Amplifier Skill

/monitor

A turn that doesn't end until the thing you're watching is actually done.

Shipped · amplifier-bundle-skills

July 2026

The Problem

“I'll keep working on it and let you know when it's done.”

Then the turn ends. The prompt sits there. Nothing is watching anything. You come back forty minutes later to find out the agent has been waiting on you the whole time.

The promise

The model says it will follow up when a build, a CI run, or a deploy finishes.

The reality

The turn completes immediately. There is no timer, no background process, no callback.

The gap

Nothing in the loop was ever wired to keep that promise. It was never a capability — it was a habit.

What It Is

A bounded loop that holds the turn open.

01

Sleep

Wait your interval in short bounded waits.

02

Check

Narrowest shell command that answers the question — exit code, status code, grep -c.

03

Decide

Not yet → loop again. Otherwise stop and report.

The turn ends on exactly three conditions: the thing is DONE, it has FAILED, or it genuinely needs you. Because Amplifier's notification fires on turn completion, holding the turn open is what makes the ping land at the right moment — desktop, terminal, or phone via ntfy. Walk away.

Usage

Describe the thing and the bounds.

/monitor the CI run for PR 412, check every 2m, stop after 1h /monitor the deploy job, ping me if it fails /monitor ./build.log until the migration finishes
What it can watch

Anything checkable from a shell: CI runs, log files, HTTP endpoints, background jobs, long builds, deploys.

It also invokes itself

The skill is model-invocable. When the agent is about to promise a follow-up on something bounded and checkable, it can reach for /monitor instead — turning a hollow promise into a real one without you asking.

What Lands On Your Phone

Verdict first. The verdict is the notification.

The notification body is the first ~100 characters of the final message — everything past that is invisible on a phone. So line one is a state token plus the single most decision-relevant fact.

DONE: CI green on PR 412, 318 tests passed
NEEDS YOU: deploy blocked on manual approval gate, job 4471
GAVE UP: 2h elapsed, endpoint still returning 502
PAUSED: stopped watching at your request -- run still going

PAUSED means it stopped watching — not that it stopped the thing being watched.

Cost

Measured on real runs.

~1.2¢
per check
$0.16
total for a 13-check monitor
<$2
8 hours at 5-minute intervals
~15K
input tokens per check — flat, however long it runs

By default the polling loop is delegated to a sub-agent with clean context on a fast model, so each check pays the bundle's system-prompt floor rather than re-sending your whole session. Prompt caching holds that input flat. It can also run inline when you want to watch progress happen or steer it mid-flight.

How It Ships
1

SKILL.md file. That's the whole feature.

No new module. No new code. No new mechanism. It uses the existing bash tool and the orchestrator's normal turn behavior — the same primitives that were already there.

The missing piece was never capability. It was discipline.

Try It

Next long-running thing you kick off, hand it over.

# update your bundle, then: /monitor the CI run for PR 412, check every 2m, stop after 1h
Sources

Where these numbers come from.

Data as of 2026-07-28 · Status shipped, active · Source microsoft/amplifier-bundle-skills, skills/monitor/SKILL.md — one file, 276 lines, PR #46.

Verified against the shipped SKILL.md
  • All four verdict tokens and their example lines, verbatim.
  • The ~100-character notification-body limit and the verdict-first rule.
  • Delegated-by-default execution: sub-agent, clean context, fast model role. Inline is documented as the alternative.
  • Single-file delivery, no accompanying module or source code.
Cost figures & caveats
  • ~1.2¢/check and the $0.16 / 13-check total are measured from real monitor runs.
  • The <$2 / 8-hour figure is projected from that per-check cost (96 checks), not a completed 8-hour run.
  • Per-check cost varies with bundle size and model role — representative, not guaranteed.
← → or click to navigate
More Amplifier Stories