ci-watcherlisted

CI watcher: always-on background process that monitors CI and notifies on both failure and pass through the `Monitor` tool's stdout event stream. Claude must never stop the watcher on its own initiative. The watcher itself may still exit on a terminal condition (PR closed without merge, no CI on the default branch, main-CI timeout, main CI resolved) — that is fine and is not a Claude-initiated kill. One session can run SEVERAL watchers at once — one per branch. Every `/tmp` file is keyed on a per-branch `SLOT`, so launching a watcher for `feat/b` never disturbs the one already watching `feat/a`. **Tool availability:** `Monitor` and `TaskStop` are deferred tools in this harness. If they are not already available, load them first with `ToolSearch` query `select:Monitor,TaskStop`. # **CRITICAL RULE: DO NOT KILL THE CI WATCHER AUTOMATICALLY** **THE CI WATCHER MUST NEVER BE KILLED BY CLAUDE ON ITS OWN.** This is an absolute rule with no exceptions: - **DO NOT** kill the watcher when CI passes. - **DO NOT** kill the watcher when CI fails. - **DO NOT** kill the watcher when a PR is merged. - **DO NOT** kill the watcher on merge to `main`. - **DO NOT** kill the watcher when the feature looks "done". - **DO NOT** kill the watcher on any other "natural completion" event. - **DO NOT** call `/ci-watcher stop`, `TaskStop` on the watcher's task id, `kill <pid>`, `pkill ci_watch`, or any equivalent on your own initiative. **THE ONLY WAY TO STOP THE WATCHER IS AN EXPLICIT USER REQUEST*