22824bd35f
The SessionStart announce no longer forces a watch-loop bootstrap on every session — it now emits a short pointer instructing Claude to invoke the new mailbox-collaborate skill (or /collaborate slash command) when the user wants peers to wake them mid-task. Messages still surface on the next user prompt via the UserPromptSubmit hook even without the watcher, so nothing is lost; idle sessions just stop burning relaunch tokens. The watch-loop protocol (exit codes, rename handling, mail handling) moves from the hook prose into the new skill body, where it only loads when actually needed. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2.2 KiB
name, description
| name | description |
|---|---|
| mailbox-collaborate | Start the Claude-Mailbox push-delivery watch loop for this session so peers can wake Claude mid-task. Use when the user says "let's collaborate", "wait for mail", "watch the mailbox", "listen for peers", "enter collaboration mode", or invokes /collaborate. Without this skill the watcher is OFF — peers can still leave messages but Claude only sees them on the next user prompt. Do NOT launch the watcher without this opt-in. |
You are now in mailbox collaboration mode. Peers can wake you mid-task by sending mail to your mailbox.
Launch the watcher
Your mailbox name was announced at SessionStart (format: <project>-<short-session-id>). Use it in place of <NAME> below.
Launch the watcher as a background bash task immediately:
Bash(command="claude-mailbox watch --block --name <NAME>", run_in_background=true)
The watcher is short-lived by design — it exits after exactly one event (delivered message, 25-second timeout, rename, or daemon error). Claude Code's BashOutput notifications fire only on task completion, so every exit MUST be followed by a fresh background relaunch with the same command. Gaps in the loop = missed pushes.
Handling watcher completions
When a watcher task completes, inspect its stdout and exit code:
- stdout contains
[Claude-Mailbox] Mail from <peer>: <body>→ treat the body as an inbound interrupt with priority over your current plan. Address it, then relaunch the watcher. - stdout contains
[Claude-Mailbox] Mailbox renamed to '<new>'→ relaunch with--name <new>, and use<new>for all futuremcp__mailbox__*calls (update your identity). - exit code 3 with no stdout → silent timeout, just relaunch.
- exit code 2 → daemon unreachable; wait ~5 seconds, then relaunch.
- any other exit code → report it to the user, then relaunch.
Stopping
Keep the loop running until the user says "stop watching", "stop collaborating", "end collaboration", or similar. When they do:
- Stop relaunching after the next completion.
- If a watcher is currently mid-poll and the user wants it killed immediately, use
TaskStopon its task id.
Do not re-enter collaboration mode on your own after stopping — wait for the user to invoke this skill again.