fix: sanitize tool schemas for strict mode

[greysonlalonde]

Summary

Pydantic schemas intermittently fail strict tool-use on openai, anthropic, and bedrock. All three reject nested objects missing additionalProperties: false, and anthropic also rejects minLength/maximum/top-level anyOf. Adds per-provider sanitizers wired into each provider's tool conversion.

Gemini and azure are follow-ups.

Verification

Ran the original bug-report schema against real APIs:

Provider	Raw + strict	Sanitized + strict
Bedrock Sonnet 4.5	additionalProperties must be explicitly set to false	ok, all fields populated
Anthropic native	same 400	ok, all fields populated
OpenAI gpt-4o	same 400	ok, chat_id: null preserved

Test plan

• Manual repro on real bedrock, anthropic, openai
• Unit tests for the sanitizers

---

Note

Medium Risk
Changes tool schema generation in strict mode across three providers, which can affect tool-calling behavior and may surface new validation edge cases for existing tool definitions.

Overview
Fixes intermittent strict tool-use failures by sanitizing tool JSON schemas when strict is enabled for OpenAI, Anthropic, and Bedrock.

Adds provider-specific sanitizers in pydantic_schema_utils that inline refs, force additionalProperties: false, require all properties, and strip unsupported/metadata keys (plus Anthropic/Bedrock-specific removals like top-level anyOf and numeric/string constraints). Providers now apply these sanitizers during tool conversion instead of passing raw Pydantic schemas.

^{Reviewed by Cursor Bugbot for commit e3fb1b8. Bugbot is set up for automated code reviews on this repo. Configure here.}

[iris-clawd]

This is the proper fix for the strict mode issue I flagged on #5388. Clean architecture:

Common pipeline (_common_strict_pipeline): resolve refs → strip $defs → oneOf→anyOf → ensure types → additionalProperties: false → require all properties → strip metadata — covers what all three providers need.

Per-provider sanitizers:

• OpenAI: common + strip unsupported formats
• Anthropic: common + lift_top_level_anyof (Claude rejects top-level unions) + strip Claude-specific keywords (numeric/string constraints, pattern, etc.)
• Bedrock: delegates to Anthropic (same grammar compiler) — correct and DRY

Integration:

• All three providers now conditionally sanitize only when strict is enabled — non-strict tools pass through unchanged
• OpenAI's old force_additional_properties_false replaced with the fuller pipeline
• Lazy imports moved to module-level (cleaner)

Verified against real Bedrock Sonnet 4.5, Anthropic, and OpenAI APIs. Unit tests noted as TODO — worth adding but not blocking.

LGTM 🚀 💬 221

[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.}

^{Reviewed by Cursor Bugbot for commit e3fb1b8. Configure here.}