Skip to content

feat(pipeline): judge-driven REFINE retries + single-pass extraction - #10

Merged
robotlearning123 merged 1 commit into
masterfrom
feat/refine-loop
Aug 7, 2026
Merged

robotlearning123 merged 1 commit into
masterfrom
feat/refine-loop

Conversation

@robotlearning123

Copy link
Copy Markdown
Contributor

Improve-7: the judge's verdicts now drive regeneration — REFINE retries with a bumped seed, best take kept, history in receipts/--json. Frame extraction collapsed to one decode pass (efficiency-review follow-up). Stub-judge behavior unchanged. 52/52 incl. a new e2e retry test.

…tion

- _run_shot: REFINE verdict regenerates with seed+100 up to
  OPEN_VIDEO_JUDGE_RETRIES extra takes (default 1); best-scoring take kept;
  full take history in shot.receipt['takes'] (surfaces through --json). With
  the stub judge take 1 PASSes — zero behavior change for unwired users.
- QualityJudge.extract_frames: one ffmpeg select pass for all N frames (was
  one full decode per frame — the assess hot path now that the judge is real)
- README + skill env table tell the new truth (refine is shipped; best-of-N
  stays roadmap)
- e2e test: REFINE→retry→best-take against the fake ComfyUI server (52/52)
@robotlearning123
robotlearning123 merged commit a3b80bf into master Aug 7, 2026
5 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant