fix(media-use): report a failed transcription as an anomaly - #5819

Thanks for this. The diagnosis is right and the test is the right shape: it spawns the real script with a fake npx that exits the way a missing ASR engine does, so it enters through the production path rather than a lifted helper. I checked transcribeWav and it does only ever return null or a non-empty array, so the null test is the correct trigger for the engines you named, and the run still exits 0 with words left empty, as described.

The fix stops one line short of the root, though. In lib/tts.mjs the HeyGen branch returns words: [] whenever the response has no word_timestamps, or when every entry is dropped by the sentinel and finite-number filters. In audio.mjs the new anomaly sits inside if (!wordArr), and an empty array is truthy, so the HeyGen path skips both the transcription fallback and the anomaly. A HeyGen run that comes back with audio and no timings still writes an empty words list and finishes silently, which is the same symptom on what is usually the default provider.

Keeping the fallback trigger on null and testing emptiness for the anomaly covers both in one place, something like if (!wordArr?.length) after the fallback. A second test case with a HeyGen response that has no word_timestamps would lock it in.

Separately, the branch needs a rebase before it can merge. skills-manifest.json conflicts: your change moves the media-use entry to a new hash at 113 files, and main is now at a different hash with 116 files after another change landed under that skill. The manifest should be regenerated after the rebase rather than hand-merged, otherwise the in-sync check will fail on the merge commit. CI is otherwise genuinely green at this head, 58 runs with nothing failing.

One tiny thing: the anomaly is pushed before the duration check, so a line with no timings and a bad duration reports both, including the one saying the voice was omitted.

The fix is worth landing, but the branch needs a rebase and there is still one gap. Nothing has been pushed since Miguel's change request, so both of his points stand.

Rebase. skills-manifest.json conflicts: main's media-use entry is now hash 1082bd1c4e3f033b with 116 files. Rebase and regenerate with gen-skills-manifest.ts rather than resolving the hash by hand.

The HeyGen path is still silent. I checked main: the HeyGen branch of lib/tts.mjs (around line 343) returns words: [] when the response has no word_timestamps, or when every entry is filtered out. An empty array is truthy, so in audio.mjs it skips both the transcription fallback and the new anomaly. Keeping the fallback on null and raising the anomaly on !wordArr?.length after it covers both providers in one place. A second test with a HeyGen response that has no word_timestamps would pin it.

The Gemini test itself is good: it drives the real script with a fake npx that fails the way a missing ASR engine does.

Need: yes, missing word timings are silent today · Size: ok · Reuse: uses the existing anomalies channel · Simpler: one emptiness check after the fallback covers both providers

— Rames

Thanks for the quick turnaround. Everything from the earlier round is addressed at this head:

• The anomaly now fires on an empty result after the fallback, so the HeyGen path that returns words: [] is reported too, and the fallback still only runs on null.
• It is pushed after the duration check, so a line with a bad duration reports only the omission.
• The manifest was regenerated after merging main; only the media-use hash changes and the in-sync check passes.

I ran the pipeline tests at this head (6 pass) and then swapped in main's audio.mjs: both new tests fail and the four existing ones still pass, so they guard the change. Both drive the real script, which is the right way in.

Two small things, neither blocking:

• On the HeyGen path the transcription fallback never runs (an empty array skips it), but the message still says "transcription failed ... run npx hyperframes transcribe to see why". Wording that covers both causes, for example "no word timings (the provider returned none, or transcription failed)", would avoid sending HeyGen users to a command that probably works. The fake npx in the HeyGen test is never called for the same reason, so its comment could say so.
• All four commits on the branch are unsigned, and main requires signed commits, so they need re-signing before this can be queued.

Verdict: APPROVE
Reasoning: Both earlier findings are fixed at this head, the two new tests fail without the change and pass with it, and CI is green; only a wording nit and commit signing remain.

Re-checked at bf51001. I ran gemini-pipeline.test.mjs: 6 passed. With audio.mjs put back to the merge base, 2 fail, so the tests guard the change. Together with #5568 this covers #5564.

Verdict: APPROVE
Reasoning: reports a failed or empty transcription where the pipeline already reports anomalies; tests fail without it; CI green.

Reviewed at 65a7ce627685f12f7fdb3b3a7319f80e2bd31a8b.

The empty-word check is now after the fallback and covers both a null transcription and HeyGen returning an empty timings array. This fixes the earlier review finding without changing successful audio output.

At this exact head, node --test skills/media-use/audio/scripts/gemini-pipeline.test.mjs passed all 6 tests with controlled network and media executables. No paid transcription was sent.

Need: yes · Size: ok · Reuse: existing implementation and tests · Simpler: already bounded · Structure: fine
Risk: low - bounded change; validation limits are stated above.

Verdict: APPROVE

Reasoning: The empty-word check is now after the fallback and covers both a null transcription and HeyGen returning an empty timings array. This fixes the earlier review finding without changing successful audio output.