Video Generation

Video Generation

Default path (minimize tool calls)

Do not read_file / ls this skill on every request. Do not run models.py unless handling fallback after failed.

1. python <skill_dir>/scripts/submit.py "<compact_json>" → get task_id
2. python <skill_dir>/scripts/poll_until_done.py <task_id> — one exec, timeout=600
3. python <skill_dir>/scripts/download.py "<video_url>" "<short_name>" "" "<task_id>" — must pass task_id as 4th arg
4. Single message(media) with local file path — see Silent policy

Exactly 4 tool calls. Do NOT add read_file, list_dir, write_file, or any other calls. Do NOT write your own scripts or call APIs directly — always use the provided .py scripts.

Image upload is automatic. When user sends an image for image-to-video, submit.py handles all URL resolution internally — no env vars or manual URL construction needed.

Nanobot limits: exec is capped at 600s per call. poll_until_done.py defaults to 540s internal wait. If it times out while still processing, run python <skill_dir>/scripts/poll_until_done.py <same_task_id> 480 again (vary the second arg so the command is not identical — do not re-submit). Consider raising agents.defaults.maxToolIterations in ~/.deskclaw/nanobot/config.json from 20 to 35–40 so long flows do not hit max iterations.

Silent policy

No user-visible status text until the final delivery: no "please wait", "polling", or intermediate narration. Only the final message() with media (and minimal content if required).

Token & URLs

• Token: auto from ~/.deskclaw/deskclaw-settings.json → authToken, or NODESTUDIO_TOKEN.
• NodeStudio base: NODESTUDIO_URL (default https://nostudio-api.deskclaw.me).
• Naming: Models are Seedance 2.0 (pro), Seedance 2.0 Fast (fast), and Seedance 1.5 Pro (1.5-pro).

Models (no pre-flight models.py)

Selection rules (in order):

1. User explicitly names a model → use it.
2. User asks for "best quality" / "highest" / "最好" / "最高画质" → use pro.
3. User asks for "1080p" or "高清" → use 1.5-pro (only model supporting 1080p).
4. User has no preference → use fast.
5. If the chosen model fails, fall back: fast → pro → 1.5-pro; only call models.py if all fail.

Do not run models.py before the first submit.py.

Build JSON (compact, no spaces after : or ,)

Required: prompt.

Image-to-video

When the user sends an image, the message contains ![image](media/<session>/<file>). Pass the relative path as image_url:

{"prompt":"...","image_url":"media/agent_main_desk-xxx/abc123.jpg"}

submit.py automatically resolves local paths: public API → presign-upload to CDN; local API → gateway /files/.... Do not set DESKCLAW_GATEWAY_HOST, construct URLs, or retry with different env vars — the script handles everything.

Parameters

Submit

python <skill_dir>/scripts/submit.py "<json_body>"

Windows note: Use " (double quotes) for the JSON argument, with internal \" escaping. Single quotes do not work on Windows cmd.

Poll

python <skill_dir>/scripts/poll_until_done.py <task_id> [max_seconds]

• Default max_seconds 540 (stays under 600s nanobot exec kill).
• Pass exec(..., timeout=600).
• Do not run poll.py in parallel with poll_until_done.py.
• progress 50 while processing is normal (placeholder), not a stall.

Resume: If output contains poll_until_done_timeout and task not failed, re-run with same task_id and a different second arg (e.g. 400). Never duplicate the exact same command string (gateway may block).

Fallback (only on failed)

1. python <skill_dir>/scripts/models.py → pick next alias from list (fast → pro).
2. submit.py again → poll_until_done.py new task_id.

Download

NANOBOT_SESSION_KEY='agent:main:desk-xxx' python <skill_dir>/scripts/download.py "<video_url>" "<short_name>" "" "<task_id>"

• 4th arg task_id is required — enables re-poll for signed URL and API proxy fallback when CDN returns Access Denied.
• 3rd arg is output_dir (empty string "" = auto from session key).
• Script saves to workspace outputs/ and prints {"status":"ok","path":"..."}.
• If download fails, returns {"status":"download_failed",...}.

Present (single message)

Silent until here. message({ "content": "…", "media": ["<local_path>"] }).

Critical rules:

• media must contain the local file path returned by download.py (path field). Never put a remote URL in media.
• Never put CDN URLs in content — they are not publicly accessible and will show Access Denied to the user.
• If download failed, say "视频生成成功但下载失败，请稍后重试" in content, leave media empty. Do NOT paste the raw URL.
• Model name: Use Seedance 2.0, Seedance 2.0 Fast, or Seedance 1.5 Pro.

Error Handling

Rules

• Python scripts only: submit.py, poll_until_done.py, poll.py, download.py, models.py (fallback only).
• Do NOT write custom scripts, Python files, or call APIs with curl / urllib yourself.
• Do NOT read_file the scripts — just exec() them.
• Do NOT set or change DESKCLAW_GATEWAY_HOST — the scripts auto-detect the correct upload method.
• Do NOT call mcp_deskclaw_restart_gateway — gateway restarts break the current session.
• If submit.py returns an error about image upload, report it to the user and stop. Do not retry with different env vars.
• JSON compact in arguments.