fix(streaming): respect auxiliary.title_generation config for session titles

[nesquena-hermes]

Summary

Fixes title generation ignoring auxiliary.title_generation config. When a user configures a dedicated title model in config.yaml, it was silently overridden by the chat session's model.

Changes

api/streaming.py:

• _aux_title_configured() — detects whether auxiliary.title_generation has any meaningful config (provider ≠ auto, model, or base_url set)
• _aux_title_timeout() — reads per-task timeout from config; previously hardcoded to 15.0 s
• _generate_llm_session_title_via_aux() — adds use_agent_model kwarg to preserve existing behavior when no dedicated config exists
• _run_background_title_update() — when aux title is configured, routes directly to aux; adds missing llm_invalid_aux fallback path

tests/test_title_aux_routing.py — 23 new tests covering all branches.

Test results

23/23 pass. Full suite: 2088 passing, 0 failed.

Co-authored by @starship-s. Replaces PR #925 (branch had been behind master).

[nesquena]

Review — end-to-end ✅ (clean, no fixes needed)

Traced against upstream hermes-agent

Fresh nousresearch/hermes-agent tarball pulled. Walked the whole routing chain:

_resolve_task_provider_model at agent/auxiliary_client.py:2439-2489 — priority order is:

1. Explicit base_url arg → forces "custom" provider with that endpoint
2. Explicit provider arg → wins over task config
3. cfg_base_url (from auxiliary.<task>.base_url)
4. cfg_provider (if not "auto")
5. Auto-detection fallback

The PR relies on the semantics that passing None for provider/model/base_url lets the resolver read them from config.yaml's auxiliary.title_generation section. Line 2472: resolved_model = model or cfg_model — explicit model wins, otherwise config.

The PR correctly handles this with provider or None / model or None / base_url or None at api/streaming.py:305-307, converting empty strings (the default when use_agent_model=False) to None so the resolver falls through to config-based resolution.

_get_auxiliary_task_config at auxiliary_client.py:2495-2506 — returns {} on ImportError; always a dict. Matches webui _aux_title_configured's safety wrapper.

call_llm at auxiliary_client.py:2693-2735 — accepts explicit timeout: float = None (None → reads from _get_task_timeout(task)). When the webui passes timeout=_timeout (a concrete float), that explicit value wins. ✓

End-to-end trace of the four fixes

#1 _aux_title_configured() (streaming.py:236-246). Returns True when any of provider (non-auto), model, or base_url is meaningfully set. Three or-gated conditions. Exception-safe (returns False on ImportError). Correctly case-insensitive on provider via .lower() != 'auto'. ✓

#2 _aux_title_timeout() (streaming.py:248-270). Reads auxiliary.title_generation.timeout, coerces to float, rejects None / non-numeric / zero / negative — falls back to default (15.0s) with a debug log. Slightly stricter than upstream _get_task_timeout (which accepts zero/negative). ✓

#3 use_agent_model kwarg on _generate_llm_session_title_via_aux (streaming.py:429-456).

• use_agent_model=False (new default): pass provider='', model='', base_url='' → provider or None → None → upstream resolver reads from config
• use_agent_model=True (legacy fallback): pass agent's attrs → explicit args win, config ignored — matches pre-PR behavior for the no-aux-config case

Correctly preserves the previous "try agent model first" behavior when auxiliary.title_generation isn't configured.

#4 Routing change in _run_background_title_update (streaming.py:578-588).

Case matrix traced:

State	Route tried	Fallback
No aux config + agent present	agent-model title	aux with agent's model (legacy)
Aux configured + agent present	aux via config	agent-model title
No agent (regardless of aux config)	aux via config	(none — agent unavailable)

Also adds llm_invalid_aux to the fallback trigger set (was only llm_error_aux before). Now both llm_error_aux (network/API error) and llm_invalid_aux (response didn't parse to a valid title) correctly trigger the agent-model fallback. ✓

Cross-tool (CLI) check

auxiliary.title_generation is a shared config.yaml section. The PR reads the SAME keys the hermes-agent itself reads for its own title generation via _get_auxiliary_task_config('title_generation'). Both tools converge on the same interpretation:

• provider='openai', model='gpt-4o-mini' → both use OpenAI gpt-4o-mini for title generation
• base_url='http://local-llm:8080/v1' → both route through that endpoint

No cross-tool divergence. The PR actually improves cross-tool consistency — before, webui ignored the auxiliary config while the CLI honored it. Now both honor it. ✓

Security audit

No new user input surface, no new CDN deps, no auth-related changes. The new code reads config.yaml (a trusted file on disk) and passes primitive values (string, float) to call_llm. No injection vectors. ✓

Edge-case trace

Scenario	Behaviour
Fresh install, no auxiliary.title_generation in config	_aux_title_configured() returns False → agent-model path wins (same as master) ✅
auxiliary.title_generation: {provider: openai, model: gpt-4o-mini}	Aux route fires → config.yaml values resolve via upstream _resolve_task_provider_model ✅
auxiliary.title_generation: {provider: auto}	_aux_title_configured() returns False (auto = not meaningful) → agent-model path. Intentional — auto means "use default routing" ✅
auxiliary.title_generation: {base_url: http://localhost:1234/v1}	Returns True (base_url is set) → aux route uses that endpoint as "custom" per upstream line 2483 ✅
auxiliary.title_generation: {provider: openai} (model blank)	Returns True → aux route uses openai with provider-default model ✅
auxiliary.title_generation: {timeout: 30} (int)	_aux_title_timeout() coerces to 30.0 ✅
auxiliary.title_generation: {timeout: 0} or -1 or 'abc' or missing	Falls back to 15.0 default ✅
Aux route returns llm_error_aux (API error)	Agent fallback triggers ✅
Aux route returns llm_invalid_aux (title didn't parse)	Agent fallback triggers ✅ (was missing before this PR)
Aux route succeeds	Agent route NOT called ✅
agent.auxiliary_client import fails	_aux_title_configured() returns False → agent path; call_llm also fails with ImportError at call-site → caught by existing except in generate_title_raw_via_aux ✅
Concurrent writes to config.yaml during title generation	Stale read possible but acceptable — title gen is a background task; worst case is using previous config for one turn
use_agent_model=True (legacy fallback)	Passes agent's provider/model/base_url as explicit overrides → they win per upstream priority 1 → exact pre-PR behavior preserved ✅

Tests

• 23/23 pass in new tests/test_title_aux_routing.py
• Full local suite: 2041 passed, 47 skipped, 0 failed

Test coverage is thorough — _aux_title_configured tested across 10 config shapes (model-only, base_url-only, provider-only, auto case, empty, provider+blank-model, etc.); timeout-in-config verified via a fake call_llm that captures the kwarg; edge cases for zero/negative/non-numeric/empty-string; llm_error_aux and llm_invalid_aux fallback cases both covered via mocked session fixture + event capture.

Other audit — things that are correct already

• Upstream semantics verified: I confirmed _resolve_task_provider_model at auxiliary_client.py:2439-2489 treats None / empty provider-or-base_url as "fall through to task config" — exactly what the PR relies on. ✓
• Timeout plumbing: webui resolves timeout locally (stricter semantics than upstream's _get_task_timeout) and passes the concrete value to call_llm. Upstream uses the explicit value unchanged. ✓
• Exception safety: both helpers (_aux_title_configured, _aux_title_timeout) wrap in try/except and fall back to safe defaults. ✓
• Case-insensitive auto check: provider.lower() != 'auto' covers 'auto', 'Auto', 'AUTO', etc. ✓
• Agent lock: the session s.title write at streaming.py:592 is already wrapped in _get_session_agent_lock(session_id) from PR #910. The PR just removed some dead code comments around that section but the locking semantics are preserved. ✓
• Agent path unchanged: _generate_llm_session_title_for_agent and the status codes llm_error / llm_invalid are unchanged — the PR only routes around them. ✓

Minor observations (non-blocking)

• Duplicated timeout logic: _aux_title_timeout (stricter: rejects zero/negative) duplicates upstream _get_task_timeout (accepts any numeric). Could be simplified to from agent.auxiliary_client import _get_task_timeout; _get_task_timeout('title_generation', default=15.0) and just pass timeout=None to call_llm to let upstream read it natively. Trade-off: webui's stricter semantics are lost. Non-urgent — the duplication is clearly documented and the tests lock the behavior.

• Default drift: webui default = 15s, upstream _DEFAULT_AUX_TIMEOUT = 30s. Intentional — titles should be quick. Non-issue.

• use_agent_model is a keyword-only arg (*, use_agent_model: bool = False) — good practice; can't be passed positionally by accident.

• _aux_title_configured() is called once per title update, which is fine, but each call re-imports agent.auxiliary_client and re-parses config.yaml. For high-volume title updates the load_config() call might show up in profiles. Not urgent — it's a background path and upstream likely caches the config.

Recommendation

Clean, correctly scoped fix for a real config-ignored bug. End-to-end trace against the upstream _resolve_task_provider_model semantics confirms the provider or None pattern correctly defers to config.yaml when no explicit args are passed. Tests cover all the branches the routing change touches. No cross-tool or security concerns. Nothing to push back on.

✅ Approved. Ready for merge + v0.50.182 tag.