Irrlicht: Skill for Claude Code

create-scenario

You run as a focused subagent with no parent context. Everything you need is in this file and the repo. Do the research yourself (web + file access) — don't bounce work back to the dispatcher. This task spends NO API tokens on agent CLIs and runs NO recording. When done, return only the summary in the "Return contract" section.

What this does

Adds one new agent-agnostic scenario so later verbs (assess, then record) have a row to fill. A scenario is defined ONCE, agnostic to any particular agent; per-agent verdicts, recipes, specs, and recordings come later from assess and record. After you run, the new scenario is a row of unknown cells — nothing claimed about any agent yet.

This is the matrix-ROW counterpart to create-agent, which adds a COLUMN.

The scenario schema (5 fields, nothing else)

of scenario add writes exactly this shape into replaydata/agents/scenarios.json:

id                   <section>.<index>  — e.g. "2.22". Stable; orders the matrix.
name                 kebab slug         — e.g. "mid-turn-message-queued". The FK.
description          one paragraph      — what behavior, why it matters.
acceptance_criteria  markdown           — what a recording must show to pass.
process              markdown           — how to drive an agent to elicit it.

There is no section/feature/requires/verify/idle_only any more — those were dropped in the factory cutover. Applicability is decided per-cell by assess, not by a requires gate on the row.

Inputs

• <slug> — kebab-case scenario name (stable; becomes the FK and the recording folder stem). E.g. mid-turn-message-queued.
• A one-paragraph description of the behavior. If the dispatcher didn't pass one, derive it from the slug and state your assumption in the summary.

Steps

1. Pick the id

List the catalog through the factory (never read the file directly):

of status --json | jq -r '.scenarios[].id' | sort -t. -k1,1n -k2,2n