!
Purpose & Capability
The declared purpose (local ComfyUI image generation) matches the included files and scripts (ComfyUI client, workflow templates, prompt agents). However some files hardcode host.docker.internal:8188 rather than consistently using the declared COMFYUI_API_URL primary env var, and helper scripts (switch-painter-model.sh) call docker exec on openclaw-gateway which touches other platform components—these behaviors go beyond a simple local renderer and are not fully justified by the skill description.
!
Instruction Scope
Runtime instructions and included scripts perform many privileged actions: cloning repositories, pip installing PyTorch and dependencies, downloading large model files, copying tools into the OpenClaw workspace, and running docker exec / docker restart on openclaw-gateway. SKILL.md and SOUL.md also contain prohibitions (e.g., '禁止修改 workspace') that conflict with the setup script which writes to the workspace, indicating internal inconsistency in what the agent is allowed to do.
ℹ
Install Mechanism
There is no formal install spec but a comprehensive setup.sh is included which clones GitHub repos and downloads models from HuggingFace (and mirrored hosts). Hosts are recognizable (github.com, huggingface.co, hf-mirror.com, ghfast.top) so not obviously malicious, but the script triggers heavy network activity and large downloads and installs (torch, requirements) without further sandboxing—this is expected for a local ComfyUI installer but high-impact.
ℹ
Credentials
The skill declares a single required env var (COMFYUI_API_URL), which is appropriate. However several scripts ignore that env var and hardcode http://host.docker.internal:8188; other parts expect particular filesystem layout (/home/node/.openclaw, ~/ai-studio) and Docker containers (openclaw-gateway). The skill does not request API keys, but it references external LLM aggregator services (ofox.ai) and a Telegram bot—these references imply external network interactions though no credentials are requested.
!
Persistence & Privilege
always:false (good). Nevertheless the included scripts perform persistent changes: copying tools/workflows into the agent's workspace, creating start scripts, downloading models, and scripts that invoke docker exec/docker restart on openclaw-gateway (which can alter running gateway/container state). Modifying other platform components (openclaw-gateway) is significant and not limited solely to the skill's own files.
What to consider before installing
This skill is plausible for local ComfyUI-based image generation, but review and take precautions before running its setup scripts:
- Inspect scripts (setup.sh, paint-dispatch.sh, quick-generate.sh, switch-painter-model.sh) line-by-line. They will clone repos, pip install torch and other packages, and download multi-gigabyte model files.
- Note that some scripts hardcode http://host.docker.internal:8188 instead of using COMFYUI_API_URL; if your ComfyUI runs elsewhere update those invocations first.
- The setup and helper scripts copy files into ~/.openclaw/workspace and create ~/ai-studio, and switch-painter-model.sh runs docker exec / docker restart on the openclaw-gateway container—only run these if you trust the environment and have backups (or run in an isolated VM/container).
- The skill references external hosts (GitHub, HuggingFace and mirrors, ghfast.top, hf-mirror.com, ofox.ai, a Telegram bot). Expect network traffic to those endpoints when installing or when invoking model-switching features.
- If you only want the prompt-generation/runtime behavior, avoid running the full setup; instead run the specific scripts you reviewed or adapt the Python tools to point to your existing ComfyUI instance.
- If unsure, run the installer in a disposable VM/container, or manually perform the steps after verifying each remote URL and command. Do not provide unrelated credentials—no extra API keys are required by the skill itself, but gateway/container operations may rely on credentials held by your system.
