feat(CUST-WP-0015): implement agent inbox for inter-agent coordination

Adds a message-passing layer to state-hub so Claude instances can
coordinate across sessions without polling shared progress events.

- Migration f3a4b5c6d7e8: agent_messages table with thread support
- FastAPI router: POST/GET /messages/, thread view, mark-read, archive, reply
- 4 MCP tools: send_message, get_messages, mark_message_read, reply_to_message
- Observable dashboard: /inbox page with unread/read/archived sections + KPI
- CLAUDE.md updates: global, custodian, marki-docx, activity-core, template
- TOOLS.md: Agent Inbox tools section documented

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

mcp_server/server.py

@@ -1434,6 +1434,74 @@ def get_repo_dispatch(repo_slug: str) -> str:
     return json.dumps(_get(f"/repos/{repo_slug}/dispatch"), indent=2)
 
 
+# ---------------------------------------------------------------------------
+# Agent Inbox (inter-agent message passing)
+# ---------------------------------------------------------------------------
+
+@mcp.tool()
+def send_message(from_agent: str, to_agent: str, subject: str, body: str, thread_id: str | None = None) -> str:
+    """Send a message from one agent to another (or 'broadcast' for all).
+
+    Use this to coordinate with other Claude instances — e.g. a worker agent
+    reporting status back to the orchestrator, or the hub agent dispatching
+    instructions to a domain agent.
+
+    Args:
+        from_agent: Sender identifier (e.g. 'hub', 'marki-docx', 'railiance')
+        to_agent: Recipient identifier or 'broadcast' for all agents
+        subject: Short subject line (max 500 chars)
+        body: Full message body (markdown supported)
+        thread_id: UUID of the root message to create a thread (optional)
+    """
+    payload: dict = {"from_agent": from_agent, "to_agent": to_agent, "subject": subject, "body": body}
+    if thread_id:
+        payload["thread_id"] = thread_id
+    msg = _post("/messages/", payload)
+    return json.dumps(msg, indent=2)
+
+
+@mcp.tool()
+def get_messages(to_agent: str | None = None, from_agent: str | None = None, unread_only: bool = False, limit: int = 20) -> str:
+    """List messages in the agent inbox.
+
+    Call this at session start to check for pending coordination messages.
+
+    Args:
+        to_agent: Filter by recipient (your agent name, or omit for all)
+        from_agent: Filter by sender (optional)
+        unread_only: Return only unread messages (default: False)
+        limit: Maximum number of messages to return (default: 20)
+    """
+    params: dict = {"limit": limit, "unread_only": unread_only}
+    if to_agent:
+        params["to_agent"] = to_agent
+    if from_agent:
+        params["from_agent"] = from_agent
+    return json.dumps(_get("/messages/", params), indent=2)
+
+
+@mcp.tool()
+def mark_message_read(message_id: str) -> str:
+    """Mark an inbox message as read.
+
+    Args:
+        message_id: UUID of the message to mark as read
+    """
+    return json.dumps(_patch(f"/messages/{message_id}/read", {}), indent=2)
+
+
+@mcp.tool()
+def reply_to_message(message_id: str, from_agent: str, body: str) -> str:
+    """Reply to a message. Marks the original as read and creates a reply in the same thread.
+
+    Args:
+        message_id: UUID of the message to reply to
+        from_agent: Your agent identifier
+        body: Reply body (markdown supported)
+    """
+    return json.dumps(_post(f"/messages/{message_id}/reply", {"from_agent": from_agent, "body": body}), indent=2)
+
+
 # ---------------------------------------------------------------------------
 # Entry point
 # ---------------------------------------------------------------------------