agentleFS
DeepusleepyDeepusleepy/t3-tots/skills/t3-watch/SKILL.md

t3-watch

Skill110 starsChanged 18 days agoMITView on GitHub

Watch a T3 thread, log, process, benchmark run or PR and ping the user when it finishes, needs them, stalls or fails, or keep threads moving while the user is away. Use for "ping me when <X> finishes", "watch <X>", "tell me when <X> needs me", "let me know when <X> is done", "monitor <X>", "keep it going overnight", "I'm away, handle it".

The file

We split this skill into the rules it gives an agent

4 rules. About 0 are common, 3 are distinctive, and 1 is very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈3

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 1

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Create a routine that runs every 10 minutes, or every 5 if they say "soon", with the prompt below filled in.

    line 21
  2. Distinctive:

    Don't set an end time unless the user gives one.

    line 24
  3. Distinctive:

    Keep a short log of what you decided and why, and send it to the user when they're back or when the watch ends.

    line 57
  4. Very specific:

    Start it when the user hands over control (t3-tots-rules "Decisions").

    line 60

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Follow the t3-tots-rules skill (skills/t3-tots-rules/SKILL.md). Steps that use the box, ListMachines, routines, cards or SendToUser are Grok Bot only; other agents follow t3-tots-rules "Grok Bot only". A watch is a routine you create only when the user asks for one. It reports by default. In delegated mode (t3-tots-rules "Decisions") it may also answer, approve, nudge and resume.

Pick the check

  • A T3 thread on a computer where t3listen runs with phone answers on: finished, failed and cancelled events already reach chat (t3-phone-answers). A watch is still useful for stalls, approvals and queue problems, which t3listen doesn't report.
  • A T3 thread: t3ctl watch --thread '<id>[,<id2>]' --state watch-<name>.json --json. One call covers several threads and prints nothing when nothing changed. Give each watch routine its own state file, so two watches on one thread don't hide changes from each other.
  • A log: tail -n 30 <log>. A process: pgrep -fl <name>.
  • A PR: gh pr view <n> -R <owner/repo> --json state,statusCheckRollup,reviews,comments.
  • T3 scheduled tasks: t3ctl tasks lists them. Create one only on an explicit ask, with --upsert ... --disabled first, then --enable once the user confirms.

Start a watch

  1. Resolve the target and take one snapshot now with the same command and state file the routine will use. It prints a first event with the state, and the routine's first run then reports only real changes. Tell the user the state, including any queue-held or limit it shows.
  2. Create a routine that runs every 10 minutes, or every 5 if they say "soon", with the prompt below filled in.
  3. Reply in one line: what you're watching, how often, and the current state.

Don't set an end time unless the user gives one. A watch runs until the user stops it. When the work finishes, say so and offer to stop it.

Routine prompt

Watch <TARGET> on <COMPUTER> (machineId <ID>). Follow t3-tots-rules and t3-watch.
Mode: <report | delegated until <end condition>>.
Check: <the command from "Pick the check">
Rules:
- No changes in the output: send nothing and write nothing to memory or a journal.
- finished: tell the user "<TARGET> finished at <time>", 2 lines of result, offer files (t3-get-output) and offer to stop this watch. Report the finish only once, and skip it if t3listen already reported it in chat. A finish can also mean the thread is asking something or has a plan ready; say which.
- question, approval, limit, error, queue-held, queue-stuck, stale-request, quiet, stalled, unreadable: follow t3-watch "Handle an event".
- first, state, reply: send nothing, unless the user asked for progress updates.
- Computer unreachable, or T3 closed there (`T3_NOT_RUNNING`, `T3_NOT_REACHABLE`): tell the user once, then keep checking quietly until it is back.

Handle an event

In report mode, send the user one line per event: what it needs and the last reply in 200 characters or fewer. For limit, offer t3ctl resume-at-reset. Keep watching.

In delegated mode, act first and report after:

EventWhat to do
questionAnswer it or escalate it, as t3-tots-rules "Decisions" says.
approvalDecline it with t3ctl answer --thread <t> --request <id> --decision decline --redirect, so the thread skips it or takes a safer way and keeps going. On REDIRECT_NOT_DELIVERED, run the send command it prints. Log the prompt and its reason for the report. Never accept one yourself.
queue-stuckt3ctl queue --thread <t> --resume once.
queue-heldLeave it. Someone held it on purpose. Tell the user.
limitIf autoResume isn't set, run t3ctl resume-at-reset --thread <t>.
stalledRead the last reply. If it is waiting on a background job, check the job. Send one nudge, for example "The capture finished; continue." Don't nudge the same stall twice.
quietCheck t3ctl subtasks --thread <t> and t3ctl read --thread <t> --last 2. If it is truly stuck, tell the user. Don't interrupt it yourself.
stale-requestClear it in T3 if it belongs to a finished run, else tell the user.
errorTell the user with the error text.
unreadableTell the user once with the error. Keep checking. The watch reports it again only if the error changes.

Keep a short log of what you decided and why, and send it to the user when they're back or when the watch ends.

Delegated mode

Start it when the user hands over control (t3-tots-rules "Decisions"). Before the first check:

  1. Save the scope and the end condition to memory.
  2. On a Mac, keep it awake for the run (t3-tots-rules "Computers").
  3. Confirm in one line which threads, until when, and what you'll still ask about. Approvals are declined with a redirect and listed in the report, never accepted.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.