/harness-help

You don't remember every skill, so ask. Most work travels one main flow; everything else is standalone. Each skill below is documented in its own skills/<name>/SKILL.md (core) or extras/skills/<name>/SKILL.md — this router only says when to reach for which.

The main flow: idea → shipped

1. /spec <slug> — turn a fuzzy request into spec.md + plan.md under .agent/plans/<slug>/, then route to plan approval (ExitPlanMode). Add --interview when the request is fuzzy enough that a wrong guess would commit the spec to the wrong shape, or --score-candidates when the field of problems/approaches is wide enough that the pick should be scored, not asserted. Approval is what unlocks editing — spec-gate stops asking once the plan-approval flag exists.
2. /supervise <slug> — run the approved plan wave by wave, dispatching specialists per the delegation contract and auditing after each wave. --goal-mode for budgeted, resumable runs.
3. /verify-completion <claim> — before accepting any "done": a fresh context re-checks the claim mechanically, then a refute-by-default judge. /supervise calls this at its completion step; reach for it directly whenever a builder's word is the only evidence.
4. /wrap — commit + PR with the safety gates (gitleaks, risk areas) intact. --auto-push / --auto-merge are opt-in escalations.

A single small edit needs none of this — just make the edit; the gates scope trivial work out by design.

Standalone

• /council-review [--staged|--head|<range>] — multi-vendor second opinion on a diff (Claude reviewer plus codex/gemini lanes), synthesized with citation checks.
• /project-init (command) — scaffold a consumer project with CLAUDE.md, rules, and gitleaks.toml.
• Security — diff-level security review is the security-reviewer agent (auto-dispatched on risk paths); a whole-repository or branch vulnerability scan is the official /claude-security plugin when installed (/plugin install claude-security@claude-plugins-official).

Extras plugin (agent-harness-extras, optional)

Rarely used skills live in a second plugin so the core stays light. Install once with /plugin install agent-harness-extras@agent; its skills appear namespaced as /agent-harness-extras:<name> (the short /<name> works when unambiguous — note loop collides with Claude Code's built-in /loop, so use the namespaced form). Each is documented in its own extras/skills/<name>/SKILL.md:

• /harness-audit — read-only health check of the harness itself: one verify-all.sh dry-run, a per-check table, the doc-reality verdict named.
• /manager-audit <slug> — meta-audit of a /supervise run (restatement, routing waste, token spend, role compliance).
• /persona-review <target> — a panel of grounded Korean citizen personas reacts to a piece of UX/copy/content; a user-perspective lens, not a code review.
• /reorg-sync <old> <new> — after a directory move, sweep orphaned absolute-path references; dry-run first.
• /code-tour [path|range] — write a human reading path through a codebase or a just-finished build; edits no source.
• /worker-setup — install → auth → verify onboarding for the cross-vendor worker lanes (codex, antigravity), with a cost briefing before anything installs.
• /loop <mission> and /harness-loop — bounded-attempt improvement loop (generic, and specialized to this harness's reviewer prompts).
• /brain-ingest — distill quarantined brain raw/ captures into curated notes.

The gates underneath

Skills are the methodology; hooks are the enforcement. When one interrupts you mid-flow, it names its own escape:

• spec-gate asks for a plan on a substantive edit → run /spec.
• tdd-guard blocks new prod code with no failing test → write the test first.
• pre-tool-guard / risk-area mutexes stop destructive or contested operations → the user drives those; don't route around them.

Keeping this router honest

Any skill added to or removed from skills/ or extras/skills/ updates this file in the same commit — a router that lies is worse than no router.