Open items¶
Known limitations and pending decisions, so they are not rediscovered. Items leave this page when they are fixed or rejected, with the reason recorded in the changelog.
Known limitations¶
| Priority | Item | Detail |
|---|---|---|
| P2 | Large acquire can starve |
A large request waits behind a stream of small ones. A fair queue would change what available and try_acquire mean. See rate limiting. |
| P2 | Sanitization needs arguments | InputSanitizationHook only sees tool arguments passed through args= or hook_ctx, while the Article 15 self-check item does not reflect that. |
| P2 | Audit truncation after a restart | A restarted sink cannot detect a deleted tail without an externally stored expected_tip. See the audit log guide. |
| P2 | Purge markers without a key | Without a signing key a purge marker proves nothing to someone who can write to the database, and it weakens expected_tip. Use a signing key where that matters. |
| P2 | Event data copying | AgentEvent.with_data is slow on deeply nested tuples and frozensets, and rejects non-string keys, before a sink is reached. |
| P2 | Slow __repr__ |
Audit coercion contains exceptions from an object's __repr__, but not the time it takes. |
| P2 | Oversight replay window | Replay protection is per process and covers the latest 65,536 ids. A retry after a failed audit write may get a different answer from the review channel; the log shows both. |
| P2 | Homoglyphs in tool names | A Cyrillic letter inside delete is a different name, so a deny list misses it. Use allowed_tools for tools that matter. |
| P2 | Spaced status codes | A bare number counts as a status, so took 401 ms classifies as a provider failure, while 401ms does not. |
| P2 | Replay scan depth | Turn replay looks at the latest 100 checkpoints of a thread. |
| P2 | Redis recipe | The recipe still scores by created_at, so a clock step can reorder it. The PostgreSQL recipe's new ordering has not been run against a real server. |
| P2 | Recovery events | recovery.succeeded means the recipe ran, not that the call worked, so it also appears for a failed attempt. |
| P2 | MCP closed session | On SDK 1.x a closed session raises the transport's ClosedResourceError rather than MCPToolError. |
| P2 | Binary scanning | An escape sequence hidden in mostly unreadable bytes past the first 64 is not detected, because scanning all binary would reject ordinary files. |
| P2 | Audit lock wait | busy_timeout_seconds blocks the caller, and so the event loop for an async session, while waiting. |
| P2 | Release-candidate guards | The production-rc-* guards still measure the 0.3.0 snapshot, so they only run on those branches. |