* Comment alongside _AUTH_FINGERPRINT_VOLATILE_KEYS notes client_secret
is treated as rotation-only on purpose, not a model-cache
differentiator (maintainer §Concerns 3).
* _write_auth docstring at tests/test_issue_t16551f61_auth_token_churn_
fingerprint.py:108 now describes actual behavior (writes + monkeypatch,
no sleep+restat) — Copilot r3302471228.
Refs #2242. PR #2964 review.
auth.json is rewritten by credential-pool/OAuth token refresh roughly every
14 minutes. _models_cache_source_fingerprint() hashed it via mtime/size
(#1699 _models_cache_file_fingerprint), so every token refresh churned the
fingerprint and the 24h /api/models cache was effectively dead -- the hot
GET /api/session?resolve_model=1 path paid a cold ~11.5s rebuild every few
minutes (RCA t_d127953d residual #2, t_16551f61).
Add _auth_store_semantic_fingerprint(): content-hash auth.json with a
DENY-list of known credential-rotation-only keys (access/refresh token,
expiry, per-credential status/telemetry, request_count, save updated_at)
stripped. Deny-list (not allow-list) is deliberate -- any unknown field, or
a real provider/endpoint/model-set change (active_provider, a new
credential_pool entry, base_url, source, label, auth_type, the providers{}
block, ...) stays in the fingerprint and still correctly busts the cache.
Conservative fallbacks: missing file -> marked; unreadable/corrupt ->
stat-based fallback (never less safe than pre-fix). config.yaml keeps the
cheap stat fingerprint (deliberate edits, no timer churn).
Bidirectional invariant regression test (non-tautological -- the
end-to-end churn test flips RED when the auth_json axis is reverted to
stat-based): token-only churn keeps fingerprint byte-identical AND keeps a
valid disk cache loadable; active_provider change / new credential_pool
entry / changed base_url each flip the fingerprint AND reject the stale
disk cache. Measured: 5/5 cold rebuilds per 5 refresh cycles -> 0/5.
Tests: 9 new pass; 28 adjacent (#1699/#1633/display-resolver) pass;
54 models_cache/fingerprint suite pass.
Two functions on the /api/session/handoff-summary hot path were opening
sqlite3.connect(...) inside a bare `with` statement, which commits the
transaction at scope exit but does NOT close the connection. Per-turn
invocations accumulated state.db / state.db-wal file descriptors and
CPython heap pages on long-lived worker threads, surfacing as the
multi-GB VmRSS and 6x duplicated state.db fds observed on the live
instance (D0 pre-restart baseline: VmRSS 1,334,248 kB, 55 fds; cold
baseline after restart: VmRSS 136,668 kB, 10 fds).
Wrap both call sites with contextlib.closing(...) (already imported and
used at seven other sites in the same files) so the connection is
closed deterministically:
- api/models.py :: count_conversation_rounds
- api/routes.py :: _persist_handoff_summary_to_state_db
Regression test:
tests/test_issue2233_sqlite_connection_leak.py loops both functions
20 times against a tmp state.db and asserts /proc/<pid>/fd count
does not grow more than 2. Linux-only via sys.platform skip.
D1 live soak against a freshly-built worktree server (port 8799,
isolated HERMES_HOME / HERMES_WEBUI_STATE_DIR) hitting
/api/session/handoff-summary 20 times:
fd_before = 5
fd_after = 5 (growth 0, threshold < 5)
vmrss_before = 52636 kB
vmrss_after = 52636 kB (growth 0 kB, threshold < 30 MB)
The patched fix curve trends below the leak curve.
Rollback: single git revert <this-sha> reverts both file edits.
Refs #2233.
Isla-Liu