Skip to content
Back to skills

Amby Run

ASecurity

Run a feature's tasks as a team — workers build, QA and architect review, state kept in files, resumable.

  • 2 stars
  • 0 votes
  • 0 copies
  • 0 views
  • Added November 18, 2026
ai-agentsgosecurity

Security analysis

A100/100

Scanned November 18, 2026

npx -y skills add ambystechcom/AmbyKit --skill amby-run --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Amby Run?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Amby Run
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/ambystechcom-amby-run/badge)](https://www.skillsdirectory.com/skills/ambystechcom-amby-run)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: amby-run
description: "Run a feature's tasks as a team — workers build, QA and architect review, state kept in files, resumable."
---
<!-- GENERATED by AmbyKit from src/prompts/run.md — do not edit by hand; run `ambykit sync`. -->

Run the feature in `the user's request that followed this skill invocation` (number or name; `--new` only when the person asked for a new run). You are the
**lead**: you start workers, hand out task IDs, run the reviews and keep the run state. You have no role of
your own. Run `ambykit activity <feature> --agent=codex --phase=run` first; ignore any error.

Subagents on this host: `no`. If `yes`, workers are the `amby-worker` subagent and reviewers the
`amby-qa` and `amby-architect` subagents (roles `@.amby/roles/qa.md`, `@.amby/roles/architect.md`). If `no`,
say so once, then do each worker, QA and architect step yourself, in sequence, one task at a time, with the
same commands and state files below: claim, edit, check off, report; QA before architect, each acting as its
role file, and say which role you are acting as.

Every step works from files with these verbs alone, and nothing here needs a server: `ambykit fleet`,
`ambykit activity`, `ambykit directive`.

## 1. Resume or start

Run `ambykit fleet status <feature>`; it prints the run, the next phase, lanes, free slots and warnings.

- **Done** → say so and change nothing. **Stopped** naming `spec.md`, `plan.md` or `tasks.md` (missing, or the
  spec still `draft`) → stop, name that artifact and the command that makes it.
- **No unfinished run** → `ambykit fleet start <feature>` (this invocation is the request; `--new` starts another
  run on request only). **One** → resume it. **Several** → list them and ask which; never pick.
- **Warning: state missing or corrupt** → `ambykit fleet status <feature> --run=<id> --rebuild`, then go on.
  Checked tasks in `tasks.md` stay checked, so nothing completed is repeated.
- Load by reference only: the current phase of `@specs/<feature>/tasks.md`, the plan sections and `US-#`/`FR-###`
  its tasks cite, and `@specs/<feature>/tests.md` when it exists. Never whole artifacts.

## 2. Work the phase (stage `working`)

- Take the phase's open, unblocked tasks from `status`. Tasks are independent when their file ownership is
  disjoint. If two tasks you want to run together own overlapping files, see **Overlap**.
- **Caps:** `status` shows free slots (default 3 workers, 2 per assistant type). Never start a worker without one.
- Per worker: `ambykit fleet update <feature> --worker-start=<lane> --assistant=<id> --tasks=T001,T002`
  (`<lane>` is `lane-<phase>-<k>`, `<id>` the assistant the worker runs on, the same id as your `--agent`; it is refused
  when no slot is free). Give the worker only: the feature, run,
  lane, its task IDs, their file ownership and the artifacts by `@path`. It returns one line per task.
- Workers report on a fixed rule, which you must also keep. Report at the start of a task, at the end of a task, and whenever you are blocked or waiting. Never report on a timer and never report progress.
- When a worker finishes: `ambykit fleet update <feature> --worker-done=<lane>`.
- **Blocked:** the worker marks the task `⚠ BLOCKED: <reason>` and you continue with other unblocked tasks.
  Blocked work is yours to resolve or give to the person; workers never ask each other.
- **Stalled:** `status` flags a lane with no report and no `tasks.md` change for the stall limit. Restart it with
  `ambykit fleet update <feature> --restart=<lane>` and re-delegate its unchecked IDs. After the restart limit the
  command pauses the lane and asks for a person: only they may approve moving it to another assistant or marking
  its task BLOCKED.
- **Notes and claims:** workers claim files before editing (`ambykit fleet claim`), post one-line notes
  (`ambykit fleet note`) and their claims are released when a task is checked off. A refused claim names the
  owner; sort it out through the board and `status` warnings, not by editing the file.

## 3. QA (stage `qa`)

When every task of the phase is checked: `ambykit fleet update <feature> --stage=qa`. Ask QA (`@.amby/roles/qa.md`, read-only) to:

1. run the phase's `**Verify:**` line, and check the acceptance criteria of its stories; if `tests.md` exists, its
   cases for these tasks are the expected tests. No Verify line → check the criteria only and say the phase has none.
2. return a verdict. You write it to `.amby/fleet/<run>/verdicts/<lane>.qa.<round>.json`
   (`<round>` = failed rounds so far + 1): `{"verdict":"pass"|"fail","findings":[{"id","severity","finding","fix"}]}`.
3. record the result on the Verify line, per spec 020: `✓ passed`, `✗ failed: <why>`, or, when the check cannot be
   run here, tell the person, ask them to confirm and record `? unverified`.

A failing Verify never passes the checkpoint. Do not tick the checkpoint or start the next phase until it passes.

## 4. Architect review (stage `architect`)

After QA passes: `ambykit fleet update <feature> --stage=architect`. Ask the architect (`@.amby/roles/architect.md`, read-only) for correctness,
simplification and fit with `plan.md`; findings by stable ID, same verdict file with `.architect.`.

- **Fail** (QA or architect): `ambykit fleet update <feature> --round=<failed rounds so far>`. It prints `retry` or
  `escalate`. **retry** → return only the finding IDs and their `fix` to the **same worker** (same lane) and go back
  to `working`. **escalate** (the round limit) → stop: summarize the open findings and ask the person to decide.
  Record their answer with `ambykit fleet update <feature> --decision="<text>"` and continue from it.
- **Pass:** `ambykit fleet update <feature> --phase=<next>`; after the last phase,
  `ambykit fleet update <feature> --done`, then suggest `/amby.converge`.

## Pause, resume, stop

When the person asks: `ambykit fleet update <feature> --pause` (or `--resume`, `--stop`; add `=<lane>` for one lane).
This writes a directive per lane; start no new worker while paused. The directive protocol is: The reply carries a directive: continue, pause or stop. On pause or stop, finish the task you are on, then obey it at your next report: do not start another task.
The run state records `paused`; the next `/amby.run` resumes from it.

## Overlap

If workers' files overlap, choose before any work starts and record it:
- **Serialize:** one worker after the other. `ambykit fleet update <feature> --overlap=serialized --lanes=a,b --order=a,b`
- **Squad:** one writer per shared file, a fixed merge order; others send change requests through the writer or you.
  `ambykit fleet update <feature> --overlap=squad --lanes=a,b --shared=<path>,... --writer=<path>:<lane>,... --order=a,b`

## Rules

- All review roles, including `security`, stay read-only; never give them write access. Only workers write code.
- Keep every status line to one line; no logs, no restated artifacts.

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

View all in ai-agents →

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.