video-workbench/handoff/2026-06-24-on-the-road-styl...

187 lines
8.7 KiB
Markdown

# On The Road Style / Reference Repair Handoff
## Status
- Date: 2026-06-24.
- Project: `projects/2026-06-23-在路上`.
- Current branch: `master`.
- Recent commits:
- `9b1a3c3` `visual-system/characters规则及清理`
- `4329653` `Broad-run images are candidates, not accepted finals. Identity/style reference workflow needs repair before further generation.`
- `c37e7ab` `Add accepted visual anchors and five-shot trial`
- Current handoff context index:
- `handoff/2026-06-24-on-the-road-style-reference-context/review-context.md`
- `handoff/2026-06-24-on-the-road-style-reference-context/review-file-manifest.json`
## User Decision
Use a fresh session. Do not rely on chat compression for the next phase.
The user corrected the reference policy:
- Person identity work must use image references.
- Do not downgrade identity-critical shots to text-only generation.
- The next identity test is not whether to use a reference at all; it is which active reference to use: master character anchor, front crop, side crop, back crop, walking crop, or transparent slice.
- Style references are different. The style board is useful, but style continuity may be better carried by a reusable prompt master plus visual calibration images rather than mandatory image-to-image for every shot.
## Must Read First
Read these in order before editing or generating:
1. `AGENTS.md`
- UTF-8 reading/writing rules.
- `gpt-image-2` workflow rule.
- Approval gate for project-bound image generation.
2. `VIDEO_WORKBENCH.md`
- `slides/sNN/sNN-vN-type.ext` naming rule.
- Character cleanup / archive rule.
- Project-bound image prompt rules.
3. `projects/2026-06-23-在路上/project.md`
- Project-local status and scope.
4. `projects/2026-06-23-在路上/execution-plan.md`
- Current execution strategy.
5. `projects/2026-06-23-在路上/visual-system/visual-system.md`
- Current visual system and A/B/C/D anchor definitions.
6. `projects/2026-06-23-在路上/visual-system/characters/character-reference-registry.md`
- Active character references and selection rules.
7. `projects/2026-06-23-在路上/slides/broad-generation-self-audit-v1.md`
- Why broad generation is paused.
- What failed in the prior broad run.
8. Five-shot prompt/review evidence:
- `projects/2026-06-23-在路上/slides/s01/s01-v1-prompt.md`
- `projects/2026-06-23-在路上/slides/s01/s01-v1-review.md`
- `projects/2026-06-23-在路上/slides/s10/s10-v1-prompt.md`
- `projects/2026-06-23-在路上/slides/s10/s10-v1-review.md`
- `projects/2026-06-23-在路上/slides/s46/s46-v1-prompt.md`
- `projects/2026-06-23-在路上/slides/s46/s46-v1-review.md`
- `projects/2026-06-23-在路上/slides/s47/s47-v1-prompt.md`
- `projects/2026-06-23-在路上/slides/s47/s47-v1-review.md`
- `projects/2026-06-23-在路上/slides/s51/s51-v1-prompt.md`
- `projects/2026-06-23-在路上/slides/s51/s51-v1-review.md`
9. Style source files:
- `projects/2026-06-23-在路上/visual-system/style/style-anchor-brief-v1.md`
- `projects/2026-06-23-在路上/visual-system/style/style-anchor-board-review-v1.md`
- `projects/2026-06-23-在路上/visual-system/style/style-anchor-c-repair-review-v1.md`
The generated context index lists the same reading set and included/excluded files.
## Confirmed Evidence
Five core shot tests were not true image-reference generation. They used Mode B standalone prompt generation. The references were used for prompt construction, not attached as actual image inputs.
Evidence in review files:
| Shot | Evidence |
| --- | --- |
| `s01` | `s01-v1-review.md` says no character reference was used and image references were not attached. |
| `s10` | `s10-v1-review.md` says character/style references were read or viewed locally, and explicit image-reference attachment was not available. |
| `s46` | `s46-v1-review.md` says character/style references were read or viewed locally, and explicit image-reference attachment was not available. |
| `s47` | `s47-v1-review.md` says character/style references were read or viewed locally, and explicit image-reference attachment was not available. |
| `s51` | `s51-v1-review.md` says character/style references were read or viewed locally, and explicit image-reference attachment was not available. |
This means the five-shot visual trial can still be useful as visually approved-looking image results, but it does not prove that the character anchor or style board can control future scene generation through actual image-reference input.
## Completed Since The Broad-Run Failure
- Broad-run candidates were committed as candidates, not accepted finals.
- `slides/sNN/` files were renamed to the `sNN-v1-type.ext` convention.
- `s34` bookkeeping was repaired:
- `slides.md` marks `s34` as `generated`.
- `broad-generation-run-v1.md` has a backfilled `s34` entry.
- The project image hash matched the host output.
- `visual-system/characters/` was cleaned:
- Active references remain there.
- Historical/process files moved to `archive/characters/`.
- `character-reference-registry.md` now defines active references and selection rules.
## Current Style Directory State
`visual-system/style/` still contains both active assets and process files:
- `style-anchor-board-v1.png`
- `style-anchor-board-v1.html`
- `style-anchor-board-spec-v1.md`
- `style-anchor-board-review-v1.md`
- `style-anchor-brief-v1.md`
- `style-anchor-c-repair-v1.png`
- `style-anchor-c-repair-prompt-v1.md`
- `style-anchor-c-repair-review-v1.md`
- `style-anchor-trial-plan.md`
- `style-blueprints-contactsheet.png`
Likely active assets:
- `style-anchor-board-v1.png`: common style calibration board.
- `style-anchor-c-repair-v1.png`: visually useful C repair sample.
- A new style prompt master / registry file to be created.
Likely archive candidates after registry/master exists:
- trial plan, brief, board spec, board review, C repair prompt/review, HTML rendering file, and contactsheet if not needed as an active visual reference.
Do not archive style files before deciding and writing the active style registry / prompt master.
## Next Recommended Sequence
1. Build style registry / prompt master.
- Suggested path: `projects/2026-06-23-在路上/visual-system/style/style-prompt-master.md`
- It should define A/B/C/D modules, global style kernel, negative drift list, and how each module maps to shot types.
- It should also state which visual files remain active calibration references.
2. Then clean `visual-system/style/`.
- Create `archive/style/`.
- Keep only active style references and the new style prompt master in `visual-system/style/`.
- Move process files to `archive/style/`.
3. Repair prompt rules before any generation resumes.
- Each `sNN-vN-prompt.md` must include explicit `Reference Inputs`.
- Character references must be selected from `character-reference-registry.md`.
- Style prompt module must be selected from the new style master.
4. Run character reference validation tests.
- Test master anchor versus front/side/back/walking crops/slices.
- Required test set:
- side or half-face identity;
- back walking;
- side walking;
- medium human in environment.
- Do not continue broad generation until this test proves which active references work.
5. Only after the above, repair or regenerate high-risk shots from the broad run:
- `s14`
- `s33`
- `s21`
- `s05`
- `s25`
6. Then resume remaining missing shots.
## Do Not Do Next
- Do not continue from `s35` yet.
- Do not treat the five-shot trial as proof of character-reference control.
- Do not use archived character files as default references.
- Do not create identity-critical scene images using text-only character descriptions.
- Do not skip per-shot prompt/spec artifacts.
- Do not ask GPT V2 for final prompts; Codex owns local final prompt construction through the `gpt-image-2` workflow.
## Open Questions For Next Session
- Which actual generation path can attach local character reference images for identity-critical tests?
- Should `style-anchor-board-v1.html` remain active as an editable board source, or move to `archive/style/` after `style-prompt-master.md` exists?
- Should `style-blueprints-contactsheet.png` remain active for quick visual inspection, or become archived process evidence?
- Should style modules reference source images from `intake/imgs/` directly, or only reference the accepted style board and C repair image?
## Verification Commands Already Run In This Handoff
```powershell
git status --short
```
Result before creating this handoff: clean.
```powershell
python C:\Users\wangq\.agents\skills\review-context-builder\scripts\review_context_builder.py ...
```
Result: wrote the context index and manifest under:
```text
handoff/2026-06-24-on-the-road-style-reference-context/
```