Add reasoning content normalization middleware

[snimu]

Description

Quick fix for missing reasoning content in GLM 4.7-Flash.

Normalize reasoning from provider-specific fields (reasoning_content, reasoning) into tags in content, and strip them on outbound requests to providers that use separate reasoning fields.

• New verifiers/utils/reasoning_utils.py with pure utility functions
• Inbound: extract reasoning from responses, prepend as tags
• Outbound: strip tags from assistant messages for providers that use reasoning_content/reasoning fields
• Auto-detect reasoning format from first response
• Accept reasoning-only responses in validation (no EmptyModelResponseError)

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Documentation update
• Test improvement

Testing

• All existing tests pass when running uv run pytest locally.
• New tests have been added to cover the changes

Checklist

• My code follows the style guidelines of this project as outlined in AGENTS.md
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• Any dependent changes have been merged and published

---

Note

Medium Risk
Touches core request/response handling and validation for model calls; behavior changes depend on provider response shape and could affect prompts or error handling if detection/stripping misfires.

Overview
Normalizes provider-specific reasoning fields into a consistent <think>...</think> prefix on inbound chat responses, so models that return reasoning_content/reasoning no longer drop reasoning.

Adds reasoning_format handling to Environment: strips <think> blocks from outbound assistant messages for providers that expect separate reasoning fields, auto-detects the format from the first response, and relaxes response validation to accept reasoning-only replies. Includes a new reasoning_utils module and comprehensive unit/integration tests covering extraction, normalization, stripping, and provider message prep.

^{Written by Cursor Bugbot for commit fcf95fa. This will update automatically on new commits. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{Bugbot Autofix is OFF. To automatically fix reported issues with Cloud Agents, enable Autofix in the Cursor dashboard.}

[cursor]

Think tags not stripped for interleaved rollout path

Medium Severity

prepare_messages_for_provider is called in get_model_response_with_messages to strip <think> tags from assistant messages before sending to providers that use separate reasoning fields, but get_model_response_with_tokens (used when interleaved_rollouts=True and trajectory is non-empty) sends prompt directly as messages without any such stripping. In multi-turn interleaved rollouts with providers using reasoning_content/reasoning, assistant messages will still contain <think> tags the provider doesn't expect.

Additional Locations (1)

• verifiers/envs/environment.py#L693-L704