AGENT SKILL

setup

Check the multi-agent role map (model and effort per role and the layer behind each value) or write confirmed keys into the user, team or local multi-agent.yaml layer.

License 许可
MIT
最近更新
2026-11-15
Stars
★ 30
Forks
7

01 前言 Front Matter

shell
bash
description
Check the multi-agent role map (model and effort per role and the layer behind each value) or write confirmed keys into the user, team or local multi-agent.yaml layer.
allowed-tools
[ "Bash(${CLAUDE_SKILL_DIR}/scripts/setup.sh:*)" ]
argument-hint
[check|apply] [--layer user|team|local] [<key>=<value> ...] [session=<alias>]
user-invocable
true
disable-model-invocation
true

02 资源文件 2

03 SKILL.md 正文

Purpose

/multi-agent:route resolves the role map through four layers: the bundled defaults, ~/.claude/multi-agent.yaml, the team layer and .claude/multi-agent.local.yaml. This skill shows what those layers resolve to and writes one of them. All three consumer layers absent is a valid state: the bundled defaults apply. Keys, values, layer paths and the fan-out guard are defined in ${CLAUDE_PLUGIN_ROOT}/reference/config.md.

No argument means check. It installs nothing; the resolver needs only Bash, awk and git.

check (read-only)

${CLAUDE_SKILL_DIR}/scripts/setup.sh check session=<alias>

Pass session=<alias> with the family in your own model ID (opus, sonnet, haiku, fable); omit it when you cannot tell, and the map is resolved as under a frontier session. The JSON is the one /multi-agent:route all prints. Show:

  1. one line per role: single <model> @ <effort>; fanout <model> @ <effort>, with inherit written as inherit: omit the model option, and the layer in source for each value;
  2. the layers rows (read, absent, skipped, or not applicable at a home or non-repository root) and every entry of notes as written; each note (a skipped layer, a rejected value, an ignored key) is a finding to report;
  3. the personal-override row: PASS when .claude/multi-agent.local.yaml is gitignored, else WARN with the recommended line .claude/**/*.local.*. Recommend the line; do not edit .gitignore;
  4. each WARN team: row: a retired team file (the config block in docs/conventions/multi-agent.md, or .claude/multi-agent.yaml) that the resolver no longer reads. Relay the row's next step as written.

It modifies nothing.

apply

  1. Run check and show the map. This is the before state.

  2. Settle the layer with the user, recommendation first:

    • local (.claude/multi-agent.local.yaml): this checkout only, for this person. The default for a personal preference.
    • team: docs/conventions/multi-agent.yaml (under the convention home the repository binds). Shared with everyone who clones the repository.
    • user (~/.claude/multi-agent.yaml): every repository on this machine.

    team and local are refused at a home or non-repository root.

  3. Settle the keys. Ask only for what the user wants changed; each value comes from the table in config.md. Pass every key to set as <key>=<value>, and <key>= to remove one from that layer. Existing keys in the layer are kept.

  4. Preview, writing nothing:

    ${CLAUDE_SKILL_DIR}/scripts/setup.sh apply --layer <layer> <key>=<value> ...
    

    It prints the target path and a diff. It exits 1 without writing when the resolver would reject or ignore a key; relay the reason and return to step 3.

  5. Show the diff and ask for an explicit yes. Write only on that yes:

    ${CLAUDE_SKILL_DIR}/scripts/setup.sh apply --layer <layer> --write <key>=<value> ...
    

    It prints the path written and the keys it read back from the file.

  6. Run check again and show the after map beside the before one, naming each value that changed and the layer that now supplies it. The file is what the next /multi-agent:route run reads; nothing about the running session changes. For team, offer git add and a commit and run each only when the user accepts.

The fan-out guard

With fanout.frontier_guard: true, the bundled value, a stage that runs more than one agent never runs them on a frontier model: under a Fable session, or a session the caller did not name, each fan-out variant that would inherit or name a frontier model takes fanout.model (opus) instead. A single synthesis or judge agent keeps the role's own model. The basis is recorded beside the key in ${CLAUDE_PLUGIN_ROOT}/reference/defaults.yaml.

fanout.frontier_guard=false is an explicit opt-in to frontier fan-outs: every parallel worker then runs on the session model, so a frontier session runs every agent of the stage on the frontier model (rates: Pricing: model pricing; usage: Manage costs: Track your costs). Before writing it, say this and ask the user to confirm that specific key, separately from the rest of the change. Recommend the narrowest layer that covers what they mean: local for one checkout, user for one machine, team only when the whole team has agreed.

What this skill does NOT do

  • Edit .gitignore, CLAUDE.md or AGENTS.md.
  • Write a model ID. Only aliases are accepted, so the map stays portable across providers.
  • Change a named agent's model or effort. An agent such as implementation:implementer or implementation:scoped-implementer owns its tier through its own frontmatter and whatever its dispatcher passes on the call; the role map governs generic agent() and Agent dispatches only.
  • Recheck a default against upstream. That is /multi-agent:audit-defaults.

Next

/multi-agent:route all Prints the map the next workflow reads from args.roles.

Gotchas

  • A layer file is rewritten whole, so its comments are not kept; the preview says so when the file has any.
  • An existing layer that does not parse stops apply with the line and the reason. Fix or remove the file first; setup never overwrites content it could not read.
  • The first team write creates the YAML. A retired team file, which the resolver no longer reads, seeds its keys and stays as it was; check names it until it is deleted.
  • Team and local layers resolve against the repository root of the working directory. Inside a second worktree, run from that worktree.

04 同仓库更多 Skills

melodic-software
melodic-software/claude-code-plugins

track

Backlog reads through the bound tracker: search open and closed items, view one item with its comments and state history. Use when another skill searches or reads tracker items.

★ 30
melodic-software
melodic-software/claude-code-plugins

reviewer-media

Reviewer media for one pull request: decides whether it needs a narrated walkthrough, builds the storyboard from /review:explain-change's digest, has /explainer-video:produce render it and hands the manifest to /source-control:pull-request media. Use when asked for 'reviewer media', 'a walkthrough video for this PR', 'a produced video of this change', or when pull-request ready or monitor hands a PR over.

★ 30
melodic-software
melodic-software/claude-code-plugins

route

Model and effort per agent role (orchestrator, worker, verifier, retrieval) for a multi-agent run, resolved through the bundled defaults and the user, team and personal override config layers into fenced JSON with single-agent and fan-out variants and the layer behind each value. Use when asked 'which model for the subagents', 'route the roles' or 'model routing', or before a skill passes a role's model on a spawn.

★ 30
melodic-software
melodic-software/claude-code-plugins

route

Model and effort per agent role (orchestrator, worker, verifier, retrieval) for a multi-agent run, resolved through the bundled defaults and the user, team and personal override config layers into fenced JSON with single-agent and fan-out variants and the layer behind each value. Use when asked 'which model for the subagents', 'route the roles' or 'model routing', or before a skill passes a role's model on a spawn.

★ 30
melodic-software
melodic-software/claude-code-plugins

principles

Module-design vocabulary for architecture findings, and the checks built on it. Use when: 'what is a deep module', 'what is a seam', 'deep vs shallow module', 'what is the deletion test', 'leverage and locality', 'module design vocabulary', 'which word do we use for X in architecture findings'. Scanning a codebase for friction is /architecture:improve.

★ 30
melodic-software
melodic-software/claude-code-plugins

write

Structured five-field bug report (title, steps to reproduce, expected vs actual, justified severity, suggested fix location) from an informal description, read-only. Use when the user names a defect they observed, or asks for one written up ('report a bug', 'file a bug', 'write this up as a bug').

★ 30