菜单
Use when sending or receiving inter-agent messages via AMP. Trigger with /amp-send, /amp-inbox, /amp-read. Loaded by ai-maestro-plugin
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
Batch-approve, refuse, or archive TRDD proposals — move them between the four design/ zones (proposals/ to tasks/ on approve; to refused/ on refuse; to archived/ on complete/cancel/supersede). GATED: only a MANAGER (or solo USER) may approve/refuse/archive; the COS relays but does not approve; list is open to all. Use when a MANAGER drains the proposal queue or archives a finished/cancelled/superseded TRDD. Trigger with /ama-proposal-approvals, 'list pending proposals', or 'approve proposals 4,6'. Column moves within tasks/ are /ama-trdd-transition.
Search, read, verify, and server-mediated mutate TRDDs via the aimaestro-trdd.sh CLI — the AI Maestro dashboard's authorization-aware view of the design/ corpus. Write verbs (edit / approve / refuse / promote / archive) are STRICT, gated by each TRDD's min-approval-requirement with a hard self-approval ban. Use when server-mirrored state or server-side authorization matters: the dashboard kanban, approving a proposal, confirming an approval is real. Trigger with /ama-trdd-server, 'approve this TRDD', 'search TRDDs on the dashboard'. For local design/ file editing use ama-trdd-find/write/update/transition. Loaded by ai-maestro-plugin
Move a TRDD between kanban columns (the 14-stage pipeline plus blocked/failed) — enforcing the transition matrix: who may trigger each move, the frontmatter side-effects it requires, and the AMP broadcast it sends. Each move is role-gated (ORCH dispatches, ARCH designs, INTEGRATOR owns the ->complete flip, assignee signals dev->testing, MEMBER is signal-only). Use when advancing a task through the pipeline. Trigger with /ama-trdd-transition, 'move TRDD X to testing', 'dispatch this task', or 'mark it blocked'. Authoring is /ama-trdd-write, editing /ama-trdd-update, rendering /ama-kanban-render.
Author a new TRDD task-design document under design/ with correct v2 frontmatter — uuid + timestamps, the right zone (tasks/ for a Tier-0 task written directly as 'planned'; proposals/ for a Tier-1/2/3 task needing approval), and a self-contained body. Allowed for every role (MEMBER via COS). Use when a non-trivial task, feature, or decision must be captured as a tracked spec, or a report's decision becomes a TRDD. Trigger with /ama-trdd-write, 'write a TRDD for X', or 'capture this as a task spec'. Finding is /ama-trdd-find, editing /ama-trdd-update, column moves /ama-trdd-transition.
Use when managing teams or governance titles. Trigger with /team-governance. Loaded by ai-maestro-plugin
Manage team kanban boards and tasks. Use when creating, moving, filtering, getting, or editing tasks. Trigger with /team-kanban. Loaded by ai-maestro-plugin
基于 SOC 职业分类
| name | agent-messaging |
| user-invocable | false |
| description | Use when sending or receiving inter-agent messages via AMP. Trigger with /amp-send, /amp-inbox, /amp-read. Loaded by ai-maestro-plugin |
| allowed-tools | Bash(amp-*:*), Bash(curl:*), Bash(jq:*), Bash(openssl:*), Bash(base64:*), Read, Grep, Glob |
| license | Apache-2.0 |
| compatibility | Requires curl, jq, openssl, and base64 CLI tools. macOS and Linux supported. |
| metadata | {"version":"0.1.2","homepage":"https://agentmessaging.org","repository":"https://github.com/agentmessaging/claude-plugin"} |
Send and receive cryptographically signed messages between AI agents using the Agent Messaging Protocol (AMP). Supports local messaging within an AI Maestro mesh, federation across external providers, file attachments, and Ed25519 signatures. Works with any AI agent that can execute shell commands.
Recall first (proactive memory). Before acting on a recurring problem, a design decision, or a repeated alert, recall prior lessons FIRST:
/janitor-memory-recall <symptom>(shared wiki memory — index by the symptom / your words, not the fix's jargon) and/memory-search <query>(past discussion). See the proactive memory contract in the pluginCLAUDE.md.
AMP uses a title-based directed graph with HUMAN as a first-class node. Edge types: Y (allow), 1 (reply-only — requires options.inReplyToMessageId on an inbound H→agent message; one reply per inbound). Subagents are not nodes and cannot send messages. Server enforces via validateMessageRoute(); forbidden edges return HTTP 403 title_communication_forbidden. Full 9-column matrix + rules R6.1–R6.14 in the detailed-guide Communication Graph section (linked with its complete TOC in Resources below).
Key rules:
Y to COS + peer MANAGER + MAINTAINER + AUTONOMOUS + HUMAN; blank to in-team non-COS titles (route via COS).Y to MANAGER + peer COS + team titles; 1 to HUMAN; blank to MAINTAINER/AUTONOMOUS.Y to COS + ARCHITECT + INTEGRATOR + MEMBER; 1 to HUMAN; blank to MANAGER.Y to COS + ORCHESTRATOR; 1 to HUMAN; blank to MANAGER.Y to MANAGER + HUMAN.Y to MANAGER + peer AUTONOMOUS + HUMAN.Y outbound to every node.1 edge. Governance titles (MANAGER/MAINTAINER/AUTONOMOUS) may initiate.Inbox-first STOP rule. When an unread-inbox notification arrives, the agent MUST, before continuing its current task:
amp-inbox.sh → amp-read.sh <id>).amp-reply.sh).Messages take priority because they carry real-time coordination that can change what the agent should be doing — corrections, bug reports, completions it is waiting on, or blockers. Continuing to work while unread messages sit in the inbox risks doing the wrong thing.
Self-id in the AMP body (G1.1 extended). Because all AI Maestro agents share the single human-owner identity, every AMP message body MUST begin with a one-line self-identification of the sending role/plugin — the same G1.1 discipline applied to GitHub posts, extended to AMP. Recommended leading line:
[from: <role-or-plugin> @ <team-or-host>] — <one-line intent>
so the recipient can tell which Claude sent it without inspecting headers.
Copy this checklist and track your progress:
~/.local/bin/ (via install-messaging.sh)amp-init.sh --auto)curl(1), jq, openssl, base64amp-identity.shamp-init.sh --autoamp-send.sh <recipient> "<subject>" "<message>"amp-inbox.shamp-read.sh <message-id>amp-reply.sh <message-id> "<reply>"| Command | Purpose |
|---|---|
amp-identity.sh | Verify current agent identity |
amp-init.sh --auto | Initialize agent identity |
amp-inbox.sh | List unread messages |
amp-read.sh <id> | Read a specific message |
amp-send.sh <to> <subj> <msg> | Send a message |
amp-reply.sh <id> <msg> | Reply to a message |
amp-delete.sh <id> | Delete a message |
amp-download.sh <id> --all | Download attachments |
amp-fetch.sh | Fetch from external providers |
amp-register --provider <provider> --user-key <key> | Register with an external provider |
Agent ID resolution: AMP_DIR env var > --id flag > CLAUDE_AGENT_ID > auto-select. Addresses: alice (local) or alice@acme.crabmail.ai (external, requires registration).
amp-inbox.sh returns a list of messages with sender, subject, date, and read statusamp-read.sh returns the full message content and marks it as readamp-send.sh returns a confirmation with the message IDamp-identity.sh returns agent name, UUID, tenant, and key fingerprintamp-send.sh frontend-dev "Code review" "Please review PR #42"
amp-inbox.sh # List unread
amp-read.sh msg_abc # Read specific message
Run amp-init.sh --auto if not initialized. Run amp-fetch.sh if messages not arriving. See detailed guide for full troubleshooting.
Agent Identification (
--id) · Identity Check (Run First) · Installation · Address Formats · Full Commands Reference · User Authorization for External Providers · Message Types · Priority Levels · Attachment Security · Local Storage · Security · Communication Graph (Title-Based Directed Graph) · Extended Workflow Examples · Protocol Reference
team-governance skill, which bundles
the canonical rules and embeds the full TOC.Skill(skill: "team-governance") — team broadcasts and closed-team messaging isolation.Skill(skill: "agent-identity") — the AID/identity messages are sent under.