agent-messenger

Instruction file imported from wolfego/agent-messenger (.cursor/rules/agent-messenger.mdc). Copyright stays with the author.

Agent Messenger

You have access to the agent-messenger MCP server for communicating with other AI agents (e.g. Claude Code running in a terminal).

Auto-check on conversation start

At the start of each new conversation (your very first response), call check_inbox once — UNLESS the user's first message is a shortcut command (#id, #sm, #dm, #ch, #wi, #la, #ct, #lt, #st, #rt). In that case, execute the command directly without checking inbox.

• If the inbox is empty, say nothing about it — proceed with the user's request normally.
• If there are unread messages, summarize them briefly to the user (e.g. "You have a message from claude-code about X") and act on any that have an action field (review, brainstorm, implement, reply). Messages are automatically marked read when fetched.
• Do NOT check inbox on subsequent turns unless the user asks.

Shortcuts

The user may type these short commands instead of full sentences:

• #help — Show this list of available commands.
• #cm — Check messages. Call check_inbox, act on any unread messages. If empty, say "No new messages."
• #sm — Send message. Ask the user who to send to (default: "claude-code") and what to say. Use send_message.
• #dm — Delete message. Ask the user for the message ID to retract. Call delete_message. Only messages you sent can be deleted.
• #ch — Set channel. Ask the user which channel to join. Call set_channel.
• #id — Set identity. Ask the user for a name (e.g. "cursor-design"). Call set_identity.
• #wi — Who am I. Call whoami and report identity, base ID, and channel.
• #ct — Create task. Ask for title and optional details, then call create_task.
• #lt — List tasks. Call list_tasks with sensible defaults (open tasks, sorted by priority).
• #st — Show task. Ask for task ID, then call show_task.
• #rt — Ready tasks. Call list_tasks with ready_only: true to show tasks with no blockers.
• #la — List agents. Call list_agents to see who is currently online in this project.
• #log — Message history. Call query_beads with type: "message" and limit: 20. If the user specifies an agent name, pass it as from. Show a concise summary: timestamp, from, to, subject.
• #status — Status overview. Run agent-messenger status in the terminal and show the output to the user.
• #orchestrate <feature> — Start the orchestrator workflow. Call scaffold_workflow with name: "orchestrate" (creates the workflow doc if it doesn't exist). Read the returned path end-to-end. Follow its phases. Record phase transitions with workflow_checkpoint. Ask user approval before each send to CC. Actions to send: challenge, verify-spec, implement, review.
• #debug <description> — Start the debug workflow. Call scaffold_workflow with name: "debug" (creates the workflow doc if it doesn't exist). Read the returned path end-to-end — especially the Diagnostic Resources section. Follow its phases. Actions to send: investigate, reproduce, fix, verify-fix.
• #ws — Show active workflow progress. Call workflow_status and present the current phase for each active feature.

Sending messages

When the user asks you to send something to another agent, or when your workflow would benefit from another agent's input, use send_message. Always include:

• A clear subject summarizing the request
• The action field so the recipient knows what to do
• context_files if the recipient needs to read specific files

Replying

When acting on a received message, always use reply (not send_message) to maintain the conversation thread.

Channels (multi-agent isolation)

If the user has multiple agent windows/terminals open in the same project, use set_channel to pair with a specific counterpart. Both agents must join the same channel name. When a channel is active, only messages on that channel appear in your inbox.

If the user says something like "set channel to design-review" or "join channel impl", call set_channel with that name.

Identity

Each agent gets a unique session ID on startup (e.g. cursor-a3f2). The base ID (cursor) is shared across all instances of the same agent type — messages addressed to the base ID are received by all instances. Use set_identity to give yourself a memorable name (e.g. cursor-design).

Use whoami to see your current identity, base ID, and channel. You do not need to self-identify in message bodies — the from label is set automatically.