Parallel Build Orchestrator: Run Agent Waves, Verify Work

Most "parallel" agent setups quietly run one worker at a time: launch, wait, launch, wait. And when a worker says "done," it gets believed. This skill turns your agent into a disciplined build orchestrator that does neither.

What it does

• Detects your host's subagent facility (one-message parallel Task calls in Claude Code, multi-agent delegation in Codex CLI, or spawn-and-await tools) and uses the wave form that actually runs in parallel.
• Verifies the wave plan by intersecting every concurrent pair's file leases, path by path. Anything that could collide is serialised.
• Launches every eligible session at once, then collects. Refills slots as soon as the binding allows.
• Checks every handoff instead of trusting it: counts checkpoint commits in git, flags any commit that wrote outside its lease, preserves uncommitted work, and re-runs the gates itself, scoped to the lease, reading the selected-test count.
• Keeps going when a worker blocks. Only dependents wait. Mechanical gaps (lockfile, manifest entry, test registration, adapter glue) go to small owner-correction workers or a bounded lease revision. You're only interrupted for product-design or destructive decisions.
• Recovers with limits: recovery, replan and checkpoint re-slice workers, at most two attempts per failure shape, and never git reset --hard on shared history.
• Finishes honestly: one read-only verifier scores every requirement verified / present-unverified / failed / blocked, every failed row is routed to an owner, and a committed FINAL-REPORT.md records waves, lease violations, checkpoint shortfalls, granularity feedback and a follow-up closure ledger.

How it works

1. Read the plan, verify leases and gates, print the ledger, pause once for objection.
2. Loop: launch the whole wave → collect → receive each result completely (a strict critical section) → refill.
3. Route blockers through the Human Interruption Gate; auto-decide everything mechanical, and log it.
4. Verification pass → final report → commit.

Frameworks & tools

Any stack. Needs git and a host with subagents (Claude Code, Codex CLI, or any agent with spawn and await tools). Consumes the plan format written by Parallel Build Planner.

Why this beats prompting it yourself

"Run these tasks in parallel" gets you sequential runs, trusted self-reports, and a whole build stalled on one missing dependency. This skill encodes the wave law, lease reconciliation from git, gate re-verification, standing authority for mechanical fixes, and strict verdict rules, so "done" means checked.

Use cases

• Execute a multi-session build plan unattended overnight.
• Run 3–5 Claude Code subagents on one feature without collisions.
• Audit what parallel agents actually committed versus what they claimed.
• Get a final report of what's verified, unproven and still owed.

Known limitations

Requires a host with subagents; without one, it offers to run sessions sequentially. Requires a plan with file leases (use Parallel Build Planner). It doesn't write application code or change product requirements. Wall-clock parallelism depends on your host's concurrency limits.