fix(memhub): keep long raw hits from clearing search

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

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

@@ -1,6 +1,7 @@
 """Authenticated SDK context search route, without generation or reindex side effects."""
 
 import asyncio
+import logging
 import time
 from collections.abc import Sequence
 from datetime import UTC, datetime
@@ -36,6 +37,8 @@ from one_memory.semantic_deadline import SemanticMemoryHit
 from one_memory.user_scope import owns
 
 CALL_EVENT_RESPONSE_BUDGET_MS = 250
+SDK_CONTEXT_CONTENT_LIMIT = 20_000
+logger = logging.getLogger(__name__)
 
 #: Pass-through filters the raw lane honours. Scene, layers, reader and the
 #: schema/working floors describe memory lanes; the raw lane is always the
@@ -90,6 +93,13 @@ def hit_namespace(hit: SemanticMemoryHit) -> dict[str, str]:
     return {}
 
 
+def bounded_context_content(content: str) -> tuple[str, bool]:
+    """Fit SDK text into its wire contract before Pydantic validates it."""
+    if len(content) <= SDK_CONTEXT_CONTENT_LIMIT:
+        return content, False
+    return content[:SDK_CONTEXT_CONTENT_LIMIT], True
+
+
 def sdk_context_router(dependencies: Sequence[Depends]) -> APIRouter:
     """SDK-native context: no Evidence/Governance ids or fabricated revisions."""
     router = APIRouter(prefix="/api/v2/context", dependencies=dependencies)
@@ -170,34 +180,39 @@ def sdk_context_router(dependencies: Sequence[Depends]) -> APIRouter:
                 )
                 turns = raw_display_turns(hit) if value.raw_unit == "turn" else ()
                 if not turns:
+                    content, truncated = bounded_context_content(hit.content)
                     primary.append(
                         SdkRawContextItem(
                             raw_id=hit.raw_id,
-                            content=hit.content,
+                            content=content,
                             source=hit.source,
                             session_id=hit.session_id,
                             memory_at=memory_at,
                             score=hit.score,
+                            truncated=truncated,
                         )
                     )
                     continue
                 # The embedding is the batch's: the hit is shown by the turns
                 # that contain the query, or else by its closest turn, each
                 # with the hit's score.
-                shown = [
-                    SdkRawContextItem(
-                        raw_id=hit.raw_id,
-                        turn_id=turn.turn_id,
-                        turn_index=turn.turn_index,
-                        turn_count=turn.turn_count,
-                        content=turn.content,
-                        source=hit.source,
-                        session_id=hit.session_id,
-                        memory_at=memory_at,
-                        score=hit.score,
+                shown: list[SdkRawContextItem] = []
+                for turn in turns_for_hit(turns, value.query):
+                    content, truncated = bounded_context_content(turn.content)
+                    shown.append(
+                        SdkRawContextItem(
+                            raw_id=hit.raw_id,
+                            turn_id=turn.turn_id,
+                            turn_index=turn.turn_index,
+                            turn_count=turn.turn_count,
+                            content=content,
+                            source=hit.source,
+                            session_id=hit.session_id,
+                            memory_at=memory_at,
+                            score=hit.score,
+                            truncated=truncated,
+                        )
                     )
-                    for turn in turns_for_hit(turns, value.query)
-                ]
                 primary.append(shown[0])
                 siblings.extend(shown[1:])
             chosen = primary[: value.limit]
@@ -250,16 +265,15 @@ def sdk_context_router(dependencies: Sequence[Depends]) -> APIRouter:
                     or not owns(memory.user_id, value.user_id)
                 ):
                     continue
+                content, truncated = bounded_context_content(
+                    hit.formatted_content
+                    if hit.sdk_formatted and hit.formatting_complete and hit.formatted_content
+                    else memory.content
+                )
                 items.append(
                     SdkMemoryContextItem(
                         memory_id=memory.memory_id,
-                        content=(
-                            hit.formatted_content
-                            if hit.sdk_formatted
-                            and hit.formatting_complete
-                            and hit.formatted_content
-                            else memory.content
-                        ),
+                        content=content,
                         layer=memory.layer,
                         status=memory.status,
                         source=memory.source,
@@ -269,6 +283,7 @@ def sdk_context_router(dependencies: Sequence[Depends]) -> APIRouter:
                             else None
                         ),
                         score=hit.score,
+                        truncated=truncated,
                     )
                 )
             return tuple(items)
@@ -318,7 +333,23 @@ def sdk_context_router(dependencies: Sequence[Depends]) -> APIRouter:
             if isinstance(lane, EmbeddingUnavailableError):
                 # No lexical fallback: the query could not be embedded.
                 raise HTTPException(503, EMBEDDING_UNAVAILABLE)
-        partial = isinstance(raw_items, Exception) or isinstance(memory_items, Exception)
+        requested_lanes = {
+            "raw": raw_items,
+            "memory": memory_items,
+        }
+        failed_lanes = {
+            name: lane
+            for name, lane in requested_lanes.items()
+            if name in value.kinds and isinstance(lane, Exception)
+        }
+        for name, error in failed_lanes.items():
+            logger.warning(
+                "sdk_context_lane_failed lane=%s error_class=%s",
+                name,
+                type(error).__name__,
+            )
+        all_requested_failed = len(failed_lanes) == len(value.kinds)
+        partial = bool(failed_lanes)
         raw_candidates = (
             () if isinstance(raw_items, Exception) else cast(tuple[SdkContextItem, ...], raw_items)
         )
@@ -347,11 +378,20 @@ def sdk_context_router(dependencies: Sequence[Depends]) -> APIRouter:
             if len(content) > remaining:
                 item = item.model_copy(update={"content": content[:remaining], "truncated": True})
                 truncated = True
+            truncated = truncated or item.truncated
             selected.append(item)
             used += len(item.content)
         final = SdkContextSearchResponse(
             request_id=str(uuid4()),
-            state=("partial" if partial else "grounded" if selected else "empty"),
+            state=(
+                "unavailable"
+                if all_requested_failed
+                else "partial"
+                if partial
+                else "grounded"
+                if selected
+                else "empty"
+            ),
             items=tuple(selected),
             used_characters=used,
             truncated=truncated,

memhub/apps/core/tests/test_sdk_api_surface.py

@@ -469,6 +469,101 @@ def test_lite_raw_search_is_plain_rag_over_active_and_shadow_raws(tmp_path, monk
     run_with_sdk(tmp_path, monkeypatch, scenario)
 
 
+@pytest.mark.fresh_sdk
+def test_long_semantic_raw_hit_cannot_erase_an_earlier_match(tmp_path, monkeypatch) -> None:
+    async def scenario(app, sdk, client):
+        app.state.memory_mode = lambda: "lite"
+
+        def search(query, **kwargs):
+            sdk._record("search", {"query": query, **kwargs})
+            memories = [
+                {
+                    "memory_id": "raw-target",
+                    "content": "[user]: lite 模式的 3d 图需要调整",
+                    "layer": "l1_raw",
+                    "status": "active",
+                    "user_id": "local",
+                    "agent_id": "all-apps",
+                    "session_id": "target-session",
+                    "source": "cursor",
+                    "memory_at": 1_790_100_000,
+                    "score": 0.97,
+                },
+                {
+                    "memory_id": "raw-long",
+                    "content": f"[user]: {'x' * 20_001}",
+                    "layer": "l1_raw",
+                    "status": "active",
+                    "user_id": "local",
+                    "agent_id": "all-apps",
+                    "session_id": "long-session",
+                    "source": "cursor",
+                    "memory_at": 1_790_000_000,
+                    "score": 0.9,
+                },
+            ]
+            return {
+                "request_id": "long-raw",
+                "memories": memories[1:] if query == "oversized" else memories,
+            }
+
+        sdk.search = search
+        response = await client.post(
+            "/api/v2/context/search",
+            json={
+                "query": "3d",
+                "kinds": ["raw"],
+                "raw_unit": "turn",
+                "limit": 20,
+                "character_budget": 20_000,
+            },
+        )
+
+        assert response.status_code == 200, response.text
+        body = response.json()
+        assert body["state"] == "grounded"
+        assert body["items"][0]["raw_id"] == "raw-target"
+        assert "3d" in body["items"][0]["content"]
+        assert body["items"][1]["raw_id"] == "raw-long"
+        assert body["items"][1]["truncated"] is True
+        assert body["truncated"] is True
+
+        single = await client.post(
+            "/api/v2/context/search",
+            json={
+                "query": "oversized",
+                "kinds": ["raw"],
+                "raw_unit": "turn",
+                "character_budget": 20_000,
+            },
+        )
+        assert single.json()["items"][0]["truncated"] is True
+        assert single.json()["truncated"] is True
+
+    run_with_sdk(tmp_path, monkeypatch, scenario)
+
+
+@pytest.mark.fresh_sdk
+def test_raw_only_lane_failure_is_unavailable_not_an_empty_partial(tmp_path, monkeypatch) -> None:
+    async def scenario(app, sdk, client):
+        app.state.memory_mode = lambda: "lite"
+
+        def search(_query, **_kwargs):
+            raise ValueError("private conversation text must not reach logs")
+
+        sdk.search = search
+        response = await client.post(
+            "/api/v2/context/search",
+            json={"query": "3d", "kinds": ["raw"], "raw_unit": "turn"},
+        )
+
+        assert response.status_code == 200, response.text
+        assert response.json()["state"] == "unavailable"
+        assert response.json()["items"] == []
+
+    run_with_sdk(tmp_path, monkeypatch, scenario)
+
+
 @pytest.mark.fresh_sdk
 def test_every_recall_and_ask_option_reaches_the_sdk(tmp_path, monkeypatch) -> None:
     async def scenario(app, sdk, client):

memhub/apps/web/src/api.raw.test.ts

@@ -41,7 +41,7 @@ it("lists raw conversations for the selected application", async () => {
 });
 
 it("keeps the application filter when searching raw conversations", async () => {
-  post.mockResolvedValue({ data: { items: [] } });
+  post.mockResolvedValue({ data: { state: "empty", items: [] } });
   const signal = new AbortController().signal;
 
   await searchSdkRaw("release plan", { source: "claude-code", signal });
@@ -60,6 +60,14 @@ it("keeps the application filter when searching raw conversations", async () =>
   );
 });
 
+it("does not present an unavailable raw search as no matches", async () => {
+  post.mockResolvedValue({
+    data: { state: "unavailable", items: [] },
+  });
+
+  await expect(searchSdkRaw("3d")).rejects.toThrow("raw_search_unavailable");
+});
+
 it("pages the independent lightweight raw graph", async () => {
   get.mockResolvedValue({
     data: {

memhub/apps/web/src/api.ts

@@ -1052,6 +1052,8 @@ export async function searchSdkRaw(
     },
     { signal: options.signal, timeout: 10_000 },
   );
+  if (response.data?.state !== "grounded" && response.data?.state !== "empty")
+    throw new Error("raw_search_unavailable");
   const items = response.data?.items;
   return Array.isArray(items)
     ? items.filter(