agent-comms（Agent 通讯）

Agent Comms

Overview

Use this skill to initialize and operate a single shared communication channel for multiple agents and humans in a repository.

If the repo does not already contain the comms files, bootstrap them first with scripts/setup.sh. After that, use the normal workflow: read the rules first, then check the active log, then reply or append at the bottom.

Quick Start

Install:

npx skills add Abbiirr/agent-comms

For Claude Code with non-interactive install:

npx skills add Abbiirr/agent-comms --agent claude-code --yes --copy

After install, invoke the skill normally.

• Codex: Use $agent-comms to bootstrap this repository.
• Claude Code: /agent-comms bootstrap this repository

On first invocation, if the repo is missing the comms files, this skill should bootstrap them automatically before doing anything else.

Local development install flow:

npx skills add /path/to/agent-comms --agent codex claude-code --yes --copy

Fallback manual bootstrap flow:

bash .agents/skills/agent-comms/scripts/setup.sh

What bootstrap does:

• bootstraps AGENT_COMMUNICATION_RULES.md
• bootstraps AGENTS_CONVERSATION.MD
• ensures docs/communication/old/ exists
• injects a standard comms section into AGENTS.md by default

After setup:

1. Read AGENT_COMMUNICATION_RULES.md
2. Read AGENTS_CONVERSATION.MD
3. Report:

- active entries - messages requiring a response - threads ready for archival

1. Before changing code or docs, append a pre-task intent entry
2. Reply or append new entries at the bottom

Bootstrap Workflow

Use scripts/setup.sh as the public entrypoint. It wraps init_agent_comms.py and gives a cleaner CLI.

Use bootstrap when the repo does not already have:

• AGENT_COMMUNICATION_RULES.md
• AGENTS_CONVERSATION.MD
• docs/communication/old/

Default commands:

# Codex
bash .agents/skills/agent-comms/scripts/setup.sh

# Claude Code
bash .claude/skills/agent-comms/scripts/setup.sh

Useful flags:

• --root <path> to target a different repo root
• --no-inject-agents-md to avoid modifying AGENTS.md
• --force overwrites the template files
• --inject-agents-md explicitly enables AGENTS.md injection
• --python <bin> chooses the Python interpreter
• --rules-file, --conversation-file, and --archive-dir let you override the defaults

Bootstrap should be conservative:

• do not overwrite existing files unless --force is set
• do not read archives by default
• do not auto-resolve threads

First Invocation Behavior

When this skill is invoked in a repo that does not yet contain the comms files, bootstrap the repo first, then continue with the requested comms work.

Missing bootstrap files means one or more of these is absent:

• AGENT_COMMUNICATION_RULES.md
• AGENTS_CONVERSATION.MD
• docs/communication/old/

Use the first valid setup script path that exists:

1. .agents/skills/agent-comms/scripts/setup.sh
2. .claude/skills/agent-comms/scripts/setup.sh
3. ${CLAUDE_SKILL_DIR}/scripts/setup.sh if that environment variable exists

Run the bootstrap script from repo root:

bash .agents/skills/agent-comms/scripts/setup.sh

or, for Claude Code if needed:

bash .claude/skills/agent-comms/scripts/setup.sh

After bootstrap:

1. tell the user what files were created
2. read AGENT_COMMUNICATION_RULES.md
3. read AGENTS_CONVERSATION.MD
4. continue with the requested comms task

If the files already exist, do not re-run bootstrap unless the user explicitly asks to reinitialize or overwrite with --force.

Operating Workflow

1. Identify the agent

Use your actual environment identity, such as Codex, Claude, OpenCode, or User.

2. Read the rules first

Always read AGENT_COMMUNICATION_RULES.md before AGENTS_CONVERSATION.MD.

3. Check for pending messages

Look for:

• Directed to: <your name>
• Replying to: <your name>
• open questions or task handoffs without a response
• resolved threads that the original author can archive

Report:

• number of active entries
• messages requiring a response
• threads ready for archival

4. Take action

Before any code or doc changes:

• append a pre-task intent entry

When replying:

• use the correct message type
• append to the bottom of AGENTS_CONVERSATION.MD
• preserve the identity header format from the rules file

When archiving:

• only the original author archives, unless the user explicitly overrides
• move the full resolved thread to docs/communication/old/
• never delete archived files

Guardrails

• Never read docs/communication/old/ unless the user explicitly asks, or an active thread points to a specific archive file.
• Never self-resolve another agent's entry.
• Never let a bootstrap command silently overwrite an existing comms setup without --force.
• Keep AGENTS_CONVERSATION.MD lean and append-only during normal operation.

Resources

• scripts/setup.sh — production-facing bootstrap entrypoint
• scripts/init_agent_comms.py — bootstrap a repo-local comms setup
• assets/AGENT_COMMUNICATION_RULES.md — default rules template
• assets/AGENTS_CONVERSATION.MD — default active-log template
• assets/AGENTS_MD_SNIPPET.md — optional snippet to add to a repo's AGENTS.md