hermes - 💡(How to fix) Fix feat(config): support per-model reasoning_effort defaults via custom_providers [1 pull requests]

Problem

reasoning_effort is currently a flat global setting — it can only be set at agent.reasoning_effort (main agent) or delegation.reasoning_effort (subagents). There is no mechanism to set a different default reasoning effort per model.

This is a gap because:

• Different models have very different reasoning characteristics. Some models (e.g. claude-opus-4.6, gpt-5.5, deepseek-v4-pro) benefit from xhigh reasoning effort, while others (e.g. gemini-flash, deepseek-v4-flash) should run at medium or none.
• Switching models via /model does not change the reasoning effort — the global setting persists, which can mean expensive reasoning effort on a cheap/fast model, or insufficient reasoning on a premium model.
• The custom_providers mechanism already supports per-provider overrides for context_length and arbitrary extra_body fields, but reasoning_effort bypasses this entirely and always resolves to the global value.

Proposed Solution

Extend the custom_providers config schema to support a top-level reasoning_effort key at both the provider level and (optionally) the per-model override level:

model:
  default: openrouter/anthropic/claude-sonnet-4-6
  provider: openrouter

custom_providers:
  - name: openrouter
    base_url: https://openrouter.ai/api/v1
    api_key: ${OPENROUTER_API_KEY}
    reasoning_effort: high          # default for all models from this provider

  - name: deepseek
    base_url: https://api.deepseek.com
    api_key: ${DEEPSEEK_API_KEY}
    models:
      deepseek-v4-pro:
        reasoning_effort: xhigh
      deepseek-v4-flash:
        reasoning_effort: none

The resolution chain should be:

1. Per-model override (if custom_providers[N].models.<model>.reasoning_effort is set)
2. Provider-level default (if custom_providers[N].reasoning_effort is set)
3. Global agent.reasoning_effort (existing fallback)

The existing delegation.reasoning_effort remains unchanged — it only governs subagent defaults, not the main agent.

Alternatives Considered

Model Presets (#20249) — that proposal is complementary (per-turn model escalation) rather than a static per-model default. This request is narrower: a static default reasoning effort per model that persists across all sessions and turns.

Per-session /reasoning override — exists today and is useful for one-off changes, but requires manual toggling every time the model changes and does not survive session boundaries.

References

• Bug: #20576 (thinking_token_budget not forwarded to vLLM/custom providers — related to incomplete custom_providers reasoning plumbing)
• PR: #18737 (propagating custom_providers thinking config to DeepSeek API — same pattern as this request)
• Feature: #20249 (Model Presets — complementary, not overlapping)

Labels

area/config, comp/agent, type/feature