Record gpt-oss harmony tool calls + stream/recorder hardening

[jpr5]

Summary

aimock could not record fixtures from local gpt-oss models (e.g. via Ollama/vLLM/OpenRouter) that emit OpenAI harmony channel tokens — the raw control tokens (<|channel|>commentary to=functions.NAME <|constrain|>json<|message|>{...}<|call|>) leaked into recorded content instead of being parsed into tool calls. This PR adds harmony parsing plus a round of stream/recorder hardening surfaced while fixing it.

Reported in Slack: local gpt-oss (gpt-5.x class) recordings were unusable because hosted OpenAI pre-parses harmony, but local runtimes pass it through raw.

What changed (by concern)

1. Harmony channel parser (src/harmony.ts) — a two-phase lexer + state-machine that routes harmony analysis/commentary/final channels into reasoning/tool-calls/content. Uniform all-or-nothing fail-safe: on any structural deviation it returns the original content verbatim and signals harmonyUnparsed/harmonyNote — it never mangles, leaks a control token, or glues message bodies together. Wired as fallback-only (used only when no structured tool calls were already parsed), so it can never produce phantom/merged calls.

2. Stream collapser hardening (src/stream-collapse.ts) — multi-data:-line and CRLF SSE handling; missing/uncorrelated tool_call index guards with symmetric dropped-chunk accounting across OpenAI/Anthropic/Bedrock/Cohere (uncorrelated arg deltas now increment droppedChunks + capture a firstDroppedSample instead of vanishing silently); Bedrock EventStream header-bounds validation (malformed frames degrade to truncated instead of throwing).

3. Recorder (src/recorder.ts, src/types.ts) — incremental multibyte UTF-8 decoding (StringDecoder) so characters split across stream chunks are no longer corrupted on the frame-timing path; CRLF-tolerant frame-timing splitter; webSearches propagation into fixtures; audio-branch companion fields (toolCalls/content/reasoning) persisted; firstDroppedSample logged alongside the dropped-chunk warning.

4. Gemini audio companion replay (src/gemini.ts) — the audio replay builders now re-emit companion tool-call/text/thought parts (mirroring the non-audio builder) instead of dropping them, completing the record→replay round-trip for audio turns that interleave tool calls.

Tests

+~3.4k lines of tests: harmony structural acceptance matrix + boundary/fail-safe regressions; collapser robustness (index guards, dropped-chunk accounting, CRLF); multibyte decode (both decode paths, byte-by-byte split); Gemini audio companion record→replay. Full suite: 3308 passed / 0 failed, tsc clean, tsdown build green.

Documented limitations

• Harmony quoted-structure: deeply ambiguous quoted control-token-like sequences inside message bodies fall to the verbatim fail-safe rather than being heuristically parsed — by design (no corruption over best-effort parsing).
• Audio companions are Gemini-replay-only: audioB64 collapse is currently Gemini-only, so cross-provider audio fixtures would not replay companion modalities. Documented at the type and builder.

Test plan

• npx vitest run — 3308 passed, 37 skipped, 0 failed
• npx tsc --noEmit — clean
• pnpm build (tsdown) — clean
• CI green

[pkg-pr-new]

Open in StackBlitz

npm i https://pkg.pr.new/@copilotkit/aimock@245

commit: 75c1066