HY_MEMORY_NAMESPACE
+    repaired = 0
+    for name in ("coding.sqlite3", "coding_cv.sqlite3"):
+        path = root / name
+        if not path.is_file():
+            continue
+        with closing(sqlite3.connect(path)) as database:
+            tables = {
+                str(row[0])
+                for row in database.execute("SELECT name FROM sqlite_master WHERE type='table'")
+            }
+            for table, identifier in (
+                ("coding_memory_meta", "experience_id"),
+                ("working_fact", "fact_id"),
+            ):
+                if table not in tables:
+                    continue
+                rows = database.execute(
+                    f"SELECT {identifier},source,created_at FROM {table}"
+                ).fetchall()
+                for memory_id, source, created_at in rows:
+                    desired = mapping.get(str(memory_id))
+                    if desired is None:
+                        continue
+                    updates: list[str] = []
+                    values: list[object] = []
+                    if desired.source and source in {None, "", "default"}:
+                        updates.append("source=?")
+                        values.append(desired.source)
+                    if desired.created_at is not None:
+                        desired_created_at = (
+                            datetime.fromtimestamp(desired.created_at, tz=UTC)
+                            .replace(tzinfo=None)
+                            .isoformat()
+                        )
+                        if created_at != desired_created_at:
+                            updates.append("created_at=?")
+                            values.append(desired_created_at)
+                    if not updates:
+                        continue
+                    values.append(memory_id)
+                    database.execute(
+                        f"UPDATE {table} SET {','.join(updates)} WHERE {identifier}=?",
+                        values,
+                    )
+                    repaired += 1
+            database.commit()
+    return repaired
+
+
+def _batches[T](values: list[T] | tuple[T, ...]) -> Iterator[tuple[T, ...]]:
+    for start in range(0, len(values), _BATCH):
+        yield tuple(values[start : start + _BATCH])
@@ -0,0 +1,446 @@
+"""HY SDK client construction and shared-runtime leasing.
+
+This module deliberately depends only on the SDK's package-root exports.  A
+lease owns one client while the process shares one ``SharedRuntime`` per
+physical MemHub data directory.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import hashlib
+import os
+import threading
+from collections.abc import Callable, Coroutine
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any, ClassVar
+
+from hy_memory import (  # type: ignore[import-untyped]
+    AmbiguousMemoryIdError as AmbiguousMemoryIdError,  # re-export for api.py
+)
+from hy_memory import (
+    HyMemoryClient,
+    MemoryConfig,
+    SharedRuntime,
+)
+
+from one_memory.embedding_assets import (
+    BUILTIN_EMBEDDING_DIMENSIONS,
+    BUILTIN_EMBEDDING_MODEL,
+)
+from one_memory.hy_graph_admin import (
+    GraphWidthMismatchError,
+    graph_store_enabled,
+    inspect_graph_width_mismatch,
+)
+from one_memory.hy_sdk_types import HySdkRuntimeError
+from one_memory.local_embedding import LocalEmbedder
+from one_memory.model_profiles import ModelProfileSnapshot, ProviderKind
+
+_RUNTIME_LOCK = threading.RLock()
+BUILTIN_RETRIEVAL_PROFILE_ID = "builtin-retrieval"
+_LOCAL_ONNX_PLACEHOLDER_URL = "http://127.0.0.1/local-onnx"
+DEFAULT_RELEVANCE_FLOOR = 0.35
+
+
+@dataclass
+class _RuntimeEntry:
+    runtime: SharedRuntime
+    references: int
+
+
+_RUNTIMES: dict[Path, _RuntimeEntry] = {}
+
+
+class LocalOnnxEmbeddingService:
+    """Public ``EmbeddingService`` implementation backed by MemHub's ONNX model."""
+
+    _embedder: ClassVar[LocalEmbedder | None] = None
+    _embedder_lock: ClassVar[threading.Lock] = threading.Lock()
+
+    def __init__(self, embedder: LocalEmbedder | None = None) -> None:
+        if embedder is None:
+            with self._embedder_lock:
+                if self._embedder is None:
+                    self.__class__._embedder = LocalEmbedder()
+                embedder = self._embedder
+        if embedder is None:  # pragma: no cover - guarded by the assignment above
+            raise HySdkRuntimeError("hy_sdk_local_embedding_unavailable")
+        self._local_embedder = embedder
+
+    async def embed(self, text: str, use_cache: bool = True) -> list[float]:
+        del use_cache
+        vectors = await asyncio.to_thread(self._local_embedder.encode, [text])
+        return vectors[0]
+
+    async def embed_with_meta(
+        self,
+        text: str,
+        use_cache: bool = True,
+    ) -> tuple[list[float], dict[str, Any]]:
+        return await self.embed(text, use_cache), {"provider": "local-onnx"}
+
+    async def embed_batch(
+        self,
+        texts: list[str],
+        use_cache: bool = True,
+    ) -> list[list[float]]:
+        del use_cache
+        if not texts:
+            return []
+        return await asyncio.to_thread(self._local_embedder.encode, list(texts))
+
+    async def embed_queued(self, text: str, use_cache: bool = True) -> list[float]:
+        return await self.embed(text, use_cache)
+
+    async def embed_queued_with_meta(
+        self,
+        text: str,
+        use_cache: bool = True,
+    ) -> tuple[list[float], dict[str, Any]]:
+        return await self.embed_with_meta(text, use_cache)
+
+    def get_embedding_dims(self) -> int:
+        return BUILTIN_EMBEDDING_DIMENSIONS
+
+
+class ClientLease:
+    """Own a client and release its shared runtime exactly once."""
+
+    def __init__(self, client: HyMemoryClient, data_dir: Path) -> None:
+        self.client = client
+        self.data_dir = data_dir.resolve()
+        self._closed = False
+        self._close_lock = threading.Lock()
+
+    @property
+    def closed(self) -> bool:
+        return self._closed
+
+    def close(self) -> None:
+        with self._close_lock:
+            if self._closed:
+                return
+            self._closed = True
+            try:
+                self.client.close()
+            finally:
+                _release_runtime(self.data_dir)
+
+    def __enter__(self) -> ClientLease:
+        return self
+
+    def __exit__(self, *_exc: object) -> None:
+        self.close()
+
+
+class SdkClientHandle:
+    """Lazy async owner for one profile-bound public SDK client lease."""
+
+    def __init__(
+        self,
+        profile: ModelProfileSnapshot,
+        *,
+        api_key: str | None,
+        data_dir: Path,
+        http_transport_factory: Callable[[], Any] | None = None,
+    ) -> None:
+        self._profile = profile
+        self._api_key = api_key
+        self._data_dir = data_dir
+        self._http_transport_factory = http_transport_factory
+        self._lease: ClientLease | None = None
+        self._lock = asyncio.Lock()
+
+    async def client(self) -> HyMemoryClient:
+        if self._lease is not None:
+            return self._lease.client
+        async with self._lock:
+            if self._lease is None:
+                self._lease = await asyncio.to_thread(
+                    acquire_client,
+                    self._profile,
+                    api_key=self._api_key,
+                    data_dir=self._data_dir,
+                    http_transport_factory=self._http_transport_factory,
+                )
+            return self._lease.client
+
+    async def close(self) -> None:
+        async with self._lock:
+            lease, self._lease = self._lease, None
+        if lease is not None:
+            await asyncio.to_thread(lease.close)
+
+
+def builtin_retrieval_snapshot() -> ModelProfileSnapshot:
+    """Return the model-free profile used by SDK-native read operations."""
+    return ModelProfileSnapshot(
+        profile_id=BUILTIN_RETRIEVAL_PROFILE_ID,
+        revision=0,
+        provider_kind=ProviderKind.LOCAL_ONNX,
+        base_url=_LOCAL_ONNX_PLACEHOLDER_URL,
+        model=BUILTIN_EMBEDDING_MODEL,
+        embedding_provider_kind=ProviderKind.LOCAL_ONNX,
+        embedding_model=BUILTIN_EMBEDDING_MODEL,
+        embedding_dimensions=BUILTIN_EMBEDDING_DIMENSIONS,
+        timeout_seconds=60,
+        max_concurrency=1,
+        context_window=None,
+        has_api_key=False,
+    )
+
+
+def sdk_relevance_floor() -> float:
+    raw = os.environ.get("ONE_MEMORY_HY_RELEVANCE_FLOOR", "").strip()
+    if not raw:
+        return DEFAULT_RELEVANCE_FLOOR
+    try:
+        value = float(raw)
+    except ValueError:
+        return DEFAULT_RELEVANCE_FLOOR
+    return min(1.0, max(0.0, value))
+
+
+def build_sdk_config(
+    profile: ModelProfileSnapshot,
+    *,
+    api_key: str | None,
+    data_dir: Path,
+) -> MemoryConfig:
+    """Map a model snapshot to SDK models, endpoints, backends, and local paths."""
+    generation_endpoint, generation_key = _generation_access(profile, api_key)
+    embedding_provider = profile.embedding_provider_kind or profile.provider_kind
+    embedding_service_is_local = embedding_provider is ProviderKind.LOCAL_ONNX
+    embedding_model = profile.embedding_model or (
+        BUILTIN_EMBEDDING_MODEL if embedding_service_is_local else None
+    )
+    if not embedding_model:
+        raise HySdkRuntimeError("hy_sdk_embedding_model_required")
+    dimensions: int | None
+    if embedding_service_is_local:
+        embedding_endpoint = None
+        embedding_key = None
+        dimensions = profile.embedding_dimensions or BUILTIN_EMBEDDING_DIMENSIONS
+    else:
+        embedding_endpoint = _openai_endpoint(
+            embedding_provider,
+            profile.embedding_base_url or profile.base_url,
+        )
+        embedding_key = "ollama" if embedding_provider is ProviderKind.OLLAMA else api_key
+        if embedding_provider is ProviderKind.OPENAI_COMPATIBLE and not embedding_key:
+            raise HySdkRuntimeError("hy_sdk_embedding_api_key_missing")
+        if embedding_provider not in {
+            ProviderKind.OPENAI_COMPATIBLE,
+            ProviderKind.OLLAMA,
+        }:
+            raise HySdkRuntimeError("hy_sdk_embedding_provider_unsupported")
+        dimensions = profile.embedding_dimensions
+
+    root = data_dir.resolve() / "hy-memory" / "governed-v2"
+    root.mkdir(parents=True, exist_ok=True)
+    collection = _collection_base(
+        embedding_provider, embedding_endpoint, embedding_model, dimensions
+    )
+
+    return MemoryConfig.from_dict(
+        {
+            "llm": {
+                "provider": "openai",
+                "model": profile.model,
+                "api_key": generation_key,
+                "base_url": generation_endpoint,
+            },
+            "embedder": {
+                "provider": "openai",
+                "model": (
+                    BUILTIN_EMBEDDING_MODEL if embedding_service_is_local else embedding_model
+                ),
+                "api_key": embedding_key,
+                "base_url": embedding_endpoint,
+                "embedding_dims": dimensions,
+            },
+            "vector_store": {
+                "provider": "chroma",
+                "collection_name": collection,
+                "persist_directory": str(root / "vector"),
+                "embedding_dims": dimensions,
+                "host": None,
+                "port": None,
+                "api_key": None,
+            },
+            "graph_store": {
+                "provider": "kuzu",
+                "db_path": str(root / "graph"),
+                "url": None,
+                "username": None,
+                "password": None,
+                "database": None,
+            },
+            "cache": {
+                "backend": "sqlite",
+                "db_path": str(root / "cache.sqlite3"),
+            },
+            "history": {"db_path": str(root / "history.sqlite3")},
+            "coding": {"db_path": str(root / "coding.sqlite3")},
+        }
+    )
+
+
+def acquire_client(
+    profile: ModelProfileSnapshot,
+    *,
+    api_key: str | None,
+    data_dir: Path,
+    http_transport_factory: Callable[[], Any] | None = None,
+) -> ClientLease:
+    """Construct a public SDK client and return its explicit runtime lease."""
+    config = build_sdk_config(profile, api_key=api_key, data_dir=data_dir)
+    config.http_transport_factory = http_transport_factory
+    mismatch = inspect_graph_width_mismatch(config)
+    if mismatch is not None:
+        raise GraphWidthMismatchError(mismatch)
+    key = data_dir.resolve()
+    runtime = _acquire_runtime(config, key)
+    embedding_service: LocalOnnxEmbeddingService | None = None
+    if (profile.embedding_provider_kind or profile.provider_kind) is ProviderKind.LOCAL_ONNX:
+        embedding_service = LocalOnnxEmbeddingService()
+    try:
+        client = HyMemoryClient(
+            config=config,
+            mode="ultra" if graph_store_enabled() else "pro",
+            runtime=runtime,
+            embedding_service=embedding_service,
+        )
+    except BaseException:
+        _release_runtime(key)
+        raise
+    return ClientLease(client, key)
+
+
+def close_all_shared_runtimes() -> None:
+    """Close process runtimes only after every client lease has been released."""
+    with _RUNTIME_LOCK:
+        if _RUNTIMES:
+            raise HySdkRuntimeError("hy_sdk_runtime_leases_active")
+
+
+def _generation_access(
+    profile: ModelProfileSnapshot,
+    api_key: str | None,
+) -> tuple[str, str]:
+    if profile.provider_kind is ProviderKind.LOCAL_ONNX:
+        return _LOCAL_ONNX_PLACEHOLDER_URL, "local-read-only"
+    if profile.provider_kind not in {
+        ProviderKind.OPENAI_COMPATIBLE,
+        ProviderKind.OLLAMA,
+    }:
+        raise HySdkRuntimeError("hy_sdk_generation_provider_unsupported")
+    endpoint = _openai_endpoint(profile.provider_kind, profile.base_url)
+    if profile.provider_kind is ProviderKind.OLLAMA:
+        return endpoint, "ollama"
+    if not api_key:
+        raise HySdkRuntimeError("hy_sdk_model_api_key_missing")
+    return endpoint, api_key
+
+
+def _collection_base(
+    provider: ProviderKind,
+    endpoint: str | None,
+    model: str,
+    dimensions: int | None,
+) -> str:
+    identity = "\0".join((provider.value, endpoint or "local", model, str(dimensions or "auto")))
+    return "memhub_" + hashlib.sha256(identity.encode()).hexdigest()[:16]
+
+
+def vector_collection_bases(profile: ModelProfileSnapshot) -> tuple[str, tuple[str, ...]]:
+    """The SDK vector collection base for this embedding space, and the older ones.
+
+    The SDK derives every vector collection (chat, Coding keys, Working facts,
+    each variant) by suffixing this base. Builds before #145 named the same
+    embedding space `one_memory_<12 hex>`, with or without a declared width;
+    the offline migration renames those so their vectors stay readable.
+    """
+    provider = profile.embedding_provider_kind or profile.provider_kind
+    local = provider is ProviderKind.LOCAL_ONNX
+    model = profile.embedding_model or (BUILTIN_EMBEDDING_MODEL if local else None)
+    if not model:
+        raise HySdkRuntimeError("hy_sdk_embedding_model_required")
+    endpoint: str | None
+    dimensions: int | None
+    if local:
+        endpoint, dimensions = None, profile.embedding_dimensions or BUILTIN_EMBEDDING_DIMENSIONS
+        legacy_endpoint = "local-onnx"
+    else:
+        endpoint = _openai_endpoint(provider, profile.embedding_base_url or profile.base_url)
+        dimensions = profile.embedding_dimensions
+        legacy_endpoint = endpoint
+    legacy_model = profile.embedding_model or "bge-small-zh-v1.5"
+    legacy = tuple(
+        "one_memory_"
+        + hashlib.sha256(
+            ":".join((provider.value, legacy_endpoint, legacy_model, *declared)).encode()
+        ).hexdigest()[:12]
+        for declared in ((), ("auto",), (str(profile.embedding_dimensions or "auto"),))
+    )
+    return _collection_base(provider, endpoint, model, dimensions), tuple(dict.fromkeys(legacy))
+
+
+def _openai_endpoint(provider: ProviderKind, endpoint: str | None) -> str:
+    value = (endpoint or "").rstrip("/")
+    if not value:
+        raise HySdkRuntimeError("hy_sdk_model_endpoint_missing")
+    if provider is ProviderKind.OLLAMA and not value.endswith("/v1"):
+        return f"{value}/v1"
+    return value
+
+
+def _acquire_runtime(config: MemoryConfig, key: Path) -> SharedRuntime:
+    with _RUNTIME_LOCK:
+        entry = _RUNTIMES.get(key)
+        if entry is None:
+            runtime = _run_sync(SharedRuntime.create(config))
+            _RUNTIMES[key] = _RuntimeEntry(runtime=runtime, references=1)
+            return runtime
+        entry.references += 1
+        return entry.runtime
+
+
+def _release_runtime(key: Path) -> None:
+    runtime: SharedRuntime | None = None
+    with _RUNTIME_LOCK:
+        entry = _RUNTIMES.get(key)
+        if entry is None:
+            return
+        entry.references -= 1
+        if entry.references == 0:
+            runtime = entry.runtime
+            del _RUNTIMES[key]
+    if runtime is not None:
+        runtime.close()
+
+
+def _run_sync[T](coroutine: Coroutine[Any, Any, T]) -> T:
+    """Run SDK async construction from either sync or async caller threads."""
+    try:
+        asyncio.get_running_loop()
+    except RuntimeError:
+        return asyncio.run(coroutine)
+
+    result: list[T] = []
+    failure: list[BaseException] = []
+
+    def run() -> None:
+        try:
+            result.append(asyncio.run(coroutine))
+        except BaseException as error:
+            failure.append(error)
+
+    thread = threading.Thread(target=run, name="hy-sdk-runtime-create")
+    thread.start()
+    thread.join()
+    if failure:
+        raise failure[0]
+    return result[0]
@@ -0,0 +1,694 @@
+"""Validated records and response parsers for public HY SDK operations."""
+
+from __future__ import annotations
+
+from collections.abc import Mapping
+from datetime import UTC, datetime
+from typing import Literal, NamedTuple, cast
+
+from pydantic import BaseModel, ConfigDict, Field
+
+from one_memory.semantic_deadline import SemanticMemoryHit
+
+
+class HySdkError(RuntimeError):
+    """Base error carrying a stable, product-facing code."""
+
+    def __init__(self, code: str) -> None:
+        super().__init__(code)
+        self.code = code
+
+
+class HySdkRuntimeError(HySdkError):
+    """The public SDK runtime could not be configured or leased."""
+
+
+class HySdkMigrationError(HySdkError):
+    """The privileged offline SDK data-plane migration failed."""
+
+
+class HySdkReadError(HySdkError):
+    """A public SDK read operation failed or returned an invalid shape."""
+
+
+MemoryScene = Literal["chat", "working", "coding"]
+MemoryKindName = Literal["chat", "experience", "fact"]
+MemoryVariantName = Literal["working", "coding"]
+
+
+class HySdkMemoryRecord(BaseModel):
+    """One memory of any scene, in the unified shape the library lists.
+
+    Chat memories keep every field they always had. Working/coding memories
+    (experiences and facts) carry ``scene``/``kind``/``variant`` so a client
+    can open, edit and delete them again without guessing the namespace; an
+    experience also keeps ``task`` and ``solution`` apart, and ``content`` is
+    the readable ``task\nsolution`` text the SDK lists.
+    """
+
+    model_config = ConfigDict(frozen=True, extra="forbid")
+
+    memory_id: str
+    content: str
+    layer: str
+    status: str
+    is_latest: bool
+    supersedes: tuple[str, ...] = ()
+    superseded_by: tuple[str, ...] = ()
+    tags: tuple[str, ...] = ()
+    owner: str | None = None
+    source_raw_ids: tuple[str, ...] = ()
+    user_id: str
+    agent_id: str
+    session_id: str | None = None
+    source: str | None = None
+    memory_at: int | None = None
+    gmt_created: int | None = None
+    scene: MemoryScene = "chat"
+    kind: MemoryKindName = "chat"
+    # The productivity namespace ("working" | "coding"); None for chat. Pass it
+    # back with `kind` on get/update/delete.
+    variant: MemoryVariantName | None = None
+    task: str | None = None
+    solution: str | None = None
+    # Epoch seconds. `updated_at` falls back to `created_at`.
+    created_at: int | None = None
+    updated_at: int | None = None
+
+    @classmethod
+    def from_sdk(cls, raw: object) -> HySdkMemoryRecord:
+        if not isinstance(raw, Mapping):
+            raise HySdkReadError("hy_memory_record_shape_unknown")
+        if raw.get("kind") in {"experience", "fact"}:
+            return cls._from_productivity(raw)
+        if any(
+            not isinstance(raw.get(key), str) or not str(raw[key]).strip()
+            for key in ("memory_id", "content")
+        ):
+            raise HySdkReadError("hy_memory_record_shape_unknown")
+        try:
+            gmt_created = (
+                raw.get("gmt_created") if isinstance(raw.get("gmt_created"), int) else None
+            )
+            created_at = _epoch(raw.get("created_at"))
+            if created_at is None:
+                created_at = gmt_created
+            updated_at = _epoch(raw.get("updated_at"))
+            return cls(
+                memory_id=str(raw["memory_id"]),
+                content=str(raw["content"]),
+                layer=str(raw.get("layer") or ""),
+                status=str(raw.get("status") or ""),
+                is_latest=bool(raw.get("is_latest", True)),
+                supersedes=_string_tuple(raw.get("supersedes")),
+                superseded_by=_string_tuple(raw.get("superseded_by")),
+                tags=_string_tuple(raw.get("tags")),
+                owner=(str(raw["owner"]) if raw.get("owner") in {"user", "agent"} else None),
+                source_raw_ids=_source_raw_ids(raw),
+                user_id=str(raw.get("user_id") or ""),
+                agent_id=str(raw.get("agent_id") or ""),
+                session_id=(
+                    str(raw["session_id"]) if isinstance(raw.get("session_id"), str) else None
+                ),
+                source=str(raw["source"]) if isinstance(raw.get("source"), str) else None,
+                memory_at=raw.get("memory_at") if isinstance(raw.get("memory_at"), int) else None,
+                gmt_created=gmt_created,
+                created_at=created_at,
+                updated_at=updated_at if updated_at is not None else created_at,
+            )
+        except (KeyError, TypeError, ValueError) as error:
+            raise HySdkReadError("hy_memory_record_shape_unknown") from error
+
+    @classmethod
+    def _from_productivity(cls, raw: Mapping[str, object]) -> HySdkMemoryRecord:
+        """An experience or fact from list/search(scene="all") or get().
+
+        get() names the namespace in ``variant`` (its ``scene`` is the legacy
+        "productivity"); list/search name it in ``scene``.
+        """
+        kind = str(raw["kind"])
+        variant = next(
+            (
+                value
+                for value in (raw.get("variant"), raw.get("scene"))
+                if value in {"working", "coding"}
+            ),
+            None,
+        )
+        id_key = "experience_id" if kind == "experience" else "fact_id"
+        memory_id = raw.get("memory_id") or raw.get(id_key)
+        if variant is None or not isinstance(memory_id, str) or not memory_id.strip():
+            raise HySdkReadError("hy_memory_record_shape_unknown")
+        task: str | None = None
+        solution: str | None = None
+        if kind == "experience":
+            task = str(raw.get("task") or "").strip()
+            solution = str(raw.get("solution") or "").strip()
+            content = "\n".join(part for part in (task, solution) if part)
+        else:
+            statement = raw.get("statement")
+            content = str(statement if isinstance(statement, str) else raw.get("content") or "")
+        if not content.strip():
+            raise HySdkReadError("hy_memory_record_shape_unknown")
+        status = str(raw.get("status") or "active")
+        created_at = _epoch(raw.get("created_at"))
+        updated_at = _epoch(raw.get("updated_at"))
+        try:
+            return cls(
+                memory_id=memory_id,
+                content=content,
+                layer=kind,
+                status=status,
+                is_latest=status == "active",
+                tags=_string_tuple(raw.get("tags")),
+                user_id=str(raw.get("user_id") or ""),
+                agent_id=str(raw.get("agent_id") or ""),
+                session_id=(
+                    str(raw["session_id"]) if isinstance(raw.get("session_id"), str) else None
+                ),
+                source=str(raw["source"]) if isinstance(raw.get("source"), str) else None,
+                gmt_created=created_at,
+                scene=cast(MemoryScene, variant),
+                kind=cast(MemoryKindName, kind),
+                variant=cast(MemoryVariantName, variant),
+                task=task,
+                solution=solution,
+                created_at=created_at,
+                updated_at=updated_at if updated_at is not None else created_at,
+            )
+        except (KeyError, TypeError, ValueError) as error:
+            raise HySdkReadError("hy_memory_record_shape_unknown") from error
+
+
+class HySdkMemoryPage(BaseModel):
+    model_config = ConfigDict(frozen=True, extra="forbid")
+
+    memories: tuple[HySdkMemoryRecord, ...]
+    total: int = Field(ge=0)
+    limit: int = Field(ge=1)
+    offset: int = Field(ge=0)
+    # Set on a search whose relevance window was full: matches were looked for
+    # only among this many most relevant memories, so `total` may undercount.
+    search_window: int | None = None
+    # Keyword search only: the server-side result window this page was cut
+    # from. Pass it back with the next offset to page the same results.
+    window_id: str | None = None
+    # True when the window_id the client sent had expired or did not match the
+    # query: the search ran again and this is page 1 of the new window.
+    window_reset: bool = False
+    # Empty-query listing only: how many memories each scene holds under the
+    # same filters (`total` is their sum).
+    totals: dict[str, int] | None = None
+
+    @classmethod
+    def from_sdk(cls, raw: object, *, limit: int, offset: int) -> HySdkMemoryPage:
+        page, _skipped = cls._parse(raw, limit=limit, offset=offset, skip_invalid=False)
+        return page
+
+    @classmethod
+    def from_sdk_skipping_invalid(
+        cls, raw: object, *, limit: int, offset: int
+    ) -> tuple[HySdkMemoryPage, int]:
+        """Parse a page, leaving out rows outside the public record contract.
+
+        Returns the page and how many rows were left out, so the caller can
+        report them. The rows themselves are never touched.
+        """
+        return cls._parse(raw, limit=limit, offset=offset, skip_invalid=True)
+
+    @classmethod
+    def _parse(
+        cls,
+        raw: object,
+        *,
+        limit: int,
+        offset: int,
+        skip_invalid: bool,
+    ) -> tuple[HySdkMemoryPage, int]:
+        if not isinstance(raw, Mapping):
+            raise HySdkReadError("hy_memory_list_shape_unknown")
+        if raw.get("scene") == "all":
+            return cls._parse_all_scenes(raw, limit=limit, offset=offset, skip_invalid=skip_invalid)
+        rows: list[object] = []
+        total = 0
+        for bucket_name, rows_name in (("vdb", "memories"), ("graph", "nodes")):
+            bucket = raw.get(bucket_name)
+            if bucket is None and bucket_name == "graph":
+                continue
+            if not isinstance(bucket, Mapping) or not isinstance(bucket.get(rows_name), list):
+                raise HySdkReadError("hy_memory_list_shape_unknown")
+            bucket_total = bucket.get("total")
+            if (
+                not isinstance(bucket_total, int)
+                or isinstance(bucket_total, bool)
+                or bucket_total < 0
+            ):
+                raise HySdkReadError("hy_memory_list_shape_unknown")
+            total += bucket_total
+            rows.extend(bucket[rows_name])
+        deduplicated: dict[str, HySdkMemoryRecord] = {}
+        skipped = 0
+        for row in rows:
+            try:
+                record = HySdkMemoryRecord.from_sdk(row)
+            except HySdkReadError:
+                if not skip_invalid:
+                    raise
+                skipped += 1
+                continue
+            deduplicated.setdefault(record.memory_id, record)
+        page = cls(
+            memories=tuple(deduplicated.values()),
+            # `total` is the SDK's stable result-set count. Subtracting only
+            # invalid rows observed on this page would make it change as the
+            # caller paged. Until the SDK exposes a quarantined count, keep the
+            # authoritative total and skip malformed rows only in this page.
+            total=total,
+            limit=limit,
+            offset=offset,
+        )
+        return page, skipped
+
+    @classmethod
+    def _parse_all_scenes(
+        cls,
+        raw: Mapping[str, object],
+        *,
+        limit: int,
+        offset: int,
+        skip_invalid: bool,
+    ) -> tuple[HySdkMemoryPage, int]:
+        """``list_memories(scene="all")``: chat, working and coding, newest first."""
+        rows = raw.get("memories")
+        total = raw.get("total")
+        if (
+            not isinstance(rows, list)
+            or not isinstance(total, int)
+            or isinstance(total, bool)
+            or total < 0
+        ):
+            raise HySdkReadError("hy_memory_list_shape_unknown")
+        totals_raw = raw.get("totals")
+        totals = (
+            {
+                str(scene): count
+                for scene, count in totals_raw.items()
+                if isinstance(count, int) and not isinstance(count, bool) and count >= 0
+            }
+            if isinstance(totals_raw, Mapping)
+            else None
+        )
+        memories: list[HySdkMemoryRecord] = []
+        seen: set[tuple[str, str, str | None]] = set()
+        skipped = 0
+        for row in rows:
+            try:
+                record = HySdkMemoryRecord.from_sdk(row)
+            except HySdkReadError:
+                if not skip_invalid:
+                    raise
+                skipped += 1
+                continue
+            key = (record.memory_id, record.kind, record.variant)
+            if key in seen:
+                continue
+            seen.add(key)
+            memories.append(record)
+        return (
+            cls(
+                memories=tuple(memories),
+                total=total,
+                limit=limit,
+                offset=offset,
+                totals=totals,
+            ),
+            skipped,
+        )
+
+
+class HySdkRawRecord(BaseModel):
+    model_config = ConfigDict(frozen=True, extra="forbid")
+
+    raw_id: str
+    content: str
+    user_id: str
+    agent_id: str
+    session_id: str
+    source: str
+    memory_at: int | None = None
+    gmt_created: int | None = None
+    custom: dict[str, object] = Field(default_factory=dict)
+
+    @classmethod
+    def from_sdk(cls, raw: object) -> HySdkRawRecord:
+        if not isinstance(raw, Mapping):
+            raise HySdkReadError("hy_memory_raw_shape_unknown")
+        if any(
+            not isinstance(raw.get(key), str) or not str(raw[key]).strip()
+            for key in ("raw_id", "content", "user_id", "agent_id", "session_id", "source")
+        ):
+            raise HySdkReadError("hy_memory_raw_shape_unknown")
+        custom = raw.get("custom")
+        if not isinstance(custom, Mapping):
+            custom = {}
+        try:
+            return cls(
+                raw_id=str(raw["raw_id"]),
+                content=str(raw["content"]),
+                user_id=str(raw["user_id"]),
+                agent_id=str(raw["agent_id"]),
+                session_id=str(raw["session_id"]),
+                source=str(raw["source"]),
+                memory_at=raw.get("memory_at") if isinstance(raw.get("memory_at"), int) else None,
+                gmt_created=(
+                    raw.get("gmt_created") if isinstance(raw.get("gmt_created"), int) else None
+                ),
+                custom=dict(custom),
+            )
+        except (KeyError, TypeError, ValueError) as error:
+            raise HySdkReadError("hy_memory_raw_shape_unknown") from error
+
+
+class HySdkRawHit(HySdkRawRecord):
+    score: float = Field(ge=0, le=1)
+
+    @classmethod
+    def from_sdk_search(cls, raw: object) -> HySdkRawHit:
+        if not isinstance(raw, Mapping):
+            raise HySdkReadError("hy_memory_raw_search_shape_unknown")
+        raw_id = raw.get("memory_id")
+        score = raw.get("score")
+        if (
+            not isinstance(raw_id, str)
+            or not isinstance(score, int | float)
+            or isinstance(score, bool)
+        ):
+            raise HySdkReadError("hy_memory_raw_search_shape_unknown")
+        base = HySdkRawRecord.from_sdk({**raw, "raw_id": raw_id})
+        return cls(
+            **base.model_dump(),
+            score=max(0.0, min(1.0, float(score))),
+        )
+
+
+class HySdkRawPage(BaseModel):
+    model_config = ConfigDict(frozen=True, extra="forbid")
+
+    raw: tuple[HySdkRawRecord, ...]
+    total: int = Field(ge=0)
+    limit: int = Field(ge=1)
+    offset: int = Field(ge=0)
+
+    @classmethod
+    def from_memory_list(
+        cls,
+        raw: object,
+        *,
+        limit: int,
+        offset: int,
+    ) -> HySdkRawPage:
+        """Parse ``list_memories(layers=['l1_raw'])`` into the App raw DTO."""
+        if not isinstance(raw, Mapping):
+            raise HySdkReadError("hy_memory_raw_list_shape_unknown")
+        bucket = raw.get("vdb")
+        if not isinstance(bucket, Mapping):
+            raise HySdkReadError("hy_memory_raw_list_shape_unknown")
+        rows = bucket.get("memories")
+        total = bucket.get("total")
+        if (
+            not isinstance(rows, list)
+            or not isinstance(total, int)
+            or isinstance(total, bool)
+            or total < 0
+        ):
+            raise HySdkReadError("hy_memory_raw_list_shape_unknown")
+        converted = []
+        for row in rows:
+            if not isinstance(row, Mapping) or row.get("layer") != "l1_raw":
+                raise HySdkReadError("hy_memory_raw_list_shape_unknown")
+            converted.append({**row, "raw_id": row.get("memory_id")})
+        return cls(
+            raw=tuple(HySdkRawRecord.from_sdk(row) for row in converted),
+            total=total,
+            limit=limit,
+            offset=offset,
+        )
+
+
+class HySdkUserProfile(BaseModel):
+    model_config = ConfigDict(frozen=True, extra="forbid")
+
+    request_id: str
+    profile: dict[str, object] | None
+    elapsed_ms: float = Field(ge=0)
+
+    @classmethod
+    def from_sdk(cls, raw: object) -> HySdkUserProfile:
+        if not isinstance(raw, Mapping):
+            raise HySdkReadError("hy_memory_profile_shape_unknown")
+        request_id = raw.get("request_id")
+        profile = raw.get("profile")
+        elapsed_ms = raw.get("elapsed_ms")
+        if (
+            not isinstance(request_id, str)
+            or (profile is not None and not isinstance(profile, Mapping))
+            or not isinstance(elapsed_ms, int | float)
+            or isinstance(elapsed_ms, bool)
+        ):
+            raise HySdkReadError("hy_memory_profile_shape_unknown")
+        return cls(
+            request_id=request_id,
+            profile=None if profile is None else dict(profile),
+            elapsed_ms=float(elapsed_ms),
+        )
+
+
+def parse_sdk_search(raw: object, *, error_code: str) -> tuple[SemanticMemoryHit, ...]:
+    rows = _reply_rows(raw, error_code)
+    found: list[SemanticMemoryHit] = []
+    seen: set[str] = set()
+    for item in sorted(rows, key=_score, reverse=True):
+        memory_id = item.get("memory_id") or item.get("id")
+        if not isinstance(memory_id, str) or not memory_id or memory_id in seen:
+            continue
+        seen.add(memory_id)
+        content = item.get("content")
+        member_ids, formatting_complete = _formatted_member_ids(item)
+        found.append(
+            SemanticMemoryHit(
+                memory_id=memory_id,
+                score=_score(item),
+                formatted_content=(
+                    content.strip() if isinstance(content, str) and content.strip() else None
+                ),
+                member_ids=member_ids,
+                sdk_formatted=True,
+                formatting_complete=formatting_complete,
+            )
+        )
+    return tuple(found)
+
+
+class LibraryHit(NamedTuple):
+    """One keyword match in a library search window, with its namespace."""
+
+    memory_id: str
+    kind: MemoryKindName
+    variant: MemoryVariantName | None
+
+
+def library_keyword_hits(raw: object, query: str) -> tuple[tuple[LibraryHit, ...], int]:
+    """Memories in an SDK search window whose text contains every query term.
+
+    The local SDK store has no keyword channel (Chroma's keyword search is
+    empty), so search is pure vector similarity: a word in the library can
+    score below the floor, and a word in none of it still scores above. The
+    library search asks the SDK for its most relevant window and keeps only the
+    rows that actually contain the words, most relevant first. Returns those
+    hits and how many distinct rows the window held.
+
+    A ``search(scene="all")`` window mixes chat memories with working/coding
+    experiences and facts; their text is the unified ``content`` (an
+    experience's task and solution, a fact's statement). The fused order is
+    kept as the SDK ranked it.
+    """
+    terms = query.casefold().split()
+    if isinstance(raw, Mapping) and raw.get("scene") == "all":
+        rows = raw.get("memories")
+        if not isinstance(rows, list) or any(not isinstance(row, Mapping) for row in rows):
+            raise HySdkReadError("hy_memory_search_shape_unknown")
+        ordered: list[Mapping[str, object]] = list(rows)
+    else:
+        ordered = sorted(
+            _reply_rows(raw, "hy_memory_search_shape_unknown"), key=_score, reverse=True
+        )
+    seen: set[LibraryHit] = set()
+    matched: list[LibraryHit] = []
+    for item in ordered:
+        hit = _library_hit(item)
+        if hit is None or hit in seen:
+            continue
+        seen.add(hit)
+        tags = item.get("tags")
+        parts = [item.get(key) for key in ("content", "task", "solution", "statement")] + list(
+            tags if isinstance(tags, list) else ()
+        )
+        text = " ".join(part for part in parts if isinstance(part, str)).casefold()
+        if all(term in text for term in terms):
+            matched.append(hit)
+    return tuple(matched), len(seen)
+
+
+def _library_hit(item: Mapping[str, object]) -> LibraryHit | None:
+    kind = item.get("kind")
+    if kind in {"experience", "fact"}:
+        variant = item.get("variant") if item.get("variant") in {"working", "coding"} else None
+        variant = variant or (
+            item.get("scene") if item.get("scene") in {"working", "coding"} else None
+        )
+        memory_id = item.get("memory_id") or item.get(
+            "experience_id" if kind == "experience" else "fact_id"
+        )
+        if variant is None or not isinstance(memory_id, str) or not memory_id:
+            return None
+        return LibraryHit(
+            memory_id,
+            "experience" if kind == "experience" else "fact",
+            "coding" if variant == "coding" else "working",
+        )
+    memory_id = item.get("memory_id") or item.get("id")
+    if not isinstance(memory_id, str) or not memory_id:
+        return None
+    return LibraryHit(memory_id, "chat", None)
+
+
+def parse_sdk_raw_search(raw: object) -> tuple[HySdkRawHit, ...]:
+    return tuple(
+        HySdkRawHit.from_sdk_search(row)
+        for row in _reply_rows(raw, "hy_memory_raw_search_shape_unknown")
+    )
+
+
+def _epoch(value: object) -> int | None:
+    """Epoch seconds from an int or an ISO-8601 string (naive means UTC)."""
+    if isinstance(value, bool):
+        return None
+    if isinstance(value, int):
+        return value
+    if isinstance(value, float):
+        return int(value)
+    if isinstance(value, str) and value.strip():
+        try:
+            parsed = datetime.fromisoformat(value.strip())
+        except ValueError:
+            return None
+        if parsed.tzinfo is None:
+            parsed = parsed.replace(tzinfo=UTC)
+        return int(parsed.timestamp())
+    return None
+
+
+def _string_tuple(value: object) -> tuple[str, ...]:
+    if not isinstance(value, list):
+        return ()
+    return tuple(
+        dict.fromkeys(item.strip() for item in value if isinstance(item, str) and item.strip())
+    )
+
+
+def _source_raw_ids(raw: Mapping[str, object]) -> tuple[str, ...]:
+    custom = raw.get("custom")
+    values: list[object] = []
+    if isinstance(custom, Mapping):
+        listed = custom.get("source_raw_memory_ids")
+        if isinstance(listed, list):
+            values.extend(listed)
+    singular = raw.get("source_raw_memory_id")
+    if isinstance(singular, str):
+        values.append(singular)
+    return tuple(
+        dict.fromkeys(value.strip() for value in values if isinstance(value, str) and value.strip())
+    )
+
+
+def _reply_rows(raw: object, error_code: str) -> tuple[Mapping[str, object], ...]:
+    if not isinstance(raw, Mapping):
+        raise HySdkReadError(error_code)
+    keys = tuple(key for key in ("memories", "coding_memories") if key in raw)
+    if "scene" in raw:
+        scene = raw["scene"]
+        if not isinstance(scene, str):
+            raise HySdkReadError(error_code)
+        if scene in ("normal", "chat"):
+            selected = "memories"
+        elif scene in ("productivity", "working"):
+            selected = "coding_memories"
+        else:
+            raise HySdkReadError(error_code)
+        if selected not in keys:
+            raise HySdkReadError(error_code)
+    elif len(keys) == 1:
+        selected = keys[0]
+    else:
+        raise HySdkReadError(error_code)
+
+    result: list[Mapping[str, object]] = []
+    for key in keys:
+        value = raw[key]
+        if key == "memories" and isinstance(value, Mapping):
+            if set(value) != {"schema", "proactive", "normal"}:
+                raise HySdkReadError(error_code)
+            channels = tuple(value.values())
+        else:
+            channels = (value,)
+        for channel in channels:
+            if not isinstance(channel, list) or any(
+                not isinstance(row, Mapping) for row in channel
+            ):
+                raise HySdkReadError(error_code)
+            if key == selected:
+                result.extend(channel)
+    return tuple(result)
+
+
+def _score(memory: Mapping[str, object]) -> float:
+    score = memory.get("score")
+    return float(score) if isinstance(score, int | float) and not isinstance(score, bool) else 0.0
+
+
+def _formatted_member_ids(memory: Mapping[str, object]) -> tuple[tuple[str, ...], bool]:
+    represented: dict[str, None] = {}
+    composite = False
+    raw_members = memory.get("member_ids")
+    if raw_members is not None:
+        composite = True
+        if not isinstance(raw_members, list | tuple) or any(
+            not isinstance(member_id, str) or not member_id for member_id in raw_members
+        ):
+            return (), False
+        for member_id in raw_members:
+            represented.setdefault(member_id, None)
+    chain = memory.get("evolution_chain")
+    if chain is not None:
+        composite = True
+        if not isinstance(chain, list | tuple):
+            return (), False
+        for entry in chain:
+            if not isinstance(entry, Mapping):
+                return (), False
+            member_id = entry.get("node_id") or entry.get("memory_id")
+            if not isinstance(member_id, str) or not member_id:
+                return (), False
+            represented.setdefault(member_id, None)
+    if memory.get("capsule_memories") is not None or memory.get("timeline") is not None:
+        composite = True
+        if raw_members is None:
+            return (), False
+    content = memory.get("content")
+    if isinstance(content, str) and content.lstrip().startswith(
+        ("Related memories:", "Matched memories:", "Memory timeline:")
+    ):
+        composite = True
+        if not represented:
+            return (), False
+    return tuple(represented), not composite or bool(represented)
@@ -1,153 +0,0 @@
-"""Per-write original Evidence and where each HY memory came from; never a memory source."""
-
-import json
-from collections.abc import Set as AbstractSet
-from importlib import import_module
-from typing import Any, Literal
-
-from pydantic import BaseModel, ConfigDict
-
-from one_memory.imports.models import EvidenceContentKind, EvidenceRecord
-from one_memory.input_sanitization import sanitize_model_input
-
-
-class HySupportedMemory(BaseModel):
-    model_config = ConfigDict(frozen=True, extra="forbid")
-
-    hy_memory_id: str
-    governed_memory_id: str
-    revision: int
-    content: str
-    owner: Literal["user", "agent"]
-    evidence_ids: tuple[str, ...]
-
-
-class HySourceContext(BaseModel):
-    model_config = ConfigDict(frozen=True, extra="forbid")
-
-    current: tuple[EvidenceRecord, ...]
-    history: tuple[EvidenceRecord, ...] = ()
-    supporting: tuple[EvidenceRecord, ...] = ()
-    known_memories: tuple[HySupportedMemory, ...] = ()
-
-
-class HyWriteBindings:
-    def __init__(self, context: HySourceContext) -> None:
-        if not context.current:
-            raise ValueError("hy_memory_write_verification_context_scope")
-        self.context = context
-        seed = context.current[0]
-        records = (*context.current, *context.history, *context.supporting)
-        workspaces = {record.workspace for record in records}
-        # A missing workspace is a legacy identity, not a wildcard that may
-        # absorb a named project. Keep NULL and named workspaces disjoint.
-        if None in workspaces and len(workspaces) != 1:
-            raise ValueError("hy_memory_write_verification_context_scope")
-        ids = [record.evidence_id for record in records]
-        if len(ids) != len(set(ids)):
-            raise ValueError("hy_memory_write_verification_context_scope")
-        for record in records:
-            if (
-                (
-                    record.user_id,
-                    record.source_id,
-                    record.conversation_id,
-                    record.workspace,
-                )
-                != (
-                    seed.user_id,
-                    seed.source_id,
-                    seed.conversation_id,
-                    seed.workspace,
-                )
-                or record.content_kind is not EvidenceContentKind.RAW_CONVERSATION
-                or record.role not in {"user", "assistant", "tool"}
-            ):
-                raise ValueError("hy_memory_write_verification_context_scope")
-        self.all_sources: list[dict[str, Any]] = [
-            {
-                "id": record.evidence_id,
-                "role": record.role,
-                "content": sanitize_model_input(record.verbatim),
-                "new": index < len(context.current),
-                "revision": record.revision,
-                "created_at": record.created_at.isoformat() if record.created_at else None,
-            }
-            for index, record in enumerate(records)
-        ]
-        self.sources = self.all_sources[: len(context.current) + len(context.history)]
-        self._supports: dict[tuple[str, str], tuple[str, ...]] = {}
-        self.profile_order: list[str] | None = None
-        self.verified_operations: dict[str, tuple[str, str]] = {}
-        self.known = {memory.hy_memory_id: memory for memory in context.known_memories}
-        for memory in context.known_memories:
-            if not memory.evidence_ids or not set(memory.evidence_ids) <= set(ids):
-                raise ValueError("hy_memory_write_verification_context_scope")
-            self._supports[(memory.content, memory.owner)] = memory.evidence_ids
-
-    def check_content(self, content: str) -> None:
-        try:
-            parsed = json.loads(content)
-        except ValueError as error:
-            raise ValueError("hy_memory_write_verification_input_mismatch") from error
-        expected = [
-            {"role": record.role, "content": sanitize_model_input(record.verbatim)}
-            for record in self.context.current
-        ]
-        # HY strips tool messages itself before chat extraction when it judges a turn chat.
-        without_tools = [item for item in expected if item["role"] != "tool"]
-        if parsed != expected and (without_tools == expected or parsed != without_tools):
-            raise ValueError("hy_memory_write_verification_input_mismatch")
-
-    def history_json(self) -> str:
-        return json.dumps(
-            [
-                {
-                    "evidence_id": source["id"],
-                    **{k: source[k] for k in ("role", "content", "revision", "created_at")},
-                }
-                for source in self.sources
-                if not source["new"]
-            ],
-            ensure_ascii=False,
-        )
-
-    def turn_support(
-        self, owner: str | None, targets: AbstractSet[str] = frozenset()
-    ) -> tuple[str, ...]:
-        """This turn's words by the memory's author, plus what a merged memory already cited."""
-        roles = {"user": {"user"}, "agent": {"assistant"}}.get(owner or "")
-        current = [source for source in self.all_sources if source["new"]]
-        ids = [source["id"] for source in current if roles is None or source["role"] in roles]
-        if not ids:
-            ids = [source["id"] for source in current]
-        for target in sorted(targets):
-            if target in self.known:
-                ids.extend(self.known[target].evidence_ids)
-        return tuple(dict.fromkeys(ids))
-
-    def record_turn_supports(self, items: list[dict[str, Any]]) -> None:
-        for item in items:
-            key = (item["content"], item["owner"] or "")
-            self._supports.setdefault(key, self.turn_support(item["owner"]))
-
-    def record_operation_support(self, content: str, owner: str | None, targets: set[str]) -> None:
-        key = (content, owner or "")
-        merged = (*self._supports.get(key, ()), *self.turn_support(owner, targets))
-        self._supports[key] = tuple(dict.fromkeys(merged))
-
-    def support_for(self, content: str, owner: str | None) -> tuple[str, ...]:
-        return self._supports.get((content, owner or ""), ())
-
-    def profile_support(self, content: str, values: dict[str, Any]) -> tuple[str, ...]:
-        render = import_module("hy_memory.agent.tools.basic_profile")._render_content
-
-        if not values or render(values, field_order=self.profile_order) != content:
-            return ()
-        ids: dict[str, None] = {}
-        for key, value in values.items():
-            support = self.support_for(f"{key}: {value}", "user")
-            if not support:
-                return ()
-            ids.update(dict.fromkeys(support))
-        return tuple(ids)
@@ -1,53 +0,0 @@
-"""Input checks around HY's native extractor; HY's output is published as it returns it."""
-
-import asyncio
-from typing import Any
-
-from one_memory.hy_source_context import HyWriteBindings
-
-
-def _extracted_items(info: dict[str, Any]) -> list[dict[str, Any]]:
-    items: list[dict[str, Any]] = []
-    for section in ("memory", "intentions"):
-        for item in info.get(section) or []:
-            if isinstance(item, dict) and isinstance(item.get("content"), str):
-                items.append({"content": item["content"], "owner": item.get("owner")})
-    for name, value in (info.get("basic_info") or {}).items():
-        items.append({"content": f"{name}: {value}", "owner": "user"})
-    return items
-
-
-class HyExtractionGuard:
-    def __init__(self, extractor: Any) -> None:
-        self.extractor = extractor
-        self.failures: list[str] = []
-        self._lock = asyncio.Lock()
-        self.bindings: HyWriteBindings | None = None
-
-    def __getattr__(self, name: str) -> Any:
-        return getattr(self.extractor, name)
-
-    async def extract(self, *args: Any, **kwargs: Any) -> Any:
-        async with self._lock:
-            try:
-                return await self._extract(args, kwargs)
-            except ValueError as error:
-                if str(error).startswith("hy_memory_write_verification"):
-                    self.failures.append(str(error))
-                raise
-
-    async def _extract(self, args: tuple[Any, ...], kwargs: dict[str, Any]) -> Any:
-        content = kwargs.get("content", args[0] if args else "")
-        bindings = self.bindings
-        if bindings is not None:
-            bindings.check_content(content)
-            fields = kwargs.get("basic_profile_fields")
-            bindings.profile_order = list(fields) if fields else None
-            kwargs = {**kwargs, "history_context": bindings.history_json()}
-        else:
-            # No raw-history identity was supplied; do not trust HY's text prefixes.
-            kwargs = {**kwargs, "history_context": ""}
-        result = await self.extractor.extract(*args, **kwargs)
-        if bindings is not None and result.success:
-            bindings.record_turn_supports(_extracted_items(result.info))
-        return result
@@ -50,6 +50,10 @@ class AuthorizedSourceFile:
    size: int
    modified_ns: int
    changed_ns: int
+    # Source-specific change token (for example Cursor's WAL identity). It is
+    # cursor metadata only; modified_ns remains the real filesystem mtime used
+    # as a fallback conversation timestamp.
+    cursor_version: str | None = None
    root_identity: str | None = None
    requested_root: Path | None = None
    access_check: Callable[[], None] | None = field(default=None, repr=False, compare=False)
@@ -557,12 +557,10 @@ class PollingLiveSyncSupervisor:
        )

    def _note_internal_error(self, source_id: str, error: Exception) -> None:
-        # Type alone used to be all this said. It is not enough to act on:
-        # `OperationalError` covers a locked database, a missing table and a
-        # corrupt page alike, and without the message and the frames two dozen
-        # of these in a log could only be guessed at. The stack is the
-        # redacted, `one_memory`-relative one, so it names the call site
-        # without printing anyone's paths or query text.
+        # The ordinary local log keeps the exception message for debugging.
+        # The structured diagnostic event deliberately keeps only the stable
+        # error code, source id, type, and relative stack: exception messages
+        # can contain paths, query text, or credentials and are exportable data.
        logging.getLogger(__name__).warning(
            "live_sync_internal_error %s: %s", type(error).__name__, error
        )
@@ -572,10 +570,7 @@ class PollingLiveSyncSupervisor:
            event="live_sync.internal_error",
            outcome="failed",
            error_code="live_sync_internal_error",
-            # `reason` is the listed key that can carry the message; see
-            # `diagnostic_log._DETAIL_KEYS`. Truncated because a detail value
-            # longer than 256 characters is rejected.
-            details={"source_id": source_id, "reason": str(error)[:200]},
+            details={"source_id": source_id},
            **diagnostic_exception(error),
        )
        if isinstance(error, sqlite3.Error):
@@ -867,10 +867,7 @@ class ImportRunner:
            outcome="failed",
            job_id=job_id,
            error_code="import_heartbeat_failed",
-            # `reason` is the listed key for this; see
-            # `diagnostic_log._DETAIL_KEYS`. Truncated: a detail value
-            # over 256 characters is rejected.
-            details={"reason": str(error)[:200]},
+            details={"phase": "heartbeat"},
            **diagnostic_exception(error),
        )

@@ -644,7 +644,7 @@ def is_unchanged(source_file: AuthorizedSourceFile, cursors: dict[str, SourceCur
    return (
        previous is not None
        and previous.read_to_end
-        and previous.file_identity == source_file.file_identity
+        and previous.file_identity == _cursor_file_identity(source_file)
        and previous.observed_size == source_file.size
        and previous.observed_modified_ns == source_file.modified_ns
        # Chmod alone leaves size, mtime and inode untouched, so a file that
@@ -691,7 +691,7 @@ def touched_cursor(
    previous = options.cursors.get(source_file.safe_id)
    keep = (
        previous is not None
-        and previous.file_identity == source_file.file_identity
+        and _same_file(previous, source_file)
        and previous.observed_size <= source_file.size
    )
    # A kept cursor means the next scan resumes exactly where this one ended:
@@ -871,7 +871,7 @@ def resume_checkpoint(
    previous = options.cursors.get(source_file.safe_id)
    if previous is None:
        return None
-    if previous.file_identity != source_file.file_identity:
+    if not _same_file(previous, source_file):
        return None
    if source_file.size < previous.observed_size:
        return None
@@ -1046,7 +1046,7 @@ def source_cursor(
 ) -> SourceCursor:
    return SourceCursor(
        file_key=source_file.safe_id,
-        file_identity=source_file.file_identity,
+        file_identity=_cursor_file_identity(source_file),
        checkpoint=checkpoint,
        checkpoint_anchor=checkpoint_anchor,
        observed_size=source_file.size,
@@ -1054,6 +1054,27 @@ def source_cursor(
    )


+def _cursor_file_identity(source_file: AuthorizedSourceFile) -> str:
+    return (
+        f"{source_file.file_identity}{_VERSION_SEPARATOR}{source_file.cursor_version}"
+        if source_file.cursor_version
+        else source_file.file_identity
+    )
+
+
+_VERSION_SEPARATOR = ":v:"
+
+
+def _same_file(previous: SourceCursor, source_file: AuthorizedSourceFile) -> bool:
+    """Whether a stored cursor still points into this file.
+
+    The stored identity may carry a change token (Cursor's WAL version). The
+    token says the file changed, which `is_unchanged` handles; it does not mean
+    the file was replaced, so resuming compares the bare `dev:ino` only.
+    """
+    return previous.file_identity.partition(_VERSION_SEPARATOR)[0] == source_file.file_identity
+
+
 _POSITION = re.compile(r"(?<=:)\d+(?=:|$)")


@@ -533,8 +533,8 @@ def _with_write_ahead_log(database: AuthorizedSourceFile) -> AuthorizedSourceFil
    version = hashlib.blake2b(
        f"{database.modified_ns}:{journal.st_mtime_ns}:{journal.st_size}:{journal.st_ino}".encode(),
        digest_size=8,
-    ).digest()
-    return replace(database, modified_ns=int.from_bytes(version, "big") & ((1 << 63) - 1))
+    ).hexdigest()
+    return replace(database, cursor_version=version)


 def _live_baseline(listing: _CursorListing) -> dict[str, SourceCursor]:
@@ -0,0 +1,35 @@
+"""Passive lifecycle projection used only by retired Governance readers."""
+
+from __future__ import annotations
+
+from collections.abc import Mapping
+from dataclasses import dataclass, field
+from types import MappingProxyType
+
+
+@dataclass(frozen=True, eq=False)
+class HyLifecycle:
+    hidden: Mapping[str, tuple[str, ...]] = field(default_factory=dict)
+
+    def __post_init__(self) -> None:
+        object.__setattr__(self, "hidden", MappingProxyType(dict(self.hidden)))
+
+    def __eq__(self, other: object) -> bool:
+        return isinstance(other, HyLifecycle) and dict(self.hidden) == dict(
+            other.hidden
+        )
+
+    def __hash__(self) -> int:
+        return hash(tuple(sorted(self.hidden.items())))
+
+    def replaced_by(self, memory_id: str) -> tuple[str, ...]:
+        return self.hidden.get(memory_id, ())
+
+    def replacing(self, memory_id: str) -> tuple[str, ...]:
+        return tuple(
+            sorted(
+                old
+                for old, replacements in self.hidden.items()
+                if memory_id in replacements
+            )
+        )
@@ -0,0 +1,110 @@
+"""Short-lived result windows for the memory library's keyword search.
+
+A library search asks the SDK once for a bounded relevance window and keeps
+the keyword matches. Paging must cut every page from that same list: running
+the search again per page can rank differently and duplicate or skip memories
+between pages. The matches are kept here under an opaque ``window_id`` that
+the client passes back with the next offset.
+
+A window is bound to the filters it was searched with (query, source,
+statuses). It expires after ``ttl_seconds`` without being read, and the store
+keeps at most ``max_windows`` windows, dropping the least recently read first.
+Only memory identities are kept, never content: each page is read fresh from
+the SDK, so an edit or a delete shows on the next page read.
+"""
+
+from __future__ import annotations
+
+import secrets
+import threading
+import time
+from collections import OrderedDict
+from collections.abc import Callable
+from dataclasses import dataclass, field
+
+from one_memory.hy_sdk_types import LibraryHit
+
+DEFAULT_WINDOW_TTL_SECONDS = 300.0
+DEFAULT_MAX_WINDOWS = 32
+
+WindowKey = tuple[str, str | None, tuple[str, ...]]
+
+
+def window_key(query: str, source: str | None, statuses: list[str]) -> WindowKey:
+    return (query, source, tuple(sorted(set(statuses))))
+
+
+@dataclass(frozen=True)
+class SearchWindow:
+    window_id: str
+    key: WindowKey
+    hits: tuple[LibraryHit, ...]
+    # Set when the SDK window was full, so matches outside it were not looked
+    # for and `total` may undercount.
+    search_window: int | None
+
+
+@dataclass
+class _Entry:
+    window: SearchWindow
+    last_read: float = field(default=0.0)
+
+
+class LibrarySearchWindows:
+    def __init__(
+        self,
+        *,
+        ttl_seconds: float = DEFAULT_WINDOW_TTL_SECONDS,
+        max_windows: int = DEFAULT_MAX_WINDOWS,
+        clock: Callable[[], float] = time.monotonic,
+    ) -> None:
+        self._ttl = ttl_seconds
+        self._max = max(1, max_windows)
+        self._clock = clock
+        self._entries: OrderedDict[str, _Entry] = OrderedDict()
+        self._lock = threading.Lock()
+
+    def get(self, window_id: str, key: WindowKey) -> SearchWindow | None:
+        """The window, if it is still live and was searched with `key`."""
+        with self._lock:
+            now = self._clock()
+            self._expire(now)
+            entry = self._entries.get(window_id)
+            if entry is None or entry.window.key != key:
+                return None
+            entry.last_read = now
+            self._entries.move_to_end(window_id)
+            return entry.window
+
+    def put(
+        self,
+        key: WindowKey,
+        hits: tuple[LibraryHit, ...],
+        search_window: int | None,
+    ) -> SearchWindow:
+        window = SearchWindow(
+            window_id=secrets.token_urlsafe(12),
+            key=key,
+            hits=hits,
+            search_window=search_window,
+        )
+        with self._lock:
+            now = self._clock()
+            self._expire(now)
+            self._entries[window.window_id] = _Entry(window=window, last_read=now)
+            while len(self._entries) > self._max:
+                self._entries.popitem(last=False)
+        return window
+
+    def __len__(self) -> int:
+        with self._lock:
+            return len(self._entries)
+
+    def _expire(self, now: float) -> None:
+        stale = [
+            window_id
+            for window_id, entry in self._entries.items()
+            if now - entry.last_read >= self._ttl
+        ]
+        for window_id in stale:
+            del self._entries[window_id]
@@ -1,6 +1,7 @@
 """Write visible hook text into Evidence. Tool output stays out."""

 import json
+import os
 import time
 from datetime import UTC, datetime
 from enum import StrEnum
@@ -74,12 +75,13 @@ _FILE_FALLBACK = {
 }
 _FALLBACK_EVENTS = frozenset({"Stop", "stop", "SubagentStop", "StopFailure"})
 _TEXT_BLOCKS = frozenset({"text"})
+SDK_MIGRATION_CAPTURE_FENCE = ".sdk-migration-capture-fence.json"


 def database_path() -> Path:
-    from one_memory.workbuddy_boundary import database_path as boundary_database
+    from one_memory.config import load_settings

-    return boundary_database()
+    return load_settings().data_dir / "one-memory.sqlite3"


 def capture_stdio(source_id: str, stdin: TextIO, output: TextIO) -> int:
@@ -88,17 +90,49 @@ def capture_stdio(source_id: str, stdin: TextIO, output: TextIO) -> int:
    del output
    if source_id not in CAPTURE_SOURCES:
        return 2
+    database = database_path()
+    if sdk_migration_capture_fenced(database.parent):
+        return 0
    try:
        payload = json.loads(stdin.read() or "")
    except json.JSONDecodeError:
        return 0
    try:
-        capture_hook(source_id, payload, database_path())
+        capture_hook(source_id, payload, database)
    except Exception:
        return 0
    return 0


+def sdk_migration_capture_fenced(data_dir: Path) -> bool:
+    """Ignore stale fences; a live migration or a failed one pauses hook writes.
+
+    A failed upgrade migration leaves the fence `frozen` (no pid): the home was
+    put back as it was before the upgrade and takes no writes until the next
+    migration attempt, which replaces the fence and lifts it on success.
+    """
+    marker = data_dir / SDK_MIGRATION_CAPTURE_FENCE
+    try:
+        document = json.loads(marker.read_text(encoding="utf-8"))
+        if isinstance(document, dict) and document.get("frozen") is True:
+            return True
+        pid = int(document["pid"])
+        if pid < 1:
+            raise ValueError("invalid migration fence pid")
+        os.kill(pid, 0)
+        return True
+    except FileNotFoundError:
+        return False
+    except ProcessLookupError:
+        marker.unlink(missing_ok=True)
+        return False
+    except PermissionError:
+        return True
+    except (KeyError, TypeError, ValueError, json.JSONDecodeError, OSError):
+        marker.unlink(missing_ok=True)
+        return False
+
+
 def capture_hook(source_id: str, payload: object, database: Path) -> bool:
    if source_id not in CAPTURE_SOURCES or not isinstance(payload, dict):
        return False
@@ -224,7 +258,7 @@ def record_capture_rejection(
 def _submit_live_memory(store: EvidenceStore, record: EvidenceRecord, *, changed: bool) -> None:
    if not changed or not _sync_open(store.database_path, record.source_id):
        return
-    from one_memory.memory_processing import _conversation_chunks
+    from one_memory.conversation_processing import _conversation_chunks
    from one_memory.memory_server.durable_queue import DurableWriteQueue
    from one_memory.memory_server.unified_entry import UnifiedWriteIntake

@@ -10,16 +10,12 @@ weights alone.
 from __future__ import annotations

 import threading
-from dataclasses import dataclass
-from hashlib import sha256
 from importlib import import_module
 from typing import TYPE_CHECKING, Any

 from one_memory.embedding_assets import (
    BUILTIN_EMBEDDING_DIMENSIONS,
    BUILTIN_EMBEDDING_MAX_TOKENS,
-    EMBEDDING_ASSETS,
-    AssetVerificationError,
    MissingEmbeddingAssetsError,
    assets_present,
    resolve_asset_dir,
@@ -34,28 +30,6 @@ _FETCH_HINT = (
 )


-@dataclass(frozen=True)
-class TokenOffset:
-    """Half-open Python Unicode code point offsets; synthetic specials use [0,0)."""
-
-    start_cp: int
-    end_cp: int
-    special: bool = False
-
-
-@dataclass(frozen=True)
-class TokenizedText:
-    """Untruncated, unpadded encoding, including tokenizer postprocessor specials."""
-
-    token_ids: tuple[int, ...]
-    offsets: tuple[TokenOffset, ...]
-
-    def __post_init__(self) -> None:
-        # Defensive copies also keep runtime callers from retaining mutable lists.
-        object.__setattr__(self, "token_ids", tuple(self.token_ids))
-        object.__setattr__(self, "offsets", tuple(self.offsets))
-
-
 class LocalEmbedder:
    """Encodes text with the built-in model, loading it on first use."""

@@ -68,43 +42,6 @@ class LocalEmbedder:
        self._session: Any | None = None
        self._tokenizer: Any | None = None
        self._input_names: frozenset[str] = frozenset()
-        self._offset_lock = threading.Lock()
-        self._offset_tokenizer: Any | None = None
-
-    def tokenize_with_offsets(self, text: str) -> TokenizedText:
-        """Tokenize original text without encode()'s truncation or batch padding.
-
-        The Python tokenizers binding returns code point offsets into the input,
-        even when its model normalizer changes spelling. This separate lazy
-        tokenizer uses the exact locked asset; it never loads an ONNX session or
-        mutates the tokenizer shared by encode(). Offsets are not UTF-16 indices.
-        """
-        if not isinstance(text, str):
-            raise TypeError("embedding_offsets_require_text")
-        with self._offset_lock:
-            if self._offset_tokenizer is None:
-                directory = resolve_asset_dir()
-                path = directory / "tokenizer.json"
-                if not path.is_file():
-                    raise MissingEmbeddingAssetsError(f"{_FETCH_HINT} Looked in {directory}.")
-                data = path.read_bytes()
-                expected = next(a.sha256 for a in EMBEDDING_ASSETS if a.filename == path.name)
-                if sha256(data).hexdigest() != expected:
-                    raise AssetVerificationError("tokenizer.json: locked digest mismatch")
-                from tokenizers import Tokenizer
-
-                tokenizer = Tokenizer.from_str(data.decode("utf-8"))
-                tokenizer.no_truncation()
-                tokenizer.no_padding()
-                self._offset_tokenizer = tokenizer
-            encoding = self._offset_tokenizer.encode(text, add_special_tokens=True)
-            offsets = tuple(
-                TokenOffset(start, end, bool(special))
-                for (start, end), special in zip(
-                    encoding.offsets, encoding.special_tokens_mask, strict=True
-                )
-            )
-            return TokenizedText(tuple(encoding.ids), offsets)

    def _load(self) -> tuple[Any, Any]:
        # Double-checked under a lock: memory processing encodes from worker
@@ -121,7 +58,7 @@ class LocalEmbedder:
                raise MissingEmbeddingAssetsError(f"{_FETCH_HINT} Looked in {directory}.")

            # Imported by name for the same reason kuzu and hy_memory are in
-            # hy_memory_bridge: onnxruntime ships no py.typed, and a dynamic
+            # onnxruntime ships no py.typed, and a dynamic
            # import keeps the strict checker on everything around it.
            ort = import_module("onnxruntime")
            from tokenizers import Tokenizer
@@ -14,6 +14,7 @@ stuck in it.
 """

 import logging
+import re
 import sys
 import threading
 import time
@@ -23,6 +24,7 @@ from types import FrameType
 logger = logging.getLogger("one_memory.loop_watchdog")

 DEFAULT_STALL_SECONDS = 2.0
+_SAFE_FALLBACK_FILE = re.compile(r"^[A-Za-z0-9_][A-Za-z0-9_.-]*\.py$")


 class LoopWatchdog:
@@ -120,22 +122,28 @@ def _safe_stack_for_main_thread() -> tuple[str, ...]:


 def _safe_stacks_for_all_threads() -> tuple[str, ...]:
-    """Return one content-free, bounded stack summary for every live thread."""
-    names = {thread.ident: thread.name for thread in threading.enumerate()}
+    """Return bounded stack summaries without recording thread names."""
    lines: list[str] = []
-    for ident, frame in sorted(sys._current_frames().items()):
+    for _ident, frame in sorted(sys._current_frames().items()):
        stack: list[str] = []
-        # Most specific frame first so the 256-character privacy bound keeps
-        # the likely culprit rather than only the thread bootstrap frames.
+        # Most specific frame first; append only complete frames so truncation
+        # can never turn a valid location into free text.
        for entry in reversed(traceback.extract_stack(frame)[-8:]):
            parts = entry.filename.replace("\\", "/").split("/")
            try:
                root = parts.index("one_memory")
                relative = "/".join(parts[root:])
            except ValueError:
-                relative = parts[-1]
+                fallback = parts[-1]
+                relative = fallback if _SAFE_FALLBACK_FILE.fullmatch(fallback) else "unknown.py"
            name = entry.name if entry.name.isidentifier() else "call"
-            stack.append(f"{relative}:{entry.lineno}:{name}")
-        thread_name = names.get(ident, "unknown").replace("\n", " ").replace("\r", " ")
-        lines.append(f"thread={thread_name[:48]}|{' > '.join(stack)}"[:256])
+            candidate = f"{relative}:{entry.lineno}:{name}"
+            separator = "" if not stack else " > "
+            if len("thread|") + len(" > ".join(stack)) + len(separator) + len(candidate) > 256:
+                break
+            stack.append(candidate)
+        if stack:
+            lines.append(f"thread|{' > '.join(stack)}")
+        if len(lines) == 16:
+            break
    return tuple(lines)
@@ -22,7 +22,7 @@ from pydantic import BaseModel, ConfigDict, Field

 from one_memory.database import enable_write_ahead_log
 from one_memory.database_access import main_database
-from one_memory.hy_lifecycle import HyLifecycle
+from one_memory.legacy_lifecycle_types import HyLifecycle
 from one_memory.library_upgrade import LibraryUpgradeError, PreparedLibrary, holds_legacy_rows
 from one_memory.memory_server.migrations import governance_ddl

@@ -1,342 +0,0 @@
-"""Read-only local library projection over the existing governed memory store."""
-
-import json
-import os
-import sqlite3
-import time
-from pathlib import Path
-from typing import Literal
-
-from pydantic import BaseModel, ConfigDict, Field
-
-from one_memory.database_access import main_database
-from one_memory.hy_lifecycle import HyLifecycle
-from one_memory.memory_evidence_freshness import register_memory_freshness
-from one_memory.memory_governance import (
-    GovernedMemory,
-    MemoryGovernanceStore,
-    MemoryStatus,
-    lifecycle_predicate,
-)
-
-#: Read budget for one library page. 2026-10-12: raised from 5s (which starved
-#: under a concurrent history import and surfaced as 503) to 15s, still well
-#: under the 30s client timeout; past it the read degrades to a `stale` page.
-READ_BUDGET_SECONDS = 15
-
-
-class MemorySourceTime(BaseModel):
-    """When the conversation a memory came from happened (D96).
-
-    `basis` is `conversation` when the cited messages carry their own time,
-    `import` when only the import time is known; the UI words them apart.
-    """
-
-    model_config = ConfigDict(frozen=True, extra="forbid")
-
-    at: str
-    basis: Literal["conversation", "import"]
-
-
-class MemoryLibraryPage(BaseModel):
-    model_config = ConfigDict(frozen=True, extra="forbid")
-
-    items: list[GovernedMemory]
-    total: int = Field(ge=0)
-    offset: int = Field(ge=0)
-    limit: int = Field(ge=1, le=500)
-    source_ids_by_memory: dict[str, list[str]] = Field(default_factory=dict)
-    source_time_by_memory: dict[str, MemorySourceTime] = Field(default_factory=dict)
-    #: True when the read budget ran out mid-page (typically during a history
-    #: import): the counts and rows above are a partial snapshot, not an
-    #: authoritative view. Clients should show "importing" instead of an error.
-    stale: bool = False
-
-
-def _sort_by_conversation() -> bool:
-    """D99: the list reads as a timeline of the conversations, newest first.
-
-    A memory generated today from a three-week-old dialogue belongs next to
-    that week, not at the top. ONE_MEMORY_LIBRARY_SORT=updated brings back the
-    old "newest generated first" order without a code change.
-    """
-    return os.environ.get("ONE_MEMORY_LIBRARY_SORT", "").strip().lower() != "updated"
-
-
-def _origin_join() -> str:
-    """The newest cited message per memory, as one joined aggregate.
-
-    A correlated subquery would re-scan the link table for every candidate row;
-    this is a single pass. Memories with no readable citation keep their old
-    position through COALESCE.
-    """
-    if not _sort_by_conversation():
-        return ""
-    return """LEFT JOIN (
-        SELECT link.memory_id AS memory_id,
-               MAX(COALESCE(e.created_at, e.imported_at)) AS origin_at
-        FROM governed_memory_evidence link
-        JOIN evidence e ON e.evidence_id = link.evidence_id
-        WHERE link.active = 1
-        GROUP BY link.memory_id
-    ) origin ON origin.memory_id = m.memory_id"""
-
-
-def _order_by() -> str:
-    if not _sort_by_conversation():
-        return "m.updated_at DESC, m.memory_id ASC"
-    # Conversations that share a timestamp keep the generated order between
-    # themselves; ties would otherwise fall back to the id and look arbitrary.
-    return "COALESCE(origin.origin_at, m.updated_at) DESC, m.updated_at DESC, m.memory_id ASC"
-
-
-class _LibraryQuery(BaseModel):
-    model_config = ConfigDict(frozen=True, extra="forbid")
-
-    user_id: str = Field(min_length=1, max_length=256)
-    query: str | None = None
-    status: MemoryStatus | None = None
-    active_only: bool = False
-    source: str | None = Field(default=None, min_length=1)
-    offset: int = Field(default=0, ge=0, le=2**63 - 1, strict=True)
-    limit: int = Field(default=100, ge=1, le=500, strict=True)
-
-
-def read_conversation_memories(
-    database_path: Path,
-    *,
-    user_id: str,
-    source_id: str,
-    conversation_id: str,
-    workspace: str | None,
-    limit: int = 100,
-) -> list[GovernedMemory]:
-    """Current memories backed by any active Evidence in one conversation."""
-    with main_database(database_path).read() as db:
-        db.row_factory = sqlite3.Row
-        freshness = register_memory_freshness(db)
-        rows = db.execute(
-            f"""
-            SELECT DISTINCT m.* FROM governed_memories m
-            JOIN governed_memory_evidence link ON link.memory_id=m.memory_id AND link.active=1
-            JOIN evidence e ON e.evidence_id=link.evidence_id
-            WHERE m.user_id=? AND m.status='confirmed' AND m.deleted_at IS NULL
-              AND e.user_id=? AND e.source_id=? AND e.conversation_id=?
-              AND e.workspace IS ? AND {freshness}
-            ORDER BY m.updated_at DESC, m.memory_id
-            LIMIT ?
-            """,
-            (user_id, user_id, source_id, conversation_id, workspace, limit),
-        ).fetchall()
-        return MemoryGovernanceStore.memories_from_rows(db, rows)
-
-
-def read_memory_library(
-    database_path: Path,
-    *,
-    user_id: str = "local",
-    query: str | None = None,
-    status: MemoryStatus | None = None,
-    active_only: bool = False,
-    source: str | None = None,
-    offset: int = 0,
-    limit: int = 100,
-    lifecycle: HyLifecycle | None = None,
-) -> MemoryLibraryPage:
-    """Read one page and its exact readable total from the same SQLite snapshot.
-
-    This is the local UI policy, not an external agent authorization entry point.
-    Missing grants do not revoke already stored content; explicit revocations do.
-    Read/parse failures propagate rather than masquerading as an empty library.
-    Which memories are current is HY's `lifecycle` when given, as in
-    `MemoryGovernanceStore.list`.
-    """
-    request = _LibraryQuery(
-        user_id=user_id,
-        query=query,
-        status=status,
-        active_only=active_only,
-        source=source,
-        offset=offset,
-        limit=limit,
-    )
-    if request.active_only and request.status is not None:
-        raise ValueError("active_only_conflicts_with_status")
-    clauses = ["m.user_id = ?"]
-    values: list[str | int] = [request.user_id]
-    selected, selected_values = lifecycle_predicate(
-        lifecycle, MemoryStatus.CONFIRMED if request.active_only else request.status, alias="m"
-    )
-    if selected is not None:
-        clauses.append(selected)
-        values.extend(selected_values)
-    if request.status != MemoryStatus.DELETED:
-        clauses.append("m.deleted_at IS NULL")
-
-    # SourceReadAccess hides the whole memory if any active citation is revoked.
-    # Do this before COUNT/LIMIT; a readable citation cannot rescue a mixed memory.
-    clauses.append("""NOT EXISTS (
-        SELECT 1 FROM governed_memory_evidence link
-        JOIN evidence e ON e.evidence_id = link.evidence_id
-        JOIN source_grants g ON g.user_id = e.user_id AND g.source_id = e.source_id
-        WHERE link.memory_id = m.memory_id AND +link.active = 1
-          AND g.revoked_at IS NOT NULL
-    )""")
-    # Governance normally guarantees ownership and referential integrity. Fail
-    # closed for broken active links instead of exposing another user's citation.
-    clauses.append("""NOT EXISTS (
-        SELECT 1 FROM governed_memory_evidence link
-        LEFT JOIN evidence e ON e.evidence_id = link.evidence_id
-        WHERE link.memory_id = m.memory_id AND +link.active = 1
-          AND (e.evidence_id IS NULL OR e.user_id != m.user_id)
-    )""")
-    if request.query:
-        clauses.append("m.content LIKE ? ESCAPE '\\'")
-        escaped = request.query.replace("\\", "\\\\").replace("%", "\\%").replace("_", "\\_")
-        values.append(f"%{escaped}%")
-    if request.source is not None:
-        # Start from the compact active-link index, then resolve each Evidence
-        # identity by primary key. The old correlated EXISTS let SQLite scan
-        # every Evidence row for every candidate memory on large libraries.
-        clauses.append("""m.memory_id IN (
-            SELECT link.memory_id FROM governed_memory_evidence link
-            JOIN evidence e ON e.evidence_id = link.evidence_id
-            WHERE link.active = 1 AND e.user_id = ? AND e.source_id = ?
-        )""")
-        values.extend((request.user_id, request.source))
-    predicate = " AND ".join(clauses)
-
-    def _interrupted(error: Exception) -> bool:
-        return isinstance(error, sqlite3.OperationalError) and "interrupted" in str(error).lower()
-
-    with main_database(database_path).read() as db:
-        # 2026-10-12: 5s starved the page whenever a history import saturated
-        # SQLite, and the caller turned the deadline abort into a 503, so the
-        # memory page read as broken for the whole import. 15s stays well under
-        # the 30s client timeout; past it we return a `stale` page instead of
-        # an error — an incomplete list is honest, an error is not.
-        deadline = time.monotonic() + READ_BUDGET_SECONDS
-        db.set_progress_handler(lambda: int(time.monotonic() >= deadline), 1000)
-        db.row_factory = sqlite3.Row
-        db.execute("BEGIN")
-        # The total counts with the cheap permission filters only: the per-row
-        # freshness UDF scales with the whole library and can burn the whole
-        # read budget on its own (D22). Product accepted a small drift here;
-        # the rows that are actually displayed are still checked row by row.
-        try:
-            total = db.execute(
-                f"SELECT COUNT(*) FROM governed_memories m WHERE {predicate}", values
-            ).fetchone()[0]
-        except sqlite3.OperationalError as error:
-            if not _interrupted(error):
-                raise
-            return MemoryLibraryPage(
-                items=[],
-                total=0,
-                offset=request.offset,
-                limit=request.limit,
-                stale=True,
-            )
-        predicate += " AND " + register_memory_freshness(db)
-        try:
-            if request.query:
-                current = db.execute(
-                    f"""SELECT m.* FROM governed_memories m {_origin_join()}
-                    WHERE {predicate}
-                    ORDER BY {_order_by()}""",
-                    values,
-                ).fetchall()
-                rows = current[request.offset : request.offset + request.limit]
-            else:
-                rows = db.execute(
-                    f"""SELECT m.* FROM governed_memories m {_origin_join()}
-                    WHERE {predicate}
-                    ORDER BY {_order_by()} LIMIT ? OFFSET ?""",
-                    [*values, request.limit, request.offset],
-                ).fetchall()
-        except sqlite3.OperationalError as error:
-            if not _interrupted(error):
-                raise
-            # The count already succeeded: report it, flag the page stale.
-            return MemoryLibraryPage(
-                items=[],
-                total=total,
-                offset=request.offset,
-                limit=request.limit,
-                stale=True,
-            )
-        # Static read helper only: constructing the Store would initialize schema.
-        items = MemoryGovernanceStore.memories_from_rows(db, rows, lifecycle=lifecycle)
-        # Only the authorized page participates; source filters select memories,
-        # not a subset of their provenance. Keep this in the count/page snapshot.
-        source_ids_by_memory: dict[str, list[str]] = {item.memory_id: [] for item in items}
-        # D96: the list shows when the cited conversation happened, not when
-        # the memory was generated (which is "today" for a fresh import). Read
-        # in the same query as the sources, so the page costs no extra round
-        # trip and only this page's IDs are looked at, never the whole library.
-        conversation_at: dict[str, str] = {}
-        imported_at: dict[str, str] = {}
-        if items:
-            try:
-                for link in db.execute(
-                    # D104: the conversation time is when the person spoke, not
-                    # when the turn ended. A memory's evidence covers the whole
-                    # round (question, tool calls, reply), so MAX() lands on the
-                    # reply and reads as a later, contradictory time next to the
-                    # cited source. Prefer the earliest person-side evidence; only
-                    # fall back to the earliest of any kind when the round has no
-                    # person-authored row at all.
-                    """SELECT link.memory_id, e.source_id,
-                           COALESCE(
-                               MIN(
-                                   CASE WHEN e.role = 'user'
-                                         AND e.role = 'user'
-                                        THEN e.created_at END
-                               ),
-                               MIN(e.created_at)
-                           ) AS conversation_at,
-                           MAX(e.imported_at) AS imported_at
-                    FROM governed_memory_evidence link
-                    JOIN evidence e ON e.evidence_id = link.evidence_id
-                    WHERE link.active = 1 AND e.user_id = ?
-                      AND link.memory_id IN (SELECT value FROM json_each(?))
-                    GROUP BY link.memory_id, e.source_id
-                    ORDER BY link.memory_id, e.source_id""",
-                    (request.user_id, json.dumps(list(source_ids_by_memory))),
-                ):
-                    memory_id = link["memory_id"]
-                    source_ids_by_memory[memory_id].append(link["source_id"])
-                    for seen, value in (
-                        (conversation_at, link["conversation_at"]),
-                        (imported_at, link["imported_at"]),
-                    ):
-                        if value and str(value) > seen.get(memory_id, ""):
-                            seen[memory_id] = str(value)
-            except sqlite3.OperationalError as error:
-                if not _interrupted(error):
-                    raise
-                # Page rows are already read; ship them without source maps
-                # rather than failing the whole request at the last step.
-                return MemoryLibraryPage(
-                    items=items,
-                    total=total,
-                    offset=request.offset,
-                    limit=request.limit,
-                    source_ids_by_memory=source_ids_by_memory,
-                    stale=True,
-                )
-        source_time_by_memory = {
-            memory_id: MemorySourceTime(at=conversation_at[memory_id], basis="conversation")
-            if memory_id in conversation_at
-            else MemorySourceTime(at=imported_at[memory_id], basis="import")
-            for memory_id in source_ids_by_memory
-            if memory_id in conversation_at or memory_id in imported_at
-        }
-        return MemoryLibraryPage(
-            items=items,
-            total=total,
-            offset=request.offset,
-            limit=request.limit,
-            source_ids_by_memory=source_ids_by_memory,
-            source_time_by_memory=source_time_by_memory,
-        )
@@ -1,180 +0,0 @@
-"""Bounded, permission-aware receipt projection; never initialize a write store."""
-
-import sqlite3
-from pathlib import Path
-from typing import Literal
-
-from pydantic import BaseModel, ConfigDict
-
-from one_memory.database_access import main_database
-
-#: `legacy` is a confirmation recorded before memories stopped waiting for
-#: one; it changed nothing that still exists, and `migrated_from` says so.
-Operation = Literal["create", "correct", "set_scopes", "undo", "delete", "hy_reconcile", "legacy"]
-ResultStatus = Literal["confirmed", "deleted"]
-MigratedFrom = Literal["confirm", "reject"]
-OPERATIONS = {"create", "correct", "set_scopes", "undo", "delete", "hy_reconcile", "legacy"}
-STATUSES = {"confirmed", "deleted"}
-#: Which operation each pre-v2 one now reads as.
-MIGRATED_OPERATIONS = {"confirm": "legacy", "reject": "delete"}
-#: A memory may move on to these after an earlier receipt was written; the
-#: receipt still reports what that request committed, not the current status.
-CURRENT_STATUSES = STATUSES | {"superseded"}
-
-
-class MemoryMutationReceiptForbidden(RuntimeError):
-    def __init__(self) -> None:
-        super().__init__("memory_mutation_receipt_forbidden")
-
-
-class MemoryMutationReceiptUnavailable(RuntimeError):
-    def __init__(self) -> None:
-        super().__init__("memory_mutation_receipt_unavailable")
-
-
-class MemoryMutationReceipt(BaseModel):
-    model_config = ConfigDict(frozen=True, extra="forbid")
-    schema_version: Literal[2] = 2
-    request_id: str
-    state: Literal["not_found", "committed"]
-    operation: Operation | None = None
-    memory_id: str | None = None
-    result_revision: int | None = None
-    result_status: ResultStatus | None = None
-    memory_state: Literal["present", "deleted"] | None = None
-    #: The operation as it was stored before v2, when it has no v2 name of its own.
-    migrated_from: MigratedFrom | None = None
-
-
-def _identity(value: str, maximum: int) -> bool:
-    return isinstance(value, str) and 0 < len(value) <= maximum
-
-
-def _read_snapshot(
-    database_path: Path, request_id: str, user_id: str
-) -> tuple[MemoryMutationReceipt, tuple[object, ...]]:
-    with main_database(database_path).read(row_factory=sqlite3.Row) as db:
-        db.execute("PRAGMA query_only=ON")
-        db.execute("BEGIN")
-        rows = db.execute(
-            "SELECT user_id,request_id,memory_id,operation,result_revision "
-            "FROM governed_memory_mutations WHERE user_id=? AND request_id=? LIMIT 2",
-            (user_id, request_id),
-        ).fetchall()
-        if not rows:
-            return MemoryMutationReceipt(request_id=request_id, state="not_found"), ()
-        if len(rows) != 1:
-            raise MemoryMutationReceiptUnavailable()
-        mutation = rows[0]
-        memory_id, operation, revision = (
-            mutation["memory_id"],
-            mutation["operation"],
-            mutation["result_revision"],
-        )
-        if (
-            not _identity(memory_id, 512)
-            or operation not in OPERATIONS
-            or type(revision) is not int
-            or not 1 <= revision <= 2**53 - 1
-        ):
-            raise MemoryMutationReceiptUnavailable()
-        current = db.execute(
-            "SELECT user_id,revision,status,deleted_at FROM governed_memories WHERE memory_id=?",
-            (memory_id,),
-        ).fetchone()
-        if (
-            current is None
-            or current["user_id"] != user_id
-            or type(current["revision"]) is not int
-            or current["revision"] < revision
-            or current["status"] not in CURRENT_STATUSES
-            or (current["status"] == "deleted") != (current["deleted_at"] is not None)
-        ):
-            raise MemoryMutationReceiptUnavailable()
-
-        # Versions have no historical evidence snapshot. Check every retained link,
-        # not only active ones, before disclosing even operation/ID metadata.
-        links = db.execute(
-            "SELECT l.evidence_id,l.active,e.user_id,e.source_id,e.revision,"
-            "g.granted_at,g.revoked_at FROM governed_memory_evidence l "
-            "LEFT JOIN evidence e ON e.evidence_id=l.evidence_id "
-            "LEFT JOIN source_grants g ON g.user_id=e.user_id AND g.source_id=e.source_id "
-            "WHERE l.memory_id=? ORDER BY l.evidence_id LIMIT 10001",
-            (memory_id,),
-        ).fetchall()
-        if not links or len(links) > 10000:
-            raise MemoryMutationReceiptUnavailable()
-        for link in links:
-            if (
-                link["user_id"] != user_id
-                or link["source_id"] is None
-                or link["active"] not in (0, 1)
-            ):
-                raise MemoryMutationReceiptUnavailable()
-            if link["revoked_at"] is not None:
-                raise MemoryMutationReceiptForbidden()
-        versions = db.execute(
-            "SELECT memory_id,revision,action,status,migrated_from FROM governed_memory_versions "
-            "WHERE memory_id=? AND revision=? LIMIT 2",
-            (memory_id, revision),
-        ).fetchall()
-        if len(versions) != 1:
-            raise MemoryMutationReceiptUnavailable()
-        version = versions[0]
-        if version["action"] != operation or version["status"] not in STATUSES:
-            raise MemoryMutationReceiptUnavailable()
-        migrated_from = version["migrated_from"]
-        if migrated_from is not None and MIGRATED_OPERATIONS.get(migrated_from) != operation:
-            raise MemoryMutationReceiptUnavailable()
-        if (operation == "legacy") != (migrated_from == "confirm"):
-            raise MemoryMutationReceiptUnavailable()
-        required_status = {
-            "create": "confirmed",
-            "legacy": "confirmed",
-            "delete": "deleted",
-        }
-        if operation in required_status and version["status"] != required_status[operation]:
-            raise MemoryMutationReceiptUnavailable()
-        if (operation == "create") != (revision == 1):
-            raise MemoryMutationReceiptUnavailable()
-        if (
-            operation in {"correct", "set_scopes", "hy_reconcile", "undo"}
-            and version["status"] != "confirmed"
-        ):
-            raise MemoryMutationReceiptUnavailable()
-        receipt = MemoryMutationReceipt(
-            request_id=request_id,
-            state="committed",
-            operation=operation,
-            memory_id=memory_id,
-            result_revision=revision,
-            result_status=version["status"],
-            memory_state="deleted" if current["status"] == "deleted" else "present",
-            migrated_from=migrated_from,
-        )
-        return receipt, (
-            tuple(mutation),
-            tuple(current),
-            tuple(version),
-            tuple(tuple(link) for link in links),
-        )
-
-
-def read_memory_mutation_receipt(
-    database_path: Path, *, request_id: str, user_id: str = "local"
-) -> MemoryMutationReceipt:
-    """History only, never current authorization, HY cleanup or safe-retry proof."""
-    if not _identity(request_id, 256):
-        raise ValueError("invalid_memory_mutation_request_id")
-    if not _identity(user_id, 256):
-        raise ValueError("invalid_memory_mutation_user_id")
-    try:
-        first = _read_snapshot(database_path, request_id, user_id)
-        current = _read_snapshot(database_path, request_id, user_id)
-        if first != current:
-            raise MemoryMutationReceiptUnavailable()
-        return current[0]
-    except MemoryMutationReceiptForbidden:
-        raise
-    except (sqlite3.Error, ValueError, TypeError, KeyError, IndexError, OSError):
-        raise MemoryMutationReceiptUnavailable() from None
@@ -3,7 +3,7 @@ import json
 import sqlite3
 from collections.abc import Sequence
 from datetime import UTC, datetime
-from typing import Literal, cast
+from typing import Any, Literal, cast
 from uuid import UUID

 from fastapi import APIRouter, BackgroundTasks, HTTPException, Request
@@ -21,11 +21,10 @@ from one_memory.imports.models import (
 )
 from one_memory.imports.storage import EvidenceStore
 from one_memory.input_sanitization import sanitize_model_input
-from one_memory.memory_governance import MemoryStatus
-from one_memory.memory_processing import MemoryProcessingSupervisor
+from one_memory.memory_governance import MemoryGovernanceStore, MemoryStatus
 from one_memory.memory_server.durable_queue import DurableWriteQueue
 from one_memory.memory_server.unified_entry import UnifiedWriteIntake
-from one_memory.memory_tasks import MemoryTaskStatus
+from one_memory.memory_tasks import MemoryTaskStatus, MemoryTaskStore
 from one_memory.secrets import SecretStoreUnavailableError

 SOURCE_ID = "one-memory-manual"
@@ -68,25 +67,23 @@ class MemoryNoteReceipt(BaseModel):
    error: str | None = None


-class SavedMemoryNoteEvidence(BaseModel):
-    model_config = ConfigDict(frozen=True, extra="forbid")
+class LegacyMemoryNoteReader:
+    """Read pre-cutover note receipts without retaining any write runtime."""

-    available: Literal[True] = True
-    evidence_id: str
-    source_id: str
-    revision: int
-    imported_at: datetime
-    role: str
-    written_by: WrittenBy
-    created_at: datetime | None
-    verbatim: str
+    def __init__(
+        self,
+        tasks: MemoryTaskStore,
+        governance: MemoryGovernanceStore,
+    ) -> None:
+        self.tasks = tasks
+        self.governance = governance


 class MemoryNoteStore:
    def __init__(
        self,
        evidence: EvidenceStore,
-        processor: MemoryProcessingSupervisor,
+        processor: Any | None,
        intake: UnifiedWriteIntake | None = None,
        *,
        sdk_authority: bool = False,
@@ -252,6 +249,8 @@ class MemoryNoteStore:
                memory_ids=(list(durable.sdk_memory_ids) if durable is not None else []),
                error=durable.detail if durable is not None else "write_missing",
            )
+        if self.processor is None:
+            raise RuntimeError("legacy_memory_note_processor_unavailable")
        task = self.processor.tasks.task_for_evidence(evidence_id)
        memory_ids = []
        if task:
@@ -356,6 +355,8 @@ def memory_note_router(dependencies: Sequence[Depends]) -> APIRouter:
        try:
            if notes.sdk_authority:
                return await asyncio.to_thread(notes.receipt, record.evidence_id)
+            if notes.processor is None:
+                raise RuntimeError("legacy_memory_note_processor_unavailable")
            if notes.intake is None:
                # Isolated tests construct the store without a queue.
                await notes.processor.enqueue_conversation_off_thread(record)
@@ -426,48 +427,6 @@ def memory_note_router(dependencies: Sequence[Depends]) -> APIRouter:
        except ValueError as error:
            raise HTTPException(409, str(error)) from error

-    @router.get("/{evidence_id}/evidence", response_model=SavedMemoryNoteEvidence)
-    def evidence(evidence_id: str, request: Request) -> SavedMemoryNoteEvidence:
-        try:
-            notes = store(request)
-            if notes.sdk_authority:
-                raise HTTPException(409, "memory_note_evidence_retired")
-            with notes.evidence._connect() as connection:
-                record = notes.evidence._get_with_connection(connection, evidence_id)
-                metadata = record.origin_meta.get(EvidenceOrigin.MANUAL_IMPORT) if record else None
-                if (
-                    record is None
-                    or record.user_id != "local"
-                    or record.source_id != SOURCE_ID
-                    or metadata is None
-                    or metadata.raw_meta.get("input_mode") != "direct_memory_note"
-                ):
-                    raise KeyError(evidence_id)
-                grant = connection.execute(
-                    "SELECT revoked_at FROM source_grants WHERE user_id='local' AND source_id=?",
-                    (SOURCE_ID,),
-                ).fetchone()
-                if grant is None or grant[0] is not None:
-                    raise ValueError("memory_note_source_revoked")
-                result = SavedMemoryNoteEvidence(
-                    evidence_id=record.evidence_id,
-                    source_id=record.source_id,
-                    revision=record.revision,
-                    imported_at=record.imported_at,
-                    role=record.role,
-                    written_by=record.written_by,
-                    created_at=record.created_at,
-                    verbatim=record.verbatim,
-                )
-            # The returned body and reference share one snapshot. Recheck authority
-            # after that read so a revoked source is not released from a stale view.
-            notes.require(evidence_id)
-            return result
-        except KeyError as error:
-            raise HTTPException(404, "memory_note_not_found") from error
-        except ValueError as error:
-            raise HTTPException(409, str(error)) from error
-
    @router.post("/{evidence_id}/process", response_model=MemoryNoteReceipt)
    async def process(evidence_id: str, request: Request) -> MemoryNoteReceipt:
        notes = store(request)
@@ -69,8 +69,8 @@ _TYPE_NAME = re.compile(r"^[A-Za-z_][A-Za-z0-9_]{0,63}$")
 def safe_error_code(error: BaseException) -> str:
    """Reduce an exception to a storable code, never its message text.

-    A recognised code (the message *is* the code, as `HyMemoryBridgeError`
-    and the permission guard raise them) is kept so callers can classify it.
+    A recognised code (the message *is* a stable domain code) is kept so
+    callers can classify it.
    Everything else collapses to its exception type, which is a fixed
    identifier rather than content.
    """
@@ -239,8 +239,8 @@ PERMISSION_REQUIRED_CODES = frozenset(
    }
 )

-#: Model-side codes `process_records` raises *before* it can open the bridge
-#: (no active profile, missing credential, secure storage down). They prove
+#: Model-side codes the SDK write coordinator raises before it can lease a
+#: client (no active profile, missing credential, secure storage down). They prove
 #: nothing was sent, so the write parks instead of being lost. Transport codes
 #: that can arrive after a request was in flight stay `unknown`.
 MODEL_NOT_READY_CODES = frozenset(
@@ -251,9 +251,9 @@ MODEL_NOT_READY_CODES = frozenset(
    }
 )

-#: Deterministic, pre-bridge refusals: the frozen input no longer matches
+#: Deterministic, pre-SDK refusals: the frozen input no longer matches
 #: current Evidence, or a turn exceeds the model's input window. Both are
-#: raised while validating the records, before any bridge or HY call, so
+#: raised while validating the records, before any SDK call, so
 #: nothing was sent and retrying the same intent changes nothing — a terminal
 #: `failed`, not `unknown` (which would wrongly imply a possibly-applied
 #: write) and not a retry (which would fail identically).
@@ -316,6 +316,12 @@ LEGACY_MEMORY_WRITE_JOB_COLUMNS = (

 _FIFO_BLOCKING_STATES = (PENDING, RUNNING, WAITING_FOR_PERMISSION, WAITING_FOR_MODEL)
 EVENT_TIME_WINDOW = timedelta(minutes=10)
+#: Sources whose jobs are a person adding a memory by hand ("添加记忆"), not
+#: history. History is digested strictly old to new through the event-time
+#: window; a note is stamped "now", so inside that window it waited behind the
+#: whole unprocessed history. These jobs are claimed ahead of the window,
+#: still one at a time per source and in order within their session.
+DIRECT_WRITE_SOURCES = frozenset({"one-memory-manual"})


@dataclass(frozen=True)
@@ -358,12 +364,10 @@ class UnknownWriteJob:

    `request_id` is the durable queue's idempotency key. It identifies the row
    for queue inspection and is the value `DurableWriteQueue.get()` accepts.
-    It is **not** the HY request id: `MemoryProcessingSupervisor` derives the
-    HY handle from the task it creates, so an inquiry implementation must
-    obtain that task-layer handle separately rather than passing this value to
-    HY. This is discovery only: it does not decide whether HY has the write,
-    and it deliberately exposes no transition method that could turn
-    `unknown` back into retryable work.
+    It is the public SDK request id used for exact outcome inquiry. This is
+    discovery only: it does not decide whether HY has the write, and it
+    deliberately exposes no transition method that could turn `unknown` back
+    into retryable work.
    """

    job_id: str
@@ -1759,8 +1763,8 @@ class DurableWriteQueue:
        """Retire an `unknown` job after an inquiry, never by guessing.

        `unknown` means HY may already hold the write; the only safe way out is
-        to ask by request id (`hy_capabilities` measured that the SDK supports
-        this) and then record what the inquiry found. `found=True` closes it as
+        to call the public SDK `get_write_outcome(request_id)` and then record
+        what the inquiry found. `found=True` closes it as
        `completed`; `found=False` gives up and marks it `dead`. The state
        machine's legal transitions are the authority, so this refuses any
        change that `job_state.next_state` would reject — in particular it can
@@ -2258,12 +2262,11 @@ async def execute_job(
    admit_hy_invocation: Callable[[str, Callable[[], Awaitable[bool]]], Awaitable[bool]]
    | None = None,
 ) -> None:
-    """Run one durable job through the existing authorized processing path.
+    """Run one durable job through the SDK-authority processor.

-    The executor does not call HY directly: `process_records` owns the
-    permission guard, the runner and governance. `hy_invoked` is recorded
-    before that call, so a crash mid-write is later reclaimed to `unknown`
-    instead of being retried blindly.
+    ``process_sdk_job`` owns the one public SDK write. ``hy_invoked`` is
+    recorded immediately before that call, so a crash mid-write is later
+    reclaimed to ``unknown`` instead of being retried blindly.
    """
    # Every store call here is synchronous SQLite work on the single writer or
    # a reader; made from the loop it would block health checks and the API for
@@ -2336,38 +2339,13 @@ async def execute_job(
    try:
        level = ServiceLevel.REALTIME if request.priority == "live" else ServiceLevel.BACKGROUND
        with service_level_context(level):
-            authority = getattr(supervisor, "sdk_authority", None)
-            if callable(authority) and authority():
-                sdk_result = await supervisor.process_sdk_job(
-                    records,
-                    request_id=request.idempotency_key,
-                    before_hy_invocation=mark_hy_invocation,
-                )
-                raw_ids = sdk_result.raw_ids
-                memory_ids = tuple(memory.memory_id for memory in sdk_result.memories)
-            else:
-                completed_tasks = (
-                    await supervisor.process_records(
-                        records,
-                        before_hy_invocation=mark_hy_invocation,
-                        source_admitted=True,
-                    )
-                    or ()
-                )
-                receipts = []
-                for task in completed_tasks:
-                    receipt = await in_the_background(supervisor.tasks.sdk_receipt, task.task_id)
-                    if receipt is None:
-                        raise RuntimeError("sdk_write_receipt_missing")
-                    receipts.append(receipt)
-                raw_ids = tuple(
-                    dict.fromkeys(raw_id for receipt in receipts for raw_id in receipt.raw_ids)
-                )
-                memory_ids = tuple(
-                    dict.fromkeys(
-                        memory_id for receipt in receipts for memory_id in receipt.memory_ids
-                    )
-                )
+            sdk_result = await supervisor.process_sdk_job(
+                records,
+                request_id=request.idempotency_key,
+                before_hy_invocation=mark_hy_invocation,
+            )
+            raw_ids = sdk_result.raw_ids
+            memory_ids = sdk_result.memory_ids
    except Exception as error:  # noqa: BLE001 - classified below
        # Never persist the raw message: it may be a provider response or the
        # user's own text, and `last_error` is shown over HTTP as `detail`.
@@ -2382,7 +2360,7 @@ async def execute_job(
        ):
            await in_the_background(queue.release_unstarted, job_id, owner=owner)
            return
-        # The task layer classifies these before any bridge is opened, so HY
+        # The write coordinator classifies these before any SDK client is leased, so HY
        # was provably not called. They must not read as `failed`:
        # - authorization stop -> cancelled (the user turned it off);
        # - no permission yet -> waiting_for_permission (parked, re-admitted
@@ -2445,13 +2423,10 @@ async def execute_job(
        sdk_raw_ids=raw_ids,
        sdk_memory_ids=memory_ids,
    )
-    authority = getattr(supervisor, "sdk_authority", None)
    release = getattr(evidence, "release_sdk_staging", None)
    durable_receipt = await in_the_background(queue.sdk_receipt, job_id)
    if (
-        callable(authority)
-        and authority()
-        and callable(release)
+        callable(release)
        and durable_receipt is not None
        and (durable_receipt.raw_ids or durable_receipt.memory_ids)
    ):
@@ -2632,6 +2607,39 @@ class DurableWriteDrain:
                placeholders = ",".join("?" for _ in enabled_sources)
                source_clause = f"AND candidate.source_key IN ({placeholders}) "
                source_params = tuple(sorted(enabled_sources))
+            direct_sources = ",".join("?" for _ in DIRECT_WRITE_SOURCES)
+            # Same eligibility as the window claim below (free lease, no
+            # running job of its source, nothing earlier in its session), but
+            # not bound to the window: a note must not wait for history.
+            direct_rows = connection.execute(
+                "SELECT candidate.job_id,candidate.source_key "
+                "FROM memory_write_jobs AS candidate "
+                "WHERE candidate.state=? "
+                "AND (candidate.lease_expires_at IS NULL OR candidate.lease_expires_at < ?) "
+                + source_clause
+                + f"AND candidate.source_key IN ({direct_sources}) "
+                "AND NOT EXISTS ("
+                "SELECT 1 FROM memory_write_jobs AS held "
+                "WHERE held.source_key=candidate.source_key AND held.state=? "
+                "AND held.lease_owner IS NOT NULL AND held.lease_expires_at>=?) "
+                "AND NOT EXISTS ("
+                "SELECT 1 FROM memory_write_jobs AS earlier "
+                "WHERE earlier.session_key=candidate.session_key "
+                f"AND earlier.state IN ({blocking}) "
+                "AND (earlier.event_at<candidate.event_at OR "
+                "(earlier.event_at=candidate.event_at "
+                "AND earlier.queue_seq<candidate.queue_seq))) "
+                "ORDER BY candidate.queue_seq",
+                (
+                    PENDING,
+                    now.isoformat(),
+                    *source_params,
+                    *sorted(DIRECT_WRITE_SOURCES),
+                    RUNNING,
+                    now.isoformat(),
+                    *_FIFO_BLOCKING_STATES,
+                ),
+            ).fetchall()
            window = connection.execute(
                "SELECT window_anchor,window_end FROM memory_write_schedule_state WHERE singleton=1"
            ).fetchone()
@@ -2685,6 +2693,7 @@ class DurableWriteDrain:
                if active is None:
                    connection.execute("DELETE FROM memory_write_schedule_state WHERE singleton=1")
                    window = None
+            anchor: Any = None
            if window is None:
                anchor = connection.execute(
                    "SELECT candidate.event_at "
@@ -2713,8 +2722,10 @@ class DurableWriteDrain:
                        *_FIFO_BLOCKING_STATES,
                    ),
                ).fetchone()
-                if anchor is None:
+                if anchor is None and not direct_rows:
                    return ()
+            has_window = window is not None or anchor is not None
+            if window is None and anchor is not None:
                anchor_at = datetime.fromisoformat(str(anchor[0]))
                if anchor_at.tzinfo is None:
                    anchor_at = anchor_at.replace(tzinfo=UTC)
@@ -2726,39 +2737,41 @@ class DurableWriteDrain:
                    "VALUES (1,?,?,?)",
                    (anchor_text, end_text, now.isoformat()),
                )
-            rows = connection.execute(
-                "SELECT candidate.job_id,candidate.source_key "
-                "FROM memory_write_jobs AS candidate "
-                "WHERE candidate.state=? "
-                "AND (candidate.lease_expires_at IS NULL OR candidate.lease_expires_at < ?) "
-                + source_clause
-                + "AND candidate.event_at>=? AND candidate.event_at<? "
-                "AND NOT EXISTS ("
-                "SELECT 1 FROM memory_write_jobs AS held "
-                "WHERE held.source_key=candidate.source_key AND held.state=? "
-                "AND held.lease_owner IS NOT NULL AND held.lease_expires_at>=?) "
-                "AND NOT EXISTS ("
-                "SELECT 1 FROM memory_write_jobs AS earlier "
-                "WHERE earlier.session_key=candidate.session_key "
-                f"AND earlier.state IN ({blocking}) "
-                "AND (earlier.event_at<candidate.event_at OR "
-                "(earlier.event_at=candidate.event_at "
-                "AND earlier.queue_seq<candidate.queue_seq))) "
-                "ORDER BY candidate.event_at,candidate.queue_seq",
-                (
-                    PENDING,
-                    now.isoformat(),
-                    *source_params,
-                    anchor_text,
-                    end_text,
-                    RUNNING,
-                    now.isoformat(),
-                    *_FIFO_BLOCKING_STATES,
-                ),
-            ).fetchall()
+            rows: list[Any] = []
+            if has_window:
+                rows = connection.execute(
+                    "SELECT candidate.job_id,candidate.source_key "
+                    "FROM memory_write_jobs AS candidate "
+                    "WHERE candidate.state=? "
+                    "AND (candidate.lease_expires_at IS NULL OR candidate.lease_expires_at < ?) "
+                    + source_clause
+                    + "AND candidate.event_at>=? AND candidate.event_at<? "
+                    "AND NOT EXISTS ("
+                    "SELECT 1 FROM memory_write_jobs AS held "
+                    "WHERE held.source_key=candidate.source_key AND held.state=? "
+                    "AND held.lease_owner IS NOT NULL AND held.lease_expires_at>=?) "
+                    "AND NOT EXISTS ("
+                    "SELECT 1 FROM memory_write_jobs AS earlier "
+                    "WHERE earlier.session_key=candidate.session_key "
+                    f"AND earlier.state IN ({blocking}) "
+                    "AND (earlier.event_at<candidate.event_at OR "
+                    "(earlier.event_at=candidate.event_at "
+                    "AND earlier.queue_seq<candidate.queue_seq))) "
+                    "ORDER BY candidate.event_at,candidate.queue_seq",
+                    (
+                        PENDING,
+                        now.isoformat(),
+                        *source_params,
+                        anchor_text,
+                        end_text,
+                        RUNNING,
+                        now.isoformat(),
+                        *_FIFO_BLOCKING_STATES,
+                    ),
+                ).fetchall()
            selected: list[str] = []
            selected_sources: set[str] = set()
-            for row in rows:
+            for row in (*direct_rows, *rows):
                source_key = str(row["source_key"])
                if source_key in selected_sources:
                    continue
@@ -1,182 +0,0 @@
-"""A0-03: measured answer to "can a HY write be made idempotent?".
-
-The result decides whether the `unknown` job state can ever be retired, so it
-is probed against the installed SDK rather than written down from memory.
-
-What is verified against `hy-memory==2.0.5.dev3`:
-
- `HyMemoryClient.add(..., request_id=...)` accepts a caller-supplied id and
-  atomically claims it in the SDK cache before invoking a write. Replays return
-  the durable terminal result; running/unknown claims are not invoked again.
- `HyMemoryClient.get_write_outcome(request_id)` reads that durable outcome
-  directly; it does not depend on optional pipeline tracing.
-
-A first pass at this probe got both answers wrong, and the failure mode is
-worth keeping in mind: searching SDK source for the string "idempotent" hits a
-docstring about *episode* nodes ("calls are idempotent") and a "dedup" comment
-about search over-sampling. Neither is write idempotency. So the probe checks
-the `add` path specifically and names the inspector class explicitly, instead
-of grepping for a word that appears in unrelated prose.
-
-Net: MemHub can query and settle an uncertain request without blind replay.
-`unknown` remains a legitimate operational state while the SDK claim itself is
-non-terminal, but the same request id cannot write twice.
-"""
-
-from __future__ import annotations
-
-import importlib
-import inspect
-from enum import StrEnum
-from typing import Any
-
-
-class HyIdempotencySupport(StrEnum):
-    """Whether the SDK can dedupe a replayed write."""
-
-    #: A caller key is honoured and a replay is rejected or folded together.
-    KEY = "key"
-    #: Same content is folded by the engine, without an explicit key.
-    CONTENT_HASH = "content_hash"
-    #: A replay writes a second time. Retrying after a crash duplicates.
-    NONE = "none"
-    #: Never probed. Must not survive A0-03.
-    UNKNOWN = "unknown"
-
-
-class HyResultLookupSupport(StrEnum):
-    """Whether the outcome of a past write can be queried afterwards."""
-
-    BY_REQUEST_ID = "by_request_id"
-    BY_MEMORY_ID_ONLY = "by_memory_id_only"
-    NONE = "none"
-    UNKNOWN = "unknown"
-
-
-def _module(name: str) -> Any | None:
-    try:
-        return importlib.import_module(name)
-    except ImportError:
-        return None
-
-
-def _has_callable(obj: Any, name: str) -> bool:
-    return obj is not None and callable(getattr(obj, name, None))
-
-
-def hy_write_accepts_request_id() -> bool:
-    """True when the write path carries a caller-supplied request id."""
-    client_mod = _module("hy_memory.client")
-    if client_mod is None:
-        return False
-    client = getattr(client_mod, "HyMemoryClient", None)
-    if client is None:
-        return False
-    add = getattr(client, "add", None)
-    if add is None:
-        return False
-    return "request_id" in inspect.signature(add).parameters
-
-
-def _add_path_declares_idempotency() -> bool:
-    """Whether the *write* path de-duplicates, judged on its own source.
-
-    Deliberately scoped to `add`/`async_add`. A whole-module search matches
-    unrelated prose: "calls are idempotent" describes episode-node upserts, and
-    a "dedup" comment describes search over-sampling. Both are about other
-    features and would make this probe answer `key` for the wrong reason.
-    """
-    client_mod = _module("hy_memory.client")
-    if client_mod is None:
-        return False
-    client = getattr(client_mod, "HyMemoryClient", None)
-    if client is None:
-        return False
-    sources: list[str] = []
-    for name in ("add", "async_add"):
-        func = getattr(client, name, None)
-        if func is None:
-            continue
-        try:
-            sources.append(inspect.getsource(func).lower())
-        except (OSError, TypeError):
-            continue
-    if not sources:
-        return False
-    if _has_callable(client, "get_write_outcome"):
-        claim = getattr(client, "_claim_write_outcome", None)
-        if claim is not None:
-            return True
-    # Older SDKs may expose an explicit key under one of these names.
-    markers = ("idempotency_key", "idempotent_key", "dedupe_key", "dedup_key")
-    return any(marker in source for source in sources for marker in markers)
-
-
-def _inspector_can_lookup_by_request_id() -> bool:
-    """True when pipeline logs can be queried by the write's request id."""
-    inspector = _module("hy_memory.inspector")
-    if inspector is None:
-        return False
-    for cls_name in ("MemoryInspector", "Inspector", "PipelineInspector"):
-        cls = getattr(inspector, cls_name, None)
-        if cls is None:
-            continue
-        for method in ("get_pipeline_logs", "async_get_pipeline_logs"):
-            func = getattr(cls, method, None)
-            if func is None:
-                continue
-            # The parameter is `request_id` on `MemoryInspector`; older
-            # docstrings call the same thing `n`. Accept either so a rename
-            # does not silently downgrade this to "no lookup available".
-            if {"request_id", "n"} & set(inspect.signature(func).parameters):
-                return True
-    return False
-
-
-def _probe_idempotency() -> HyIdempotencySupport:
-    if not hy_write_accepts_request_id():
-        # No handle at all: not even a correlation id to investigate later.
-        return HyIdempotencySupport.NONE
-    if _add_path_declares_idempotency():
-        return HyIdempotencySupport.KEY
-    # A request_id is present but nothing de-duplicates on it.
-    return HyIdempotencySupport.NONE
-
-
-def _probe_result_lookup() -> HyResultLookupSupport:
-    client_mod = _module("hy_memory.client")
-    client = getattr(client_mod, "HyMemoryClient", None) if client_mod is not None else None
-    if _has_callable(client, "get_write_outcome"):
-        return HyResultLookupSupport.BY_REQUEST_ID
-    if _inspector_can_lookup_by_request_id():
-        return HyResultLookupSupport.BY_REQUEST_ID
-    if _has_callable(client, "get"):
-        return HyResultLookupSupport.BY_MEMORY_ID_ONLY
-    return HyResultLookupSupport.NONE
-
-
-IDEMPOTENCY_SUPPORT: HyIdempotencySupport = _probe_idempotency()
-RESULT_LOOKUP_SUPPORT: HyResultLookupSupport = _probe_result_lookup()
-
-
-def probe_hy_write_idempotency() -> dict[str, Any]:
-    """Full probe report, for STATUS evidence and for debugging a bumped SDK.
-
-    Re-probed on every call so a dependency upgrade that changes the answer is
-    visible, rather than being frozen at import time.
-    """
-    client_mod = _module("hy_memory.client")
-    client = getattr(client_mod, "HyMemoryClient", None)
-    add = getattr(client, "add", None)
-    return {
-        "sdk_importable": client_mod is not None,
-        "add_signature": str(inspect.signature(add)) if add is not None else "",
-        "accepts_request_id": hy_write_accepts_request_id(),
-        "add_declares_idempotency": _add_path_declares_idempotency(),
-        "durable_write_outcome": (
-            client is not None and _has_callable(client, "get_write_outcome")
-        ),
-        "inspector_lookup_by_request_id": _inspector_can_lookup_by_request_id(),
-        "idempotency": _probe_idempotency().value,
-        "result_lookup": _probe_result_lookup().value,
-    }
@@ -2217,34 +2217,6 @@ CREATE TABLE IF NOT EXISTS retrieval_chunks (
 """


-def retrieval_documents_v2_ddl(dialect: str = "mysql") -> str:
-    """The v2 documents table under its temporary `_v2` name.
-
-    `index_store.migrate_v2` creates this, copies rows from the v1 table, drops
-    the v1 table and renames. The rename is the only reason a `_v2` name exists
-    at all, and it is derived from the single definition so the copy-across
-    cannot drift from what fresh databases get.
-    """
-    documents = next(
-        statement
-        for statement in _split_statements(retrieval_index_ddl(dialect))
-        if "CREATE TABLE IF NOT EXISTS retrieval_documents (" in statement
-    )
-    return documents.replace(
-        "CREATE TABLE IF NOT EXISTS retrieval_documents (",
-        "CREATE TABLE retrieval_documents_v2 (",
-    )
-
-
-def retrieval_build_maps_ddl(dialect: str = "mysql") -> str:
-    """The build-maps table added by v2."""
-    return next(
-        statement
-        for statement in _split_statements(retrieval_index_ddl(dialect))
-        if "retrieval_build_maps" in statement
-    )
-
-
 async def _033_retrieval_index(cursor: Any) -> None:
    for statement in _split_statements(retrieval_index_ddl(_dialect(cursor))):
        await cursor.execute(statement)
@@ -1,7 +1,7 @@
 """A3-01: the HTTP entry to the unified write path.

-`POST /api/v1/memory-writes` accepts a write and returns a request id;
-`GET /api/v1/memory-writes/{id}` reports its outcome.
+`POST /api/v2/memory-writes` accepts OpenAI-style messages and returns a
+request id; `GET /api/v2/memory-writes/{id}` reports its outcome.

 Two rules shape this endpoint, both consequences of A0-03's measurement that HY
 has no idempotency key:
@@ -45,17 +45,6 @@ from one_memory.memory_server.registry import WriteRegistry, WriteState
 from one_memory.memory_server.submit import WriteConflictError


-class MemoryWriteInput(BaseModel):
-    model_config = ConfigDict(extra="forbid")
-
-    internal_session_id: str = Field(min_length=1, max_length=128)
-    first_round: int = Field(ge=1)
-    last_round: int = Field(ge=1)
-    evidence_ids: list[str] = Field(min_length=1)
-    priority: Literal["live", "normal"] = "normal"
-    idempotency_key: str | None = Field(default=None, max_length=256)
-
-
 class MemoryWriteReceipt(BaseModel):
    request_id: str
    state: WriteState
@@ -121,67 +110,6 @@ def _registry(
    return cast(WriteRegistry | DurableWriteRegistry, registry)


-def router(dependencies: list[Any] | None = None) -> APIRouter:
-    instance = APIRouter(
-        prefix="/api/v1/memory-writes",
-        tags=["memory-writes"],
-        dependencies=dependencies or [],
-    )
-
-    @instance.post("", response_model=MemoryWriteReceipt, status_code=202)
-    def submit(payload: MemoryWriteInput, request: Request) -> MemoryWriteReceipt:
-        registry = _registry(request)
-        batch = BatchIdentity(
-            internal_session_id=payload.internal_session_id,
-            first_round=payload.first_round,
-            last_round=payload.last_round,
-        )
-        request_contract = WriteRequest(
-            internal_session_id=payload.internal_session_id,
-            batch_identity=batch,
-            evidence_ids=payload.evidence_ids,
-            idempotency_key=payload.idempotency_key or batch.batch_id,
-            priority=payload.priority,
-        )
-        try:
-            record, enqueued = registry.submit(request_contract)
-        except WriteConflictError as error:
-            raise HTTPException(409, str(error)) from error
-        except LookupError as error:
-            # Evidence-first: the write points at Evidence the business
-            # database does not hold. Rejecting it here is a caller error, not
-            # a server fault, and it must not create a queue row.
-            raise HTTPException(422, str(error)) from error
-        return MemoryWriteReceipt(
-            request_id=record.request_id,
-            state=record.state,
-            created_at=record.created_at,
-            updated_at=record.updated_at,
-            detail="" if enqueued else "already accepted; not enqueued again",
-        )
-
-    @instance.get("/{request_id}", response_model=MemoryWriteReceipt)
-    def status(request_id: str, request: Request) -> MemoryWriteReceipt:
-        registry = _registry(request, write=False)
-        try:
-            record = registry.get(request_id)
-        except DurableQueueSchemaUnavailableError as error:
-            raise HTTPException(503, str(error)) from error
-        if record is None:
-            raise HTTPException(404, "unknown request id")
-        return MemoryWriteReceipt(
-            request_id=record.request_id,
-            state=record.state,
-            created_at=record.created_at,
-            updated_at=record.updated_at,
-            detail=record.detail,
-            raw_ids=record.raw_ids,
-            memory_ids=record.memory_ids,
-        )
-
-    return instance
-
-
 def sdk_router(dependencies: list[Any] | None = None) -> APIRouter:
    instance = APIRouter(
        prefix="/api/v2/memory-writes",
@@ -64,7 +64,6 @@ EXEMPT_PATHS = frozenset(
        "memory_server/sqlite_semantics.py",
        "memory_tasks.py",
        "sdk_data_plane_migration.py",
-        "sdk_migration_inventory.py",
        "target_schema.py",
        # D93 A: the visibility cache is SQLite-only by construction. It is a
        # derived table with triggers on the six tables the read check reads —
@@ -9,9 +9,8 @@ Evidence→queue crash window stayed open.
 This adapter is deliberately small.  It turns one committed Evidence record
 into the one-record complete-turn request, submits it once under an evidence
 derived idempotency key, and returns the durable record.  The existing
-``DurableWriteDrain`` remains the only consumer; it calls
-``process_records``, which preserves the complete-turn/debounce/settle rules
-in ``MemoryProcessingSupervisor``.
+``DurableWriteDrain`` remains the only consumer; it calls the SDK write
+coordinator once and persists only the public receipt ids.
 """

 from __future__ import annotations
@@ -477,7 +477,7 @@ class MemoryTaskStore:
    #: Statuses whose input snapshot nothing reads back afterwards.
    #:
    #: Deliberately narrow. A `completed` task's inputs are read after the fact:
-    #: `memory_processing` compares them against the evidence actually sent to
+    #: SDK write processing compares them against the evidence actually sent to
    #: HY to prove coverage, and `matters/discovery` reads them to report how
    #: many records HY really saw. Clearing those would silently turn a
    #: verification into a pass and a report into a lie, which is far worse
@@ -10,7 +10,6 @@ from contextlib import contextmanager
 from contextvars import ContextVar
 from dataclasses import dataclass
 from datetime import UTC, datetime
-from importlib import import_module
 from pathlib import Path
 from typing import TypedDict
 from uuid import uuid4
@@ -46,14 +45,7 @@ def audit_precipitation_request(request_id: str) -> Iterator[None]:


 def _current_prompt_stage() -> str:
-    try:
-        resolver = getattr(  # noqa: B009 - optional SDK version boundary
-            import_module("hy_memory.utils.model_call_context"),
-            "current_model_call_stage",
-        )
-    except (ImportError, AttributeError):
-        return "other"
-    return str(resolver())
+    return "other"


 class ModelRequestAuditStore:
@@ -1,623 +0,0 @@
-"""Read-only process: serves the same read routes as Core without write access.
-
-Opens the database in `mode=ro` with `PRAGMA query_only=ON`; every store
-operation here goes through `main_database().read()`, which already enforces
-this. No HY SDK, no Chroma, no ONNX runtime, no keychain: the imports that
-would pull them in are simply not reached, and the routes below never call
-into a write path.
-
-Reads that need HY (semantic search, graph overlay) are delegated to the
-Core's internal endpoints; the Core is the only process that owns the
-vector stores. Side effects (call events, feedback) are queued and sent
-asynchronously, never blocking the read response.
-"""
-
-from __future__ import annotations
-
-import asyncio
-import builtins
-import contextlib
-import logging
-import os
-import sqlite3
-import time
-from collections.abc import AsyncIterator, Collection, Mapping, Sequence
-from contextlib import asynccontextmanager
-from datetime import UTC, datetime
-from pathlib import Path
-from typing import Any, cast
-
-from fastapi import Depends, FastAPI, HTTPException, Query, Request
-from pydantic import BaseModel, ConfigDict, SecretStr
-
-from one_memory.database_access import main_database
-from one_memory.home_summary import (
-    HomeSummary,
-    HomeSummaryUnavailable,
-    read_home_summary,
-)
-from one_memory.memory_governance import GovernedMemory, MemoryStatus
-from one_memory.memory_library import MemoryLibraryPage, read_memory_library
-from one_memory.recall import RecallRequest, _KeywordCandidates
-from one_memory.semantic_deadline import SemanticMemoryHit
-from one_memory.source_read_access import SourceReadAccess
-
-logger = logging.getLogger(__name__)
-
-MAX_READER_RSS_BYTES = 250 * 1024 * 1024
-
-_PROBE_QUERY = "__om_probe__"
-
-
-def _check_imports() -> None:
-    """Fail fast if any forbidden module is already loaded."""
-    forbidden = ("hy_memory", "chromadb", "onnxruntime", "keyring")
-    import sys
-
-    loaded = [name for name in forbidden if name in sys.modules]
-    if loaded:
-        raise ImportError(f"reader_process_loaded_forbidden_modules: {', '.join(loaded)}")
-
-
-def _memory_readable_sql() -> str:
-    return """
-    m.user_id = ? AND m.status = 'confirmed' AND m.deleted_at IS NULL
-    AND NOT EXISTS (
-        SELECT 1 FROM governed_memory_evidence link
-        JOIN evidence e ON e.evidence_id = link.evidence_id
-        JOIN source_grants g ON g.user_id=e.user_id AND g.source_id=e.source_id
-        WHERE link.memory_id=m.memory_id AND +link.active=1
-          AND g.revoked_at IS NOT NULL
-    )
-    AND NOT EXISTS (
-        SELECT 1 FROM governed_memory_evidence link
-        LEFT JOIN evidence e ON e.evidence_id = link.evidence_id
-        WHERE link.memory_id=m.memory_id AND +link.active=1
-          AND (e.evidence_id IS NULL OR e.user_id != m.user_id)
-    )
-    """
-
-
-class ReaderSettings(BaseModel):
-    model_config = ConfigDict(frozen=True)
-
-    data_dir: Path
-    host: str = "127.0.0.1"
-    port: int = 0  # 0 = ephemeral
-    api_token: SecretStr
-    core_port: int
-    core_api_token: str
-
-
-class SideEffectForwarder:
-    """Queues read-path side effects and forwards them to Core.
-
-    Reads in the Reader must stay read-only: the database is opened with
-    `query_only`, so a call event cannot be written here. Instead each read
-    that would have recorded one enqueues it, and a background task posts it
-    to Core's internal endpoint. The response never waits for it; a Core that
-    is unavailable drops it and counts the drop.
-    """
-
-    def __init__(self, settings: ReaderSettings) -> None:
-        self.settings = settings
-        self.queue: asyncio.Queue[dict[str, object]] = asyncio.Queue(maxsize=256)
-        self.dropped = 0
-        self.task: asyncio.Task[None] | None = None
-
-    def enqueue(self, event: dict[str, object]) -> None:
-        try:
-            self.queue.put_nowait(event)
-        except asyncio.QueueFull:
-            self.dropped += 1
-
-    def start(self) -> None:
-        if self.task is None or self.task.done():
-            self.task = asyncio.ensure_future(self._drain())
-
-    def stop(self) -> None:
-        if self.task is not None and not self.task.done():
-            self.task.cancel()
-
-    async def _drain(self) -> None:
-        import httpx
-
-        while True:
-            event = await self.queue.get()
-            try:
-                async with httpx.AsyncClient(
-                    headers={
-                        "Authorization": f"Bearer {self.settings.core_api_token}",
-                        "Content-Type": "application/json",
-                    },
-                    timeout=5.0,
-                    proxy=None,
-                ) as client:
-                    await client.post(
-                        f"http://127.0.0.1:{self.settings.core_port}/internal/call-event",
-                        json=event,
-                    )
-            except Exception:
-                # A dropped side effect is never a failed read: recall already
-                # answered. Counted so a persistently unavailable Core is
-                # visible rather than silent.
-                self.dropped += 1
-
-
-async def _core_get(
-    settings: ReaderSettings,
-    path: str,
-    timeout_s: float = 5.0,
-    params: Mapping[
-        str, str | int | float | bool | None | Sequence[str | int | float | bool | None]
-    ]
-    | None = None,
-) -> Any:
-    """Call the Core's internal read endpoint for data only Core can serve.
-
-    `params` is passed to httpx rather than interpolated into the path: a
-    query containing `&`, `#` or a space would otherwise be cut short or
-    read as a fragment, silently searching for the wrong thing.
-    """
-    import httpx
-
-    url = f"http://127.0.0.1:{settings.core_port}{path}"
-    async with httpx.AsyncClient(
-        headers={"Authorization": f"Bearer {settings.core_api_token}"},
-        timeout=timeout_s,
-        proxy=None,
-    ) as client:
-        response = await client.get(url, params=params)
-        response.raise_for_status()
-        return response.json()
-
-
-class _ReaderGovernance:
-    """Minimal read interface over the governance store for recall."""
-
-    def __init__(self, database_path: Path) -> None:
-        self.database_path = database_path
-
-    def list(
-        self,
-        *,
-        user_id: str,
-        status: MemoryStatus | None = None,
-        query: str | None = None,
-        limit: int = 100,
-        offset: int = 0,
-    ) -> list[GovernedMemory]:
-        """One page of memories, read through the Core's own read function.
-
-        `MemoryGovernanceStore.list` cannot be used directly: constructing the
-        store runs schema DDL on the write connection, and the Reader opens
-        with `query_only`. `read_memory_library` is the same read expressed as
-        a static SQL read (revoked sources excluded, per-row freshness checked
-        for the rows it returns, cheap total for the rest), so the Reader and
-        the Core cannot drift on what "readable" means.
-        """
-        page = read_memory_library(
-            self.database_path,
-            user_id=user_id,
-            query=query,
-            status=status,
-            offset=offset,
-            limit=limit,
-        )
-        return list(page.items)
-
-    def find_confirmed_matching_terms(
-        self,
-        *,
-        user_id: str,
-        terms: Collection[str],
-        limit: int = 200,
-        scopes: Collection[str] | None = None,
-    ) -> tuple[builtins.list[GovernedMemory], bool]:
-        if not 1 <= limit <= 2_000:
-            raise ValueError("invalid candidate limit")
-        cleaned = tuple(dict.fromkeys(term.strip() for term in terms if term.strip()))
-        if not cleaned or scopes is not None and not scopes:
-            return [], False
-        clauses = ["m.user_id=?", "m.status='confirmed'", "m.deleted_at IS NULL"]
-        values: builtins.list[object] = [user_id]
-        matches = []
-        for term in cleaned[:32]:
-            matches.append("m.content LIKE ? ESCAPE '\\'")
-            escaped = term.replace("\\", "\\\\").replace("%", "\\%").replace("_", "\\_")
-            values.append(f"%{escaped}%")
-        clauses.append(f"({' OR '.join(matches)})")
-        if scopes is not None:
-            clauses.append(
-                "EXISTS (SELECT 1 FROM json_each(m.scopes_json) "
-                f"WHERE json_each.value IN ({','.join('?' for _ in scopes)}))"
-            )
-            values.extend(scopes)
-        values.append(limit + 1)
-        with main_database(self.database_path).read(row_factory=sqlite3.Row) as db:
-            rows = db.execute(
-                f"SELECT m.* FROM governed_memories m WHERE {' AND '.join(clauses)} "
-                "ORDER BY m.updated_at DESC,m.memory_id LIMIT ?",
-                values,
-            ).fetchall()
-            items = self._memories_from_rows(db, rows[:limit])
-        return items, len(rows) > limit or len(cleaned) > 32
-
-    def get(self, memory_id: str, *, user_id: str) -> GovernedMemory | None:
-        memory = _get_memory_direct(self.database_path, memory_id, user_id)
-        return memory if isinstance(memory, GovernedMemory) else None
-
-    def sources_behind_memories(
-        self,
-        *,
-        user_id: str,
-        memory_ids: Collection[str],
-    ) -> dict[str, frozenset[str]]:
-        if not memory_ids:
-            return {}
-        identifiers = tuple(memory_ids)
-        placeholders = ",".join("?" for _ in identifiers)
-        with main_database(self.database_path).read(row_factory=sqlite3.Row) as db:
-            rows = db.execute(
-                "SELECT link.memory_id,e.source_id FROM governed_memory_evidence link "
-                "JOIN evidence e ON e.evidence_id=link.evidence_id "
-                f"WHERE link.active=1 AND e.user_id=? AND link.memory_id IN ({placeholders})",
-                (user_id, *identifiers),
-            ).fetchall()
-        found: dict[str, set[str]] = {}
-        for row in rows:
-            found.setdefault(str(row["memory_id"]), set()).add(str(row["source_id"]))
-        return {key: frozenset(values) for key, values in found.items()}
-
-    @staticmethod
-    def _memories_from_rows(
-        db: sqlite3.Connection, rows: Sequence[sqlite3.Row]
-    ) -> builtins.list[GovernedMemory]:
-        from one_memory.memory_governance import MemoryGovernanceStore
-
-        return MemoryGovernanceStore.memories_from_rows(db, builtins.list(rows))
-
-
-def _conversation_hit(row: Any) -> dict[str, Any]:
-    """One stored row as a search hit.
-
-    The author travels with the row. A client that is not told who wrote a row
-    has to invent an author, and the invention is always "the user", which
-    files an assistant turn under the person who was talking to it.
-    """
-    return {
-        "evidence_id": row["evidence_id"],
-        "conversation_id": row["conversation_id"],
-        "source_id": row["source_id"],
-        "workspace": row["workspace"],
-        "project": row["project"],
-        "created_at": row["created_at"],
-        "excerpt": row["verbatim"][:200] if row["verbatim"] else "",
-        # Both columns are NOT NULL on `evidence`, so a stored row always
-        # knows; a summary of a whole conversation has no single author and
-        # never reaches here.
-        "role": row["role"],
-        "written_by": row["written_by"],
-    }
-
-
-def create_reader_app(settings: ReaderSettings) -> FastAPI:
-    """A read-only FastAPI application for the Reader process."""
-    _check_imports()
-    database_path = settings.data_dir / "one-memory.sqlite3"
-    if not database_path.exists():
-        raise FileNotFoundError(f"reader_database_missing: {database_path}")
-
-    # Validate schema version without writing.
-    from one_memory.memory_server.migrations import SCHEMA_TABLE
-    from one_memory.schema_gate import BUILD_SCHEMA_VERSION
-
-    with main_database(database_path).read() as db:
-        tables = {
-            str(row[0]) for row in db.execute("SELECT name FROM sqlite_master WHERE type='table'")
-        }
-        if SCHEMA_TABLE in tables:
-            row = db.execute(f"SELECT MAX(version) FROM {SCHEMA_TABLE}").fetchone()
-            found = int(row[0]) if row and row[0] is not None else 0
-            if found > BUILD_SCHEMA_VERSION:
-                raise RuntimeError(f"reader_schema_too_new: {found} > {BUILD_SCHEMA_VERSION}")
-        elif not tables:
-            raise RuntimeError("reader_database_empty")
-
-    access = SourceReadAccess(database_path)
-    forwarder = SideEffectForwarder(settings)
-    governance = _ReaderGovernance(database_path)
-
-    @asynccontextmanager
-    async def lifespan(_: FastAPI) -> AsyncIterator[None]:
-        monitor = asyncio.get_running_loop().create_task(_memory_monitor())
-        forwarder.start()
-        yield
-        monitor.cancel()
-        forwarder.stop()
-
-    async def _memory_monitor() -> None:
-        import resource
-        import sys
-
-        while True:
-            await asyncio.sleep(30)
-            usage = resource.getrusage(resource.RUSAGE_SELF).ru_maxrss
-            rss_bytes = usage if sys.platform == "darwin" else usage * 1024
-            if rss_bytes > MAX_READER_RSS_BYTES:
-                logger.error(
-                    "reader_memory_limit_exceeded: %d bytes > %d",
-                    rss_bytes,
-                    MAX_READER_RSS_BYTES,
-                )
-                os._exit(1)
-
-    app = FastAPI(title="MemHub Reader", lifespan=lifespan)
-    app.state.settings = settings
-    app.state.database_path = database_path
-
-    expected_token = settings.api_token.get_secret_value()
-
-    def require_auth(request: Request) -> None:
-        import hmac
-
-        origin = request.headers.get("origin")
-        if origin is not None and origin not in (
-            "one-memory://app",
-            "http://localhost:5173",
-        ):
-            raise HTTPException(403, "origin_not_allowed")
-        authorization = request.headers.get("authorization", "")
-        scheme, _, token = authorization.partition(" ")
-        if not (scheme.lower() == "bearer" and hmac.compare_digest(token, expected_token)):
-            raise HTTPException(401, "invalid_local_token")
-
-    protected = [Depends(require_auth)]
-
-    @app.get("/api/v1/home-summary", response_model=HomeSummary, dependencies=protected)
-    def home_summary() -> HomeSummary:
-        try:
-            return read_home_summary(database_path)
-        except HomeSummaryUnavailable:
-            # Belt and braces: read_home_summary answers stale itself, but if
-            # that ever changes, never let the homepage 500.
-            return HomeSummary(
-                memory_count=0, latest=None, observed_at=datetime.now(UTC), stale=True
-            )
-
-    @app.get("/api/v1/memory-library", response_model=MemoryLibraryPage, dependencies=protected)
-    def memory_library(
-        query: str | None = Query(default=None, max_length=512),
-        status: str | None = None,
-        active_only: bool = False,
-        source: str | None = Query(default=None, min_length=1),
-        offset: int = Query(default=0, ge=0),
-        limit: int = Query(default=100, ge=1, le=500),
-    ) -> MemoryLibraryPage:
-        from one_memory.memory_governance import MemoryStatus as MS
-
-        ms = MS(status) if status else None
-        return read_memory_library(
-            database_path,
-            user_id="local",
-            query=query,
-            status=ms,
-            active_only=active_only,
-            source=source,
-            offset=offset,
-            limit=limit,
-        )
-
-    @app.get("/api/v1/memories", dependencies=protected)
-    def list_memories(
-        status: str | None = None,
-        query: str | None = None,
-        limit: int = Query(default=100, ge=1, le=500),
-        offset: int = Query(default=0, ge=0),
-    ) -> list[dict[str, Any]]:
-        from one_memory.memory_governance import MemoryStatus as MS
-
-        ms = MS(status) if status else None
-        memories = governance.list(
-            user_id="local", status=ms, query=query, limit=limit, offset=offset
-        )
-        visible = access.memories(memories)
-        return [m.model_dump(mode="json") for m in visible]
-
-    @app.get("/api/v1/memories/graph", dependencies=protected)
-    async def memory_graph(
-        limit: int = Query(default=160, ge=1, le=300),
-    ) -> dict[str, Any]:
-        memories = governance.list(user_id="local", status=MemoryStatus.CONFIRMED, limit=500)
-        projected = access.memories(memories[:limit])
-        nodes: dict[str, dict[str, Any]] = {}
-        edges: list[dict[str, Any]] = []
-        source_ids: set[str] = set()
-        with main_database(database_path).read(row_factory=sqlite3.Row) as db:
-            for memory in projected:
-                node_id = f"memory:{memory.memory_id}"
-                nodes[node_id] = {
-                    "node_id": node_id,
-                    "node_type": "memory",
-                    "label": memory.content,
-                    "memory_id": memory.memory_id,
-                    "kind": memory.kind.value
-                    if hasattr(memory.kind, "value")
-                    else str(memory.kind),
-                    "status": memory.status.value
-                    if hasattr(memory.status, "value")
-                    else str(memory.status),
-                }
-                evidence_ids = list(memory.evidence_ids)
-                if evidence_ids:
-                    placeholders = ",".join("?" for _ in evidence_ids)
-                    sql = (
-                        f"SELECT DISTINCT source_id FROM evidence"
-                        f" WHERE evidence_id IN ({placeholders})"
-                    )
-                    for row in db.execute(
-                        sql,
-                        evidence_ids,
-                    ):
-                        sid = str(row["source_id"])
-                        source_ids.add(sid)
-            for sid in sorted(source_ids):
-                source_node_id = f"source:{sid}"
-                nodes[source_node_id] = {
-                    "node_id": source_node_id,
-                    "node_type": "source",
-                    "label": sid,
-                    "source_id": sid,
-                }
-                for memory in projected:
-                    if any(_source_for_evidence(db, eid) == sid for eid in memory.evidence_ids):
-                        memory_node_id = f"memory:{memory.memory_id}"
-                        edges.append(
-                            {
-                                "edge_id": f"{source_node_id}->{memory_node_id}",
-                                "source_node_id": source_node_id,
-                                "target_node_id": memory_node_id,
-                            }
-                        )
-        # Fetch HY graph overlay from Core (not available in Reader).
-        hy_overlay = None
-        with contextlib.suppress(Exception):
-            hy_overlay = await _core_get(
-                settings,
-                "/internal/graph-snapshot",
-                timeout_s=5.0,
-                params={"limit": limit},
-            )
-        if hy_overlay and not hy_overlay.get("consulted", False):
-            hy_overlay = None
-        return {
-            "limit": limit,
-            "total": len(projected),
-            "nodes": tuple(nodes.values()),
-            "edges": edges,
-            **({"hy_overlay": hy_overlay} if hy_overlay else {}),
-        }
-
-    @app.get("/api/v1/memories/{memory_id}", dependencies=protected)
-    def get_memory(memory_id: str) -> dict[str, Any]:
-        memory = _get_memory_direct(database_path, memory_id, "local")
-        if memory is None:
-            raise HTTPException(404, "memory_not_found")
-        if not access.memories([memory]):
-            raise HTTPException(403, "memory_source_revoked")
-        return cast(dict[str, Any], memory.model_dump(mode="json"))
-
-    @app.post("/api/v1/recall", response_model=None, dependencies=protected)
-    async def recall(request: RecallRequest) -> dict[str, Any]:
-        from one_memory.recall import GovernedRecallTarget
-
-        started = time.monotonic()
-        target = GovernedRecallTarget(
-            governance=governance,
-            semantic_search=None,
-            resolve_workspace=None,
-            filter_memories=access.memories,
-        )
-
-        if request.query == _PROBE_QUERY:
-            return target.recall(request, user_id="local").model_dump(mode="json")
-        deadline = time.monotonic() + request.timeout_ms / 1_000
-        semantic_ids: tuple[str, ...] = ()
-        try:
-            remaining = max(0.1, deadline - time.monotonic())
-            result = await asyncio.wait_for(
-                _core_get(
-                    settings,
-                    "/internal/semantic-ids",
-                    timeout_s=remaining,
-                    params={"query": request.query, "limit": request.limit},
-                ),
-                timeout=remaining,
-            )
-            semantic_ids = tuple(str(value) for value in result.get("ids", []) if value)
-        except Exception:
-            pass
-        readable = target._readable_scopes(request)
-        # Keyword candidates from read-only SQL.
-        keywords = target._keyword_candidates(
-            request, user_id="local", readable=readable, deadline=deadline
-        )
-        receipt = target._rank(
-            request,
-            user_id="local",
-            readable=readable,
-            semantic_hits={
-                memory_id: SemanticMemoryHit(
-                    memory_id,
-                    1.0 - index / max(1, len(semantic_ids)),
-                )
-                for index, memory_id in enumerate(semantic_ids)
-            },
-            keywords=keywords
-            if isinstance(keywords, _KeywordCandidates)
-            else _KeywordCandidates((), False, False),
-            deadline=deadline,
-            timed_out=False,
-        )
-        forwarder.enqueue(
-            {
-                "tool": "om_recall",
-                "query": request.query,
-                "scope": request.scope,
-                "state": receipt.state.value,
-                "memory_count": len(receipt.hits),
-                "elapsed_ms": (time.monotonic() - started) * 1000,
-            }
-        )
-        return receipt.model_dump(mode="json")
-
-    @app.get("/api/v1/conversations/search", dependencies=protected)
-    def search_conversations(
-        query: str = Query(min_length=1, max_length=512),
-        limit: int = Query(default=20, ge=1, le=100),
-    ) -> dict[str, Any]:
-        if not database_path.exists():
-            raise HTTPException(503, "conversation_search_unavailable")
-        with main_database(database_path).read(row_factory=sqlite3.Row) as db:
-            try:
-                rows = db.execute(
-                    """SELECT * FROM evidence
-                    WHERE user_id = 'local' AND verbatim LIKE ? ESCAPE '\\'
-                    ORDER BY imported_at DESC LIMIT ?""",
-                    (f"%{query}%", limit),
-                ).fetchall()
-            except Exception:
-                raise HTTPException(503, "conversation_search_unavailable") from None
-        hits = [_conversation_hit(row) for row in rows]
-        return {
-            "hits": hits,
-            "empty_reason": None if hits else "no_matches",
-            "imported_messages": len(hits),
-        }
-
-    @app.get("/health")
-    def health() -> dict[str, str]:
-        return {"status": "ok", "service": "one-memory-reader"}
-
-    return app
-
-
-def _get_memory_direct(database_path: Path, memory_id: str, user_id: str) -> Any:
-    from one_memory.memory_governance import _memory_from_row, _memory_links
-
-    with main_database(database_path).read(row_factory=sqlite3.Row) as db:
-        row = db.execute(
-            "SELECT * FROM governed_memories WHERE memory_id = ? AND user_id = ?",
-            (memory_id, user_id),
-        ).fetchone()
-        if row is None:
-            return None
-        links = _memory_links(db, [memory_id])
-        return _memory_from_row(row, *(tuple(vals) for vals in links[memory_id]))
-
-
-def _source_for_evidence(db: sqlite3.Connection, evidence_id: str) -> str | None:
-    row = db.execute(
-        "SELECT source_id FROM evidence WHERE evidence_id = ?", (evidence_id,)
-    ).fetchone()
-    return str(row["source_id"]) if row else None
@@ -12,7 +12,7 @@ from typing import Any, Literal, Protocol

 from pydantic import BaseModel, ConfigDict, Field

-from one_memory.hy_memory_bridge import HySdkMemoryRecord
+from one_memory.hy_sdk_types import HySdkMemoryRecord
 from one_memory.memory_governance import GovernedMemory, MemoryStatus
 from one_memory.semantic_deadline import (
    SemanticMemoryHit,
@@ -81,7 +81,11 @@ class RecallHit(BaseModel):
    score: float = Field(ge=0, le=1)
    #: HY already rendered memory time/evolution/capsule context into content.
    #: Hosts must not append their own governance timestamp in this case.
-    content_format: Literal["plain", "hy-sdk"] = "plain"
+    #: "raw" is an excerpt of an imported conversation, not a memory.
+    content_format: Literal["plain", "hy-sdk", "raw"] = "plain"
+    #: Raw excerpts only: the app and conversation the excerpt came from.
+    source: str | None = None
+    session_id: str | None = None


 class RecallReceipt(BaseModel):
@@ -103,6 +107,11 @@ class RecallReceipt(BaseModel):
    #: whether the answer is that nothing was found or that reading was not
    #: allowed.
    excluded_sources: tuple[str, ...] = ()
+    #: "raw" when recall searched the imported conversations instead of
+    #: memories, because no generation model is configured or no memory
+    #: exists yet (`fallback_reason`). Raw hits are excerpts, not memories.
+    mode: Literal["memory", "raw"] = "memory"
+    fallback_reason: Literal["model_unavailable", "no_memories"] | None = None


 class RecallTarget(Protocol):
@@ -0,0 +1,149 @@
+"""Recall that degrades to the original conversations instead of answering empty.
+
+MemHub_recall searches the memories the SDK distilled from conversations. That
+needs a configured generation model to have produced any, and a library that
+holds some. Without either, the memory search can only come back empty, so
+recall searches the SDK's raw layer (the imported conversations themselves)
+instead: plain retrieval over what was imported, the same lane
+`/api/v2/context/search` reads, under the same raw-source authority.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import time
+from collections.abc import Awaitable, Callable
+from datetime import UTC, datetime
+from typing import Literal
+
+from one_memory.hy_sdk_types import HySdkRawHit
+from one_memory.recall import (
+    PROBE_QUERY,
+    RecallHit,
+    RecallReceipt,
+    RecallRequest,
+    RecallState,
+    TruncationBudget,
+)
+
+FallbackReason = Literal["model_unavailable", "no_memories"]
+
+#: Raw sources this caller may read, or None for every source (a trusted local
+#: caller). An empty set means the caller may read no raw conversation.
+RawSources = frozenset[str] | None
+
+
+async def recall_with_fallback(
+    request: RecallRequest,
+    *,
+    model_ready: Callable[[], Awaitable[bool]],
+    recall_memories: Callable[[RecallRequest], Awaitable[RecallReceipt]],
+    has_memories: Callable[[], Awaitable[bool]],
+    raw_sources: Callable[[], Awaitable[RawSources]],
+    search_raw: Callable[[str, int, RawSources], Awaitable[tuple[HySdkRawHit, ...]]],
+) -> RecallReceipt:
+    """Memory recall when a model and memories exist, otherwise the raw layer.
+
+    The library check runs only after an empty memory search: with memories
+    present and nothing relevant, the answer stays an honest empty memory
+    result rather than a different kind of hit.
+    """
+    if request.query == PROBE_QUERY:
+        return await recall_memories(request)
+    started = time.monotonic()
+    if await model_ready():
+        receipt = await recall_memories(request)
+        if receipt.hits or await has_memories():
+            return receipt
+        reason: FallbackReason = "no_memories"
+    else:
+        reason = "model_unavailable"
+    remaining = request.timeout_ms / 1_000 - (time.monotonic() - started)
+    return await _raw_recall(
+        request,
+        reason=reason,
+        remaining_seconds=remaining,
+        raw_sources=raw_sources,
+        search_raw=search_raw,
+    )
+
+
+async def _raw_recall(
+    request: RecallRequest,
+    *,
+    reason: FallbackReason,
+    remaining_seconds: float,
+    raw_sources: Callable[[], Awaitable[RawSources]],
+    search_raw: Callable[[str, int, RawSources], Awaitable[tuple[HySdkRawHit, ...]]],
+) -> RecallReceipt:
+    def receipt(
+        hits: tuple[RecallHit, ...] = (),
+        *,
+        used: int = 0,
+        truncated_by: TruncationBudget | None = None,
+        unexamined: int = 0,
+    ) -> RecallReceipt:
+        return RecallReceipt(
+            state=RecallState.GROUNDED if hits else RecallState.EMPTY,
+            query=request.query,
+            scope=request.scope,
+            hits=hits,
+            used_characters=used,
+            truncated=truncated_by is not None,
+            truncated_by=truncated_by,
+            unexamined_count=unexamined,
+            mode="raw",
+            fallback_reason=reason,
+        )
+
+    sources = await raw_sources()
+    if sources is not None and not sources:
+        # The caller may read no original conversation: an empty answer, not
+        # a wider grant than recall had.
+        return receipt()
+    if remaining_seconds <= 0:
+        return receipt(truncated_by=TruncationBudget.TIMEOUT)
+    try:
+        found = await asyncio.wait_for(
+            search_raw(request.query, request.limit, sources),
+            timeout=remaining_seconds,
+        )
+    except TimeoutError:
+        return receipt(truncated_by=TruncationBudget.TIMEOUT)
+    hits: list[RecallHit] = []
+    used = 0
+    truncated_by: TruncationBudget | None = None
+    for index, hit in enumerate(found):
+        if sources is not None and hit.source not in sources:
+            continue
+        if len(hits) >= request.limit:
+            truncated_by = TruncationBudget.COUNT
+            return receipt(
+                tuple(hits),
+                used=used,
+                truncated_by=truncated_by,
+                unexamined=len(found) - index,
+            )
+        if used + len(hit.content) > request.character_budget:
+            truncated_by = TruncationBudget.CHARACTER
+            continue
+        hits.append(
+            RecallHit(
+                memory_id=hit.raw_id,
+                content=hit.content,
+                kind="raw",
+                scopes=(),
+                evidence_ids=(),
+                updated_at=(
+                    datetime.fromtimestamp(hit.memory_at, tz=UTC)
+                    if hit.memory_at is not None
+                    else datetime(1970, 1, 1, tzinfo=UTC)
+                ),
+                score=hit.score,
+                content_format="raw",
+                source=hit.source,
+                session_id=hit.session_id,
+            )
+        )
+        used += len(hit.content)
+    return receipt(tuple(hits), used=used, truncated_by=truncated_by)
@@ -22,6 +22,13 @@ class RetiredEvidenceStore:
        del limit
        return 0

+    def release_sdk_staging(self, evidence_ids: tuple[str, ...]) -> int:
+        # The write drain releases staged Evidence after every SDK write. On
+        # SDK authority nothing is staged; without this, `__getattr__` raised
+        # past `getattr(..., None)` and every completed write logged an error.
+        del evidence_ids
+        return 0
+
    def get(self, _identifier: str) -> None:
        return None

@@ -1,167 +0,0 @@
-"""Single-Evidence identity chunking; no reads, persistence, normalization or grants.
-
-Callers must first establish eligibility/sanitization/permission. This module
-does not strip wrappers or secrets and must not be used as a content safety gate.
-"""
-
-from dataclasses import dataclass
-from typing import Protocol
-
-from one_memory.embedding_assets import BUILTIN_EMBEDDING_MAX_TOKENS
-from one_memory.local_embedding import TokenizedText, TokenOffset
-
-
-class OffsetTokenizer(Protocol):
-    def tokenize_with_offsets(self, text: str) -> TokenizedText: ...
-
-
-class ChunkingError(ValueError):
-    """Invalid offsets, uncovered text or insufficient encoding budget; never truncate."""
-
-
-@dataclass(frozen=True)
-class TextChunk:
-    ordinal: int
-    start_cp: int
-    end_cp: int
-    token_count: int
-    text: str
-
-
-@dataclass(frozen=True)
-class _Span:
-    start: int
-    end: int
-    tokens: int
-
-
-def _spans(text: str, encoded: TokenizedText) -> tuple[_Span, ...]:
-    if not isinstance(encoded, TokenizedText) or len(encoded.token_ids) != len(encoded.offsets):
-        raise ChunkingError("chunk_token_offset_count_mismatch")
-    spans: list[_Span] = []
-    previous_start = previous_end = 0
-    for token_id, offset in zip(encoded.token_ids, encoded.offsets, strict=True):
-        if type(token_id) is not int or token_id < 0 or not isinstance(offset, TokenOffset):
-            raise ChunkingError("chunk_invalid_token_offset")
-        if (
-            type(offset.start_cp) is not int
-            or type(offset.end_cp) is not int
-            or type(offset.special) is not bool
-        ):
-            raise ChunkingError("chunk_invalid_token_offset")
-        start, end = offset.start_cp, offset.end_cp
-        if offset.special:
-            if start != 0 or end != 0:
-                raise ChunkingError("chunk_invalid_special_offset")
-            continue
-        if not 0 <= start < end <= len(text) or start < previous_start or end < previous_end:
-            raise ChunkingError("chunk_invalid_offset_order_or_bounds")
-        if start > previous_end and not text[previous_end:start].isspace():
-            raise ChunkingError("chunk_uncovered_nonwhitespace")
-        # Overlapping subpieces (including several pieces for one code point)
-        # form an indivisible interval; slicing inside it loses token identity.
-        if spans and start < spans[-1].end:
-            prior = spans[-1]
-            spans[-1] = _Span(prior.start, end, prior.tokens + 1)
-        else:
-            spans.append(_Span(start, end, 1))
-        previous_start, previous_end = start, end
-    if text[previous_end:] and not text[previous_end:].isspace():
-        raise ChunkingError("chunk_uncovered_nonwhitespace")
-    return tuple(spans)
-
-
-def _encoding(text: str, tokenizer: OffsetTokenizer) -> tuple[TokenizedText, tuple[_Span, ...]]:
-    try:
-        result = tokenizer.tokenize_with_offsets(text)
-    except Exception as error:
-        raise ChunkingError("chunk_tokenization_failed") from error
-    return result, _spans(text, result)
-
-
-def chunk_evidence(
-    text: str,
-    tokenizer: OffsetTokenizer,
-    *,
-    prefix: str = "",
-    target_tokens: int = 384,
-    overlap_tokens: int = 64,
-    max_tokens: int = BUILTIN_EMBEDDING_MAX_TOKENS,
-) -> tuple[TextChunk, ...]:
-    """Chunk one original with code point [start,end), preserving exact slices.
-
-    target/overlap count original non-special tokens. Overlap is up to the
-    requested token count at indivisible offset boundaries; forward progress is
-    mandatory. token_count is re-tokenized prefix + chunk.text INCLUDING specials,
-    never an additive estimate. Prefix is caller-provided, unchanged, and omitted
-    from chunk.text/offsets. Empty/whitespace-only inputs have no index chunks.
-    """
-    if not isinstance(text, str) or not isinstance(prefix, str):
-        raise ChunkingError("chunk_requires_text")
-    if (
-        any(type(value) is not int for value in (target_tokens, overlap_tokens, max_tokens))
-        or not 0 <= overlap_tokens < target_tokens
-        or not 1 <= max_tokens <= BUILTIN_EMBEDDING_MAX_TOKENS
-        or target_tokens > max_tokens
-    ):
-        raise ChunkingError("chunk_invalid_token_budget")
-    if not text or text.isspace():
-        return ()
-    _, spans = _encoding(text, tokenizer)
-    if not spans:
-        raise ChunkingError("chunk_uncovered_nonwhitespace")
-    chunks: list[TextChunk] = []
-    start_index = 0
-    previous_end_cp = 0
-    while start_index < len(spans):
-        end_index = start_index
-        body_tokens = 0
-        while end_index < len(spans):
-            candidate = body_tokens + spans[end_index].tokens
-            if end_index > start_index and candidate > target_tokens:
-                break
-            body_tokens = candidate
-            end_index += 1
-        # Prefer the last paragraph boundary in the latter half of the target.
-        # Never trade away an entire overlap window or split a token interval.
-        if end_index < len(spans):
-            count = 0
-            preferred = None
-            for boundary in range(start_index + 1, end_index + 1):
-                count += spans[boundary - 1].tokens
-                if count >= max(target_tokens // 2, overlap_tokens + 1):
-                    left, right = spans[boundary - 1].end, spans[boundary].start
-                    if "\n" in text[max(spans[boundary - 1].start, left - 1) : right]:
-                        preferred = boundary
-            if preferred is not None:
-                end_index = preferred
-        start_cp = 0 if start_index == 0 else spans[start_index].start
-        while end_index > start_index:
-            # Include gaps verbatim, including whitespace and combining marks;
-            # source coverage is never reconstructed by joining decoded tokens.
-            end_cp = spans[end_index].start if end_index < len(spans) else len(text)
-            encoded, _ = _encoding(prefix + text[start_cp:end_cp], tokenizer)
-            token_count = len(encoded.token_ids)
-            if token_count <= max_tokens:
-                break
-            end_index -= 1
-        else:
-            raise ChunkingError("chunk_indivisible_span_exceeds_token_budget")
-        if end_cp <= previous_end_cp or start_cp > previous_end_cp:
-            raise ChunkingError("chunk_coverage_cannot_progress")
-        chunks.append(TextChunk(len(chunks), start_cp, end_cp, token_count, text[start_cp:end_cp]))
-        previous_end_cp = end_cp
-        if end_index == len(spans):
-            break
-        next_start = end_index
-        overlap = 0
-        while next_start > start_index + 1:
-            candidate = overlap + spans[next_start - 1].tokens
-            if candidate > overlap_tokens:
-                break
-            next_start -= 1
-            overlap = candidate
-        start_index = next_start
-    if previous_end_cp != len(text):
-        raise ChunkingError("chunk_incomplete_coverage")
-    return tuple(chunks)