Pipeline
Nine scripts, one per step. Each reads the previous one’s file and writes its own to work/, so only the one that crashed gets rerun.
Replay of 7 November 2026, read from report and work
- Source video
- 5 February 2021, 7:00,
JZMgB1UItkk - Storyboard
- 106 keyframes, 95 caption lines of the video
- H sessions
- 12 runs in
report/bench.csv - Bench cost
- 0.55 USD over the runs that recorded a cost
The chain, after the run of 7 November
- Fetchyt-dlpdone
- ListenGladiadone
- Cutffmpegdone
- Pairlocal scriptdone
- ReadHolo4, H Models APIdone
- RegisterH Agents APIdone
- RunH agent in Chromedone, 200 s
- VerifySQL readdone
- Repairlocal scriptnot run
The steps
| No. | What | Who acts | Script | Output | State |
|---|---|---|---|---|---|
| 01 | Fetch, downloads the public video | yt-dlp | 10_download.sh | data/videos/<id>.mp4 | done |
| 02 | Listen, the voice as sentences with timestamps | Gladia | 20_transcribe.py | transcript.json | done |
| 03 | Cut, a keyframe each time the screen changes | ffmpeg | 30_keyframes.py | frames/*.jpg | done |
| 04 | Pair, each sentence on its keyframe | local script | 40_storyboard.py | storyboard.json | done |
| 05 | Read, steps, fields and rules from frames and voice | Holo4, H Models API | 50_induce.py | SKILL.<variant>.md | done |
| 06 | Register, pins the skill revision, creates the two agents | H Agents API | 60_register.py | agent.json | done |
| 07 | Run, the agent does the task in Dolibarr | H agent in Chrome | 70_run_one.py | work/runs/<run_id>.json | done |
| 08 | Verify, what the database says, not the agent | SQL read | 80_verify.py | verdict in the run | done |
| 09 | Repair, the next revision of the skill from the drifts | local script | 75_repair.py | repairs.json | built, not run |