[哈基米 (BlueX888)]

Summary

MCPToolWrapper (the external HTTPS URL resolution path used by mcp/tool_resolver.py::_resolve_external) returned the response text unconditionally from _execute_tool and never inspected the isError flag an MCP server sets on an otherwise successful tools/call response. The error text reached the framework as a plain str, so the call was recorded as a success: no ToolFailureDetectedEvent and TaskOutput.has_tool_failures stayed false. MCPNativeTool already handles the same server answer by returning ToolFailure(reason=MCP_ERROR) (lib/crewai/src/crewai/tools/mcp_native_tool.py:136-146).

Changes

_execute_tool now checks result.isError after extracting the content and returns ToolFailure(message=content, reason=ToolFailureReason.MCP_ERROR, details={"server": ..., "tool": ...}), matching MCPNativeTool. Signatures along the retry path are widened to str | ToolFailure. Successful calls are unchanged.

Rebased onto main 38db11e. The conflict with #8300 is resolved by keeping that PR's content_blocks_to_text(result.content) extraction and wrapping it in the isError check (lib/crewai/src/crewai/tools/mcp_tool_wrapper.py:194-208), so multi-block results and error reporting both hold.

Verification

• Tests added or updated for the changed behavior

• Relevant tests and quality checks pass locally

• python -m pytest lib/crewai/tests/tools/test_mcp_tool_wrapper.py lib/crewai/tests/mcp/test_tool_result_content.py -q -p no:randomly — 9 passed

• python -m pytest lib/crewai/tests/tools/ lib/crewai/tests/mcp/ -q -p no:randomly — 347 passed

• With the source change reverted and the new tests kept, test_is_error_becomes_a_tool_failure fails with AssertionError: assert False + where False = isinstance('Error: file not found', ToolFailure), confirming the old code reported the server error as a success

• ruff check lib/crewai/src/crewai/tools/mcp_tool_wrapper.py lib/crewai/tests/tools/test_mcp_tool_wrapper.py — All checks passed

• mypy lib/crewai/src/crewai/tools/mcp_tool_wrapper.py — Success: no issues found in 1 source file

Additional context

This PR was authored by an AI coding agent. Per .github/CONTRIBUTING.md it should carry the llm-generated label; I don't have permission to apply labels, so a maintainer would need to add it.

[coderabbitai]

Important

Review skipped

Review was skipped as selected files did not have any reviewable changes.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Advanced

Run ID: d264cd3b-9600-4ff6-91a7-c66e4000b7f4

📥 Commits

Reviewing files that changed from the base of the PR and between 0f1d751 and a3e5334.

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Advanced

Run ID: 3c532ff0-d892-45f8-951f-8634fc5d65bc

📥 Commits

Reviewing files that changed from the base of the PR and between 0374c63 and 74a1d99.

📒 Files selected for processing (2)

• lib/crewai/src/crewai/tools/mcp_tool_wrapper.py
• lib/crewai/tests/tools/test_mcp_tool_wrapper.py

Included review availability: Your plan provides up to 10 included reviews per hour; 9 remain after this review.

---

📝 Walkthrough

Walkthrough

MCPToolWrapper now returns structured ToolFailure objects for MCP error responses. Successful responses still return text. Execution and retry helpers now propagate the wider return type, with tests covering both outcomes.

Changes

MCP error handling

Layer / File(s)	Summary
ToolFailure return propagation lib/crewai/src/crewai/tools/mcp_tool_wrapper.py	Execution, retry, timeout, and MCP call methods now accept and return str | ToolFailure.
MCP result conversion and validation lib/crewai/src/crewai/tools/mcp_tool_wrapper.py, lib/crewai/tests/tools/test_mcp_tool_wrapper.py	MCP results now extract content before checking isError. Error results become ToolFailure objects with MCP_ERROR, message content, and server/tool details. Tests cover error and successful responses.

Sequence Diagram(s)

sequenceDiagram
  participant MCPToolWrapper
  participant MCPServer
  participant ToolFailure
  MCPToolWrapper->>MCPServer: call_tool
  MCPServer-->>MCPToolWrapper: CallToolResult with content and isError
  MCPToolWrapper->>ToolFailure: create MCP_ERROR failure when isError is true
  ToolFailure-->>MCPToolWrapper: structured failure

Loading

Priority: ⬇️ Low

🚥 Pre-merge checks | ✅ 3 | ❌ 2

❌ Failed checks (2 warnings)

Check name	Status	Explanation	Resolution
Description check	⚠️ Warning	The description clearly explains the problem, solution, tests, and quality checks. However, it omits the required Related issue section and does not provide an issue reference.	Add the Related issue section with a valid linked open issue, for example: "Fixes #123". If no issue applies, confirm the repository policy with a maintainer before merging.
Docstring Coverage	⚠️ Warning	Docstring coverage is 38.89% which is insufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 18 functions across 2 files.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (3 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly and concisely describes the main change: reporting MCP isError responses as ToolFailure on the external URL path.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create a new PR

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[哈基米 (BlueX888)]

The pip-audit check on this PR has been failing since Sep 26 with a network-level error reaching PyPI while resolving packages, rather than any vulnerability finding — the audit never completed, and the dependency diff in this PR (reporting MCP isError as ToolFailure on the external URL path) doesn't add or change dependencies. Could someone re-run the scan when convenient? Happy to rebase or retrigger anything that helps it get a clean pass.