---
name: "Agent playtest harness"
description: "Use when a build needs scripted agent play sessions with screenshots, logs, and a bug list with repro."
---

# Agent playtest harness

## When
Use when a game or interactive build should be playtested by an agent: boot, play N minutes, capture evidence, file bugs. Also use before calling a vertical slice "done."

## Stop
- Separate **builder** and **tester** roles — the agent that authored the build must not be the sole tester of its own happy path.
- Do not file bugs without repro steps and at least one Screenshot or log excerpt.
- Do not "fix while testing" in the same role; hand bugs back to builder.
- Stop on license/auth/paywalls; do not sidestep.

## OVEN
Default: Grok Bot tools — GenerateImage / computerUse / browserUse / Shell / Blender MCP if connected / Screenshot. (Swap this one line for Astra / Fable / Claude.)

## Steps
1. Receive build path, boot command, target platform, and play script (or write a minimal script: boot → move → interact → pause menu → quit).
2. Spawn/assume **tester** role with a clean run folder `playtest/run-YYYYMMDD-HHMM/`.
3. Boot the build via Shell or computerUse. Confirm first frame with Screenshot. Record boot time.
4. Execute the scripted session for N minutes (default 5–10). On each beat: action → wait → Screenshot → note.
5. Capture logs (stdout, engine log, browser console) alongside screenshots.
6. File bugs in `playtest/BUGS.md`: title, severity, repro steps, expected vs actual, evidence paths.
7. Smoke checklist: boot, critical path, pause/quit, no hard crash. Mark pass/fail.
8. Hand `BUGS.md` + evidence to the builder role; tester does not patch in-session.

## Done when
- Run folder contains screenshots, logs, and `BUGS.md`.
- Smoke checklist filled.
- Builder/tester separation respected in the notes.
- Crashes have repro or "non-repro + evidence" explicitly marked.

## Cross-model
Recipe is fixed: scripted boot → timed play → evidence → bugs → handoff to builder. Only the OVEN line swaps — how the host launches builds and takes screenshots changes; role separation and repro requirements stay identical across Astra, Grok Bot, Fable, and Claude.