google-colab-runner
Run video-director productions on Google Colab. It packages a manifest into a self-contained Google Colab bundle (a generated .ipynb that installs ComfyUI and custom nodes at pinned commits, downloads the open-weights models with the user's own Hugging Face token from Google Colab secrets, renders every shot, detects the GPU tier and collects outputs to Drive or a zip), or it lets the director drive a running Google Colab ComfyUI through a tunnel URL with a compute-unit estimate, the spend gate, a hard cap, drop-and-resume and teardown. Use it when the user has no local GPU, or asks to run on Google Colab.
How do I install this agent skill?
npx skills add https://github.com/mojomast/open-higgsfield-skills --skill google-colab-runnerIs this agent skill safe to install?
- Gen Agent Trust Hubpass
This skill enables AI agents to run video-director productions on Google Colab by packaging projects into notebooks or orchestrating remote sessions. It follows security best practices by using user-provided secrets for model access, implementing budget spend gates, and fetching all dependencies from well-known repositories like GitHub and Hugging Face.
- Socketwarn
2 alerts: gptAnomaly
- Snykpass
Risk: LOW · No issues
What does this agent skill do?
google-colab-runner
Two modes, both from video-director:
| mode | command | who presses Run | QA loop |
|---|---|---|---|
| package | direct.py package --target google-colab <manifest> | the user, in Google Colab | after the run: colab.py ingest, then still-qa / shot-qc locally |
| orchestrate | direct.py run <manifest> --target google-colab --url <tunnel> | the user starts the notebook in orchestrate mode; the director submits every job | live: the director's normal judge, retry and continuity loop |
Nothing here starts a Google Colab session or spends compute units by itself, and no credential is read or stored. The Hugging Face token is the user's own, read at run time inside Google Colab from the secret HF_TOKEN.
Files
| path | what |
|---|---|
scripts/colab.py | estimate, package, orchestrate, teardown, ingest (stdlib only) |
scripts/notebook.py | builds the nbformat-4 notebook and the bundle README |
templates/runner.py | the bundle runner: GPU detection, tier pick, upload, submit, poll, download, resumable state.json |
scripts/comfy_client.py | copy of comfyui-connect's client, shipped inside every bundle |
references/models.json | every model file the API workflows load: target folder, pinned Hugging Face URL, size, gated flag, licence note |
references/custom-nodes.lock.json | ComfyUI and node-pack commits the notebook checks out |
references/gpu-tiers.json | T4 / L4 / A100 / H100 → tier, compute units per hour, speed factor (unverified) |
references/google-colab.md | the runbook: step by step for an agent |
references/ui-to-api.md + ui-to-api.report.json | how the Comfy-Org UI templates were converted to API format, and the result per template |
scripts/ui_to_api/ | the converter (headless ComfyUI frontend graphToPrompt) and the /object_info validator |
tests/mock_comfy.py, tests/test_static.py | stdlib mock ComfyUI + VLM and the static end-to-end tests |
Rules
- Ask before any start. Show
colab.py estimate(session minutes and compute units, including setup and the collect-and-stop reserve) and wait for a yes.orchestratestops at thespendgate unless it isapproved, even when the manifest saysauto. - Hard cap.
orchestraterefuses to start without a compute-unit cap (target.budget.max_creditsor--max-cu) and refuses an estimate over it. During the run, the wall-clock meter plus the reserve is checked before every render; at the cap it logscap_hitand stops. The notebook watchdog enforcesMAX_SESSION_MINUTESon the runtime side. - Tear down. Always end by offering
colab.py teardown --url <tunnel>(the watchdog releases the runtime), or use--teardown auto|on-cap. Tell the user that Google Colab bills while the runtime is connected. - Drops resume. A dropped session exits with code 5 and the manifest saved. The user restarts the notebook and gives the new URL; re-running the same command resumes. Finished outputs stay on the user's machine and are re-uploaded to the new session.
- Open weights and licences.
models.jsoncarries each file's licence note. LTX-2 family weights are under the Lightricks LTX-2 Community License (not OSI). The one UI template that needs a closed hosted API node (GeminiImage2Node) is excluded.
Exit codes: 0 done, 3 waiting at a gate, 4 stopped at a stage (awaiting approval or blocked), 5 Google Colab ComfyUI unreachable or dropped, 6 refused by the spend cap.
How can the creator link this skill?
Add the canonical catalog link to the repository README so users can inspect current installs and available audits. The publishing guide covers the complete discovery path.
<a href="https://skillzs.dev/skills/mojomast/open-higgsfield-skills/google-colab-runner">View google-colab-runner on skillZs</a>