feat(memhub): add authenticated generic local MCP read connection

* fix: prevent authority-invalid HY retries

Co-authored-by: Cursor <cursoragent@cursor.com>

* fix: align coding authority with conversations

Co-authored-by: Cursor <cursoragent@cursor.com>

---------

Co-authored-by: alvinfei <alvinfei@tencent.com>
Co-authored-by: Cursor <cursoragent@cursor.com>

CHANGELOG.md

@@ -2,8 +2,37 @@
 
 Hy Memory and MemHub are versioned and released independently.
 
+## Hy Memory 2.0.3 — 2026-10-17
+
+- Added an optional runtime HTTP transport factory so embedded products can
+  enforce request budgets without wrapping or restricting SDK local-store
+  access.
+- Published
+  [`hy-memory==2.0.3`](https://pypi.org/project/hy-memory/2.0.3/) with
+  unchanged memory extraction, reconciliation, storage, and Coding/Working
+  behavior.
+
 ## Unreleased
 
+- 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.
@@ -26,7 +55,21 @@ 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 1.0.2 — Unreleased
+## MemHub 1.0.3 — Unreleased
+
+- Updated the source version to `1.0.3` (build 7).
+- Pinned the public SDK to `hy-memory==2.0.3`.
+- Removed the HyPermission store wrappers and global model-content consent
+  switch; the trusted embedded SDK now uses native local stores.
+- Prevented uncertain writes from blocking later source work and added one
+  reconciled automatic retry after ten seconds.
+- Added product-facing memory pipeline counters and the log troubleshooting
+  runbook.
+- Allowed a verified packaged Desktop Runtime from a different commit to take
+  over an equal-build install atomically while preserving downgrade and
+  tamper refusal.
+
+## MemHub 1.0.2 — 2026-10-17
 
 - Updated the source version to `1.0.2` (build 6).
 - Pinned the public SDK to `hy-memory==2.0.2`.

README.md

@@ -32,7 +32,8 @@ English | [中文](README.zh.md)
 
 ## Latest News
 
-* *<mark>[new]</mark>* 🔥🔥🔥 **[2026-10]**: 🚀 **Released** **Hy Memory 2.0 (`hy-memory==2.0.2`)** on [PyPI](https://pypi.org/project/hy-memory/2.0.2/) and open-sourced the **Hy-Memory** monorepo together with **[MemHub](memhub/README.md)**!
+* *<mark>[new]</mark>* 🔥🔥🔥 **[2026-10]**: 📦 **Released [MemHub v1.0.4](https://github.com/Tencent-Hunyuan/Hy-Memory/releases/tag/v1.0.4)** and **[memhub-cli 1.0.5](https://www.npmjs.com/package/memhub-cli/v/1.0.5)**, with a signed macOS desktop app and verified macOS/Linux Runtimes!
+* *<mark>[new]</mark>* 🔥🔥🔥 **[2026-10]**: 🚀 **Released** **Hy Memory 2.0 (`hy-memory==2.0.4`)** on [PyPI](https://pypi.org/project/hy-memory/2.0.4/) and open-sourced the **Hy-Memory** monorepo together with **[MemHub](memhub/README.md)**!
 * *<mark>[new]</mark>* 🔥🔥🔥 **[2026-10]**: 🚀 **Open-sourced** **[Hy-Memory-30B-A3B](https://huggingface.co/tencent/Hy-Memory-30B-A3B)** on Hugging Face along with its [technical report](https://arxiv.org/abs/2610.22196) on attributed reinforcement learning for long-term memory writing!
 * *<mark>[new]</mark>* 🔥🔥 **[2026-09]**: 📊 **Released** **[CMB: A Coding Memory Benchmark for Long-Horizon and Interleaved Coding Agents](docs/papers/CMB_A_Coding_Memory_Benchmark_for_Long_Horizon_and_Interleaved_Coding_Agents.pdf)** (submitted to **ICLR 2027**).
 * *<mark>[new]</mark>* 🔥🔥 **[2026-08]**: 🎉 **Accepted** at **EMNLP 2026**: **[Memory Beyond Recall: A Dual-Process Cognitive Memory System for Self-Evolving LLM Agents](https://arxiv.org/abs/2606.09483)**!
@@ -63,13 +64,13 @@ Welcome to the **Hy Memory Research Family**, covering cognitive chat memory, lo
 ### 1. Installation
 
 ```bash
-pip install hy-memory==2.0.2              # core (includes ChromaDB and Kuzu)
-pip install "hy-memory[qdrant]==2.0.2"    # + Qdrant
-pip install "hy-memory[faiss]==2.0.2"     # + FAISS
-pip install "hy-memory[graph]==2.0.2"     # + Neo4j
-pip install "hy-memory[mysql]==2.0.2"     # + MySQL cache
-pip install "hy-memory[rerank]==2.0.2"    # + local multilingual cross-encoder rerank
-pip install "hy-memory[all]==2.0.2"       # everything
+pip install hy-memory==2.0.3              # core (includes ChromaDB and Kuzu)
+pip install "hy-memory[qdrant]==2.0.3"    # + Qdrant
+pip install "hy-memory[faiss]==2.0.3"     # + FAISS
+pip install "hy-memory[graph]==2.0.3"     # + Neo4j
+pip install "hy-memory[mysql]==2.0.3"     # + MySQL cache
+pip install "hy-memory[rerank]==2.0.3"    # + local multilingual cross-encoder rerank
+pip install "hy-memory[all]==2.0.3"       # everything
 ```
 
 ### 2. Configuration

README.zh.md

@@ -32,7 +32,8 @@
 
 ## 最新动态
 
-* *<mark>[new]</mark>* 🔥🔥🔥 **[2026-10]**: 🚀 **正式发布** **Hy Memory 2.0 (`hy-memory==2.0.2`)**（已上线 [PyPI](https://pypi.org/project/hy-memory/2.0.2/)），并开源 **Hy-Memory** monorepo 与 **[MemHub](memhub/README.md)**！
+* *<mark>[new]</mark>* 🔥🔥🔥 **[2026-10]**: 📦 **正式发布 [MemHub v1.0.4](https://github.com/Tencent-Hunyuan/Hy-Memory/releases/tag/v1.0.4)** 与 **[memhub-cli 1.0.5](https://www.npmjs.com/package/memhub-cli/v/1.0.5)**，提供已签名的 macOS 桌面应用及经过验证的 macOS/Linux Runtime！
+* *<mark>[new]</mark>* 🔥🔥🔥 **[2026-10]**: 🚀 **正式发布** **Hy Memory 2.0 (`hy-memory==2.0.4`)**（已上线 [PyPI](https://pypi.org/project/hy-memory/2.0.4/)），并开源 **Hy-Memory** monorepo 与 **[MemHub](memhub/README.md)**！
 * *<mark>[new]</mark>* 🔥🔥🔥 **[2026-10]**: 🚀 **模型开源** 在 Hugging Face 发布 **[Hy-Memory-30B-A3B](https://huggingface.co/tencent/Hy-Memory-30B-A3B)** 长期记忆写入模型及[技术报告](https://arxiv.org/abs/2610.22196)！
 * *<mark>[new]</mark>* 🔥🔥 **[2026-09]**: 📊 **基准与方法发布** 推出长周期与交错式编程记忆评测基准 **[CMB: A Coding Memory Benchmark for Long-Horizon and Interleaved Coding Agents](docs/papers/CMB_A_Coding_Memory_Benchmark_for_Long_Horizon_and_Interleaved_Coding_Agents.pdf)**（已投稿 **ICLR 2027**）。
 * *<mark>[new]</mark>* 🔥🔥 **[2026-08]**: 🎉 **论文录用** 双系统认知记忆架构论文 **[Memory Beyond Recall: A Dual-Process Cognitive Memory System for Self-Evolving LLM Agents](https://arxiv.org/abs/2606.09483)** 被 **EMNLP 2026** 接收！
@@ -63,13 +64,13 @@
 ### 1. 安装
 
 ```bash
-pip install hy-memory==2.0.2              # 核心（含 ChromaDB 与 Kuzu）
-pip install "hy-memory[qdrant]==2.0.2"    # + Qdrant
-pip install "hy-memory[faiss]==2.0.2"     # + FAISS
-pip install "hy-memory[graph]==2.0.2"     # + Neo4j
-pip install "hy-memory[mysql]==2.0.2"     # + MySQL 缓存
-pip install "hy-memory[rerank]==2.0.2"    # + 本地多语言 cross-encoder rerank
-pip install "hy-memory[all]==2.0.2"       # 全部
+pip install hy-memory==2.0.3              # 核心（含 ChromaDB 与 Kuzu）
+pip install "hy-memory[qdrant]==2.0.3"    # + Qdrant
+pip install "hy-memory[faiss]==2.0.3"     # + FAISS
+pip install "hy-memory[graph]==2.0.3"     # + Neo4j
+pip install "hy-memory[mysql]==2.0.3"     # + MySQL 缓存
+pip install "hy-memory[rerank]==2.0.3"    # + 本地多语言 cross-encoder rerank
+pip install "hy-memory[all]==2.0.3"       # 全部
 ```
 
 ### 2. 配置

docs/monorepo.md

@@ -37,7 +37,7 @@ do not belong in the public source tree.
 - Root GitHub workflows run SDK and MemHub checks from their respective working
   directories. Native lifecycle workflows produce reviewable Actions artifacts;
   tag releases remain fail-closed on public identity, signing and notarization.
-- Hy Memory is published as `hy-memory==2.0.2`. MemHub pins that public
+- Hy Memory is published as `hy-memory==2.0.3`. MemHub pins that public
   distribution; its selected npm package is `memhub-cli`, which is not yet
   published.
 - Hy-Memory source uses Apache-2.0 at the repository root. Third-party

docs/usage.md

@@ -5,7 +5,7 @@
 ## 安装与快速开始
 
 ```bash
-pip install hy-memory==2.0.2
+pip install hy-memory==2.0.3
 
 # 可选后端
 pip install 'hy-memory[qdrant]==2.0.1'
@@ -106,6 +106,15 @@ result = client.add(
 )
 ```
 
+`metadata` 与 `meta` 不是同一字段：
+
+- `metadata` 是兼容旧调用方保留的 SDK 处理选项，会进入内部
+  `WriteRequest.extra`，不保证持久化或随查询结果返回。
+- `meta` 是调用方的业务上下文，会持久化并在 `search()`、`get()` 等结果中返回。
+  同一记忆被多次合并时，各次 `meta` 会按写入顺序保存在列表中。
+
+需要以后读回的数据必须放在 `meta`，不要放在 `metadata`。
+
 `scene` 的公开取值：
 
 - `None`：普通文本直接走 chat；messages 含 tool 时执行快速场景判断。
@@ -263,6 +272,8 @@ 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)
 result = client.update("memory_id", "更新后的内容")
 result = client.delete("memory_id")
 result = client.delete_all(user_id="user_001")
@@ -282,6 +293,10 @@ 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。
+
 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`。
 

hy_memory/agent/extractor.py

@@ -1123,6 +1123,7 @@ class Extractor:
                 max_tokens=self._llm_config.agent_max_tokens,
                 tools=None,
                 tool_choice=None,
+                audit_stage="extract",
             )
             self._call_count += 1
             total_tokens += response.tokens_used
@@ -1233,6 +1234,7 @@ class Extractor:
                         max_tokens=self._llm_config.agent_max_tokens,
                         tools=None,
                         tool_choice=None,
+                        audit_stage="extract",
                     )
                     self._call_count += 1
                     total_tokens += response.tokens_used

hy_memory/agent/llm_provider.py

@@ -34,6 +34,7 @@ import os
 import weakref
 
 from ..config import MemoryConfig
+from ..utils.model_call_context import model_call_stage
 from ..core.workload import llm_slot
 
 logger = logging.getLogger(__name__)
@@ -348,6 +349,7 @@ class LLMProvider:
         stop: List[str] = None,
         tools: Optional[List[Dict[str, Any]]] = None,
         tool_choice: Optional[Any] = None,
+        audit_stage: Optional[str] = None,
         **kwargs
     ) -> LLMResponse:
         """
@@ -386,15 +388,16 @@ class LLMProvider:
                         len(prompt), _content_fingerprint(prompt),
                     )
 
-                response = await self._call_llm(
-                    prompt=prompt,
-                    max_tokens=max_tokens,
-                    temperature=temperature,
-                    stop=stop,
-                    tools=tools,
-                    tool_choice=tool_choice,
-                    **kwargs
-                )
+                with model_call_stage(audit_stage):
+                    response = await self._call_llm(
+                        prompt=prompt,
+                        max_tokens=max_tokens,
+                        temperature=temperature,
+                        stop=stop,
+                        tools=tools,
+                        tool_choice=tool_choice,
+                        **kwargs
+                    )
 
                 self._total_calls += 1
                 self._total_tokens += response.tokens_used
@@ -595,6 +598,7 @@ class LLMProvider:
         temperature: Optional[float] = None,
         tools: Optional[List[Dict[str, Any]]] = None,
         tool_choice: Optional[Any] = None,
+        audit_stage: Optional[str] = None,
     ) -> LLMResponse:
         """
         多轮 messages 调用（支持 tool results 回传）。
@@ -622,13 +626,14 @@ class LLMProvider:
                         f"msgs={len(messages)} tools={len(tools) if tools else 0}"
                     )
 
-                response = await self._chat_completions(
-                    messages,
-                    max_tokens=max_tokens,
-                    temperature=temperature,
-                    tools=tools,
-                    tool_choice=tool_choice,
-                )
+                with model_call_stage(audit_stage):
+                    response = await self._chat_completions(
+                        messages,
+                        max_tokens=max_tokens,
+                        temperature=temperature,
+                        tools=tools,
+                        tool_choice=tool_choice,
+                    )
 
                 self._total_calls += 1
                 self._total_tokens += response.tokens_used

hy_memory/agent/reconciler.py

@@ -669,6 +669,7 @@ class MemoryReconciler:
             response = await llm.complete(
                 prompt=prompt,
                 max_tokens=500,
+                audit_stage="reconcile",
             )
             text = response.content.strip()
             if "```json" in text:
@@ -1383,6 +1384,7 @@ class MemoryReconciler:
                     response = await llm.complete(
                         prompt=prompt,
                         max_tokens=dyn_max_tokens,
+                        audit_stage="reconcile",
                     )
                 except Exception as e:
                     # LLM 彻底不可用（provider 已按 max_retries 重试过，网关 500 /

hy_memory/client.py

@@ -1597,7 +1597,9 @@ class HyMemoryClient:
             session_id:   Session ID（三级隔离 key）
             source:       调用来源，例如 openclaw / cursor / codex；与 agent_id 正交。
                           写入前执行 strip().lower()，默认 "default"。
-            metadata:     自定义元数据
+            metadata:     旧版 SDK 处理选项，会并入内部 WriteRequest.extra。
+                          不保证持久化或随查询返回；调用方需要保存并读回的业务字段
+                          请使用 meta。
             memory_at:    记忆时间戳（不传则用当前时间），用于导入历史记忆
             scene:        显式指定写入路由，跳过 LLM 自动判定。
                           - None（默认）：含 tool 时 LLM 判 chat/working/coding；
@@ -1623,8 +1625,8 @@ class HyMemoryClient:
             branch:       Coding memory 路径专用。分支名。同 workspace_id。chat 路径忽略。
             meta:         调用方自定义业务 meta，随本次产出的记忆存储、查询时原样返回。
                           不 embed、不参与召回、不作过滤条件。
-                          与 metadata 不同：metadata 会并入 extra（含 SDK 内部控制键），
-                          meta 独立存放，只做透传。
+                          与 metadata 不同：metadata 是旧版内部处理选项，
+                          meta 独立持久化，只做业务透传。
                           语义为「该次 write 时刻的上下文」，故存为 list——同一节点被
                           多次写入合并时会追加而非覆盖，详见 get() 返回的 meta 字段。
 
@@ -1791,17 +1793,26 @@ class HyMemoryClient:
             query_embedding=query_embedding,
         ))
 
-    def get(self, memory_id: str) -> Optional[Dict[str, Any]]:
+    def get(
+        self,
+        memory_id: str,
+        *,
+        include_raw: bool = False,
+    ) -> Optional[Dict[str, Any]]:
         """
         获取单条记忆。
 
         Args:
             memory_id: 记忆 ID
+            include_raw: 是否允许返回 L1_RAW 原始记忆。默认 False，保持原有隐藏行为；
+                         lite 模式的记忆均为 L1_RAW，需显式设为 True 才能按 ID 读取。
 
         Returns:
             {"memory_id", "content", "layer", "request_id", "elapsed_ms"} 或 None（不存在时）
         """
-        return self._loop_thread.run(self.async_get(memory_id))
+        return self._loop_thread.run(
+            self.async_get(memory_id, include_raw=include_raw)
+        )
 
     def update(
         self,
@@ -3650,12 +3661,21 @@ class HyMemoryClient:
     # ================================================================
 
     @_ensure_internal_loop
-    async def async_get(self, memory_id: str) -> Optional[Dict[str, Any]]:
+    async def async_get(
+        self,
+        memory_id: str,
+        *,
+        include_raw: bool = False,
+    ) -> Optional[Dict[str, Any]]:
         """
         获取单条记忆（异步）。
 
         VDB + Graph 双路并行查询，哪个有结果返回哪个。
 
+        Args:
+            memory_id: 记忆 ID
+            include_raw: 是否允许返回 L1_RAW 原始记忆，默认 False。
+
         Returns:
             {"memory_id", "content", "layer", "source", "request_id", "elapsed_ms"} 或 None
         """
@@ -3683,17 +3703,24 @@ class HyMemoryClient:
 
         vdb_node, graph_node = await _asyncio_get.gather(_get_vdb(), _get_graph())
 
-        # L1_RAW 不对外可见：按 ID 直取走主键查询、绕过向量库的 layer filter，
-        # 故在此统一挡掉。async_get 是 GET /v1/memories/{id} 与
-        # /v1/memories/batch 两个对外接口的共同必经处，而内部写链路不经过它
-        # （writer 写后校验直接调 vector_store.get_by_id），挡在此处不影响内部。
+        # L1_RAW 默认不对外可见：按 ID 直取走主键查询、绕过向量库的 layer
+        # filter，故在此统一挡掉。SDK 调用方可通过 include_raw=True 显式读取；
+        # HTTP 接口不传该参数，继续保持隐藏。内部写链路不经过此方法。
         #
         # 置于 gather 之后、埋点之前：命中 raw 仍照常落 GET 埋点（保留可观测性，
         # 便于发现有调用方拿 raw id 试探），只是不把内容返回给调用方。
         # 对调用方的表现与"该 id 不存在"完全一致，不额外暴露"存在但被挡"的信息。
-        if vdb_node is not None and vdb_node.layer == MemoryLayer.L1_RAW:
+        if (
+            not include_raw
+            and vdb_node is not None
+            and vdb_node.layer == MemoryLayer.L1_RAW
+        ):
             vdb_node = None
-        if graph_node is not None and getattr(graph_node, "layer", None) == MemoryLayer.L1_RAW:
+        if (
+            not include_raw
+            and graph_node is not None
+            and getattr(graph_node, "layer", None) == MemoryLayer.L1_RAW
+        ):
             graph_node = None
 
         elapsed = round((time.perf_counter() - t0) * 1000, 2)

hy_memory/coding/extractor.py

@@ -919,6 +919,7 @@ class CodingMemoryExtractor:
             resp = await self.llm_provider.complete(
                 prompt=prompt,
                 max_tokens=self.max_tokens,
+                audit_stage="coding_extract",
             )
         except Exception as e:
             logger.warning(f"[coding-extract] LLM call failed: {e}")

hy_memory/coding/fact_reconciler.py

@@ -466,7 +466,13 @@ class WorkingFactReconciler:
             return [], []
 
         candidates = await self._retrieve_candidates(
-            drafts, user_id=user_id, source=source,
+            drafts,
+            user_id=user_id,
+            agent_id=agent_id,
+            session_id=session_id,
+            workspace_id=workspace_id,
+            branch=branch,
+            source=source,
         )
 
         try:
@@ -499,7 +505,15 @@ class WorkingFactReconciler:
     # ------------------------------------------------------------
 
     async def _retrieve_candidates(
-        self, drafts: List[WorkingFactDraft], *, user_id: str, source: str
+        self,
+        drafts: List[WorkingFactDraft],
+        *,
+        user_id: str,
+        agent_id: str,
+        session_id: Optional[str],
+        workspace_id: Optional[str],
+        branch: Optional[str],
+        source: str,
     ) -> List[List[WorkingFact]]:
         """每条 draft 用 statement 召回同 user 的 top-N active 事实。
 
@@ -512,6 +526,10 @@ class WorkingFactReconciler:
         关掉时整份召回原样交给 LLM。适用于 subject 不是稳定实体名的数据集（自由文本、
         位置标签），那里精确字符串相等会把同实体判成不同主体，反而拦掉真实更新。
         """
+        # Empty/legacy session metadata remains readable, but cannot prove
+        # conversation authority for a destructive SUPERSEDE.
+        if not session_id:
+            return [[] for _ in drafts]
         try:
             embeddings = await self.embed_service.embed_batch(
                 [d.statement for d in drafts]
@@ -525,6 +543,10 @@ class WorkingFactReconciler:
             try:
                 hits = await self.store.search_facts(
                     emb, user_id=user_id, top=self.candidates_per_draft,
+                    agent_id=agent_id,
+                    session_id=session_id,
+                    workspace_id=workspace_id,
+                    branch=branch,
                     # 写入侧同样接词面通道：候选选错了 LLM 就没有机会判对。这里不传
                     # min_score——写入侧要的是"有哪些可能的前身"，让 LLM 看更宽的名单，
                     # 阈值过滤是读取侧的事。
@@ -571,6 +593,7 @@ class WorkingFactReconciler:
         resp = await self.llm_provider.complete(
             prompt=prompt,
             max_tokens=self.max_tokens,
+            audit_stage="coding_fact_reconcile",
         )
         return self._parse_ops(resp.content)
 
@@ -637,6 +660,7 @@ class WorkingFactReconciler:
             resp = await self.llm_provider.complete(
                 prompt=prompt,
                 max_tokens=_CHAIN_SUMMARY_MAX_TOKENS,
+                audit_stage="coding_fact_summary",
             )
         except Exception as e:
             logger.warning(f"[fact-reconcile] chain summary failed: {e}")
@@ -792,6 +816,15 @@ class WorkingFactReconciler:
                 old = await self.store.get_fact(old_id, user_id=user_id)
                 if old is None:
                     continue
+                if (
+                    old.agent_id != agent_id
+                    or old.session_id != session_id
+                    or old.workspace_id != workspace_id
+                ):
+                    logger.warning(
+                        f"[fact-reconcile] refused cross-authority SUPERSEDE: {old.fact_id}"
+                    )
+                    continue
                 # 闸门在候选阶段已经拦过一道，这里再拦一次：候选是按 statement 召回的，
                 # 而 LLM 只能覆盖候选名单里的 id，理论上到不了这儿。真到了就说明前面
                 # 哪一环漏了，宁可让这次覆盖不生效——多留一条旧事实是可恢复的，删掉

hy_memory/coding/judge.py

@@ -29,6 +29,9 @@ logger = logging.getLogger(__name__)
 
 _JUDGE_MAX_TURNS = 20
 _JUDGE_HEAD_TURNS = 5
+_JUDGE_USER_CONTENT_LIMIT = 1_000
+_JUDGE_USER_CONTENT_HEAD = 200
+_JUDGE_USER_CONTENT_TAIL = 200
 _JUDGE_MAX_QUERY_CHARS = 300
 _JUDGE_MAX_QUERIES = 20
 
@@ -186,6 +189,7 @@ async def classify_messages_scene(
             prompt=prompt,
             max_tokens=max_tokens,
             temperature=temperature,
+            audit_stage="coding_judge",
         )
         _elapsed = round((time.perf_counter() - _t0) * 1000, 2)
         result = _parse_scene_json(resp.content)
@@ -631,9 +635,13 @@ def _format_turns(summary: List[Dict[str, Any]]) -> str:
     rows = []
     for t in summary:
         user = t.get("user", "") or ""
-        # 单 turn 用户文本截到 200 字符
-        if len(user) > 200:
-            user = user[:200] + "..."
+        if len(user) > _JUDGE_USER_CONTENT_LIMIT:
+            omitted = len(user) - _JUDGE_USER_CONTENT_HEAD - _JUDGE_USER_CONTENT_TAIL
+            user = (
+                user[:_JUDGE_USER_CONTENT_HEAD]
+                + f"\n[...{omitted} chars omitted...]\n"
+                + user[-_JUDGE_USER_CONTENT_TAIL:]
+            )
         tools = t.get("tools", []) or []
         rows.append(
             f"- turn {t.get('turn', '?')}: user={user!r}, tools={tools}"

hy_memory/coding/reconciler.py

@@ -240,7 +240,7 @@ class CodingMemoryReconciler:
         )
 
         # 3. 校验 + 执行
-        executed = await self._execute(ops, drafts)
+        executed = await self._execute(ops, drafts, candidates_per_draft)
         logger.info(
             f"[coding-reconcile] drafts={len(drafts)} ops={len(executed)} "
             + ", ".join(f"{a}={sum(1 for o in executed if o.action == a)}"
@@ -263,6 +263,12 @@ class CodingMemoryReconciler:
         """
         result: List[List[CodingMemory]] = []
         for d in drafts:
+            # A coding mutation must be attributable to one conversation.
+            # Empty/legacy session metadata stays readable through ordinary
+            # search, but it cannot authorize UPDATE/DELETE.
+            if not d.session_id:
+                result.append([])
+                continue
             try:
                 # task 为主入口，search_keys 为附加入口（去重、去空）
                 queries = [d.task] + [k for k in (d.search_keys or []) if k and k != d.task]
@@ -274,6 +280,8 @@ class CodingMemoryReconciler:
                     hits = await self.store.search_by_query_embedding(
                         emb,
                         user_id=d.user_id,
+                        agent_id=d.agent_id,
+                        session_id=d.session_id,
                         workspace_id=d.workspace_id,
                         branch=d.branch,
                         top=self.candidates_per_draft * 4,
@@ -310,6 +318,7 @@ class CodingMemoryReconciler:
         resp = await self.llm_provider.complete(
             prompt=prompt,
             max_tokens=self.max_tokens,
+            audit_stage="coding_reconcile",
         )
         return self._parse_ops(resp.content), prompt, resp.content, resp
 
@@ -414,6 +423,7 @@ class CodingMemoryReconciler:
         self,
         ops: List[ReconcileOp],
         drafts: List[CodingMemoryDraft],
+        candidates_per_draft: List[List[CodingMemory]],
     ) -> List[ReconcileOp]:
         """
         校验 + 执行 ops。返回成功执行的 ops（filter 掉无效 op）。
@@ -453,6 +463,11 @@ class CodingMemoryReconciler:
                 if not op.target_experience_id:
                     logger.warning(f"[coding-reconcile] drop UPDATE missing target: {op.to_dict()}")
                     continue
+                if not self._target_was_offered(op, candidates_per_draft):
+                    logger.warning(
+                        f"[coding-reconcile] drop UPDATE outside authority scope: {op.to_dict()}"
+                    )
+                    continue
                 if op.draft_idx in consumed:
                     logger.warning(f"[coding-reconcile] drop UPDATE on consumed draft: {op.to_dict()}")
                     continue
@@ -475,6 +490,11 @@ class CodingMemoryReconciler:
                 if not op.target_experience_id:
                     logger.warning(f"[coding-reconcile] drop DELETE missing target: {op.to_dict()}")
                     continue
+                if not self._target_was_offered(op, candidates_per_draft):
+                    logger.warning(
+                        f"[coding-reconcile] drop DELETE outside authority scope: {op.to_dict()}"
+                    )
+                    continue
                 # DELETE 需要知道 user_id；通过 draft_idx（如有）或第一个 draft 推断
                 user_id = self._infer_user_id(op, drafts)
                 if not user_id:
@@ -511,6 +531,19 @@ class CodingMemoryReconciler:
 
         return executed
 
+    @staticmethod
+    def _target_was_offered(
+        op: ReconcileOp,
+        candidates_per_draft: List[List[CodingMemory]],
+    ) -> bool:
+        """Only mutate a target frozen into this draft's scoped candidate set."""
+        if op.draft_idx is None or not (0 <= op.draft_idx < len(candidates_per_draft)):
+            return False
+        return any(
+            candidate.experience_id == op.target_experience_id
+            for candidate in candidates_per_draft[op.draft_idx]
+        )
+
     @staticmethod
     def _infer_user_id(
         op: ReconcileOp, drafts: List[CodingMemoryDraft]

hy_memory/coding/store.py

@@ -529,10 +529,20 @@ class CodingMemoryStore:
         """覆盖式更新一条 coding memory。"""
         # 拿旧 keys 长度，先删旧 VDB rows
         prev = await self.get_by_id(memory.experience_id, user_id=memory.user_id)
-        prev_key_count = (1 + len(prev.search_keys)) if prev else 0
+        if prev is None:
+            raise ValueError("coding_memory_update_target_missing")
+        if (
+            not prev.session_id
+            or not memory.session_id
+            or prev.agent_id != memory.agent_id
+            or prev.session_id != memory.session_id
+            or prev.workspace_id != memory.workspace_id
+        ):
+            raise ValueError("coding_memory_update_authority_mismatch")
+        prev_key_count = 1 + len(prev.search_keys)
 
         memory.updated_at = datetime.utcnow()
-        if memory.created_at is None and prev is not None:
+        if memory.created_at is None:
             memory.created_at = prev.created_at
         await _run_in_pool(self._upsert_sql_sync, memory)
         # 删旧 VDB
@@ -671,6 +681,8 @@ class CodingMemoryStore:
         query_embedding: List[float],
         user_id: str,
         *,
+        agent_id: Optional[str] = None,
+        session_id: Optional[str] = None,
         workspace_id: Optional[str] = None,
         branch: Optional[str] = None,
         top: int = 40,
@@ -688,21 +700,28 @@ class CodingMemoryStore:
                 },
                 ...
             ]
-        workspace_id / branch 是作用域过滤，规则见 _scope_where：不传则纯语义召回
-        （行为与加过滤前完全一致）；传了则只保留 SQLite 里落在作用域内的经验，并按
-        作用域的选择率放大向量取数，避免"先取 top-N 再过滤"把召回吃掉。
+        agent_id / session_id 是写授权作用域，传入时精确匹配；workspace_id / branch
+        的规则见 _scope_where。不传任何作用域则维持纯语义召回；传了则只保留 SQLite
+        里落在作用域内的经验，并按选择率放大向量取数，避免"先取 top-N 再过滤"把
+        召回吃掉。
         """
         if self._vec is None:
             return []
 
         allowed: Optional[Set[str]] = None
-        if workspace_id:
+        if agent_id or session_id or workspace_id:
             allowed, total = await _run_in_pool(
-                self._scoped_experience_ids_sync, user_id, workspace_id, branch
+                self._authority_scoped_experience_ids_sync,
+                user_id,
+                agent_id,
+                session_id,
+                workspace_id,
+                branch,
             )
             if not allowed:
                 logger.info(
                     f"[coding-store] no experience in scope: "
+                    f"agent_id={agent_id!r} session_id={session_id!r} "
                     f"workspace_id={workspace_id!r} branch={branch!r} "
                     f"(user has {total}); skip vdb search"
                 )
@@ -778,6 +797,8 @@ class CodingMemoryStore:
         query_embedding: List[float],
         user_id: str,
         *,
+        agent_id: Optional[str] = None,
+        session_id: Optional[str] = None,
         workspace_id: Optional[str] = None,
         branch: Optional[str] = None,
         top: int = 5,
@@ -791,8 +812,9 @@ class CodingMemoryStore:
         SUPERSEDE 时向量行已被删掉，理论上召不到 superseded；这里再按 SQLite 的
         status 过一道，避免两边不一致时把过期结论放出去。
 
-        workspace_id / branch 是作用域过滤，规则见 _scope_where，语义与经验层一致：
-        不传不过滤，传了按 SQLite 里的值筛，并按选择率放大向量取数。
+        agent_id / session_id 是写授权作用域，传入时精确匹配；workspace_id / branch
+        的规则见 _scope_where。不传任何作用域则不过滤，传了按 SQLite 里的值筛，并
+        按选择率放大向量取数。
 
         向量库按 top 的 4 倍过采样：向量行和 SQLite 行不是一个事务里写的，中间挂掉
         就会留下"向量还在、SQLite 没了"的孤儿，软删时删向量失败也会留下 superseded
@@ -809,13 +831,19 @@ class CodingMemoryStore:
             return []
 
         fetch = top * 4
-        if workspace_id:
+        if agent_id or session_id or workspace_id:
             scoped, total = await _run_in_pool(
-                self._count_active_facts_sync, user_id, workspace_id, branch
+                self._count_active_facts_sync,
+                user_id,
+                agent_id,
+                session_id,
+                workspace_id,
+                branch,
             )
             if not scoped:
                 logger.info(
                     f"[coding-store] no active fact in scope: "
+                    f"agent_id={agent_id!r} session_id={session_id!r} "
                     f"workspace_id={workspace_id!r} branch={branch!r} "
                     f"(user has {total}); skip fact vdb search"
                 )
@@ -852,6 +880,8 @@ class CodingMemoryStore:
         alive = [
             (f, scores[f.fact_id]) for f in facts
             if f.status == "active"
+            and (not agent_id or f.agent_id == agent_id)
+            and (not session_id or f.session_id == session_id)
             and _scope_matches(f.workspace_id, f.branch, workspace_id, branch)
         ]
         kept = [(f, s) for f, s in alive if s >= min_score]
@@ -1050,14 +1080,25 @@ class CodingMemoryStore:
             by_id = {r["experience_id"]: _row_to_memory(r) for r in rows}
             return [by_id[eid] for eid in experience_ids if eid in by_id]
 
-    def _scoped_experience_ids_sync(
-        self, user_id: str, workspace_id: str, branch: Optional[str]
+    def _authority_scoped_experience_ids_sync(
+        self,
+        user_id: str,
+        agent_id: Optional[str],
+        session_id: Optional[str],
+        workspace_id: Optional[str],
+        branch: Optional[str],
     ) -> Tuple[Set[str], int]:
-        """返回 (落在作用域内的 experience_id 集合, 该 user 的经验总数)。
+        """返回 (落在授权作用域内的 experience_id 集合, 该 user 的经验总数)。
 
         一次查询同时拿到过滤结果和选择率分母：前者供向量命中做交集，后者定放大倍数。
         """
         scope_sql, scope_params = _scope_where(workspace_id, branch)
+        if agent_id:
+            scope_sql = f"agent_id = ? AND ({scope_sql})"
+            scope_params = [agent_id, *scope_params]
+        if session_id:
+            scope_sql = f"session_id = ? AND ({scope_sql})"
+            scope_params = [session_id, *scope_params]
         with self._lock:
             assert self._conn is not None
             cur = self._conn.execute(
@@ -1069,13 +1110,24 @@ class CodingMemoryStore:
         return {r["experience_id"] for r in rows if r["in_scope"]}, len(rows)
 
     def _count_active_facts_sync(
-        self, user_id: str, workspace_id: str, branch: Optional[str]
+        self,
+        user_id: str,
+        agent_id: Optional[str],
+        session_id: Optional[str],
+        workspace_id: Optional[str],
+        branch: Optional[str],
     ) -> Tuple[int, int]:
         """返回 (作用域内 active 事实数, 该 user 的 active 事实总数)。
 
         事实层的过滤在 get_facts 拿到对象之后用 _scope_matches 做，这里只要选择率。
         """
         scope_sql, scope_params = _scope_where(workspace_id, branch)
+        if agent_id:
+            scope_sql = f"agent_id = ? AND ({scope_sql})"
+            scope_params = [agent_id, *scope_params]
+        if session_id:
+            scope_sql = f"session_id = ? AND ({scope_sql})"
+            scope_params = [session_id, *scope_params]
         with self._lock:
             assert self._conn is not None
             row = self._conn.execute(

hy_memory/data/cache_sqlite.py

@@ -7,7 +7,7 @@ Agent Memory V2 - SqliteCache
 设计模式跟随 history_store.py:
 - 标准库 sqlite3 + _run_in_sqlite_pool()
 - WAL 模式 + check_same_thread=False
-- threading.Lock 保护写操作
+- 每实例 threading.RLock 保护共享连接的全部操作与关闭
 
 三张表:
 - memory_operations / pipeline_logs / system_metrics: 审计与观测
@@ -122,7 +122,10 @@ class SqliteCache(CacheBase):
     def __init__(self, config: MemoryConfig):
         self.config = config
         self._conn: Optional[sqlite3.Connection] = None
-        self._lock = threading.Lock()
+        # One connection is shared by the executor workers. Use the same lock
+        # for reads, writes, initialization and close so close() cannot free
+        # the connection while another worker is inside sqlite3.
+        self._lock = threading.RLock()
 
         # 读取 db_path
         cache_cfg = getattr(config, "cache", None)
@@ -144,35 +147,37 @@ class SqliteCache(CacheBase):
         logger.debug(f"SqliteCache initialized: {self._db_path}")
 
     def _init_sync(self) -> None:
-        db_dir = os.path.dirname(self._db_path)
-        if db_dir:
-            os.makedirs(db_dir, exist_ok=True)
-        self._conn = sqlite3.connect(self._db_path, check_same_thread=False)
-        self._conn.row_factory = sqlite3.Row
-        self._conn.execute("PRAGMA journal_mode=WAL;")
-        self._conn.executescript(_CREATE_TABLES_SQL)
-        self._migrate_pipeline_logs_columns()
-        self._conn.commit()
+        with self._lock:
+            db_dir = os.path.dirname(self._db_path)
+            if db_dir:
+                os.makedirs(db_dir, exist_ok=True)
+            self._conn = sqlite3.connect(self._db_path, check_same_thread=False)
+            self._conn.row_factory = sqlite3.Row
+            self._conn.execute("PRAGMA journal_mode=WAL;")
+            self._conn.executescript(_CREATE_TABLES_SQL)
+            self._migrate_pipeline_logs_columns()
+            self._conn.commit()
 
     def _migrate_pipeline_logs_columns(self) -> None:
         """给旧库的 pipeline_logs_v2 补 namespace/session/event time 列。"""
-        try:
-            cols = {
-                row["name"]
-                for row in self._conn.execute("PRAGMA table_info(pipeline_logs_v2)")
-            }
-            _text_cols = ("namespace", "session_id")
-            for col in _text_cols:
-                if col not in cols:
+        with self._lock:
+            try:
+                cols = {
+                    row["name"]
+                    for row in self._conn.execute("PRAGMA table_info(pipeline_logs_v2)")
+                }
+                _text_cols = ("namespace", "session_id")
+                for col in _text_cols:
+                    if col not in cols:
+                        self._conn.execute(
+                            f"ALTER TABLE pipeline_logs_v2 ADD COLUMN {col} TEXT NOT NULL DEFAULT ''"
+                        )
+                if "event_at_ms" not in cols:
                     self._conn.execute(
-                        f"ALTER TABLE pipeline_logs_v2 ADD COLUMN {col} TEXT NOT NULL DEFAULT ''"
+                        "ALTER TABLE pipeline_logs_v2 ADD COLUMN event_at_ms INTEGER NOT NULL DEFAULT 0"
                     )
-            if "event_at_ms" not in cols:
-                self._conn.execute(
-                    "ALTER TABLE pipeline_logs_v2 ADD COLUMN event_at_ms INTEGER NOT NULL DEFAULT 0"
-                )
-        except Exception as e:
-            logger.debug(f"[sqlite] pipeline_logs_v2 column migration skipped: {e}")
+            except Exception as e:
+                logger.debug(f"[sqlite] pipeline_logs_v2 column migration skipped: {e}")
 
     async def close(self) -> None:
         if self._conn is not None:
@@ -180,9 +185,10 @@ class SqliteCache(CacheBase):
         logger.info("SqliteCache closed")
 
     def _close_sync(self) -> None:
-        if self._conn:
-            self._conn.close()
-            self._conn = None
+        with self._lock:
+            if self._conn:
+                self._conn.close()
+                self._conn = None
 
     # ================================================================
     # 统计
@@ -255,39 +261,40 @@ class SqliteCache(CacheBase):
         """
         try:
             def _query():
-                conditions = []
-                params = []
-                if request_id:
-                    conditions.append("request_id = ?")
-                    params.append(request_id)
-                if memory_id:
-                    conditions.append("memory_id = ?")
-                    params.append(memory_id)
-                if user_id:
-                    conditions.append("user_id = ?")
-                    params.append(user_id)
+                with self._lock:
+                    conditions = []
+                    params = []
+                    if request_id:
+                        conditions.append("request_id = ?")
+                        params.append(request_id)
+                    if memory_id:
+                        conditions.append("memory_id = ?")
+                        params.append(memory_id)
+                    if user_id:
+                        conditions.append("user_id = ?")
+                        params.append(user_id)
 
-                if not conditions:
-                    return []
+                    if not conditions:
+                        return []
 
-                where = " AND ".join(conditions)
-                rows = self._conn.execute(
-                    f"SELECT * FROM memory_operations_v2 WHERE {where} "
-                    f"ORDER BY id DESC LIMIT ?",
-                    params + [limit],
-                ).fetchall()
-                result = []
-                for r in rows:
-                    row = dict(r)
-                    # 反序列化 supersedes JSON 字符串 → list
-                    import json as _json
-                    raw_sup = row.get("supersedes", "")
-                    try:
-                        row["supersedes"] = _json.loads(raw_sup) if raw_sup else []
-                    except Exception:
-                        row["supersedes"] = []
-                    result.append(row)
-                return result
+                    where = " AND ".join(conditions)
+                    rows = self._conn.execute(
+                        f"SELECT * FROM memory_operations_v2 WHERE {where} "
+                        f"ORDER BY id DESC LIMIT ?",
+                        params + [limit],
+                    ).fetchall()
+                    result = []
+                    for r in rows:
+                        row = dict(r)
+                        # 反序列化 supersedes JSON 字符串 → list
+                        import json as _json
+                        raw_sup = row.get("supersedes", "")
+                        try:
+                            row["supersedes"] = _json.loads(raw_sup) if raw_sup else []
+                        except Exception:
+                            row["supersedes"] = []
+                        result.append(row)
+                    return result
 
             return await _run_in_sqlite_pool(_query)
         except Exception as e:
@@ -369,37 +376,38 @@ class SqliteCache(CacheBase):
         """
         try:
             def _query():
-                conditions = []
-                params = []
-                if request_id:
-                    conditions.append("request_id = ?")
-                    params.append(request_id)
-                if user_id:
-                    conditions.append("user_id = ?")
-                    params.append(user_id)
-                if step:
-                    conditions.append("step = ?")
-                    params.append(step)
+                with self._lock:
+                    conditions = []
+                    params = []
+                    if request_id:
+                        conditions.append("request_id = ?")
+                        params.append(request_id)
+                    if user_id:
+                        conditions.append("user_id = ?")
+                        params.append(user_id)
+                    if step:
+                        conditions.append("step = ?")
+                        params.append(step)
 
-                if not conditions:
-                    return []
+                    if not conditions:
+                        return []
 
-                where = " AND ".join(conditions)
-                rows = self._conn.execute(
-                    f"SELECT * FROM pipeline_logs_v2 WHERE {where} "
-                    f"ORDER BY id ASC LIMIT ?",
-                    params + [limit],
-                ).fetchall()
-                out = [dict(r) for r in rows]
-                # 派生可读的 event_at（ISO ms）供查询接口直接用；event_at_ms 为 0 时留空
-                from datetime import datetime as _dt
-                for _r in out:
-                    _ms = _r.get("event_at_ms") or 0
-                    _r["event_at"] = (
-                        _dt.fromtimestamp(_ms / 1000.0).isoformat(timespec="milliseconds")
-                        if _ms else ""
-                    )
-                return out
+                    where = " AND ".join(conditions)
+                    rows = self._conn.execute(
+                        f"SELECT * FROM pipeline_logs_v2 WHERE {where} "
+                        f"ORDER BY id ASC LIMIT ?",
+                        params + [limit],
+                    ).fetchall()
+                    out = [dict(r) for r in rows]
+                    # 派生可读的 event_at（ISO ms）供查询接口直接用；event_at_ms 为 0 时留空
+                    from datetime import datetime as _dt
+                    for _r in out:
+                        _ms = _r.get("event_at_ms") or 0
+                        _r["event_at"] = (
+                            _dt.fromtimestamp(_ms / 1000.0).isoformat(timespec="milliseconds")
+                            if _ms else ""
+                        )
+                    return out
 
             return await _run_in_sqlite_pool(_query)
         except Exception as e:
@@ -414,26 +422,27 @@ class SqliteCache(CacheBase):
         """存储一个分钟桶的增量指标数据（UPSERT: 同一分钟多次写入会合并）"""
         try:
             def _store():
-                now = datetime.now().isoformat()
-                # UPSERT: 如果同一分钟已存在，合并数据
-                existing = self._conn.execute(
-                    "SELECT data FROM system_metrics WHERE minute_ts = ?",
-                    (minute_ts,),
-                ).fetchone()
-                if existing:
-                    import json as _json
-                    old_data = _json.loads(existing["data"])
-                    merged = self._merge_metric_buckets(old_data, data)
-                    self._conn.execute(
-                        "UPDATE system_metrics SET data = ?, created_at = ? WHERE minute_ts = ?",
-                        (_json.dumps(merged, default=str), now, minute_ts),
-                    )
-                else:
-                    self._conn.execute(
-                        "INSERT INTO system_metrics (minute_ts, data, created_at) VALUES (?, ?, ?)",
-                        (minute_ts, json.dumps(data, default=str), now),
-                    )
-                self._conn.commit()
+                with self._lock:
+                    now = datetime.now().isoformat()
+                    # UPSERT: 如果同一分钟已存在，合并数据
+                    existing = self._conn.execute(
+                        "SELECT data FROM system_metrics WHERE minute_ts = ?",
+                        (minute_ts,),
+                    ).fetchone()
+                    if existing:
+                        import json as _json
+                        old_data = _json.loads(existing["data"])
+                        merged = self._merge_metric_buckets(old_data, data)
+                        self._conn.execute(
+                            "UPDATE system_metrics SET data = ?, created_at = ? WHERE minute_ts = ?",
+                            (_json.dumps(merged, default=str), now, minute_ts),
+                        )
+                    else:
+                        self._conn.execute(
+                            "INSERT INTO system_metrics (minute_ts, data, created_at) VALUES (?, ?, ?)",
+                            (minute_ts, json.dumps(data, default=str), now),
+                        )
+                    self._conn.commit()
             await _run_in_sqlite_pool(_store)
         except Exception as e:
             logger.warning(f"store_metrics_minute failed: {e}")
@@ -442,17 +451,18 @@ class SqliteCache(CacheBase):
         """读取指定时间范围内的分钟指标数据"""
         try:
             def _load():
-                rows = self._conn.execute(
-                    "SELECT data FROM system_metrics WHERE minute_ts >= ? AND minute_ts <= ? ORDER BY minute_ts",
-                    (start_ts, end_ts),
-                ).fetchall()
-                results = []
-                for row in rows:
-                    try:
-                        results.append(json.loads(row["data"]))
-                    except (json.JSONDecodeError, TypeError):
-                        pass
-                return results
+                with self._lock:
+                    rows = self._conn.execute(
+                        "SELECT data FROM system_metrics WHERE minute_ts >= ? AND minute_ts <= ? ORDER BY minute_ts",
+                        (start_ts, end_ts),
+                    ).fetchall()
+                    results = []
+                    for row in rows:
+                        try:
+                            results.append(json.loads(row["data"]))
+                        except (json.JSONDecodeError, TypeError):
+                            pass
+                    return results
             return await _run_in_sqlite_pool(_load)
         except Exception as e:
             logger.warning(f"load_metrics_range failed: {e}")
@@ -462,11 +472,12 @@ class SqliteCache(CacheBase):
         """删除指定时间之前的 metrics 数据"""
         try:
             def _cleanup():
-                self._conn.execute(
-                    "DELETE FROM system_metrics WHERE minute_ts < ?",
-                    (before_ts,),
-                )
-                self._conn.commit()
+                with self._lock:
+                    self._conn.execute(
+                        "DELETE FROM system_metrics WHERE minute_ts < ?",
+                        (before_ts,),
+                    )
+                    self._conn.commit()
             await _run_in_sqlite_pool(_cleanup)
         except Exception as e:
             logger.warning(f"cleanup_old_metrics failed: {e}")

hy_memory/data/history_store.py

@@ -16,6 +16,7 @@ import json
 import logging
 import os
 import sqlite3
+import threading
 from datetime import datetime, timezone
 from typing import Dict, List, Any, Optional
 
@@ -89,6 +90,10 @@ class HistoryStore:
         """
         self._db_path = config.history.db_path
         self._conn: Optional[sqlite3.Connection] = None
+        # sqlite3.Connection.close() must never race with an execute() on the
+        # same connection. The lock also serializes operations on this shared
+        # connection; separate HistoryStore instances remain independent.
+        self._conn_lock = threading.RLock()
 
     async def initialize(self) -> None:
         """创建表和索引（如果不存在）"""
@@ -97,20 +102,21 @@ class HistoryStore:
 
     def _init_sync(self) -> None:
         """同步初始化：创建目录、连接数据库、建表建索引"""
-        # 确保目录存在
-        db_dir = os.path.dirname(self._db_path)
-        if db_dir:
-            os.makedirs(db_dir, exist_ok=True)
+        with self._conn_lock:
+            # 确保目录存在
+            db_dir = os.path.dirname(self._db_path)
+            if db_dir:
+                os.makedirs(db_dir, exist_ok=True)
 
-        self._conn = sqlite3.connect(self._db_path, check_same_thread=False)
-        self._conn.row_factory = sqlite3.Row
-        # 开启 WAL 模式，提高并发读写性能
-        self._conn.execute("PRAGMA journal_mode=WAL;")
+            self._conn = sqlite3.connect(self._db_path, check_same_thread=False)
+            self._conn.row_factory = sqlite3.Row
+            # 开启 WAL 模式，提高并发读写性能
+            self._conn.execute("PRAGMA journal_mode=WAL;")
 
-        self._conn.execute(_CREATE_TABLE_SQL)
-        for idx_sql in _CREATE_INDEXES_SQL:
-            self._conn.execute(idx_sql)
-        self._conn.commit()
+            self._conn.execute(_CREATE_TABLE_SQL)
+            for idx_sql in _CREATE_INDEXES_SQL:
+                self._conn.execute(idx_sql)
+            self._conn.commit()
 
     # ================================================================
     # 记录操作
@@ -218,13 +224,14 @@ class HistoryStore:
         return await _run_in_sqlite_pool(self._get_history_sync, memory_id)
 
     def _get_history_sync(self, memory_id: str) -> List[Dict[str, Any]]:
-        if self._conn is None:
-            return []
-        cursor = self._conn.execute(
-            "SELECT * FROM memory_history WHERE memory_id = ? ORDER BY created_at ASC",
-            (memory_id,),
-        )
-        return [dict(row) for row in cursor.fetchall()]
+        with self._conn_lock:
+            if self._conn is None:
+                return []
+            cursor = self._conn.execute(
+                "SELECT * FROM memory_history WHERE memory_id = ? ORDER BY created_at ASC",
+                (memory_id,),
+            )
+            return [dict(row) for row in cursor.fetchall()]
 
     async def get_recent(
         self,
@@ -244,14 +251,15 @@ class HistoryStore:
         return await _run_in_sqlite_pool(self._get_recent_sync, isolation_key, limit)
 
     def _get_recent_sync(self, isolation_key: str, limit: int) -> List[Dict[str, Any]]:
-        if self._conn is None:
-            return []
-        cursor = self._conn.execute(
-            "SELECT * FROM memory_history WHERE isolation_key = ? "
-            "ORDER BY created_at DESC LIMIT ?",
-            (isolation_key, limit),
-        )
-        return [dict(row) for row in cursor.fetchall()]
+        with self._conn_lock:
+            if self._conn is None:
+                return []
+            cursor = self._conn.execute(
+                "SELECT * FROM memory_history WHERE isolation_key = ? "
+                "ORDER BY created_at DESC LIMIT ?",
+                (isolation_key, limit),
+            )
+            return [dict(row) for row in cursor.fetchall()]
 
     async def count(self, isolation_key: str = "") -> int:
         """
@@ -266,16 +274,17 @@ class HistoryStore:
         return await _run_in_sqlite_pool(self._count_sync, isolation_key)
 
     def _count_sync(self, isolation_key: str) -> int:
-        if self._conn is None:
-            return 0
-        if isolation_key:
-            cursor = self._conn.execute(
-                "SELECT COUNT(*) FROM memory_history WHERE isolation_key = ?",
-                (isolation_key,),
-            )
-        else:
-            cursor = self._conn.execute("SELECT COUNT(*) FROM memory_history")
-        return cursor.fetchone()[0]
+        with self._conn_lock:
+            if self._conn is None:
+                return 0
+            if isolation_key:
+                cursor = self._conn.execute(
+                    "SELECT COUNT(*) FROM memory_history WHERE isolation_key = ?",
+                    (isolation_key,),
+                )
+            else:
+                cursor = self._conn.execute("SELECT COUNT(*) FROM memory_history")
+            return cursor.fetchone()[0]
 
     async def close(self) -> None:
         """关闭数据库连接"""
@@ -284,9 +293,10 @@ class HistoryStore:
             logger.info("[HistoryStore] closed")
 
     def _close_sync(self) -> None:
-        if self._conn:
-            self._conn.close()
-            self._conn = None
+        with self._conn_lock:
+            if self._conn:
+                self._conn.close()
+                self._conn = None
 
     # ================================================================
     # 内部
@@ -339,33 +349,34 @@ class HistoryStore:
         role: str = "",
         extra: Optional[Dict[str, Any]] = None,
     ) -> int:
-        if self._conn is None:
-            return -1
-        now = _now_iso()
-        extra_json = json.dumps(extra or {}, ensure_ascii=False)
-        cursor = self._conn.execute(
-            """INSERT INTO memory_history
-               (memory_id, isolation_key, event,
-                old_memory, new_memory, old_status, new_status,
-                change_reason, layer, actor_id, role, extra,
-                created_at, updated_at)
-               VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)""",
-            (
-                memory_id,
-                isolation_key,
-                event,
-                old_memory,
-                new_memory,
-                old_status,
-                new_status,
-                change_reason,
-                layer,
-                actor_id,
-                role,
-                extra_json,
-                now,
-                now,
-            ),
-        )
-        self._conn.commit()
-        return cursor.lastrowid
+        with self._conn_lock:
+            if self._conn is None:
+                return -1
+            now = _now_iso()
+            extra_json = json.dumps(extra or {}, ensure_ascii=False)
+            cursor = self._conn.execute(
+                """INSERT INTO memory_history
+                   (memory_id, isolation_key, event,
+                    old_memory, new_memory, old_status, new_status,
+                    change_reason, layer, actor_id, role, extra,
+                    created_at, updated_at)
+                   VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)""",
+                (
+                    memory_id,
+                    isolation_key,
+                    event,
+                    old_memory,
+                    new_memory,
+                    old_status,
+                    new_status,
+                    change_reason,
+                    layer,
+                    actor_id,
+                    role,
+                    extra_json,
+                    now,
+                    now,
+                ),
+            )
+            self._conn.commit()
+            return cursor.lastrowid

hy_memory/utils/model_call_context.py

@@ -0,0 +1,24 @@
+"""Context propagated to host-provided HTTP transports for cost accounting."""
+
+from collections.abc import Iterator
+from contextlib import contextmanager
+from contextvars import ContextVar, Token
+
+_model_call_stage: ContextVar[str] = ContextVar(
+    "hy_memory_model_call_stage",
+    default="other",
+)
+
+
+def current_model_call_stage() -> str:
+    return _model_call_stage.get()
+
+
+@contextmanager
+def model_call_stage(stage: str | None) -> Iterator[None]:
+    normalized = (stage or "other").strip().lower() or "other"
+    token: Token[str] = _model_call_stage.set(normalized)
+    try:
+        yield
+    finally:
+        _model_call_stage.reset(token)

hy_memory/utils/pipeline_log_writer.py

@@ -80,7 +80,11 @@ class PipelineLogWriter:
                 split_namespace_user,
             )
 
-            tokens = (prompt_tokens or 0) + (completion_tokens or 0) + (total_tokens or 0)
+            # Providers define total_tokens as prompt+completion. Adding all
+            # three fields doubled every cost estimate in this JSONL.
+            tokens = (total_tokens or 0) or (
+                (prompt_tokens or 0) + (completion_tokens or 0)
+            )
             namespace, uid = split_namespace_user(user_id or "")
             if not is_llm_content_logging_enabled():
                 prompt = redact_content(prompt)

memhub/MEMORY.md

@@ -15,12 +15,12 @@ write、search，并把搜索结果格式化为可交给模型的 memory text ch
 
 ```text
 distribution: hy-memory
-version:      2.0.2
+version:      2.0.3
 mode:         按运行时配置选择
 reader:       hybrid
 ```
 
-仓库 pin、lockfile 和运行时版本守卫均固定到公开的 `hy-memory==2.0.2`。
+仓库 pin、lockfile 和运行时版本守卫均固定到公开的 `hy-memory==2.0.3`。
 MemHub 在本地 Runtime 内嵌入 SDK，但不绕过自身的来源授权、模型授权、Evidence、
 Governance 和撤权边界。
 

memhub/README.md

@@ -1,18 +1,30 @@
-# MemHub / One Memory
+# MemHub
 
-MemHub 是一个本地优先的共享记忆产品：把用户授权的多个 AI 产品中的历史对话和新增内容
-汇总到同一份记忆，并让这些 AI 在独立授权后复用它。Hy Memory 是 MemHub 使用的 Python
-SDK 和记忆引擎。One Memory / `one-memory` 是历史模块、数据目录、服务和兼容命令名称。
+MemHub是一个本地优先的共享记忆产品。它把用户明确授权的AI对话和编码记录汇总到一份本地记忆中，再由用户分别授权给支持的AI工具调用。你可以继续使用熟悉的客户端，同时保留跨工具延续的上下文。Hy Memory是MemHub使用的Python SDK和记忆引擎。
+
+# 💡❤️ MemHub 刚刚开源，欢迎一起优化共建！ ❤️💡 
+
+**# 💭 MemHub处于preview阶段，当前公开版本可以下载和体验，部分平台制品、安装路径与宿主体验仍在持续优化验证。安装前请先查看Release里的实际制品和版本说明。**
+
+> 🙋 如果你遇到问题/有建议：欢迎提交Issue，并附上MemHub版本、操作系统与芯片架构、复现步骤、预期结果、实际结果和经过脱敏的日志。请不要上传APIKey、真实对话、个人路径或其他敏感信息。
+
+> 我们会公开跟进可复现的问题，持续发布修复，并根据真实反馈完善安装、同步和召回体验。
+
+**# 感谢每一条反馈，期待你和我们一起打磨MemHub的体验 ：）**
+
+> 💗 MemHub has just been open-sourced and is under active development. The current public release is available for download andevaluation. Some platform packages, installation paths, and host integrations are still being completed and verified. Check the assets
+and release notes before installing.
+If you run into a problem, please open an Issue with the version, platform,reproduction steps, and sanitized logs, We wilreviewreports, publish fixes, and keep the support matrix up to date.
 
 ## 发行与平台合同
 
-公共 Python SDK 为 `hy-memory==2.0.2`。MemHub 源码版本为 `1.0.2`
-（build 6），GitHub 仓库为
+公共 Python SDK 为 `hy-memory==2.0.3`。MemHub 源码版本为 `1.0.3`
+（build 7），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)。
 
-`v1.0.2` 的发行目标是 Linux x64/arm64 Runtime 与 macOS arm64/x64 App 和 Runtime。
+`v1.0.3` 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
@@ -101,7 +113,7 @@ MemHub 的 SQLite 数据默认位于 `~/.one-memory`，不在 npm 包或 Runtime
 
 ## Hy Memory
 
-当前开发依赖固定为公开 PyPI 包 `hy-memory==2.0.2`。SDK 配置、权限边界及 search
+当前开发依赖固定为公开 PyPI 包 `hy-memory==2.0.3`。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`。
 

memhub/SBOM.spdx.json

@@ -3,9 +3,9 @@
   "dataLicense": "CC0-1.0",
   "SPDXID": "SPDXRef-DOCUMENT",
   "name": "one-memory",
-  "documentNamespace": "https://one-memory.local/spdx/2026-10-15T17:26:47Z",
+  "documentNamespace": "https://one-memory.local/spdx/2026-10-17T13:52:48Z",
   "creationInfo": {
-    "created": "2026-10-15T17:26:47Z",
+    "created": "2026-10-17T13:52:48Z",
     "creators": [
       "Tool: one-memory-generate-sbom"
     ]
@@ -2335,7 +2335,7 @@
     {
       "SPDXID": "SPDXRef-Package-hy-memory",
       "name": "hy-memory",
-      "versionInfo": "2.0.2",
+      "versionInfo": "2.0.4",
       "downloadLocation": "NOASSERTION",
       "filesAnalyzed": false,
       "licenseConcluded": "Apache-2.0",

memhub/THIRD_PARTY_NOTICES.md

@@ -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 2026-10-15T17:26:47Z
+Generated from the installed dependency tree on 2026-10-17T13:52:48Z
 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 2.0.2
+## hy-memory 2.0.4
 
 License: Apache-2.0
 

memhub/apps/cli/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@one-memory/cli",
-  "version": "1.0.2",
+  "version": "1.0.4",
   "description": "MemHub bootstrap CLI for installing and managing the local-first shared memory runtime",
   "license": "Apache-2.0",
   "repository": {

memhub/apps/cli/scripts/prepare-publish-manifest.mjs

@@ -12,7 +12,6 @@
  * It is generated rather than hand-written so the two cannot drift: this reads
  * the real package.json and only strips what is provably not needed at runtime.
  */
-import { execFileSync } from "node:child_process";
 import { mkdirSync, readFileSync, writeFileSync } from "node:fs";
 import { createRequire } from "node:module";
 import { dirname, resolve } from "node:path";
@@ -20,12 +19,14 @@ import { fileURLToPath } from "node:url";
 
 const here = dirname(fileURLToPath(import.meta.url));
 const root = resolve(here, "..");
-const repositoryRoot = resolve(root, "../..");
 const require = createRequire(import.meta.url);
-const { resolveBuildVersion } = require("../../../scripts/release-version.cjs");
-const releaseVersion = resolveBuildVersion(process.env, repositoryRoot);
+const { assertSemver } = require("../../../scripts/release-version.cjs");
 const pkgPath = resolve(root, "package.json");
 const pkg = JSON.parse(readFileSync(pkgPath, "utf8"));
+const publishVersion = assertSemver(
+  process.env.MEMHUB_NPM_VERSION?.trim() || pkg.version,
+  "MEMHUB_NPM_VERSION",
+);
 const publishName = process.env.MEMHUB_NPM_PACKAGE?.trim();
 if (!publishName || !/^(?:@[a-z0-9._-]+\/)?[a-z0-9._-]+$/.test(publishName)) {
   throw new Error(
@@ -53,7 +54,7 @@ mkdirSync(outDir, { recursive: true });
 
 const published = { ...pkg };
 published.name = publishName;
-published.version = releaseVersion;
+published.version = publishVersion;
 for (const section of ["dependencies", "devDependencies", "peerDependencies"]) {
   if (!published[section]) continue;
   published[section] = Object.fromEntries(

memhub/apps/cli/src/cli.ts

@@ -81,6 +81,7 @@ import {
   readMemoryWrite,
   readScanRoots,
   removeScanRoot,
+  resetOnboarding,
   submitBackup,
   submitMemoryWrite,
   submitScan,
@@ -330,7 +331,7 @@ import {
   renderVersion,
   renderVersionDocument,
 } from "./usage.js";
-import { CLI_VERSION, PRODUCT } from "./version.js";
+import { CLI_VERSION, PRODUCT, RUNTIME_RELEASE_VERSION } from "./version.js";
 import { runWatchLoop, type WatchEvent, type WatchRead } from "./watch-loop.js";
 
 export interface RunControls {
@@ -416,7 +417,8 @@ async function selectInstallSource(
     throw new Error("release_manifest_url_unavailable");
   }
   const manifestUrl =
-    explicitManifestUrl ?? template.replaceAll("{version}", CLI_VERSION);
+    explicitManifestUrl ??
+    template.replaceAll("{version}", RUNTIME_RELEASE_VERSION);
   const allowedHosts = parseAllowedReleaseHosts(
     environment.MEMHUB_RELEASE_ALLOWED_HOSTS,
   );
@@ -461,7 +463,7 @@ async function resolveRemoteInstallPreview(
 ): Promise<ResolvedRuntimeRelease> {
   return (controls.resolveRuntimeRelease ?? resolveRuntimeRelease)({
     manifestUrl: selection.manifestUrl,
-    expectedVersion: CLI_VERSION,
+    expectedVersion: RUNTIME_RELEASE_VERSION,
     channel: selection.channel,
     platform: process.platform,
     arch: process.arch,
@@ -475,7 +477,7 @@ async function acquireRemoteInstallSource(
 ): Promise<AcquiredRuntimeDistribution> {
   return (controls.acquireRuntimeDistribution ?? acquireRuntimeDistribution)({
     manifestUrl: selection.manifestUrl,
-    expectedVersion: CLI_VERSION,
+    expectedVersion: RUNTIME_RELEASE_VERSION,
     channel: selection.channel,
     platform: process.platform,
     arch: process.arch,
@@ -688,6 +690,36 @@ async function runCommand(
   // `path list` and `adapter list` are read-only: they have no plan to confirm,
   // so they are answered here rather than falling into the two-phase branch
   // (which would reject them for having nothing to apply).
+  // Back to the cold-start screens on purpose. Deleting the app cannot do this
+  // (the completion flag lives with the memories, by design), and deleting the
+  // data to get there would cost the user everything they have stored.
+  if (named === "onboarding" && arguments_.includes("reset")) {
+    if (
+      arguments_.filter((item) => item === "--json").length > 1 ||
+      arguments_.filter((item) => item !== "--json").join(" ") !==
+        "onboarding reset"
+    ) {
+      return usage();
+    }
+    const attachment = await attachToRuntime(
+      resolveRuntimeCliConfig(environment, process.argv[1]).dataDir,
+    );
+    if (attachment === null) {
+      return runtimeUnavailable(descriptor, asJson);
+    }
+    const state = await resetOnboarding(attachment);
+    if (state === null) {
+      return reject(descriptor, asJson, "引导状态端点不可用。");
+    }
+    if (asJson) {
+      process.stdout.write(`${JSON.stringify(state)}\n`);
+      return EXIT_OK;
+    }
+    process.stdout.write(
+      `引导流程已回到首次运行（status=${state.status}），记忆数据未改动。\n`,
+    );
+    return EXIT_OK;
+  }
   if (named === "path" && arguments_.includes("list")) {
     if (
       arguments_.filter((item) => item === "--json").length > 1 ||
@@ -6489,7 +6521,7 @@ function parseInstallOptions(
         ...(releaseManifestUrl ? { releaseManifestUrl } : {}),
         releaseChannel:
           releaseChannel ??
-          (CLI_VERSION.includes("-") ? "prerelease" : "stable"),
+          (RUNTIME_RELEASE_VERSION.includes("-") ? "prerelease" : "stable"),
         installRoot: resolvedInstallRoot,
         home: resolvedHome,
         uid: uid ?? process.getuid?.(),

memhub/apps/cli/src/command-catalog.ts

@@ -651,6 +651,14 @@ export const COMMANDS: readonly CommandDescriptor[] = [
     ],
     exit_codes: [EXIT_OK, EXIT_FAILED, EXIT_USAGE, EXIT_PLAN_BLOCKED],
   },
+  {
+    name: "onboarding",
+    summary:
+      "onboarding reset 回到首次运行的引导流程。只清除引导完成状态：不删记忆数据，也不需要卸载重装。",
+    usage: "memhub onboarding reset [--json]",
+    options: [JSON_OPTION],
+    exit_codes: [EXIT_OK, EXIT_FAILED, EXIT_USAGE],
+  },
   {
     name: "help",
     summary: "显示总帮助或单个命令的用法。",

memhub/apps/cli/src/command-contract.test.ts

@@ -92,6 +92,7 @@ const COMMANDS = [
   "uninstall",
   "upgrade",
   "delete-data",
+  "onboarding",
   "help",
   "version",
   // A6 product commands. Appended, not inserted: the catalog is a stable

memhub/apps/cli/src/context-search.test.ts

@@ -384,6 +384,101 @@ describe("unified context search", () => {
     expect(streams.stdout()).not.toContain("must-not-print");
   });
 
+  it("counts characters as code points so an emoji hit stays grounded", async () => {
+    // Core reports used_characters with Python's len(), which counts code
+    // points: this text is 10 characters there and 12 UTF-16 units here.
+    const text = "消费保 📍 扩展 📡";
+    expect(text.length).toBe(12);
+    expect(Array.from(text).length).toBe(10);
+    transport.post.mockResolvedValue({
+      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,
+              },
+            ],
+          },
+        ],
+        used_characters: 10,
+      },
+    });
+    const streams = captureStreams();
+    try {
+      expect(await run(["recall", "test", "--json"])).toBe(0);
+    } finally {
+      streams.restore();
+    }
+    expect(JSON.parse(streams.stdout())).toMatchObject({
+      state: "grounded",
+      used_characters: 10,
+      items: [{ kind: "raw", messages: [{ text }] }],
+    });
+  });
+
+  it("still rejects a response whose character total does not match", async () => {
+    // The same text counted as UTF-16 units is exactly what Core does not mean,
+    // so the cross-check must keep failing instead of accepting both totals.
+    const text = "消费保 📍 扩展 📡";
+    transport.post.mockResolvedValue({
+      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,
+              },
+            ],
+          },
+        ],
+        used_characters: 12,
+      },
+    });
+    const streams = captureStreams();
+    try {
+      expect(await run(["recall", "test", "--json"])).toBe(1);
+    } finally {
+      streams.restore();
+    }
+    expect(JSON.parse(streams.stdout())).toMatchObject({
+      state: "invalid",
+      code: "response_invalid",
+    });
+  });
+
   it("reports endpoint denial without exposing the response body", async () => {
     transport.post.mockResolvedValue({
       status: 403,

memhub/apps/cli/src/context-search.ts

@@ -181,9 +181,9 @@ function validateContextResponse(
     items.push(item);
     characters +=
       item.kind === "memory"
-        ? String(item.content).length
+        ? characterCount(item.content)
         : (item.messages as Array<Record<string, unknown>>).reduce(
-            (sum, message) => sum + String(message.text).length,
+            (sum, message) => sum + characterCount(message.text),
             0,
           );
   }
@@ -330,6 +330,17 @@ function record(value: unknown): Record<string, unknown> | null {
     : null;
 }
 
+/**
+ * Count characters the way Core does. Python's len() counts Unicode code
+ * points; String.prototype.length counts UTF-16 code units, so an emoji outside
+ * the BMP is one character to Core and two here. Comparing the two totals made
+ * every response containing an emoji fail the used_characters cross-check and
+ * surface as response_invalid.
+ */
+function characterCount(value: unknown): number {
+  return Array.from(String(value)).length;
+}
+
 function boundedText(
   value: unknown,
   maximum: number,

memhub/apps/cli/src/core-client.ts

@@ -397,6 +397,41 @@ export async function readMemoryWrite(
   }
 }
 
+export interface OnboardingState {
+  status: string;
+  selected_source_ids: string[];
+  started_at: string | null;
+  completed_at: string | null;
+}
+
+/** Send onboarding back to its first-run state.
+ *
+ * Returns null when the endpoint is not there, so an older runtime degrades to
+ * "not available" instead of failing the command. Only the completion flag
+ * moves: memories, sources and settings live in other tables, which is what
+ * makes this the honest way to see the cold-start screens again.
+ */
+export async function resetOnboarding(
+  attachment: RuntimeAttachmentLike,
+): Promise<OnboardingState | null> {
+  try {
+    const response = await axios.post<OnboardingState>(
+      `${baseUrlOf(attachment)}/api/v1/onboarding/reset`,
+      undefined,
+      {
+        headers: { Authorization: `Bearer ${attachment.apiToken}` },
+        timeout: 15_000,
+        proxy: false,
+        maxRedirects: 0,
+        maxContentLength: MAX_RESPONSE_BYTES,
+      },
+    );
+    return response.data;
+  } catch {
+    return null;
+  }
+}
+
 export async function readScanRoots(
   attachment: RuntimeAttachmentLike,
 ): Promise<{

memhub/apps/cli/src/publish-manifest.test.ts

@@ -48,7 +48,11 @@ describe("publish manifest is production-safe", () => {
     execFileSync("node", ["scripts/prepare-publish-manifest.mjs"], {
       cwd: root,
       stdio: "pipe",
-      env: { ...process.env, MEMHUB_NPM_PACKAGE: "@example/memhub-cli" },
+      env: {
+        ...process.env,
+        MEMHUB_NPM_PACKAGE: "@example/memhub-cli",
+        MEMHUB_NPM_VERSION: "1.0.5",
+      },
     });
     const published = JSON.parse(
       readFileSync(path.join(root, "dist-publish", "package.json"), "utf8"),
@@ -56,6 +60,7 @@ describe("publish manifest is production-safe", () => {
     expect(published.dependencies).toBeUndefined();
     expect(published.private).toBe(false);
     expect(published.name).toBe("@example/memhub-cli");
+    expect(published.version).toBe("1.0.5");
     expect(published.bin).toMatchObject({
       memhub: "dist/one-memory.cjs",
       "memhub-cli": "dist/one-memory.cjs",

memhub/apps/cli/src/version.ts

@@ -1,5 +1,9 @@
 declare const __ONE_MEMORY_VERSION__: string;
 declare const __ONE_MEMORY_BUILD_NUMBER__: string;
+declare const __MEMHUB_DEFAULT_RUNTIME_VERSION__: string;
+
+const SEMVER =
+  /^(0|[1-9]\d*)\.(0|[1-9]\d*)\.(0|[1-9]\d*)(?:-[0-9A-Za-z-]+(?:\.[0-9A-Za-z-]+)*)?(?:\+[0-9A-Za-z-]+(?:\.[0-9A-Za-z-]+)*)?$/;
 
 const embeddedVersion =
   typeof __ONE_MEMORY_VERSION__ === "string"
@@ -9,13 +13,18 @@ const embeddedBuildNumber =
   typeof __ONE_MEMORY_BUILD_NUMBER__ === "string"
     ? __ONE_MEMORY_BUILD_NUMBER__.trim()
     : "";
+const embeddedRuntimeVersion =
+  typeof __MEMHUB_DEFAULT_RUNTIME_VERSION__ === "string"
+    ? __MEMHUB_DEFAULT_RUNTIME_VERSION__.trim()
+    : "";
 
-export const CLI_VERSION =
-  /^(0|[1-9]\d*)\.(0|[1-9]\d*)\.(0|[1-9]\d*)(?:-[0-9A-Za-z-]+(?:\.[0-9A-Za-z-]+)*)?(?:\+[0-9A-Za-z-]+(?:\.[0-9A-Za-z-]+)*)?$/.test(
-    embeddedVersion,
-  )
-    ? embeddedVersion
-    : "development";
+export const CLI_VERSION = SEMVER.test(embeddedVersion)
+  ? embeddedVersion
+  : "development";
+
+export const RUNTIME_RELEASE_VERSION = SEMVER.test(embeddedRuntimeVersion)
+  ? embeddedRuntimeVersion
+  : CLI_VERSION;
 
 export const CLI_BUILD_NUMBER = /^\d+$/.test(embeddedBuildNumber)
   ? Number(embeddedBuildNumber)

memhub/apps/cli/tsup.config.ts

@@ -2,11 +2,22 @@ import { createRequire } from "node:module";
 import { defineConfig } from "tsup";
 
 const require = createRequire(import.meta.url);
-const { resolveBuildVersion } =
+const { assertSemver, resolveBuildVersion } =
   require("../../scripts/release-version.cjs") as {
+    assertSemver(value: string, label?: string): string;
     resolveBuildVersion(environment?: NodeJS.ProcessEnv, root?: string): string;
   };
 const releaseVersion = resolveBuildVersion(process.env);
+const packageVersion = (require("./package.json") as { version: string })
+  .version;
+const cliVersion = assertSemver(
+  process.env.MEMHUB_NPM_VERSION?.trim() || packageVersion,
+  "MEMHUB_NPM_VERSION",
+);
+const runtimeReleaseVersion = assertSemver(
+  process.env.MEMHUB_DEFAULT_RUNTIME_VERSION?.trim() || releaseVersion,
+  "MEMHUB_DEFAULT_RUNTIME_VERSION",
+);
 
 export default defineConfig({
   entry: { "one-memory": "src/entry.ts" },
@@ -15,7 +26,7 @@ export default defineConfig({
   outDir: "dist",
   clean: true,
   define: {
-    __ONE_MEMORY_VERSION__: JSON.stringify(releaseVersion),
+    __ONE_MEMORY_VERSION__: JSON.stringify(cliVersion),
     __ONE_MEMORY_BUILD_NUMBER__: JSON.stringify(
       process.env.ONE_MEMORY_BUILD_NUMBER ?? "unknown",
     ),
@@ -25,6 +36,7 @@ export default defineConfig({
     __MEMHUB_DEFAULT_RELEASE_MANIFEST_URL_TEMPLATE__: JSON.stringify(
       process.env.MEMHUB_DEFAULT_RELEASE_MANIFEST_URL_TEMPLATE ?? "",
     ),
+    __MEMHUB_DEFAULT_RUNTIME_VERSION__: JSON.stringify(runtimeReleaseVersion),
   },
   // The shared operations package ships TypeScript sources, so it must be
   // inlined into the CJS bundle the same way runtime-control is; otherwise the

memhub/apps/cli/vitest.config.ts

@@ -11,6 +11,7 @@ export default defineConfig({
     __ONE_MEMORY_BUILD_NUMBER__: JSON.stringify("unknown"),
     __ONE_MEMORY_COMMIT__: JSON.stringify("unknown"),
     __MEMHUB_DEFAULT_RELEASE_MANIFEST_URL_TEMPLATE__: JSON.stringify(""),
+    __MEMHUB_DEFAULT_RUNTIME_VERSION__: JSON.stringify(version),
   },
   test: {
     testTimeout: 120_000,

memhub/apps/core/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "uv_build"
 
 [project]
 name = "one-memory"
-version = "1.0.2"
+version = "1.0.4"
 description = "MemHub local-first memory core"
 readme = "README.md"
 requires-python = ">=3.12"
@@ -12,7 +12,7 @@ dependencies = [
     "fastapi>=0.141.1",
     "filelock>=3.32.4",
     # Public SDK version validated against MemHub's permission and recall contracts.
-    "hy-memory==2.0.2",
+    "hy-memory==2.0.4",
     # Retrieval's derived vector index directly consumes these locked versions.
     "chromadb==1.5.9",
     "numpy==2.5.2",

memhub/apps/core/src/one_memory/api.py

@@ -100,7 +100,13 @@ from one_memory.imports.access import (
     SourceAccessError,
     root_identity_matches,
 )
-from one_memory.imports.adapters import SourceAdapter, SyncCompleteness
+from one_memory.imports.adapters import (
+    ScanFailure,
+    ScannedEvidence,
+    ScanOptions,
+    SourceAdapter,
+    SyncCompleteness,
+)
 from one_memory.imports.candidates import (
     SourceCandidateError,
     SourceCandidateSelection,
@@ -111,7 +117,11 @@ from one_memory.imports.connections import (
     ConnectionIntentReceipt,
     SourceConnectionStore,
 )
-from one_memory.imports.grants import SourceGrantReceipt, SourceGrantStore
+from one_memory.imports.grants import (
+    SourceAuthorizationState,
+    SourceGrantReceipt,
+    SourceGrantStore,
+)
 from one_memory.imports.host_packages import CapabilityState, HostPackageCapabilities
 from one_memory.imports.jobs import (
     ImportIssueRecord,
@@ -175,7 +185,12 @@ from one_memory.memory_queue import (
     resolve_extraction_state,
 )
 from one_memory.memory_server.config_state import ConfigStore
-from one_memory.memory_tasks import MemoryTask, MemoryTaskStatus, MemoryTaskStore
+from one_memory.memory_tasks import (
+    NON_RETRYABLE_POST_WRITE_CODES,
+    MemoryTask,
+    MemoryTaskStatus,
+    MemoryTaskStore,
+)
 from one_memory.model_presets import seed_preset_profile
 from one_memory.model_probe import ModelProber, ModelProbeResult
 from one_memory.model_profiles import (
@@ -190,6 +205,7 @@ from one_memory.model_profiles import (
     ModelProfileStore,
     ModelProfileUnavailableError,
 )
+from one_memory.model_request_audit import AuditedModelTransport, ModelRequestAuditStore
 from one_memory.model_request_budget import BudgetedModelTransport, ModelRequestBudget
 from one_memory.onboarding import OnboardingState, OnboardingStore
 from one_memory.recall import GovernedRecallTarget, RecallReceipt, RecallRequest
@@ -281,6 +297,9 @@ class SourceResponse(BaseModel):
     candidate_revision: int = 0
     candidate_change_allowed: bool = False
     candidate_change_reason: str | None = None
+    #: Why `authorized` is false, so the UI can ask for the one thing that
+    #: would fix it instead of offering every recovery at once.
+    authorization_state: SourceAuthorizationState = SourceAuthorizationState.OK
 
 
 class PreviewSourceCandidateRequest(BaseModel):
@@ -313,6 +332,22 @@ class SourceSyncState(BaseModel):
     sync_enabled: bool
 
 
+class SedimentationRequest(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+
+    enabled: StrictBool
+
+
+class SedimentationState(BaseModel):
+    model_config = ConfigDict(frozen=True, extra="forbid")
+
+    enabled: bool
+    #: What the switch means for the model bill and for what still runs. Read
+    #: by the app so the setting can say so without hardcoding the wording.
+    pending_retained: bool = True
+    imports_unaffected: bool = True
+
+
 class DetectionResponse(BaseModel):
     source_id: str
     detected: bool
@@ -339,7 +374,7 @@ class ConnectSourcesRequest(BaseModel):
     import_upload_id: UUID | None = None
 
 
-ONBOARDING_LARGE_HISTORY_SESSIONS = 200
+ONBOARDING_LARGE_HISTORY_USER_TURNS = 200
 ONBOARDING_HISTORY_WINDOW = timedelta(days=3)
 
 
@@ -886,6 +921,7 @@ def create_app(
             )
             app.state.import_job_store = ImportJobStore(database_path)
             app.state.onboarding_store = OnboardingStore(database_path)
+            app.state.history_import_gate_released = False
             app.state.extraction_settings_store = ExtractionSettingsStore(database_path)
             app.state.source_grant_store = SourceGrantStore(
                 database_path,
@@ -908,6 +944,7 @@ def create_app(
             app.state.source_permission_locks = {
                 source_id: asyncio.Lock() for source_id in app.state.source_catalog
             }
+            app.state.sedimentation_lock = asyncio.Lock()
             _follow_workbuddy_home(app, runtime_settings.source_home)
             app.state.source_detection = {}
             app.state.secret_store = secret_store or KeyringSecretStore()
@@ -974,6 +1011,7 @@ def create_app(
                 else None
             )
             app.state.model_request_budget = model_request_budget
+            app.state.model_request_audit_store = ModelRequestAuditStore(database_path)
             app.state.memory_answerer = MemoryAnswerer(
                 governance=app.state.memory_governance_store,
                 filter_memories=app.state.source_read_access.memories,
@@ -1016,10 +1054,17 @@ def create_app(
                     profile=profile,
                     api_key=api_key,
                     data_dir=data_dir,
-                    http_transport_factory=(
-                        (lambda: BudgetedModelTransport(model_request_budget))
-                        if model_request_budget is not None
-                        else None
+                    http_transport_factory=lambda: AuditedModelTransport(
+                        app.state.model_request_audit_store,
+                        profile_id=profile.profile_id,
+                        profile_revision=profile.revision,
+                        provider_kind=str(profile.provider_kind),
+                        model=profile.model,
+                        inner=(
+                            BudgetedModelTransport(model_request_budget)
+                            if model_request_budget is not None
+                            else None
+                        ),
                     ),
                 ),
                 # `None` means "whatever the active profile allows". Pinned to
@@ -1032,6 +1077,7 @@ def create_app(
                 max_concurrent_tasks=runtime_settings.precipitation_max_concurrent_tasks,
                 source_writes_permitted=lambda source_id: _source_writes_on(app, source_id),
                 service_levels=app.state.service_levels,
+                sedimentation_enabled=lambda: _config_store(app).sedimentation_enabled(),
             )
             # C2 owns the single production budget (handle 5s + reserve 250ms,
             # total 5.25s, lease 10s, renew 1s, claim 250ms). The composition
@@ -1041,26 +1087,41 @@ def create_app(
             drain_budget = production_drain_budget()
 
             async def reconcile_unknown_writes() -> None:
-                await app.state.memory_processing_supervisor.reconcile_unknown_outputs()
-                unknown = await asyncio.to_thread(memory_write_queue.unknown_jobs, limit=128)
+                await app.state.memory_processing_supervisor.reconcile_unknown_outputs(
+                    exclude_error_codes=NON_RETRYABLE_POST_WRITE_CODES,
+                )
+                unknown = await asyncio.to_thread(
+                    memory_write_queue.unknown_jobs,
+                    limit=128,
+                    exclude_error_codes=NON_RETRYABLE_POST_WRITE_CODES,
+                )
                 for job in unknown:
                     await asyncio.to_thread(
                         memory_write_queue.resolve_unknown_from_linked_output,
                         job.job_id,
                     )
                 await asyncio.to_thread(
-                    memory_write_queue.retry_timed_out_unknown_after_reconcile,
-                    older_than=datetime.now(UTC) - timedelta(minutes=5),
+                    memory_write_queue.retry_first_unknown_after_reconcile,
+                    older_than=datetime.now(UTC) - timedelta(seconds=10),
                     limit=32,
                 )
+                await asyncio.to_thread(
+                    app.state.memory_task_store.retry_first_unknown_after_reconcile,
+                    older_than=datetime.now(UTC) - timedelta(seconds=10),
+                    limit=128,
+                )
 
             async def admit_hy_invocation(
                 source_id: str,
                 mark_invoked: Callable[[], Awaitable[bool]],
             ) -> bool:
                 """Serialize pause with the last boundary before an HY call."""
-                lock = app.state.source_permission_locks.setdefault(source_id, asyncio.Lock())
-                async with lock:
+                source_lock = app.state.source_permission_locks.setdefault(
+                    source_id, asyncio.Lock()
+                )
+                async with app.state.sedimentation_lock, source_lock:
+                    if not _config_store(app).sedimentation_enabled():
+                        raise RuntimeError("model_sedimentation_disabled")
                     if not _source_writes_on(app, source_id):
                         raise RuntimeError("model_content_source_disabled")
                     return await mark_invoked()
@@ -1098,8 +1159,18 @@ def create_app(
                     + 10
                 ),
                 reconcile_unknown=reconcile_unknown_writes,
-                source_writes_permitted=lambda source_id: _source_writes_on(app, source_id),
-                source_claim_enabled=lambda source_id: _config_store(app).sync_enabled(source_id),
+                source_writes_permitted=lambda source_id: _sedimentation_claim_enabled(
+                    app, source_id
+                ),
+                # The claim gate has to be the same question the write gate
+                # asks, not a second answer to it. It used to be
+                # `sync_enabled`, the per-app toggle, and `one-memory-manual` is
+                # not a connected app: it has no entry in the configuration
+                # document, so that predicate answers False for it while the
+                # write gate admits it by its grant. Every manual note's job was
+                # therefore skipped on every drain tick — the note stayed
+                # "saved, waiting" with no task behind it, forever.
+                source_claim_enabled=lambda source_id: _sedimentation_claim_enabled(app, source_id),
             )
             app.state.memory_write_drain = memory_write_drain
             app.state.memory_write_drain_task = watched_task(
@@ -1128,10 +1199,7 @@ def create_app(
                 for chunk in _conversation_chunks(conversation, max_turns=12, completed_only=True):
                     if (
                         record not in chunk
-                        or not any(
-                            item.role == "user" and item.written_by.value == "person"
-                            for item in chunk
-                        )
+                        or not any(item.role == "user" for item in chunk)
                         or not any(item.role == "assistant" for item in chunk)
                     ):
                         continue
@@ -1744,6 +1812,30 @@ def create_app(
             "events": events,
         }
 
+    @app.get("/api/v1/diagnostics/model-usage", dependencies=protected)
+    def diagnostic_model_usage(
+        since: datetime | None = Query(default=None),  # noqa: B008
+        request_id: str | None = Query(  # noqa: B008
+            default=None, min_length=1, max_length=512
+        ),
+        limit: int = Query(default=200, ge=1, le=2_000),  # noqa: B008
+    ) -> dict[str, object]:
+        """Return content-free precipitation usage for provider reconciliation."""
+
+        normalized_since = (
+            (since if since.tzinfo is not None else since.replace(tzinfo=UTC))
+            if since is not None
+            else None
+        )
+        report = app.state.model_request_audit_store.report(
+            since=normalized_since.astimezone(UTC).isoformat()
+            if normalized_since is not None
+            else None,
+            request_id=request_id,
+            limit=limit,
+        )
+        return {"schema_version": 1, **report}
+
     @app.get(
         "/api/v1/runtime/identity",
         response_model=HealthResponse,
@@ -1881,6 +1973,20 @@ def create_app(
         """Record UI completion only; finishing or skipping is never model consent."""
         return await asyncio.to_thread(_onboarding_store(app).complete)
 
+    @app.post(
+        "/api/v1/onboarding/reset",
+        response_model=OnboardingState,
+        dependencies=protected,
+    )
+    async def reset_onboarding() -> OnboardingState:
+        """Go back to the first-run screens. Only the completion flag moves.
+
+        Memories, sources and settings live in other tables and are untouched,
+        which is the point: someone who wants to see the cold-start flow again
+        should not have to delete anything to do it.
+        """
+        return await asyncio.to_thread(_onboarding_store(app).reset)
+
     @app.get(
         "/api/v1/sources",
         response_model=list[SourceResponse],
@@ -1918,6 +2024,11 @@ def create_app(
                     candidate_change_reason=app.state.source_candidate_store.change_reason(
                         registration
                     ),
+                    authorization_state=grant_store.authorization_state(
+                        user_id="local",
+                        source_id=source_id,
+                        candidate=registration.candidate.path,
+                    ),
                 )
                 for source_id, registration in _catalog(app).items()
             ]
@@ -2229,24 +2340,39 @@ def create_app(
                 raise HTTPException(status_code=409, detail=error.code) from None
             if grant is None:
                 raise HTTPException(status_code=409, detail="grant_required")
-            await _prepare_source_resume(source_id)
+            try:
+                await _prepare_source_resume(source_id)
+            except BaseException as error:
+                raise _sync_refused("sync_prepare_failed") from error
         else:
-            await _live_sync(app).stop_source(source_id)
+            try:
+                await _live_sync(app).stop_source(source_id)
+            except BaseException as error:
+                raise _sync_refused("sync_prepare_failed") from error
         try:
             await asyncio.to_thread(_config_store(app).set_sync, source_id, request.enabled)
-        except BaseException:
+        except BaseException as error:
             if request.enabled:
                 await _live_sync(app).stop_source(source_id)
             else:
                 await _prepare_source_resume(source_id)
                 await _activate_source_resume(source_id)
             await _refresh_live_watch(app)
-            raise
+            raise _sync_refused("sync_config_write_failed") from error
         try:
             if request.enabled:
-                await _activate_source_resume(source_id)
-            await _refresh_live_watch(app)
+                try:
+                    await _activate_source_resume(source_id)
+                except BaseException as error:
+                    raise _sync_refused("sync_resume_failed") from error
+            try:
+                await _refresh_live_watch(app)
+            except BaseException as error:
+                raise _sync_refused("sync_watcher_failed") from error
         except BaseException:
+            # Rolls back the desired state and settles the watcher exactly as
+            # before; the only change is that the raise above now carries a
+            # code saying which stage failed.
             await asyncio.to_thread(_config_store(app).set_sync, source_id, previous)
             if request.enabled:
                 await _live_sync(app).stop_source(source_id)
@@ -2377,6 +2503,8 @@ def create_app(
     async def connect_sources(
         request: ConnectSourcesRequest,
     ) -> ConnectSourcesResponse:
+        if request.onboarding:
+            app.state.history_import_gate_released = False
         source_ids = tuple(dict.fromkeys(request.source_ids))
         registrations = [_require_registration(app, source_id) for source_id in source_ids]
         if request.request_id is None:
@@ -2676,6 +2804,12 @@ def create_app(
         # D75: work queued for the previous model continues with this one.
         # Only tasks that never reached the model are re-bound.
         await asyncio.to_thread(app.state.memory_task_store.rebind_unstarted_profile, profile)
+        # A missing profile is an event-driven wait, not a provider outage.
+        # Release those durable rows only now that activation has changed the
+        # condition; the ordinary drain poll will claim them without a restart.
+        await asyncio.to_thread(
+            app.state.memory_write_queue.readmit_model_profile_waiters
+        )
         # Not awaited: the catch-up reads every conversation ever imported, and
         # the answer to "which profile is active" does not depend on it.
         _start_backfill(app)
@@ -3542,6 +3676,48 @@ def create_app(
     def put_extraction_settings(request: ExtractionWindowUpdate) -> ExtractionSettings:
         return _memory_processing_supervisor(app).set_extraction_window(request.window_preset)
 
+    @app.get(
+        "/api/v1/sedimentation",
+        response_model=SedimentationState,
+        dependencies=protected,
+    )
+    def get_sedimentation() -> SedimentationState:
+        """Whether the model may still turn conversations into memories."""
+        return SedimentationState(enabled=_config_store(app).sedimentation_enabled())
+
+    @app.put(
+        "/api/v1/sedimentation",
+        response_model=SedimentationState,
+        dependencies=protected,
+    )
+    async def put_sedimentation(request: SedimentationRequest) -> SedimentationState:
+        """Start or stop model-driven sedimentation, keeping the configuration.
+
+        Stopping is the user's cost control: nothing further goes to the model
+        and already queued work waits rather than failing, so starting again
+        resumes it. Importing conversations is not affected either way.
+        """
+        store = _config_store(app)
+        changed = False
+        async with app.state.sedimentation_lock:
+            previous = store.sedimentation_enabled()
+            if previous is not request.enabled:
+                await asyncio.to_thread(store.set_sedimentation_enabled, request.enabled)
+                changed = True
+                diagnostic_event(
+                    level="info",
+                    component="memory.sedimentation",
+                    event="memory_sedimentation.toggled",
+                    outcome="succeeded",
+                    details={"enabled": request.enabled},
+                )
+        if changed and request.enabled:
+            # Work parked by the stop is picked up again rather than waiting
+            # for the next natural dispatch, but not on this request's time:
+            # see _resume_sedimentation.
+            _resume_sedimentation(app)
+        return SedimentationState(enabled=store.sedimentation_enabled())
+
     @app.post(
         "/api/v1/memory-processing/manual-extract",
         response_model=ManualExtractReceipt,
@@ -3902,6 +4078,7 @@ async def _authorize_answer_generation(
     # The explicit question is this operation's input, not a standing permission for history.
     if not question.strip() or not grounding:
         raise MemoryAnswerError("memory_answer_generation_not_authorized")
+
     def check() -> None:
         if _model_profile_store(app).active_identity() != (
             profile.profile_id,
@@ -3925,6 +4102,7 @@ async def _authorize_answer_generation(
     await recheck()
     return recheck
 
+
 class _LazyASGIApp:
     """Compatibility ASGI app that is built only when a server imports it.
 
@@ -3956,6 +4134,16 @@ class _LazyASGIApp:
 app = _LazyASGIApp()
 
 
+def _sync_refused(code: str) -> HTTPException:
+    """A sync refusal that says which stage refused.
+
+    Turning the switch on takes four steps: authorize, prepare, write, resume.
+    The UI used to be told only that the switch had sprung back, and every
+    stage needs a different fix, so each one gets its own code.
+    """
+    return HTTPException(status_code=409, detail=code)
+
+
 def _granted_roots(
     source_ids: Iterable[str],
     grant_store: SourceGrantStore,
@@ -4007,6 +4195,7 @@ def _source_response(
     totals: tuple[int, int],
     *,
     candidate_change_reason: str | None = None,
+    authorization_state: SourceAuthorizationState = SourceAuthorizationState.OK,
 ) -> SourceResponse:
     location = inspect_source_location(
         registration.descriptor.source_id, registration.candidate.path
@@ -4219,7 +4408,7 @@ async def _onboarding_history_after(
     home: Path,
     now: datetime | None = None,
 ) -> datetime | None:
-    """Return one shared cutoff when any selected harness has large history.
+    """Return one shared cutoff when any selected harness has large, old history.
 
     The cutoff is decided before any history job is created, then persisted on
     every accepted job. This is deliberately global: independently limiting
@@ -4228,6 +4417,8 @@ async def _onboarding_history_after(
     """
     if not request.onboarding:
         return None
+    anchor = now or datetime.now(UTC)
+    cutoff = anchor - ONBOARDING_HISTORY_WINDOW
     revisions = request.source_candidate_revisions or {}
     for registration in registrations:
         source_id = registration.descriptor.source_id
@@ -4255,29 +4446,52 @@ async def _onboarding_history_after(
                 home=home,
             )
             adapter = registration.create_adapter(grant)
-            count_sessions = cast(
-                Callable[..., Awaitable[int]] | None,
-                getattr(adapter, "count_sessions", None),
-            )
-            if count_sessions is not None:
-                session_count = await count_sessions(limit=ONBOARDING_LARGE_HISTORY_SESSIONS + 1)
-            else:
-                pending = _install_state(adapter)
-                if pending is None:
-                    continue
-                session_count = (await pending).session_count
+            if await _has_large_old_history(adapter, cutoff=cutoff):
+                return cutoff
         except (OSError, SourceAccessError):
             # A supported source whose metadata count cannot be completed is
             # treated conservatively. The later connection still owns the
             # precise authorization error, while onboarding never falls back
             # to an unbounded history import.
-            session_count = ONBOARDING_LARGE_HISTORY_SESSIONS + 1
-        if session_count > ONBOARDING_LARGE_HISTORY_SESSIONS:
-            anchor = now or datetime.now(UTC)
-            return anchor - ONBOARDING_HISTORY_WINDOW
+            return cutoff
     return None
 
 
+async def _has_large_old_history(adapter: SourceAdapter, *, cutoff: datetime) -> bool:
+    """Probe one authorized Harness with the same parsers the import will use.
+
+    Stop as soon as both policy facts are proven: more than 200 genuine person
+    turns and at least one record older than the three-day cutoff. Counting
+    parsed Evidence instead of transcript files prevents one giant coding
+    session from bypassing the limit, while reusing the adapters keeps host
+    envelopes and tool-shaped user records out of the count.
+    """
+
+    user_turns = 0
+    has_old_data = False
+    async for event in adapter.scan(ScanOptions()):
+        if isinstance(event, ScanFailure):
+            # A partial probe must never turn into an unbounded first import.
+            return True
+        if not isinstance(event, ScannedEvidence):
+            continue
+        observed = event.evidence.created_at
+        if observed is None and event.cursor.observed_modified_ns > 0:
+            observed = datetime.fromtimestamp(
+                event.cursor.observed_modified_ns / 1_000_000_000,
+                tz=UTC,
+            )
+        if observed is not None:
+            if observed.tzinfo is None:
+                observed = observed.replace(tzinfo=UTC)
+            has_old_data = has_old_data or observed.astimezone(UTC) < cutoff
+        if event.evidence.role == "user":
+            user_turns += 1
+        if user_turns > ONBOARDING_LARGE_HISTORY_USER_TURNS and has_old_data:
+            return True
+    return False
+
+
 def _install_state(adapter: object) -> Awaitable[SourceInstallState] | None:
     """Ask an adapter for the two-level answer when it has one.
 
@@ -4569,7 +4783,7 @@ async def _keep_loop_heartbeat(watchdog: LoopWatchdog) -> None:
 async def _emit_resource_snapshots(interval_seconds: float = 60.0) -> None:
     while True:
         diagnostic_event(
-            level="info",
+            level="debug",
             component="core.resources",
             event="core.resource_snapshot",
             outcome="observed",
@@ -4594,6 +4808,33 @@ def _start_backfill(app: FastAPI) -> None:
     )
 
 
+def _resume_sedimentation(app: FastAPI) -> None:
+    """Pick the work parked by a stop back up, off the caller's clock.
+
+    Turning the switch back on is a decision, not a job the caller waits for:
+    the choice is already stored, and what follows it is maintenance over every
+    unfinished task followed by a scan of the whole queue. On a real backlog
+    that runs for many minutes, so awaiting it here left the PUT unanswered
+    until the client gave up — the renderer showed "启用未能确认" for a switch
+    that had in fact been saved, and the switch only appeared to turn on once
+    the follow-up read came back. Measured: one toggle, twenty-plus minutes,
+    no response, client timeout at thirty seconds.
+
+    Repeated toggles reuse the running walk; it reads the switch as it goes, so
+    a stop part-way through is honoured on the next batch instead of needing a
+    cancel.
+    """
+    running = cast("asyncio.Task[None] | None", getattr(app.state, "sedimentation_resume", None))
+    if running is not None and not running.done():
+        return
+    app.state.sedimentation_resume = watched_task(
+        asyncio.get_running_loop(),
+        lambda: _memory_processing_supervisor(app).resume_allowed(),
+        component="memory.sedimentation",
+        name="one-memory-sedimentation-resume",
+    )
+
+
 def _start_visibility_recompute(app: FastAPI) -> None:
     """Keep the stored visibility conclusions current, in the background.
 
@@ -4837,6 +5078,52 @@ def _source_writes_on(app: FastAPI, source_id: str) -> bool:
     return _source_processing_enabled(app, source_id)
 
 
+def _sedimentation_claim_enabled(app: FastAPI, source_id: str) -> bool:
+    """Claim queued model work only when policy, source and model all admit it."""
+
+    return (
+        _config_store(app).sedimentation_enabled()
+        and _model_profile_store(app).active_identity() is not None
+        and _history_import_gate_released(app)
+        and _source_writes_on(app, source_id)
+    )
+
+
+def _history_import_gate_released(app: FastAPI) -> bool:
+    """Hold first sedimentation until every accepted history import is terminal."""
+
+    if getattr(app.state, "history_import_gate_released", False):
+        return True
+    onboarding = getattr(app.state, "onboarding_store", None)
+    jobs = getattr(app.state, "import_job_store", None)
+    if onboarding is None or jobs is None:
+        return True
+    state = onboarding.get()
+    if state.status != "importing" or not state.selected_source_ids:
+        app.state.history_import_gate_released = True
+        return True
+    selected = set(state.selected_source_ids)
+    matching = [
+        job
+        for job in jobs.list_jobs()
+        if job.operation == "history"
+        and job.source_id in selected
+        and (state.started_at is None or job.created_at >= state.started_at)
+    ]
+    if {job.source_id for job in matching} != selected:
+        return False
+    active = {
+        ImportStatus.PENDING,
+        ImportStatus.RUNNING,
+        ImportStatus.PAUSED,
+        ImportStatus.INTERRUPTED,
+    }
+    released = all(job.status not in active for job in matching)
+    if released:
+        app.state.history_import_gate_released = True
+    return released
+
+
 def _model_prober(app: FastAPI) -> ModelProber:
     return cast(ModelProber, app.state.model_prober)
 

memhub/apps/core/src/one_memory/background_work.py

@@ -175,6 +175,22 @@ async def in_interactive_read[T](
     return await in_read_pool(interactive_read_pool(), work, *args, **kwargs)
 
 
+def _retrieve[T](future: "asyncio.Future[T]") -> None:
+    """Mark a shielded read's outcome as seen.
+
+    `in_read_pool` shields the wrapped worker, so a caller that gives up and
+    cancels leaves the read running to completion. When that read then fails,
+    nothing awaits the future any more: asyncio reports the orphan at GC as
+    "Future exception was never retrieved", the loop handler turns it into a
+    `core.loop.task_exception` error diagnostic, and a lane that degraded on
+    its own budget -- correctly, and already reported -- looks like a crashed
+    task on top of it. Retrieving here keeps the diagnostics honest. Awaiting
+    still raises for callers that are still there, so no failure is hidden.
+    """
+    if not future.cancelled():
+        future.exception()
+
+
 async def in_read_pool[T](
     pool: ThreadPoolExecutor,
     work: Callable[..., T],
@@ -193,7 +209,9 @@ async def in_read_pool[T](
             _interactive_reads -= 1
 
     worker.add_done_callback(released)
-    return await asyncio.shield(asyncio.wrap_future(worker))
+    pending = asyncio.wrap_future(worker)
+    pending.add_done_callback(_retrieve)
+    return await asyncio.shield(pending)
 
 
 async def wait_for_interactive_reads() -> None:

memhub/apps/core/src/one_memory/cli.py

@@ -266,7 +266,13 @@ def serve(*, maintenance: bool = False, reader: bool = False) -> None:
         )
         application = create_reader_app(reader_settings)
         print(f"reader_port:{port}", file=sys.stderr, flush=True)
-        uvicorn.run(application, host=settings.host, port=port, reload=False)
+        uvicorn.run(
+            application,
+            host=settings.host,
+            port=port,
+            reload=False,
+            access_log=False,
+        )
         return
 
     from one_memory.api import create_app
@@ -297,9 +303,21 @@ def serve(*, maintenance: bool = False, reader: bool = False) -> None:
     )
     if maintenance:
         with CoreRuntime(settings, publish_receipt=False):
-            uvicorn.run(application, host=settings.host, port=settings.port, reload=False)
+            uvicorn.run(
+                application,
+                host=settings.host,
+                port=settings.port,
+                reload=False,
+                access_log=False,
+            )
     else:
-        uvicorn.run(application, host=settings.host, port=settings.port, reload=False)
+        uvicorn.run(
+            application,
+            host=settings.host,
+            port=settings.port,
+            reload=False,
+            access_log=False,
+        )
 
 
 if __name__ == "__main__":

memhub/apps/core/src/one_memory/context_access.py

@@ -20,6 +20,9 @@ from one_memory.memory_server.store_binding import store_for
 REVOKED_SOURCES_SQL = (
     "SELECT source_id FROM source_grants WHERE user_id='local' AND revoked_at IS NOT NULL"
 )
+ACTIVE_SOURCES_SQL = (
+    "SELECT source_id FROM source_grants WHERE user_id='local' AND revoked_at IS NULL"
+)
 
 
 @dataclass(frozen=True)
@@ -52,13 +55,16 @@ def resolve_context_authority(
     declared_sources = frozenset(
         scope.removeprefix("source:") for scope in access.scopes if scope.startswith("source:")
     )
+    generic_raw = access.host_id == "generic-local-mcp" and "raw-imported-all" in access.scopes
     raw_allowed = (
         raw_requested
         and "om_conversations" in access.tools
         and "all-apps" in access.scopes
-        and bool(declared_sources)
+        and (generic_raw if access.host_id == "generic-local-mcp" else bool(declared_sources))
     )
-    if explicit is not None and (not raw_allowed or not explicit <= declared_sources):
+    if explicit is not None and (
+        not raw_allowed or (not generic_raw and not explicit <= declared_sources)
+    ):
         raise HTTPException(403, "external_context_scope_denied")
 
     raw_sources: frozenset[str] = frozenset()
@@ -70,10 +76,13 @@ def resolve_context_authority(
             # the same code runs on both engines without a branch per engine.
             if hasattr(db, "set_progress_handler"):
                 db.set_progress_handler(lambda: int(time.monotonic() >= deadline), 1000)
-            revoked = frozenset(str(row[0]) for row in db.execute(REVOKED_SOURCES_SQL))
+            sources = frozenset(
+                str(row[0])
+                for row in db.execute(ACTIVE_SOURCES_SQL if generic_raw else REVOKED_SOURCES_SQL)
+            )
         if time.monotonic() >= deadline:
             raise TimeoutError
-        raw_sources = declared_sources - revoked
+        raw_sources = sources if generic_raw else declared_sources - sources
         if explicit is not None:
             if not explicit <= raw_sources:
                 raise HTTPException(403, "external_context_scope_denied")

memhub/apps/core/src/one_memory/context_raw.py

@@ -196,7 +196,7 @@ def read_raw_context(
         if _prefers_declared_facts(request.query)
         else "1"
     )
-    person_authored = "CASE WHEN e.written_by='person' THEN 1 ELSE 0 END"
+    person_authored = "CASE WHEN e.role='user' THEN 1 ELSE 0 END"
     query_echo = (
         "CASE WHEN trim(e.verbatim)=? OR trim(e.verbatim)=? THEN 0 ELSE 1 END"
         if _prefers_declared_facts(request.query)
@@ -433,12 +433,18 @@ def revalidate_raw_context(
     *,
     source_ids: frozenset[str] | None,
     deadline: float,
+    require_active_grant: bool = False,
 ) -> tuple[RawContextItem, ...]:
     if not items or source_ids == frozenset():
         return ()
     if len(items) > 21:
         raise ValueError("raw_context_candidate_limit")
     readable, authority = _readable(source_ids)
+    if require_active_grant:
+        readable += """ AND EXISTS (
+            SELECT 1 FROM source_grants g
+            WHERE g.user_id=e.user_id AND g.source_id=e.source_id AND g.revoked_at IS NULL
+        )"""
     retained: list[RawContextItem] = []
     with _connection(database_path, deadline) as db:
         for item in items:

memhub/apps/core/src/one_memory/context_resume.py

@@ -66,12 +66,12 @@ def _field(text: str) -> ResumeField | None:
     return None
 
 
-def _raw_class(written_by: str, evidence_id: str, accepted: frozenset[str]) -> EvidenceClass:
-    if written_by == "person":
+def _raw_class(role: str, evidence_id: str, accepted: frozenset[str]) -> EvidenceClass:
+    if role == "user":
         return "person_statement"
-    if written_by == "agent":
+    if role == "assistant":
         return "agent_report"
-    if written_by == "tool" and evidence_id in accepted:
+    if role == "tool" and evidence_id in accepted:
         return "system_acceptance"
     return "tool_output"
 
@@ -118,7 +118,7 @@ def assemble_resume_packet(
                     field=_field(message.text) or "unclassified",
                     text=message.text,
                     evidence_class=_raw_class(
-                        message.written_by,
+                        message.role,
                         message.evidence_id,
                         verified_system_evidence_ids,
                     ),

memhub/apps/core/src/one_memory/context_routes.py

@@ -308,6 +308,9 @@ def context_router(dependencies: Sequence[Depends]) -> APIRouter:
                     tuple(item for item in items if isinstance(item, RawContextItem)),
                     source_ids=authority.raw_sources,
                     deadline=end,
+                    require_active_grant=(
+                        access is not None and access.host_id == "generic-local-mcp"
+                    ),
                 )
             return await in_interactive_read(
                 revalidate_memory_context,

memhub/apps/core/src/one_memory/context_v2_coordinator.py

@@ -345,7 +345,7 @@ def _lexical_raw_rows(
         if context_raw._prefers_declared_facts(query.query)
         else "1"
     )
-    person_authored = "CASE WHEN e.written_by='person' THEN 1 ELSE 0 END"
+    person_authored = "CASE WHEN e.role='user' THEN 1 ELSE 0 END"
     query_echo = (
         "CASE WHEN trim(e.verbatim)=? OR trim(e.verbatim)=? THEN 0 ELSE 1 END"
         if context_raw._prefers_declared_facts(query.query)
@@ -507,7 +507,7 @@ def _raw_item(row: sqlite3.Row, start1: int, matched_by: tuple[str, ...]) -> Raw
             "source_id": row["source_id"],
             "conversation_id": row["conversation_id"],
             "role": row["role"],
-            "written_by": "person" if row["written_by"] == "person" else "agent",
+            "written_by": "person" if row["role"] == "user" else "agent",
             "event_time": event_time.isoformat() if event_time else None,
             "spans": [{"start_cp": start, "end_cp": start + len(text), "text": text}],
             "matched_by": list(matched_by),

memhub/apps/core/src/one_memory/database_access.py

@@ -1039,7 +1039,7 @@ def close_main_databases() -> None:
 def log_main_database_stats(moment: str) -> None:
     """One line in core.log: how many connections are open and who waited."""
     stats = main_database_stats()
-    logger.info(
+    logger.debug(
         "main_database_connections moment=%s %s",
         moment,
         " ".join(f"{key}={value}" for key, value in stats.items()),

memhub/apps/core/src/one_memory/diagnostic_log.py

@@ -19,6 +19,7 @@ from __future__ import annotations
 
 import atexit
 import json
+import logging
 import os
 import re
 import sys
@@ -139,6 +140,10 @@ _DETAIL_KEYS = frozenset(
         "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",
     }
 )
 _trace_id: ContextVar[str | None] = ContextVar("one_memory_trace_id", default=None)
@@ -649,12 +654,24 @@ 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.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.
+        return
 
 
 def current_session_id() -> str | None:

memhub/apps/core/src/one_memory/external_memory_access.py

@@ -153,6 +153,7 @@ def external_access(headers: Mapping[str, str]) -> ExternalMemoryAccess | None:
             not isinstance(item, str)
             or (
                 item != "all-apps"
+                and not (item == "raw-imported-all" and host_id == "generic-local-mcp")
                 and not re.fullmatch(
                     r"workspace:v1:[0-9a-f]{64}|source:[a-z0-9][a-z0-9-]{0,127}", item
                 )
@@ -174,7 +175,13 @@ def require_external_endpoint(request: Request) -> None:
     access = external_access(request.headers)
     if access is None:
         return
-    if (request.method, request.url.path) == ("POST", "/api/v1/context/search"):
+    route = (request.method, request.url.path)
+    if access.host_id == "generic-local-mcp" and route not in {
+        ("POST", "/api/v1/recall"),
+        ("POST", "/api/v1/context/search"),
+    }:
+        raise HTTPException(403, "external_endpoint_denied")
+    if route == ("POST", "/api/v1/context/search"):
         # 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"}):

memhub/apps/core/src/one_memory/hy_memory_bridge.py

@@ -33,6 +33,7 @@ from one_memory.hy_runtime_config import (
 from one_memory.hy_source_context import HySourceContext, HyWriteBindings
 from one_memory.hy_store_concurrency import native_store_read_slot
 from one_memory.model_profiles import ModelProfileSnapshot, ProviderKind, attach_disable_thinking
+from one_memory.model_request_audit import audit_precipitation_request
 from one_memory.semantic_deadline import SemanticMemoryHit
 
 #: D102: how relevant HY has to call a hit before it counts. Raised from 0.25
@@ -143,6 +144,22 @@ class HyMemoryBridgeError(RuntimeError):
     pass
 
 
+MEMHUB_DEFAULT_CODING_TOOL_RESULT_MAX_BYTES = 512
+
+
+def coding_tool_result_max_bytes() -> int:
+    """Return the per-tool payload ceiling used by MemHub's HY client."""
+
+    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
+
+
 def _exclusive[**P, T](
     method: Callable[P, Coroutine[Any, Any, T]],
 ) -> Callable[P, Coroutine[Any, Any, T]]:
@@ -414,8 +431,9 @@ class HyMemoryBridge:
 
     @_exclusive
     async def ingest(self, request: HyMemoryIngestRequest) -> HyMemoryIngestResult:
-        async with self._write_lock:
-            return await self._ingest(request)
+        with audit_precipitation_request(request.request_id):
+            async with self._write_lock:
+                return await self._ingest(request)
 
     async def _ingest(self, request: HyMemoryIngestRequest) -> HyMemoryIngestResult:
         started = monotonic()
@@ -431,9 +449,7 @@ class HyMemoryBridge:
         guards = getattr(client, "_one_memory_write_guards", ())
         from one_memory.hy_response_guard import HyReconcileGuard
 
-        reconcile_guards = tuple(
-            guard for guard in guards if isinstance(guard, HyReconcileGuard)
-        )
+        reconcile_guards = tuple(guard for guard in guards if isinstance(guard, HyReconcileGuard))
         for guard in reconcile_guards:
             guard.begin_write()
         before = [len(guard.failures) for guard in guards]
@@ -544,8 +560,9 @@ class HyMemoryBridge:
         request: HyMemoryIngestRequest,
         contents: Sequence[str],
     ) -> HyMemoryIngestResult:
-        async with self._write_lock:
-            return await self._ingest_extracted(request, contents)
+        with audit_precipitation_request(request.request_id):
+            async with self._write_lock:
+                return await self._ingest_extracted(request, contents)
 
     async def _ingest_extracted(
         self,
@@ -1110,7 +1127,7 @@ class HyMemoryBridge:
                                 self.profile,
                                 api_key=self.api_key,
                                 data_dir=self.data_dir,
-                        http_transport_factory=self._http_transport_factory,
+                                http_transport_factory=self._http_transport_factory,
                             )
 
                     await _joined_thread(initialize)
@@ -1137,7 +1154,7 @@ class HyMemoryBridge:
                     )
                     raise HyMemoryBridgeError("hy_memory_init_failed") from error
                 diagnostic_event(
-                    level="info",
+                    level="debug",
                     component="hy.client",
                     event="hy.client_initialization.completed",
                     outcome="succeeded",
@@ -1328,9 +1345,7 @@ async def _with_native_mutation_delta(
     """
 
     retired = set(result.retired_memory_ids)
-    superseded = {
-        key: tuple(values) for key, values in result.superseded_memory_ids.items()
-    }
+    superseded = {key: tuple(values) for key, values in result.superseded_memory_ids.items()}
     chat_targets = {
         key
         for guard in reconcile_guards
@@ -1338,7 +1353,8 @@ async def _with_native_mutation_delta(
         if isinstance(key, str) and key
     }
 
-    for item in raw.get("ops", ()) if isinstance(raw.get("ops"), list) else ():
+    operations = raw.get("ops")
+    for item in operations if isinstance(operations, list) else ():
         if not isinstance(item, Mapping):
             continue
         if item.get("action") == "DELETE":
@@ -1346,7 +1362,8 @@ async def _with_native_mutation_delta(
             if isinstance(key, str) and key:
                 retired.add(key)
 
-    for item in raw.get("fact_ops", ()) if isinstance(raw.get("fact_ops"), list) else ():
+    fact_operations = raw.get("fact_ops")
+    for item in fact_operations if isinstance(fact_operations, list) else ():
         if not isinstance(item, Mapping) or item.get("action") != "SUPERSEDE":
             continue
         old, new = item.get("fact_id"), item.get("superseded_by")
@@ -1367,9 +1384,7 @@ async def _with_native_mutation_delta(
         if isinstance(replacement, str) and replacement:
             superseded[key] = (replacement,)
         elif isinstance(replacement, list):
-            values = tuple(
-                value for value in replacement if isinstance(value, str) and value
-            )
+            values = tuple(value for value in replacement if isinstance(value, str) and value)
             if values:
                 superseded[key] = values
 
@@ -1697,6 +1712,7 @@ def build_hy_memory_config(
     config.history.db_path = str(memory_root / "history.sqlite3")
     config.coding.enable = True
     config.coding.db_path = str(memory_root / "coding.sqlite3")
+    config.coding.tool_result_max_bytes = coding_tool_result_max_bytes()
     config.recall.episode_recall_enabled = True
     config.recall.episode_summary_enabled = False
     config.recall.dynamic_episode_enabled = True

memhub/apps/core/src/one_memory/hy_runtime_config.py

@@ -14,7 +14,7 @@ from pathlib import Path
 from typing import Literal, cast
 
 HY_MEMORY_DISTRIBUTION = "hy-memory"
-HY_MEMORY_VERSION = "2.0.2"
+HY_MEMORY_VERSION = "2.0.4"
 HY_MEMORY_NAMESPACE = "governed-v2"
 DiagnosticLogLevel = Literal["error", "warning", "info", "debug"]
 
@@ -31,11 +31,12 @@ _COLLECTION_LOGGERS = (
     "httpx",
     "httpcore",
 )
+_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 public SDK 2.0.2.
+# compatibility with public SDK 2.0.4.
 HY_MEMORY_MODE = "ultra"
 # The SDK only recognises `legacy` and `hybrid` (see
 # hy_memory/pipelines/_retrieval/config.py: `ALL_READERS = (READER_LEGACY,
@@ -139,6 +140,18 @@ def apply_diagnostic_log_level(level: DiagnosticLogLevel) -> None:
     for name in tuple(logging.Logger.manager.loggerDict):
         if any(name == root or name.startswith(f"{root}.") for root in _COLLECTION_LOGGERS):
             logging.getLogger(name).setLevel(selected)
+    # Successful HTTP exchanges, file-watch notifications and tokenizer
+    # initialization are useful only while actively debugging. At INFO they
+    # dominated core.log and the in-app panel without describing product work.
+    noisy_level = selected if selected == logging.DEBUG else max(selected, logging.WARNING)
+    for root in _NOISY_LIBRARY_LOGGERS:
+        logging.getLogger(root).setLevel(noisy_level)
+    for name in tuple(logging.Logger.manager.loggerDict):
+        if any(
+            name == root or name.startswith(f"{root}.")
+            for root in _NOISY_LIBRARY_LOGGERS
+        ):
+            logging.getLogger(name).setLevel(noisy_level)
 
     os.environ["MEMORY_LOG_LEVEL"] = logging.getLevelName(selected)
     os.environ["MEMORY_PIPELINE_TRACE_ENABLED"] = "true" if level == "debug" else "false"

memhub/apps/core/src/one_memory/hy_source_context.py

@@ -7,7 +7,7 @@ from typing import Any, Literal
 
 from pydantic import BaseModel, ConfigDict
 
-from one_memory.imports.models import EvidenceContentKind, EvidenceRecord, WrittenBy
+from one_memory.imports.models import EvidenceContentKind, EvidenceRecord
 from one_memory.input_sanitization import sanitize_model_input
 
 
@@ -61,17 +61,7 @@ class HyWriteBindings:
                     seed.workspace,
                 )
                 or record.content_kind is not EvidenceContentKind.RAW_CONVERSATION
-                or (record.role, record.written_by)
-                not in {
-                    # D98: the same change as the turn assembly. A sub agent's
-                    # question is a `user` message marked written by the
-                    # agent; excluding it here would make the source context
-                    # disagree with what was actually sent.
-                    ("user", WrittenBy.PERSON),
-                    ("user", WrittenBy.AGENT),
-                    ("assistant", WrittenBy.AGENT),
-                    ("tool", WrittenBy.TOOL),
-                }
+                or record.role not in {"user", "assistant", "tool"}
             ):
                 raise ValueError("hy_memory_write_verification_context_scope")
         self.all_sources: list[dict[str, Any]] = [

memhub/apps/core/src/one_memory/hy_step_timing.py

@@ -152,7 +152,7 @@ class HyStepTimer:
             if value is None:
                 continue
             diagnostic_event(
-                level="info",
+                level="debug",
                 component=STEP_COMPONENTS[name],
                 event="hy.step.completed",
                 outcome="succeeded",
@@ -168,7 +168,7 @@ class HyStepTimer:
             )
         model = measured.get("model")
         diagnostic_event(
-            level="info",
+            level="debug",
             component="hy.bridge",
             event="hy.steps.summary",
             outcome="succeeded",

memhub/apps/core/src/one_memory/imports/async_bridge.py

@@ -105,7 +105,7 @@ async def iterate_in_worker[T](
         # was unwinding with, and a diagnostic is not worth an import.
         try:
             diagnostic_event(
-                level="info",
+                level="debug",
                 component="import.async_bridge",
                 event="import.source_read_finished",
                 outcome="succeeded",

memhub/apps/core/src/one_memory/imports/grants.py

@@ -1,6 +1,7 @@
 import sqlite3
 from contextlib import AbstractContextManager
 from datetime import UTC, datetime
+from enum import StrEnum
 from pathlib import Path
 
 from pydantic import BaseModel, ConfigDict
@@ -20,6 +21,23 @@ CURSOR_MAX_SOURCE_FILE_BYTES = 64 * 1024 * 1024 * 1024
 GRANT_SCHEMA = source_grants_ddl("sqlite")
 
 
+class SourceAuthorizationState(StrEnum):
+    """Why a source cannot be read right now, or `ok` when it can.
+
+    The API used to answer `authorized=false` for all of these at once, which
+    is how a revoked grant, a rebuilt history directory and a moved candidate
+    all looked like the same bug and all got the same advice. The values carry
+    no paths and no user content, because they are written to diagnostics.
+    """
+
+    OK = "ok"
+    MISSING_GRANT = "missing_grant"
+    REVOKED_GRANT = "revoked_grant"
+    SOURCE_IDENTITY_CHANGED = "source_identity_changed"
+    SOURCE_ROOT_UNAVAILABLE = "source_root_unavailable"
+    CANDIDATE_CHANGED = "candidate_changed"
+
+
 class SourceGrantReceipt(BaseModel):
     model_config = ConfigDict(frozen=True, extra="forbid")
 
@@ -146,6 +164,46 @@ class SourceGrantStore:
             raise SourceAccessError(SourceAccessErrorCode.SOURCE_IDENTITY_CHANGED)
         return authorized
 
+    def authorization_state(
+        self,
+        *,
+        user_id: str,
+        source_id: str,
+        candidate: Path,
+    ) -> SourceAuthorizationState:
+        """Why reading this source would be refused.
+
+        Read-only and fails the same way `authorized_root` does: this names a
+        state for the UI and for diagnostics, it does not widen access. The
+        candidate is compared by path, so a grant that still points somewhere
+        else reads as a moved candidate rather than as a rebuilt directory —
+        the two need different fixes.
+        """
+        with self._connect() as connection:
+            row = connection.execute(
+                "SELECT * FROM source_grants WHERE user_id=? AND source_id=?",
+                (user_id, source_id),
+            ).fetchone()
+        if row is None:
+            return SourceAuthorizationState.MISSING_GRANT
+        if row["revoked_at"] is not None or row["revision"] <= 0:
+            return SourceAuthorizationState.REVOKED_GRANT
+        granted = Path(row["requested_path"])
+        if not granted.exists():
+            return SourceAuthorizationState.SOURCE_ROOT_UNAVAILABLE
+        try:
+            if not root_identity_matches(granted, row["root_identity"]):
+                return SourceAuthorizationState.SOURCE_IDENTITY_CHANGED
+        except SourceAccessError:
+            # Unreadable or gone between the two calls: there is no grant to
+            # repair, only a directory to make available again.
+            return SourceAuthorizationState.SOURCE_ROOT_UNAVAILABLE
+        # Same directory, but the app now looks somewhere else. Re-authorizing
+        # the old grant would not help; the candidate has to be granted.
+        if granted != candidate:
+            return SourceAuthorizationState.CANDIDATE_CHANGED
+        return SourceAuthorizationState.OK
+
     def revoke(self, *, user_id: str, source_id: str) -> SourceGrantReceipt:
         now = datetime.now(UTC).isoformat()
         with self._connect(writing=True) as connection:

memhub/apps/core/src/one_memory/imports/live_sync.py

@@ -106,7 +106,7 @@ class PollingLiveSyncSupervisor:
             level=(
                 "warn"
                 if value.state in {LiveSyncState.DEGRADED, LiveSyncState.UNAVAILABLE}
-                else "info"
+                else "debug"
             ),
             component="import.live_sync",
             event="live_sync.state_changed",
@@ -282,11 +282,24 @@ class PollingLiveSyncSupervisor:
                 and job.status
                 in {ImportStatus.PENDING, ImportStatus.RUNNING, ImportStatus.INTERRUPTED}
             ]
+            bounded_history_incomplete = any(
+                job.source_id == source_id
+                and job.operation == "history"
+                and job.history_after is not None
+                and job.status is not ImportStatus.COMPLETED
+                for job in source_jobs
+            )
             # Establish the realtime EOF independently even while history owns
             # this source. Otherwise the P1 job starts from the P2 checkpoint
             # (or from zero) and becomes a second full-history scan.
             adapter = self.import_supervisor.adapters[source_id]
-            await self._live_cursors(source_id, adapter)
+            live_cursors = await self._live_cursors(source_id, adapter)
+            # A bounded onboarding history is the authority for everything up
+            # to its EOF. An adapter without a native live baseline has no safe
+            # realtime cursor until that history finishes; starting a sync from
+            # zero here would bypass `history_after` and import the old corpus.
+            if bounded_history_incomplete and not live_cursors:
+                return None
             # Only one scan of a source may be in flight: they share the
             # cursor, and two readers means every file read twice. With a
             # catch-up still holding the source there is no free moment to ask

memhub/apps/core/src/one_memory/imports/live_watch.py

@@ -156,7 +156,7 @@ class FileWatchLiveSync:
             if source is not None
         ):
             diagnostic_event(
-                level="info",
+                level="debug",
                 component="import.live_watch",
                 event="live_sync.hook_received",
                 outcome="queued",
@@ -190,7 +190,7 @@ class FileWatchLiveSync:
                 try:
                     await self._trigger_past(source_id, observed)
                     diagnostic_event(
-                        level="info",
+                        level="debug",
                         component="import.live_watch",
                         event="live_sync.hook_completed",
                         outcome="written",

memhub/apps/core/src/one_memory/imports/models.py

@@ -79,6 +79,8 @@ class EvidenceInput(BaseModel):
     source_ordinal: str | None = Field(default=None, min_length=1, max_length=256)
     content_kind: EvidenceContentKind = EvidenceContentKind.RAW_CONVERSATION
     role: str = Field(min_length=1, max_length=128)
+    # Legacy import/storage metadata only. Protocol `role` is authoritative;
+    # this value must not affect identity, admission, scheduling or ownership.
     written_by: WrittenBy
     created_at: datetime | None = None
     verbatim: str = Field(min_length=1, max_length=MAX_VERBATIM_CHARACTERS)
@@ -149,17 +151,22 @@ class EvidenceInput(BaseModel):
                 )
             )
         else:
+            # `role` is the protocol-level speaker identity. `written_by` was a
+            # parser guess and differed between two imports of the same host
+            # message (notably sub-agent/user and unknown/user), minting two
+            # Evidence rows. Keep it out of every new identity.
+            #
             # No workspace and no source_ordinal in the derived key: the hook
             # lane has neither (it reports no workspace and a constant ordinal),
             # so including either gave the same message a different key per lane
             # and stored it twice. Both are record metadata here, not identity.
             identity = "\0".join(
                 (
-                    "derived-v3",
+                    "derived-v4",
                     self.user_id,
                     self.source_id,
                     self.conversation_id,
-                    self.written_by,
+                    self.role,
                     self.content_hash,
                 )
             )
@@ -168,6 +175,54 @@ class EvidenceInput(BaseModel):
             # two rows. The capture time is record metadata, not identity.
         return hashlib.sha256(identity.encode()).hexdigest()
 
+    @property
+    def written_by_derived_v3_idempotency_key(self) -> str:
+        """The last identity that treated the parser's author guess as canonical.
+
+        New builds key by protocol role. During an upgrade this key lets a
+        rescan find and reuse the row already stored by derived-v3 instead of
+        inserting the same host message under a second Evidence id.
+        """
+        identity = "\0".join(
+            (
+                "derived-v3",
+                self.user_id,
+                self.source_id,
+                self.conversation_id,
+                self.written_by,
+                self.content_hash,
+            )
+        )
+        return hashlib.sha256(identity.encode()).hexdigest()
+
+    @property
+    def written_by_derived_v3_idempotency_keys(self) -> tuple[str, ...]:
+        """Every legacy author guess this protocol role may have carried."""
+        guesses = {
+            "user": (
+                WrittenBy.PERSON,
+                WrittenBy.AGENT,
+                WrittenBy.UNKNOWN,
+            ),
+            "assistant": (WrittenBy.AGENT, WrittenBy.UNKNOWN, WrittenBy.PERSON),
+            "tool": (WrittenBy.TOOL, WrittenBy.UNKNOWN),
+        }.get(self.role, tuple(WrittenBy))
+        ordered = (self.written_by, *guesses)
+        keys: list[str] = []
+        for written_by in dict.fromkeys(ordered):
+            identity = "\0".join(
+                (
+                    "derived-v3",
+                    self.user_id,
+                    self.source_id,
+                    self.conversation_id,
+                    written_by,
+                    self.content_hash,
+                )
+            )
+            keys.append(hashlib.sha256(identity.encode()).hexdigest())
+        return tuple(keys)
+
     @property
     def legacy_message_idempotency_key(self) -> str | None:
         """The pre-workspace key for a message-shaped input, when one exists.
@@ -212,7 +267,13 @@ class EvidenceInput(BaseModel):
         """
         return self.derived_v2_idempotency_key("derived-v2", workspace_agnostic=False)
 
-    def derived_v2_idempotency_key(self, label: str, *, workspace_agnostic: bool) -> str:
+    def derived_v2_idempotency_key(
+        self,
+        label: str,
+        *,
+        workspace_agnostic: bool,
+        written_by: WrittenBy | None = None,
+    ) -> str:
         identity = "\0".join(
             (
                 label,
@@ -220,7 +281,7 @@ class EvidenceInput(BaseModel):
                 self.source_id,
                 *([] if workspace_agnostic else [self.workspace or "\uffffnull"]),
                 self.conversation_id,
-                self.written_by,
+                written_by or self.written_by,
                 self.created_at.isoformat() if self.created_at else "",
                 self.source_ordinal or "",
                 # v2 hashed the cleaned body; v1 predates raw_verbatim entirely,
@@ -230,6 +291,26 @@ class EvidenceInput(BaseModel):
         )
         return hashlib.sha256(identity.encode()).hexdigest()
 
+    @property
+    def older_written_by_idempotency_keys(self) -> tuple[str, ...]:
+        keys: list[str] = []
+        for written_by in WrittenBy:
+            keys.append(
+                self.derived_v2_idempotency_key(
+                    "derived-v1",
+                    workspace_agnostic=True,
+                    written_by=written_by,
+                )
+            )
+            keys.append(
+                self.derived_v2_idempotency_key(
+                    "derived-v2",
+                    workspace_agnostic=False,
+                    written_by=written_by,
+                )
+            )
+        return tuple(keys)
+
     @property
     def evidence_id(self) -> str:
         return str(uuid5(EVIDENCE_NAMESPACE, self.idempotency_key))
@@ -249,6 +330,7 @@ class EvidenceRecord(BaseModel):
     source_ordinal: str | None
     content_kind: EvidenceContentKind
     role: str
+    # Retained while old databases and export formats still carry the column.
     written_by: WrittenBy
     created_at: datetime | None
     verbatim: str

memhub/apps/core/src/one_memory/imports/runner.py

@@ -199,7 +199,7 @@ class ImportRunner:
             raise ValueError("Import job source does not match adapter source")
         job, runner_token = await in_the_background(self.job_store.claim, job_id)
         diagnostic_event(
-            level="info",
+            level="debug",
             component="import.parser",
             event="import.parse.started",
             outcome="started",
@@ -310,7 +310,7 @@ class ImportRunner:
                 if event.kind == "progress" and str(event.phase) != last_phase:
                     phase = str(event.phase)
                     diagnostic_event(
-                        level="info",
+                        level="debug",
                         component="import.parser",
                         event="import.phase_changed",
                         outcome="changed",
@@ -516,7 +516,7 @@ class ImportRunner:
     async def _emit_parse_result(self, job: ImportJob, started: float) -> None:
         succeeded = job.status is ImportStatus.COMPLETED
         diagnostic_event(
-            level="info" if succeeded else "warn",
+            level="debug" if succeeded else "warn",
             component="import.parser",
             event="import.parse.completed",
             outcome="succeeded" if succeeded else str(job.status),

memhub/apps/core/src/one_memory/imports/sources/codex.py

@@ -141,11 +141,18 @@ def parse_codex_record(
     if role == "user":
         written_by = _user_author(record, payload, session)
         # Machine inputs may quote a whole transcript, including user_query tags.
-        # Session provenance is not proof of who wrote any particular future turn.
-        if written_by is WrittenBy.PERSON:
+        # Protocol role is authoritative; host author metadata does not change
+        # how a user message is cleaned.
+        if not (
+            content.startswith(
+                "The following is the Codex agent history whose request action you are assessing."
+            )
+            and ">>> TRANSCRIPT START" in content
+            and ">>> TRANSCRIPT END" in content
+        ):
             content = strip_host_wrappers(content)
-            if not content:
-                return ParsedSkip(ScanIssueCode.NO_TEXT)
+        if not content:
+            return ParsedSkip(ScanIssueCode.NO_TEXT)
     return ParsedMessage(
         conversation_id=conversation_from_relative_path(context.relative_path),
         message_id=first_string(payload.get("id"), record.get("id")),

memhub/apps/core/src/one_memory/imports/sources/common.py

@@ -355,7 +355,7 @@ class JsonlSourceAdapter:
                 source_file, checkpoint, header = await asyncio.to_thread(
                     self._prepare, listed, options
                 )
-                _log.info(
+                _log.debug(
                     "import_file_read source=%s file=%d/%d file_bytes=%d planned_bytes=%d",
                     self.descriptor.source_id,
                     index,
@@ -376,7 +376,7 @@ class JsonlSourceAdapter:
                     else DEFAULT_IMPORT_MAX_FILE_BYTES
                 )
                 if source_file.size > budget:
-                    _log.info(
+                    _log.debug(
                         "import_file_budgeted source=%s file=%d/%d file_bytes=%d budget=%d path=%s",
                         self.descriptor.source_id,
                         index,

memhub/apps/core/src/one_memory/imports/sources/cursor_unified.py

@@ -7,7 +7,7 @@ import sqlite3
 import stat
 from collections.abc import AsyncIterator, Iterator
 from dataclasses import dataclass, replace
-from datetime import UTC, datetime, timedelta
+from datetime import UTC, datetime, timedelta, timezone
 from functools import partial
 
 from pydantic import JsonValue
@@ -82,6 +82,11 @@ _CURSOR_HOST_CONTEXT_START = re.compile(
 )
 _CURSOR_USER_QUERY = re.compile(r"<user_query\b")
 _CURSOR_TIMESTAMP_PREFIX = re.compile(r"<timestamp>[\s\S]*?</timestamp>\s*")
+_CURSOR_TIMESTAMP_VALUE = re.compile(
+    r"<timestamp>\s*"
+    r"([A-Za-z]+,\s+[A-Za-z]{3}\s+\d{1,2},\s+\d{4},\s+\d{1,2}:\d{2}\s+[AP]M)"
+    r"\s+\(UTC([+-]\d{1,2})(?::(\d{2}))?\)\s*</timestamp>",
+)
 _CURSOR_IMAGE_FILES_PREFIX = re.compile(r"\[Image\]\s*<image_files>[\s\S]*?</image_files>\s*")
 _CURSOR_BARE_IMAGE_PREFIX = re.compile(r"(?:\[Image\]\s*)+")
 
@@ -123,6 +128,24 @@ def _strip_cursor_timestamp(content: str | None) -> str | None:
     return _CURSOR_TIMESTAMP_PREFIX.sub("", content, count=1) if content else None
 
 
+def _cursor_host_timestamp(content: str | None) -> datetime | None:
+    """Parse Cursor's user-visible timestamp envelope into UTC."""
+    if not content:
+        return None
+    match = _CURSOR_TIMESTAMP_VALUE.search(content)
+    if match is None:
+        return None
+    try:
+        local = datetime.strptime(match[1], "%A, %b %d, %Y, %I:%M %p")
+        hours = int(match[2])
+        minutes = int(match[3] or "0")
+        sign = -1 if hours < 0 else 1
+        offset = timedelta(hours=hours, minutes=sign * minutes)
+        return local.replace(tzinfo=timezone(offset)).astimezone(UTC)
+    except (ValueError, OverflowError):
+        return None
+
+
 def _strip_cursor_host_prefix(content: str | None) -> str | None:
     if not content:
         return content
@@ -398,10 +421,7 @@ class CursorUnifiedSourceAdapter:
             if (
                 isinstance(event, ScannedEvidence)
                 and event.evidence.conversation_id in transcript_ids
-                and not (
-                    event.evidence.written_by is WrittenBy.AGENT
-                    and is_host_generated_user_turn(event.evidence.verbatim)
-                )
+                and not is_host_generated_user_turn(event.evidence.verbatim)
             ):
                 yield ScanSkip(
                     cursor=event.cursor,
@@ -426,10 +446,7 @@ def _shadowed_by_transcript(event: ScanEvent, transcript_ids: set[str]) -> ScanE
     if (
         isinstance(event, ScannedEvidence)
         and event.evidence.conversation_id in transcript_ids
-        and not (
-            event.evidence.written_by is WrittenBy.AGENT
-            and is_host_generated_user_turn(event.evidence.verbatim)
-        )
+        and not is_host_generated_user_turn(event.evidence.verbatim)
     ):
         return ScanSkip(
             cursor=event.cursor,
@@ -703,6 +720,7 @@ def _parse_transcript_record(
     project_name = path_parts[path_parts.index("projects") + 1] if "projects" in path_parts else ""
     session_name = context.relative_path.rsplit("/", 1)[-1].removesuffix(".jsonl")
     is_subagent = f"{project_name}\0{session_name}" in subagent_ids
+    host_timestamp = _cursor_host_timestamp(content) if role == "user" else None
     created_at = _transcript_fallback_time(context)
     native_message_id = record.get("id") or message.get("id")
     message_id = (
@@ -728,7 +746,16 @@ def _parse_transcript_record(
             created_at=created_at,
             workspace=workspace,
             record_type="cursor_transcript",
-            origin_metadata=OriginMetadata(raw_meta={"time_basis": "file_mtime"}),
+            origin_metadata=OriginMetadata(
+                raw_meta={
+                    "time_basis": "file_mtime",
+                    **(
+                        {"turn_event_at": host_timestamp.isoformat()}
+                        if host_timestamp is not None
+                        else {}
+                    ),
+                }
+            ),
         ),
         tool_blocks_evidence(message.get("content"), block_types=TOOL_BLOCKS),
     )
@@ -819,6 +846,13 @@ def _read_store_messages(
                         reason_code=ScanIssueCode.UNKNOWN_RECORD,
                     )
                     continue
+                host_timestamp = (
+                    _cursor_host_timestamp(content) if role == "user" else None
+                )
+                fallback_timestamp = datetime.fromtimestamp(
+                    source_file.modified_ns / 1_000_000_000,
+                    tz=UTC,
+                ) + timedelta(microseconds=rowid)
                 parsed = ParsedMessage(
                     conversation_id=conversation_id,
                     message_id=str(record.get("id") or row["id"]),
@@ -829,8 +863,19 @@ def _read_store_messages(
                         else written_by_user
                     ),
                     verbatim=cleaned,
+                    created_at=fallback_timestamp,
                     workspace=_store_workspace(source_file),
                     record_type="cursor_store_blob",
+                    origin_metadata=OriginMetadata(
+                        raw_meta={
+                            "time_basis": "file_mtime",
+                            **(
+                                {"turn_event_at": host_timestamp.isoformat()}
+                                if host_timestamp is not None
+                                else {}
+                            ),
+                        }
+                    ),
                 )
                 yield ScannedEvidence(
                     cursor=cursor,

memhub/apps/core/src/one_memory/imports/storage.py

@@ -57,7 +57,8 @@ _UNANSWERED_PREFIXES = (
 
 
 def _is_visible_user_row(role: str, written_by: str, verbatim: str) -> bool:
-    if role != "user" or written_by != WrittenBy.PERSON.value:
+    del written_by  # Legacy storage metadata; protocol role is authoritative.
+    if role != "user":
         return False
     return not verbatim.lstrip().lower().startswith(_UNANSWERED_PREFIXES)
 
@@ -468,16 +469,54 @@ class EvidenceStore:
             if anchored is not None:
                 self._write_origins(connection, anchored, evidence.origin_meta, now)
                 workspace, project = self._resolve_conversation_attribution(connection, evidence)
-                connection.execute(
-                    """
-                    UPDATE evidence
-                    SET workspace = COALESCE(workspace, ?),
-                        project = COALESCE(project, ?),
-                        updated_at = ?
-                    WHERE evidence_id = ?
-                    """,
-                    (workspace, project, now, anchored),
-                )
+                existing = connection.execute(
+                    "SELECT role,revision FROM evidence WHERE evidence_id=?",
+                    (anchored,),
+                ).fetchone()
+                if existing is None:
+                    raise RuntimeError("anchored evidence disappeared")
+                if str(existing["role"]) != evidence.role:
+                    was_revised = True
+                    revision = int(existing["revision"]) + 1
+                    connection.execute(
+                        """
+                        UPDATE evidence
+                        SET workspace=COALESCE(workspace,?),
+                            project=COALESCE(project,?),
+                            content_kind=?,role=?,created_at=COALESCE(?,created_at),
+                            verbatim=?,content_hash=?,revision=?,updated_at=?
+                        WHERE evidence_id=?
+                        """,
+                        (
+                            workspace,
+                            project,
+                            evidence.content_kind,
+                            evidence.role,
+                            evidence.created_at.isoformat() if evidence.created_at else None,
+                            evidence.verbatim,
+                            evidence.content_hash,
+                            revision,
+                            now,
+                            anchored,
+                        ),
+                    )
+                    self._insert_current_version(
+                        connection,
+                        anchored,
+                        revision=revision,
+                        changed_at=now,
+                    )
+                else:
+                    connection.execute(
+                        """
+                        UPDATE evidence
+                        SET workspace=COALESCE(workspace,?),
+                            project=COALESCE(project,?),
+                            updated_at=?
+                        WHERE evidence_id=?
+                        """,
+                        (workspace, project, now, anchored),
+                    )
                 record = self._get_with_connection(connection, anchored)
                 if record is None:
                     raise RuntimeError("anchored evidence disappeared")
@@ -485,7 +524,11 @@ class EvidenceStore:
                     from one_memory import ingest_journal
 
                     ingest_journal.record(connection, anchored, cause="import")
-                return PutEvidenceResult(record=record, was_created=False, was_revised=False)
+                return PutEvidenceResult(
+                    record=record,
+                    was_created=False,
+                    was_revised=was_revised,
+                )
             insert = connection.execute(
                 """
                 INSERT INTO evidence (
@@ -803,27 +846,20 @@ class EvidenceStore:
                 and same_scope[0]["idempotency_key"] != evidence.idempotency_key
             ):
                 return str(same_scope[0]["evidence_id"])
-        elif (legacy_id := claim_legacy_key(evidence.legacy_derived_idempotency_key)) is not None:
-            return legacy_id
-        if not evidence.message_id and not evidence.correlation_id:
-            # BUG-03 moved the derived key off the capture time and onto the
-            # pre-cleaning text. A replay of a row stored under the old scheme
-            # has to be recognized as the same message.
-            replayed = connection.execute(
-                """
-                SELECT evidence_id FROM evidence
-                WHERE idempotency_key = ? AND user_id = ? AND source_id = ?
-                  AND conversation_id = ?
-                """,
+        elif (
+            legacy_id := next(
                 (
-                    evidence.prior_derived_idempotency_key,
-                    evidence.user_id,
-                    evidence.source_id,
-                    evidence.conversation_id,
+                    anchored
+                    for key in (
+                        *evidence.written_by_derived_v3_idempotency_keys,
+                        *evidence.older_written_by_idempotency_keys,
+                    )
+                    if (anchored := claim_legacy_key(key)) is not None
                 ),
-            ).fetchone()
-            if replayed is not None:
-                return str(replayed["evidence_id"])
+                None,
+            )
+        ) is not None:
+            return legacy_id
         if (
             EvidenceOrigin.HOOK in evidence.origins
             and evidence.message_id is None
@@ -838,7 +874,7 @@ class EvidenceStore:
                   ON hook.evidence_id=e.evidence_id AND hook.origin='hook'
                 WHERE e.user_id=? AND e.source_id=? AND e.conversation_id=?
                   AND (e.workspace IS ? OR e.workspace IS NULL OR ? IS NULL)
-                  AND e.role=? AND e.written_by=? AND e.content_hash=?
+                  AND e.role=? AND e.content_hash=?
                   AND e.message_id IS NOT NULL AND hook.evidence_id IS NULL
                   AND ABS((julianday(e.created_at)-julianday(?))*86400.0) <= 5.0
                 ORDER BY ABS((julianday(e.created_at)-julianday(?))*86400.0), e.evidence_id
@@ -851,7 +887,6 @@ class EvidenceStore:
                     evidence.workspace,
                     evidence.workspace,
                     evidence.role,
-                    evidence.written_by,
                     evidence.content_hash,
                     evidence.created_at.isoformat(),
                     evidence.created_at.isoformat(),
@@ -872,7 +907,7 @@ class EvidenceStore:
               AND (e.workspace IS ? OR e.workspace IS NULL OR ? IS NULL)
               -- Unary plus keeps the coverage index from looking better than
               -- the message index, which narrows to this exact text.
-              AND e.role = ? AND +e.written_by = ? AND e.content_hash = ?
+              AND e.role = ? AND e.content_hash = ?
               AND e.message_id IS NULL
             """,
             (
@@ -882,7 +917,6 @@ class EvidenceStore:
                 evidence.workspace,
                 evidence.workspace,
                 evidence.role,
-                evidence.written_by,
                 evidence.content_hash,
             ),
         ).fetchall()
@@ -1247,7 +1281,7 @@ class EvidenceStore:
             rows = connection.execute(
                 """
                 SELECT evidence_id FROM evidence
-                WHERE user_id = 'local' AND role = 'user' AND written_by = 'person'
+                WHERE user_id = 'local' AND role = 'user'
                 ORDER BY imported_at, evidence_id
                 LIMIT ?
                 """,
@@ -1301,7 +1335,7 @@ class EvidenceStore:
             rows = connection.execute(
                 f"""
                 SELECT evidence_id FROM evidence
-                WHERE user_id = 'local' AND role = 'user' AND written_by = 'person'
+                WHERE user_id = 'local' AND role = 'user'
                   AND ({predicates})
                 ORDER BY imported_at, evidence_id
                 LIMIT ?
@@ -1333,7 +1367,7 @@ class EvidenceStore:
             rows = connection.execute(
                 f"""
                 SELECT evidence_id FROM evidence
-                WHERE user_id = 'local' AND role = 'user' AND written_by = 'person'
+                WHERE user_id = 'local' AND role = 'user'
                   AND workspace IS NOT NULL
                   AND ({predicates})
                 ORDER BY imported_at, evidence_id
@@ -1476,7 +1510,7 @@ class EvidenceStore:
         )
         values: list[object] = [user_id, *(source_ids or ())]
         visible_user = """
-            role='user' AND written_by='person' AND NOT (
+            role='user' AND NOT (
               lower(ltrim(verbatim)) LIKE '<codex_internal_context%' OR
               lower(ltrim(verbatim)) LIKE '<environment_context>%' OR
               lower(ltrim(verbatim)) LIKE '<heartbeat>%' OR
@@ -1485,7 +1519,7 @@ class EvidenceStore:
             )
         """
         selected_visible_user = """
-            e.role='user' AND e.written_by='person' AND NOT (
+            e.role='user' AND NOT (
               lower(ltrim(e.verbatim)) LIKE '<codex_internal_context%' OR
               lower(ltrim(e.verbatim)) LIKE '<environment_context>%' OR
               lower(ltrim(e.verbatim)) LIKE '<heartbeat>%' OR
@@ -1638,7 +1672,7 @@ class EvidenceStore:
         if seed is None:
             return ConversationTurnPage((), {}, 0, offset, limit)
         visible_user = """
-          role='user' AND written_by='person' AND NOT (
+          role='user' AND NOT (
             lower(ltrim(verbatim)) LIKE '<codex_internal_context%' OR
             lower(ltrim(verbatim)) LIKE '<environment_context>%' OR
             lower(ltrim(verbatim)) LIKE '<heartbeat>%' OR
@@ -1731,7 +1765,7 @@ class EvidenceStore:
                       AND first.conversation_id = evidence.conversation_id
                       AND first.workspace IS evidence.workspace
                       AND (
-                        (first.role = 'user' AND first.written_by = 'person')
+                        first.role = 'user'
                         OR first.content_kind = 'memory_content'
                       )
                     ORDER BY julianday(COALESCE(first.created_at, first.imported_at)),
@@ -2125,8 +2159,6 @@ class EvidenceStore:
             (
                 existing["content_kind"] != evidence.content_kind,
                 existing["role"] != evidence.role,
-                evidence.written_by is not WrittenBy.UNKNOWN
-                and existing["written_by"] != evidence.written_by,
                 existing["content_hash"] != evidence.content_hash,
             )
         )

memhub/apps/core/src/one_memory/imports/supervisor.py

@@ -111,7 +111,7 @@ class ImportSupervisor:
                 )
             )
             diagnostic_event(
-                level="info",
+                level="debug",
                 component="import.supervisor",
                 event="import.queued",
                 outcome="queued",
@@ -198,7 +198,7 @@ class ImportSupervisor:
         started = monotonic()
         failure_type: str | None = None
         diagnostic_event(
-            level="info",
+            level="debug",
             component="import.supervisor",
             event="import.started",
             outcome="started",

memhub/apps/core/src/one_memory/ingest_journal.py

@@ -85,15 +85,14 @@ def _state(connection: sqlite3.Connection, evidence_id: str) -> dict[str, str |
     value = evidence.model_dump(mode="json")
     # The stored normalized content_hash is insufficient for exact-content leases.
     content = _digest(value["verbatim"])
-    material = _digest(
-        {key: value[key] for key in ("verbatim", "role", "written_by", "content_kind")}
-    )
+    material = _digest({key: value[key] for key in ("verbatim", "role", "content_kind")})
     origins = _digest(value["origin_meta"])
     identity = _digest(
         {
             key: item
             for key, item in value.items()
-            if key not in {"verbatim", "origin_meta", "imported_at", "updated_at"}
+            if key
+            not in {"verbatim", "written_by", "origin_meta", "imported_at", "updated_at"}
         }
     )
     return dict(

memhub/apps/core/src/one_memory/live_capture.py

@@ -3,6 +3,7 @@
 import json
 import time
 from datetime import UTC, datetime
+from enum import StrEnum
 from pathlib import Path
 from typing import TextIO, cast
 
@@ -104,7 +105,13 @@ def capture_hook(source_id: str, payload: object, database: Path) -> bool:
     # Cursor also runs ~/.claude/settings.json hooks; its own hook already files the turn.
     if source_id != "cursor" and "cursor_version" in payload:
         return False
-    if not database.is_file() or not _sync_open(database, source_id):
+    if not database.is_file():
+        return False
+    rejection = capture_rejection(database, source_id)
+    if rejection is not None:
+        # Silent, and the reason is counted rather than printed: a hook that
+        # spoke would put this into the host's turn as text.
+        record_capture_rejection(database, source_id, rejection)
         return False
     if source_id == "pi" and _is_automated_session(first_string(payload.get("cwd"))):
         return False
@@ -139,17 +146,79 @@ def capture_hook(source_id: str, payload: object, database: Path) -> bool:
 
 
 def _sync_open(database: Path, source_id: str) -> bool:
+    return capture_rejection(database, source_id) is None
+
+
+class CaptureRejection(StrEnum):
+    """Why a hook's turn was dropped, or nothing at all.
+
+    Capture fails open on purpose: a hook that raises or prints would be read
+    back into the host's turn, so a refusal has to be silent and exit 0. It was
+    also completely invisible, which is why "持续同步开着但一条都没进" could not be
+    told apart from "capture 坏了". These values name the reason without any
+    payload, path, session or message — the only part that may be recorded.
+    """
+
+    SYNC_DISABLED = "sync_disabled"
+    GRANT_MISSING = "grant_missing"
+    GRANT_REVOKED = "grant_revoked"
+
+
+def capture_rejection(database: Path, source_id: str) -> CaptureRejection | None:
+    """None when this turn may be captured, otherwise why it may not."""
     from one_memory.memory_server.config_state import ConfigStore
 
     if not ConfigStore(database.parent).sync_enabled(source_id):
-        return False
+        return CaptureRejection.SYNC_DISABLED
     with main_database(database).read() as connection:
         row = connection.execute(
-            "SELECT 1 FROM source_grants WHERE user_id='local' AND source_id=? "
-            "AND revoked_at IS NULL",
+            "SELECT revoked_at FROM source_grants WHERE user_id='local' AND source_id=?",
             (source_id,),
         ).fetchone()
-    return row is not None
+    if row is None:
+        return CaptureRejection.GRANT_MISSING
+    # Indexed, not named: this connection has no row factory, and a hook is
+    # not the place to add one.
+    if row[0] is not None:
+        return CaptureRejection.GRANT_REVOKED
+    return None
+
+
+def record_capture_rejection(
+    database: Path,
+    source_id: str,
+    reason: CaptureRejection,
+    *,
+    window_seconds: float = 600.0,
+) -> None:
+    """Note one dropped turn, at most once per window per source and reason.
+
+    Best effort by construction: a hook runs inside somebody's turn, so a sink
+    that is slow, full or read-only must never surface here. The record is a
+    count and nothing else — no payload, no path, no session or conversation
+    id — because this file is meant to be handed to support as is.
+
+    One small file per source and reason, rather than a line per turn: capture
+    runs on every prompt, and a log that grew with the conversation would be a
+    privacy problem wearing a diagnosis as a disguise.
+    """
+    try:
+        directory = database.parent / "logs"
+        directory.mkdir(parents=True, exist_ok=True)
+        marker = directory / f"capture-rejection-{source_id}-{reason.value}.json"
+        now = time.time()
+        payload: dict[str, object] = {"source_id": source_id, "reason": reason.value}
+        if marker.exists() and now - marker.stat().st_mtime < window_seconds:
+            payload = json.loads(marker.read_text(encoding="utf8"))
+            previous = payload.get("count", 1)
+            payload["count"] = int(previous) + 1 if isinstance(previous, (str, int, float)) else 2
+        else:
+            payload["count"] = 1
+            payload["window_started_at"] = datetime.fromtimestamp(now, UTC).isoformat()
+            payload["last_at"] = payload["window_started_at"]
+        marker.write_text(json.dumps(payload, sort_keys=True), encoding="utf8")
+    except Exception:
+        return
 
 
 def _submit_live_memory(store: EvidenceStore, record: EvidenceRecord, *, changed: bool) -> None:
@@ -167,7 +236,7 @@ def _submit_live_memory(store: EvidenceStore, record: EvidenceRecord, *, changed
     ):
         if (
             record in chunk
-            and any(item.role == "user" and item.written_by is WrittenBy.PERSON for item in chunk)
+            and any(item.role == "user" for item in chunk)
             and any(item.role == "assistant" for item in chunk)
         ):
             intake.submit_records(chunk, priority="live")

memhub/apps/core/src/one_memory/matters/discovery.py

@@ -160,7 +160,9 @@ class MatterDiscovery:
         task_by_id: dict[str, MemoryTask] = {}
         for chunk in chunks:
             for record in chunk:
-                task = await asyncio.to_thread(processor.tasks.task_for_evidence, record.evidence_id)
+                task = await asyncio.to_thread(
+                    processor.tasks.task_for_evidence, record.evidence_id
+                )
                 if task is not None and task.status is MemoryTaskStatus.COMPLETED:
                     task_by_id[task.task_id] = task
         tasks = tuple(sorted(task_by_id.values(), key=lambda task: (task.created_at, task.task_id)))

memhub/apps/core/src/one_memory/matters/synthesis.py

@@ -270,7 +270,7 @@ async def reconcile_current(
         (
             r
             for r in records
-            if r.written_by == "person"
+            if r.role == "user"
             and r.created_at
             and (r.source_id, r.conversation_id) in related
         ),

memhub/apps/core/src/one_memory/matters/validation.py

@@ -51,14 +51,19 @@ def validate_claim(claim: Claim, records: Mapping[str, EvidenceRecord | SampleRe
         text = record.text if isinstance(record, SampleRecord) else record.verbatim
         if not citation.quote.strip() or citation.quote not in text:
             raise ValueError("citation quote is not in the source")
-        authors.add(str(record.written_by))
+        author = {
+            "user": "person",
+            "assistant": "agent",
+            "tool": "tool",
+        }.get(record.role, "unknown")
+        authors.add(author)
         created = record.created_at
         citations.append(
             citation.model_copy(
                 update={
                     "revision": record.revision,
                     "source_id": record.source_id,
-                    "written_by": str(record.written_by),
+                    "written_by": author,
                     "created_at": created.isoformat() if isinstance(created, datetime) else created,
                 }
             )

memhub/apps/core/src/one_memory/memory_library.py

@@ -290,7 +290,7 @@ def read_memory_library(
                            COALESCE(
                                MIN(
                                    CASE WHEN e.role = 'user'
-                                         AND e.written_by IN ('person', 'user')
+                                         AND e.role = 'user'
                                         THEN e.created_at END
                                ),
                                MIN(e.created_at)

memhub/apps/core/src/one_memory/memory_processing.py

@@ -45,6 +45,7 @@ from one_memory.hy_memory_bridge import (
     HyMemoryOutput,
     HyMemoryTurn,
     builtin_retrieval_snapshot,
+    coding_tool_result_max_bytes,
     decide_profile_graph_width,
     hy_memory_store_exists,
     inspect_profile_graph_width,
@@ -52,7 +53,7 @@ from one_memory.hy_memory_bridge import (
 from one_memory.hy_runtime_config import hy_memory_version
 from one_memory.hy_source_context import HySourceContext, HySupportedMemory
 from one_memory.imports.access import opaque_workspace_id
-from one_memory.imports.models import EvidenceContentKind, EvidenceOrigin, EvidenceRecord, WrittenBy
+from one_memory.imports.models import EvidenceContentKind, EvidenceOrigin, EvidenceRecord
 from one_memory.imports.runner import live_sync_arrival
 from one_memory.imports.storage import EvidenceStore
 from one_memory.input_sanitization import sanitize_model_input
@@ -162,7 +163,7 @@ def _report_stale_known_reasons(task_id: str, reasons: Mapping[str, int]) -> Non
     """Emit one schema-valid scalar diagnostic per stale-memory reason."""
     for reason, count in sorted(reasons.items()):
         diagnostic_event(
-            level="info",
+            level="debug",
             component="memory.context",
             event="known_memory_stale_omitted",
             outcome="omitted",
@@ -173,6 +174,21 @@ def _report_stale_known_reasons(task_id: str, reasons: Mapping[str, int]) -> Non
 
 MEMORY_INGEST_POLICY_VERSION = "hy-native-v23"
 UNIFIED_HY_AGENT_ID = "one-memory"
+#: Deterministic refusals raised before the payload reaches HY, so nothing was
+#: sent even though the invocation marker is already recorded by then: the
+#: write bindings are built inside `ingest` ahead of the call, and the input
+#: window is checked while the request is frozen.
+#:
+#: They are terminal rather than unknown. `unknown` here means "HY may hold a
+#: write we cannot confirm", and that is exactly what did not happen — parking
+#: the task on it hid the reason and stopped the queue behind it (issue #44).
+#: Kept in step with `durable_queue.PRE_SEND_REFUSAL_CODES`, which the durable
+#: write path applies to the same codes.
+PRE_SEND_REFUSAL_CODES = frozenset(
+    {
+        "hy_memory_input_exceeds_window",
+    }
+)
 _BACKFILL_BATCH_SIZE = 8
 _BACKFILL_DRAIN_BUDGET_SECONDS = 330.0
 _HISTORY_MERGE_TURNS = 8
@@ -280,6 +296,7 @@ class MemoryTaskRunner:
         *,
         timeout: float | None = None,
         before_hy_invocation: Callable[[], Awaitable[None]] | None = None,
+        input_records: tuple[EvidenceRecord, ...] | None = None,
     ) -> MemoryTask:
         started = monotonic()
         diagnostic_event(
@@ -290,13 +307,28 @@ class MemoryTaskRunner:
             task_id=task_id,
         )
         try:
-            body = (
-                self._run_without_content_permission(task_id)
-                if before_hy_invocation is None
-                else self._run_without_content_permission(
-                    task_id, before_hy_invocation=before_hy_invocation
+            if input_records is None:
+                body = (
+                    self._run_without_content_permission(task_id)
+                    if before_hy_invocation is None
+                    else self._run_without_content_permission(
+                        task_id,
+                        before_hy_invocation=before_hy_invocation,
+                    )
+                )
+            else:
+                body = (
+                    self._run_without_content_permission(
+                        task_id,
+                        input_records=input_records,
+                    )
+                    if before_hy_invocation is None
+                    else self._run_without_content_permission(
+                        task_id,
+                        before_hy_invocation=before_hy_invocation,
+                        input_records=input_records,
+                    )
                 )
-            )
             # The ceiling bounds how long this task holds its harness's slot:
             # a harness is serial, so one hung task starves every turn behind
             # it. It is insurance against a single bad call, not a way to run
@@ -367,7 +399,7 @@ class MemoryTaskRunner:
     @staticmethod
     def _log_step(task_id: str, step: int, action: str, started: float) -> None:
         diagnostic_event(
-            level="info",
+            level="debug",
             component="memory.task",
             event="memory_task.step_completed",
             outcome="succeeded",
@@ -381,6 +413,7 @@ class MemoryTaskRunner:
         task_id: str,
         *,
         before_hy_invocation: Callable[[], Awaitable[None]] | None = None,
+        input_records: tuple[EvidenceRecord, ...] | None = None,
     ) -> MemoryTask:
         # Every store call here is synchronous, and several are writes that
         # wait on the write lock for up to five seconds. Made from the loop,
@@ -434,7 +467,9 @@ class MemoryTaskRunner:
         step_started = monotonic()
         frozen = await in_the_background(self.tasks.inputs, task.task_id)
         chunk: tuple[EvidenceRecord, ...] | None
-        if frozen:
+        if input_records is not None:
+            chunk = input_records
+        elif frozen:
             try:
                 chunk = await in_the_background(self._frozen_chunk, task)
             except HyMemoryBridgeError as error:
@@ -467,8 +502,11 @@ class MemoryTaskRunner:
                     self._source_context, task, request.source_context
                 )
                 request = request.model_copy(update={"source_context": context})
+            uses_frozen_input = bool(frozen) or input_records is not None
             grant_token = (
-                await in_the_background(self._grant_token, task, chunk[0]) if frozen else None
+                await in_the_background(self._grant_token, task, chunk[0])
+                if uses_frozen_input
+                else None
             )
             memory_contents = tuple(
                 sanitize_model_input(record.verbatim)
@@ -511,19 +549,24 @@ class MemoryTaskRunner:
                     agent_id=request.agent_id,
                 ),
             )
-            if frozen:
-                await in_the_background(self._frozen_chunk, task)
+            if uses_frozen_input:
+                if input_records is None:
+                    await in_the_background(self._frozen_chunk, task)
                 if grant_token != await in_the_background(self._grant_token, task, chunk[0]):
                     raise HyMemoryBridgeError("hy_memory_source_changed")
             if request.source_context is not None:
-                await in_the_background(self._check_source_context, request.source_context)
+                await in_the_background(
+                    self._check_source_context,
+                    request.source_context,
+                    skip_current=input_records is not None,
+                )
             # One thread hop for the whole batch rather than a dozen writes
             # made from the loop.
             step_started = monotonic()
             await in_the_background(
                 self._record_outputs,
                 task=task,
-                chunk=self._original_records(chunk),
+                chunk=chunk,
                 outputs=outputs,
                 request=request,
                 derivation=derivation,
@@ -533,6 +576,13 @@ class MemoryTaskRunner:
             self._log_step(task_id, 6, "governance", step_started)
         except HyMemoryBridgeError as error:
             code = str(error)
+            if code == "model_sedimentation_disabled" and not hy_invoked:
+                return await in_the_background(
+                    self.tasks.release_unstarted,
+                    task_id,
+                    runner_token=runner_token,
+                    error_code=code,
+                )
             model_wait_errors = {
                 "hy_memory_unavailable",
                 "hy_memory_rate_limit",
@@ -547,6 +597,16 @@ class MemoryTaskRunner:
                 "hy_memory_model_timeout",
             }
             safe_terminal_after_invocation = {"hy_memory_import_no_readable_output"}
+            if code in PRE_SEND_REFUSAL_CODES:
+                # Raised before the call, so the marker only says "about to":
+                # fail with the reason rather than parking on a write HY never
+                # saw. See the constant and issue #44.
+                return await in_the_background(
+                    self.tasks.fail,
+                    task_id,
+                    runner_token=runner_token,
+                    error_code=code,
+                )
             if hy_invoked and (
                 code == "hy_memory_evidence_changed"
                 or (
@@ -744,8 +804,18 @@ class MemoryTaskRunner:
                 raise HyMemoryBridgeError("hy_memory_evidence_changed") from error
         return context
 
-    def _check_source_context(self, context: HySourceContext) -> None:
-        for original in (*context.current, *context.history, *context.supporting):
+    def _check_source_context(
+        self,
+        context: HySourceContext,
+        *,
+        skip_current: bool = False,
+    ) -> None:
+        checked = (
+            (*context.history, *context.supporting)
+            if skip_current
+            else (*context.current, *context.history, *context.supporting)
+        )
+        for original in checked:
             record = self.evidence.get(original.evidence_id)
             if record is None or not matches_evidence_ref(record, evidence_ref(original)):
                 raise HyMemoryBridgeError("hy_memory_evidence_changed")
@@ -846,7 +916,7 @@ class MemoryTaskRunner:
             # HY's own attribution. An undecided memory counts as the person's:
             # HY's prompt defaults that way.
             from_agent = output.owner == "agent" or all(
-                record.written_by is WrittenBy.AGENT
+                record.role == "assistant"
                 and record.content_kind is EvidenceContentKind.RAW_CONVERSATION
                 for record in support
             )
@@ -1016,15 +1086,20 @@ class MemoryTaskRunner:
         task: MemoryTask,
         chunk: tuple[EvidenceRecord, ...],
     ) -> HyMemoryIngestRequest:
+        if not chunk:
+            # A task with no inputs has nothing to name the request after and
+            # nothing to attribute an answer to. Left alone this reached
+            # `chunk[0]` below and came out as IndexError, which no caller can
+            # read: a reconcile pass catches domain errors per task and moves
+            # on, so one such task ended the whole pass and left every other
+            # unknown task unreconciled. On this machine 68 of 303 unknown
+            # tasks had no inputs, and the pass raised on the first of them.
+            raise HyMemoryBridgeError("hy_memory_task_without_inputs")
         has_raw = any(
             record.content_kind is EvidenceContentKind.RAW_CONVERSATION for record in chunk
         )
         source = next(
-            (
-                record
-                for record in chunk
-                if record.role == "user" and record.written_by is WrittenBy.PERSON
-            ),
+            (record for record in chunk if record.role == "user"),
             chunk[0],
         )
         request_turns = tuple(
@@ -1033,12 +1108,9 @@ class MemoryTaskRunner:
             if (
                 record.content_kind is EvidenceContentKind.RAW_CONVERSATION
                 and (
-                    # D98: any `user` message goes, whoever the parser says
-                    # wrote it. A sub agent's question is ordinary
-                    # user-assistant data; the author field cannot tell who
-                    # wrote a message, and filtering on it dropped the sub
-                    # agent's turns. `written_by` still labels the display and
-                    # decides owner — it no longer decides what is sent.
+                    # Protocol role is authoritative. A sub-agent question is
+                    # an ordinary user message regardless of a host parser's
+                    # former author guess.
                     record.role == "user"
                     # The assistant's side goes too. Without it 「就按你说的那个来」
                     # arrives with nothing to resolve it against, and HY's own
@@ -1049,7 +1121,7 @@ class MemoryTaskRunner:
                     or record.role == "assistant"
                     # Tool calls go as they happened; whether they make the
                     # turn a coding turn is HY's call, not ours.
-                    or (record.role == "tool" and record.written_by is WrittenBy.TOOL)
+                    or record.role == "tool"
                 )
             )
             or (not has_raw and record.content_kind is EvidenceContentKind.MEMORY_CONTENT)
@@ -1218,6 +1290,7 @@ class MemoryProcessingSupervisor:
         max_concurrent_tasks: int | None = None,
         source_writes_permitted: Callable[[str], bool] | None = None,
         service_levels: ServiceLevelScheduler | None = None,
+        sedimentation_enabled: Callable[[], bool] | None = None,
     ) -> None:
         self.tasks = tasks
         self.profiles = profiles
@@ -1230,6 +1303,12 @@ class MemoryProcessingSupervisor:
         #: absent every source is treated as writable, which is what the
         #: product did before the switch existed.
         self.source_writes_permitted = source_writes_permitted
+        #: User-level "stop turning conversations into memories" switch. Optional
+        #: so tests and a Core without a config file keep their current
+        #: behaviour; when absent sedimentation is on, which is what the product
+        #: did before the switch existed. Read per decision rather than cached,
+        #: so stopping takes effect on the next dispatch without a restart.
+        self.sedimentation_enabled = sedimentation_enabled
         self.service_levels = service_levels or ServiceLevelScheduler()
         self.extraction_settings = extraction_settings or ExtractionSettingsStore(
             data_dir / "one-memory.sqlite3"
@@ -1283,12 +1362,26 @@ class MemoryProcessingSupervisor:
         self._orphan_reclaimer: asyncio.Task[None] | None = None
         self._diagnostic_idle_reason: str | None = None
 
+    def _sedimentation_stopped(self) -> bool:
+        """Whether the user stopped model-driven sedimentation.
+
+        Stopping must not look like a broken pipeline: nothing is sent to the
+        model, so the pending count stays as it is and the scheduler reports
+        why it is idle. Starting again picks the same work up.
+        """
+        if self.sedimentation_enabled is None:
+            return False
+        if self.sedimentation_enabled():
+            return False
+        self._record_idle("sedimentation_stopped")
+        return True
+
     def _record_idle(self, reason: str) -> None:
         if self._diagnostic_idle_reason == reason:
             return
         self._diagnostic_idle_reason = reason
         diagnostic_event(
-            level="info",
+            level="debug",
             component="memory.scheduler",
             event="memory_scheduler.idle",
             outcome="idle",
@@ -1310,7 +1403,8 @@ class MemoryProcessingSupervisor:
     ) -> tuple[MemoryTask, ...]:
         """Process a bounded, authorized snapshot through the existing HY runner."""
         async with self._explicit_processing_lock(records):
-            records = await in_the_background(self._original_records, records)
+            if self._sedimentation_stopped():
+                raise HyMemoryBridgeError("model_sedimentation_disabled")
             active = await in_the_background(self.profiles.active, self.secrets)
             if active is None or (profile_id is not None and profile_id != active.profile_id):
                 raise HyMemoryBridgeError("model_profile_not_active")
@@ -1325,13 +1419,6 @@ class MemoryProcessingSupervisor:
                 raise HyMemoryBridgeError("hy_memory_source_disabled")
             conversations: dict[tuple[str, str, str | None], list[EvidenceRecord]] = {}
             for record in records:
-                stored = await in_the_background(self.evidence.get, record.evidence_id)
-                if (
-                    stored is None
-                    or stored.revision != record.revision
-                    or stored.content_hash != record.content_hash
-                ):
-                    raise ValueError("hy_memory_evidence_changed")
                 conversations.setdefault(
                     (record.source_id, record.conversation_id, record.workspace), []
                 ).append(record)
@@ -1348,7 +1435,7 @@ class MemoryProcessingSupervisor:
                 for conversation in conversations.values()
                 for chunk in _conversation_chunks(conversation, max_turns=12, completed_only=True)
             ]
-            prepared = []
+            prepared: list[tuple[MemoryTask, tuple[EvidenceRecord, ...]]] = []
             for chunk in chunks:
                 task = await in_the_background(
                     self._enqueue_chunk,
@@ -1356,12 +1443,12 @@ class MemoryProcessingSupervisor:
                     active,
                     freeze_inputs=True,
                 )
-                prepared.append(task)
+                prepared.append((task, chunk))
             runner = self.runner_factory()
             completed = []
             if on_progress:
-                on_progress(tuple(prepared), 0)
-            for task in prepared:
+                on_progress(tuple(task for task, _chunk in prepared), 0)
+            for task, chunk in prepared:
                 if task.status is not MemoryTaskStatus.COMPLETED:
                     # Another importer may already own this task. Never race its claim.
                     task = await in_the_background(self.tasks.require, task.task_id)
@@ -1390,23 +1477,41 @@ class MemoryProcessingSupervisor:
                                 work_id=task.task_id,
                                 scope_key=chunk[0].source_id,
                             ):
-                                task = (
-                                    await runner.run(task.task_id, timeout=timeout)
-                                    if before_hy_invocation is None
-                                    else await runner.run(
-                                        task.task_id,
-                                        timeout=timeout,
-                                        before_hy_invocation=before_hy_invocation,
+                                if isinstance(runner, MemoryTaskRunner):
+                                    task = (
+                                        await runner.run(
+                                            task.task_id,
+                                            timeout=timeout,
+                                            input_records=chunk,
+                                        )
+                                        if before_hy_invocation is None
+                                        else await runner.run(
+                                            task.task_id,
+                                            timeout=timeout,
+                                            before_hy_invocation=before_hy_invocation,
+                                            input_records=chunk,
+                                        )
+                                    )
+                                else:
+                                    task = (
+                                        await runner.run(task.task_id, timeout=timeout)
+                                        if before_hy_invocation is None
+                                        else await runner.run(
+                                            task.task_id,
+                                            timeout=timeout,
+                                            before_hy_invocation=before_hy_invocation,
+                                        )
                                     )
-                                )
                         finally:
                             _unmark_direct_task(self.tasks.database_path, task.task_id)
                 if task.status is not MemoryTaskStatus.COMPLETED:
                     raise HyMemoryBridgeError(task.last_error_code or "hy_memory_incomplete")
-                await in_the_background(runner._frozen_chunk, task)
                 completed.append(task)
                 if on_progress:
-                    on_progress(tuple(prepared), len(completed))
+                    on_progress(
+                        tuple(prepared_task for prepared_task, _chunk in prepared),
+                        len(completed),
+                    )
             expected = {record.evidence_id for record in records if _is_hy_turn_record(record)}
             actual = {
                 row["evidence_id"] for task in completed for row in self.tasks.inputs(task.task_id)
@@ -1417,14 +1522,24 @@ class MemoryProcessingSupervisor:
                 await self._refresh_read_bridge(completed[-1].profile_snapshot)
             return tuple(completed)
 
-    async def reconcile_unknown_outputs(self, *, limit: int = 128) -> int:
+    async def reconcile_unknown_outputs(
+        self,
+        *,
+        limit: int = 128,
+        exclude_error_codes: tuple[str, ...] = (),
+    ) -> int:
         """Publish HY outputs that can be proven by an unknown task's request id.
 
         This inquiry is read-only on HY and never invokes the model. Rows with
         no exact match remain unknown for the person's explicit retry choice.
         """
         reconciled = 0
-        for task in await in_the_background(self.tasks.unknown_tasks, limit=limit):
+        tasks = await in_the_background(
+            self.tasks.unknown_tasks,
+            limit=limit,
+            exclude_error_codes=exclude_error_codes,
+        )
+        for task in tasks:
             runner = self.runner_factory()
             bridge: MemoryBridgeProtocol | None = None
             try:
@@ -1667,6 +1782,8 @@ class MemoryProcessingSupervisor:
     def enqueue_evidence(self, record: EvidenceRecord) -> MemoryTask | None:
         if self._shutting_down:
             return None
+        if self._sedimentation_stopped():
+            return None
         if (
             record.role not in {"user", "assistant"}
             and record.content_kind is not EvidenceContentKind.MEMORY_CONTENT
@@ -1699,6 +1816,8 @@ class MemoryProcessingSupervisor:
         should_stop: Callable[[], bool] = lambda: False,
     ) -> list[QueuedWork]:
         """Synchronous compatibility entry; preparation never starts a model."""
+        if self._sedimentation_stopped():
+            return []
         work = [item for batch in self._prepare_backfill_batches(should_stop) for item in batch]
         return [] if should_stop() or self._shutting_down else work
 
@@ -1707,7 +1826,7 @@ class MemoryProcessingSupervisor:
         should_stop: Callable[[], bool] = lambda: False,
     ) -> Generator[list[QueuedWork], None, None]:
         """Read bodies as needed; yield the first task before the next conversation."""
-        if should_stop() or self._shutting_down:
+        if should_stop() or self._shutting_down or self._sedimentation_stopped():
             return
         active = self.profiles.active(self.secrets)
         if active is None:
@@ -1718,7 +1837,7 @@ class MemoryProcessingSupervisor:
         visited = 0
         seed_ids = self.evidence.list_conversation_seed_ids()
         diagnostic_event(
-            level="info",
+            level="debug",
             component="memory.scheduler",
             event="memory_scheduler.scan_inventory",
             outcome="observed",
@@ -1788,7 +1907,7 @@ class MemoryProcessingSupervisor:
         started = monotonic()
         prepared_count = 0
         diagnostic_event(
-            level="info",
+            level="debug",
             component="memory.scheduler",
             event="memory_scheduler.scan_started",
             outcome="started",
@@ -1800,6 +1919,7 @@ class MemoryProcessingSupervisor:
         stop = threading.Event()
         batches = self._prepare_backfill_batches(stop.is_set)
         exhausted = False
+        stopped_saturated = False
         rounds = 0
         scan_error: BaseException | None = None
         try:
@@ -1815,20 +1935,21 @@ class MemoryProcessingSupervisor:
                             >= self._backfill_round_budget_seconds
                         ):
                             break
-                        while (
+                        if (
                             cohort_size is not None
                             and len(self._queue) + len(self._in_flight) >= cohort_size
                         ):
-                            if self._shutting_down:
-                                return
-                            if (
-                                handled_batch
-                                and self._backfill_clock() - round_started
-                                >= self._backfill_round_budget_seconds
-                            ):
-                                break
-                            self._record_idle("queue_saturated")
-                            await asyncio.sleep(0.1)
+                            # A bounded scan stops at its ceiling instead of
+                            # waiting for the queue to drain. Waiting here slept
+                            # inside the scan lock, so whoever asked for the
+                            # scan — a switch toggle, over HTTP — waited with
+                            # it, and on a saturated backlog that is not a
+                            # moment but the whole backlog. The caller's
+                            # timeout then reported a failure for work that had
+                            # already been accepted.
+                            stopped_saturated = True
+                            exhausted = True
+                            break
                         if (
                             handled_batch
                             and self._backfill_clock() - round_started
@@ -1847,7 +1968,7 @@ class MemoryProcessingSupervisor:
                             on_first_scheduled()
                 if not exhausted and not self._shutting_down:
                     diagnostic_event(
-                        level="info",
+                        level="debug",
                         component="memory.scheduler",
                         event="memory_scheduler.scan_yielded",
                         outcome="yielded",
@@ -1877,14 +1998,18 @@ class MemoryProcessingSupervisor:
             stop.set()
             batches.close()
             if prepared_count == 0 and not self._shutting_down:
-                if scan_error is None:
+                if stopped_saturated:
+                    # Parked behind a full queue, not out of work: the status
+                    # must not read "no_tasks" while hundreds wait.
+                    self._record_idle("queue_saturated")
+                elif scan_error is None:
                     self._record_idle("no_tasks")
                 else:
                     self._record_idle("scan_failed")
             elif prepared_count:
                 self._diagnostic_idle_reason = None
             diagnostic_event(
-                level="info",
+                level="debug",
                 component="memory.scheduler",
                 event="memory_scheduler.scan_completed",
                 outcome=(
@@ -2092,6 +2217,8 @@ class MemoryProcessingSupervisor:
         Work that falls outside the extraction window is marked deferred so
         the pending count stops pretending it is about to run.
         """
+        if self._sedimentation_stopped():
+            return []
         now = datetime.now(UTC)
         window = self._extraction_window() if within_window else None
         exempt_all_apps_from_window = True
@@ -2174,7 +2301,7 @@ class MemoryProcessingSupervisor:
         if work:
             self._diagnostic_idle_reason = None
             diagnostic_event(
-                level="info",
+                level="debug",
                 component="memory.scheduler",
                 event="memory_scheduler.tasks_queued",
                 outcome="succeeded",
@@ -2239,7 +2366,7 @@ class MemoryProcessingSupervisor:
         return self.HARNESS_SERIAL
 
     def _ensure_workers(self) -> None:
-        if self._shutting_down:
+        if self._shutting_down or self._sedimentation_stopped():
             return
         load = self._harness_load()
         background_capacity = (
@@ -2317,6 +2444,8 @@ class MemoryProcessingSupervisor:
         """
         looked_for_older = False
         while not self._shutting_down:
+            if self._sedimentation_stopped():
+                return
             # The harness cap is recomputed from what is running, so a worker
             # that just finished frees its harness's slot immediately.
             work = self._queue.take(
@@ -2370,7 +2499,16 @@ class MemoryProcessingSupervisor:
                         scope_key=scope_key,
                     ),
                 ):
-                    finished = await runner.run(work.task_id, timeout=self._profile_timeout)
+
+                    async def require_sedimentation() -> None:
+                        if self._sedimentation_stopped():
+                            raise HyMemoryBridgeError("model_sedimentation_disabled")
+
+                    finished = await runner.run(
+                        work.task_id,
+                        timeout=self._profile_timeout,
+                        before_hy_invocation=require_sedimentation,
+                    )
                 if finished.status is MemoryTaskStatus.COMPLETED:
                     self._request_read_bridge_refresh(finished.profile_snapshot)
                 elif (
@@ -3049,7 +3187,7 @@ class MemoryProcessingSupervisor:
             self._read_bridges_ready.add(key)
             self._read_bridge_generations[key] = generation
             diagnostic_event(
-                level="info",
+                level="debug",
                 component="memory.semantic",
                 event="semantic_read_bridge.refreshed",
                 outcome="succeeded",
@@ -3172,9 +3310,7 @@ class MemoryProcessingSupervisor:
         if not tail:
             return False
         last = tail[-1]
-        return bool(
-            last.role == "user" and last.written_by is WrittenBy.PERSON and _opens_turn(last)
-        )
+        return bool(last.role == "user" and _opens_turn(last))
 
     def _admitted_chunks(
         self,
@@ -3519,7 +3655,6 @@ def _is_memory_note(record: EvidenceRecord) -> bool:
     meta = record.origin_meta.get(EvidenceOrigin.MANUAL_IMPORT)
     return (
         record.source_id == "one-memory-manual"
-        and record.written_by is WrittenBy.PERSON
         and meta is not None
         and meta.raw_meta.get("input_mode") == "direct_memory_note"
     )
@@ -3595,7 +3730,7 @@ _HOST_ENVELOPE = re.compile(
 #: before HY chose a path cut long answers mid-way and sent tool-heavy turns as
 #: dozens of calls. ONE_MEMORY_SPLIT_TURNS=1 restores the 12,000 / 24 split.
 _WHOLE_TURN_CHARACTERS = 60_000
-_HY_TOOL_CHARACTERS = 2_048
+_HY_TOOL_CHARACTERS = coding_tool_result_max_bytes()
 #: Hard ceiling on the UTF-8 bytes one window may hold, whatever HY counts.
 #: HY reads the payload in-process, so a turn made of enormous raw tool
 #: records is held in memory even though HY will cut each one down. Cut by
@@ -3814,14 +3949,12 @@ def _merge_historical_chunks(
 
 
 def _chunk_input_ref(chunk: tuple[EvidenceRecord, ...]) -> str:
-    person_records = tuple(
-        item for item in chunk if item.role == "user" and item.written_by is WrittenBy.PERSON
-    )
-    return person_records[-1].evidence_id if person_records else chunk[-1].evidence_id
+    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" and record.written_by is WrittenBy.PERSON
+    return record.role == "user"
 
 
 def _trailing_user_turn_is_open(turn: list[EvidenceRecord]) -> bool:
@@ -3838,10 +3971,8 @@ def _trailing_user_turn_is_open(turn: list[EvidenceRecord]) -> bool:
 def _is_hy_turn_record(record: EvidenceRecord) -> bool:
     """What HY receives of a turn: the input turns and the tool calls made in it."""
     if record.role == "tool":
-        return (
-            record.written_by is WrittenBy.TOOL
-            and record.content_kind is EvidenceContentKind.RAW_CONVERSATION
-            and bool(record.verbatim.strip())
+        return record.content_kind is EvidenceContentKind.RAW_CONVERSATION and bool(
+            record.verbatim.strip()
         )
     return _is_hy_input_turn(record)
 
@@ -3858,11 +3989,8 @@ def _is_hy_input_turn(record: EvidenceRecord) -> bool:
     # from a model.
     if record.role == "assistant":
         return True
-    # D98: the author no longer decides. A sub agent's question arrives as a
-    # `user` message marked written by the agent — it is ordinary
-    # user-assistant data, and dropping it sent Claude Code transcripts to HY
-    # with every question but the first missing. `written_by` labels the
-    # display and decides owner; it cannot tell who actually wrote a message.
+    # A sub-agent question is still a protocol `user` message. Host-specific
+    # author guesses do not decide whether it reaches HY.
     if record.role != "user":
         return False
     if source.casefold().startswith(_HARNESS_PREFIXES):

memhub/apps/core/src/one_memory/memory_server/config_state.py

@@ -30,6 +30,12 @@ class ConfigDocument(BaseModel):
 
     schema_version: int = Field(default=1, ge=1, le=1)
     sources: dict[str, SourceConfig] = Field(default_factory=dict)
+    #: Whether the configured model may turn imported conversations into
+    #: memories. Defaulting to on keeps every existing install doing what it
+    #: did before the switch existed; stopping it is the user's cost control
+    #: and leaves the model configuration itself untouched, so starting again
+    #: needs no re-entry of a key.
+    sedimentation_enabled: StrictBool = True
 
     @field_validator("sources")
     @classmethod
@@ -63,6 +69,19 @@ class ConfigStore:
             source = self._document.sources.get(source_id)
             return bool(source and source.sync_enabled)
 
+    def sedimentation_enabled(self) -> bool:
+        with self._lock:
+            return bool(self._document.sedimentation_enabled)
+
+    def set_sedimentation_enabled(self, enabled: bool) -> ConfigDocument:
+        if type(enabled) is not bool:
+            raise ValueError("invalid_sedimentation_config")
+        with self._lock:
+            updated = self._document.model_copy(update={"sedimentation_enabled": enabled})
+            self._write(updated)
+            self._document = updated
+            return updated
+
     def set_sync(self, source_id: str, enabled: bool) -> ConfigDocument:
         return self.set_sync_many({source_id: enabled})
 

memhub/apps/core/src/one_memory/memory_server/contracts.py

@@ -23,10 +23,13 @@ import hashlib
 import hmac
 import os
 import uuid
+from datetime import datetime
 from enum import StrEnum
 from typing import Annotated, Final, Literal
 
-from pydantic import BaseModel, ConfigDict, Field, field_validator
+from pydantic import BaseModel, ConfigDict, Field, ValidationInfo, field_validator
+
+from one_memory.imports.models import EvidenceRecord
 
 #: Bumped when a derivation rule changes, so old and new ids never compare equal.
 SCHEMA_VERSION: Final[str] = "ms1"
@@ -166,7 +169,6 @@ class CanonicalMessage(BaseModel):
     message_id: str | None = Field(default=None, min_length=1, max_length=512)
     ordinal: int = Field(ge=0)
     role: str = Field(min_length=1, max_length=128)
-    written_by: str = Field(min_length=1, max_length=64)
     content: str = Field(min_length=1)
     workspace: str | None = Field(default=None, max_length=1024)
 
@@ -213,8 +215,13 @@ class WriteRequest(BaseModel):
     source_id: str | None = Field(default=None, min_length=1, max_length=128)
     batch_identity: BatchIdentity
     evidence_ids: list[str] = Field(min_length=1)
+    # New jobs freeze the complete input in their own payload. The default is
+    # solely for reading pre-upgrade rows whose payload only carried ids.
+    input_records: tuple[EvidenceRecord, ...] = ()
     idempotency_key: str = Field(min_length=1, max_length=256)
     priority: Literal["live", "normal"] = "normal"
+    event_at: datetime | None = None
+    time_quality: Literal["exact", "anchored", "session", "approximate"] = "approximate"
 
     @field_validator("evidence_ids")
     @classmethod
@@ -227,6 +234,16 @@ class WriteRequest(BaseModel):
             raise ValueError("evidence_ids must be unique")
         return value
 
+    @field_validator("input_records")
+    @classmethod
+    def _input_records_match_ids(
+        cls, value: tuple[EvidenceRecord, ...], info: ValidationInfo
+    ) -> tuple[EvidenceRecord, ...]:
+        evidence_ids = info.data.get("evidence_ids")
+        if value and evidence_ids != [record.evidence_id for record in value]:
+            raise ValueError("input_records must match evidence_ids in order")
+        return value
+
     @property
     def request_key(self) -> str:
         """Stable dedup key. Same batch + same intent must not enqueue twice."""

memhub/apps/core/src/one_memory/memory_server/durable_queue.py

@@ -30,7 +30,7 @@ from uuid import uuid4
 from one_memory.background_work import in_the_background
 from one_memory.database_access import main_database
 from one_memory.diagnostic_log import diagnostic_event
-from one_memory.imports.models import EvidenceRecord
+from one_memory.imports.models import EvidenceOrigin, EvidenceRecord
 from one_memory.imports.storage import EvidenceStore
 from one_memory.memory_server.contracts import WriteRequest
 from one_memory.memory_server.data_safety import MIN_FREE_BYTES  # noqa: F401 - re-exported
@@ -38,10 +38,11 @@ from one_memory.memory_server.migrations import memory_write_jobs_sqlite_schema
 from one_memory.memory_server.queue import build_reclaim_sqlite_statements
 from one_memory.memory_server.registry import WriteRecord
 from one_memory.memory_tasks import (
-    POLICY_STOPPED_CODES as _TASK_POLICY_STOPPED_CODES,
+    NON_RETRYABLE_POST_WRITE_CODES,
+    MemoryTaskStatus,
 )
 from one_memory.memory_tasks import (
-    MemoryTaskStatus,
+    POLICY_STOPPED_CODES as _TASK_POLICY_STOPPED_CODES,
 )
 from one_memory.service_levels import ServiceLevel, service_level_context
 
@@ -249,6 +250,7 @@ MODEL_NOT_READY_CODES = frozenset(
 #: 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).
+#:
 PRE_SEND_REFUSAL_CODES = frozenset(
     {
         "hy_memory_input_exceeds_window",
@@ -282,11 +284,6 @@ _VALID_JOB_STATES = frozenset(
 #: A row in one of these states can never be moved back to a live one.
 _TERMINAL_WRITE_STATES = frozenset({COMPLETED, FAILED, DEAD, CANCELLED})
 
-#: A job in one of these states can never run again, so its payload is dead
-#: weight. `unknown` is deliberately absent: it is non-terminal and still needs
-#: its intent for reconciliation. `running` is obviously still live.
-TERMINAL_PAYLOAD_STATES = (COMPLETED, FAILED, DEAD, CANCELLED)
-
 #: Columns a current schema declares but an older `memory_write_jobs` may lack.
 #: `CREATE TABLE IF NOT EXISTS` is a no-op for an existing table, so without
 #: this an upgraded install would keep an older queue that every claim/lease
@@ -304,9 +301,12 @@ LEGACY_MEMORY_WRITE_JOB_COLUMNS = (
     ("lease_expires_at", "TEXT"),
     ("hy_invoked", "INTEGER NOT NULL DEFAULT 0"),
     ("last_error", "TEXT"),
+    ("event_at", "TEXT"),
+    ("time_quality", "TEXT NOT NULL DEFAULT 'approximate'"),
 )
 
-_FIFO_BLOCKING_STATES = (PENDING, RUNNING, WAITING_FOR_PERMISSION, WAITING_FOR_MODEL, UNKNOWN)
+_FIFO_BLOCKING_STATES = (PENDING, RUNNING, WAITING_FOR_PERMISSION, WAITING_FOR_MODEL)
+EVENT_TIME_WINDOW = timedelta(minutes=10)
 
 
 @dataclass(frozen=True)
@@ -393,6 +393,42 @@ def _matches_payload_except_priority(stored: str, payload: str) -> bool:
     return old == new
 
 
+def _matches_legacy_reference_replay(stored: str, payload: str) -> bool:
+    """Match an old id-only caller to the same job's frozen payload."""
+    try:
+        old = json.loads(stored)
+        new = json.loads(payload)
+    except (TypeError, ValueError):
+        return False
+    if not isinstance(old, dict) or not isinstance(new, dict):
+        return False
+    old_inputs = old.get("input_records")
+    new_inputs = new.get("input_records")
+    if bool(old_inputs) == bool(new_inputs):
+        return False
+    old.pop("input_records", None)
+    new.pop("input_records", None)
+    return old == new
+
+
+def _matches_reaped_legacy_payload(stored: str, payload: str) -> bool:
+    """Compare a current frozen request with a pre-snapshot reaped payload."""
+    try:
+        old = json.loads(stored)
+        current = json.loads(payload)
+    except (TypeError, ValueError):
+        return False
+    if (
+        not isinstance(old, dict)
+        or not isinstance(current, dict)
+        or not isinstance(old.get(REAPED_PAYLOAD_MARKER), str)
+    ):
+        return False
+    current.pop("input_records", None)
+    legacy_payload = json.dumps(current, sort_keys=True)
+    return bool(old[REAPED_PAYLOAD_MARKER] == _payload_digest(legacy_payload))
+
+
 class DurableQueueSchemaUnavailableError(RuntimeError):
     """The durable queue schema has not been created yet.
 
@@ -445,6 +481,7 @@ class DurableWriteQueue:
         if not initialize:
             return
         with main_database(database_path).write() as connection, connection:
+            connection.row_factory = sqlite3.Row
             # First create the base table if absent. Then widen an older table
             # before replaying indexes that reference the new columns; a bare
             # CREATE INDEX would fail on a pre-priority queue.
@@ -455,6 +492,7 @@ class DurableWriteQueue:
             for statement in memory_write_jobs_sqlite_schema()[1:]:
                 connection.execute(statement)
             self._backfill_legacy_references(connection)
+            self._backfill_legacy_event_times(connection)
         # A process that died while holding a lease cannot run its own
         # recovery. Opening the queue is the first safe point to recover any
         # rows it left behind, before any new worker claims them.
@@ -479,8 +517,7 @@ class DurableWriteQueue:
             "UPDATE memory_write_jobs SET source_key=session_key WHERE source_key=''"
         )
         legacy_rows = connection.execute(
-            "SELECT job_id FROM memory_write_jobs WHERE queue_seq=0 "
-            "ORDER BY created_at, rowid"
+            "SELECT job_id FROM memory_write_jobs WHERE queue_seq=0 ORDER BY created_at, rowid"
         ).fetchall()
         next_seq = int(
             connection.execute(
@@ -514,6 +551,49 @@ class DurableWriteQueue:
         if columns and columns != ["state", "priority", "created_at"]:
             connection.execute("DROP INDEX idx_write_jobs_claim")
 
+    @staticmethod
+    def _backfill_legacy_event_times(connection: sqlite3.Connection) -> None:
+        if (
+            connection.execute(
+                "SELECT 1 FROM sqlite_master WHERE type='table' AND name='evidence'"
+            ).fetchone()
+            is None
+        ):
+            connection.execute(
+                "UPDATE memory_write_jobs SET event_at=created_at WHERE event_at IS NULL"
+            )
+            return
+        evidence_columns = {
+            str(row[1]) for row in connection.execute("PRAGMA table_info(evidence)").fetchall()
+        }
+        if not {"created_at", "imported_at", "role"} <= evidence_columns:
+            connection.execute(
+                "UPDATE memory_write_jobs SET event_at=created_at WHERE event_at IS NULL"
+            )
+            return
+        connection.execute(
+            """
+            UPDATE memory_write_jobs
+            SET event_at = COALESCE(
+                (
+                    SELECT MAX(COALESCE(e.created_at, e.imported_at))
+                    FROM memory_write_job_evidence AS linked
+                    JOIN evidence AS e ON e.evidence_id = linked.evidence_id
+                    WHERE linked.job_id = memory_write_jobs.job_id
+                      AND e.role = 'user'
+                ),
+                (
+                    SELECT MIN(COALESCE(e.created_at, e.imported_at))
+                    FROM memory_write_job_evidence AS linked
+                    JOIN evidence AS e ON e.evidence_id = linked.evidence_id
+                    WHERE linked.job_id = memory_write_jobs.job_id
+                ),
+                created_at
+            )
+            WHERE event_at IS NULL
+            """
+        )
+
     @staticmethod
     def _backfill_legacy_references(connection: sqlite3.Connection) -> None:
         """Populate the reference table after its schema exists."""
@@ -610,8 +690,20 @@ class DurableWriteQueue:
         replacing the earlier intent.
         """
         self._refuse_if_below_reserve()
-        payload = json.dumps(request.model_dump(mode="json"), sort_keys=True)
+        request = self._freeze_request_inputs(request)
+        payload = json.dumps(
+            request.model_dump(
+                mode="json",
+                exclude={"event_at", "time_quality"},
+            ),
+            sort_keys=True,
+        )
         now = datetime.now(UTC).isoformat()
+        resolved_event, resolved_quality = self._event_time_for_request(request)
+        event = request.event_at or resolved_event or datetime.fromisoformat(now)
+        if event.tzinfo is None:
+            event = event.replace(tzinfo=UTC)
+        event_at = event.astimezone(UTC).isoformat()
         key = request.request_key
         with (
             main_database(self.database_path).write(row_factory=sqlite3.Row) as connection,
@@ -624,16 +716,22 @@ class DurableWriteQueue:
                 (key,),
             ).fetchone()
             if existing is not None:
-                # A replay of work that already finished is a hit, not a
-                # conflict. Terminal rows store a reaped digest instead of the
-                # original payload, and the caller's key is derived from the
-                # evidence's revision, which can move on between replays while
-                # the payload's evidence ids stay the same. Refusing that pair
-                # turned every startup replay of one settled job into an
-                # exception: 236 of them in a single afternoon, all one key.
                 already_settled = (
                     str(existing["state"]) == COMPLETED and int(existing["hy_invoked"]) == 1
                 )
+                stored_payload = str(existing["payload_json"])
+                priority_replay = _matches_payload_except_priority(stored_payload, payload)
+                legacy_reaped_match = bool(
+                    already_settled
+                    and _matches_reaped_legacy_payload(stored_payload, payload)
+                )
+                if (
+                    not _matches_stored_payload(stored_payload, payload)
+                    and not priority_replay
+                    and not _matches_legacy_reference_replay(stored_payload, payload)
+                    and not legacy_reaped_match
+                ):
+                    raise ValueError(f"idempotency key {key} already exists with different content")
                 if already_settled:
                     return DurableWriteRecord(
                         job_id=str(existing["job_id"]),
@@ -644,14 +742,6 @@ class DurableWriteQueue:
                         updated_at=str(existing["updated_at"]),
                         detail="already completed; replay treated as a hit",
                     )
-                priority_replay = _matches_payload_except_priority(
-                    str(existing["payload_json"]), payload
-                )
-                if (
-                    not _matches_stored_payload(str(existing["payload_json"]), payload)
-                    and not priority_replay
-                ):
-                    raise ValueError(f"idempotency key {key} already exists with different content")
                 if priority_replay and request.priority == "live":
                     connection.execute(
                         "UPDATE memory_write_jobs SET priority='live', updated_at=? "
@@ -711,11 +801,13 @@ class DurableWriteQueue:
             )
             connection.execute(
                 "INSERT INTO memory_write_jobs ("
-                " job_id, queue_seq, session_key, source_key, batch_id, idempotency_key, payload_json,"
+                " job_id, queue_seq, session_key, source_key, batch_id,"
+                " idempotency_key, payload_json,"
                 " evidence_ids_json,"
                 " state, priority, attempts, max_attempts, lease_owner,"
-                " lease_expires_at, hy_invoked, last_error, created_at, updated_at"
-                ") VALUES (?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?)",
+                " lease_expires_at, hy_invoked, last_error, event_at, time_quality,"
+                " created_at, updated_at"
+                ") VALUES (?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?)",
                 (
                     job_id,
                     queue_seq,
@@ -733,6 +825,8 @@ class DurableWriteQueue:
                     None,
                     0,
                     None,
+                    event_at,
+                    request.time_quality if request.event_at is not None else resolved_quality,
                     now,
                     now,
                 ),
@@ -741,6 +835,18 @@ class DurableWriteQueue:
                 "INSERT INTO memory_write_job_evidence (job_id, evidence_id) VALUES (?,?)",
                 ((job_id, str(evidence_id)) for evidence_id in request.evidence_ids),
             )
+        diagnostic_event(
+            level="info",
+            component="memory.write_queue",
+            event="memory_write_job.queued",
+            outcome="queued",
+            job_id=job_id,
+            details={
+                "source_id": request.source_id or request.internal_session_id,
+                "mode": request.priority,
+                "count": len(request.evidence_ids),
+            },
+        )
         return DurableWriteRecord(
             job_id=job_id,
             request_id=key,
@@ -750,6 +856,45 @@ class DurableWriteQueue:
             updated_at=now,
         )
 
+    def _freeze_request_inputs(self, request: WriteRequest) -> WriteRequest:
+        if request.input_records or self.evidence is None:
+            return request
+        records: list[EvidenceRecord] = []
+        for evidence_id in request.evidence_ids:
+            record = self.evidence.get(evidence_id)
+            if record is None:
+                return request
+            records.append(record)
+        return request.model_copy(update={"input_records": tuple(records)})
+
+    def _event_time_for_request(self, request: WriteRequest) -> tuple[datetime | None, str]:
+        records = list(request.input_records)
+        if not records and self.evidence is not None:
+            records = [
+                record
+                for evidence_id in request.evidence_ids
+                if (record := self.evidence.get(evidence_id)) is not None
+            ]
+        if not records:
+            return None, "approximate"
+        opener = next(
+            (record for record in records if record.role == "user"),
+            records[0],
+        )
+        if opener.created_at is None:
+            return opener.imported_at, "approximate"
+        origin = opener.origin_meta.get(EvidenceOrigin.FILE)
+        basis = origin.raw_meta.get("time_basis") if origin is not None else None
+        turn_event_at = origin.raw_meta.get("turn_event_at") if origin is not None else None
+        if isinstance(turn_event_at, str):
+            try:
+                return datetime.fromisoformat(turn_event_at), "exact"
+            except ValueError:
+                pass
+        if basis == "file_mtime":
+            return opener.created_at, "session"
+        return opener.created_at, "exact"
+
     def repair_pre_send_identity_failures(self, *, limit: int = 128) -> int:
         """Refresh stale Evidence journals and requeue only provably unsent jobs."""
         if self.evidence is None:
@@ -859,11 +1004,55 @@ class DurableWriteQueue:
                 raise ReconcileSchemaUnavailableError(
                     "evidence table is unavailable; run migrations before reconcile"
                 )
-            where = (
-                "WHERE evidence.imported_at IS NOT NULL "
-                "AND NOT EXISTS (SELECT 1 FROM memory_write_job_evidence ref "
-                "WHERE ref.evidence_id=evidence.evidence_id)"
+            has_versions = (
+                connection.execute(
+                    "SELECT 1 FROM sqlite_master "
+                    "WHERE type='table' AND name='evidence_versions'"
+                ).fetchone()
+                is not None
             )
+            evidence_columns = {
+                str(row[1])
+                for row in connection.execute("PRAGMA table_info(evidence)").fetchall()
+            }
+            if {"revision", "content_hash", "updated_at"} <= evidence_columns:
+                revision_time = (
+                    "COALESCE((SELECT version.changed_at FROM evidence_versions version "
+                    "WHERE version.evidence_id=evidence.evidence_id "
+                    "AND version.revision=evidence.revision),"
+                    "evidence.updated_at,evidence.imported_at)"
+                    if has_versions
+                    else "COALESCE(evidence.updated_at,evidence.imported_at)"
+                )
+                where = (
+                    "WHERE evidence.imported_at IS NOT NULL "
+                    "AND NOT EXISTS ("
+                    "SELECT 1 FROM memory_write_job_evidence linked "
+                    "JOIN memory_write_jobs job ON job.job_id=linked.job_id, "
+                    "json_each(job.payload_json,'$.input_records') frozen "
+                    "WHERE linked.evidence_id=evidence.evidence_id "
+                    "AND json_extract(frozen.value,'$.evidence_id')=evidence.evidence_id "
+                    "AND json_extract(frozen.value,'$.revision')=evidence.revision "
+                    "AND json_extract(frozen.value,'$.content_hash')=evidence.content_hash"
+                    ") "
+                    # Pre-snapshot builds only recorded an Evidence id. Treat
+                    # that as coverage only when its Job was created after this
+                    # exact revision.
+                    "AND NOT EXISTS ("
+                    "SELECT 1 FROM memory_write_job_evidence legacy "
+                    "JOIN memory_write_jobs legacy_job ON legacy_job.job_id=legacy.job_id "
+                    "WHERE legacy.evidence_id=evidence.evidence_id "
+                    "AND COALESCE(json_array_length("
+                    "json_extract(legacy_job.payload_json,'$.input_records')),0)=0 "
+                    f"AND legacy_job.created_at>=({revision_time}))"
+                )
+            else:
+                # A genuinely old/minimal schema has no revision identity.
+                where = (
+                    "WHERE evidence.imported_at IS NOT NULL "
+                    "AND NOT EXISTS (SELECT 1 FROM memory_write_job_evidence legacy "
+                    "WHERE legacy.evidence_id=evidence.evidence_id)"
+                )
             params: tuple[Any, ...] = ()
             cursor_clause = ""
             if after_cursor is not None:
@@ -1095,6 +1284,7 @@ class DurableWriteQueue:
         with main_database(self.database_path).write() as connection, connection:
             updated = connection.execute(
                 "UPDATE memory_write_jobs SET state=?,lease_owner=NULL,"
+                "attempts=CASE WHEN attempts>0 THEN attempts-1 ELSE 0 END,"
                 "lease_expires_at=NULL,updated_at=? WHERE " + where,
                 (PENDING, now, *params),
             )
@@ -1162,45 +1352,10 @@ class DurableWriteQueue:
         return state
 
     def clear_terminal_payloads(self, *, limit: int = 512) -> int:
-        """Replace terminal rows' request bodies with a content digest.
-
-        A completed/failed/dead/cancelled row can never run again, so the full
-        serialized `WriteRequest` is unbounded dead weight. A digest keeps the
-        idempotency contract intact: a replay of the same intent still resolves
-        to the same row, while the same key with different content still
-        conflicts. `unknown` is left alone because it is non-terminal and its
-        intent is needed for reconciliation.
-
-        Bounded by `limit` so a large backlog drains over several calls instead
-        of one long write on the single writer.
-        """
+        """Deprecated compatibility hook; durable inputs are immutable audit data."""
         if type(limit) is not int or limit < 1:
             raise ValueError("limit must be positive")
-        placeholders = ",".join("?" for _ in TERMINAL_PAYLOAD_STATES)
-        with (
-            main_database(self.database_path).write(row_factory=sqlite3.Row) as connection,
-            connection,
-        ):
-            rows = connection.execute(
-                "SELECT job_id, payload_json FROM memory_write_jobs "
-                f"WHERE state IN ({placeholders}) "
-                "AND instr(payload_json, ?) = 0 "
-                "LIMIT ?",
-                (*TERMINAL_PAYLOAD_STATES, REAPED_PAYLOAD_MARKER, limit),
-            ).fetchall()
-            if not rows:
-                return 0
-            for row in rows:
-                connection.execute(
-                    "UPDATE memory_write_jobs SET payload_json=? "
-                    f"WHERE job_id=? AND state IN ({placeholders})",
-                    (
-                        _reaped_payload(_payload_digest(str(row["payload_json"]))),
-                        str(row["job_id"]),
-                        *TERMINAL_PAYLOAD_STATES,
-                    ),
-                )
-        return len(rows)
+        return 0
 
     def readmit_parked(self, *, states: tuple[str, ...] = PARKED_STATES) -> int:
         """Move parked jobs back to `pending` for a fresh claim.
@@ -1250,11 +1405,12 @@ class DurableWriteQueue:
     ) -> int:
         """Re-admit `waiting_for_model` jobs whose wait has elapsed.
 
-        A model outage clears on its own, so this state is retried on a timer
+        A model outage clears on its own, so transient waiters retry on a timer
         rather than waiting for a grant event like `waiting_for_permission`.
-        Keyed on `updated_at` so repeated calls while a job is still parked do
-        not reset its own clock; the attempt consumed by the parking claim is
-        returned, the same way `readmit_parked` does.
+        `model_profile_not_active` is excluded: only explicit activation can
+        change that condition. Keyed on `updated_at` so repeated calls while a
+        job is still parked do not reset its own clock; the attempt consumed by
+        the parking claim is returned, the same way `readmit_parked` does.
         """
         now = now or datetime.now(UTC)
         threshold = (now - after).isoformat()
@@ -1263,8 +1419,33 @@ class DurableWriteQueue:
                 "UPDATE memory_write_jobs SET state=?, lease_owner=NULL, "
                 "lease_expires_at=NULL, attempts=CASE WHEN attempts>0 "
                 "THEN attempts-1 ELSE 0 END, last_error=NULL, updated_at=? "
-                "WHERE state=? AND updated_at <= ? AND hy_invoked=0",
-                (PENDING, now.isoformat(), WAITING_FOR_MODEL, threshold),
+                "WHERE state=? AND updated_at <= ? AND hy_invoked=0 "
+                "AND (last_error IS NULL OR last_error<>?)",
+                (
+                    PENDING,
+                    now.isoformat(),
+                    WAITING_FOR_MODEL,
+                    threshold,
+                    "model_profile_not_active",
+                ),
+            )
+        return changed.rowcount
+
+    def readmit_model_profile_waiters(self) -> int:
+        """Re-admit no-profile work only after model activation changes the condition."""
+        now = datetime.now(UTC).isoformat()
+        with main_database(self.database_path).write() as connection, connection:
+            changed = connection.execute(
+                "UPDATE memory_write_jobs SET state=?, lease_owner=NULL, "
+                "lease_expires_at=NULL, attempts=CASE WHEN attempts>0 "
+                "THEN attempts-1 ELSE 0 END, last_error=NULL, updated_at=? "
+                "WHERE state=? AND last_error=? AND hy_invoked=0",
+                (
+                    PENDING,
+                    now,
+                    WAITING_FOR_MODEL,
+                    "model_profile_not_active",
+                ),
             )
         return changed.rowcount
 
@@ -1362,9 +1543,13 @@ class DurableWriteQueue:
             connection,
         ):
             row = connection.execute(
-                "SELECT state FROM memory_write_jobs WHERE job_id=?", (job_id,)
+                "SELECT state,last_error FROM memory_write_jobs WHERE job_id=?", (job_id,)
             ).fetchone()
-            if row is None or str(row["state"]) != UNKNOWN:
+            if (
+                row is None
+                or str(row["state"]) != UNKNOWN
+                or str(row["last_error"]) in NON_RETRYABLE_POST_WRITE_CODES
+            ):
                 return 0
             connection.execute(
                 """
@@ -1402,7 +1587,13 @@ class DurableWriteQueue:
         cross back to pending, so this cannot duplicate a possibly-sent write.
         """
         now = datetime.now(UTC).isoformat()
-        payload = json.dumps(request.model_dump(mode="json"), sort_keys=True)
+        payload = json.dumps(
+            request.model_dump(
+                mode="json",
+                exclude={"event_at", "time_quality"},
+            ),
+            sort_keys=True,
+        )
         with (
             main_database(self.database_path).write(row_factory=sqlite3.Row) as connection,
             connection,
@@ -1492,6 +1683,66 @@ class DurableWriteQueue:
                 )
         return len(job_ids)
 
+    def retry_first_unknown_after_reconcile(
+        self,
+        *,
+        older_than: datetime,
+        limit: int = 32,
+    ) -> int:
+        """Retry one unresolved unknown outcome once, after reconciliation.
+
+        The caller first asks HY and checks linked governed output. Rows still
+        unknown may have lost the response after the SDK call, so one bounded
+        retry is useful. `attempts=1` is the guard: the retry claim increments
+        it to two, and a second unknown is left for explicit user recovery.
+        """
+        if type(limit) is not int or limit < 1:
+            raise ValueError("limit must be positive")
+        now = datetime.now(UTC).isoformat()
+        with (
+            main_database(self.database_path).write(row_factory=sqlite3.Row) as connection,
+            connection,
+        ):
+            rows = connection.execute(
+                "SELECT job_id FROM memory_write_jobs "
+                "WHERE state=? AND attempts=1 AND attempts < max_attempts "
+                f"AND COALESCE(last_error, '') NOT IN "
+                f"({','.join('?' for _ in NON_RETRYABLE_POST_WRITE_CODES)}) "
+                "AND updated_at <= ? ORDER BY updated_at, job_id LIMIT ?",
+                (
+                    UNKNOWN,
+                    *NON_RETRYABLE_POST_WRITE_CODES,
+                    older_than.isoformat(),
+                    limit,
+                ),
+            ).fetchall()
+            job_ids = tuple(str(row["job_id"]) for row in rows)
+            if not job_ids:
+                return 0
+            placeholders = ",".join("?" for _ in job_ids)
+            connection.execute(
+                f"UPDATE memory_write_jobs SET state=?, lease_owner=NULL, "
+                f"lease_expires_at=NULL, hy_invoked=0, "
+                f"last_error='model_content_unknown_auto_retry', updated_at=? "
+                f"WHERE state=? AND attempts=1 AND job_id IN ({placeholders})",
+                (PENDING, now, UNKNOWN, *job_ids),
+            )
+            if connection.execute(
+                "SELECT 1 FROM sqlite_master WHERE type='table' AND name='memory_tasks'"
+            ).fetchone():
+                connection.execute(
+                    f"UPDATE memory_tasks SET status=?, runner_token=NULL, "
+                    f"last_error_code='model_content_unknown_auto_retry', updated_at=? "
+                    f"WHERE status=? AND task_id IN ("
+                    f"SELECT inputs.task_id FROM memory_task_inputs AS inputs "
+                    f"JOIN memory_write_job_evidence AS linked ON EXISTS ("
+                    f"SELECT 1 FROM json_each(inputs.refs_json) AS ref "
+                    f"WHERE json_extract(ref.value, '$.evidence_id')=linked.evidence_id) "
+                    f"WHERE linked.job_id IN ({placeholders}))",
+                    (MemoryTaskStatus.PENDING, now, MemoryTaskStatus.UNKNOWN, *job_ids),
+                )
+        return len(job_ids)
+
     def resolve_unknown_from_linked_output(self, job_id: str) -> bool:
         """Close an unknown durable row when local governance output proves success."""
         with main_database(self.database_path).read(row_factory=sqlite3.Row) as connection:
@@ -1527,6 +1778,7 @@ class DurableWriteQueue:
         *,
         limit: int = 128,
         after_cursor: tuple[str, str] | None = None,
+        exclude_error_codes: tuple[str, ...] = (),
     ) -> tuple[UnknownWriteJob, ...]:
         """List unknown rows that still need an inquiry, in stable order.
 
@@ -1543,6 +1795,9 @@ class DurableWriteQueue:
         """
         if type(limit) is not int or limit < 1:
             raise ValueError("limit must be positive")
+        excluded = tuple(dict.fromkeys(exclude_error_codes))
+        if not all(isinstance(code, str) and code for code in excluded):
+            raise ValueError("exclude_error_codes must contain non-empty strings")
         cursor_clause = ""
         params: tuple[Any, ...] = ()
         if after_cursor is not None:
@@ -1555,14 +1810,20 @@ class DurableWriteQueue:
             updated_at, job_id = after_cursor
             cursor_clause = "AND (updated_at, job_id) > (?, ?)"
             params = (updated_at, job_id)
+        if excluded:
+            placeholders = ",".join("?" for _ in excluded)
+            excluded_clause = f"AND COALESCE(last_error, '') NOT IN ({placeholders})"
+        else:
+            excluded_clause = ""
         with main_database(self.database_path).read(row_factory=sqlite3.Row) as connection:
             rows = connection.execute(
                 "SELECT job_id, idempotency_key, created_at, updated_at, last_error "
                 "FROM memory_write_jobs "
                 "WHERE state=? "
                 f"{cursor_clause} "
+                f"{excluded_clause} "
                 "ORDER BY updated_at, job_id LIMIT ?",
-                (UNKNOWN, *params, limit),
+                (UNKNOWN, *params, *excluded, limit),
             ).fetchall()
         return tuple(
             UnknownWriteJob(
@@ -1688,9 +1949,12 @@ class DurableWriteRegistry:
 def evidence_for_request(store: EvidenceStore, request: WriteRequest) -> tuple[EvidenceRecord, ...]:
     """The Evidence a stored write points at, in the request's own order.
 
-    A write may only point at Evidence that still exists; a missing id is a
-    rejection rather than a silent partial send.
+    Current jobs are self-contained: their immutable records are the model
+    input even if a Harness later rewinds or edits its live transcript. The
+    store lookup remains only for payloads written by pre-upgrade builds.
     """
+    if request.input_records:
+        return request.input_records
     records = []
     for evidence_id in request.evidence_ids:
         record = store.get(evidence_id)
@@ -1797,7 +2061,14 @@ async def execute_job(
         # 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`.
         code = safe_error_code(error)
-        if code == "model_content_source_disabled" and not hy_invoked:
+        if (
+            code
+            in {
+                "model_content_source_disabled",
+                "model_sedimentation_disabled",
+            }
+            and not hy_invoked
+        ):
             await in_the_background(queue.release_unstarted, job_id, owner=owner)
             return
         # The task layer classifies these before any bridge is opened, so HY
@@ -1808,7 +2079,11 @@ async def execute_job(
         # - transient model/provider trouble -> waiting_for_model (parked).
         # Everything else may have crossed into HY, and HY has no idempotency
         # key, so it is `unknown` rather than a blind retry.
-        if hy_invoked:
+        if code in PRE_SEND_REFUSAL_CODES:
+            # The frozen input window is checked before the payload reaches
+            # HY. This one refusal therefore has a proven pre-send boundary.
+            state, invoked_marker = FAILED, False
+        elif hy_invoked:
             # Error names do not prove execution phase: permission and
             # evidence rechecks also run after HY returns. Once the durable
             # marker is set, every failure is possibly-sent and must stay out
@@ -1820,10 +2095,6 @@ async def execute_job(
             state, invoked_marker = WAITING_FOR_PERMISSION, False
         elif code in MODEL_NOT_READY_CODES:
             state, invoked_marker = WAITING_FOR_MODEL, False
-        elif code in PRE_SEND_REFUSAL_CODES:
-            # Provably before the bridge: the frozen input is stale or too
-            # large. Nothing was sent, so this is terminal rather than unknown.
-            state, invoked_marker = FAILED, False
         elif code == "model_content_transport_request_failed":
             # The generation or embedding HTTP exchange can fail after HY has
             # already persisted a node. Treating that as a pre-send failure
@@ -1848,7 +2119,12 @@ async def execute_job(
         # Never `logger.exception` here: it prints the exception's message,
         # which may be a provider response or the user's own text. Log the
         # already-sanitised code instead, matching `hy-failure-classification`.
-        logger.error("durable_write_job_failed job_id=%s error_code=%s", job_id, code)
+        if code == "model_profile_not_active":
+            logger.debug("durable_write_job_parked job_id=%s error_code=%s", job_id, code)
+        elif state in PARKED_STATES:
+            logger.info("durable_write_job_parked job_id=%s error_code=%s", job_id, code)
+        else:
+            logger.error("durable_write_job_failed job_id=%s error_code=%s", job_id, code)
         return
     await in_the_background(queue.settle, job_id, state=COMPLETED, owner=owner)
 
@@ -1896,6 +2172,13 @@ class DurableWriteDrain:
         #: Reaping is a write; it runs at most once per this interval on the
         #: idle path so an empty queue does not issue one write per poll.
         self._reap_interval = reap_interval
+        # Unknown recovery is user-visible and lightweight enough to check
+        # sooner than general database housekeeping. Tests that set the common
+        # interval to zero keep immediate maintenance.
+        self._unknown_reconcile_interval = min(
+            reap_interval,
+            timedelta(seconds=10),
+        )
         #: The maintenance writes below are recoverability/retention work, not
         #: idle housekeeping, so each has its own interval and runs at the start
         #: of every drain tick. Sharing one timestamp would let a busy tick
@@ -1905,7 +2188,6 @@ class DurableWriteDrain:
         #: sustained backlog must not keep finished rows' request bodies in the
         #: business database indefinitely. Its own gate so neither interval can
         #: starve the other.
-        self._last_terminal_reap: datetime | None = None
         #: `waiting_for_model` re-admission is also time-based recoverability:
         #: a model outage clears on its own, and a queue that never goes idle
         #: must still re-admit the write once the wait has elapsed.
@@ -1956,7 +2238,7 @@ class DurableWriteDrain:
         now: datetime | None = None,
         budget_ms: int | None = None,
     ) -> tuple[str, ...]:
-        """Claim at most one FIFO head per enabled Harness."""
+        """Claim event-time heads, with one running job per enabled Harness."""
         if type(limit) is not int or limit < 1:
             raise ValueError("limit must be positive")
         now = now or datetime.now(UTC)
@@ -1998,9 +2280,7 @@ class DurableWriteDrain:
                     (PENDING,),
                 ).fetchall()
             enabled_sources = {
-                str(row[0])
-                for row in source_rows
-                if self.source_writes_permitted(str(row[0]))
+                str(row[0]) for row in source_rows if self.source_writes_permitted(str(row[0]))
             }
             if not enabled_sources:
                 return ()
@@ -2019,38 +2299,142 @@ class DurableWriteDrain:
                 placeholders = ",".join("?" for _ in enabled_sources)
                 source_clause = f"AND candidate.source_key IN ({placeholders}) "
                 source_params = tuple(sorted(enabled_sources))
+            window = connection.execute(
+                "SELECT window_anchor,window_end FROM memory_write_schedule_state WHERE singleton=1"
+            ).fetchone()
+            if window is not None:
+                anchor_text, end_text = str(window[0]), str(window[1])
+                active_source_clause = ""
+                active_source_params: tuple[Any, ...] = ()
+                if enabled_sources is not None:
+                    placeholders = ",".join("?" for _ in enabled_sources)
+                    active_source_clause = f"AND active.source_key IN ({placeholders}) "
+                    active_source_params = tuple(sorted(enabled_sources))
+                # The contract retires a window when every *eligible*
+                # pending/running job in it has settled or parked. A pending row
+                # that scheduling already holds back (an unexpired lease, a
+                # busy Harness, or an earlier event in its Session) is not eligible, so it
+                # must not keep the anchor alive: those blockers usually sit
+                # outside this ten-minute range and can never be claimed from
+                # inside it, which pinned the anchor and starved the whole
+                # queue. Reuse the claim's own predicate so "the window is
+                # busy" and "the window can still be worked" agree.
+                active = connection.execute(
+                    "SELECT 1 FROM memory_write_jobs AS active "
+                    "WHERE active.event_at>=? AND active.event_at<? AND ("
+                    "active.state=? OR (active.state=? "
+                    "AND (active.lease_expires_at IS NULL "
+                    "OR active.lease_expires_at < ?) " + active_source_clause + "AND NOT EXISTS ("
+                    "SELECT 1 FROM memory_write_jobs AS held "
+                    "WHERE held.source_key=active.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=active.session_key "
+                    f"AND earlier.state IN ({blocking}) "
+                    "AND (earlier.event_at<active.event_at OR "
+                    "(earlier.event_at=active.event_at "
+                    "AND earlier.queue_seq<active.queue_seq))) "
+                    ")) LIMIT 1",
+                    (
+                        anchor_text,
+                        end_text,
+                        RUNNING,
+                        PENDING,
+                        now.isoformat(),
+                        *active_source_params,
+                        RUNNING,
+                        now.isoformat(),
+                        *_FIFO_BLOCKING_STATES,
+                    ),
+                ).fetchone()
+                if active is None:
+                    connection.execute("DELETE FROM memory_write_schedule_state WHERE singleton=1")
+                    window = None
+            if window is None:
+                anchor = connection.execute(
+                    "SELECT candidate.event_at "
+                    "FROM memory_write_jobs AS candidate "
+                    "WHERE candidate.state=? "
+                    "AND (candidate.lease_expires_at IS NULL "
+                    "OR candidate.lease_expires_at < ?) " + source_clause + "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 LIMIT 1",
+                    (
+                        PENDING,
+                        now.isoformat(),
+                        *source_params,
+                        RUNNING,
+                        now.isoformat(),
+                        *_FIFO_BLOCKING_STATES,
+                    ),
+                ).fetchone()
+                if anchor is None:
+                    return ()
+                anchor_at = datetime.fromisoformat(str(anchor[0]))
+                if anchor_at.tzinfo is None:
+                    anchor_at = anchor_at.replace(tzinfo=UTC)
+                anchor_text = anchor_at.astimezone(UTC).isoformat()
+                end_text = (anchor_at + EVENT_TIME_WINDOW).astimezone(UTC).isoformat()
+                connection.execute(
+                    "INSERT INTO memory_write_schedule_state "
+                    "(singleton,window_anchor,window_end,updated_at) "
+                    "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 +
+                + 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.source_key=candidate.source_key "
+                "WHERE earlier.session_key=candidate.session_key "
                 f"AND earlier.state IN ({blocking}) "
-                "AND earlier.queue_seq<candidate.queue_seq) "
-                "ORDER BY candidate.queue_seq",
+                "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:
+                source_key = str(row["source_key"])
+                if source_key in selected_sources:
+                    continue
                 if self.source_claim_enabled is not None and not self.source_claim_enabled(
-                    str(row["source_key"])
+                    source_key
                 ):
                     continue
                 selected.append(str(row["job_id"]))
+                selected_sources.add(source_key)
                 if len(selected) == min(limit, self.max_concurrent_sources):
                     break
             if not selected:
@@ -2279,19 +2663,6 @@ class DurableWriteDrain:
                         self.queue.reclaim_expired,
                         operation_name="reclaim_expired",
                     )
-                # Terminal-payload reaping is a bounded retention write: a
-                # finished row's full request body must not sit in the business
-                # database forever just because the queue never went idle. Its
-                # own interval keeps it off every poll.
-                if (
-                    self._last_terminal_reap is None
-                    or tick_started - self._last_terminal_reap >= self._reap_interval
-                ):
-                    self._last_terminal_reap = tick_started
-                    await self._run_idle_maintenance(
-                        self.queue.clear_terminal_payloads,
-                        operation_name="clear_terminal_payloads",
-                    )
                 # A model outage clears on its own, so `waiting_for_model` jobs
                 # are re-admitted on a timer. That must not depend on idleness: a
                 # queue busy for longer than the wait window would otherwise keep
@@ -2308,7 +2679,8 @@ class DurableWriteDrain:
                     )
                 if self.reconcile_unknown is not None and (
                     self._last_unknown_reconcile is None
-                    or tick_started - self._last_unknown_reconcile >= self._reap_interval
+                    or tick_started - self._last_unknown_reconcile
+                    >= self._unknown_reconcile_interval
                 ):
                     self._last_unknown_reconcile = tick_started
                     try:

memhub/apps/core/src/one_memory/memory_server/hy_capabilities.py

@@ -3,7 +3,7 @@
 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 was verified against `hy-memory==2.0.2`:
+What was verified against `hy-memory==2.0.4`:
 
 - `HyMemoryClient.add(..., request_id=...)` accepts a caller-supplied id and
   echoes it back. It is described as a request/receipt handle, and the SDK

memhub/apps/core/src/one_memory/memory_server/migrations.py

@@ -151,6 +151,8 @@ def memory_write_jobs_ddl(dialect: str = "mysql") -> str:
             lease_expires_at TEXT,
             hy_invoked INTEGER NOT NULL DEFAULT 0,
             last_error TEXT,
+            event_at TEXT NOT NULL DEFAULT CURRENT_TIMESTAMP,
+            time_quality TEXT NOT NULL DEFAULT 'approximate',
             created_at TEXT NOT NULL,
             updated_at TEXT NOT NULL
         );
@@ -164,6 +166,8 @@ def memory_write_jobs_ddl(dialect: str = "mysql") -> str:
         ON memory_write_jobs(source_key, state, queue_seq);
         CREATE INDEX IF NOT EXISTS idx_write_jobs_fifo
         ON memory_write_jobs(queue_seq);
+        CREATE INDEX IF NOT EXISTS idx_write_jobs_event_claim
+        ON memory_write_jobs(state, event_at, queue_seq);
         CREATE INDEX IF NOT EXISTS idx_write_jobs_source_recent
         ON memory_write_jobs(source_key, updated_at DESC);
         CREATE TABLE IF NOT EXISTS memory_write_job_evidence (
@@ -172,7 +176,13 @@ def memory_write_jobs_ddl(dialect: str = "mysql") -> str:
             PRIMARY KEY (job_id, evidence_id)
         );
         CREATE INDEX IF NOT EXISTS idx_write_job_evidence_evidence
-        ON memory_write_job_evidence(evidence_id)
+        ON memory_write_job_evidence(evidence_id);
+        CREATE TABLE IF NOT EXISTS memory_write_schedule_state (
+            singleton INTEGER PRIMARY KEY CHECK (singleton = 1),
+            window_anchor TEXT NOT NULL,
+            window_end TEXT NOT NULL,
+            updated_at TEXT NOT NULL
+        )
         """
     return """
         CREATE TABLE IF NOT EXISTS memory_write_jobs (
@@ -192,6 +202,8 @@ def memory_write_jobs_ddl(dialect: str = "mysql") -> str:
             lease_expires_at DATETIME(6) NULL,
             hy_invoked TINYINT(1) NOT NULL DEFAULT 0,
             last_error TEXT NULL,
+            event_at DATETIME(6) NOT NULL DEFAULT CURRENT_TIMESTAMP(6),
+            time_quality VARCHAR(32) NOT NULL DEFAULT 'approximate',
             created_at DATETIME(6) NOT NULL DEFAULT CURRENT_TIMESTAMP(6),
             updated_at DATETIME(6) NOT NULL DEFAULT CURRENT_TIMESTAMP(6)
                 ON UPDATE CURRENT_TIMESTAMP(6),
@@ -201,6 +213,7 @@ def memory_write_jobs_ddl(dialect: str = "mysql") -> str:
             KEY idx_write_jobs_source_claim (state, source_key, priority, created_at),
             KEY idx_write_jobs_source_fifo (source_key, state, queue_seq),
             KEY idx_write_jobs_fifo (queue_seq),
+            KEY idx_write_jobs_event_claim (state, event_at, queue_seq),
             KEY idx_write_jobs_source_recent (source_key, updated_at)
         ) ENGINE=InnoDB DEFAULT CHARSET=utf8mb4;
 
@@ -209,6 +222,13 @@ def memory_write_jobs_ddl(dialect: str = "mysql") -> str:
             evidence_id VARCHAR(128) NOT NULL,
             PRIMARY KEY (job_id, evidence_id),
             KEY idx_write_job_evidence_evidence (evidence_id)
+        ) ENGINE=InnoDB DEFAULT CHARSET=utf8mb4;
+
+        CREATE TABLE IF NOT EXISTS memory_write_schedule_state (
+            singleton INT PRIMARY KEY CHECK (singleton = 1),
+            window_anchor DATETIME(6) NOT NULL,
+            window_end DATETIME(6) NOT NULL,
+            updated_at DATETIME(6) NOT NULL
         ) ENGINE=InnoDB DEFAULT CHARSET=utf8mb4
         """
 
@@ -1231,6 +1251,68 @@ CREATE TABLE IF NOT EXISTS model_request_budget_events (
 """
 
 
+def model_request_audit_ddl(dialect: str = "mysql") -> str:
+    """Content-free accounting for physical precipitation model requests."""
+    if dialect == "sqlite":
+        return """
+CREATE TABLE IF NOT EXISTS model_request_audit (
+    audit_id TEXT PRIMARY KEY,
+    request_id TEXT NOT NULL,
+    profile_id TEXT NOT NULL,
+    profile_revision INTEGER NOT NULL,
+    provider_kind TEXT NOT NULL,
+    model TEXT NOT NULL,
+    request_kind TEXT NOT NULL,
+    prompt_stage TEXT NOT NULL DEFAULT 'other',
+    started_at TEXT NOT NULL,
+    completed_at TEXT,
+    outcome TEXT NOT NULL CHECK(outcome IN ('in_flight','succeeded','failed')),
+    http_status INTEGER,
+    error_type TEXT,
+    provider_request_id TEXT,
+    prompt_tokens INTEGER,
+    completion_tokens INTEGER,
+    total_tokens INTEGER,
+    cached_prompt_tokens INTEGER,
+    reasoning_tokens INTEGER
+);
+CREATE INDEX IF NOT EXISTS model_request_audit_request
+ON model_request_audit(request_id, started_at);
+CREATE INDEX IF NOT EXISTS model_request_audit_started
+ON model_request_audit(started_at);
+"""
+    return """
+CREATE TABLE IF NOT EXISTS model_request_audit (
+    audit_id VARCHAR(128) PRIMARY KEY,
+    request_id VARCHAR(512) NOT NULL,
+    profile_id VARCHAR(128) NOT NULL,
+    profile_revision INT NOT NULL,
+    provider_kind VARCHAR(64) NOT NULL,
+    model VARCHAR(512) NOT NULL,
+    request_kind VARCHAR(64) NOT NULL,
+    prompt_stage VARCHAR(128) NOT NULL DEFAULT 'other',
+    started_at DATETIME(6) NOT NULL,
+    completed_at DATETIME(6) NULL,
+    outcome VARCHAR(32) NOT NULL CHECK(outcome IN ('in_flight','succeeded','failed')),
+    http_status INT NULL,
+    error_type VARCHAR(128) NULL,
+    provider_request_id VARCHAR(512) NULL,
+    prompt_tokens BIGINT NULL,
+    completion_tokens BIGINT NULL,
+    total_tokens BIGINT NULL,
+    cached_prompt_tokens BIGINT NULL,
+    reasoning_tokens BIGINT NULL,
+    KEY model_request_audit_request (request_id, started_at),
+    KEY model_request_audit_started (started_at)
+) ENGINE=InnoDB DEFAULT CHARSET=utf8mb4
+"""
+
+
+async def _052_model_request_audit(cursor: Any) -> None:
+    for statement in _split_statements(model_request_audit_ddl(_dialect(cursor))):
+        await cursor.execute(statement)
+
+
 def model_probe_operations_ddl(dialect: str = "mysql") -> str:
     """Probe operations; `state` is CHECKed to a known set.
 
@@ -1414,6 +1496,97 @@ async def _051_drop_hy_node_permissions(cursor: Any) -> None:
     await cursor.execute("DROP TABLE IF EXISTS hy_node_permissions")
 
 
+async def _053_memory_write_event_time(cursor: Any) -> None:
+    """Add event-time scheduling without re-keying any existing write."""
+
+    dialect = _dialect(cursor)
+    columns = await cursor.columns("memory_write_jobs")
+    added_event_at = "event_at" not in columns
+    if added_event_at:
+        await cursor.execute(
+            "ALTER TABLE memory_write_jobs ADD COLUMN event_at "
+            + ("TEXT" if dialect == "sqlite" else "DATETIME(6) NULL")
+        )
+    if "time_quality" not in columns:
+        await cursor.execute(
+            "ALTER TABLE memory_write_jobs ADD COLUMN time_quality "
+            + (
+                "TEXT NOT NULL DEFAULT 'approximate'"
+                if dialect == "sqlite"
+                else "VARCHAR(32) NOT NULL DEFAULT 'approximate'"
+            )
+        )
+    await cursor.execute(
+        """
+        UPDATE memory_write_jobs
+        SET event_at = COALESCE(
+            (
+                SELECT MAX(COALESCE(e.created_at, e.imported_at))
+                FROM memory_write_job_evidence AS linked
+                JOIN evidence AS e ON e.evidence_id = linked.evidence_id
+                WHERE linked.job_id = memory_write_jobs.job_id
+                  AND e.role = 'user'
+            ),
+            (
+                SELECT MIN(COALESCE(e.created_at, e.imported_at))
+                FROM memory_write_job_evidence AS linked
+                JOIN evidence AS e ON e.evidence_id = linked.evidence_id
+                WHERE linked.job_id = memory_write_jobs.job_id
+            ),
+            created_at
+        )
+        WHERE event_at IS NULL
+        """
+    )
+    if dialect == "sqlite":
+        await cursor.execute(
+            "CREATE INDEX IF NOT EXISTS idx_write_jobs_event_claim "
+            "ON memory_write_jobs(state, event_at, queue_seq)"
+        )
+        await cursor.execute(
+            """
+            CREATE TABLE IF NOT EXISTS memory_write_schedule_state (
+                singleton INTEGER PRIMARY KEY CHECK (singleton = 1),
+                window_anchor TEXT NOT NULL,
+                window_end TEXT NOT NULL,
+                updated_at TEXT NOT NULL
+            )
+            """
+        )
+    else:
+        if added_event_at:
+            await cursor.execute(
+                "CREATE INDEX idx_write_jobs_event_claim "
+                "ON memory_write_jobs(state, event_at, queue_seq)"
+            )
+        await cursor.execute(
+            """
+            CREATE TABLE IF NOT EXISTS memory_write_schedule_state (
+                singleton INT PRIMARY KEY CHECK (singleton = 1),
+                window_anchor DATETIME(6) NOT NULL,
+                window_end DATETIME(6) NOT NULL,
+                updated_at DATETIME(6) NOT NULL
+            ) ENGINE=InnoDB DEFAULT CHARSET=utf8mb4
+            """
+        )
+
+
+async def _054_legacy_possibly_sent_tasks(cursor: Any) -> None:
+    """Preserve the old sent marker before removing its retired permission table."""
+    columns = await cursor.columns("memory_task_continuous_permissions")
+    if "task_id" not in columns:
+        return
+    await cursor.execute(
+        """
+        UPDATE memory_tasks
+        SET hy_invocation_count=CASE
+            WHEN hy_invocation_count<1 THEN 1 ELSE hy_invocation_count END
+        WHERE task_id IN (SELECT task_id FROM memory_task_continuous_permissions)
+        """
+    )
+    await cursor.execute("DROP TABLE memory_task_continuous_permissions")
+
+
 def model_profiles_ddl(dialect: str = "mysql") -> str:
     """Model profiles and their version history.
 
@@ -2680,7 +2853,11 @@ MIGRATIONS: tuple[Migration, ...] = (
         2,
         "memory_write_jobs",
         _002_write_jobs,
-        ("memory_write_jobs", "memory_write_job_evidence"),
+        (
+            "memory_write_jobs",
+            "memory_write_job_evidence",
+            "memory_write_schedule_state",
+        ),
     ),
     Migration(3, "source_file_cursors", _003_scanner_cursors, ("source_file_cursors",)),
     Migration(4, "change_outbox", _004_change_outbox, ("change_outbox",)),
@@ -2849,4 +3026,21 @@ MIGRATIONS: tuple[Migration, ...] = (
         _050_model_probe_operations_reserved_at,
     ),
     Migration(51, "drop_hy_node_permissions", _051_drop_hy_node_permissions),
+    Migration(
+        52,
+        "model_request_audit",
+        _052_model_request_audit,
+        ("model_request_audit",),
+    ),
+    Migration(
+        53,
+        "memory_write_event_time",
+        _053_memory_write_event_time,
+        ("memory_write_schedule_state",),
+    ),
+    Migration(
+        54,
+        "legacy_possibly_sent_tasks",
+        _054_legacy_possibly_sent_tasks,
+    ),
 )

memhub/apps/core/src/one_memory/memory_server/submit.py

@@ -107,6 +107,8 @@ def build_request(
     first_round: int,
     last_round: int,
     evidence_ids: list[str],
+    input_records: tuple[EvidenceRecord, ...] = (),
+    source_id: str | None = None,
     priority: Literal["live", "normal"] = "normal",
     idempotency_key: str | None = None,
 ) -> WriteRequest:
@@ -125,6 +127,8 @@ def build_request(
         internal_session_id=internal_session_id,
         batch_identity=batch,
         evidence_ids=evidence_ids,
+        input_records=input_records,
+        source_id=source_id,
         idempotency_key=idempotency_key or batch.batch_id,
         priority=priority,
     )
@@ -271,6 +275,8 @@ def build_request_from_evidence(
         first_round=first_round,
         last_round=last_round,
         evidence_ids=list(ordered),
+        input_records=tuple(ordered.values()),
+        source_id=seed.source_id,
         priority=priority,
         idempotency_key=idempotency_key,
     )

memhub/apps/core/src/one_memory/memory_server/unified_entry.py

@@ -18,10 +18,11 @@ from __future__ import annotations
 
 from collections.abc import Callable
 from dataclasses import dataclass
+from datetime import datetime
 from hashlib import sha256
 from typing import Literal
 
-from one_memory.imports.models import EvidenceRecord
+from one_memory.imports.models import EvidenceOrigin, EvidenceRecord
 from one_memory.memory_server.contracts import (
     BatchIdentity,
     WriteRequest,
@@ -124,6 +125,12 @@ class UnifiedWriteIntake:
         priority: WritePriority,
     ) -> WriteRequest:
         session_id = internal_session_id_for(record.conversation_id, record.source_id)
+        event_at, time_quality = _job_time((record,))
+        request_key = (
+            f"entry:{record.evidence_id}"
+            if record.revision == 1
+            else (f"entry:v2:{record.evidence_id}:{record.revision}:{record.content_hash}")
+        )
         return WriteRequest(
             internal_session_id=session_id,
             source_id=record.source_id,
@@ -133,8 +140,11 @@ class UnifiedWriteIntake:
                 last_round=last_round,
             ),
             evidence_ids=[record.evidence_id],
-            idempotency_key=f"entry:{record.evidence_id}",
+            input_records=(record,),
+            idempotency_key=request_key,
             priority=priority,
+            event_at=event_at,
+            time_quality=time_quality,
         )
 
     def submit_records(
@@ -161,6 +171,7 @@ class UnifiedWriteIntake:
         )
         request_id = f"turn:v1:{sha256(material.encode()).hexdigest()}"
         session_id = internal_session_id_for(first.conversation_id, first.source_id)
+        event_at, time_quality = _job_time(records)
         durable = self.queue.submit(
             WriteRequest(
                 internal_session_id=session_id,
@@ -171,8 +182,11 @@ class UnifiedWriteIntake:
                     last_round=1,
                 ),
                 evidence_ids=[record.evidence_id for record in records],
+                input_records=records,
                 idempotency_key=request_id,
                 priority=priority,
+                event_at=event_at,
+                time_quality=time_quality,
             )
         )
         return UnifiedWriteReceipt(
@@ -189,3 +203,29 @@ class SourceWriteDisabledError(RuntimeError):
     def __init__(self, evidence_id: str) -> None:
         super().__init__("model_content_source_disabled")
         self.evidence_id = evidence_id
+
+
+def _job_time(
+    records: tuple[EvidenceRecord, ...],
+) -> tuple[datetime, Literal["exact", "anchored", "session", "approximate"]]:
+    """Use the turn opener's wall clock, with an explicit fallback quality."""
+
+    opener = next(
+        (record for record in records if record.role == "user"),
+        records[0],
+    )
+    if opener.created_at is None:
+        return opener.imported_at, "approximate"
+    file_origin = opener.origin_meta.get(EvidenceOrigin.FILE)
+    basis = file_origin.raw_meta.get("time_basis") if file_origin is not None else None
+    turn_event_at = file_origin.raw_meta.get("turn_event_at") if file_origin is not None else None
+    if isinstance(turn_event_at, str):
+        try:
+            return datetime.fromisoformat(turn_event_at), "exact"
+        except ValueError:
+            pass
+    if basis == "file_mtime":
+        return opener.created_at, "session"
+    if EvidenceOrigin.HOOK in opener.origins:
+        return opener.created_at, "exact"
+    return opener.created_at, "exact"

memhub/apps/core/src/one_memory/memory_tasks.py

@@ -17,7 +17,7 @@ from pydantic import BaseModel, ConfigDict, Field
 
 from one_memory.database import take_the_write_lock_first
 from one_memory.database_access import main_database
-from one_memory.imports.models import EvidenceRecord
+from one_memory.imports.models import EvidenceRecord, WrittenBy
 from one_memory.memory_server.migrations import memory_tasks_ddl
 from one_memory.model_profiles import ModelProfile, ModelProfileSnapshot
 
@@ -67,6 +67,12 @@ POLICY_STOPPED_CODES = (
     "hy_memory_evidence_changed",
 )
 
+#: HY completed a write, but Core could not prove that the selected mutation
+#: target belonged to this task's frozen authority. Reconciliation may inspect
+#: HY by request id, but neither automatic nor user-confirmed generic retry may
+#: invoke the model again for these codes.
+NON_RETRYABLE_POST_WRITE_CODES = ("hy_memory_write_verification_authority_invalid",)
+
 
 class MemoryTask(BaseModel):
     model_config = ConfigDict(frozen=True, extra="forbid")
@@ -116,6 +122,24 @@ PROFILE_STOP_CODES = (
 ACTIVE_WINDOW_DEFER_LIMIT = 2048
 
 
+def _task_only_unknown_clause(connection: sqlite3.Connection) -> str:
+    """Exclude tasks whose Evidence is already owned by the durable queue."""
+    linked_table = connection.execute(
+        "SELECT 1 FROM sqlite_master WHERE type='table' AND name='memory_write_job_evidence'"
+    ).fetchone()
+    if linked_table is None:
+        return ""
+    return """
+        AND NOT EXISTS (
+            SELECT 1
+            FROM memory_task_inputs AS inputs, json_each(inputs.refs_json) AS ref
+            JOIN memory_write_job_evidence AS linked
+              ON linked.evidence_id=json_extract(ref.value, '$.evidence_id')
+            WHERE inputs.task_id=memory_tasks.task_id
+        )
+    """
+
+
 @dataclass(frozen=True, slots=True)
 class ActiveMemoryTask:
     """One task plus only the evidence fields the scheduler needs."""
@@ -174,29 +198,60 @@ def evidence_ref(record: EvidenceRecord) -> TaskEvidenceRef:
             "segment_end": int(marker["end"]),
             "segment_hash": str(marker["hash"]),
         }
-    identity = record.model_dump(mode="json", exclude={"origin_meta", "imported_at", "updated_at"})
     return {
         "evidence_id": record.evidence_id,
         "revision": record.revision,
         "content_hash": record.content_hash,
-        "identity_hash": sha256(json.dumps(identity, sort_keys=True).encode()).hexdigest(),
+        "identity_hash": _evidence_identity_hash(record, include_written_by=False),
     }
 
 
+def _evidence_identity_hash(record: EvidenceRecord, *, include_written_by: bool) -> str:
+    excluded = {"origin_meta", "imported_at", "updated_at"}
+    if not include_written_by:
+        excluded.add("written_by")
+    identity = record.model_dump(mode="json", exclude=excluded)
+    return sha256(json.dumps(identity, sort_keys=True).encode()).hexdigest()
+
+
+def _identity_hash_matches(record: EvidenceRecord, expected: object) -> bool:
+    candidates = {_evidence_identity_hash(record, include_written_by=False)}
+    # Tasks created by pre-role-only builds keep their original proof. The
+    # legacy metadata could also have changed in place without a revision, so
+    # accept the old hash under every historical author guess.
+    for written_by in WrittenBy:
+        candidates.add(
+            _evidence_identity_hash(
+                record.model_copy(update={"written_by": written_by}),
+                include_written_by=True,
+            )
+        )
+    return expected in candidates
+
+
 def matches_evidence_ref(record: EvidenceRecord, ref: TaskEvidenceRef) -> bool:
     if "segment_hash" in ref:
         base = evidence_ref(record)
         if any(
             base.get(key) != ref.get(key)
-            for key in ("evidence_id", "revision", "content_hash", "identity_hash")
+            for key in ("evidence_id", "revision", "content_hash")
         ):
             return False
+        if not _identity_hash_matches(record, ref.get("identity_hash")):
+            return False
         start, end = ref.get("segment_start"), ref.get("segment_end")
         if not isinstance(start, int) or not isinstance(end, int) or not (0 <= start < end):
             return False
         return sha256(record.verbatim[start:end].encode()).hexdigest() == ref["segment_hash"]
     current = evidence_ref(record)
-    return all(current.get(key) == value for key, value in ref.items())
+    return all(
+        (
+            _identity_hash_matches(record, value)
+            if key == "identity_hash"
+            else current.get(key) == value
+        )
+        for key, value in ref.items()
+    )
 
 
 def segment_evidence_record(
@@ -977,6 +1032,44 @@ class MemoryTaskStore:
             error_code=error_code,
         )
 
+    def release_unstarted(
+        self,
+        task_id: str,
+        *,
+        runner_token: str,
+        error_code: str,
+    ) -> MemoryTask:
+        """Return a claimed task to pending when a user pause wins before HY.
+
+        Claim increments ``attempts``. A pause is not an execution failure, so
+        the attempt is returned as well; repeatedly pausing and resuming can
+        never exhaust an otherwise healthy task.
+        """
+        now = datetime.now(UTC).isoformat()
+        with self._connect(writing=True) as connection:
+            changed = connection.execute(
+                """
+                UPDATE memory_tasks
+                SET status=?, runner_token=NULL,
+                    attempts=CASE WHEN attempts>0 THEN attempts-1 ELSE 0 END,
+                    last_error_code=?, duration_seconds=NULL, updated_at=?
+                WHERE task_id=? AND status=? AND runner_token=?
+                """,
+                (
+                    MemoryTaskStatus.PENDING,
+                    error_code,
+                    now,
+                    task_id,
+                    MemoryTaskStatus.RUNNING,
+                    runner_token,
+                ),
+            )
+            if changed.rowcount != 1:
+                raise MemoryTaskOwnershipError(
+                    "Memory task runner no longer owns this task"
+                )
+        return self.require(task_id)
+
     def park_for_permission(self, task_id: str, *, error_code: str) -> bool:
         """Park unstarted work until a later permission snapshot admits it."""
         now = datetime.now(UTC).isoformat()
@@ -1245,7 +1338,7 @@ class MemoryTaskStore:
         live_clause = f"task.task_id IN ({live_placeholders})" if live_ids else "0"
         not_live_clause = f"task.task_id NOT IN ({live_placeholders})" if live_ids else "1"
         excluded_clause = f"task.task_id NOT IN ({excluded_placeholders})" if excluded else "1"
-        common = f"""
+        common_base = f"""
             task.status IN (?,?,?,?)
             AND coalesce(task.last_error_code, '') NOT IN (
                 'hy_memory_input_exceeds_window', 'model_content_outcome_unknown'
@@ -1264,7 +1357,33 @@ class MemoryTaskStore:
         """
         moment = "coalesce(evidence.created_at, evidence.imported_at)"
 
+        def common_for(connection: sqlite3.Connection) -> str:
+            has_durable_queue = connection.execute(
+                "SELECT COUNT(*) FROM sqlite_master WHERE type='table' "
+                "AND name IN ('memory_write_jobs','memory_write_job_evidence')"
+            ).fetchone()[0]
+            if int(has_durable_queue) != 2:
+                return common_base
+            return (
+                common_base
+                + """
+                AND NOT EXISTS (
+                    SELECT 1
+                    FROM memory_task_inputs AS inputs, json_each(inputs.refs_json) AS ref
+                    JOIN memory_write_job_evidence AS linked
+                      ON linked.evidence_id=json_extract(ref.value, '$.evidence_id')
+                    JOIN memory_write_jobs AS job ON job.job_id=linked.job_id
+                    WHERE inputs.task_id=task.task_id
+                      AND job.state IN (
+                          'pending', 'running', 'unknown',
+                          'waiting_for_permission', 'waiting_for_model'
+                      )
+                )
+            """
+            )
+
         def run(connection: sqlite3.Connection, *, defer: bool) -> list[sqlite3.Row]:
+            common = common_for(connection)
             in_window = (
                 "1"
                 if extraction_since is None
@@ -1361,6 +1480,7 @@ class MemoryTaskStore:
         def has_defer_candidate(connection: sqlite3.Connection) -> bool:
             if extraction_since is None:
                 return False
+            common = common_for(connection)
             row = connection.execute(
                 f"""
                 SELECT 1
@@ -2077,6 +2197,7 @@ class MemoryTaskStore:
         *,
         limit: int = 128,
         after_cursor: tuple[str, str] | None = None,
+        exclude_error_codes: tuple[str, ...] = (),
     ) -> tuple[MemoryTask, ...]:
         """List `unknown` tasks that still need an external inquiry.
 
@@ -2091,6 +2212,9 @@ class MemoryTaskStore:
         """
         if type(limit) is not int or limit < 1:
             raise ValueError("limit must be positive")
+        excluded = tuple(dict.fromkeys(exclude_error_codes))
+        if not all(isinstance(code, str) and code for code in excluded):
+            raise ValueError("exclude_error_codes must contain non-empty strings")
         if after_cursor is not None:
             if (
                 not isinstance(after_cursor, tuple)
@@ -2104,13 +2228,19 @@ class MemoryTaskStore:
         else:
             cursor_clause = ""
             params = ()
+        if excluded:
+            placeholders = ",".join("?" for _ in excluded)
+            excluded_clause = f"AND COALESCE(last_error_code, '') NOT IN ({placeholders})"
+        else:
+            excluded_clause = ""
         with self._connect() as connection:
             rows = connection.execute(
                 "SELECT * FROM memory_tasks "
                 "WHERE status=? "
                 f"{cursor_clause} "
+                f"{excluded_clause} "
                 "ORDER BY updated_at, task_id LIMIT ?",
-                (MemoryTaskStatus.UNKNOWN, *params, limit),
+                (MemoryTaskStatus.UNKNOWN, *params, *excluded, limit),
             ).fetchall()
         return tuple(self._to_task(row) for row in rows)
 
@@ -2142,20 +2272,68 @@ class MemoryTaskStore:
         """
         now = datetime.now(UTC).isoformat()
         with self._connect(writing=True) as connection:
+            task_only = _task_only_unknown_clause(connection)
+            blocked = ",".join("?" for _ in NON_RETRYABLE_POST_WRITE_CODES)
             changed = connection.execute(
-                """
+                f"""
                 UPDATE memory_tasks
                 SET status=?, runner_token=NULL, attempts=0,
                     last_error_code='model_content_manual_retry_requested', updated_at=?
                 WHERE task_id IN (
                     SELECT task_id FROM memory_tasks
-                    WHERE status=? ORDER BY updated_at, task_id LIMIT ?
+                    WHERE status=?
+                      AND COALESCE(last_error_code, '') NOT IN ({blocked})
+                      {task_only}
+                    ORDER BY updated_at, task_id LIMIT ?
                 )
                 """,
                 (
                     MemoryTaskStatus.PENDING,
                     now,
                     MemoryTaskStatus.UNKNOWN,
+                    *NON_RETRYABLE_POST_WRITE_CODES,
+                    limit,
+                ),
+            )
+        return changed.rowcount
+
+    def retry_first_unknown_after_reconcile(
+        self,
+        *,
+        older_than: datetime,
+        limit: int = 128,
+    ) -> int:
+        """Automatically retry a task-only unknown exactly once.
+
+        Reconciliation runs before this transition. Attempts are preserved, so
+        the retry claim moves `attempts=1` to two and another unknown is left
+        for explicit user recovery.
+        """
+        if type(limit) is not int or limit < 1:
+            raise ValueError("limit must be positive")
+        now = datetime.now(UTC).isoformat()
+        with self._connect(writing=True) as connection:
+            task_only = _task_only_unknown_clause(connection)
+            blocked = ",".join("?" for _ in NON_RETRYABLE_POST_WRITE_CODES)
+            changed = connection.execute(
+                f"""
+                UPDATE memory_tasks
+                SET status=?, runner_token=NULL,
+                    last_error_code='model_content_unknown_auto_retry', updated_at=?
+                WHERE task_id IN (
+                    SELECT task_id FROM memory_tasks
+                    WHERE status=? AND attempts=1 AND updated_at <= ?
+                      AND COALESCE(last_error_code, '') NOT IN ({blocked})
+                      {task_only}
+                    ORDER BY updated_at, task_id LIMIT ?
+                )
+                """,
+                (
+                    MemoryTaskStatus.PENDING,
+                    now,
+                    MemoryTaskStatus.UNKNOWN,
+                    older_than.isoformat(),
+                    *NON_RETRYABLE_POST_WRITE_CODES,
                     limit,
                 ),
             )

memhub/apps/core/src/one_memory/memory_visibility.py

@@ -200,10 +200,12 @@ def _trigger_sql() -> list[tuple[str, str]]:
         CREATE TRIGGER IF NOT EXISTS memory_visibility_dirty_on_link_update
         AFTER UPDATE ON governed_memory_evidence
         BEGIN
-            INSERT OR IGNORE INTO memory_visibility_dirty (memory_id)
-            VALUES (NEW.memory_id);
-            INSERT OR IGNORE INTO memory_visibility_dirty (memory_id)
-            VALUES (OLD.memory_id);
+            INSERT INTO memory_visibility_dirty (memory_id)
+            VALUES (NEW.memory_id)
+            ON CONFLICT(memory_id) DO NOTHING;
+            INSERT INTO memory_visibility_dirty (memory_id)
+            VALUES (OLD.memory_id)
+            ON CONFLICT(memory_id) DO NOTHING;
         END;""",
         ),
         (
@@ -294,6 +296,10 @@ def install_triggers(db: sqlite3.Connection) -> None:
         str(row["name"])
         for row in cursor.execute("SELECT name FROM sqlite_master WHERE type='table'")
     }
+    # CREATE TRIGGER IF NOT EXISTS cannot repair the legacy definition. Replace
+    # this derived-cache trigger transactionally on every startup so old
+    # libraries get the row-level conflict handling without a schema migration.
+    cursor.execute("DROP TRIGGER IF EXISTS memory_visibility_dirty_on_link_update")
     for table, statement in _trigger_sql():
         if table not in present:
             continue
@@ -303,11 +309,15 @@ def install_triggers(db: sqlite3.Connection) -> None:
 def startup(database_path: Path) -> None:
     """Idempotent: create the derived tables and triggers. No version change.
 
-    Runs through the write connection because the triggers are schema. It is
-    one short transaction and happens once per start.
+    Runs through the write connection because the triggers are schema. The
+    trigger replacement is one short transaction and happens once per start.
     """
-    with main_database(database_path).write() as db:
+    with main_database(database_path).write() as db, db:
         ensure_tables(db)
+        # ensure_tables uses executescript(), which commits any pending legacy
+        # transaction. Begin only afterward so replacing the installed trigger
+        # is one atomic schema change and a failed CREATE restores the old one.
+        db.execute("BEGIN IMMEDIATE")
         install_triggers(db)
 
 

memhub/apps/core/src/one_memory/model_request_audit.py

@@ -0,0 +1,404 @@
+"""Persistent, content-free accounting for precipitation model requests."""
+
+from __future__ import annotations
+
+import asyncio
+import json
+import logging
+from collections.abc import Iterator
+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
+
+import httpx
+
+from one_memory.database_access import main_database
+from one_memory.memory_server.migrations import model_request_audit_ddl
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass(frozen=True)
+class ModelRequestAuditContext:
+    request_id: str
+
+
+_audit_context: ContextVar[ModelRequestAuditContext | None] = ContextVar(
+    "one_memory_model_request_audit_context",
+    default=None,
+)
+
+
+@contextmanager
+def audit_precipitation_request(request_id: str) -> Iterator[None]:
+    """Enable physical-request accounting for one memory precipitation."""
+
+    token = _audit_context.set(ModelRequestAuditContext(request_id=request_id))
+    try:
+        yield
+    finally:
+        _audit_context.reset(token)
+
+
+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())
+
+
+class ModelRequestAuditStore:
+    def __init__(self, database: Path) -> None:
+        self.database = database
+        with main_database(database).write() as connection, connection:
+            for statement in model_request_audit_ddl("sqlite").split(";"):
+                if statement.strip():
+                    connection.execute(statement)
+
+    def start(
+        self,
+        *,
+        request_id: str,
+        profile_id: str,
+        profile_revision: int,
+        provider_kind: str,
+        model: str,
+        request_kind: str,
+        prompt_stage: str,
+    ) -> str:
+        audit_id = str(uuid4())
+        with main_database(self.database).write() as connection, connection:
+            connection.execute(
+                """
+                INSERT INTO model_request_audit (
+                    audit_id,request_id,profile_id,profile_revision,provider_kind,
+                    model,request_kind,prompt_stage,started_at,outcome
+                ) VALUES (?,?,?,?,?,?,?,?,?,?)
+                """,
+                (
+                    audit_id,
+                    request_id,
+                    profile_id,
+                    profile_revision,
+                    provider_kind,
+                    model,
+                    request_kind,
+                    prompt_stage,
+                    datetime.now(UTC).isoformat(),
+                    "in_flight",
+                ),
+            )
+        return audit_id
+
+    def finish(
+        self,
+        audit_id: str,
+        *,
+        outcome: str,
+        http_status: int | None,
+        error_type: str | None = None,
+        provider_request_id: str | None = None,
+        prompt_tokens: int | None = None,
+        completion_tokens: int | None = None,
+        total_tokens: int | None = None,
+        cached_prompt_tokens: int | None = None,
+        reasoning_tokens: int | None = None,
+    ) -> None:
+        with main_database(self.database).write() as connection, connection:
+            connection.execute(
+                """
+                UPDATE model_request_audit
+                SET completed_at=?,outcome=?,http_status=?,error_type=?,
+                    provider_request_id=?,prompt_tokens=?,completion_tokens=?,
+                    total_tokens=?,cached_prompt_tokens=?,reasoning_tokens=?
+                WHERE audit_id=?
+                """,
+                (
+                    datetime.now(UTC).isoformat(),
+                    outcome,
+                    http_status,
+                    error_type,
+                    provider_request_id,
+                    prompt_tokens,
+                    completion_tokens,
+                    total_tokens,
+                    cached_prompt_tokens,
+                    reasoning_tokens,
+                    audit_id,
+                ),
+            )
+
+    def report(
+        self,
+        *,
+        since: str | None = None,
+        request_id: str | None = None,
+        limit: int = 200,
+    ) -> dict[str, object]:
+        clauses: list[str] = []
+        params: list[object] = []
+        if since is not None:
+            clauses.append("started_at>=?")
+            params.append(since)
+        if request_id is not None:
+            clauses.append("request_id=?")
+            params.append(request_id)
+        where = f"WHERE {' AND '.join(clauses)}" if clauses else ""
+        with main_database(self.database).read() as connection:
+            summary = connection.execute(
+                f"""
+                SELECT COUNT(*) AS request_count,
+                       SUM(CASE WHEN outcome='succeeded' THEN 1 ELSE 0 END) AS succeeded,
+                       SUM(CASE WHEN outcome='failed' THEN 1 ELSE 0 END) AS failed,
+                       SUM(CASE WHEN outcome='in_flight' THEN 1 ELSE 0 END) AS in_flight,
+                       SUM(prompt_tokens) AS prompt_tokens,
+                       SUM(completion_tokens) AS completion_tokens,
+                       SUM(total_tokens) AS total_tokens,
+                       SUM(cached_prompt_tokens) AS cached_prompt_tokens,
+                       SUM(reasoning_tokens) AS reasoning_tokens,
+                       AVG(prompt_tokens) AS avg_prompt_tokens,
+                       AVG(completion_tokens) AS avg_completion_tokens
+                FROM model_request_audit {where}
+                """,
+                tuple(params),
+            ).fetchone()
+            rows = connection.execute(
+                f"""
+                SELECT audit_id,request_id,profile_id,profile_revision,provider_kind,
+                       model,request_kind,prompt_stage,started_at,completed_at,outcome,http_status,
+                       error_type,provider_request_id,prompt_tokens,completion_tokens,
+                       total_tokens,cached_prompt_tokens,reasoning_tokens
+                FROM model_request_audit {where}
+                ORDER BY started_at DESC,audit_id DESC LIMIT ?
+                """,
+                (*params, limit),
+            ).fetchall()
+            minute_rows = connection.execute(
+                f"""
+                SELECT substr(started_at,1,16) || ':00Z' AS minute,
+                       prompt_stage,
+                       COUNT(*) AS request_count,
+                       SUM(prompt_tokens) AS prompt_tokens,
+                       SUM(completion_tokens) AS completion_tokens,
+                       SUM(total_tokens) AS total_tokens,
+                       AVG(prompt_tokens) AS avg_prompt_tokens,
+                       AVG(completion_tokens) AS avg_completion_tokens
+                FROM model_request_audit {where}
+                GROUP BY substr(started_at,1,16),prompt_stage
+                ORDER BY minute,prompt_stage
+                """,
+                tuple(params),
+            ).fetchall()
+        keys = (
+            "audit_id",
+            "request_id",
+            "profile_id",
+            "profile_revision",
+            "provider_kind",
+            "model",
+            "request_kind",
+            "prompt_stage",
+            "started_at",
+            "completed_at",
+            "outcome",
+            "http_status",
+            "error_type",
+            "provider_request_id",
+            "prompt_tokens",
+            "completion_tokens",
+            "total_tokens",
+            "cached_prompt_tokens",
+            "reasoning_tokens",
+        )
+        summary_keys = (
+            "request_count",
+            "succeeded",
+            "failed",
+            "in_flight",
+            "prompt_tokens",
+            "completion_tokens",
+            "total_tokens",
+            "cached_prompt_tokens",
+            "reasoning_tokens",
+            "avg_prompt_tokens",
+            "avg_completion_tokens",
+        )
+        return {
+            "summary": {
+                key: (
+                    round(float(value), 2)
+                    if key.startswith("avg_") and value is not None
+                    else int(value)
+                    if value is not None
+                    else 0
+                )
+                for key, value in zip(summary_keys, summary, strict=True)
+            },
+            "requests": [
+                dict(zip(keys, tuple(row), strict=True))
+                for row in rows
+            ],
+            "by_minute": [
+                {
+                    "minute": str(row[0]),
+                    "prompt_stage": str(row[1]),
+                    "request_count": int(row[2]),
+                    "prompt_tokens": int(row[3] or 0),
+                    "completion_tokens": int(row[4] or 0),
+                    "total_tokens": int(row[5] or 0),
+                    "avg_prompt_tokens": round(float(row[6] or 0), 2),
+                    "avg_completion_tokens": round(float(row[7] or 0), 2),
+                }
+                for row in minute_rows
+            ],
+        }
+
+
+def _non_negative_int(value: object) -> int | None:
+    return value if type(value) is int and value >= 0 else None
+
+
+class _ResponseUsage(TypedDict):
+    provider_request_id: str | None
+    prompt_tokens: int | None
+    completion_tokens: int | None
+    total_tokens: int | None
+    cached_prompt_tokens: int | None
+    reasoning_tokens: int | None
+
+
+def _empty_usage() -> _ResponseUsage:
+    return {
+        "provider_request_id": None,
+        "prompt_tokens": None,
+        "completion_tokens": None,
+        "total_tokens": None,
+        "cached_prompt_tokens": None,
+        "reasoning_tokens": None,
+    }
+
+
+def _response_usage(content: bytes) -> _ResponseUsage:
+    try:
+        body = json.loads(content)
+    except (json.JSONDecodeError, UnicodeDecodeError):
+        return _empty_usage()
+    if not isinstance(body, dict):
+        return _empty_usage()
+    usage = body.get("usage")
+    if not isinstance(usage, dict):
+        usage = {}
+    prompt_details = usage.get("prompt_tokens_details")
+    completion_details = usage.get("completion_tokens_details")
+    cached = (
+        prompt_details.get("cached_tokens")
+        if isinstance(prompt_details, dict)
+        else usage.get("prompt_cache_hit_tokens")
+    )
+    reasoning = (
+        completion_details.get("reasoning_tokens")
+        if isinstance(completion_details, dict)
+        else None
+    )
+    return {
+        "provider_request_id": body.get("id") if isinstance(body.get("id"), str) else None,
+        "prompt_tokens": _non_negative_int(usage.get("prompt_tokens")),
+        "completion_tokens": _non_negative_int(usage.get("completion_tokens")),
+        "total_tokens": _non_negative_int(usage.get("total_tokens")),
+        "cached_prompt_tokens": _non_negative_int(cached),
+        "reasoning_tokens": _non_negative_int(reasoning),
+    }
+
+
+def _request_kind(request: httpx.Request) -> str:
+    path = request.url.path.rstrip("/")
+    if path.endswith("/chat/completions"):
+        return "chat_completion"
+    if path.endswith("/embeddings"):
+        return "embedding"
+    return "model_http"
+
+
+def _request_model(content: bytes, fallback: str) -> str:
+    try:
+        body = json.loads(content)
+    except (json.JSONDecodeError, UnicodeDecodeError):
+        return fallback
+    model = body.get("model") if isinstance(body, dict) else None
+    return model if isinstance(model, str) and model else fallback
+
+
+class AuditedModelTransport(httpx.AsyncBaseTransport):
+    """Record one row for each physical precipitation HTTP attempt."""
+
+    def __init__(
+        self,
+        store: ModelRequestAuditStore,
+        *,
+        profile_id: str,
+        profile_revision: int,
+        provider_kind: str,
+        model: str,
+        inner: httpx.AsyncBaseTransport | None = None,
+    ) -> None:
+        self.store = store
+        self.profile_id = profile_id
+        self.profile_revision = profile_revision
+        self.provider_kind = provider_kind
+        self.model = model
+        self.inner = inner or httpx.AsyncHTTPTransport(retries=0, trust_env=False)
+
+    async def handle_async_request(self, request: httpx.Request) -> httpx.Response:
+        context = _audit_context.get()
+        if context is None:
+            return await self.inner.handle_async_request(request)
+        content = await request.aread()
+        audit_id = await asyncio.to_thread(
+            self.store.start,
+            request_id=context.request_id,
+            profile_id=self.profile_id,
+            profile_revision=self.profile_revision,
+            provider_kind=self.provider_kind,
+            model=_request_model(content, self.model),
+            request_kind=_request_kind(request),
+            prompt_stage=_current_prompt_stage(),
+        )
+        try:
+            response = await self.inner.handle_async_request(request)
+        except BaseException as error:
+            await asyncio.to_thread(
+                self.store.finish,
+                audit_id,
+                outcome="failed",
+                http_status=None,
+                error_type=type(error).__name__,
+            )
+            raise
+        content = await response.aread()
+        usage = _response_usage(content)
+        try:
+            await asyncio.to_thread(
+                self.store.finish,
+                audit_id,
+                outcome="succeeded" if response.is_success else "failed",
+                http_status=response.status_code,
+                error_type=None if response.is_success else f"http_{response.status_code}",
+                **usage,
+            )
+        except Exception:
+            # The durable in_flight row is more truthful than hiding the model
+            # response after it has already incurred cost.
+            logger.exception("model_request_audit_finish_failed audit_id=%s", audit_id)
+        return response
+
+    async def aclose(self) -> None:
+        await self.inner.aclose()

memhub/apps/core/src/one_memory/onboarding.py

@@ -68,6 +68,16 @@ COMPLETE_INSERT_MYSQL = (
     "singleton, status, selected_source_ids_json, started_at, completed_at, updated_at"
     ") VALUES (1, 'completed', '[]', NULL, %s, %s)"
 )
+RESET_UPDATE_SQLITE = (
+    "UPDATE onboarding_state SET status = 'not_started', "
+    "selected_source_ids_json = '[]', started_at = NULL, completed_at = NULL, "
+    "updated_at = ? WHERE singleton = 1"
+)
+RESET_UPDATE_MYSQL = (
+    "UPDATE onboarding_state SET status = 'not_started', "
+    "selected_source_ids_json = '[]', started_at = NULL, completed_at = NULL, "
+    "updated_at = %s WHERE singleton = 1"
+)
 
 
 class OnboardingState(BaseModel):
@@ -140,6 +150,30 @@ class OnboardingStore:
                 )
         return self.get()
 
+    def reset(self) -> OnboardingState:
+        """Return onboarding to its first-run state, touching nothing else.
+
+        The completion flag is stored beside the memories on purpose: a fresh
+        install of the app must not greet someone who already has memories as a
+        first-time user. That makes the way back to the cold-start screens an
+        explicit action rather than a side effect of reinstalling, and it is
+        the whole reason this exists: QA and long-time users need to see those
+        screens again without giving up a single memory.
+
+        Sources, memories and settings are separate tables and are not read or
+        written here.
+        """
+        now = datetime.now(UTC).isoformat()
+        mysql = is_mysql(self._store)
+        with self._connect(writing=True) as connection:
+            execute(
+                connection,
+                mysql,
+                RESET_UPDATE_MYSQL if mysql else RESET_UPDATE_SQLITE,
+                (now,),
+            )
+        return self.get()
+
     @contextmanager
     def _connect(
         self,

memhub/apps/core/src/one_memory/quality_acceptance.py

@@ -87,7 +87,7 @@ def build_acceptance_pack(
             """
             SELECT DISTINCT e.evidence_id
             FROM evidence e
-            WHERE e.role = 'user' AND e.written_by = 'person'
+            WHERE e.role = 'user'
               AND LENGTH(e.verbatim) <= 300
               AND (
                 e.verbatim LIKE '%我是一名%' OR e.verbatim LIKE '%我是一个%'
@@ -118,7 +118,7 @@ def build_acceptance_pack(
         for row in connection.execute(
             """
             SELECT evidence_id FROM evidence
-            WHERE role = 'user' AND written_by != 'person'
+            WHERE role != 'user'
             """
         ):
             strata["attribution_edge"].add(row["evidence_id"])
@@ -138,11 +138,7 @@ def build_acceptance_pack(
             quota,
         )
     if len(selected) < sample_size:
-        eligible = (
-            evidence_id
-            for evidence_id, row in evidence.items()
-            if row["role"] == "user" and row["written_by"] == "person"
-        )
+        eligible = (evidence_id for evidence_id, row in evidence.items() if row["role"] == "user")
         _extend_unique(selected, _stable_order(eligible, seed=f"{seed}:fill"), sample_size)
     selected = selected[:sample_size]
     if len(selected) < sample_size:
@@ -163,7 +159,11 @@ def build_acceptance_pack(
                 row_id=f"{index:03d}",
                 evidence_id=evidence_id,
                 source_id=source["source_id"],
-                written_by=source["written_by"],
+                written_by=(
+                    "person"
+                    if source["role"] == "user"
+                    else ("agent" if source["role"] == "assistant" else source["role"])
+                ),
                 character_count=len(source["verbatim"]),
                 content_hash=source["content_hash"],
                 strata=tuple(

memhub/apps/core/src/one_memory/retrieval/content.py

@@ -2,7 +2,7 @@
 
 from collections.abc import Sequence
 
-from one_memory.imports.models import EvidenceContentKind, EvidenceRecord, WrittenBy
+from one_memory.imports.models import EvidenceContentKind, EvidenceRecord
 from one_memory.imports.sources.parsing import is_host_generated_user_turn, strip_host_wrappers
 from one_memory.input_sanitization import sanitize_model_input
 from one_memory.memory_processing import _is_hy_input_turn
@@ -15,10 +15,7 @@ def identity_content_ready(record: EvidenceRecord) -> bool:
     # cleaning requires a source map and remains explicitly uncovered for now.
     return (
         record.content_kind is EvidenceContentKind.RAW_CONVERSATION
-        and (
-            (record.role == "user" and record.written_by is WrittenBy.PERSON)
-            or (record.role == "assistant" and record.written_by is WrittenBy.AGENT)
-        )
+        and record.role in {"user", "assistant"}
         and _is_hy_input_turn(record)
         and not is_host_generated_user_turn(record.verbatim)
         and strip_host_wrappers(record.verbatim) == record.verbatim.strip()

memhub/apps/core/src/one_memory/retrieval/mapped_content.py

@@ -10,7 +10,6 @@ from one_memory.imports.models import (
     MAX_VERBATIM_CHARACTERS,
     EvidenceContentKind,
     EvidenceRecord,
-    WrittenBy,
 )
 from one_memory.imports.sources.parsing import (
     is_host_generated_user_turn,
@@ -164,10 +163,10 @@ def prepare_mapped_content(record: EvidenceRecord) -> PreparedMappedContent:
     The returned text is transient; only metadata is suitable for an outer manifest.
     That manifest must independently bind Evidence identity/revision/hash and grants.
     """
-    if record.content_kind is not EvidenceContentKind.RAW_CONVERSATION or not (
-        (record.role == "user" and record.written_by is WrittenBy.PERSON)
-        or (record.role == "assistant" and record.written_by is WrittenBy.AGENT)
-    ):
+    if record.content_kind is not EvidenceContentKind.RAW_CONVERSATION or record.role not in {
+        "user",
+        "assistant",
+    }:
         raise MappedContentError("mapped_content_ineligible")
     original = record.verbatim
     if len(original) > MAX_VERBATIM_CHARACTERS:

memhub/apps/core/src/one_memory/retrieval/raw_evidence.py

@@ -11,7 +11,7 @@ from datetime import UTC
 
 from one_memory import ingest_journal as journal
 from one_memory.context_raw import _check, _connection
-from one_memory.imports.models import EvidenceRecord, WrittenBy
+from one_memory.imports.models import EvidenceRecord
 from one_memory.imports.storage import EvidenceStore
 from one_memory.retrieval.index_store import RetrievalIndexStore
 from one_memory.retrieval.models import EvidenceSpan, RawEvidence
@@ -166,7 +166,7 @@ def materialize_raw_candidates(
                         source_id=record.source_id,
                         conversation_id=record.conversation_id,
                         role=record.role,
-                        written_by="person" if record.written_by is WrittenBy.PERSON else "agent",
+                        written_by="person" if record.role == "user" else "agent",
                         event_time=event_time,
                         spans=tuple(spans),
                         matched_by=("semantic",),

memhub/apps/core/src/one_memory/service_levels.py

@@ -264,7 +264,7 @@ class ServiceLevelScheduler:
             self._quiet_admission_logged_at[key] = now
             if last is not None:
                 diagnostic_event(
-                    level="info",
+                    level="debug",
                     component="service_level.scheduler",
                     event="service_level.summary",
                     outcome="succeeded",
@@ -276,7 +276,7 @@ class ServiceLevelScheduler:
                 )
                 return
         diagnostic_event(
-            level="info",
+            level="debug",
             component="service_level.scheduler",
             event="service_level.admitted",
             outcome="succeeded",

memhub/apps/core/src/one_memory/source_import_receipts.py

@@ -56,10 +56,10 @@ def build_source_import_receipt(
             "SELECT count(*) AS messages, "
             "count(DISTINCT conversation_id || char(0) || COALESCE(workspace, '')) "
             "AS conversations, "
-            "coalesce(sum(written_by='person'),0) AS person_messages, "
-            "coalesce(sum(written_by='agent'),0) AS agent_messages, "
-            "coalesce(sum(written_by='tool'),0) AS tool_messages, "
-            "coalesce(sum(written_by NOT IN ('person','agent','tool')),0) "
+            "coalesce(sum(role='user'),0) AS person_messages, "
+            "coalesce(sum(role='assistant'),0) AS agent_messages, "
+            "coalesce(sum(role='tool'),0) AS tool_messages, "
+            "coalesce(sum(role NOT IN ('user','assistant','tool')),0) "
             "AS unknown_author_messages, "
             "coalesce(sum(created_at IS NULL),0) AS unknown_time_messages "
             "FROM evidence WHERE user_id=? AND source_id=?",
@@ -68,7 +68,9 @@ def build_source_import_receipt(
         # Bound text reads before materializing snippets. These are samples, not a
         # recency ranking; sorting every large original body is unnecessary here.
         examples = db.execute(
-            "SELECT evidence_id, role, written_by, created_at, "
+            "SELECT evidence_id, role, CASE role "
+            "WHEN 'user' THEN 'person' WHEN 'assistant' THEN 'agent' "
+            "WHEN 'tool' THEN 'tool' ELSE 'unknown' END AS written_by, created_at, "
             "substr(verbatim,1,600) AS excerpt, length(verbatim)>600 AS truncated "
             "FROM evidence WHERE evidence_id IN ("
             "SELECT evidence_id FROM evidence WHERE user_id=? AND source_id=? LIMIT 3)",
@@ -80,10 +82,13 @@ def build_source_import_receipt(
         # Qualify before LIMIT: agent/system rows must not consume the three samples.
         query_filter = " AND evidence_id=?" if query_evidence_id is not None else ""
         queries = db.execute(
-            f"""SELECT evidence_id, revision AS evidence_revision, role, written_by, created_at,
-                CASE WHEN user_id=? AND source_id=? AND role='user' AND written_by='person'
+            f"""SELECT evidence_id, revision AS evidence_revision, role,
+                CASE role WHEN 'user' THEN 'person' WHEN 'assistant' THEN 'agent'
+                    WHEN 'tool' THEN 'tool' ELSE 'unknown' END AS written_by,
+                created_at,
+                CASE WHEN user_id=? AND source_id=? AND role='user'
                     AND length(verbatim)<=65536
-                    THEN display_query_projection(verbatim,role,written_by)
+                    THEN display_query_projection(verbatim,role,'person')
                     ELSE NULL END AS projection
                 FROM evidence WHERE user_id=? AND source_id=?{query_filter}
                     AND projection IS NOT NULL

memhub/apps/core/tests/TEST_INDEX.md

@@ -250,6 +250,7 @@
 | `test_model_profile_identity.py` | `contract` | Stable model editing and configured-vs-available credentials; no Keychain/network. |
 | `test_model_profile_queue_handover.py` | `contract` | Queued work when the model changes or its configuration is deleted (D75). |
 | `test_model_profiles.py` | `contract` | 覆盖 profile uses one model and keeps key out of database、api key updates reuse one stable secret reference、startup cleanup removes only secrets not referenced by current profiles 等 15 个场景。 |
+| `test_model_request_audit.py` | `contract` | 覆盖 failure is recorded without error message、missing usage stays unknown in detail and zero in summary、records one physical request with provider usage and no content 等 5 个场景。 |
 | `test_model_request_budget.py` | `contract` | 覆盖 persists failed request reservations without content、parallel leases cannot exceed request limit、unknown or excessive output and body are denied 等 6 个场景。 |
 | `test_mysql_lifecycle.py` | `contract` | A1-07: the decisions Runtime makes around the bundled MySQL. |
 | `test_mysql_platform.py` | `contract` | A1-07: the machine-facing half of managing the bundled MySQL. |
@@ -327,6 +328,7 @@
 | `test_secret_store_error_codes.py` | `integration` | 3925a kernel item 3: keychain failure classification and backoff that heals. |
 | `test_secure_storage.py` | `integration` | 覆盖 isolated keychain service namespace、custom keychain service is refused outside isolated tests、packaged keychain budget covers loaded process startup 等 20 个场景。 |
 | `test_sedimentation_concurrency.py` | `contract` | Sedimentation runs several harnesses at once, but one harness at a time. |
+| `test_sedimentation_switch.py` | `contract` | Stopping model-driven sedimentation without losing the configuration. |
 | `test_service_level_pressure_acceptance.py` | `contract` | 覆盖 workbuddy recall remains available during 100k import 场景。 |
 | `test_service_levels.py` | `contract` | 覆盖 background never consumes the realtime reservation、realtime is next for the same source after background yields、import history gets one slice after nine realtime admissions 等 8 个场景。 |
 | `test_session_inbox.py` | `contract` | A session inbox is not a shared mailbox and a pull is not adoption. |
@@ -336,6 +338,7 @@
 | `test_snapshot_pool.py` | `contract` | A whole-directory snapshot must not hold a thread everything else is waiting for. |
 | `test_soak_probe_effectiveness.py` | `integration` | The soak probe must observe the load it reports on. |
 | `test_source_access.py` | `contract` | 覆盖 source root requires explicit user grant、sensitive home directory is rejected without requesting os access、authorized file cannot escape root through a symlink 等 24 个场景。 |
+| `test_source_authorization_state.py` | `unit` | Why a source cannot be read, named separately from whether it can. |
 | `test_source_candidate_api.py` | `integration` | Candidate location changes must not implicitly grant or import anything. |
 | `test_source_candidate_connections.py` | `contract` | 覆盖 empty candidate revisions preserve default fingerprint、persisted candidate requires exact revision、connection resolves candidate in transaction and prior replay never regrants 等 7 个场景。 |
 | `test_source_candidates.py` | `contract` | 覆盖 preview is metadata only and cancel changes nothing、select persists only candidate and restarts with revision、compare and swap and aba cannot reuse old preview 等 9 个场景。 |
@@ -385,4 +388,4 @@
 | `test_workspace_recall.py` | `contract` | Workspace memories were unreachable from every host; these pin the way in. |
 | `test_write_queue_lease_renewal.py` | `contract` | Rule 8: a lease must outlive the work it covers, and be renewed while it runs. |
 
-共 380 个测试文件。
+共 383 个测试文件。

memhub/apps/core/tests/core_test_classification.py

@@ -14,8 +14,8 @@ from pathlib import Path
 import pytest
 
 PRIMARY_MARKERS = ("unit", "contract", "integration", "performance", "live")
-EXPECTED_TEST_FILE_COUNT = 380
-EXPECTED_TEST_FILES_SHA256 = "9f075d18403328882c25325d29ebe77d0e3365d675ba19d56134a6a0fb9b64b4"
+EXPECTED_TEST_FILE_COUNT = 383
+EXPECTED_TEST_FILES_SHA256 = "3c1a3d3febaf478e1ba005fbcd22ad23d76fc5925b32b8d8740e4e65c554b922"
 
 # Pure transformations, validation models, parity checks and state machines.
 # This list is intentionally small: uncertain files default to contract.
@@ -35,6 +35,7 @@ UNIT_FILES = frozenset(
         "test_memory_server_job_state.py",
         "test_scan_roots.py",
         "test_scanner_classifier.py",
+        "test_source_authorization_state.py",
         "test_scanner_sessionizer.py",
         "test_source_envelope_positions.py",
         "test_source_lifecycle.py",

memhub/apps/core/tests/test_background_work.py

@@ -1,12 +1,17 @@
 import asyncio
+import gc
 import os
 import threading
 import time
+from concurrent.futures import ThreadPoolExecutor
+
+import pytest
 
 from one_memory.background_work import (
     BULK_WORKERS,
     bulk_pool,
     in_interactive_read,
+    in_read_pool,
     in_the_background,
     shutdown_bulk_pool,
 )
@@ -110,3 +115,73 @@ def test_interactive_read_does_not_wait_behind_saturated_default_pool() -> None:
         shutdown_bulk_pool()
 
     assert waited < 0.25
+
+
+def test_an_abandoned_read_does_not_report_an_unretrieved_failure() -> None:
+    """A read that outlives its caller must not leave its failure lying about.
+
+    `in_read_pool` shields the worker, so giving up on an answer does not yank
+    a query out from under SQLite. The caller cancels; the read runs on. When
+    such an abandoned read then failed, nothing awaited its future any more and
+    asyncio reported an exception nobody retrieved -- which the loop handler
+    files as `core.loop.task_exception`, a red error sitting next to a lane
+    that had already degraded on its own budget and already said so.
+    """
+
+    def open_pool() -> ThreadPoolExecutor:
+        return ThreadPoolExecutor(max_workers=1, thread_name_prefix="test-read-pool")
+
+    async def scenario() -> list[BaseException]:
+        loop = asyncio.get_running_loop()
+        reported: list[BaseException] = []
+        previous = loop.get_exception_handler()
+        loop.set_exception_handler(lambda _loop, context: reported.append(context.get("exception")))
+        gate = threading.Event()
+
+        def work() -> None:
+            gate.wait(5.0)
+            raise TimeoutError("raw_context_timeout")
+
+        pool = open_pool()
+        try:
+            read = asyncio.create_task(in_read_pool(pool, work))
+            await asyncio.sleep(0.05)
+            read.cancel()  # the lane ran out of budget; the read is still going
+            await asyncio.sleep(0.10)  # the cancellation settles first
+            gate.set()  # only now does the abandoned read fail
+            await asyncio.gather(read, return_exceptions=True)
+            # Let the shielded worker finish, then let asyncio notice anything
+            # that nobody ever retrieved. The reference has to go: asyncio only
+            # reports an unretrieved outcome when the future is collected.
+            del read
+            for _ in range(5):
+                await asyncio.sleep(0.02)
+            gc.collect()
+            await asyncio.sleep(0)
+        finally:
+            pool.shutdown(wait=True)
+            loop.set_exception_handler(previous)
+        return [error for error in reported if isinstance(error, TimeoutError)]
+
+    assert asyncio.run(scenario()) == []
+
+
+def test_an_awaited_read_still_raises_its_failure() -> None:
+    """Marking the outcome seen must not swallow it for a caller who waited."""
+
+    def open_pool() -> ThreadPoolExecutor:
+        return ThreadPoolExecutor(max_workers=1, thread_name_prefix="test-read-pool")
+
+    async def scenario() -> str:
+        def work() -> None:
+            raise TimeoutError("raw_context_timeout")
+
+        pool = open_pool()
+        try:
+            with pytest.raises(TimeoutError):
+                await in_read_pool(pool, work)
+        finally:
+            pool.shutdown(wait=True)
+        return "raised"
+
+    assert asyncio.run(scenario()) == "raised"

memhub/apps/core/tests/test_context_access.py

@@ -16,8 +16,10 @@ def database(tmp_path):
     path = tmp_path / "context.sqlite3"
     with closing(sqlite3.connect(path)) as db, db:
         db.execute("CREATE TABLE source_grants(user_id TEXT, source_id TEXT, revoked_at TEXT)")
+        db.execute("INSERT INTO source_grants VALUES('local','cursor',NULL)")
         db.execute("INSERT INTO source_grants VALUES('local','revoked','2026-09-19')")
         db.execute("INSERT INTO source_grants VALUES('other','cursor','2026-09-19')")
+        db.execute("INSERT INTO source_grants VALUES('other','foreign',NULL)")
     return path
 
 
@@ -111,6 +113,53 @@ def test_implicit_revoked_raw_path_can_be_reported_beside_allowed_memory(databas
     assert result.memory_allowed and not result.raw_allowed
 
 
+def test_generic_raw_uses_only_current_local_positive_grants(database):
+    generic = ExternalMemoryAccess(
+        "00000000-0000-0000-0000-000000000001",
+        1,
+        frozenset({"all-apps", "raw-imported-all"}),
+        frozenset({"om_recall", "om_conversations"}),
+        "generic-local-mcp",
+    )
+    result = resolve(database, generic)
+    assert result.raw_allowed and result.raw_sources == frozenset({"cursor"})
+    assert result.memory_allowed and result.memory_scopes == frozenset({"all-apps"})
+    assert resolve(database, generic, raw_source_ids=("cursor",)).raw_sources == {"cursor"}
+    for source in ("revoked", "foreign", "missing"):
+        with pytest.raises(HTTPException) as error:
+            resolve(database, generic, raw_source_ids=(source,))
+        assert error.value.status_code == 403
+
+    with closing(sqlite3.connect(database)) as db, db:
+        db.execute("DELETE FROM source_grants WHERE user_id='local' AND source_id='cursor'")
+    result = resolve(database, generic)
+    assert result.memory_allowed and not result.raw_allowed and result.raw_sources == frozenset()
+    with pytest.raises(HTTPException) as error:
+        resolve(database, generic, raw_source_ids=("cursor",))
+    assert error.value.status_code == 403
+    with pytest.raises(HTTPException):
+        resolve(database, generic, kinds=("raw",))
+
+
+def test_existing_per_source_authority_is_not_replaced_by_generic_grants(database):
+    with closing(sqlite3.connect(database)) as db, db:
+        db.execute("DELETE FROM source_grants WHERE user_id='local' AND source_id='cursor'")
+    result = resolve(database, access(["om_conversations"], ["all-apps", "source:cursor"]))
+    assert result.raw_allowed and result.raw_sources == frozenset({"cursor"})
+
+
+def test_generic_raw_needs_its_dedicated_scope(database):
+    generic = ExternalMemoryAccess(
+        "00000000-0000-0000-0000-000000000001",
+        1,
+        frozenset({"all-apps", "source:cursor"}),
+        frozenset({"om_recall", "om_conversations"}),
+        "generic-local-mcp",
+    )
+    result = resolve(database, generic)
+    assert result.memory_allowed and not result.raw_allowed
+
+
 def test_local_selection_does_not_create_or_initialize_a_database(tmp_path):
     path = tmp_path / "absent.sqlite3"
     result = resolve(path, None)

memhub/apps/core/tests/test_context_api.py

@@ -154,6 +154,69 @@ def test_revocation_during_other_lane_is_rechecked_before_response(app, database
     assert response.json()["state"] == "partial"
 
 
+def test_generic_context_tracks_active_local_sources_without_per_source_scopes(app, database):
+    generic = headers(
+        ["om_recall", "om_conversations"], scopes=("all-apps", "raw-imported-all")
+    ) | {"x-one-memory-host-id": "generic-local-mcp"}
+    with closing(sqlite3.connect(database)) as db, db:
+        grant(db, "cursor", revoked=None)
+        grant(db, "workbuddy", user="other", revoked=None)
+    response = post(app, auth=generic)
+    assert response.status_code == 200, response.text
+    assert {item["source_id"] for item in response.json()["items"] if item["kind"] == "raw"} == {
+        "cursor"
+    }
+    assert response.json()["paths"]["memory"]["status"] != "not_authorized"
+    for source in ("workbuddy", "missing"):
+        denied = post(app, {"query": "needle", "raw_source_ids": [source]}, generic)
+        assert denied.status_code == 403
+
+    with closing(sqlite3.connect(database)) as db, db:
+        grant(db, "workbuddy", revoked=None)
+    response = post(app, auth=generic)
+    assert {item["source_id"] for item in response.json()["items"] if item["kind"] == "raw"} == {
+        "cursor",
+        "workbuddy",
+    }
+    with closing(sqlite3.connect(database)) as db, db:
+        grant(db, "cursor", revoked="2026-09-20")
+    response = post(app, auth=generic)
+    assert {item["source_id"] for item in response.json()["items"] if item["kind"] == "raw"} == {
+        "workbuddy"
+    }
+    with closing(sqlite3.connect(database)) as db, db:
+        db.execute("DELETE FROM source_grants WHERE user_id='local' AND source_id='workbuddy'")
+    response = post(app, auth=generic)
+    assert response.status_code == 200
+    assert response.json()["paths"]["raw"]["status"] == "not_authorized"
+    assert response.json()["paths"]["memory"]["status"] != "not_authorized"
+
+
+def test_generic_raw_grant_removed_after_read_is_dropped(app, database, monkeypatch):
+    from one_memory import context_routes
+
+    with closing(sqlite3.connect(database)) as db, db:
+        grant(db, "cursor", revoked=None)
+    generic = headers(
+        ["om_recall", "om_conversations"], scopes=("all-apps", "raw-imported-all")
+    ) | {"x-one-memory-host-id": "generic-local-mcp"}
+    original = context_routes.read_raw_indexed_context
+
+    def remove_grant_after_read(*args, **kwargs):
+        found = original(*args, **kwargs)
+        assert any(item.source_id == "cursor" for item in found.items)
+        with closing(sqlite3.connect(database)) as db, db:
+            db.execute("DELETE FROM source_grants WHERE user_id='local' AND source_id='cursor'")
+        return found
+
+    monkeypatch.setattr(context_routes, "read_raw_indexed_context", remove_grant_after_read)
+    response = post(app, {"query": "needle", "kinds": ["raw"]}, generic)
+    assert response.status_code == 200, response.text
+    assert response.json()["items"] == []
+    assert response.json()["paths"]["raw"]["status"] == "degraded"
+    assert response.json()["paths"]["raw"]["reason"] == "evidence_changed"
+
+
 def test_missing_storage_is_unavailable_not_empty(app, tmp_path):
     app.state.evidence_store.database_path = tmp_path / "missing.sqlite3"
     response = post(app)

memhub/apps/core/tests/test_context_endpoint_guard.py

@@ -8,7 +8,11 @@ from one_memory.external_memory_access import require_external_endpoint
 
 
 def request(
-    tools, scopes=("all-apps", "source:cursor"), method="POST", path="/api/v1/context/search"
+    tools,
+    scopes=("all-apps", "source:cursor"),
+    method="POST",
+    path="/api/v1/context/search",
+    host_id=None,
 ):
     headers = {
         "x-one-memory-read-context": "external",
@@ -17,6 +21,8 @@ def request(
         "x-one-memory-allowed-scopes": json.dumps(scopes),
         "x-one-memory-allowed-tools": json.dumps(tools),
     }
+    if host_id is not None:
+        headers["x-one-memory-host-id"] = host_id
     return Request(
         {
             "type": "http",
@@ -56,6 +62,60 @@ def test_composite_read_does_not_open_other_routes(method, path):
         )
 
 
+def test_generic_scope_requires_exact_host_identity():
+    from one_memory.external_memory_access import external_access
+
+    scopes = ("all-apps", "raw-imported-all")
+    generic = request(["om_recall", "om_conversations"], scopes=scopes, host_id="generic-local-mcp")
+    assert external_access(generic.headers).scopes == frozenset(scopes)
+    for host_id in (None, "cursor", "generic-local-mcp-other"):
+        with pytest.raises(HTTPException) as error:
+            require_external_endpoint(
+                request(["om_recall", "om_conversations"], scopes=scopes, host_id=host_id)
+            )
+        assert error.value.status_code == 403
+
+
+@pytest.mark.parametrize(
+    "method,path",
+    [
+        ("GET", "/api/v1/conversations/search"),
+        ("POST", "/api/v1/memories/ask"),
+        ("POST", "/api/v1/matters/context/search"),
+        ("POST", "/api/v1/diagnostics/trace"),
+        ("POST", "/api/v1/boundary/peek"),
+        ("POST", "/api/v1/call-events/123e4567-e89b-42d3-a456-426614174000/delivery"),
+        ("GET", "/api/v1/context/search"),
+        ("POST", "/api/v1/memories"),
+    ],
+)
+def test_generic_host_rejects_all_non_recall_and_context_routes(method, path):
+    with pytest.raises(HTTPException) as error:
+        require_external_endpoint(
+            request(
+                ["om_recall", "om_conversations"],
+                scopes=("all-apps", "raw-imported-all"),
+                method=method,
+                path=path,
+                host_id="generic-local-mcp",
+            )
+        )
+    assert error.value.status_code == 403
+    assert error.value.detail == "external_endpoint_denied"
+
+
+@pytest.mark.parametrize("path", ["/api/v1/recall", "/api/v1/context/search"])
+def test_generic_host_allows_only_its_read_routes(path):
+    require_external_endpoint(
+        request(
+            ["om_recall", "om_conversations"],
+            scopes=("all-apps", "raw-imported-all"),
+            path=path,
+            host_id="generic-local-mcp",
+        )
+    )
+
+
 def test_legacy_recall_still_requires_its_own_tool():
     with pytest.raises(HTTPException):
         require_external_endpoint(request(["om_conversations"], path="/api/v1/recall"))

memhub/apps/core/tests/test_context_memory.py

@@ -74,8 +74,9 @@ def put(
 def revoke(db, source, *, user="local"):
     db.execute(
         """INSERT OR REPLACE INTO source_grants
-        (user_id,source_id,root_path,candidate_kind,granted_at,revoked_at)
-        VALUES (?,?,'/synthetic','default',?,?)""",
+        (user_id,source_id,root_path,requested_path,root_identity,candidate_kind,
+         granted_at,revoked_at,revision)
+        VALUES (?,?,'/synthetic','/synthetic','synthetic','default',?,?,1)""",
         (user, source, NOW, NOW),
     )
 

memhub/apps/core/tests/test_context_raw.py

@@ -46,7 +46,11 @@ def put(
 
 def revoke(db, source, *, user="local"):
     db.execute(
-        "INSERT INTO source_grants VALUES (?,?,'/synthetic','default','2026-09-19','revoked')",
+        "INSERT INTO source_grants "
+        "(user_id,source_id,root_path,requested_path,root_identity,candidate_kind,"
+        "granted_at,revoked_at,revision) "
+        "VALUES (?,?,'/synthetic','/synthetic','synthetic','default',"
+        "'2026-09-19','revoked',1)",
         (user, source),
     )
 
@@ -488,6 +492,45 @@ def test_final_revalidation_drops_changed_items_without_replacement(database, ch
     assert all(item in initial for item in final)
 
 
+def test_generic_finalizer_rechecks_each_result_against_current_positive_grant(database):
+    with closing(sqlite3.connect(database)) as db, db:
+        for source in ("cursor", "workbuddy", "missing", "revoked"):
+            put(db, source, f"needle {source}", source=source)
+        put(db, "foreign", "needle foreign", source="foreign", user="other")
+        db.executemany(
+            "INSERT INTO source_grants "
+            "(user_id,source_id,root_path,requested_path,root_identity,candidate_kind,"
+            "granted_at,revision) "
+            "VALUES (?,?,'/synthetic','/synthetic','synthetic','default','2026-09-19',1)",
+            (("local", "cursor"), ("local", "workbuddy"), ("other", "foreign")),
+        )
+        revoke(db, "revoked")
+    initial = read(database).items
+    assert {item.source_id for item in initial} == {"cursor", "workbuddy", "missing"}
+
+    def positive():
+        return revalidate_raw_context(
+            database,
+            initial,
+            source_ids=frozenset({"cursor", "workbuddy", "missing", "foreign", "revoked"}),
+            deadline=time.monotonic() + 5,
+            require_active_grant=True,
+        )
+
+    assert {item.source_id for item in positive()} == {"cursor", "workbuddy"}
+    with closing(sqlite3.connect(database)) as db, db:
+        db.execute("DELETE FROM source_grants WHERE user_id='local' AND source_id='cursor'")
+    assert {item.source_id for item in positive()} == {"workbuddy"}
+    with closing(sqlite3.connect(database)) as db, db:
+        db.execute(
+            "UPDATE source_grants SET revoked_at='2026-09-20' "
+            "WHERE user_id='local' AND source_id='workbuddy'"
+        )
+    assert positive() == ()
+    # Trusted internal reads retain their existing missing-grant semantics.
+    assert {item.source_id for item in revalidate(database, initial)} == {"cursor", "missing"}
+
+
 def test_neighbor_revision_change_drops_group_and_empty_authority_reads_nothing(
     database, monkeypatch
 ):

memhub/apps/core/tests/test_diagnostic_log.py

@@ -1,4 +1,5 @@
 import json
+import logging
 import os
 import sqlite3
 from datetime import UTC, datetime, timedelta
@@ -13,6 +14,7 @@ from one_memory.diagnostic_log import (
     configure_diagnostics,
     current_trace_id,
     database_error_code,
+    diagnostic_event,
     diagnostic_exception,
     diagnostics_enabled,
     diagnostics_requested,
@@ -23,6 +25,32 @@ from one_memory.diagnostic_log import (
 )
 
 
+def test_debug_events_are_collected_only_at_debug_level(tmp_path, monkeypatch) -> None:
+    monkeypatch.setenv("ONE_MEMORY_DIAGNOSTICS", "1")
+    root = logging.getLogger()
+    previous = root.level
+    root.setLevel(logging.INFO)
+    receipt = configure_diagnostics(tmp_path)
+    try:
+        diagnostic_event(
+            level="debug",
+            component="core.poll",
+            event="poll.completed",
+            outcome="succeeded",
+        )
+        diagnostic_event(
+            level="info",
+            component="memory.task",
+            event="memory_task.finished",
+            outcome="completed",
+        )
+    finally:
+        reset_diagnostics()
+        root.setLevel(previous)
+    events = [json.loads(line) for line in receipt.path.read_text().splitlines()]
+    assert [event["event"] for event in events] == ["memory_task.finished"]
+
+
 def test_high_frequency_success_events_are_summarized_but_warnings_are_not(
     tmp_path, monkeypatch
 ) -> None:
@@ -97,6 +125,53 @@ def test_arbitrary_detail_and_absolute_stack_are_rejected(tmp_path):
         )
 
 
+def test_a_rejected_payload_does_not_break_the_caller(tmp_path, monkeypatch) -> None:
+    """Rejection protects the log; it must not become the caller's outage.
+
+    Two of our own nodes shipped with unlisted `details` keys. Both sat on a
+    request path, so the rejection surfaced as a 500 to the person using the
+    app — including the sedimentation switch, whose every request failed
+    because logging the toggle raised before the endpoint could answer.
+    """
+    monkeypatch.setenv("ONE_MEMORY_DIAGNOSTICS", "1")
+    receipt = configure_diagnostics(tmp_path)
+    try:
+        diagnostic_event(
+            level="info",
+            component="memory.sedimentation",
+            event="memory_sedimentation.toggled",
+            outcome="succeeded",
+            details={"prompt": "private"},
+        )
+    finally:
+        reset_diagnostics()
+    # Rejected, so nothing is written — but the call returned.
+    assert not receipt.path.exists() or receipt.path.read_text() == ""
+
+
+def test_the_sedimentation_toggle_payload_is_recorded(tmp_path, monkeypatch) -> None:
+    """`enabled` is a bounded boolean and has to be a legal detail key.
+
+    Omitting it from the allow-list is what made the switch's own endpoint
+    answer 500 while still saving the new value.
+    """
+    monkeypatch.setenv("ONE_MEMORY_DIAGNOSTICS", "1")
+    receipt = configure_diagnostics(tmp_path)
+    try:
+        diagnostic_event(
+            level="info",
+            component="memory.sedimentation",
+            event="memory_sedimentation.toggled",
+            outcome="succeeded",
+            details={"enabled": False},
+        )
+    finally:
+        reset_diagnostics()
+    events = [json.loads(line) for line in receipt.path.read_text().splitlines()]
+    assert [event["event"] for event in events] == ["memory_sedimentation.toggled"]
+    assert events[0]["details"]["enabled"] is False
+
+
 def test_exception_diagnostic_keeps_code_location_not_message():
     secret = "private-provider-response-918273"
     log = DiagnosticLog(Path.cwd())

memhub/apps/core/tests/test_diagnostic_payload_keys.py

@@ -16,16 +16,15 @@ import pytest
 
 from one_memory.diagnostic_log import _DETAIL_KEYS
 
-SOURCES = (
-    "api.py",
-    "imports/async_bridge.py",
-    "imports/live_sync.py",
-    "imports/runner.py",
-    "memory_processing.py",
-)
-
 ROOT = Path(__file__).resolve().parent.parent / "src" / "one_memory"
 
+#: Every module, not a hand-picked list. The two payloads that caused real
+#: outages were in files nobody thought to add here, so the only shape of this
+#: check that can actually hold the line is "every module, always".
+SOURCES = tuple(
+    sorted(str(path.relative_to(ROOT)) for path in ROOT.rglob("*.py") if path.name != "__init__.py")
+)
+
 
 def _literal_calls(path: Path):
     tree = ast.parse(path.read_text())

memhub/apps/core/tests/test_diagnostic_trace.py

@@ -162,8 +162,6 @@ def test_import_request_trace_reaches_the_background_job_without_original_text(t
     ]
     names = [event["event"] for event in events]
     assert "import.request.started" in names
-    assert "import.queued" in names
-    assert "import.started" in names
     assert "import.finished" in names
     finished = next(event for event in events if event["event"] == "import.finished")
     assert {
@@ -175,18 +173,17 @@ def test_import_request_trace_reaches_the_background_job_without_original_text(t
     assert secret not in written
 
 
-def test_import_run_logs_parse_stages_and_core_resources_without_content(tmp_path):
+def test_default_info_omits_parser_and_resource_internals(tmp_path):
     written, job_id = _run_codex_import(tmp_path)
     events = list(map(json.loads, written.splitlines()))
     job_events = [event for event in events if event.get("job_id") == job_id]
     names = [event["event"] for event in job_events]
-    assert names.index("import.parse.started") < names.index("import.parse.completed")
-    assert "import.phase_changed" in names
-    completed = next(event for event in job_events if event["event"] == "import.parse.completed")
-    assert completed["outcome"] == "succeeded"
-    assert {"stored_count", "skipped_count", "error_count"} <= completed["details"].keys()
+    assert "import.parse.started" not in names
+    assert "import.parse.completed" not in names
+    assert "import.phase_changed" not in names
+    assert "import.finished" in names
     everything = [event["event"] for event in events]
-    assert "core.resource_snapshot" in everything
+    assert "core.resource_snapshot" not in everything
     assert everything.index("core.ready") < everything.index("core.stopped")
     assert IMPORT_SECRET not in written
 
@@ -230,4 +227,5 @@ def test_a_failing_import_finished_log_does_not_escape_the_supervisor_task(tmp_p
 
     monkeypatch.setattr(supervisor, "diagnostic_event", rejecting)
     written, _job_id = _run_codex_import(tmp_path)
-    assert "import.started" in written
+    assert "core.stopped" in written
+    assert "import.finished" not in written

memhub/apps/core/tests/test_diagnostic_trace_api.py

@@ -33,6 +33,42 @@ def test_trace_lookup_is_disabled_until_detailed_logging_is_enabled(tmp_path):
     assert response.json() == {"schema_version": 1, "state": "disabled", "events": []}
 
 
+def test_model_usage_route_returns_content_free_totals(tmp_path):
+    app = create_app(
+        Settings(
+            _env_file=None,
+            data_dir=tmp_path,
+            source_home=tmp_path / "home",
+            api_token="trace-token",
+            imports_enabled=False,
+            automatic_memory_processing=False,
+            live_sync_enabled=False,
+        ),
+        source_catalog={},
+    )
+    with TestClient(app, headers={"Authorization": "Bearer trace-token"}) as client:
+        response = client.get("/api/v1/diagnostics/model-usage")
+    assert response.status_code == 200
+    assert response.json() == {
+        "schema_version": 1,
+        "summary": {
+            "request_count": 0,
+            "succeeded": 0,
+            "failed": 0,
+            "in_flight": 0,
+            "prompt_tokens": 0,
+            "completion_tokens": 0,
+            "total_tokens": 0,
+            "cached_prompt_tokens": 0,
+            "reasoning_tokens": 0,
+            "avg_prompt_tokens": 0,
+            "avg_completion_tokens": 0,
+        },
+        "requests": [],
+        "by_minute": [],
+    }
+
+
 def test_trace_lookup_returns_only_safe_structured_steps(tmp_path, monkeypatch):
     monkeypatch.setenv("ONE_MEMORY_DIAGNOSTICS", "1")
     set_detailed_logging(tmp_path, True)

memhub/apps/core/tests/test_evidence_store.py

@@ -2,6 +2,7 @@ import json
 import sqlite3
 from concurrent.futures import ThreadPoolExecutor
 from datetime import UTC, datetime
+from hashlib import sha256
 
 import pytest
 from pydantic import ValidationError
@@ -14,6 +15,7 @@ from one_memory.imports.models import (
     WrittenBy,
 )
 from one_memory.imports.storage import EvidenceStore
+from one_memory.memory_tasks import evidence_ref, matches_evidence_ref
 from one_memory.workspace_resolution import KnownWorkspace
 
 
@@ -25,6 +27,7 @@ def evidence(
     verbatim: str = "Remember the source, not the guess.",
     origin: EvidenceOrigin = EvidenceOrigin.FILE,
     written_by: WrittenBy = WrittenBy.PERSON,
+    role: str = "user",
     source_ordinal: str | None = None,
 ) -> EvidenceInput:
     return EvidenceInput(
@@ -35,7 +38,7 @@ def evidence(
         source_ordinal=source_ordinal
         if source_ordinal is not None
         else ("line:1" if message_id is None and correlation_id is None else None),
-        role="user",
+        role=role,
         written_by=written_by,
         created_at=datetime(2026, 8, 31, 12, 0, tzinfo=UTC),
         verbatim=verbatim,
@@ -87,7 +90,7 @@ def test_rescanning_an_unchanged_file_does_not_write_the_message_again(tmp_path)
     """A rescan is idempotent because the identity is derived from the content.
 
     `EvidenceInput.idempotency_key` is a hash of the source, conversation,
-    author, ordinal and content hash, and `evidence_id` is a uuid5 over that
+    protocol role and content hash, and `evidence_id` is a uuid5 over that
     key, so replaying an unchanged file reproduces the same primary key and
     `ON CONFLICT` absorbs it. Verified: removing the anchoring logic still
     leaves this green, so this pins the contract, not one implementation.
@@ -102,6 +105,87 @@ def test_rescanning_an_unchanged_file_does_not_write_the_message_again(tmp_path)
     assert store.count() == 1
 
 
+def test_derived_identity_uses_role_not_host_author_guess(tmp_path) -> None:
+    store = EvidenceStore(tmp_path / "one-memory.sqlite3")
+    first = store.put(evidence(message_id=None, written_by=WrittenBy.UNKNOWN))
+    replay = store.put(evidence(message_id=None, written_by=WrittenBy.PERSON))
+    assistant = store.put(evidence(message_id=None, role="assistant", written_by=WrittenBy.AGENT))
+
+    assert replay.record.evidence_id == first.record.evidence_id
+    assert replay.was_created is False
+    assert assistant.record.evidence_id != first.record.evidence_id
+    assert store.count() == 2
+
+
+def test_task_proof_ignores_legacy_written_by_metadata(tmp_path) -> None:
+    store = EvidenceStore(tmp_path / "one-memory.sqlite3")
+    record = store.put(evidence()).record
+    proof = evidence_ref(record)
+
+    changed_author = record.model_copy(update={"written_by": WrittenBy.AGENT})
+    legacy_identity = record.model_dump(
+        mode="json",
+        exclude={"origin_meta", "imported_at", "updated_at"},
+    )
+    legacy_proof = {
+        **proof,
+        "identity_hash": sha256(
+            json.dumps(legacy_identity, sort_keys=True).encode()
+        ).hexdigest(),
+    }
+
+    assert matches_evidence_ref(changed_author, proof) is True
+    assert matches_evidence_ref(changed_author, legacy_proof) is True
+
+
+def test_rescan_reuses_a_row_stored_under_written_by_derived_v3(tmp_path) -> None:
+    database = tmp_path / "one-memory.sqlite3"
+    store = EvidenceStore(database)
+    item = evidence(message_id=None, written_by=WrittenBy.UNKNOWN)
+    original = store.put(item).record
+    with sqlite3.connect(database) as connection, connection:
+        connection.execute(
+            "UPDATE evidence SET idempotency_key=? WHERE evidence_id=?",
+            (item.written_by_derived_v3_idempotency_key, original.evidence_id),
+        )
+
+    replay = store.put(evidence(message_id=None, written_by=WrittenBy.PERSON))
+
+    assert replay.was_created is False
+    assert replay.record.evidence_id == original.evidence_id
+    assert store.count() == 1
+
+
+def test_legacy_identity_role_change_revises_the_existing_evidence(tmp_path) -> None:
+    database = tmp_path / "one-memory.sqlite3"
+    store = EvidenceStore(database)
+    old = evidence(
+        message_id=None,
+        role="assistant",
+        written_by=WrittenBy.AGENT,
+    )
+    original = store.put(old).record
+    with sqlite3.connect(database) as connection, connection:
+        connection.execute(
+            "UPDATE evidence SET idempotency_key=? WHERE evidence_id=?",
+            (old.written_by_derived_v3_idempotency_key, original.evidence_id),
+        )
+
+    replay = store.put(
+        evidence(
+            message_id=None,
+            role="user",
+            written_by=WrittenBy.AGENT,
+        )
+    )
+
+    assert replay.record.evidence_id == original.evidence_id
+    assert replay.was_revised is True
+    assert replay.record.revision == original.revision + 1
+    assert replay.record.role == "user"
+    assert store.count() == 1
+
+
 def test_changed_text_at_one_position_is_a_new_message_not_a_silent_edit(tmp_path) -> None:
     """An edited line becomes new evidence, never an overwrite of the old one.
 
@@ -146,6 +230,7 @@ def test_personal_candidate_prefilter_excludes_unrelated_user_turns(tmp_path) ->
         evidence(
             message_id="assistant",
             verbatim="我偏好简洁文案。",
+            role="assistant",
             written_by=WrittenBy.AGENT,
         )
     )

memhub/apps/core/tests/test_hy_memory_bridge.py

@@ -310,6 +310,7 @@ def test_ollama_profile_ignores_ambient_remote_storage_and_models(
     monkeypatch.setenv("MEMORY_VECTOR_HOST", "remote.example.com")
     monkeypatch.setenv("MEMORY_COLLECTION_NAME", "hostile-collection")
     monkeypatch.setenv("MEMORY_EMBEDDING_DIMS", "99")
+    monkeypatch.delenv("MEMORY_CODING_TOOL_RESULT_MAX_BYTES", raising=False)
     profile = snapshot(
         ProviderKind.OLLAMA,
         base_url="http://127.0.0.1:11434",
@@ -333,6 +334,20 @@ def test_ollama_profile_ignores_ambient_remote_storage_and_models(
     assert config.vector_store.embedding_dims == 768
     assert config.graph_store.provider == "kuzu"
     assert str(tmp_path) in config.graph_store.db_path
+    assert config.coding.tool_result_max_bytes == 512
+
+
+def test_memhub_coding_tool_result_ceiling_is_configurable(tmp_path, monkeypatch) -> None:
+    monkeypatch.setenv("MEMORY_CODING_TOOL_RESULT_MAX_BYTES", "1024")
+    profile = snapshot(
+        ProviderKind.OLLAMA,
+        base_url="http://127.0.0.1:11434",
+        has_api_key=False,
+    )
+
+    config = build_hy_memory_config(profile, api_key=None, data_dir=tmp_path)
+
+    assert config.coding.tool_result_max_bytes == 1024
 
 
 def _kuzu_graph_with_embedding_width(db_path: Path, width: int) -> None:
@@ -1326,7 +1341,7 @@ def test_hy_runtime_matches_public_sdk_without_reranker_or_service_backends(
     config = build_hy_memory_config(profile, api_key=None, data_dir=tmp_path)
 
     assert HY_MEMORY_DISTRIBUTION == "hy-memory"
-    assert HY_MEMORY_VERSION == "2.0.2"
+    assert HY_MEMORY_VERSION == "2.0.4"
     assert HY_MEMORY_MODE == "ultra"
     assert HY_MEMORY_READER == "hybrid"
     assert os.environ["MEMORY_MODE"] == "ultra"

memhub/apps/core/tests/test_hy_runtime_config.py

@@ -5,6 +5,7 @@ Checking that contract automatically prevents a silently flipped rerank flag
 from going unnoticed.
 """
 
+import logging
 import re
 from pathlib import Path
 
@@ -16,6 +17,7 @@ from one_memory.hy_runtime_config import (
     HY_MEMORY_MODE,
     HY_MEMORY_READER,
     HY_MEMORY_VERSION,
+    apply_diagnostic_log_level,
     hy_memory_version,
 )
 
@@ -69,3 +71,19 @@ def test_the_pin_matches_the_lock_file_and_its_wheel():
     )
     assert f"hy_memory-{HY_MEMORY_VERSION}-" in lock
     assert f'specifier = "=={HY_MEMORY_VERSION}"' in lock
+
+
+def test_info_level_suppresses_noisy_third_party_success_logs():
+    names = ("", "one_memory", "httpx", "httpcore", "watchfiles", "jieba")
+    previous = {name: logging.getLogger(name).level for name in names}
+    try:
+        apply_diagnostic_log_level("info")
+        assert logging.getLogger("one_memory").level == logging.INFO
+        for name in ("httpx", "httpcore", "watchfiles", "jieba"):
+            assert logging.getLogger(name).level == logging.WARNING
+        apply_diagnostic_log_level("debug")
+        for name in ("one_memory", "httpx", "httpcore", "watchfiles", "jieba"):
+            assert logging.getLogger(name).level == logging.DEBUG
+    finally:
+        for name, level in previous.items():
+            logging.getLogger(name).setLevel(level)