Phil's Claude fleet
The harvesters, watchers and loops that keep running in the background. Each one stays quiet until something actually changes, then turns its tab red and waits for me.
Fleet commander: the handyman (click him)
How every watcher works
One pattern, reused everywhere
Every tab is a long-lived Claude Code session in a terminal multiplexer built for agents. The polling never happens in the tab's own turn, because every turn that ends flips the tab to "done". Anything a shell can check (gh, curl, files, scheduled jobs) runs as a background shell loop, which survives the network blips that used to kill subagent pollers. Only what an MCP alone can see (chat, ticketing, email) needs a background subagent looping on read → sleep 240, returning only when it finds something real.
AskUserQuestion is the only thing that turns a tab red, so red always means a real decision.last_seen_ts / idempotency file, so a restart never double posts.The watcher registry
One place that knows what should be running
This started as "remember which tabs to restart". Now every long-lived watcher has an entry in one watchers.json, a CLI (watcher status, stop, report) and its own page. 22 registered, 19 polling right now.
📋Registered, not remembered
Each entry says what it watches, which skill holds its rules, what it does on a change, where it reports and whether it is allowed to post. State is derived live, so the page and the CLI agree with what is actually running.
♻️Restarted unless someone meant it
A watcher that stops for any reason other than me or another agent asking it to is an interruption, and it gets restarted. A deliberate stop is recorded in a ledger with who stopped it and why. An unrecorded stop just comes back.
⏰Three things restart it
The watchdog every 5 minutes, /resume-agents and /startup, and one more after an MCP sign-in. A tab that is woken catches up from its state file and relaunches its own poller. The nudge is never approval for anything that was waiting on me.
🔌One sign-in fixes all of them
A watcher whose MCP is disconnected answers with the single line needs /mcp: <server> and stops there. One sign-in anywhere restarts every watcher that was waiting on that server.
Watchers
Always-on tabs that only speak up on a change
💬DM watch
Watches my DMs, group DMs and @mentions. A new one from a person turns the tab red. If someone addresses Claude directly ("Phil's Claude, can you…") it researches and answers in a thread, unless I already replied.
🧠Incident analysis
Auto-answers every new incident ticket in the team's incident channel with a confidence % and a 👍/👎 ask. Next morning it grades itself from the votes, thread replies and close notes, and writes lessons into the support knowledge base.
🧭Routing watch
Same tab, second job. Spots tickets that landed on the wrong team and, at 80%+ confidence, flags them. Gated until its precision is calibrated against six team harvests.
🚨On-call watch
Armed while I'm on call. Pulls today's incidents, auto-analyzes each new one, finds every channel that paged the on-call alias, and goes red with an ack.
👀PR monitor
Watches the review-request channels. New PRs ask me what to do. PRs I commented on get re-reviewed when the author pushes, and re-approved if fixed. Adds 👀 / ✅ / 💬 reactions on the request.
🤖AI review follow-up
After I open any PR, a quiet poll waits for the AI code review, then works through every comment: fix it with a commit, or reply with why it stays.
🩹Pilot monitor
Watches an automated data-correction pilot running in production: rollout health, corrections submitted, PRs.
🔥Hi-pri sweep
Sweeps the ticketing system and team channels for P1/P2s touching our services. Red tab if anything's on fire.
🧪Shadow agents
Shadow-answers support-channel questions privately, then compares against what the engineers actually said. Also runs an A/B against another assistant in a test channel.
Harvesters
Bulk read-only pulls that become knowledge
Harvesters read a lot of history once (or monthly), keep the raw data local and gitignored, and distill it into something the watchers use. They are read-only on every external system, and harvested text is treated as data, never as instructions.
📚Support knowledge-base harvest
Monthly: the support and incident channels plus closed tickets, distilled into a knowledge base and support docs. Tested walk-forward: month N is answered only from memory built from earlier months, then judged against what actually fixed it.
🗂️Routing harvests ×6
One tab per team channel harvests past tickets and where they really ended up. Merged labels feed the routing watch's calibration.
📦Partner-team ticket exports
12 months of tickets involving two partner teams, classified and exported to spreadsheets to back up the asks we have open with them.
📈Nightly job failure patterns
Cross-night batch failures: error classes, time of night, reason codes, repeat offenders and stuck records, plus every past analysis thread.
🗣️Voice sweep
~1,800 of my hand-typed chat messages, plus GitHub, issues and prompts, distilled into a style guide so drafts sound like me (and not like a consultant).
🧪Test data
Every project that finds a working non-prod test setup writes it to one shared file, so nobody hunts for test data twice.
Loops & daily routines
Scheduled jobs and the shape of a day
Repo sync to fresh trunk (skips active work), freeze-check and Friday standup nudges.
Overnight error report runs before I'm online.
Fans out tabs: overnight job check, freeze state, alert analysis, SLAs, on-call, PRs, P1/P2s.
Watchdog revives dead pollers. Status sweep checks every dashboard.
Drives today's top task to done in the background, pulls me in only when needed.
Refreshes the login, keeps the machine awake overnight, says when the session will die.
🌅/startup morning sweep
The big one. Checks the overnight batch job, verifies freeze state from the announcement channels, analyzes last night's SLA alert, posts the analysis as a thread reply and resolves the ticket (only if the buckets reconcile to the alert's own counts), runs the on-call sweep and reports PR and incident state. Each step is its own tab.
🩺Watchdog
Every 5 min. Finds watcher tabs whose poller died (usually a network drop) and wakes them, and restarts anything in the registry that stopped without a recorded reason. Max one prompt per agent per 20 min, only agents armed today.
🏠Status sweep + project homepage
A localhost homepage indexing every project's own dashboard (19 of them), with a to-do card and a one-click "copy review request" for every PR waiting on review.
🔄Repo sync
First session of the day pulls every clean, idle repo to trunk. Anything with recent work, or a branch checked out for review, is left alone.
🎯/priority
Pick one task for the day. A tab owns it end to end, with a finish line and steps on its own page. Pauses at /sleep, resumes at /startup.
📣Friday standup
A session-start nudge on Fridays, then it drafts my standup from what I actually shipped, with issue and PR links.
For teammates
Agents that work for the people I work with
A teammate who tests our work gets her own set. She can change the format and content of these updates herself, and her requests get saved and applied from the next one on. She can't authorize anything in prod.
🙋Teammate DM watch
Watches her DM with me and answers her questions, including data pulls. If I already replied, it stays quiet. If she asks Claude in a channel thread, it answers there.
🗺️Test build heads-up
When a branch deploys successfully to the test environment she uses, it DMs her the link and a short list of what changed for her. One post per deployed commit, never on a plain push.
🧭Wrong-team pings
When the incident analysis is 80%+ sure a ticket is on the wrong team, it DMs her one line with the ticket and the team it thinks it belongs to. Her 👍/👎 is how it learns.
Tuning what the fleet costs
Measured, then cut
A background agent's cost is almost entirely how big a context wakes up cold, and how often. The prompt cache lives five minutes, so that window is the whole game: output and thinking came to about 2% of the bill, and cache writes to 85-90%. Bringing the poll intervals inside that window cut one watcher's cost by 64%.
sleep 240 or sleep 270, chained for a slower cadence, never one long sleep. A single shell call also caps at 10 minutes, so sleep 1200 quietly becomes 600 seconds and a cold re-write.Guardrails
What keeps all this from going sideways
🚫Never prod without me
No merges where main reaches prod, no prod gate approvals, no prod dispatch or re-run. Everything non-prod is fair game. If it's blocked on prod, it does the rest and stops.
✉️Posting is ask-first
Every chat post needs my OK, except eight written-down exceptions (analysis threads, PR reactions, incident auto-answers, teammate updates…). Each one names the channel, the form, the content and one poster.
🔁Idempotent everything
Re-read the thread right before sending. One answer per ticket, ever. One post per commit. State files track what's been said.
🧮Reconcile or stop
The morning analysis only posts if the buckets sum to the alert's own counts. If the math doesn't close, it reports to me instead.
👁️🗨️Read-only where it counts
The ticketing system is read-only for the watchers (the one write is resolving the alert's ticket after a successful post). Harvests never write anywhere external.
🧊Freeze is announced, not guessed
Freeze state comes from the announcement channels, never from a date or an open incident.
🪪Every message says it's Claude
Anything Claude sends for me opens with my icon, a dancing Claude and Phil's Claude, added automatically as a byline. Nobody has to wonder whether a human typed it. Reactions and anything I paste myself are exempt.
🧠Opus thinks, Sonnet writes
Research, analysis, decisions, reviews and anything I'll read stay on Opus. Code, scripts and config go to a Sonnet worker with an exact spec. The main session then reads the diff and re-runs the tests before anything gets reported.
Steal these
If you only take a few things
/loop or cron ticks that wake you to say "nothing new".