Resolve the canonical repository path, ticket and requirements, acceptance criteria, applicable instructions and ADRs, target branch, remote, and the user's existing authority for implementation, commits and publication. Record the authorized delivery endpoint (for example, reviewed commit or pushed target). Readiness alone does not authorize a ticket. If a blocking decision or required authority is missing, prepare a concrete handoff and ask only for what is missing.
Inspect repository status and worktrees; fetch the target remote and pin its commit as the starting base. Record a unique run ID, canonical checkout status, resolved remote URL, target branch and initial review base. Use unique run-owned branches and worktree paths outside the canonical checkout.
Run bd prime in the beads tracker, claim the ticket with bd update --claim, and record the run in beads. This workflow requires bd: beads notes are the durable handoff store every stage reads and writes. Confirm ownership before dispatch. Discover the available agent tools at runtime. Prefer distinct stage agents; for capacity constraints read recovery.

Preparation is complete when scope, endpoint, ownership, base and required human decisions are settled and recorded.

Dispatch and gate

Send each stage the handoff below and its reference when it starts:

Implementer: require a committed candidate and checks.
Reviewer: require independent assessments against the exact candidate and review base.
Review-fixer: required findings enter the fix loop; otherwise pass through the approved commit. Every changed candidate returns to reviewer. Allow three fix-and-review rounds after initial review; persist the count across interruptions. Remaining required findings after round three block delivery: preserve the candidate and report them.
Merger: verify the authorized endpoint before closing.

For interruption, target movement, conflicts, capacity limits or cleanup, load recovery only when needed.

Handoff record

The coordinator writes each transition to the ticket's beads notes or comments; workers return evidence for that record. Temporary logs supplement these durable facts. Every dispatch and return includes:

Ticket ID and pointers to canonical requirements, instructions, ADRs and gates.
Run ID, repository, stage branch/worktree, remote, target and authorized endpoint.
Starting base, current review base, candidate SHA and approval SHA/base (or pending).
Agent roles and authorship, stage status, required findings with stable IDs, advisory observations, dispositions and unresolved decisions/limitations.
Exact check commands, working directory, candidate SHA, outcome (passed, failed, unrun), and evidence location. Mark evidence fresh or cached; cached evidence names its original commit/environment and why it still applies. Record why any check was unrun. Mandatory missing evidence blocks approval.
Fix-round and target-recovery counters, owned worktrees/processes, next action.

Finish

Report the delivered commit and endpoint, Standards/Spec/platform outcomes, verification limits, preserved work and next ticket (from bd ready, if available). Starting that ticket, deploying, or publishing anything further requires its own authorization.

Companion skills

The stages invoke companion skills by role, not by path, so the pipeline runs with whatever the repository has installed. Nothing here is bundled; the gating discipline above holds regardless. Skills that fill these roles well:

Implementation standards — a skill encoding the repository's architecture and conventions. Without one, the repository's own documented conventions are the standard.
Code review — supplies the independent Standards and Spec axes the reviewer stage requires, for example code-review (npx skills@latest add mattpocock/skills).
Platform/architecture review — an extra axis against acceptance gates, if the repository defines one.
Merge conflicts — used only by recovery on target movement, for example resolving-merge-conflicts from the same collection.

afk-workflow

AFK worktree workflow

Prepare