Developer

Developer (orchestrator)

Drives the triage → build → review → fix → merge pipeline across isolated subagent workers, looping over every sub-issue of a spec unattended. Each worker gets a clean context — the only thing it knows is the arguments you pass in its prompt. You (the orchestrator) hold the state between steps.

Invoke

/developer <issue>              # spec with sub-issues → deliver them all
                                # plain issue → deliver just that one
/developer <spec> <subissue>    # deliver a single specific sub-issue

Flags (override the repo defaults — see Run configuration):
  --parallel | --sequential     # spec mode: waves vs one-at-a-time
  --auto-merge | --no-auto-merge  # merge CLEAN PRs vs leave them ready
  --build-oversized             # build `oversized` tickets instead of escalating

If no issue number is given, ask for it and stop. Do not guess issue numbers. The execution flags only change spec mode; in single mode they are a no-op. Accept the bare words parallel / sequential as synonyms for the flags.

Namespacing. Installed as a Claude Code plugin, skills and agents carry the plugin prefix: the skills appear as developer-skills:<name> and the subagents as developer-skills:dispatcher / developer-skills:code-author / developer-skills:diff-reviewer. Use the names exactly as they appear in your available-skills and available-agents lists; the short names below refer to whichever form is installed.

Contract docs (tracker + code host)

The pipeline is agnostic about where issues and changes live. Two committed docs define the mechanics for this repo, and every worker reads them in its own context:

• docs/agents/issue-tracker.md — issue operations (read an issue, enumerate children of a parent, check a blocker, comment, label, close) in its ## Delivery operations section.
• docs/agents/code-host.md — change operations (publish, check out in a worktree, review, mark ready, reply, merge, auto-close semantics).

Read those two once at the start (they are short — an allowed exception to "never read bodies yourself").

Their annexes are deferred, and every worker inherits that. Each core doc links phase annexes naming the phase that opens them: code-host-ci.md is opened at the checks gate, when a change's CI has to be waited on, read or classified — nowhere earlier; issue-authoring.md is for whatever creates issues (/to-tickets) and this pipeline never opens it at all. Read an annex at the step that names it and not before. A run that opens them at the start pays the whole contract, three workers deep, per sub-issue, to use a fraction of it. Every command block below shows the GitHub factory default (gh); when a contract doc defines a different mechanic for the same operation, the doc wins. If a doc is missing, the GitHub defaults apply as-is — suggest /setup-developer-skills if that looks wrong.

Note one capability flag from docs/agents/code-host.md: issue auto-close on merge? If not (e.g. issues on a tracker the code host can't close), the orchestrator closes the delivered issue itself per the tracker ops right after verifying the merge.

If either doc says the tracker or the code host is local (files in the repo, no remote), read LOCAL-HOST.md now, before anything else. It holds every standing adjustment a local host or tracker needs — capability overrides, tracker writes, branch discipline, cleanup and wrap-up. A run on a remote host never loads it.

Two more files are read on demand, never at the start: MERGE-FIX.md at the first merge conflict, and WRAP-UP.md when the loop ends.

(All three live next to this SKILL.md, in the skill's own directory — under the plugin root when installed as a plugin, the same place scripts/cleanup-worktrees.sh comes from. They are part of this skill: a step that says to read one is not optional, it is that step's other half.)

Run configuration

Three knobs govern a run. Resolve each one before mode detection, in this precedence order: CLI flag > repo default > factory default.

Repo defaults live in docs/agents/developer-defaults.md, written by /setup-developer-skills. Read it once at the start (it is short — this is an allowed exception to "never read bodies yourself"); if it is missing or a knob is absent, fall back to the factory default. State the resolved configuration in one line before starting, e.g. Run config: execution=parallel, merge=manual (repo defaults).

What merge means:

• auto — a CLEAN verdict triggers the code host's merge operation (Merge step; gh pr merge on GitHub). The committed merge: auto line in docs/agents/developer-defaults.md is the user's standing authorization for these merges.
• manual — the pipeline stops at CLEAN: you already marked the PR ready after the review, so record the sub-issue as ready-to-merge and leave the merge to the human. Because sub-issues only close on merge (Closes #N), anything Blocked by a ready-to-merge sub-issue stays blocked for the rest of the run — expected, not an error; it lands in the wrap-up as the human's queue.

What oversized means — what to do with a sub-issue triage scores oversized (--build-oversized sets it to build for the run):

• escalate — the default and the safe reading: the ticket is handed to a human to re-cut, and nothing is built (Triage step).
• build — build it anyway, at opus, exactly as if triage had said complex. This is the user's standing answer to "the ticket is too big": they have decided the split is not worth the round trip. Note the risk once when you resolve the config, then stop arguing it — the dispatcher's fault lines still go into the builder's prompt as its order of work, and if the builder does come back with half a feature, that PR escalates through the ordinary non-convergent path rather than a second opinion about size.

Workers (subagents)

Spawn each via the Agent tool with the matching subagent_type. Pass isolation: "worktree" to every code-author and diff-reviewer spawn. Pass model explicitly to code-author spawns (triage decides the tier) and to re-review diff-reviewer spawns (model: "sonnet") — the first review is discovery across the whole change and stays on the agent's pinned opus, while a re-review is verification of a diff the skill has already scoped down to the last fix pass. Never run the skills yourself in the main context — the point is isolation.

Every spawn is run_in_background: true, dispatchers included, in both execution modes. A foreground spawn holds your turn open for the worker's whole run, and anything that interrupts that turn — a Ctrl-C, a dropped connection — takes the worker down with it: its context, its worktree and its commits are gone for good. The same interruption leaves a background worker running and still reachable. Sequential mode is not an exception to this: it means wait for this worker's result before spawning the next, not spawn it in the foreground.

Three things at every spawn, each cheap now and each the difference between a resume and a rebuild later:

• Give the Agent tool a description that names the job and the sub-issue — Build #<N>, Review PR #<PR>, Fix #<N> cycle 2. It is how the worker is identified in the agent list and in the user's view of the run.

• Keep the agentId the tool returns for as long as that worker runs: it is the handle that picks the worker back up instead of starting it over (see Resuming the orchestrator). Drop it when its RESULT arrives.

• On a BUILD, FIX or HARVEST spawn — the jobs that hold uncommitted work and are expensive to lose — append a spawn row to this run's log as soon as the tool hands you the agentId, so the job leaves a trace that outlives your context:

  mkdir -p .scratch
  echo "$(date +%F) spec=#<spec> sub=#<N> event=spawned job=<build|fix|harvest> agent=<agentId> model=<tier>" \
    >> .scratch/developer-run-<spec>.log
  

  (sub=none on the harvest, which belongs to the whole run.) Triage and review spawns skip the row entirely: both are cheap to repeat, and pipeline step 0 reconstructs where a PR stands without them. Step 7 writes the other kind of row, the terminal one — a resume reads both, the wrap-up reads only the terminal ones.

Context economy

The loop may cover many sub-issues; your context must survive all of them.

• Never read issue or PR bodies yourself. Workers read them in their own disposable contexts. You only run the cheap listing commands below.
• Every command you run prints a bounded projection, never a blob. Ask for the fields you need (--json/--jq, --format) and cap what is left (head -20, | wc -l). Two failure modes cost the most: a streaming or watching command (gh pr checks --watch, gh run watch) repainting its progress into your transcript, and a malformed command dumping its tool's entire --help — a single mis-escaped gh pr list --search did exactly that for 4k tokens in one field run. So send both streams of any probe you are not certain of through a cap: <cmd> 2>&1 | head -20. What you need from these commands is one number, one state or one exit code.
• Every spawn costs about the same whatever it carries — roughly 750 tokens of prompt, launch metadata and result notification, against a payload that is often one word. So batch the work that can be batched (triage scores a whole wave in one dispatcher) and never spawn a worker to recover something a default already covers. Builds, reviews and fixes cannot be batched — each needs its own worktree and its own clean context — and are worth their envelope; a second dispatcher for one missing score is not.
• Never read CI logs yourself either — gh run view --log-failed and anything like it. The Merge step's --json classification is the whole diagnosis the orchestrator gets; the rest is the fixer's, in its context.
• A worker's whole final message is its RESULT line — the agents require it and every spawn prompt below restates it. A worker that reports prose before its line is spending your context, not its own; nothing it says there survives the run, so anything worth keeping belongs on the PR or the issue.
• Track per sub-issue: number, task id, chosen model, PR number, verdict, fix cycles, wave (parallel mode), the agentId of the worker running on it right now (dropped the moment its RESULT arrives), outcome (merged / ready-to-merge / escalated / blocked) — and write the row to the run log the moment the sub-issue goes terminal (delivery pipeline step 7), so the wrap-up reads facts instead of recalling them. Also keep the dispatcher's touches/hints just long enough to forward hints into that sub-issue's Build step — discard both once the build is spawned, they have no use after that.

Resuming the orchestrator

A run outlives your turn: workers keep going in the background, their notifications can arrive late, and sessions get interrupted, compacted or restarted. So while a run is in flight, every prompt that reaches you is a resume — including a bare Continue from where you left off., an empty continuation, or a notification you believe you have already handled. There is no state in which the right answer is "no response requested": either work is pending and you take its next step, or nothing is, and you go to Wrap-up. Silence is the one failure mode this pipeline cannot recover from on its own.

(This is the orchestrator's own resume. Pipeline step 0 is the per-sub-issue one — it is what step B below runs.)

On any such prompt:

1. Rebuild the picture from the three places that survive a dead context, never from recall:

   • the progress board (TaskList) — which sub-issues are in_progress;
   • .scratch/developer-run-<spec>.log — the terminal rows already recorded, and the event=spawned rows naming the worker that was running on each sub-issue still in flight. If a wrap-up already ran this spec, the rows from before it are in .scratch/archive/developer-run-<spec>-*.log;
   • ListAgents — which of those workers are still alive.

2. Recover each non-terminal sub-issue in this order, stopping at the first that works:

   A. Its worker is still alive → SendMessage to its agentId and ask it to report. Its context, its worktree and its commits are all intact, so this continues the job rather than repeating it — by far the cheapest recovery, and the only one that does not throw away work already paid for. A worker that finished while you were not looking answers here too, with the RESULT whose notification you missed.

   This works for code-author. It usually does not for diff-reviewer: a review changes no files, so its worktree is removed when it ends and a reviewer missing from ListAgents is missing for good. Try it if it is listed; otherwise go straight to B — a review is cheap to repeat, a build is not.

   B. Its worker is gone → run pipeline step 0 on that sub-issue exactly as written. It asks the code host rather than your memory, and routes the sub-issue to Review, to the Fix cycle, or back to Triage when nothing was ever opened for it.

3. Re-enter the loop: recompute the unblocked set (spec loop step 1, or the wave in parallel mode) and carry on. Nothing left → Wrap-up.

Say in one line what you recovered and how, before continuing. During an unattended run the board and that line are the user's whole window into it.

Step 0 — Publish context docs before anything else

Workers branch from origin/main, so any domain-context file that is not committed and pushed is invisible to them. Grilling/spec sessions edit these files but do not commit them. Before dispatching any worker:

git status --porcelain -- CONTEXT-MAP.md '**/CONTEXT.md' docs/adr docs/agents AGENTS.md CLAUDE.md

If anything shows up, stage only those paths (never the user's unrelated work-in-progress), commit on the current branch (must be main — if not, stop and tell the user), and push:

git add CONTEXT-MAP.md '**/CONTEXT.md' docs/adr docs/agents AGENTS.md CLAUDE.md
git commit -m "docs(domain): publish context map and ADR updates"
git push origin main

If the push is rejected, stop and report — do not rebase or force anything. This is the flow's start, before going unattended; the user is still there to resolve it.

Mode detection

Enumerate the children of the given issue per the tracker's Delivery operations. GitHub default — native sub-issues (infer OWNER/REPO from git remote -v):

gh api graphql -f query='
{
  repository(owner:"OWNER", name:"REPO") {
    issue(number: N) {
      subIssues(first: 50) {
        pageInfo { hasNextPage }
        nodes { number title state labels(first: 10) { nodes { name } } }
      }
    }
  }
}' --jq '.data.repository.issue.subIssues'

If hasNextPage is true, stop and report: a spec with more than 50 sub-issues is not sized for this pipeline — tell the user to split it and end the run. Never proceed on the first page alone: delivering 50 of 60 while reporting the spec complete is a silent failure, the one outcome worse than stopping.

Keep each sub-issue's labels from this query — the pick reads them (spec loop step 1). Where a tracker's enumeration carries no labels, get them per its read-labels operation instead.

(Throughout this skill, #<N> stands for the issue ref in the tracker's own format — a number on GitHub/GitLab, a file path on a local tracker — and #<PR> for the change ref in the code host's format.)

• Open sub-issues exist → spec mode: loop over all of them (below).
• No sub-issues → single mode: run the delivery pipeline once on the given issue, with the issue itself as spec (no separate parent spec).
• Two arguments given: run the delivery pipeline once on <subissue> with <spec> as the spec. Skip the loop. If the sub-issue ends merged (verified CLOSED), read WRAP-UP.md and run its Close the spec step (step 4) afterwards — it may have been the spec's last open sub-issue. That one step is all this mode needs from the wrap-up.

Progress board (spec mode — not optional)

The user follows the run through the harness task list. Keep it faithful at every transition; a stale board defeats its purpose.

1. Immediately after mode detection, create one task per open sub-issue with TaskCreate, in sub-issue order: subject #<N> <short>, activeForm Delivering #<N>. The whole plan must be on the board before the first worker spawns.

   <short> is the sub-issue title trimmed to about six words / 50 characters, cut at a word boundary and with no ellipsis — enough for the user to tell the rows apart, and it never grows. Keep the same <short> for that sub-issue's every later rename. The number is the identifier; the words are only a label, and the full title is one gh issue view away for anyone who needs it. This is not cosmetic: the harness re-injects the whole board into your context on a timer, so every character of every subject is re-read many times over a long run — a board of 25 full titles costs more over a run than the entire spawn traffic it is tracking.

2. When the delivery pipeline starts on a sub-issue → TaskUpdate status: in_progress. In parallel mode every wave member goes in_progress as its build spawns, so the board shows exactly what is running concurrently.

3. Terminal transitions, the moment they happen:

   • merged (sub-issue verified CLOSED) → status: completed.
   • ready-to-merge (merge: manual, verdict CLEAN) → back to status: pending and rename the subject to #<N> <short> — ready to merge: PR #<PR>. Not completed — the human still has to merge it.
   • escalated → back to status: pending and rename the subject to #<N> <short> — escalated: <one-line reason>. Never mark an escalated sub-issue completed — unchecked items at the end are the human's queue.

4. Sub-issues that never became deliverable (blocked by an escalated one, or by a ready-to-merge one the human hasn't merged yet) stay pending; rename them #<N> <short> — blocked by #<M> at wrap-up.

Single mode (no sub-issues) skips the board.

The board is the report — do not narrate the run beside it. Between the run-config line and the wrap-up, a spec run's default output is nothing: the task list already says which sub-issue is building, which is in review, which is merged and which is waiting, and it says it live, without costing a turn. Prose that restates it — "wave 1 launched", "triage complete", a table of the tier each sub-issue drew, "5 of 25 merged" — is a second, staler copy of the board, and the user has to read past it to reach the part that is not on the board. Keep the board current instead; that is the progress report.

Six things still get said, each in one or two lines, never a table:

• the resolved run config, once, before starting (Run configuration);
• what you recovered, once, after a resume (Resuming the orchestrator);
• a switch into the conflict queue (Parallel mode);
• how to merge, the first time a sub-issue lands ready-to-merge under merge: manual — once for the run, not once per PR; the wrap-up repeats it for the rest;
• anything that stops the run or needs the human: an escalation and why, a denied permission, a spec too large to size;
• a direct question from the user, answered directly — the silence rule governs unprompted narration, never a reply.

Everything else the run learns goes where it survives: the board, the PR, the issue, the run log, and the wrap-up summary at the end.

Spec loop

Repeat while open sub-issues remain:

1. Pick the next unblocked sub-issue: for each open sub-issue (lowest number first), check its blockers without reading full bodies — the "check a blocker's state" operation from the tracker doc. Blockers may be wired as the tracker's native dependency links, as a Blocked by body section, or both (/to-tickets prefers native edges where the tracker has them) — check both. GitHub default:

   # native dependencies: count of OPEN blockers (0 or absent = clear)
   gh api repos/{owner}/{repo}/issues/<N> --jq '.issue_dependencies_summary.blocked_by // 0'
   # body fallback: every blocker listed in the section must be CLOSED
   gh issue view <N> --json body --jq '.body' \
     | awk '/^##[#]* *[Bb]locked by/{f=1;next} /^#/{f=0} f'
   gh issue view <BLOCKER> --json state --jq '.state'
   

   Extract the Blocked by section, never a fixed window around the heading: a grep -A<n> reads the wrong number of lines by construction — it drops the fifth blocker of a list of six and swallows the first lines of whatever section follows a list of two.

   Take the first open sub-issue whose blockers are all closed, and:

   • Skip any sub-issue carrying the ready-for-human triage label (the repo's own string for that role if docs/agents/triage-labels.md maps it differently) — from the labels the enumeration returned, plus the ones you applied yourself while escalating this run. That label is the escalation gate: someone already gave up on this sub-issue, and picking it up again buys three more fix cycles against the same wall. The gate is symmetric and it is the whole mechanism: removing the label re-queues the sub-issue, there is no other state to reset.
   • Whatever a gated sub-issue blocks stays blocked, as with any open one.
   • With merge: manual, sub-issues you already delivered as ready-to-merge count as done for your loop but their dependents stay blocked — skip both.

2. Run the delivery pipeline on it.

3. On merged or ready-to-merge → next iteration. On escalated/blocked → record it, next iteration.

4. When no deliverable sub-issue remains (all closed or ready-to-merge, or the rest are blocked by escalated/unmerged ones) → wrap-up.

Parallel mode (execution: parallel)

Parallel is the factory default (see Run configuration). The trade-off: sequential with merge: auto delivers one sub-issue fully before the next starts, so each PR branches from a main that already contains the previous one — no merge conflicts by construction. Parallel trades that guarantee for throughput: independent sub-issues are built concurrently, and conflicts between their PRs become expected work, resolved by extra merge-fix jobs. Note that with merge: manual sibling PRs all branch from the same main regardless of execution mode — sequential buys no conflict guarantee there, so parallel costs nothing extra.

Work in waves:

1. Wave = every open sub-issue whose blockers are all closed (same check as step 1 of the spec loop), minus the ones that step's gate excludes — ready-for-human above all, whether this run applied it or an earlier one did.

2. Run the delivery pipeline on each wave member concurrently, entry points first: the pipeline's step 0 resolves where each member starts, and only the ones with no open change get triaged and built. Spawn one dispatcher for all of them at once (Triage step — one spawn per wave, not per member, capped at 5 issues a spawn), then their code-author BUILD jobs in parallel (each in its own worktree, run_in_background: true) — minus any member the Triage step escalates as oversized, which leaves the wave without a build (under oversized: build, or against a no-split directive, it is built like any other member). A resumed member goes straight into the review or fix stage alongside them. As each build reports its PR, spawn its diff-reviewer; as each reviewer reports, mark that PR ready (step 3 of the pipeline); fix cycles run per PR exactly as in the sequential pipeline. Cap concurrent build/review/fix workers at 3; queue the rest of the wave.

3. Merges stay strictly serial — never merge two PRs concurrently. With merge: auto, merge each PR as it reaches CLEAN. After every successful merge, refresh the wave's still-open PRs per the code host's update-branch operation — GitHub default, per open sibling:

   gh pr update-branch <PR>
   

   It is a remote operation — no local git. Each sibling branched from a main that did not contain this merge; left stale, its CI goes red for synchronization, not for a bug, and a full fix cycle ends up doing what this one call does. A sibling whose update fails on a conflict is left alone — but note it: that failure is what puts the PR in the conflict queue (step 4), which is the only place a conflicting PR is worked on. Learning it here rather than at the merge gate is most of the point — spec #994 skipped one refresh and met the conflict forty minutes later, with the queue idle in between.

   Every still-open member, including the ones mid-fix-cycle. A PR that is being fixed is exactly the one that will still be open in an hour and exactly the one that goes stale; skipping it because "it is not ready yet" defers the conflict to the moment you most want a clean merge.

   One PR per Bash call. Issue the refreshes as separate commands, not as a for loop over the wave: a loop that writes to the code host reads as a bulk operation to the permission classifier and gets denied wholesale (observed in spec #994 — the same two calls, run singly, went through untouched).

   Every PR in the wave branched from the same main, so any PR merged after the first may conflict: on merge failure after the Merge step's checks gate passed, run the merge-fix job (MERGE-FIX.md) and retry once. With merge: manual there is nothing to serialize — each CLEAN PR just becomes ready-to-merge.

   The first conflict of the wave closes the parallel phase. From that moment the wave finishes through a conflict queue (below), not concurrently. Conflicts are not independent work: every conflicting PR resolves against the same main, and the first one merged invalidates every resolution computed beside it. Two merge-fix workers running at once are one worker and one rewrite waiting to happen.

4. Conflict queue. Once step 3 has seen one conflict, the wave's remaining unmerged PRs form a queue, ordered oldest-PR-first, and it drains one at a time:

   • At most one merge-fix worker is alive in the whole run. Never spawn a second while one is running, whatever the worker cap allows.
   • A PR's merge-fix is spawned only when that PR is at the head of the queue — i.e. after the previous PR is merged into main. Until its turn a queued PR gets no merge-fix worker; its conflict is not stale work, it is work not yet started. (Step 3's gh pr update-branch refresh still runs on it after every merge — that call is remote, cheap, and often is the resolution.)
   • When the head PR merges, drop it from the queue and try the next one's merge before assuming it still conflicts: the winner's merge, plus the refresh, resolves most of the rest for free. Only a merge that actually fails earns a merge-fix worker.
   • Everything that is not the merge path — builds, reviews, and review fix cycles, including a queued PR's own — carries on in parallel underneath. The queue serializes conflict resolution, not the wave.

   Say the switch out loud once, in one line, e.g. #936 conflicts — wave finishing through the conflict queue: #936 → #937 → #938.

5. When every wave member is delivered (merged, ready-to-merge, or escalated), recompute the unblocked set → next wave. None left → wrap-up.

Everything else — context economy, escalation, wrap-up, rules — is unchanged.

Delivery pipeline (per sub-issue)

0. Entry point — resume, never rebuild

(Pipeline step 0, not the top-level Step 0 that publishes the context docs. This is the per-sub-issue resume, reached both on a fresh run and as step B of Resuming the orchestrator.)

A run can die at any point — a dead session, a compaction, a Ctrl-C — and the sub-issues it half-delivered are still open, so re-running /developer <spec> picks them right back up. What the tracker forgets is how far each one got: build from scratch again and you get a second PR for the same sub-issue and a second review paying for it. So before triaging, ask the code host whether a change already exists for this sub-issue, per its "open change for this issue" operation. GitHub default:

gh pr list --state open --search '"Closes #<subissue>" in:body' --json number,isDraft

• No open PR → nothing to resume: step 1 (Triage).
• One open PR, no unresolved review threads → keep its <PR>, skip Triage and Build, start at step 3 (Review). The reviewer settles its own scope from the PR's review history, so this covers both shapes: a build that was never reviewed (full scope) and one whose review was answered in full but never reached a verdict (incremental). If it comes back blocked reason=no new commits since the last review, the previous review was the last word and everything it raised is resolved: treat it as CLEAN and go to Merge — here only, because no fix pass ran this cycle to leave findings standing.
• One open PR with unresolved review threads → a review landed and its fixes did not: start at step 4 (Fix cycle), counting from cycle 1 with the fixer at opus (the build tier died with the session that chose it).
• More than one open PR matches → escalate: two open changes for one sub-issue is a human's call, never a pick.

GitHub default for the unresolved-thread count:

gh api graphql -f query='
{
  repository(owner:"OWNER", name:"REPO") {
    pullRequest(number: <PR>) {
      reviewThreads(first: 50) { nodes { isResolved } }
    }
  }
}' --jq '[.data.repository.pullRequest.reviewThreads.nodes[]
          | select(.isResolved == false)] | length'

The fix-cycle budget starts fresh on a resume: a PR that already burned cycles in the dead run gets three more here. That is deliberate — the alternative is reconstructing a counter nothing ever recorded — and the ready-for-human gate is what stops a sub-issue looping forever across runs.

1. Triage

Triage a whole wave in one spawn, not one spawn per sub-issue. In parallel mode this step runs once for the wave, covering every member the pipeline's step 0 found without an open change; in sequential mode the wave is one sub-issue and the same prompt carries a list of one. Batch at most 5 issues per dispatcher; a larger wave takes a second batched spawn, never a spawn per issue. The reason is the spawn envelope, not the dispatcher: a triage round trip costs the orchestrator ~750 tokens of prompt, launch metadata and notification whatever it carries, and it carries one word. Five scores in one spawn pay that once instead of five times, and triage is the step where it is free to do so — the scores are independent, the codebase glance is shared, and nothing downstream needs them at different times.

Spawn dispatcher with run_in_background: true:

Triage issues #<N1>, #<N2>, … Score each one's implementation complexity per your rubric, independently of the others. Your entire final message must be one RESULT issue=… complexity=… model=… touches=… hints=… why=… line per issue, in the order given — nothing before the first, nothing between them, nothing after the last. Do not explain your scoring anywhere except the why= field, which is capped at 15 words: your final message lands in my context whole and stays there for the rest of the run, so a paragraph in front of the lines is charged to every turn that follows and read by no one.

Repeat that last sentence in the spawn prompt every time. It is the rule this worker breaks most often — twice in field runs, once after the agent definition had already been hardened — and the spawn prompt is the last thing it reads before working. Parse why= and drop it: it exists to give the dispatcher's justification a 20-token home instead of a 330-token one, not because anything downstream needs it.

Match each line back to its sub-issue by issue=. A member with no line at all is treated exactly like a malformed one (below). Then, per sub-issue, parse complexity= first:

• complexity=oversized → the sub-issue does not fit in a single fresh context window, and what happens next is the oversized knob's call (Run configuration):

  • oversized: escalate (default) → Do not build it. No model tier rescues a ticket that does not fit: the builder runs out of room, the review finds half a feature, and three fix cycles burn against the same wall. Go straight to Escalation, quoting the dispatcher's hints= — they carry the fault lines — and move to the next sub-issue. No BUILD, no review, no fix cycles are spent on it.
  • oversized: build → continue to Build at opus, passing the dispatcher's hints= through as usual; the fault lines become the builder's order of work. Do not escalate, do not label, and do not re-argue the size — the knob is the answer to that argument.

  Before escalating, check the issue body once for an explicit no-split directive ("deliberately indivisible", "no dividir", "ship as one unit"). The dispatcher is told to veto its own oversized score when it finds one, so an oversized line on such a ticket means triage missed it: build it at opus as if the knob said build. This is the one place you overrule a triage verdict, and only ever in that direction — the author's directive outranks the rubric, never the reverse.

• Anything else → parse model=<tier> and continue to Build.

On any malformed or missing result, default to opus and build — a line you cannot parse, or that never arrived, is not an oversized verdict. Parse touches= and hints= too, defaulting each to none if the line predates these fields or omits them — never block the pipeline on a missing hint. Never re-spawn a dispatcher to recover one missing line: the default costs less than the round trip it would take to improve on it.

2. Build

Spawn code-author with model: <tier>, isolation: "worktree" and run_in_background: true, then log the spawn row (Workers):

BUILD job. Spec issue #<spec>, sub-issue #<subissue>. Run the implement-issue skill on the sub-issue. The sub-issue's ## Spec extract section carries the spec decisions that apply to it — read the full spec issue only if that section is missing. Triage found: <dispatcher's hints, verbatim, or omit this line entirely when hints=none>. Your entire final message must be the RESULT pr=… url=… line — no summary before it, nothing after it. Whatever deserves a record goes in the PR body, not in your reply. Report only a PR number you have confirmed exists.

• RESULT blocked … → escalate (see below) and move to the next sub-issue.

• RESULT pr=<PR> url=<URL> → confirm the PR exists, then keep <PR> and continue:

  gh pr view <PR> --json number,state,headRefName
  

  Never skip it. A build has reported pr=<N> for a number the host 404s on, with its whole implementation sitting uncommitted in its worktree because the publish step never ran at all. A worker's RESULT is a claim; this is the one cheap command that turns it into a fact, and the only thing standing between a fabricated line and a reviewer sent after a PR that was never opened. (Change metadata is read per docs/agents/code-host.md on another host.)

  On a 404, do not re-spawn the build — the work is almost certainly intact in the worker's worktree. Recover the worker per Resuming the orchestrator step A: its spawn row holds the agentId, ListAgents says whether it is still alive, and SendMessage tells it what you found and to run its publish step for real, reporting only a number it has verified. If the worker is gone, escalate naming its branch and worktree, so nobody rebuilds on top of work that still exists.

3. Review

Spawn diff-reviewer with isolation: "worktree" and run_in_background: true:

Review PR #<PR> by running the review-pr skill on it — its step 1 plus the repo's docs/agents/code-host.md give the exact checkout procedure for your worktree; follow them, not memory. Post the review (inline comments + summary) as a single COMMENT submission — never an approval event — and do not mark the PR ready or merge; those are orchestrator steps. Your entire final message must be the RESULT verdict=… line — the review itself is your output, your reply is not.

Then mark the PR ready yourself, whatever the verdict — per the code-host doc's mark-ready operation. GitHub default:

gh pr ready <PR>

Skip it whenever the PR is already ready — a re-review, or a resumed run whose step 0 found isDraft: false.

• verdict=CLEAN → go to Merge.
• verdict=NEEDS_FIXES → enter the fix cycle.
• RESULT blocked reason=no new commits since the last review … → the fix pass pushed nothing, so there is nothing to re-review. Treat it exactly as that cycle's NEEDS_FIXES: the previous findings still stand. Do not re-spawn the reviewer — go straight to the next fix cycle (or escalate if the budget is spent).
• RESULT blocked because the change branch is held by another worktree (the worker quotes git's "already used by worktree" error) → a previous worker's worktree wasn't cleaned: run Cleanup (step 6) and re-spawn the reviewer, once per sub-issue — if it blocks again, escalate.
• Any other RESULT blocked or malformed result → escalate, next sub-issue.

4. Fix cycle (max 3)

For cycle c = 1, 2, 3:

1. Fixer model: cycle 1 uses the build tier, each later cycle escalates one tier (sonnet → opus; opus stays opus). A sub-issue resumed straight into this step has no build tier — step 0 already set it to opus.

2. Spawn code-author with that model, isolation: "worktree" and run_in_background: true, then log the spawn row (Workers):

   FIX job. PR #<PR>. Run the fix-pr skill to address all review threads — its step 1 plus the repo's docs/agents/code-host.md give the exact checkout procedure for your worktree; follow them, not memory. Pushing the fixes and replying to the review th

…(truncated)