fix: use provider-agnostic schema for non-OpenAI tool conversion (#4472)

[devin-ai-integration]

fix: use provider-agnostic schema for non-OpenAI tool conversion (#4472)

Summary

extract_tool_info() in common.py called generate_model_description() when converting a tool's args_schema to parameters. That function applies OpenAI strict-mode transformations (additionalProperties: false, forcing all properties required, title on every object) that Gemini and Bedrock reject.

This PR adds generate_tool_parameters_schema() — a new provider-agnostic path that:

• Resolves $ref / $defs
• Strips anyOf null unions (for Optional fields)
• Recursively removes title, default, and additionalProperties
• Preserves the original required array (does not force every property required)

extract_tool_info() now calls this new function instead of generate_model_description() for the args_schema fallback path.

Files changed:

• lib/crewai/src/crewai/utilities/pydantic_schema_utils.py — added _strip_schema_metadata() and generate_tool_parameters_schema()
• lib/crewai/src/crewai/llms/providers/utils/common.py — switched extract_tool_info to use new function
• lib/crewai/tests/utilities/test_tool_schema_compatibility.py — 24 new tests

Review & Testing Checklist for Human

• Verify this is the correct code path for MCP tools: The fix targets the args_schema fallback in extract_tool_info() (line ~82 of common.py). This path is only hit when the tool dict is in "direct format" (no "function" key), has no inline "parameters", and does have an "args_schema". Confirm MCP tools actually flow through this branch and not the "function" or "parameters" branches above it.
• Stripping default values: The helper removes default from all properties recursively. Verify this doesn't cause issues for providers that can usefully consume defaults (e.g., OpenAI or Anthropic direct format). Note: this code path is only reached for the args_schema fallback, not for tools that already supply inline parameters.
• _strip_schema_metadata mutates in-place: The helper does schema.pop(key) on the passed dict. It's safe in generate_tool_parameters_schema (fresh dict from model_json_schema), but since it's a module-level function that tests import directly, consider whether it should defensively copy.
• End-to-end manual test: Set up an MCP tool (e.g., the firewall tool from the issue) and verify it works with gemini-2.5-flash and Bedrock Claude — this is the scenario from the bug report and cannot be fully validated by unit tests alone.

Notes

• The OpenAI schema path (generate_model_description) is unchanged and verified by TestOpenAISchemaUnchanged tests.
• All CI checks pass (lint, type-checker, tests on Python 3.10–3.13).
• Link to Devin run: https://app.devin.ai/sessions/3219aa033fef47e4b86891b65ab8b9f5
• Requested by: João

[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 and CI monitoring