run-planlisted
Install: claude install-skill atretyak1985/swarmery
# Purpose
Turn a finished plan directory into shipped code with one command. The main-session
agent acts as **controller**: it parses the plan's phase DAG, picks an execution route
per the triage table, dispatches executor subagents with the plan's own copy-paste
prompts, reviews each result, and keeps durable progress. The plan is the spec; this
skill is the runner.
**Hard constraint — main session only.** Subagents cannot spawn subagents, so this
playbook can never be delegated to an agent. It is invoked via `/run-plan` (or when
the user says "run the plan") and executed by the main loop.
# Step 0 — Locate and load the plan
1. Plan dir = `$ARGUMENTS` if given; otherwise the newest `working/**/{slug}/plan/`
under the project workspace. If more than one plausible candidate exists, ask.
2. Parse the DAG: run `${CLAUDE_PLUGIN_ROOT}/bin/plan-manifest.sh <plan-dir>`.
It prints `manifest.json` if the planner emitted one (authoritative), else a
best-effort derivation (`"source": "derived"`) from the phase docs + README
sequencing table. On a derived manifest, sanity-check `depends_on` against the
README's table before trusting parallelism.
3. Read the plan README fully (risks, Definition of Done). Do NOT bulk-read all
phase docs into context — each phase's content travels to its executor as a
file reference, not as pasted text.
4. **Pre-flight review** (from superpowers:subagent-driven-development, which this
skill follows where the two overlap): scan on