VibeCodeWorker Manual

Open a run on the target that matches the work:

• Catalog games — every 4weird game is a legal run target via the Hub.
• Cloud remotes — rent a Kasm desktop on RunPod for GPU-heavy or off-site targets.
• Your own agent — drive the same loop over HTTP with a login session or gateway key.

1. 01 — Open a run with a game slug and a plain-language goal (1–500 chars).
2. 02 — Play the game yourself in another tab while you record each observe → reason → act step as a run action.
3. 03 — File anything broken as a bug with a severity. Pause by simply stopping; complete the run with a verdict when done.

For targets that need more than your browser, provision an autoplay remote from the Cloud Run page. Open the stream URL, log in with the one-time VNC password, and open the locked game URL in the remote Chromium. Idle remotes warn, then stop, then terminate — stop yours from /runpods when finished.

For video: record the remote stream with OBS or another recorder. VibeCodeWorker supplies the supervised playtest; your recorder creates the final footage.

• Run detail — the full observe → reason → act trail plus findings; read before every next step.
• Export — the uncapped portable archive (2000 steps + 500 bugs) for offline analysis.
• Compare — two runs side by side: verdicts, step-kind mix, bug-severity mix.
• Handoff — a portable markdown brief with a next-actions checklist for any coding tool.

No games listed: the catalog needs a login session — sign in, then retry.

Run won't open: the slug must be a catalog slug and the goal 1–500 chars.

Remote won't start: the provision call answers honestly (unconfigured, no stock, over budget) — fix the reported state and retry; nothing is ever faked.

• “Open the menu, change one setting, return to the game, and check controls.”
• “Walk a short route, rotate the camera, and note frame drops or stuck movement.”
• “Test the sign-up form for validation, keyboard navigation, and visible errors.”