| name | leap-debug-pipeline |
| description | Debugs stuck or failed LEAP recording pipelines—Celery queues, pipeline stages, and log correlation. Use when a recording fails mid-flight, a Celery task errors, workers seem idle, or the user asks to trace download through upload. |
LEAP pipeline debugging
Mental model
Stages (order may vary by config): download → trim (FFmpeg) → transcribe → topics → subtitles → upload. Work runs on Celery with queues such as downloads, processing_cpu, async_operations, uploads, maintenance (see backend/Makefile targets celery-downloads, celery-cpu, celery-async, celery-uploads, etc.).
Logs (local)
| Log | Use |
|---|
backend/logs/app.log | API, orchestration, many service errors |
backend/logs/celery-async.log | async_operations worker (transcription, topics, async-side effects) |
Search by recording id (and task name if present). Do not paste secrets into chat.
Workflow
- Confirm state — recording row / API status / last known stage (from app or DB if the user has access).
- Map stage to queue — e.g. trim → CPU queue; heavy I/O async →
async_operations; platform upload → uploads.
- Scan the right log for the recording id; note the first exception and stack trace.
- Classify
- External API (Zoom, Fireworks, VK, yt-dlp): rate limits, auth, payload — check creds and guides under
backend/docs/guides/.
- DB / migration mismatch: schema vs code — check
make db-version vs expected head (see alembic-migrations.mdc).
- Worker not running: task stays pending — confirm the matching worker process is up for that queue.
- Code path — open the task module under
backend/api/tasks/ for the failing stage; follow imports; prefer fixing root cause over silencing errors.
Do not
- Commit or log access tokens, refresh tokens, or API keys.
- Assume
async_operations uses gevent — project Makefile documents threads for asyncio compatibility.
