Process Storyboard Tasks

Process Agent Storyboard Tasks

Task recovery

While processing a long render, call heartbeat_storyboard_generation_task at least every 10 minutes. Tasks expire after 30 minutes without a heartbeat. If the user released a task or its ID changed, stop attempting to complete it; never attach a late output to a replacement task automatically.

Process the local storyboard queue. The MCP tools start the bundled local app automatically when needed and default to http://127.0.0.1:43218.

If the current agent session does not expose Storyboard MCP tools such as list_storyboard_generation_tasks, claim_storyboard_generation_task, or complete_storyboard_generation_task: in Codex, use tool_search to search for agent storyboard and load the deferred tools; in Claude Code, check that the agent-storyboard MCP server is connected (/mcp). Only if the tools still cannot be found, tell the user to start a new conversation or restart the agent so plugin tools are reloaded.

Mandatory B-roll motion gate

For every B-ROLL shot routed to hyperframes or remotion, plan_broll_motion is mandatory. Do not claim the task, write a composition, download media, render, or complete the task before this gate is confirmed. The MCP server rejects both claim and completion when the confirmed plan is missing.

Reuse-first template policy (mandatory)

Treat the local template roots and the two required repositories as the primary production library. Before designing a new composition, identify a concrete reusable template, component, shot skeleton, GIF/MP4, or other directly usable media candidate from the inspected sources. If a source has no suitable candidate, record that explicit empty result and the reason. Select the closest existing candidate, reuse it directly whenever it fits, and make only the minimum changes needed for the shot's dialogue, visual hierarchy, palette, typography, and duration.

Do not start from a blank composition when an existing candidate can carry the shot's core semantic structure. If no candidate fits, adapt the closest existing skeleton and record the original path plus the mismatch and adaptation boundary. A from-scratch motion system is an exception: use it only after the plan records the local roots and both repositories searched, the concrete mismatch, why adaptation would fail, and the smallest new structure required. The confirmed plan must carry this evidence in selectedTemplate, motionSkeleton, sources, and researchNotes.

Before asking for confirmation, research the shot in this order:

1. Read every available local B-roll template under the project's template roots or the active workspace template directory. Record the roots checked, including an explicit empty result.
2. Inspect both required reference repositories, even when a local candidate exists: https://github.com/heygen-com/hyperframes-launches and https://github.com/Vincentwei1021/video-shotcraft. For each repository, record a concrete candidate path or an explicit "no suitable candidate" result with the reason.
3. Choose the best reusable candidate across the local roots and repositories. Prefer direct reuse; if it needs changes, keep the adaptation minimal and preserve the source skeleton's relationships and phase order.
4. When an external reference is useful, inspect public X posts and their attached media. Prefer a directly usable public GIF/MP4/video over a screenshot; use a screenshot only when the media cannot be accessed. Record the source URL and a rights/source note, and never invent an account, post, quote, or asset URL.

Call plan_broll_motion with approval: "proposed" and include the shot's duration, dialogue, audience takeaway, brollType (有素材 / 无素材 / 纯文字), and semanticStructure (对比 / 聚合 / 筛选 / 层级 / 因果 / 替换 / 展开). Name the selected reusable local or repository template/component/media path first; if adapting, name the source skeleton and the exact mismatch; include any direct media sources and only the minimum UI changes: background, anchor color, font, and personalization. Show that proposal to the user and wait for confirmation.

After the user confirms, call the same tool again with approval: "confirmed", researchComplete: true, both required repository URLs in reviewedSources, and complete selectedTemplate, motionSkeleton, and uiChanges. Only then continue to claim and implement the task. If no suitable local or external motion exists, report the gap and ask for a decision rather than silently inventing a new motion system.

Workflow

1. Call list_storyboard_generation_tasks with status pending.

2. Inspect the generators used by the pending tasks and verify the matching local capabilities before claiming anything:

   • image-gen requires an image generation capability. If your agent has its own (Codex: the built-in imagegen skill and tool), use it. If it does not (for example Claude Code), use the generate_storyboard_image tool, which generates through the local Codex CLI; check inspect_storyboard_environment first and confirm the "AI 生图" item is ready.
   • hyperframes requires the HyperFrames and HyperFrames CLI skills.
   • remotion requires the Remotion skill and its local rendering toolchain.

   If a required capability is unavailable, do not claim affected tasks. Report the exact missing capability and continue with tasks whose generators are available.

3. Probe the local environment with read-only checks before deciding to install anything:

   • Check Node.js, npm, and pnpm availability and versions.
   • Check whether the task outputDir, active workspace, or chosen render directory already has package.json, a lockfile, and node_modules.
   • Check for local CLIs with existing project commands such as pnpm exec ..., npm run ..., or executable files under node_modules/.bin.
   • Check for global CLIs only with non-installing commands such as command -v, where, or direct --version after the command is found.
   • Check whether Chromium, FFmpeg, or tool-specific render caches already exist when the selected renderer needs them.

   Do not use commands that can implicitly download packages as probes. Do not run npx remotion, pnpm dlx, or package-manager exec forms that will install missing packages merely to test availability.

4. Prefer existing render environments. Use local project dependencies first, then compatible global CLIs. Install only when a required tool is missing, incompatible, or no local dependency set exists. Any install must be explicit and must go into the task outputDir or a dedicated renderer cache, not into the user's unrelated project.

5. Process available tasks one at a time.

6. For eligible B-roll tasks, complete the mandatory motion gate above. For all other tasks, continue directly to claiming after capability checks.

7. Before generating, call claim_storyboard_generation_task.

8. If the claimed task has hasDesign: true, read the complete Markdown file at the exact absolute designPath before generating anything. Apply it as the project-wide visual system:

   • visualPrompt defines the concrete shot subject and requested content.
   • DESIGN.md defines shared visual style, color, typography, composition, texture, and motion language.
   • An explicit shot requirement takes precedence if it conflicts with the general visual system.

9. Route by generator:

   • image-gen: with your own image generation, use the built-in imagegen skill and tool. Without one, call generate_storyboard_image with the taskId instead: it claims a pending task, applies DESIGN.md and the reference image, keeps the task alive, verifies the file, and completes or fails the task by itself (skip steps 7, 10 and 11 for that task; expect 1 to 6 minutes per image and process tasks one at a time). Treat visualPrompt as the primary prompt and honor the task's aspectRatio. If the task includes referenceImagePath, use that local image as the visual reference/input for the generation or edit. Copy the final verified image into the active workspace before completing the task.
   • hyperframes: use the HyperFrames and HyperFrames CLI skills. Start from the selected local or repository template/component/media skeleton, reuse it directly or make the smallest documented adaptation, then create the composition with the task's width, height, duration, and visualPrompt. Lint, inspect, render to MP4, and verify the output.
   • remotion: use the Remotion skill. Start from the selected local or repository template/component/media skeleton, reuse it directly or make the smallest documented adaptation, then render an MP4 with the task's width and height matching the task duration, and verify the output.

10. Technically verify each generated video before completion: the file exists, is readable, has the requested width and height, has a duration close to the task duration, and has a valid video stream/codec. Then visually inspect representative frames to ensure the render is not blank, black, or the wrong composition.

11. Call complete_storyboard_generation_task with the exact absolute output path and correct media type.

12. If generation or verification fails, call fail_storyboard_generation_task with the concise cause.

13. Continue until no processable pending tasks remain.