fix(transport): skip outputSchema validation for isError tool responses

[kimsehwan96]

🔗 Related Issue

Closes #4135

---

📝 Summary

When an upstream MCP tool with outputSchema defined returns an error (isError: true) without structuredContent, the MCP Python SDK's server-side validation replaces the original error message with a generic "outputSchema defined but no structured output returned".

This hides the actual error cause from the user.

This PR bypasses the SDK validation by returning a CallToolResult(isError=True) directly when is_error is True, which triggers the SDK's early-return path and skips outputSchema validation. This uses the same pattern already established in the codebase (lines 1451, 1468).

The root cause is in the MCP Python SDK v1.x (mcp/server/lowlevel/server.py), where outputSchema validation runs unconditionally regardless of isError.

---

🏷️ Type of Change

• Bug fix
• Feature / Enhancement
• Documentation
• Refactor
• Chore (deps, CI, tooling)
• Other (describe below)

---

🧪 Verification

Check	Command	Status
Lint suite	make lint	pass
Unit tests	make test	pass
Coverage ≥ 80%	make coverage	pass

---

✅ Checklist

• Code formatted (make black isort pre-commit)
• Tests added/updated for changes
• Documentation updated (if applicable)
• No secrets or credentials committed

---

📓 Notes (optional)

Implementation detail

The fix is placed after the structuredContent check in streamablehttp_transport.py:call_tool():

1. if structured → return (unstructured, structured) # SDK validates against outputSchema
2. if is_error is True → return CallToolResult(isError=True) # SDK early-return, no validation
3. return unstructured # normal path

This ordering is intentional: if an error response does include structuredContent, the existing validation path is preserved. We have not observed any MCP server returning both isError: true and structuredContent simultaneously, and the MCP spec does not define this combination. If it does occur, validating the structured content is arguably the correct behavior — the server explicitly provided it, so schema conformance should be checked.

is_error is True (strict identity check)

We use getattr(result, "is_error", None) with is True rather than a truthy check. This prevents false positives when result is a MagicMock in unit tests (where unset attributes return truthy MagicMock objects).

Manual E2E verification

To validate this fix end-to-end, we added a lookup_calculation tool to the existing output_schema_test_server that raises ToolError for unknown IDs (producing isError: true with outputSchema defined). Using the docker-compose stack:

• Before fix: "Output validation error: outputSchema defined but no structured output returned" ❌
• After fix: "Calculation 'nonexistent-id' not found. Only 'calc-001' is available." ✅

The test server change is not included in this PR — unit tests provide sufficient coverage for the transport-layer logic. The E2E setup can be formalized in a follow-up if needed.