--- name: still2rig-psd description: Convert a user-provided static anime character image into a verified rig-ready PSD by coordinating the Still2Rig PSD local job CLI with a user-approved Google Colab session through Colab MCP Go. Use for Still2Rig PSD conversion, repair, import, or QA requests; do not use for unrelated PSD editing or general browser automation. --- # Still2Rig PSD Use the project CLI as the source of truth for job state and generated Colab cells. Run commands from the repository root. ## Preflight and job creation 1. Run `npm run doctor`. Resolve missing local prerequisites without modifying global Codex configuration. Installing software still requires normal user approval. 2. Run `npm run still2rig-psd -- prepare [--name ]`. 3. Confirm the copied input hash in `.still2rig-psd/jobs//job.json` before any remote execution. ## Colab consent boundary Before connecting, tell the user that the next MCP call will wait for their Colab tab. Run `npm run still2rig-psd -- colab-url` and give that URL to the user so they can open it in the intended signed-in Chrome profile and select an L4 GPU. - Use only the `colab-mcp` tools supplied by Colab MCP Go. - Do not use Chrome control, an in-app browser, Computer Use, or UI automation. - Do not select an account, accept terms, allocate a runtime, or change a GPU on the user's behalf. - Call `open_colab_browser_connection` only after clearly announcing it. The MCP approval prompt is the authorization gate. - Never call `disconnect_colab_runtime` or delete the runtime. Leave release of compute to the user unless they explicitly request it in the current turn. - Stop after three connection failures and report the exact failure; do not switch to another browser or MCP implementation silently. ## Execute the generated cells After connection, call `list_colab_tools`. Use the notebook tools it reports (normally `add_code_cell`, `run_code_cell`, and `get_cells`; names can vary). Add and run these exact generated files in order: 1. `01-upload.py` 2. `02-setup.py` 3. `03-run.py` 4. `04-download.py` Read a cell with `npm run still2rig-psd -- cell `. Do not rewrite the cell ad hoc. Wait for completion and inspect outputs after each stage. Stop on Python exceptions, input-hash mismatch, wrong GPU, revision mismatch, or a missing output PSD. The final cell starts a normal browser download; it does not transfer bytes through the MCP response. ## Import and finalize Once `still2rig-psd-.zip` is present locally, run: ```bash npm run still2rig-psd -- import /path/to/still2rig-psd-.zip npm run still2rig-psd -- finalize ``` If the user provides full-canvas, registered expression files named `mouth_open.png`, `mouth_close.png`, or `eye_close.png`, pass their directory as `--expressions `. Use `--preview-placeholders` only when the user wants a non-production preview. Never describe copied or missing expressions as real lip-sync or blink art. Choose the expression source in this order: 1. Drawn expression files from the user: pass them with `--expressions `. 2. An available image generation tool (for example, Codex image generation): generate full-canvas `eye_close.png` and `mouth_open.png` registered to the source, pass them with `--expressions `, and report them as generated art. Require `expressionRegistration=true` and `expressionArtwork=true` before calling them ready. 3. Only when neither is available, as a stopgap: run `npm run still2rig-psd -- derive-expressions ` and rebuild with `repair --preview-expressions `. The derived layers reuse the upper lash line and draw a small fixed-shape mouth, so they are less accurate than generated or drawn art. The reports list them under `previewExpressions` and keep `productionReady=false`. For mouth repair, prefer extracting the closed-mouth ink from the registered original artwork. Do not fill the component interior or carry a skin-colored rectangle into an expression layer. A derived opposite mouth state is preview art, not trustworthy production art. Render both mouth states at full-canvas scale and as a face crop, then require `expressionArtwork=true` and `mouthWithinTolerance=true` in the QA report before reporting the repair. ## Motion QA Run `npm run still2rig-psd -- motion-qa ` after the PSD is final. It records the contract captures in the built-in preview renderer and scores them. Run it again with `--negative-fixture` when the checks themselves changed; that run must FAIL. Only a passing `motion-qa` report for the current PSD SHA-256 supports a motion, seam, or hair-motion claim, and only for the built-in renderer. ## Completion report Report the job id, input SHA-256, pinned See-through revision, PSD path and SHA-256, contact sheet, structural QA result, expression readiness (including any `previewExpressions`), and the `motion-qa` result or that it was not run. A structurally valid PSD is not automatically a Live2D- or renderer-ready rig. Tell the user that `npm run preview` starts the built-in local PSD preview. It lists finalized outputs and opens the newest PSD without exposing internal job details in the UI. Do not launch a browser automatically. For failure modes and the exact trust model, read `../../../docs/codex-workflow.md` and `../../../docs/security.md` only when they are relevant.