# Scenario Runner Capture

> Implement deterministic Scenario Runner + capture outputs for agents

- Skill: `majiayu000/scenario-runner-capture` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds add majiayu000/scenario-runner-capture`
- Raw SKILL.md: https://api.skillmd.com/api/skills/majiayu000/scenario-runner-capture/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: majiayu000 (https://skillmd.com/u/majiayu000)
- Updated: 2026-09-09
- Page: https://skillmd.com/skills/majiayu000/scenario-runner-capture

---

# Skill: Scenario Runner + Capture

## Objective
Allow agents to validate gameplay and visuals without controlling the OS window.

## Steps

1) Implement a Scenario Runner autoload
- `game/autoload/ScenarioRunner.gd`
- Parse CLI args after `--`:
  - `--scenario <id>`
  - `--capture_dir <dir>`
  - `--seed <int>`
  - `--quit_after_frames <int>`

2) Define scenario scripts
- Store scenarios under `tests/scenarios/` (JSON)
- Actions should include:
  - wait_frames
  - move_to
  - interact_nearest
  - trigger_encounter
  - select_battle_command

3) Capture outputs
- Always write a `trace.json` of executed actions + timings
- Primary: use deterministic movie capture mode in Godot CLI (Movie Maker)
- Secondary: checkpoint screenshots if stable

4) CI scripts
- Implement `tools/ci/run-scenario.sh <scenario_id>`

## Verification
- Scenario runs headlessly
- Artifacts appear in `captures/` or specified capture dir
- Runs are deterministic across repeated runs on the same machine