| name | to-implement |
| description | The execution engine — runs ONE unit (an approved chat plan, a raw task, a single-slice spec, or one ticket) through its full lifecycle in an isolated worktree: ready-check, brief, verify-brief, waves of implementer+reviewer pairs, slice review, ADR, integrate. The body of /tier-1, and directly invocable to skip the tiers; /tier-2 feeds it one ticket per fresh chat. Use to build an approved plan, a task, or a ticket. |
Present what is about to run and wait for the user's explicit go before creating any worktree or dispatching any seat:
- the unit — the plan, task, spec, or ticket, in one line;
- the base commit the worktree will branch from, named as a commit the user can see.
Nothing is branched, dispatched, or integrated until the go is given.
Satisfied when a driver reached you and its gate already carried both facts:
/tier-1 — its G3 gate presented the plan and the base commit and took the go;/tier-2 — the ticket was approved at /to-tickets; present the base commit in one line as you branch, no stop.
Typed directly with a raw task, the gate is live.
Autopilot exception. If DEV_STACK_AUTOPILOT=1 is set in the environment (check with printenv DEV_STACK_AUTOPILOT), this gate is already satisfied — the ticket was approved at /to-tickets and the batch at /autopilot. The base is the integration branch's tip; landing is /finish-branch Option 1 to that branch with no menu; any state that would normally escalate to the user instead halts the run for resume.
Waves run unchanged — tickets are sequential, tasks inside one are not. One rule is stricter: you may not park a finding at a cap. Parking is a judgement the user delegates by being present, and unattended it is a bar quietly lowered — so at any cap, and on any finding that conflicts with the brief, halt without writing the sentinel. The /autopilot skill owns the exact base/branch/sentinel contract.
To Implement
One unit, start to finish. You are the orchestrator — you gate, dispatch, commit, and integrate. You never edit code, and you are the only one who touches git. Every seat below is a subagent dispatched by agent name, never an inline role prompt.
The unit is a vertical slice; the brief cuts it into horizontal tasks; each task is built by its own implementer and judged by its own reviewer. You hold the loop over tasks — the seats never see each other.
Input — one unit
Whatever you are handed is built as a single unit. The engine does not decompose.
| Input | What it is |
|---|
| a plan | the approved inline plan from chat — the body of /tier-1 |
| a task | a raw task handed straight here, skipping /route-me and the tiers |
| a spec | a single-slice spec (/tier-2 can output a spec with no tickets) |
| a ticket | one ticket |
A feature split into several tickets is run one ticket per fresh chat — the loop over tickets is the user's, not the engine's. A dependent ticket already sees its blockers' code, because they are merged into the branch this worktree cuts from: the contract is the code read from the tree, never paper from another ticket.
The lifecycle
0. Ready to go
Read-only, seconds, on an empty context — and it closes early what otherwise surfaces after an expensive dispatch or, worse, at merge:
- the working tree is clean (
git status --porcelain empty);
- no merge, rebase or cherry-pick in progress;
- the base branch is not behind its remote — if it is, the eventual merge is a rebase, and the user should know now;
- no stale worktree already exists for this unit;
- the unit resolves and reads — the ticket, the plan, the spec;
- the unit's blockers are actually merged into this tree. Today a missing blocker surfaces as the brief-writer returning BLOCKED, after an opus dispatch has been paid for. Checking here is nearly free.
Anything failing is the user's call, presented at the gate above.
1. Isolate
Cut the worktree from the tip of the current branch — /using-git-worktrees. Record the base commit: it is the review's fixed point, and it is never HEAD~1 (which drops every commit but the last of a multi-commit unit).
2. brief-writer → the brief
Dispatch the brief-writer into the worktree. It explores the code itself and writes .scratch/brief-NN.md, returning only the path — or BLOCKED. You never read its exploration; only the path travels on.
3. Verify the brief — yourself
Read the brief and confirm it is load-bearing. Never dispatch an implementer against a hollow brief: with several implementers building from it at once, a hole becomes N agents guessing in parallel.
- Blocks present — The slice, Files, Interfaces, Binding Decisions, Plan, Execution Boundary, Verify, Sweep — with
Verify named and shown already-run, Global Constraints verbatim, no placeholders.
- The cut. Is each task one functionally whole horizontal layer? A single task covering a slice that plainly has layers is the defect this engine exists to prevent — send it back. So is a task per function.
- Every task self-sufficient. Real values, named seams, no step that sends its implementer into the code to find a location, name or approach. Its
Consumes names every signature it needs, including those a sibling task produces.
- The task graph. Does every
Consumes from a sibling carry a matching Blocked by? Do two tasks with no edge between them name the same file? Either answer wrong is a brief defect, not something to work around — an undeclared edge means building against nothing, and a shared file means two writers colliding.
- Closure. Could a competent implementer make a different locally-valid semantic choice, broaden into an adjacent surface, or cross an escalation trigger without the brief stopping it? Then
Binding Decisions, a task's Out of scope, or Execution Boundary is hollow.
Send a hollow brief back to the brief-writer with the specific defect. Five rounds maximum — past that the brief-writer cannot see its own problem, and grinding is worse than asking. Escalate to the user with the rounds and what stayed broken.
4. Build the waves
The brief carries the tasks; you do not re-plan them. Read the list and execute it. If the list looks wrong, that is step 3's business — go back, don't fix it here.
Compute the wave
A wave is the set of tasks that can run side by side right now:
- take every task whose
Blocked by entries are all closed;
- from those, keep a set whose
Files are pairwise disjoint; anything left over waits for the next wave.
Both conditions matter and they catch different things: the edges stop a task building against a signature that doesn't exist yet, the file check stops two agents editing one file. A wave of one is a normal and correct outcome — layers usually chain.
Dispatch the pairs
Each task is a pair: an implementer builds it, then a task-reviewer judges it. Each dispatch carries exactly two things — the brief's path and the task number. Both seats read the brief for the rest, and the reviewer builds its own diff scoped to the task's Files, so nothing large passes through your context.
Dispatch every implementer in the wave in one message so they run concurrently. As each returns, dispatch its task-reviewer — reviewers are read-only, so they safely run alongside implementers still working.
Handle each implementer's status:
- DONE → its task-reviewer.
- DONE_WITH_CONCERNS → read the concerns first; resolve correctness or scope before the reviewer, carry observations to the slice review.
- NEEDS_CONTEXT → supply what was missing, re-dispatch — then fix the brief so the gap doesn't recur for the tasks still to come.
- BLOCKED → triage; most "too hard" is something else in disguise. A gap the brief should have closed → treat as NEEDS_CONTEXT. An open decision with several valid answers → the user's call, never the seat's. The task too large → the cut was wrong, back to step 3. The unit itself wrong → escalate. Only when none fit is the seat genuinely too small: propose the same implementer on a stronger model and wait for the user's go — spend sits behind a gate. Never force an unchanged retry.
A failed task does not stop its siblings. They are file-disjoint by construction, so they cannot corrupt each other; let the wave finish. Only closed tasks are committed. The next wave waits on its own blockers, not on the whole previous wave — if the failed task doesn't block it, it runs. A task that exhausts its fix loop is BLOCKED and escalates before any wave that depends on it starts.
The fix loop, per task
Triggered by any Critical or Important finding, or a failed conformance verdict. Minor findings never enter it — record them and hand them to the slice review to triage.
Five rounds maximum per task:
- rounds 1–3 — the same implementer, findings verbatim. Its context is intact: it knows the task, the code, and its own choices, so it pays only for the findings.
- rounds 4–5 — a fresh implementer on a stronger model, carrying the brief path, the report path, and the framing: a prior implementer attempted this N times; you own it now, read the report for what was tried. A loop that survives three resumes usually means the implementer cannot see its own problem — fresh eyes and a capability bump in one move.
- Every round ends with the same task-reviewer contract, re-run against the amended code. The reviewer verdicts each finding; "attempted" is not addressed.
At the cap, stop dispatching and adjudicate each open finding yourself:
- the reviewer is wrong, or the point is contestable → park it with your ruling, and hand it to the slice review;
- real but nothing downstream builds on it → park it the same way, ruled real and deferred;
- real and load-bearing — a later task builds on it, or it exposes a brief defect → stop. Escalate to the user with the finding, the brief text it collides with, and the fix history.
Adjudicate only at the cap. Adjudicating earlier to end a loop is pre-judging with a different name, and a silently dropped finding is forbidden.
A finding that conflicts with what the brief mandates is the user's call: present it beside the brief text and ask which governs. brief-mandated is a label, not a defence — the brief's authorship does not grade its own work.
Commit the wave — you, and only you
Every task in the wave closed? Then commit, and you are the only one who touches git. No agent stages, commits or amends anything: one worktree means one index, and two agents staging at once produces a commit silently carrying another's half-written files.
- Check the fence mechanically. Compare
git status --porcelain against the union of the wave's tasks' Files. Anything modified outside that union belongs to no task — that is a boundary violation found without an agent looking for it. Resolve it before committing.
- Stage by path — the wave's declared files, never
git add -A.
- One commit for the wave, naming the tasks:
feat(x): wave 2 — transport layer, client wiring.
Fixes land before the commit, not after: a commit of unreviewed work means "wave N" no longer marks a closed wave.
The commit list plus the task reports in .scratch/ are the run's memory. If your context is compacted mid-run, recover from those — git log for the closed waves, .scratch/report-NN-task-K.md for the tasks closed inside the current one. Trust them over your own recollection, and never re-dispatch a task whose report is already on disk.
Then compute the next wave, until every task is closed.
5. /verified-review — the slice
Raise /verified-review once, after the last wave, with the fixed point set to the base commit from step 1.
Once, and not per wave: until the final wave lands, the vertical slice does not exist yet — there is no end-to-end behaviour to drive, and the expensive part of the review (the full sweep, the runtime, a capable model on the whole diff) would be paid N times for a partial answer.
This is where everything a task reviewer structurally could not see gets judged: the full sweep, the real runtime, duplication between tasks, Shotgun Surgery across the slice, coherence between the layers, the Spec axis against the ticket, and the Minor findings parked during the waves.
- Findings → dispatch one fixer per findings list (never one per finding), carrying the line: findings received → follow
/receiving-code-review. Re-review after — scoped, per /verified-review's re-review contract. The system-wide pass happens once; rounds converge on the findings.
- A finding that conflicts with what the unit mandates is the user's call: present it beside the unit text and ask which governs.
- Fixes are committed by you, as their own commit —
fix: slice fix 1 — <what changed>.
- Green on every axis, with the runtime observed working, closes the review. No report is ever a close condition; a green sweep is a working signal, not evidence the thing runs.
6. adr-candidate → write it
If the review surfaced an adr-candidate, write the ADR now, automatically, before integrating — its own commit on the branch, so it merges with the work and cannot be lost. Use /domain-modeling's ADR format; the candidate already carries the decision and its trade-off, so no question is asked — not losing it outweighs the gate. (/finish-branch's harvest stays the final catch for anything that slipped.)
7. Integrate
Hand off to /finish-branch. It carries its own STOP-gate and the merge / PR / keep / discard menu, and asks what to do with the worktree.
Before the menu, read the branch log. wave 1 → wave 1 fix → wave 2 → slice fix 1 is a diagnostic surface that costs nothing and disappears at the squash: many rounds on one wave says the cut was wrong, the brief was thin, or the reviewers are miscalibrated. Read it while it exists.
Red flags
The six that are not stated anywhere above — the rest of the contract lives in the steps:
- Say "done" on green tests alone. A fake/mock/fixture-backed green is a working signal, not proof the thing runs; the word waits until you have watched it work.
- Dispatch a seat by inline role prompt instead of by agent name — the seat's protocol lives in its definition, and an inline prompt silently forks it.
- Paste a seat's output into the next dispatch. Dispatches carry paths, never contents.
- Tell a reviewer what not to flag, or pre-rate a finding's severity. Let it surface and adjudicate it yourself.
- Edit code yourself, or let an agent touch git. You orchestrate and you alone commit; the seats build.
- Re-dispatch a task whose report is already on disk. After a compaction,
git log and .scratch/ are the truth and your recollection is not.
Related: /tier-1 drives this for one unit · /tier-2 feeds it tickets one at a time · /brief is the brief-writer's protocol · /verified-review reviews the finished slice and drives the real runtime · /finish-branch integrates · /using-git-worktrees isolates the worktree
