Multi-agent patterns,
ready to drop in.

concurrency — one number decides the shape. 1 (default) is the original sequential runner, one ticket at a time, unchanged. N (autonomous only) runs independent tickets — the ones the dependency DAG says don't block each other — as parallel lanes, scheduled by the deterministic workflow.
/start-milestone docs/prd/01-foundation autonomous 4   /start-all autonomous 4

/breakdown-prd writes docs/prd/dag.html — every ticket in one dependency graph, colored by module. A self-contained file: double-click it, no server, no build step. It tells you the concurrency worth passing instead of making you guess, and flags a module that can only ever use one lane — that is a file-scope decomposition problem, and Gate 1 is the cheapest moment to fix it.

15 tickets, 4 modules — pick a lane count and watch the same graph re-shape. This board is computed at build time by the pipeline's own scheduler, so it is the schedule you would actually get.

The PRD document splits by phase. The ticket tree never does. Write the next phase as its own PRD and point /breakdown-prd at it — it decomposes into the same docs/prd/, because /start-all schedules one global DAG and a dependency that crosses phases only resolves inside it. Then run the same command again: everything already delivered has a closed issue and filters itself out.

Three commands, no migration:

docs/PRD-02-billing.md                    # write the next phase
/breakdown-prd docs/PRD-02-billing.md     # appends modules; delivered work is frozen
/start-all autonomous 2                   # only the new tickets run

One tree, two phases — toggle to see what the second run actually schedules. Both boards are computed at build time by the pipeline's own scheduler. Watch BIL-1: it is blocked_by a phase-1 ticket that already shipped — the edge every argument for splitting the tree would break.

The tracker is not on the critical path unless you put it there. Pass none and every ticket merges to your local default branch — no push, no PR/MR, no tracker.

/start-all autonomous 1 none

Every delivery defect this catalog has recorded lives at the forge boundary — a pipeline gate, a protected branch, a 403 MR API, squash-merge ancestry, an expired token — and each one stopped a whole run. Review is unchanged: a ticket still only merges on CLEAR. What is deferred is publication, not judgement.

Install this pattern and the Claude one in the same repo. They do not collide — .claude/ + CLAUDE.md sit beside .codex/ + .agents/ + AGENTS.md.

npx agent-templates@latest adopt three-agent-architect-builder-reviewer .
npx agent-templates@latest adopt codex-three-agent-architect-builder-reviewer .

So teammates with different tools work the same project, and one person can switch runtimes on a token budget without development stalling. No hybrid pattern is offered — it would add a third scaffold to maintain and a model table pinning two vendors at once, for no capability these two installs lack.

Read this before adopting. The hub reviews diffs written against a contract it wrote itself, in the same session. That review is not independent — it is what this pattern trades away for cost and throughput, and it will not catch a wrong brief that was faithfully implemented. If a bad merge is expensive, use the three-agent pattern instead.

One brief = one disjoint write-set. Not a feature, not an effort estimate. Two briefs may run at once exactly when the files they own do not overlap, so the decomposition is a partition of the filesystem — and the audit checks it against what the executor actually wrote.

In the committed rehearsal a 70-line PRD with 4 requirements became 4 briefs of 57–73 lines, each owning exactly one file. A brief is longer than the PRD section it implements, on purpose: the executor starts cold and is forbidden to design, so the contract has to be in the brief.

4 briefs, 2 waves — the real schedule from the committed rehearsal, recomputed here by the driver's own scheduler. Wave 2 cannot start until wave 1 is merged, because its worktrees fork from the default branch.

The commands do not carry across — adopt installs one pattern's commands, and the scripts the other pattern's commands call are not present. The names differ on purpose: a reader who knows /breakdown-prd expects an independent reviewer downstream, and reusing the name would import that expectation into a pipeline that deliberately has none.