Map Anthropic reasoning effort by provider

[xeophon]

Summary

• Map sampling_args.reasoning_effort to extra_body.verbosity for Anthropic models on OpenRouter/Prime OpenAI-compatible routes.
• Map reasoning_effort to native Anthropic output_config.effort and enable adaptive thinking for Claude 4.6/4.7 when unset.
• Add focused client tests for Claude 4.6/4.7 OpenRouter-style and native Anthropic mappings.

Verification

• uv run ruff check --fix verifiers/clients/openai_chat_completions_client.py verifiers/clients/anthropic_messages_client.py tests/test_client_multimodal_types.py
• uv run pytest tests/test_client_multimodal_types.py
• Pre-push hooks: ruff check, ruff format, ty (ci parity)
• Live Prime GPQA 1x1 checks on anthropic/claude-opus-4.7:
  • reasoning_effort="high": https://app.primeintellect.ai/dashboard/evaluations/rbej739yowcedihefu7sqykk
  • reasoning_effort="xhigh": https://app.primeintellect.ai/dashboard/evaluations/gl5n8e3cv0fcesfhjwgm1bcz
  • reasoning_effort="max": https://app.primeintellect.ai/dashboard/evaluations/ieaij5d5dpnjtcmlycby2wjg

---

Note

Medium Risk
Touches request parameter normalization for both Anthropic and OpenAI chat clients, which can change model behavior and cost/latency characteristics across providers. Risk is mitigated by being gated to reasoning_effort presence and specific Claude/OpenRouter-style routes.

Overview
Adds provider-specific translation of sampling_args.reasoning_effort in both clients.

For native Anthropic messages, it now maps reasoning_effort to output_config.effort and auto-enables thinking={"type":"adaptive"} for select Claude 4.6/4.7 models when thinking isn’t already set.

For OpenAI-compatible chat completions routed to Anthropic via OpenRouter/Pinference, it maps reasoning_effort into extra_body.verbosity (and ensures extra_body.reasoning.enabled) while leaving other models/providers to use the standard reasoning_effort field.

^{Reviewed by Cursor Bugbot for commit a4c1fc7. Bugbot is set up for automated code reviews on this repo. Configure here.}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: a4c1fc71bb

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Restrict verbosity mapping to Claude 4.6/4.7

When this client targets OpenRouter/Pinference with Claude models that still use budget-based reasoning (for example Claude 3.7/4.5), this broad model_id.startswith("claude-") branch moves reasoning_effort into verbosity and only sends reasoning.enabled, so a requested high/xhigh reasoning effort becomes OpenRouter’s default reasoning instead. OpenRouter’s Claude 4.6 migration docs distinguish verbosity from reasoning.effort and say older models continue to use reasoning.effort, so this should be gated to the 4.6/4.7 adaptive-thinking model IDs rather than every Claude model.

Useful? React with 👍 / 👎.