Carry one assessed cell to a committed, verified recording: check prerequisites, port any missing driver step, drive the live agent CLI under a recording daemon via `of record run`, verify EVERY websocket observation (state + model + cost + tokens + agent) via `of record verify`, refresh the replay golden, and commit. Backflows a correction into the cell when the live recording disagrees with the assessment. Invoked as `/ir:onboarding-factory record <agent> <scenario>`.

Add a brand-new scenario ROW to the matrix: one agent-agnostic `{id, name, description, acceptance_criteria, process}` entry written through `of scenario add`. Researches how the behavior manifests across every onboarded agent (and what the daemon would observe) before synthesizing the agent-agnostic spec. No agent CLI invocation, no recording. Invoked as `/ir:onboarding-factory create-scenario <slug>`.

Onboard a brand-new agent CLI as a matrix COLUMN: research its identity + recording prerequisites, register it via `of agent add`, scaffold its interactive driver from the template, and predict which step types each scenario's recipe will need (the driver-needs punch-list). No live recording. Invoked as `/ir:onboarding-factory create-agent <slug>`.

Judge one (agent, scenario) cell across the three pillars — agent capability, daemon sensor capture, driver capability — on cited evidence, then author the cell's recipe and machine-checkable spec. Writes the cell metadata via `of cell write` and the spec (expected.jsonl) via `of cell spec`. No live recording. Invoked as `/ir:onboarding-factory assess <agent> <scenario>`.

Maintain the canonical scenario × agent fixture matrix for irrlicht. A slim dispatcher that routes intent to four focused subagents — `create-scenario` (add a matrix row), `create-agent` (add a matrix column), `assess` (judge one cell across the three pillars and write its spec), and `record` (drive the live agent and verify every websocket observation). Every read and every write goes through the `of` factory CLI (tools/onboarding-factory) — the skill itself never touches `replaydata/`. Each subagent returns a ≤6-line summary so the parent keeps its context for strategic decisions instead of drowning in per-cell tool output. Use when the user says "/ir:onboarding-factory", "onboard agent", "add a scenario", "assess fixtures", "record fixtures", or "regenerate recordings".

Scan the web for coding agents and agent orchestrators, track GitHub stars and trends, rank by popularity+momentum, and publish a report to the irrlicht site. Shows which agents irrlicht already supports. Use when user says 'agent landscape', 'scan agents', 'coding agent tracker', 'agent popularity', '/ir:agent-landscape', or wants to see the competitive landscape of coding agents.

Triage a single open GitHub issue from a PM lens. Applies a 'Triage In-Progress' label during triage; reads code, checks for already-shipped work, returns a verdict in {close, decompose} with confidence and a flip-condition. Renders and stops — no other side effects.

Phase 1: Start — begin a new feature. Creates a worktree, upgrades dependencies, opens a PR, creates .flow-states/<branch>/state.json, and configures the workspace. Usage: /flow:flow-start <feature name words>

Display the FLOW skill catalog grouped by user role. Maintainer and Private buckets render only when invoked inside the FLOW plugin repo.

Phase 3: Review — six tenants assessed by four cognitively isolated agents (reviewer, pre-mortem, adversarial, documentation) launched in parallel. Parent session gathers context, triages findings, and fixes.

Wipe `.flow-states/` on this machine in one pass. PRs, worktrees, and branches are NOT touched — those require per-flow `/flow:flow-abort`.

One-time project setup — configure and commit workspace permissions, install bin/* stubs, and write the version marker. Run once after installing or upgrading FLOW. Usage: /flow:flow-prime

Decompose a problem statement into a pre-planned decomposed issue. Accepts either an issue reference (#N, re-plans in place) or a bare prompt (synthesizes What/Why/AC and files a new issue). Runs a Tech-Lead-default planning conversation, dispatches to PM/Tech Lead/CTO sub-agents on explicit user request, then files or edits the issue ready for /flow:flow-start. Usage: /flow:flow-plan #N or /flow:flow-plan <topic>

Process decomposed issues sequentially overnight via flow-start, tracking outcomes and generating a morning report.

Capture a correction or learning to the FLOW state file. Invoke explicitly with /flow:flow-note. Fast — captures and continues without interrupting flow.

Phase 4: Learn — audit rule compliance and identify process gaps. Routes findings to CLAUDE.md, .claude/rules/, and plugin issues.

Group open issues by label into four sections (Blocked, Other, Vanilla, Decomposed) with mechanical sort and a copy-pasteable command per row.

Audit instruction corpus health — CLAUDE.md, rules, and memory for staleness, misplacement, duplication, and contradictions.

Open a problem-statement conversation. Stays in discussion mode with PM as default voice; on user signal, files a vanilla What/Why/Acceptance Criteria issue against the current repo. Usage: /flow:flow-explore <topic>

Full codebase documentation accuracy review — reports drift between code behavior and documentation.

Clear the autonomous-flow halt set when the user spoke mid-flow. Invokes `bin/flow clear-halt` so the next assistant turn resumes execution. User-only: the model cannot invoke this skill.

Display the current FLOW configuration from .flow.json — version and per-skill autonomy settings.

Phase 5: Complete — merge the PR, remove the worktree, and delete the state file. Final phase.

Review the full diff, then git add + commit + push. Use at every commit checkpoint in the FLOW workflow.