fix(llm): make reasoning_effort model-aware (denylist + retry-on-400)

[ropoctl]

Supersedes #47 (closed) — re-ported onto the current refactored native.py.

Problem

AsyncLLMClient defaulted reasoning_effort="medium" and sent it on every OpenAI-compat request unconditionally. Non-reasoning models served over OpenAI-compat endpoints reject the param with HTTP 400:

400 Bad Request: `reasoning_effort` is not supported with this model

This kills the sourcehunt hunter loop + ranker against Llama 3.x/4, Mistral, Mixtral, Qwen2/2.5, Gemma on Groq/Together/Fireworks/etc. The default Anthropic/Claude path was never affected.

Fix — two layers, both in clearwing/llm/native.py

1. Model-name denylist with a sentinel default. The constructor default becomes "auto"; _auto_resolve_reasoning_effort() maps known non-reasoning families (_REASONING_EFFORT_UNSUPPORTED_PATTERNS) to None and everything else — including Claude — to "medium". So the default path is byte-for-byte unchanged for Anthropic. Any explicit value (including None) bypasses auto-resolution.
2. Retry-on-400 safety net. If a provider still rejects the param at call time, achat/achat_stream detect that specific 400 (_is_unsupported_reasoning_effort_error: must mention reasoning_effort and 400/unsupported), drop it, latch reasoning_effort=None for the session, and retry once. Unrelated 400s/429s/500s and the existing OpenAI aiohttp stream fallback are preserved.

Notes on the re-port

native.py was refactored since #47's branch (centralized provider routing, codex/aiohttp streaming fallback), so the original patch no longer applied. Re-implemented against current main:

• achat retry wraps the _with_rate_limit_retries call.
• achat_stream retry is layered into the existing except alongside (not replacing) the OpenAI HTTP fallback, via a small _consume(opts) helper.

Original author credited via Co-authored-by.

Verification

.venv/bin/ruff check clearwing/llm/native.py tests/test_native_reasoning_effort.py   # clean
.venv/bin/pytest tests/test_native_reasoning_effort.py tests/test_llm_openai_fallback.py -q   # 30 passed

Tests from #47 carried over verbatim (denylist resolution, constructor wiring, error classification, options rebuild, and achat/achat_stream retry-once behavior).

[msyed-01]

Thanks for the fast and thorough re-port @ropoctl, just saw this and realizing I missed the rebase window. You were fast lol. The two-layer shape is preserved exactly as designed: denylist + sentinel-default for Layer 1, retry-once with instance-state latch for Layer 2. Putting the streaming retry before the existing aiohttp fallback in achat_stream is cleaner than what I had-- nice touch. co-author credit appreciated!!

One related item that also surfaced during the end-to-end smoke test against Groq Llama 3.3 70B: the Ranker hits a second 400, this time from response_format: json_schema (since Groq Llama 3.3 doesn't support OpenAI Structured Outputs). Same shape as this fix and as #51. Was going to open a separate issue but new issues are restricted in the repo. Want me to open a Discussion for it, or somewhere else you prefer? Happy to send a follow-up PR mirroring the pattern here once this lands, if thats welcome.

[ropoctl]

Sorry, I probably would have waited a bit longer for your response on this. btw groq llama supports year of our lord 2026 features https://console.groq.com/docs/responses-api https://console.groq.com/docs/structured-outputs

[msyed-01]

lmao ur so right, my bad, kept it from the reasoning_effort repro (just ended up with it cuz its free tier and hits that bug). moving to gpt-oss / llama-4-scout for json_schema runs. thnx again chief

[ropoctl]

ah does free tier not allow V1 Responses (openai_resp) api?

[msyed-01]

honestly never tested it-- was just running openai_compat on groq, never tried the openai_resp path.

[msyed-01]

actually just tested it, free tier allows /v1/responses on both llama-3.3-70b and gpt-oss-120b. we just hadnt tried the openai_resp path.

also, something I noticed: chat-completions says reasoning_effort in the 400 but /v1/responses says reasoning.effort (with a dot). i think the _is_unsupported_reasoning_effort_error substring check in #65 might not fire on Responses API errors