wantsong
/
video-workbench
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
video-workbench/projects/2026-06-23-在路上/tasks/2026-06-24-character-consis.../task-plan.md

11 KiB
Raw Blame History

Character Consistency Test

Status

  • Task id: 2026-06-24-character-consistency
  • Worker profile: video-slide-production
  • Running mode: prompt and generate
  • Status: test complete; s21-v3, s14-v2, and s25-v2 accepted by user; continuation mechanism has a duplicate Agent S-14 defect to fix before scaling
  • Child session reuse test: reset; start a new child session
  • Deprecated child session id: 019ef77d-552b-7343-abe4-88921b35d7ea (do not reuse)
  • Session-02 child-session id: 019ef7ca-413a-7662-9633-7fb126b7ef35
  • Artifact policy: append-only next slide version; overwrite disabled by default
  • Session-02 resolved version: s21-v2 (runtime fallback; no prompt/image)
  • Session-03 resolved version: s21-v3
  • Thread-level child test: session-03
  • Session-03 child thread id: 019ef7d2-1a7f-7a90-8c04-0098239dfe0f
  • Session-03 slide-agent thread id: 019ef7d6-3ff6-7213-b68c-fd646e39f1d2
  • Session-03 continuation recorded Agent S-14: 019ef803-95d7-7411-8d3c-c7f2f68a32e7
  • Session-03 duplicate Agent S-14 observed by parent: 019ef803-328a-7e40-a0b0-0333c6b5fc6c
  • Session-03 continuation Agent S-25: 019ef803-e757-7041-8f4a-fc9dfce23bbd

User Instruction

Run the character-consistency workflow as a process and identity-consistency test.

Earlier s05 work exposed two workflow problems:

1. The child session did not start the slide-agent itself.
2. `s05-v1-image.png` was overwritten even though a broad-run candidate already existed.

The reset rule for the next run is:

  • do not reuse child session 019ef77d-552b-7343-abe4-88921b35d7ea;
  • start a new child session;
  • require the new child session to start the single-slide S-21 agent itself;
  • use a new slide artifact version resolved from existing slides/s21/s21-v*-* files;
  • do not overwrite existing slide artifacts;
  • explicitly disable review for this S-21 test so the workflow can verify that no review file is created when review is disabled.

Slide Scope

Completed previous test:

s05

Authorized next:

s21

Planned later, only after user confirms s21:

s14
s25

Child Session Split

session-01: historical previous attempt; do not reuse
session-02: s21 fallback evidence only; no prompt/image
session-03: s21 accepted -> s14/s25 accepted; duplicate S-14 worker defect recorded

Current session-03 continuation generated s14-v2 and s25-v2 after user accepted s21-v3-image.png; user then accepted both images and closed the test. The mechanism result is not a clean pass because two S-14 workers were created for one slide/version.

Continuation rule:

  • use the same child thread 019ef7d2-1a7f-7a90-8c04-0098239dfe0f;
  • child thread must start two separate one-shot slide workers: Agent S-14 and Agent S-25;
  • do not reuse or continue Agent S-21;
  • each worker handles exactly one slide;
  • each slide resolves the next unused version by scanning existing files;
  • expected next versions from current scan are s14-v2 and s25-v2;
  • do not overwrite existing broad-run v1 images.

Session id record:

session-01 child session agent id: 019ef77d-552b-7343-abe4-88921b35d7ea (deprecated / do not reuse)
session-02 child session agent id: 019ef7ca-413a-7662-9633-7fb126b7ef35
session-03 child thread id: 019ef7d2-1a7f-7a90-8c04-0098239dfe0f
session-03 Agent S-21 thread id: 019ef7d6-3ff6-7213-b68c-fd646e39f1d2
session-03 recorded Agent S-14 thread id: 019ef803-95d7-7411-8d3c-c7f2f68a32e7
session-03 duplicate Agent S-14 thread id observed by parent: 019ef803-328a-7e40-a0b0-0333c6b5fc6c
session-03 Agent S-25 thread id: 019ef803-e757-7041-8f4a-fc9dfce23bbd

Continuation result:

  • s14-v2-prompt.md, s14-v2-image.png, and s14-v2-slide-agent-return.md exist; no s14-v2-review.md exists.
  • s25-v2-prompt.md, s25-v2-image.png, and s25-v2-slide-agent-return.md exist; no s25-v2-review.md exists.
  • s14 prompt evidence records gpt-image-2 Advisor / host-native mode, visible side/front character refs, and style module B | 高位观察.
  • s25 prompt evidence records gpt-image-2 Advisor / host-native mode, visible side/front/body refs, and style module A | 清晨泥泞旷野 adapted for dust/detail.
  • User accepted both s14-v2-image.png and s25-v2-image.png.
  • Defect: child handoff records only Agent S-14 019ef803-95d7-7411-8d3c-c7f2f68a32e7, but parent thread inspection also found duplicate Agent S-14 019ef803-328a-7e40-a0b0-0333c6b5fc6c; the S-14 return/image source path records the duplicate thread folder. Treat step 2 as generated-with-defect, not as workflow pass.

Shared Context

  • AGENTS.md
  • VIDEO_WORKBENCH.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/visual-system/visual-system.md
  • projects/2026-06-23-在路上/visual-system/style/style-prompt-master.md
  • projects/2026-06-23-在路上/visual-system/characters/character-reference-registry.md
  • projects/2026-06-23-在路上/intake/阶段五-人物一致性单镜.md
  • projects/2026-06-23-在路上/slides/slides.md
  • projects/2026-06-23-在路上/slides/broad-generation-self-audit-v1.md
  • investigations/2026-06-24-s14-reference-capable-generation/reviews/summary-and-rule-implications.md

Visual-System Usage

  • Style continuity must use visual-system/style/style-prompt-master.md.
  • Character continuity must use visual-system/characters/character-reference-registry.md.
  • Reference-capable generation must use the validated host built-in reference-by-visible-context path:
    • write the prompt/spec through the gpt-image-2 Skill Advisor workflow;
    • load the real reference image into the conversation context before generation;
    • call host image_gen with the loaded image assigned a role such as Image 1;
    • record the input path, output path, and reference-path caveat in the review or return.

Rules

  • Do not invoke CCPE.
  • Do not ask GPT V2 for final prompts.
  • Do not use Minimax MCP or other external review services.
  • Do not treat old broad-run images as accepted finals.
  • Do not generate visible-character identity shots from text-only prompts.
  • Do not use archived character or style references unless explicitly reopened.
  • Each slide agent handles exactly one slide.
  • Child sessions and slide agents run sequentially.
  • One slide/version must have exactly one owning slide-agent. A child session must not create two workers with the same packet and same output targets; prompt/image/return ownership must stay with the same worker id.
  • Project-specific allocation override for identity-heavy character-consistency work: assign at most 2 visible-character slides per child session. Keep the generic docs default at maximum 3 slides per child session for non-expensive/general slide work.
  • Child session carrier rule: brief/orchestration child sessions should be Codex child/session threads visible to the user and recorded by thread id. Slide agents inside those child sessions should be internal SubAgent / one-shot workers, not long-lived reusable child sessions.
  • Main session records the new child session id so later slides can be sent to the same child session after user approval.
  • Resolve the slide artifact version before each slide run by scanning existing slides/sNN/sNN-v*-* files.
  • Do not overwrite existing slide artifacts. If any planned target file already exists, stop and report an artifact collision.
  • Do not backfill missing prompt or brief files for old broad-run images as if they were original generation inputs.
  • For the next s21 test, review is explicitly disabled. A s21-vN-review.md file should not be produced unless the task plan is changed.

Expected Outputs

Task records:

projects/2026-06-23-在路上/tasks/2026-06-24-character-consistency/task-plan.md
projects/2026-06-23-在路上/tasks/2026-06-24-character-consistency/task-log.md
projects/2026-06-23-在路上/tasks/2026-06-24-character-consistency/session-02/child-session-plan.md
projects/2026-06-23-在路上/tasks/2026-06-24-character-consistency/session-02/child-session-handoff.md
projects/2026-06-23-在路上/tasks/2026-06-24-character-consistency/session-02/packets/s21-vN-slide-agent-packet.md
projects/2026-06-23-在路上/tasks/2026-06-24-character-consistency/session-02/returns/s21-vN-slide-agent-return.md
projects/2026-06-23-在路上/tasks/2026-06-24-character-consistency/session-03/child-session-plan.md
projects/2026-06-23-在路上/tasks/2026-06-24-character-consistency/session-03/child-session-handoff.md
projects/2026-06-23-在路上/tasks/2026-06-24-character-consistency/session-03/packets/s21-v3-slide-agent-packet.md
projects/2026-06-23-在路上/tasks/2026-06-24-character-consistency/session-03/returns/s21-v3-slide-agent-return.md

Slide records:

projects/2026-06-23-在路上/slides/s21/s21-vN-brief.md
projects/2026-06-23-在路上/slides/s21/s21-vN-prompt.md
projects/2026-06-23-在路上/slides/s21/s21-vN-image.png
projects/2026-06-23-在路上/slides/s21/s21-v3-brief.md
projects/2026-06-23-在路上/slides/s21/s21-v3-prompt.md
projects/2026-06-23-在路上/slides/s21/s21-v3-image.png

User Review / Sampling Plan

After s21 generation, stop and report:

  • new child session id;
  • S-21 slide-agent id;
  • files written;
  • resolved output version;
  • whether the child session started Agent S-21 itself;
  • whether review was correctly suppressed;
  • whether the prompt used gpt-image-2 Advisor, active character reference, and the D style module;
  • whether the same child session should receive s14 and s25 next.

For the s14 / s25 continuation, report:

  • whether the existing child thread continued work;
  • Agent S-14 and Agent S-25 thread ids;
  • whether the two new agents are separate from Agent S-21;
  • files written for s14-v2 and s25-v2;
  • whether prompts used gpt-image-2 Advisor and active character/style references;
  • whether review files were or were not produced, matching the continuation packet rule;
  • whether generation completed for both slides.

This reporting gate is closed for session-03: both images were accepted by the user, review files were correctly absent, and the remaining work is workflow repair before the next batch.

Stop Conditions

  • Required reference image cannot be loaded into the conversation context.
  • gpt-image-2 mode check no longer returns a host/advisor path.
  • Host image generation cannot be called.
  • Generated image cannot be copied into the project slide folder.
  • Any planned output path already exists and overwrite is not explicitly authorized.
  • The result clearly fails the side/back dynamic character-continuity purpose.