Stop default openmemory_query from dropping the HTTP session - #212
Stop default openmemory_query from dropping the HTTP session#212vincenzopalazzo wants to merge 1 commit into
Conversation
Default openmemory_query runs HSG with no timeout, so a hung embed or vector writeback dropped the Streamable HTTP POST. Catch HSG failures as MCP tool errors, time out after 12s, and cap unpatterned factual dumps at k. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 70a92566a7
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| facts = await asyncio.wait_for( | ||
| query_facts_at_time( |
There was a problem hiding this comment.
When a factual query encounters a slow or locked SQLite database, this timeout cannot fire: query_facts_at_time() contains no suspension point and calls the synchronous db.fetchall() directly, so it blocks the same event loop that must schedule asyncio.wait_for's timer. Consequently OM_MCP_QUERY_TIMEOUT_MS remains ineffective for factual searches and the Python MCP session can still hang beyond the configured deadline; offload the database operation to a thread or make the query genuinely asynchronous.
Useful? React with 👍 / 👎.
| except asyncio.TimeoutError: | ||
| msg = f"HSG contextual search timed out after {int(timeout_s * 1000)}ms" | ||
| if qtype == "contextual": | ||
| return [TextContent(type="text", text=f"openmemory_query failed: {msg}")] |
There was a problem hiding this comment.
When the Python contextual query times out, returning a plain list of TextContent is wrapped by the low-level MCP server as a successful tool result (isError remains false). Clients therefore cannot distinguish this failure from valid output and may proceed instead of retrying or reporting a tool failure, unlike the JavaScript path added here; return an error-bearing CallToolResult or propagate an exception that the MCP server converts to isError.
Useful? React with 👍 / 👎.
1 participant
Summary
openmemory_query(type=contextual) ranhsg_querywith no try/catch or timeout. A hung embed, rate-limit throw, or long vector writeback dropped the per-request Streamable HTTP POST (Streamable HTTP error: Error POSTing to endpoint, empty status), so the first query looked like a total outage.OM_MCP_QUERY_TIMEOUT_MS) and return a normal MCP tool error (isError+ readable message) instead of letting the transport die. Unified queries fail-soft on HSG and still return facts.k(patterned queries still cap at 32). Same fail-soft + cap on the Python MCP path.Test plan
npx vitest run tests/mcp_query_failsoft.test.ts(7 tests)npm testinpackages/openmemory-js(65 tests)npm run typecheckopenmemory_querywith notypereturns a tool result orisError, never a transport droplist/get/store/delete/type=factualwithfact_patternstill work