manim-agent

Manim Agent

Use this skill to run the full Manim Agent production workflow, not a simplified one-off Manim snippet. Locate the local repository from MANIM_AGENT_HOME, the current workspace, or a user-provided path; if it is missing, clone https://github.com/gqy20/manim-agent.git before running project commands.

Operating Mode

1. Clarify the requested output only when needed: topic, target duration, audience, voice/TTS need, quality level, and final file path.
2. Run scripts/check_manim_agent_env.py before the first real render in a session, or whenever a failure suggests missing dependencies.
3. Prefer the CLI path for direct video delivery. Use the Web path only when the user asks for task history, SSE progress, browser UI, or backend persistence.
4. Preserve the repository pipeline: planning, implementation, render resolution/review, narration, TTS, and mux. Do not replace it with a handmade scene.py unless the user explicitly asks for a raw Manim scene.
5. Produce concrete artifacts: final MP4, generated scene/code location, logs or error summary, and the command used.

Required Interfaces

• A language-model interface is required for normal pipeline runs. Manim Agent uses Claude Agent SDK to plan scenes and write or fix Manim code; Manim and FFmpeg alone are not enough.
• A working local runtime is required before generation: Python 3.12+, uv, Manim, FFmpeg, claude-agent-sdk, and httpx. Run scripts/check_manim_agent_env.py instead of guessing.
• The LLM interface uses Aliyun DashScope / Bailian Model Studio through the Claude Code compatible route (https://dashscope.aliyuncs.com/apps/anthropic) and a supported model such as qwen3.7-plus. The OpenAI-compatible route is not the right path for this repository's SDK flow.
• These qwen models run in hybrid/extended-thinking mode by default on this route; there is no separate "-nothinking" model id. The repo keeps thinking enabled by default for release-quality generations. Use MANIM_AGENT_THINKING_MODE=disabled only for low-latency smoke tests or when diagnosing provider latency.
• If Phase 1 fails before rendering, check the LLM provider first: expired plan, invalid model name, missing auth token, or incompatible structured-output behavior.
• TTS is optional. Use --no-tts for smoke tests. For narrated output, configure DashScope CosyVoice with DASHSCOPE_API_KEY; do not expect the skill package to contain an API key. Apply for a DashScope/Bailian API key at https://help.aliyun.com/zh/model-studio/get-api-key.
• The speech route is Aliyun DashScope CosyVoice. Default model: cosyvoice-v3-flash; default voice: longanyang. The adapter downloads the returned audio URL and measures real duration before muxing.
• Database and R2 credentials are not required for direct CLI MP4 generation; they are only needed for the Web/backend persistence path.

Reference Routing

• Read references/repo-runtime.md for installation, environment variables, CLI/Web commands, ports, and local paths.
• Read references/pipeline-workflow.md before running or explaining the end-to-end pipeline.
• Read references/production-quality.md before generating or reviewing teaching animation content.
• Read references/recovery-and-review.md when a render, structured output, TTS, mux, or frontend/backend task fails.

Default CLI Pattern

From the local manim-agent repository:

uv run python -m manim_agent "解释傅里叶变换的核心直觉" --target-duration 30 --quality high --no-tts -o outputs/fourier.mp4

Use --no-tts for the first smoke run unless the user explicitly wants narration and a supported TTS key is available. For production narration:

uv run python -m manim_agent "证明勾股定理" --target-duration 30 --quality high --voice longanyang -o outputs/pythagorean.mp4

Default runs do not enable independent AI frame review. Add --render-review only when the user asks for strict visual review, release QA, frame-by-frame inspection, or when a previous render showed overlap, cropping, unreadable math, or other visual risk:

uv run python -m manim_agent "证明勾股定理" --target-duration 30 --quality high --voice longanyang --render-review -o outputs/pythagorean_reviewed.mp4

Delivery Rules

• State whether the run used no-TTS, TTS, render review, intro/outro, full render, or segment render.
• Never claim the video is ready until the MP4 path exists and is readable.
• If a dependency is missing, report the exact missing dependency and the next command to fix it.
• If a secret is provided in chat, use it only for the current run when necessary; do not write it into this skill, logs, examples, or user-visible output.
• If the user asks for a packaged skill or reusable workflow, update this skill rather than scattering notes into the repository.
• Do not expose API keys, database URLs, R2 credentials, or .env values in responses.

Local Repo Awareness

The upstream repo already contains a production plugin at plugins/manim-production/ with scene planning, scene building, layout safety, narration sync, render review, and intro/outro rules. Reuse those rules when working inside the repo. This Codex skill is the stable outer entrypoint: it decides when to invoke the repo, which path to run, what checks to perform, and what evidence to return.