Play ONE session of a System Shock 2 mission via the headless debug runtime, as a QA agent. SEE the world (screenshots you read) and ACT with the runtime's tools (move, look, frob, pick up, attack, follow real triggers) toward a goal - observe and react, don't run a script. Emit a structured data.json (per-step screenshot + what you saw + what you did + any bug, plus the frontier reached and the issues found) so the run renders to an HTML timeline and drives the play-through loop. This is the atomic unit; the `play-through` manager runs it in a fix-and-replay loop. Invoke with a mission (default medsci1), a start state (fresh, or resume at a frontier), and a goal.
Manager loop that hardens the game toward fully-playable end-to-end. It drives the `playtest` primitive in a PLAYTEST to REVIEW to FIX to REPLAY loop: play one session (agent plays a mission via the debug runtime, emits a data.json), adversarially REVIEW that session against a real walkthrough (did it genuinely progress, were the actions sensible, are the findings real - not artifacts of poking the wrong thing), triage the validated blocker/bugs, file + delegate a fix (PR "Fixes #n"), then REPLAY from the advancing frontier - until a mission (then the game) plays through with no new blocker. Aggregates every session into a self-contained HTML timeline report. Invoke with a mission (default medsci1) and a goal (default: reach the level's exit).
Capture a looping GIF + a still PNG of a visual/rendering change and embed them in its pull request. Uses the headless debug runtime (cargo dbgr + HTTP /v1/step + /v1/screenshot - no window interaction, deterministic fixed-timestep stepping), assembles a GIF with PIL, hosts the binaries in a gist (pushed via git, the only token-friendly host for binary images), and embeds them in the PR body. Use whenever a change adds or alters something visible (a viewmodel/HUD/ rendering/material/lighting feature, a debug scene, an animation) and you're opening or updating a PR - see AGENTS.md "Visual Changes".
Record an mp4 video of the headless debug runtime by capturing periodic screenshots and stitching them with ffmpeg. Deterministic fixed-timestep: the sim runs at 60Hz, so capturing a frame every 4 sim-frames yields real-time 15fps. Drive movement (turn/walk via /v1/control/input) while capturing for smooth footage - a pan of a scene, a walk down a corridor, or a record of a playtest session. Local artifact (not hosted), so file size is fine. Used standalone to grab a clip, or by the `playtest` / `play-through` skills to record a session.