fix(mcp): mark errored MCP tool results as failed (set error_code=-1)

[nankingjing]

Problem

MCPTool.execute (trae_agent/tools/mcp_tool.py) does not propagate a failure code when an MCP server returns an error result:

output = await self.client.call_tool(self.get_name(), arguments)
if output.isError:
    return ToolExecResult(output=None, error=output.content[0].text)   # error_code defaults to 0
else:
    return ToolExecResult(output=output.content[0].text)

ToolExecResult.error_code defaults to 0 (see trae_agent/tools/base.py). ToolExecutor.execute_tool_call derives success from that field:

success=tool_exec_result.error_code == 0,

So when an MCP tool call reports isError == True, the resulting ToolResult.success is True. The failure is silently treated as a success:

• The Anthropic client sends the result back with is_error=not success → is_error=False, telling the model the tool succeeded.
• Failure-reflection logic (BaseAgent.reflect_on_result, which keys on tool_result.success) never fires.
• The trajectory records the failed step as success: true.

Note the except branch in the same method already sets error_code=-1, and the existing test_execute_exception asserts it — the isError branch was simply missing the same signal.

Fix

Set error_code=-1 in the isError branch so failures are propagated consistently with the exception branch.

Verification

• python -m py_compile on both changed files: passed.
• Reproduced end-to-end against the real MCPTool + ToolExecutor code with a stubbed client:
  • Before: an isError result yields ToolResult.success == True (bug).
  • After: it yields ToolResult.success == False; the success path still yields success == True.
• Extended the existing test_execute_failure unit test to assert error_code == -1. (Full pytest suite not run here — repo targets Python 3.12+ and the local interpreter is 3.11; verification used a typing.override shim.)

[nankingjing]

Reviewed — the error_code fix correctly signals failures to the agent loop, and the test assertion covers it. Ready for review.

[nankingjing]

@chao-peng ready for review. Fixes MCP errored results not signaling failure to the agent loop by setting error_code=-1. One regression test covers the assertion.