video-workbench/projects/2026-06-23-在路上/execution-plan.md

Execution Plan

Current Medium Branch

• Branch: MV shot-image production for later video assembly.
• Format: 16:9 horizontal.
• Production mode: image assets first; video generation/editing remains manual or semi-manual until explicitly automated.
• Prompt ownership: Codex owns final image prompts; GPT intake is source specification, not final prompt.

Current Execution Focus

Prepare the execution base, not the full MV.

Immediate focus:

1. Build and validate the character anchor board.
2. Build and validate the style anchor board.
3. Use both anchors in a small visual test package.
4. Only after tests pass, expand into broader shot production.

Materialization Targets

Target	Required Output	Status
Character anchor board	Main 16:9 board on shallow gray-white background	accepted v2
Character transparent slices	Front full body, side full body, back full body, walking side, walking back	accepted v2-from-anchor set
Character labels	Short identity, clothing, prop, and boundary labels	pending
Style anchor board	16:9 dark-gray board with A/B/C/D samples	accepted
C style retest	Dark muddy puddle / broken star reflection sample	accepted
Shot test package	S-01, S-10, S-46, S-47, S-51	accepted

Small-Batch Strategy

Do not generate all 51 shots first.

Round 1 uses five shots because they test the system's biggest risks:

Shot	Why It Is In Round 1	Anchor
S-01	Opening landscape: proves world-before-person and main A composition.	A
S-10	Steel canyon: proves city pressure without tech/UI drift.	B
S-46	Muddy starlight: proves the soul image and C repair.	C
S-47	Night road fire: proves restrained near-fire transition after C.	D
S-51	Final wide: proves distant fire, no victory, no gathering, person remains small.	D

Round 1 success means these are true:

• The traveler stays recognizable without becoming a portrait subject.
• A/B/C/D feel like one MV, not four unrelated styles.
• C is no longer sunrise/pretty reflection and has clear muddy starlight.
• D is warm but not camping, gathering, or victory.
• The output can support later prompt batching without rewriting the visual system.

Current Round

• Round: R2 five-shot trial preparation.
• R0 completed: intake cleaned; project execution files initialized.
• R1 completed so far: v2 main character anchor accepted; identity lock approved for slice generation; five transparent slice candidates generated and rejected due identity drift from v2.
• R1 completed: v2 main character anchor accepted; v2-from-anchor transparent slice set accepted.
• R1 completed: C style retest accepted as style-anchor-c-repair-v1.png.
• R1 completed: common style anchor board accepted as style-anchor-board-v1.png.
• R2 completed: S-01 generated once and accepted by user.
• R2 completed: user waived intermediate prompt review for S-10; S-10 generated once and accepted by user.
• R2 completed: S-46 generated once and accepted by user.
• R2 completed: S-47 generated once and accepted by user.
• R2 completed: S-51 generated once and accepted by user.
• R2 completed: five-shot visual trial generated and accepted by user.
• R3 character-consistency workflow test completed: s21-v3, s14-v2, and s25-v2 generated through the Advisor / visible-reference path and accepted by user. Quality route is usable; duplicate Agent S-14 remains a historical mechanism defect from this run.
• R3 third-round workflow test completed: session-04 generated s33-v2 and s06-v2 through a clean user-visible child session plus internal one-shot SubAgents. Both images were accepted by user; review files were absent as requested; no duplicate-worker defect was observed in this two-slide scope.
• R4 planned: formal-run simulation from a clean committed baseline. Candidate scope is S-02, S-03, and S-04, using one clean child session with up to 3 slides, main-session no-audit default, child-session mandatory completeness gate, and optional review only if requested.

Active Batch

Broad generation trial remains paused until R4 formal-run simulation starts.

No single-shot gate is active. The five-shot visual trial is accepted. The character-consistency workflow test is also visually accepted for S-21, S-14, S-25, S-33, and S-06. The clean session-04 route is accepted as the basis for R4. R4 should test whether one child session can carry 3 consecutive unfinished shots while the main session only consumes the child handoff.

Current broad-run issue report:

slides/broad-generation-self-audit-v1.md

Open Decisions

Decision	Default
Character source face	Current stage does not use the real face; use a 50-year-old traveler based on the defined character system.
Character board size	16:9 page, shallow gray-white background.
Character board scope	MVP can start with front/side/back full body plus half-body state; full version should add walking side/back and prop breakdowns.
Style board scope	MVP one-page A/B/C/D board; C should be marked as repaired / accepted.
Full shot production	Paused until R4 formal-run simulation. Current default route: per-shot brief/prompt trace restored, visible character shots use actual references, child sessions may handle up to 3 slides by default, clean user-visible child sessions for orchestration, internal one-shot SubAgents for slide work, one slide/version has exactly one owning worker, and child sessions perform the completion gate before handoff.

Next Decision

Run R4 formal-run simulation:

• Recommended scope: S-02, S-03, S-04.
• Use one clean user-visible child session with up to 3 slides.
• Main session waits 10 minutes after dispatch, then polls every 1 minute up to 15 minutes total.
• Main session reads only the child handoff by default and does not audit prompt / return / ledger / image details when completion gate passes.
• Child session performs mandatory completeness gate. Instruction / visual review is optional and should run only if the task asks for review.
• Start brief/orchestration work as clean user-visible Codex child/session threads and record their thread ids.
• Inside a child session, start one internal SubAgent / one-shot slide worker per slide; do not create long-lived reusable slide agents.
• Add a duplicate-worker guard: one slide/version must have exactly one owning worker and one prompt/image/return target set.
• Keep the quality route from this test: gpt-image-2 Advisor prompt/spec, active style master, active character registry, and visible reference images loaded before generation.
• If R4 succeeds, the project does not need to stay limited to only 3 total new shots; future production can use broader approved scopes or multiple child sessions while preserving the child-session gate.