| name | agent-all |
| description | Cost-unrestricted multi-agent pipeline. Drives intent→plan→wave-dispatch→gate→PR over the .claude/agents/ roster, with optional --loop until a shell break-condition succeeds (bounded by --max-iter and --max-cost). Requires /agent-init scaffolding. |
/agent-all
Runs a complete multi-agent pipeline from a free-form prompt or an existing task file. Phase 3 fan-out delegates to superpowers:subagent-driven-development. Phase 6 optionally loops the entire run.
Usage
/agent-all "add user signup form"
/agent-all docs/tasks/12-fix-login.md
/agent-all "fix flaky test" --loop --max-iter=5
/agent-all docs/tasks/x.md --no-pr --wave-size=large
Flags
--loop — enable Phase 6 looping. Interactively prompts for the break-condition
preset (Test auto-detect / visual-qa skill / Custom shell / Composite) on first
use, then offers to save the choice to .agent-all.json.--max-iter=<N> — cap loop iterations (default from config, hard cap 50).--max-cost=<USD> — cap accumulated cost.--break-condition=<spec> — non-interactive override for the loop break
spec. Accepts either a JSON object (e.g. '{"type":"visual-qa"}') or a
plain shell string (treated as {type:"shell", cmd:<string>}).--reconfigure — force the interactive break-condition prompt even when
.agent-all.json already has a non-default value.--qa — shortcut for --break-condition='{"type":"composite","steps":[ {"type":"test-auto"},{"type":"visual-qa","mode":"comprehensive"}]}'. Tests
run first as a cheap gate; visual-qa (comprehensive mode) runs as the
final end-to-end check. Auto-scaffolds .visual-qa.json with sane
defaults if missing. Takes priority over --break-condition, the
interactive prompt, and any saved config value.--wave-size=small|medium|large — override config default.--no-pr — skip Phase 5 (PR creation).--no-brainstorm — skip Phase 1's brainstorming for free-form prompts.--resume — skip phases already complete per .agent-all-state.json.--force — wipe state and restart.--yes — skip all interactive confirms (including the break-condition prompt;
falls back to the config or built-in default).
Pipeline
| Phase | File | Purpose |
|---|
| 0 | phases/0-preflight.md | git+roster+config+input checks |
| 1 | phases/1-intent.md | brainstorming OR load task file |
| 2 | phases/2-plan.md | writing-plans for the task |
| 3 | phases/3-dispatch.md | wave-builder + subagent-driven-development |
| 4 | phases/4-gate.md | wave-level spec+quality reviews |
| 5 | phases/5-pr.md | branch push + gh pr create |
| 6 | phases/6-loop.md | breakCondition + stableIters + maxIter/Cost |
Rules
- You orchestrate; phases are source of truth. Read each phase file before running it.
- State lives in
.agent-all-state.json. Shape: {phases:[{phase,completedAt}], task, plan, waves[], iter, costUSD, prUrl, decisions:{<task-id>:{<decision-id>:{chosen_index,auto_resolved,timestamp,reasoning?}}}}. --resume resumes after max(phases[*].phase).
- Delegate, don't reimplement. Phase 1 calls
superpowers:brainstorming; Phase 2 calls superpowers:writing-plans; Phase 3 calls superpowers:subagent-driven-development. Your code is a thin coordinator.
- Loop is opt-in. Without
--loop, Phase 6 is a no-op.
- Hard caps:
--max-iter clamped to 50 server-side; --max-cost enforced after each wave.
- Orchestrator routing.
/agent-all is for durable, gated, PR-shipping code changes. Evidence-producing work (research, multi-unit audits, design/findings reports) belongs to the built-in Workflow (ultracode) tool instead — they are siblings, never nested. Phase 1 gauges this; the full decision table, the no-nesting constraint, and the governance-across-the-seam rule live in references/orchestrator-routing.md.
Lib modules
lib/config-loader.mjs — loadConfig(path) → {ok, config | errors, warning?}. Returns built-in DEFAULTS when path missing.lib/wave-builder.mjs — buildWaves(tasks, waveConfig) → array of waves.lib/gate-plan.mjs — buildGatePlan({files,gates,taskId,title}) → ordered coordinator/reviewer dispatches with audit-token, gate-reason, and pass-criteria contracts.lib/loop-evaluator.mjs — evaluateLoop(state, limits, runner) → {action: "break"|"continue"|"exhausted", consecutivePass?, exitCode?}.
On error
- Dirty git tree → abort.
.claude/agents/ missing → abort + suggest /agent-init..agent-all.json missing → warn + use built-ins.- writing-plans fails → abort.
- Wave task BLOCKED 3× → Phase 3 abort with exit code 2.
--max-cost exceeded → finish current wave, abort, preserve state.- Loop maxIter exhausted → exit code 3, last commit preserved.
When done
Print summary: phases completed, iters, cost, PR URL. Exit code 0/1/2/3 per spec.
