[wilke]

Why

Every refiner call used the owning Duet's single route. On mango (vLLM Qwen3.8-27B) the best reasoning level (xhigh) can only be reached by omitting reasoning_effort. vLLM accepts only low/medium on the wire. Measured on one hard prompt: low 6.4k tokens in 140 s; medium and default (xhigh) hit the 12k cap in about 280 s. Deep reasoning pays off for one Designer or emitter call, but it is too costly on every step of a coding loop. Operators also want to send coding work to a different model, such as Qwen3-Coder-Next.

What

openchia.refiner_roles in the Hermes config:

openchia:
  refiner_roles:
    designer:    {reasoning: default}   # omit reasoning_effort -> server default
    emitter:     {reasoning: default}   # module transcription (#81)
    coder:       {custom_provider: Mango Coder}   # Implementer coding session
    implementer: {reasoning: low}

• Fields: custom_provider (a name from custom_providers), model, and reasoning, which takes default, none, or a level.
• Role keys:
  • Episode local IDs; designer also covers launch.
  • Two call kinds: emitter (set by a role_scope context var in emit_modules) and coder (via DuetEpisodeBinding.for_role).
• Lookup order: scoped call kind, then the Episode local ID, then the Duet route.
• Overrides are resolved and validated once, when the binding is frozen at /build. They are stored in the binding record without secrets, so /build continue keeps them, and the recovery policy is recomputed for each route. Keys are read at call time from the provider's key_env.
• Each attempt receipt now records refiner_role.
• /code (agent/workflow_coding.py) is unchanged.

Tests

• New tests/agent/test_refiner_role_routes.py (11 tests) covers:
  • route resolution;
  • rejection of invalid overrides;
  • lookup order;
  • key sourcing;
  • the coder view from for_role;
  • transport dispatch with refiner_role in the receipts;
  • default reasoning leaving reasoning_effort out of the request.
• On the related suites, this branch shows the same failures and errors as origin/main (6 failed, 30 errors in tests/iterative_episode_refiner/*). The errors all come from the assignment-v1 contract fixtures and are not caused by this change. Ruff is clean.

🤖 Generated with Claude Code