fix: do not re-finalize mcp_list_tools in closeLastOutputItem

[Ali-Aleph-Alpha]

Summary

Fixes a bug where the Responses streaming layer emitted response.output_item.done twice for an mcp_list_tools output item, surfacing as a duplicate TOOL_CALL_RESULT in downstream AG-UI consumers.

Root Cause

Output-item finalization is split between two places in this codebase:

Item type	Where response.output_item.done is emitted
message	closeLastOutputItem
reasoning	closeLastOutputItem
function_call	closeLastOutputItem
mcp_call	closeLastOutputItem
mcp_approval_request	closeLastOutputItem
mcp_list_tools	listMcpToolsStream (full lifecycle)

listMcpToolsStream already emits the full lifecycle for its own item (added → in_progress → completed → output_item.done). However, the previous shared branch in closeLastOutputItem matched both mcp_approval_request and mcp_list_tools:

} else if (lastOutputItem?.type === "mcp_approval_request" || lastOutputItem?.type === "mcp_list_tools") {
    yield { type: "response.output_item.done", ... };
}

When handleOneTurnStream later called closeLastOutputItem (e.g. before transitioning to reasoning/text output) and the last output item was still that same mcp_list_tools, this branch re-emitted response.output_item.done for it.

Why this only manifests for mcp_list_tools

PydanticAI's Responses adapter maps each response.output_item.done for an McpListTools item into a BuiltinToolReturnPart, and AGUIAdapter renders each return as a TOOL_CALL_RESULT. The duplicate output_item.done therefore surfaced as two identical TOOL_CALL_RESULT events for the same toolCallId in AG-UI.

Other tool types (function_call, mcp_call, mcp_approval_request) get output_item.done exclusively from closeLastOutputItem and are never re-finalized, so they're unaffected.

Before / After

Before

response.output_item.added      item=mcp_list_tools id=mcpl_123
response.mcp_list_tools.in_progress
response.mcp_list_tools.completed
response.output_item.done       item=mcp_list_tools id=mcpl_123
...
response.output_item.done       item=mcp_list_tools id=mcpl_123  # duplicate

PydanticAI/AG-UI rendered:

TOOL_CALL_RESULT toolCallId=pyd_ai_builtin|openai|mcpl_123
TOOL_CALL_RESULT toolCallId=pyd_ai_builtin|openai|mcpl_123

After

response.output_item.added      item=mcp_list_tools id=mcpl_123
response.mcp_list_tools.in_progress
response.mcp_list_tools.completed
response.output_item.done       item=mcp_list_tools id=mcpl_123

AG-UI now emits exactly one TOOL_CALL_RESULT for the list-tools discovery.

Fix

Split the combined branch in closeLastOutputItem so:

• mcp_approval_request keeps its existing close behavior (it is genuinely closed only here).
• mcp_list_tools is treated as already finalized by listMcpToolsStream and produces no events.

Event schema is unchanged; this only prevents the redundant completion. No behavioral change for any other item type.

Test plan

• pnpm run test -- src/routes/responses/mcpStream.test.ts src/routes/responses/closeOutputItem.test.ts src/routes/responses/handleOneTurn.test.ts (all 148 tests pass)
• pnpm run check (tsc clean)
• Prettier check on touched files
• New test in closeOutputItem.test.ts: closing when last item is mcp_list_tools yields no events
• Strengthened existing mcpStream.test.ts: listMcpToolsStream emits exactly one response.output_item.done

[github-actions]

⚠️ Deprecation Warning: The deny-licenses option is deprecated for possible removal in the next major release. For more information, see issue 997.

Dependency Review

✅ No vulnerabilities or license issues or OpenSSF Scorecard issues found.

Scanned Files

None

[Copilot]

Pull request overview

Fixes duplicate response.output_item.done emission for mcp_list_tools by ensuring closeLastOutputItem does not re-finalize items whose lifecycle is already fully emitted by listMcpToolsStream, preventing duplicate downstream TOOL_CALL_RESULT events.

Changes:

• Split closeLastOutputItem handling so mcp_approval_request still emits response.output_item.done, while mcp_list_tools becomes a no-op.
• Strengthen listMcpToolsStream test to assert exactly one response.output_item.done.
• Add a regression test ensuring closeLastOutputItem yields no events when the last item is mcp_list_tools.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

File	Description
src/routes/responses/mcpStream.test.ts	Adds an assertion to ensure listMcpToolsStream emits response.output_item.done exactly once.
src/routes/responses/closeOutputItem.ts	Prevents re-emitting response.output_item.done for mcp_list_tools while keeping existing behavior for mcp_approval_request.
src/routes/responses/closeOutputItem.test.ts	Adds a regression test verifying closing a trailing mcp_list_tools item is a no-op.

Comments suppressed due to low confidence (1)

src/routes/responses/closeOutputItem.ts:242

• The StreamingError message is now misleading: this function also supports reasoning, mcp_approval_request, and explicitly no-ops mcp_list_tools, but the error says it only expects message, function_call, or mcp_call. Please update the message to reflect the full set of supported output item types (or make it generic, e.g. "unexpected output item type").

			throw new StreamingError(
				`Not implemented: expected message, function_call, or mcp_call, got ${(lastOutputItem as ResponseOutputItem)?.type}`
			);

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.