feat(config): accept STRIX_REASONING_EFFORT=max for providers that support it

[devin-ai-integration]

Summary

STRIX_REASONING_EFFORT=max failed config validation, so DeepSeek V4's top reasoning tier was unreachable — ReasoningEffort mirrored the OpenAI SDK's enum, which stops at xhigh.

Adding "max" to the literal is not enough: Reasoning(effort="max") is itself rejected by the OpenAI SDK, and on the native deepseek/ route LiteLLM's DeepSeek transform pops reasoning_effort and substitutes thinking: {"type": "enabled"}, collapsing every level to the provider default. So max skips the SDK field and rides along as a raw body field, passed through verbatim — whether the model supports it is the user's responsibility (unsupported providers reject the request):

if effort != "max":
    return ModelSettings(reasoning=Reasoning(effort=effort))
return ModelSettings(extra_args={..., "extra_body": {"reasoning_effort": "max"}})

Verified on the wire:

STRIX_LLM	effort	request body
openrouter/deepseek/deepseek-v4-flash	max	"reasoning_effort": "max" (live completion succeeded)
deepseek/deepseek-v4-flash	max	"reasoning_effort": "max" (was "thinking": {"type": "enabled"})

The Codex/ChatGPT-subscription clamp gains a max → high case alongside the existing xhigh → high one, since that backend only accepts low/medium/high.

Link to Devin session: https://app.devin.ai/sessions/1e2f7b6ab4de438ea9d6e516b57eb96c
Requested by: @0xallam

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring

[greptile-apps]

Greptile Summary

The PR adds max as a reasoning-effort configuration value and serializes it according to the selected backend.

• Sends max as a raw request-body field for LiteLLM-routed providers.
• Clamps max to xhigh for OpenAI routes and to high for ChatGPT subscription models.
• Updates validation guidance, documentation, and focused model-settings tests.

Confidence Score: 5/5

The PR appears safe to merge with no concrete changed-code failure identified.

The new value is translated before reaching SDK enums that reject it, provider-route handling is explicit, subscription models use a complete clamp map, and focused tests cover the principal LiteLLM and OpenAI paths.

Important Files Changed

Filename	Overview
strix/core/inputs.py	Adds backend-aware serialization for maximum reasoning effort while preserving timeout and prompt-cache settings.
strix/config/models.py	Adds LiteLLM route detection and a total reasoning-effort translation map for subscription-backed Codex requests.
strix/config/settings.py	Extends validated reasoning-effort configuration with the new max value.
tests/test_inputs.py	Covers raw max forwarding on DeepSeek and clamping on OpenAI routes.
docs/advanced/configuration.mdx	Documents the new effort level and its provider-dependent behavior.

_{Reviews (1): Last reviewed commit: "feat(config): accept STRIX_REASONING_EFF..." | Re-trigger Greptile}

[devin-ai-integration]

Tested end-to-end on the real LLM request path (StrixProvider → LiteLLM, wire bodies captured via a CustomLogger printing complete_input_dict).

openrouter/deepseek/deepseek-v4-flash + STRIX_REASONING_EFFORT=max — live completion succeeded (OpenRouter 400s invalid efforts, so success is server-side validation):

WIRE: {"model": "deepseek/deepseek-v4-flash", ..., "reasoning_effort": "max"}
OUTPUT: ...ResponseOutputText(text='OK')...   # response_id gen-1785628516-jp3zZQvTjn5VQ2yDKywR

high and xhigh also pass through verbatim and complete.

Native deepseek/ route: fix vs main

This branch, max (dummy key, expected 401 after send):

WIRE: {"model": "deepseek-v4-flash", ..., "reasoning_effort": "max"}   # no "thinking" key

origin/main, high — the reported bug (LiteLLM pops the effort and substitutes thinking):

WIRE: {"model": "deepseek-v4-flash", ..., "thinking": {"type": "enabled"}}

Other checks

• STRIX_REASONING_EFFORT=max loads cleanly; bogus still raises the pydantic literal error; on main, max is rejected.
• openai/gpt-5.4 + max → Reasoning(effort='xhigh'), no extra_body.
• uv run pytest: 1 failed / 653 passed — sole failure is the pre-existing test_finish_scan_bypasses_active_agent_guard_after_reserve. make check-all fails only on the pre-existing PLC0415 in strix/interface/viewer/transcript.py (both confirmed on clean main).

[devin-ai-integration]

Re-tested after the simplification (384c5aa — route detection dropped, max always sent as raw body field) with live OpenRouter completions on the real StrixProvider path:

STRIX_REASONING_EFFORT=max → WIRE: {"model": "deepseek/deepseek-v4-flash", ..., "reasoning_effort": "max"}
→ completion succeeded (gen-1785629260-SZtWPzKXuEUD3VIC51z7)

xhigh and high also pass through verbatim and complete (no regression). OpenRouter rejects invalid efforts with a 400, so success is server-side validation. tests/test_inputs.py: 32 passed; ruff check/format clean on changed files.