Summary

When an auxiliary task is configured with provider: main and no explicit auxiliary model, Hermes resolves the provider correctly but does not inherit the live session model. Instead, it falls through to the provider's baked-in auxiliary default (for Codex this is gpt-5.2-codex).

This makes session compression pick the wrong model by default and can trigger bogus threshold auto-lowering warnings even though the active agent model already has the needed context window.

Reproduction

Use config like:

auxiliary:
  compression:
    provider: main
    model: ''

Run a session with a live model such as gpt-5.4 on openai-codex.

Compression feasibility checks and compression calls resolve through provider: main, but the blank model does not pick up the live session model. It falls back to Codex's hardcoded auxiliary default.

Actual behavior

Compression resolves to gpt-5.2-codex instead of the active session model.

Expected behavior

When provider: main is selected and no explicit auxiliary model override is configured, Hermes should use:

1. the live session model
2. then the configured main/default model
3. only then a provider-specific fallback

Root cause

main_runtime is available to callers like compression, but _resolve_task_provider_model() does not accept or use it. So provider resolution can follow main, while model resolution still falls through to provider defaults.

Proposed fix

• Thread main_runtime into _resolve_task_provider_model()
• If requested provider is main and no explicit task model is set, inherit main_runtime['model']
• Add regression coverage for both:
  • _resolve_task_provider_model() precedence
  • get_text_auxiliary_client() / call_llm() forwarding main_runtime

Verification

Local regression tests added and passing:

• tests/agent/test_auxiliary_client.py
• full file run passed after the change (86 passed locally)