fix: collapse OpenAI Responses-API function-call streams into toolCalls

[jpr5]

Problem

collapseOpenAISSE() handles OpenAI Responses API text deltas, reasoning, and web-search output items, then a catch-all (if (parsed.type?.startsWith("response.")) continue;) silently skips every other response.* event — including the entire function-call sequence (response.output_item.added with a function_call item, response.function_call_arguments.delta/.done, and response.output_item.done for a function_call). Recording a tool-call-only turn therefore collapses to empty content with no toolCalls (logged: Stream collapse produced empty content — fixture may be incomplete), and replaying that fixture yields an empty assistant turn / no tool invocation.

The older Chat-Completions path already handles tool calls (choices[0].delta.tool_calls), which is why pre-existing fixtures work and only fresh Responses-API recordings break.

This surfaced downstream: the CopilotKit 1.68.1 showcase deploy shifted integrations onto the Responses-API tool-call shape, and any attempt to re-record D4 chat-smoke fixtures produced empty fixtures → whole-column "empty assistant response" flap across staging integrations.

Fix

Add a Responses-API function-call handler in collapseOpenAISSE (before the response.* catch-all), mirroring the Chat-Completions accumulation:

• response.output_item.added (item.type === "function_call") → start a tool-call accumulator keyed by output_index, capturing call_id as the tool-call id (the id a tool result references, not the internal fc_… item id) and name.
• response.function_call_arguments.delta → append delta to arguments (get-or-create, robust to delta-before-added).
• response.function_call_arguments.done → adopt full args only if deltas produced none.
• response.output_item.done (function_call) → finalize (fill id/name if .added was missed; adopt args only if empty).

Reuses the existing toolCallMap + orderAtoms and the existing emit path, so output is { id, name, arguments } (arguments normalized to valid JSON) — identical to the Chat-Completions shape and the checked-in fixtures. output_index keyspace is disjoint from Chat-Completions tool_calls[].index (a stream is never both shapes).

Proof — local red→green (unit)

New describe("collapseOpenAISSE Responses API function calls") in src/__tests__/stream-collapse.test.ts feeds a synthetic Responses-API function-call SSE (single tool, multiple tools keyed by output_index, and a text-only no-regression case) using the exact wire shape aimock emits.

• RED (src reverted to origin/main, new tests present): the two tool-call tests fail — result.toolCalls undefined. 2 failed | 1 passed.
• GREEN (fix applied): 3 passed.

Proof — live record → replay (independent, real OpenAI)

Validated independently by @mark against a real OpenAI key, aimock image built from this exact commit, on the built-in-agent D4 weather flow:

• Record (--record, Loaded 0 fixture(s)): D4 L3 greeting green, L4 "What's the weather in San Francisco?" green. Journal held exactly 3 correlated 200s (greeting, tool-call turn, post-tool-result turn). The initial weather fixture now contains real toolCalls (no empty-content fallback):

  { "toolCalls": [ { "name": "get_weather", "arguments": "{\"location\":\"San Francisco\"}", "id": "call_WTiZY5LQCFpwqkbDm6CdMGhN" } ] }

  The post-tool-result entry had turnIndex: 1, hasToolResult: true, and a non-empty final assistant response.
• Fresh replay (new process, no --record/--proxy-only, Loaded 3 fixture(s)): D4 L3+L4 green; replay journal shows the full tool chain; DOM assertions verified the weather card, tool result values (68, 55%, 10 mph), and the final assistant text in [data-testid="copilot-assistant-message"].

Checks

• stream-collapse.test.ts: 205 passed.
• Full suite: 5,271 passed / 46 skipped / 0 failed.
• pnpm build, pnpm format:check, pnpm lint: all clean.

Scope / follow-up

This PR fixes the stream collapser only. It does not address the fixture-capture race (a post-tool-result turn can still be draining after the probe exits, landing a fixture in the next run's dir). That belongs in a separate PR that makes the capture workflow wait for the request journal to drain before moving fixtures / restarting aimock. Per @mark's live findings, a replay drain barrier should key on response.status === 200 && response.fixture != null (on ordinary replay response.source is left unset; record-mode upstream completion is response.source === "proxy").

D4 fixture re-recording stays paused until that capture-workflow barrier lands.

🤖 Generated with Claude Code

https://claude.ai/code/session_01V5dSyNeU6xH9ofPf9ZLaNQ

[pkg-pr-new]

Open in StackBlitz

npm i https://pkg.pr.new/@copilotkit/aimock@380

commit: 23a8c3c

[contextablemark]

Approved. Confirmed this correctly preserves Responses API function-call events as toolCalls, with unit coverage and successful live record → fresh replay validation. The journal-drain race is separate follow-up work and is outside this PR’s scope.