auto-ai-canvas-skill-mode

Auto AI Canvas Skill Mode

This skill processes executable AI Canvas Skill requests submitted from the canvas Skill panel. Current executable scope covers xiaohongshu-cover, youtube-thumbnail, cross-platform-adapt, product-marketing-set, logo-and-brand, and marketing-brochure.

Tool Availability Gate

This workflow must use watch_skill_requests, submit_canvas_skill_request, create_image_version, import_image_asset, save_snapshot, and update_skill_request. It only uses apply_canvas_actions when an older request explicitly includes overlayActions. If these tools are not callable in the current Codex thread, stop and tell the user:

AI Canvas 插件已经识别到，但 Skill 处理工具没有加载出来。如果你刚安装或刚更新过插件，请完全退出并重新打开 Codex，再发送同一句需求；插件不需要重新安装。

Do not recreate the listener with shell scripts, curl, port checks, or manual service startup during normal use. Those actions are only for explicit plugin development/debugging.

Workflow

1. Tell the user briefly: 我会等画布里的 Skill 请求。你在画布里选中图片，选择 Skill，再点“提交给 Codex 生成”。
2. Call watch_skill_requests with waitMs around 30000-45000 and claim: true.
3. If the result times out with no request:
   • Continue polling while the user expects automatic Skill handling.
   • Do not stop because of one idle timer.
   • If you say anything while waiting, make clear that Codex is waiting for a canvas Skill submission, not generating yet.
4. When a request arrives:
   • If canAutoGenerate is false or (inputImagePath is missing and skillId is neither logo-and-brand nor marketing-brochure), call update_skill_request with needs_clarification and tell the user to re-import the image via upload, drag, or paste.
   • If skillId is not one of xiaohongshu-cover, youtube-thumbnail, cross-platform-adapt, product-marketing-set, logo-and-brand, or marketing-brochure, call update_skill_request with needs_clarification; that Skill is not part of the executable loop yet.
   • If skillId is cross-platform-adapt, product-marketing-set, logo-and-brand, or marketing-brochure and briefStatus is needs_input, do not generate images. Call update_skill_request with needs_clarification, then ask the user the clarificationQuestions in normal conversation, at most 3 questions at a time. After the user answers, merge their answers into the original brief, briefly summarize the output plan, ask for confirmation when the plan is ambiguous, then call submit_canvas_skill_request again with the same skillId and the completed brief.
   • Build jobs = generationJobs when present and non-empty; otherwise use [generationJob].
   • Process jobs sequentially. For each job:
     • Use job.prompt as the generation/editing prompt. If inputImagePath exists, use it as the reference image; if it is missing for logo-and-brand or marketing-brochure, generate from the text brief alone.
     • Generate exactly one result image for that job using the available Codex image-generation/image-editing capability.
     • Follow the job prompt carefully. For 小红书, YouTube, cross-platform adaptation, product marketing, Logo/brand, and marketing brochure jobs, generate the final designed image directly: platform-safe recomposition/cropping, title typography when requested, colors, layout, decoration, brand/campaign system details, and visual hierarchy must be baked into the raster image.
     • Do not create separate canvas text boxes for title design unless the request explicitly includes overlayActions.
     • Save the generated image at job.outputDir/job.outputName when a local output path is needed.
     • If targetShapeId exists, call create_image_version with:
       • sourceShapeId = targetShapeId
       • imagePath = generated local image path
       • placement = "right"
       • title = job.title
       • runId = runId
       • x = job.placement.x
       • y = job.placement.y
       • w = job.placement.w
       • h = job.placement.h
     • If targetShapeId is missing, call import_image_asset with:
       • inputPath = generated local image path
       • source = "codex_generation"
       • title = job.title
       • placement = "absolute"
       • x = job.placement.x
       • y = job.placement.y
       • w = job.placement.w
       • h = job.placement.h
       • selectAfterCreate = true
     • Remember the output path and returned newShapeId or shapeId.
     • For multi-job Skills, call update_skill_request with status processing after each completed job and include partial progress in result, so the user can see that the sequence is moving.
   • If the request includes overlayActions, call apply_canvas_actions with those actions for backward compatibility. New 小红书 and YouTube Skill requests normally should not include overlay actions.
   • Call save_snapshot.
   • Call update_skill_request with completed and include every new shape id, run id, output image path, job id, and overlay result in result.
5. Continue watching only if the user asked for continuous Skill handling. Stop when the user says to stop, the thread is interrupted, or a blocking clarification is required.

Stop Behavior

When the user says 停止监听, 停止处理 Skill, 先停一下, 不用继续等了, or similar:

1. Stop polling watch_skill_requests.
2. Reply clearly:

已停止处理 Skill。

你可以继续在画布里查看；但现在提交 Skill 只会保存任务，不会自动开始生成。

以后要继续时，回到 Codex 说：AI Canvas 继续处理 Skill。

User-Facing Tone

Be direct and product-like:

• "我会等画布里的 Skill 请求。"
• "收到 Skill 任务，我开始生成。"
• "这个产品套图还缺少几个关键信息，我先问清楚再生成。"
• "这个品牌简报还缺少几个关键信息，我先问清楚再生成。"
• "这个营销宣传册还缺少几个关键信息，我先问清楚再生成。"
• "这个跨平台适配还缺少几个关键信息，我先问清楚再生成。"
• "结果已经放到原图右侧，旧图保留。"
• "这张图没有本地文件路径，请用导入、拖拽或粘贴重新放入画布后再提交。"

Avoid exposing raw JSON unless the user asks to debug.