Use when manage multiple local CLI agents via tmux sessions (start/stop/monitor/assign) with cron-friendly scheduling.
한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.
Interview a project folder into AOS and write the result: a stable slug and domain in .aos/project.json, an OpenWiki wiki that is actually watched and does not index junk, memB memory bound to the slug, a Synapse map, AGENTS.md with its harness symlinks, and optionally CI and issue triage.
한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.
설치 방법 보기설치하기 전에 에이전트에게 전달되는 지시 내용을 확인할 수 있습니다.
/aos-setup prepares the machine. This prepares one folder in it — and unlike
the machine skill, this one writes. It asks what it cannot know, shows what
it intends to write, and writes it once you agree.
Everything it decides lands in .aos/project.json, so a second run knows what
was already settled and changes only what you ask it to.
aos-config show # is this machine's vocabulary present?
If it reports unvollständig, stop and run /aos-setup first. The domain list
is per-machine and nothing ships with one — guessing here would invent a
vocabulary the user never chose.
Then measure this folder:
node ~/.claude/skills/aos-project-init/scripts/aos-project-doctor.mjs
Report the failing rows before touching anything. On an existing repo some failures are deliberate: a repo public on purpose, a scratch folder with no remote, a project nobody wants the wiki daemon to visit. Ask rather than "fixing" those.
One question at a time, each with a proposal the user can accept in a word.
Skip anything already settled in .aos/project.json unless asked to change it.
Name and slug. Propose the folder name for both. The slug becomes the memB
project_id, and it is written down precisely so it survives a rename: a
binding derived from the basename orphans every memory the day the folder
moves, and two folders sharing a basename share one memory scope.
Domain. Offer the list from aos-config, preselected by what the path
implies — the first segment under the workspace root. If the project sits
outside that root, ask instead of guessing.
Wiki. Three separate questions, because they are three separate things:
initialise a wiki here at all · watch it, so the daemon refreshes it every
2h · write .openwikiignore.
Watching is not free. Every watched project with changes costs an LLM call per cycle, and a free-tier key hits quota on a handful. Recommend watching repos under active work and leaving archives out.
memB. Bind the slug, then draft a project card instead of offering a file
dump. Read the project yourself — README.md, AGENTS.md/CLAUDE.md,
package.json/pyproject.toml, and the .aos/project.json this interview
settles — and distill a project card: 3–8 facts, each ≤ 300 chars, each
prefixed with its date (YYYY-MM-DD), covering what the project is, where it
lives, stack, status, and settled decisions. Show the numbered list and save
only on the user's confirmation:
add_memory({ text: "(2026-09-24) …", category: "project_card", project_id: "<stable-slug>", infer: false })
The slug settled above is the project_id, so the hook's candidate matching
finds the card. Facts stay small and dated because cards are shown to every
future prompt in this project.
Synapse. A spatial map earns its keep on a codebase too large to hold in your head. Ask; do not default it on.
AGENTS.md. If missing, offer the template. If a real CLAUDE.md already
holds content, merge it into AGENTS.md first and only then replace it with a
pointer. Never silently overwrite it.
Subagents & Multi-Harness Pipeline. Offer to initialize .aos/pipeline.json (or set per-role, per-harness model overrides via /subagent-setup; without an override agents use the session model, see docs/agent-models.md). Ensures subagents are automatically compiled across Claude Code, Antigravity, OpenCode, and Codex CLI.
With a GitHub remote, two more. Skip both entirely without a remote — a local-only repo has nowhere to put them, and a skipped question is clearer than one answered "no".
release-please.needs-triage, needs-info,
ready-for-agent, ready-for-human, wontfix) · local markdown in
.scratch/.release-please computes the version from Conventional Commits and nothing
else: an unprefixed subject is invisible to it, and one feat: forces a minor
bump however small the change. Offering it imposes that rule on the project, so
write the rule into AGENTS.md in the same breath — and ask what the version
policy is. Some repos here move in patch steps only, where feat: needs
explicit permission.
Show every intended write as one list, then write on agreement.
{
"name": "<Project Name>",
"slug": "<stable-slug>",
"domain": "<from aos-config>",
"openwiki": { "enabled": true, "watch": true },
"memb": { "projectId": "<stable-slug>" },
"synapse": false,
"ci": "gate",
"triage": "github"
}
Then, only for what the interview enabled:
cp ~/.claude/skills/aos-project-init/assets/openwikiignore.template .openwikiignore
openwiki --init # needs provider credentials in ~/.openwiki/.env (see /aos-setup)
When the user chose to watch it, add the absolute path to projects in
~/.openwiki/projects.json — that file is the daemon's watch list. Merge into
it; never rewrite it.
Domain Registration (WORKTREE.md): Every project belongs to a domain (e.g., ~/dev/bdb-dev/). To ensure multi-agent navigation works, the parent domain directory must have a WORKTREE.md acting as a map. Check if ../WORKTREE.md exists. If it does, append a link to this new project specifying its Slug, Domain, active Harnesses, and a pointer to its .aos/project.json. If it doesn't exist, create it with a simple markdown list structure.
For memB, write the project card through the MCP with the slug as project_id
and project_card as category — domain-category rows are dead weight, since
the hook whitelist only injects project_card/godmode:
add_memory({ text: "(2026-09-24) …", category: "project_card", project_id: "<stable-slug>", infer: false })
Raw file ingest is opt-in only — if the user explicitly asks for a dump, run
python3 ~/.gemini/config/mcps/memb-mcp/memb_ingest.py <path>; the default
answer is no.
Never write credentials or high-entropy strings into memory.
node ~/.claude/skills/aos-project-init/scripts/aos-project-doctor.mjs
Report the doctor's own count, and name anything left failing on purpose. Then
commit: chore: for wiring, docs: for AGENTS.md. Check the project's
version policy before reaching for feat:.
aos-config has none. Send the user to
/aos-setup instead.AGENTS.md still full of template placeholders.release-please without writing the Conventional Commits rule into
AGENTS.md — it is a trap otherwise.아직 리뷰가 없습니다. 사용해 본 소감을 남겨 주세요.
로그인 없이 쓸 수 있습니다. 사용 후기, 질문, 작동하지 않은 환경 등을 편하게 남겨 주세요.
아직 댓글이 없습니다. 첫 댓글을 남겨 보세요.
Use when manage multiple local CLI agents via tmux sessions (start/stop/monitor/assign) with cron-friendly scheduling.
한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.
Use when a hybrid memory system that provides persistent, searchable knowledge management for AI agents (Architecture, Patterns, Decisions).
한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.
Reference for how BDB structures autonomous software engineering work — the seven-node dispatcher graph (Architect, TechLead, UI/UX, Engineering, Media/EventTech, Reviewer, Shipping) that /startcycle-graph actually runs. Use when you need the high-level lifecycle framing without inventing your own process.
한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.
Tools are how AI agents interact with the world. A well-designed tool is the difference between an agent that works and one that hallucinates, fails silently, or costs 10x more tokens than necessary. This skill covers tool design from schema to error handling.
한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.
Harness patterns for coding agents — memory, permissions, context engineering, delegation, skills, hooks, bootstrap.
한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.
agenttrail: live map of a multi-agent build in the browser: which plan component is being worked on, by which agent or harness, what is done and what is stuck. Use when a multi-agent pipeline starts (/startcycle, /startcycle-graph, /teamwork-preview) or after a plan-canvas approve, when the user asks to see what the agents are doing ("live map", "build board", "who is running now"), or by yourself whenever a multi-agent run is under way — no user prompt needed.
한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.