mirror of
https://github.com/NousResearch/hermes-agent.git
synced 2026-06-09 08:21:50 +00:00
19db9cd076
session.py _persist() bypassed SessionDB's thread-safe write path by
accessing private internals db._lock and db._conn directly:
with db._lock:
db._conn.execute("UPDATE sessions SET model_config = ? ...")
db._conn.commit()
This was fragile for three reasons:
1. It bypassed _execute_write()'s BEGIN IMMEDIATE + jitter-retry logic,
so concurrent writes could hit SQLite BUSY without retrying.
2. It called db._conn.commit() manually, breaking the transactional
contract that _execute_write() enforces.
3. Any internal rename of _lock or _conn would silently break this
call site with an AttributeError at runtime.
Fix:
- Add SessionDB.update_session_meta(session_id, model_config_json, model)
to hermes_state.py. Routes through _execute_write() for the standard
BEGIN IMMEDIATE + lock + jitter-retry guarantee. Uses COALESCE so
passing model=None leaves the stored model column unchanged.
- Replace the db._lock / db._conn block in session.py _persist() with
a single db.update_session_meta() call.
Tests (tests/acp/test_session_db_private_access.py, 11 tests):
- Unit tests for update_session_meta: updates model_config, updates
model, preserves existing model on None, routes through _execute_write,
no-op on non-existent session.
- AST checks: db._lock and db._conn not referenced in session.py;
_persist() calls update_session_meta().
- Integration round-trips: cwd and model persisted correctly; COALESCE
prevents overwriting an existing model with NULL.
|
||
|---|---|---|
| .. | ||
| __init__.py | ||
| test_approval_isolation.py | ||
| test_auth.py | ||
| test_edit_approval.py | ||
| test_entry.py | ||
| test_events.py | ||
| test_mcp_e2e.py | ||
| test_permissions.py | ||
| test_ping_suppression.py | ||
| test_registry_manifest.py | ||
| test_server.py | ||
| test_session.py | ||
| test_session_db_private_access.py | ||
| test_tools.py | ||