video-workbench/handoff/2026-06-24-character-consistency-r4-formal-run-handoff.md

R4 Formal-Run Simulation Handoff

Date: 2026-06-24

Project:

projects/2026-06-23-在路上

Status

The third session / SubAgent workflow test is complete.

R3 accepted results:

• S-33 -> slides/s33/s33-v2-image.png
• S-06 -> slides/s06/s06-v2-image.png

R3 mechanism result:

• Clean child session: 019ef876-c5b2-7202-b204-774cb5fb95fd
• Agent S-33: 019ef87a-3aa1-75a2-a97c-de8f57e234e1
• Agent S-06: 019ef886-8797-7243-9427-d45065fcfc47
• Both slide agents were distinct one-shot internal workers.
• Review files were intentionally absent.
• Prompts recorded gpt-image-2 Advisor mode plus style and character references.
• User accepted both images.

The user will manually commit the current documentation updates. R4 should start in a new clean session after that commit.

Rules Updated Before R4

Read these first in the new session:

AGENTS.md
docs/workflows/slide-task-orchestration.md
docs/workflows/slide-task-templates.md
projects/2026-06-23-在路上/project.md
projects/2026-06-23-在路上/execution-plan.md
projects/2026-06-23-在路上/slides/slides.md

Key runtime rules:

• 子会话 / session: clean user-visible Codex child thread, created with create_thread; do not use forked inherited-context threads by default.
• Agent / SubAgent: internal one-shot worker owned by the child session.
• One slide Agent handles exactly one slide/version, writes one return, then becomes terminal / closed.
• Closed Agents must not be continued or reused.
• Main session production default: dispatch, wait, read short handoff, continue. Do not audit prompt / return / ledger / image details when the child reports complete with completion gate pass.
• Waiting rule: first wait 10 minutes; if incomplete, poll every 1 minute; stop routine waiting at 15 minutes total and inspect/report timeout.
• Child-session completion gate is mandatory.
• Instruction / visual review is optional and should run only when the task asks for review.
• 无 review means no sNN-vN-review.md; return and handoff still record that review was intentionally suppressed.

Completion Gate Expected From Child Session

The child session must verify and summarize:

• each target slide version was resolved by scanning existing slides/sNN/sNN-v*-* files;
• no target file was overwritten;
• expected brief, packet, prompt, image, and return files exist for each slide;
• review files are present or absent according to the task rule;
• each slide/version has exactly one owning Agent;
• each Agent is terminal / closed after return;
• visible-context image references were loaded when required, or generation blocked before text-only identity-critical output;
• generated project image paths and hashes are recorded when available;
• handoff includes a compact parent summary.

If any item fails, return blocked, failed, or warning and name what the main session should inspect.

Slide Inventory

Project-Confirmed Complete / Keep

Total: 11.

Group	Shots	Notes
Core / anchor trial	S-01, S-10, S-46, S-47, S-51	Accepted five-shot visual trial. 阶段五-核心单镜.md only expands S-10, S-46, and S-51; S-01 and S-47 were included by the anchor-test package.
Character-consistency intake	S-05, S-14, S-21, S-25, S-33	These are the five shots from 阶段五-人物一致性单镜.md. S-05 has an overwrite workflow defect but has prompt/review/reference evidence. S-14 image is accepted but has historical duplicate-Agent mechanism defect. S-21, S-25, and S-33 are accepted.
Temporary no-single-shot-skeleton test	S-06	Accepted in R3 as s06-v2; clean session-04 continuation and mechanism pass.

Generated Before But Must Be Redone

These have broad-run v1 images but no reliable brief/prompt/review trace. They cannot prove use of the style master or active references. Keep the existing v1 files as historical evidence only; formal production should create the next append-only version.

S-02, S-03, S-04, S-07, S-08, S-09,
S-11, S-12, S-13, S-15, S-16, S-17,
S-18, S-19, S-20, S-22, S-23, S-24,
S-26, S-27, S-28, S-29, S-30, S-31,
S-32, S-34

Total: 26.

Not Started

These have no slide artifacts yet.

S-35, S-36, S-37, S-38, S-39, S-40,
S-41, S-42, S-43, S-44, S-45,
S-48, S-49, S-50

Total: 14.

Recommended R4 Scope

Use:

S-02
S-03
S-04

Expected new versions:

s02-v2
s03-v2
s04-v2

Reasoning:

• They are the first unfinished consecutive shots after accepted S-01.
• They start real production from the beginning rather than only testing isolated character shots.
• They exercise three useful cases in one child session:
  • S-02: distant small traveler / environment-led human shot.
  • S-03: body / prop detail shot.
  • S-04: landscape-led no-face road shot.
• All three already have broad-run v1 images, so append-only version handling is tested.

Recommended task settings unless user changes them:

running mode: prompt and generate
review: none
completion gate: required
main-session audit: none unless child handoff reports warning / blocked / failed
child-session capacity: 3 slides

R4 Test Goals

R4 should simulate formal production rather than another audit-heavy mechanism test.

Main-session expectations:

• create one clean child session;
• send the R4 packet for S-02, S-03, and S-04;
• wait 10 minutes before checking;
• if incomplete, poll every 1 minute up to 15 minutes total;
• when the child reports complete with completion_gate: pass, continue without opening detailed prompt / return / ledger / image files;
• only inspect details if the child returns blocked, failed, warning, missing completion evidence, duplicate worker, artifact collision, or user-requested audit.

Child-session expectations:

• handle all 3 assigned slides sequentially;
• resolve versions by scanning existing files;
• write one brief and one packet per slide;
• launch exactly one internal one-shot Agent per slide;
• ensure each Agent closes after return;
• write prompt and image for each slide;
• write no review files unless the task explicitly changes review mode;
• run the completion gate;
• return a compact handoff usable by the main session.

After R4

If R4 succeeds, production does not need to remain limited to only three total shots. The next step can either:

• continue in additional 3-shot child sessions, or
• move to a broader user-approved production scope while preserving child-session completion gates and one-shot slide Agents.

The next natural unfinished sequence after R4 would be:

S-07
S-08
S-09

because S-05 and S-06 are already in the keep set.

Non-Goals For R4

• Do not invoke CCPE.
• Do not ask GPT V2 for final prompts.
• Do not call Minimax MCP or external vision/review tools unless the user explicitly changes the task.
• Do not reuse or continue R3 Agent S-33 or Agent S-06.
• Do not backfill old broad-run v1 images as if they had original prompt traces.
• Do not overwrite existing slide artifacts.