Skip to content

Heyvhuang/copybook-skills / standalone/coordinator/skills/coordinator

coordinator

Flagged

Coordinate an ongoing software delivery run when the user assigns this session the coordinator role or asks it to manage worker handoffs, gates, release approvals and progress. Use when starting, supervising or resuming that run, including a run with one active lane. Lanes run as Herdr workers (parallel-goal-fleet) on Claude or Codex, from whichever harness coordinates; T3 Code child tasks are the fallback. Do not start a coordinator run for discussion, a read-only review of the workflow, a standalone quick second opinion or an ordinary single-session edit.

FileKindSize
SKILL.mdskill md11.8 KB
references/example-run.mdmarkdown3.4 KB
references/herdr.mdmarkdown7.7 KB
references/lessons.mdmarkdown6.9 KB
references/models.mdmarkdown4.9 KB
references/pipeline.mdmarkdown6.1 KB
references/situations.mdmarkdown3.4 KB
references/t3.mdmarkdown4.9 KB
scripts/gemini.mjsexecutablescript2.3 KB
scripts/heavy.shexecutablescript829 B
scripts/status.shexecutablescript2.0 KB
scripts/watch-ready.shexecutablescript2.0 KB
templates/lane-brief.mdmarkdown1.5 KB
templates/release-packet.mdmarkdown1.4 KB
templates/run-record.mdmarkdown1.2 KB
templates/ship.mdmarkdown1.1 KB

references/lessons.md

View on GitHub
# Lessons from the example run (3 days, about 160 PRs)

Each: what happened → the rule. Times are from the run record, gate reports and playbook.

## Authorization

- The ship lane took a gate PASS on the payment PRs as permission to ship and had the batch ready to push (01:36). Stopped before the push. → PASS is evidence, not authorization; never-batch list in the ship brief; exclusive batches only on a written `SHIP-EXCLUSIVE-AUTHORIZED` line carrying the user's words.
- A UI PR contained its back-end PR (twice). Leaving the back-end PR number out of a batch does not keep its commits out. → Before freezing a batch, check no candidate contains a held head (`git merge-base --is-ancestor`).
- The first draft of this skill turned one run's grants (ship per batch, fast rollback, imagery without a cap) into general rules. → Grants live in the run's approval record with source and scope; examples grant nothing.
- Production steps were sometimes proposed as if checks were approval. → Passing checks is not approval; every production change names the exact commit, SQL or setting and waits for the user's yes for that change.
- Creating paid or account-level resources (a database project, an error-tracker project) can be refused to an agent. → Put it on the user's list with the page to open; never work around it.

## Gate and tests

- One gate passed the payment PRs with a scope that missed the admission route suites; the ship lane's wider checks found 27 failures. → Boundary changes add every caller's suite (`rg` the changed function).
- One train ran 639 test files for an 11-file change (12 min). → Scope = changed tests + tests beside changed source; full suite once a day.
- One train stopped at the first conflict and tested nothing. → Conflict rule: abort that merge only, test the rest.
- Two trains failed on tests main had already broken; another failed a required layout-shift ceiling that main also failed. → Reproduce on the exact base, classify contract drift vs product defect; a test-only fix needs a decision behind it; baseline exceptions are explicit and scoped and never waive a required acceptance target.
- Name filters skipped every control assertion and still exited 0 with passing counts; a changed-file design check read uncommitted `git status` and was empty on a committed merge. → Verify selection against the real diff; whole-file control runs; per-assertion executed/pass/fail/skip evidence.
- One train failed on 8 stale test names; the payment batch failed on a real behaviour change. → Decide stale vs real before editing a test.
- Load made tests time out at random (a 5 s timeout failed once, passed once on plain main). → Rerun alone twice under the run's documented criteria; intermittent timeouts may be load-sensitive, a single isolated assertion failure is real.
- Stale generated types (`.next/types`) produced type errors that were not real. → Clear generated output before trusting a red check.

## Shipping and builds

- Main took about 136 pushes in 32 hours, each a production build. → One push per batch; docs ride the next batch.
- A batch kept growing while its checks ran and went 40 minutes without a push. → Freeze at check start.
- The ship monitor stopped on a failing file and sat idle three times. → Standing lanes write a STOP line; every tick reads their panes; attribute, drop the causal PR, ship the rest.
- Two merges in a row started two production builds. → Cancel the earlier one at once.
- A docs commit pushed while the previous production build was queued built in full. → Never push docs to main alone.
- 15 failed hosted preview builds in two days, each billed in full. → Local build before any preview push; at most two local builds at once.
- Hosted CI ran a 70-minute matrix per push and used the month's allowance in one morning. → No paid CI on that run; the local gate is the merge check.
- The framework's build rejected a `node_modules` symlinked to the main checkout, and removing the link needed a permission prompt. → Real install per worktree with the project's package manager (for pnpm: `pnpm install --frozen-lockfile --prefer-offline`).
- Local `main` was a day behind `origin/main` during review. → Writer lanes branch from a fetched, recorded SHA.
- The SSH agent stopped signing at 03:20. → Per-command HTTPS with the gh credential helper; change no git config.
- `git add docs/brand` committed 119 local working files (~33 MB). → Add explicit paths; ignore rules for working folders.

## Watching lanes (Herdr runtime; T3 child tasks wake the coordinator themselves)

- Herdr's `done` flickers; a hand-over is a report. → Completion = the READY line plus your own check of the diff and checks.
- A READY word inside a sentence woke the watcher; two READY lines were missed while no watcher ran; a stray second watcher ran beside the first; reviews ending `VERDICT:` and gate reports ending `GATE-COMPLETE` never woke it at all. → One completion contract that the watcher reads exactly; line-start match; seen state kept; one watcher per run.
- The watcher lapsed after a long review and lines piled up. → Re-arm after every fire; a lapse means a catch-up tick at once.
- A brief sat unsent in a lane's input; two coordinator messages stayed queued after their lanes' turns ended; a lane sat on an auto-mode dialog. → Read the pane after every send and on every tick.
- `agent_name_taken` on restarting a reviewer. → Reuse it or choose a new name.

## Machine and quota

- Load above 300, later 500–1000 on 10 cores, swap nearly full, with six heavy workers and full test runs; 32 Claude and 19 Codex sessions were alive. → `heavy.sh` slots (3, then 2; configurable), one full suite per machine, close finished lanes.
- "Wait while load > 150" left a lane waiting 2 h 45 m. → Slots, not thresholds.
- An Opus `max` round took over an hour. → `max` only on request, with the duration stated.
- Overnight the second reader's key was missing. → Decide alone, mark it, put it on the morning list; at the start of the next overnight run check that the second reader starts.

## Working with the user

- A day was lost to questions that waited for the user. → Decide routine choices, log them, keep going.
- One-PR-at-a-time merging felt slow. → Waves.
- References the user gave were not used in a footer. → Every reference into the brief; hand-overs map each one.
- Placeholder prices (3, 12, 20 credits) reached copy drafts. → Prices and limits come from the source module, never from memory.
- Clean-up removed routes a lane had just hardened (18 fixed, 9 then scheduled for removal). → Decide removals before hardening.
- On the money lane two reviewers on different models each found a high-severity defect the other missed. → Two reviewers for money and auth; adopt it per run.
- A long-lived design session was auto-archived after a PR merged. → Never archive owner sessions; unarchive and say so.
- Error-tracker issues stayed open after their fixes shipped. → Resolve on SHIPPED, same turn.
reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 472kUpdated Claude CodeCodexVerifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 322kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 322kUpdated Claude CodeCodexVerifiedInstall
executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is availableStars: 322kUpdated Claude CodeCodexVerifiedInstall