hermes - 💡(How to fix) Fix [Bug]: Fix: stale title generation request reloads unloaded Ollama model after model switch [1 participants]

When using Ollama with strict_single_load (only one model in memory at a time),
the background auto-title generation thread can cause a previously unloaded model
to be reloaded into memory.

Reproduction steps

1. Start a session with model A (e.g. Qwen3.6-27B-oQ8-fp16)
2. Hermes completes the first response and spawns a background thread for
   auto-title generation (captures main_runtime pointing to model A)
3. User immediately sends the next prompt, triggering a model switch to model B
4. Ollama's strict_single_load unloads model A
5. The background title thread still holds the old main_runtime snapshot and
   sends an HTTP request for model A → Ollama reloads model A into memory

Observed symptoms


[ollama-router] Switched general -> Qwen:Qwen3.6-27B-oQ8-fp16 (reason: default fast daily route; strategy: strict_single_load)
⚠ Auxiliary title generation failed: HTTP 400: Error code: 400 - {'error': 'Model unloaded.'}
[ollama-router] Switched general -> Qwen:Qwen3.6-27B-oQ8-fp16 (model reloaded again)


Both models end up in memory, defeating the purpose of strict_single_load and
wasting GPU VRAM.

Root cause

maybe_auto_title() is called with a snapshot of the current main_runtime
(model name, provider, base_url). The background thread holds this snapshot and
uses it unconditionally — it has no way to detect that the user has already
switched to a different model before its LLM call executes.

Fix

Add an optional runtime_validator callback to the title generation pipeline
(generate_title → auto_title_session → maybe_auto_title). All callers
(cli.py, gateway/run.py, tui_gateway/server.py, acp_adapter/server.py) now
capture the current model/provider at call time and pass a validator that checks
whether the live agent state still matches. If the model was switched, the
background thread skips the request silently — no HTTP call is made, so no
stale model reloads are triggered.

Changes

- agent/title_generator.py — new runtime_validator parameter on
  generate_title(), auto_title_session(), and maybe_auto_title();
  validator is checked before the LLM call
- cli.py — captures _title_model / _title_provider and passes validator
- gateway/run.py — same pattern for gateway callers
- tui_gateway/server.py — same pattern; also adds missing failure_callback
  and main_runtime for parity with CLI
- acp_adapter/server.py — same pattern; also adds missing parameters for parity
- tests/agent/test_title_generator.py — 3 new tests covering the validator,
  plus updated existing assertions for the new parameter

Backward compatibility

All new parameters have default values of None. Callers that don't pass
runtime_validator get the original behavior unchanged. No breaking changes.

Testing

bash
python -m pytest tests/agent/test_title_generator.py -v  # 22 passed
python -m pytest tests/test_tui_gateway_server.py -k "title" -v  # 13 passed

Fix: stale title generation request reloads unloaded Ollama model after model switch