bug: approve_or_deny_turn bypasses per-thread lock — race with concurrent chat

[G9000]

Summary

approve_or_deny_turn (service.py:177-317) performs sequence allocation, result persistence, compaction, companion history refresh, and post-turn hooks without acquiring the per-thread asyncio.Lock that _execute_agent_turn uses.

Failure Scenario

1. User sends a chat message to thread X → acquires get_thread_lock(X)
2. Simultaneously, user approves a pending tool call on the same thread X → approve_or_deny_turn runs with no lock
3. Both paths call reserve_message_sequences, persist_agent_result, compact_thread_context, and _refresh_companion_history concurrently
4. Result: sequence ID collisions, corrupted companion history cache, duplicated or interleaved messages

Fix

Wrap the body of approve_or_deny_turn in async with get_thread_lock(run.thread_id): after resolving the checkpoint and determining the thread ID.

Files

• apps/server/src/anima_server/services/agent/service.py:177-317

Severity

High — data corruption under concurrent approval + chat on same thread.