Fix H3 rejoin assembling stale clips after an image rerun

[cursor]

Resumen ejecutivo

Esta primera sección está pensada para project managers y revisores no
técnicos. Mantén el detalle técnico completo más abajo.

Qué cambia

Rejoin de producciones MiniMax H3 ahora rechaza un clip marcado video_stale en vez de montar los segmentos o la toma seleccionada.

Para qué sirve

Después de regenerar la imagen de inicio, el vídeo ya no corresponde al nuevo fotograma. En H3, Rejoin igual ensamblaba ese material y publicaba una película final desactualizada.

Impacto para el usuario

Rejoin pide regenerar el vídeo stale antes de unir. El comportamiento no-H3 no cambia.

Riesgo

• Bajo
• Medio
• Alto

Estado

• En desarrollo
• Listo para revisión
• Bloqueado por CI o revisión
• Requiere migración o acción manual

Summary

rejoin_clips only checked video_stale on the non-H3 path. Sequential H3 productions (minimax_h3 / minimax_h3_legacy) took a different branch that assembled playable h3_segments or an explicit selected_video_filename without looking at the clip-level stale flag.

rerun_clip_image marks video_stale=True and leaves those files in place for playback. A later Rejoin therefore concatenated the pre-rerun video into a new minimax_h3_*_rejoin_*.mp4 and wrote it as final_output_filename.

Hoist the existing stale-clip ValueError so both branches share it. Segment-level H3 stale checks and whole-slot selections stay as they were once the clip is current.

Overview

1. User finishes an H3 multi-clip production.
2. User reruns the start image on clip 1. The old segments remain on disk; the clip is video_stale.
3. User clicks Rejoin.
4. Before this change: ffmpeg concatenated the old H3 segments (or a Studio selection) and the dashboard showed a new final movie.
5. After this change: Rejoin raises Regenerate stale video clip(s) 1 before rejoining. Concatenate is not called.

Detailed changes

Backend

• app/services/director_pipeline.py _rejoin_clips_impl: evaluate video_stale before the H3/non-H3 split. Remove the duplicate check from the non-H3 branch.

UI and Wizard

No UI change. The dashboard already surfaces stale clips; Rejoin now matches that gate for H3.

Data, provenance and compatibility

No checkpoint migration. Existing H3 files stay playable; only assembly is blocked until video is regenerated.

Files and ownership

• app/services/director_pipeline.py — Rejoin stale gate.
• tests/test_director_h3_workflow_edits.py — H3 stale clip with current segments must not concatenate.
• Left untouched: director_review.py persist, _backfill_clip_video_attempts (covered by open fix(review): preserve recovered takes and stale-video guards #406 / merged-into-fix(review): preserve recovered takes and stale-video guards #406 Fix review/rejoin treating a selected stale take as current #409), and _select_take clearing stale on an explicit take switch (documented by the review-desk projector).

Validation

• Date (UTC): 2026-09-12

• Base SHA: 60e867a

• Head SHA: 4048ebd

• Validation scope: focused

• python scripts/verify_clean_repo.py

• python -m compileall -q app/services/director_pipeline.py

• Focused Python tests: PYTHONPATH=/workspace python -m pytest tests/test_director_h3_workflow_edits.py tests/test_director_cancellation.py::TestDirectorCancellation::test_rejoin_rejects_stale_video_instead_of_omitting_clip -q → 8 passed

• cd ui && npm run i18n:check (no UI change)

• UI tests: N/A (no UI change)

• cd ui && npm run lint -- --max-warnings=0 (no UI change)

• cd ui && npm run build (no UI change)

• git diff --check

• E2E/smoke checks: N/A — Rejoin H3 assembly is covered by the focused pytest; no live ffmpeg/provider run

Code quality

• Score: pending CI
• Complexity trend: pending CI
• Production LOC trend: pending CI
• Regression versus main: pending CI
• Ratchet: pending CI

CI and review

• CI of this HEAD: pending
• Independent agent review of this HEAD: pending
• Human merge click (operational, not code review): pending

Cursor/Bugbot — pick exactly one.

• Reviewed at current HEAD
• Earlier review; HEAD has changed since (stale)
• Pending
• Unavailable

Coste de la tarea

• Tests simulados: 0 tokens externos
• Tests reales: N/A
• Llamadas LLM externas: 0
• Tokens de prompt: N/A
• Tokens de respuesta: N/A
• Tokens totales: N/A
• Generaciones de imágenes/audio/vídeo: 0
• Tiempo transcurrido: N/A
• Proveedores/modelos: N/A

Notes and limitations

• The H3 branch still skips start-image presence checks (pre-existing). This PR only shares the video_stale gate.
• Selecting a different historical take via save_review still clears video_stale when the filename changes. That matches the review-desk projector test and was not changed here.

Follow-up work

• fix(review): preserve recovered takes and stale-video guards #406 (sidecar persist) and Fix review persist dropping Approve after notes blur #405 (Approve after notes blur) remain open.
• Fix review/rejoin treating a selected stale take as current #409 merged into fix(review): preserve recovered takes and stale-video guards #406, not into development.

Checklist

• The executive summary is understandable without reading the code.
• The detailed Summary/Overview has not been removed or shortened.
• Tests and their actual results are recorded.
• Generated assets, secrets and local-only files are not committed.
• Required CI and Cursor/Bugbot review are complete, or the PR is clearly marked as waiting for them.

  

[github-actions]

PR Review — Loreframe Studio

Risk: low
Scope: 2 file(s); +63/-11; backend services

Automated review from scripts/analyze_pr.py. This is a heuristic pass (no LLM) so humans still own the merge decision.

Findings

• No heuristic issues. Still run the CI checklist below.

Changed files

• modified: app/services/director_pipeline.py, tests/test_director_h3_workflow_edits.py

CONTRIBUTING checklist

• python scripts/verify_clean_repo.py
• python -m compileall -q app/services app/launch.py scripts
• cd ui && npm run build if the UI changed
• No weights, CivitAI sidecars, or generated guides
• Stays local-first (no required accounts / telemetry)

Posted by the repo PR review workflow. Re-runs on each push to the PR.

[github-actions]

Code health

Quality score: 63.8/100

Higher is better. The score is a trend dashboard; the independent ratchet below remains the CI gate.

Component	Weight	Current	Change
Cyclomatic health	45%	59.2	+0.0
File concentration	25%	73.7	+0.0
Oversized-file debt	20%	49.4	+0.0
Modularity	10%	88.3	+0.0

Change vs PR base: +0.0 points.

Metric	Value
Production LOC	313,652
Production files	1,052
Test LOC	127,860
Functions measured	21,367
Functions complexity ≥ 15	906
Maximum complexity	674
Policy	code-health-policy-v1
HEAD	526a692863e9726434f80c7fce0000f1d35ecd9a
Base	60e867a6aae832d5b56c5ebd44b99867bda55b26
UI measurement	complete

Markdown, JSON catalogs and tests are out of this table. Only app/ runtime + ui/src TS/JS count.

Most complex functions

Complexity	Where
674	app/wgp.py:7230 generate_video
382	ui/src/stores/useStore.ts:4129 Async method 'startGeneration'
355	app/_launch_runtime.py:23674 _run_generation
308	app/wgp.py:12372 generate_video_tab
269	ui/src/components/Sidebar/SceneAnimatorPanel.tsx:508 Function 'SceneAnimatorPanel'
264	ui/src/stores/useStore.ts:8684 Async method 'loadSettingsFromOutput'
258	app/services/director/planners/short_film.py:3433 ShortFilmPlanner._plan_story_driven
248	app/services/director_pipeline.py:13614 _run_video_generation
245	app/services/director_pipeline.py:7714 _run_pipeline
233	ui/src/features/agent/agentActions.ts:1152 Function 'parseAction'
226	app/services/director_pipeline.py:6543 update_comic_preview
225	ui/src/features/agent/agentActions.ts:2998 Async function 'executeAgentActions'

Trend vs baseline

Metric	Δ
Production LOC	+2
Test LOC	+50
Functions ≥ 15	+0
Maximum complexity	+0

Warnings

• production LOC increased by +2
• hotspot app/services/director_pipeline.py increased by +2 lines

Ratchet passed.