fix: make review repairs converge

[Arismemo]

Summary

• make storyboard repair runs read and minimally edit the existing storyboard table
• inject the previous full supervision report into repair and re-review contexts
• prevent repeated confirmations and moving review criteria
• align long-dialogue and optional-sound rules between generation and supervision
• make script-stage cross-workspace repairs update storySkeleton first and adaptationStrategy second
• enforce the missing downstream repair in code before supervision can run
• pass the current user instruction into supervision and freeze already-confirmed tradeoffs
• exempt single-episode videos up to 60 seconds from long-form paywall and reversal rules
• keep story-skeleton and adaptation-strategy executors scoped to their own workspaces
• avoid storing empty XML-only execution memories
• support OpenAI-compatible image and video generation, including storyboard references
• cap reference-video duration at the endpoint-supported 10 seconds
• cap video prompts at 4096 UTF-8 bytes without splitting multibyte content
• preserve structured upstream error codes and messages, including content moderation failures

Root causes

Storyboard review loop

The storyboard-table agent never read the current storyboardTable during repair runs. Every repair regenerated the whole table from script, assets, and plan data, so fixed items regressed while the independent reviewer found different issues each round. The reviewer also did not receive the prior report as a frozen acceptance checklist.

Script adaptation review loop

One user confirmation could require changes to both storySkeleton and adaptationStrategy, but the decision agent called only one executor and then claimed both workspaces were synchronized. The stale downstream strategy contradicted the updated skeleton. Generic per-stage instructions also allowed supervision to run after the first workspace, while the reviewer repeatedly reopened confirmed tradeoffs and applied long-form paid-series rules to a one-episode 30-second children's short.

OpenAI-compatible video failures

The adapter discarded string-valued upstream errors, so actionable failures appeared only as HTTP 400. Live endpoint reproduction confirmed three independent constraints: reference-to-video duration is limited to 10 seconds, prompt length is limited to 4096 UTF-8 bytes, and content moderation failures return a structured imagine:content-moderated response.

Verification

• yarn test:script-agent-review — 6/6 passed
• yarn test:openai-image — 5/5 passed
• yarn test:openai-video — 7/7 passed
• yarn lint
• yarn build
• git diff --check
• live ToonFlow test on project 1785157202433: story skeleton executor completed, adaptation strategy executor completed, then one final review ran
• final live script review returned A — 可直接使用 and explicitly closed the confirmed shoe-lace tradeoff instead of reopening it
• verified both runtime workspaces were synchronized and old active semantics such as aiming, staggering, and arrow-in-mud were removed
• verified runtime execution memories are non-empty after deploying the user-data runtime copy
• live endpoint reproduction for duration, prompt-length, and content-moderation failures
• retried the affected local project sequentially and verified 19/19 video tracks have a successful MP4
• validated all 19 MP4 files with ffprobe (H.264, 848×480, non-zero duration)

Target

develop

[HBAI-Ltd]

Thank you for submitting the pull request. We will test it and evaluate the merge as soon as possible.