/fleet-watch

Best used: configure always-on plan pullers for a project’s .plans/ (from the project, or from Anchor with a project name/path). See Skills overview.

Turn on always-on plan watchers for a project so eligible work under .plans/ is picked up and driven through the same kind of claim → execute → complete loop as /work—without you sitting in an interactive session.

You run the skill once (or occasionally) in a coding agent to configure those watchers. The watchers themselves then keep applying that work-style workflow on a schedule. Architecture of pull-per-endpoint fleets: Fleet workers.

Why use it

Without /fleet-watch	With /fleet-watch
You must open a session and run /work for each plan	Watchers poll, claim fit-appropriate plans, and run the work loop unattended
Idle machines stay idle until someone remembers	Mid/small/reasoner workers keep watching after reboot
Handoff depends on chat history	File-based ready → in-progress → completed continues on its own

Watchers pull one plan per tick (they do not drain the whole backlog by default and do not promote drafts). They still honor Preferred models / fit and do not replace a preferred orchestrator for architecture—lesser workers escalate hard planning.

When to run it

• You have (or want) a project with a .plans/ tree and ready work under bugs/ / features/
• You want pollers that survive reboot (user-level timers + linger)
• You are in the project folder, or in the Anchor repo and want to target another project by name

Skip it if you only work plans interactively with /work and do not want background claimers.

Usage

Invocation	Behavior
/fleet-watch	Use this project (CWD has .plans/, or its git root does). Status + recommended durable setup
/fleet-watch foo-project	Target ./foo-project, ../foo-project, or another resolved path named foo-project
/fleet-watch /abs/path/to/app	Explicit project root
/fleet-watch --status	Inspect only (lanes, timers, linger)—no install
/fleet-watch --install	After status, install recommended user timers (agent asks for consent if not already clear)
/fleet-watch --tiers mid,small	Prefer those capability tiers when proposing/installing workers

Everything after /fleet-watch is $ARGUMENTS. The skill resolves the project first, then runs the right helper under the hood—you should not need to memorize script paths.

Project resolution (what the skill does)

1. If the first token looks like a path or project name → resolve it (absolute path, ./name, sibling of CWD, or under a known workspace).
2. Else if CWD (or git root) contains .plans/ → that is the project.
3. Else if CWD is the Anchor repo and a single obvious app path was not given → ask once, or accept foo-project as a sibling/child name.
4. Prefer absolute paths in any commands the agent runs.

What “good” looks like after success

1. Project has .plans/ with known lanes (including in-progress/).
2. At least one systemd user timer is enabled for that project (name like anchor-watch-<project>-<agent>.timer), or you accepted a printed install plan to enable yourself.
3. loginctl linger is on for your user so timers fire after reboot without an open login.
4. Each watcher has a unique agent id so claims under in-progress/ do not collide.
5. You know how to check: “list my anchor-watch timers” / status via /fleet-watch --status.

Skill session vs watcher workers

	/fleet-watch skill (this chat)	Watchers it installs
Job	Resolve project, status, emit/install timers	On each tick: pick/claim a plan and run the work-style pipeline
Like /work?	No—you are wiring the fleet	Yes in spirit—claim → in-progress → execute (and complete when Done when holds)
When	You type /fleet-watch	Continuously (timers + linger)

So: the skill does not step through a plan’s table in-session; the workers it sets up are how the work workflow runs in the background.

What the skill will not do

• Promote drafts → ready (use /draft --promote; not this skill)
• Implement a plan’s steps in this chat (that is interactive /work; background execution is what the watchers do after install)
• Install system-wide units under /etc/systemd unless you explicitly demand root installs
• Enable timers without consent when you only asked for status or a dry recommendation
• Replace /work as the interactive “work this plan with me now” command

Install (platform wiring)

Same contract everywhere; only how the agent loads the skill differs:

Platform	Install
Claude Code	Scaffold installs .claude/commands/fleet-watch.md
Grok Build	Scaffold installs .grok/skills/fleet-watch/SKILL.md

Related

• /work — interactive: pick and execute one plan with you in the session
• Fleet workers — multi-tier pull model, leases, isolation (what the watchers implement)
• Utility scripts — fleet_watch.py / work_once.py if you automate outside an agent
• CLI — scaffold, preferred orchestrator