🤖 fix: map Copilot Responses function-call SSE events

[alecsg77]

Copilot Responses adapter sends tool schemas but drops function-call SSE events

Environment

• Tested Mux: v0.28.1-nightly.12-7-gd25b50ecc-dirty (d25b50ecc)
• Platform: Linux x86_64
• Source fixture: coder/mux a9a6b44fa63628a67c06d7d03ef239d668a2646a
• Component: src/node/services/copilot/copilotResponsesLanguageModel.ts

Problem

CopilotResponsesLanguageModel includes function tools in the outbound Responses request and serializes prior function calls/results. Its streaming parser, however, does not map Responses function-call events to AI SDK tool-input/tool-call stream parts.

mapStreamEvent() currently handles:

• response.created
• message/text events
• terminal events
• errors

For response.output_item.added, it explicitly returns no parts unless the item type is message. It has no cases for:

response.function_call_arguments.delta
response.function_call_arguments.done

Regression fixture

The attached patch adds a test with this SSE sequence:

response.output_item.added        item.type=function_call
response.function_call_arguments.delta
response.function_call_arguments.done
response.completed

Expected stream-part types:

stream-start
tool-input-start
tool-input-delta
tool-input-end
finish

Actual stream-part types:

stream-start
finish

Test command:

bun test src/node/services/copilot/copilotResponsesLanguageModel.test.ts \
  --test-name-pattern 'streams function-call argument events as tool input parts'

Result:

Targeted fixture: 0 pass, 1 fail.
Full adapter file: 12 existing tests pass, only the new fixture fails.
Expected tool-input-start, tool-input-delta, and tool-input-end;
received only stream-start and finish.

Impact

Any Copilot model using this custom Responses adapter can advertise/receive tool definitions but lose the streamed function call before Mux can execute the tool. This is independent of the separate early-EOF/stream_truncated behavior seen with gpt-5.3-codex.

Expected result

Map the Responses function-call lifecycle into AI SDK stream parts, maintain stable call IDs, and add regression coverage for complete and chunked argument streams.

Evidence

• evidence/copilot-responses-function-call-fixture.patch
• evidence/copilot-responses-function-call-fixture.log

---

Generated with mux • Model: openai:gpt-5.6-terra • Thinking: high • Cost: $1.39

[linear-code]

MUX-66