Skip to content

Check in on a thread every 30 minutes while it says something is running - #218

Merged
yobazy merged 5 commits into
mainfrom
moment-2026-11-12t20-10-54-642z
Nov 12, 2026
Merged

yobazy merged 5 commits into
mainfrom
moment-2026-11-12t20-10-54-642z

Conversation

@yobazy

@yobazy yobazy commented Nov 12, 2026

Copy link
Copy Markdown
Owner

From a marked moment: work an agent starts on pc (a tmux run that takes hours) leaves its thread Done with "🟠 Keep open: round 3 is running on pc", and the only way to see progress was to type "status" into it by hand.

What changes

  • Check-ins. While a finished thread's last reply ends on a 🟠 verdict that says something is running (or in progress, underway, still going), the app sends it a short check-in every 30 minutes: "Automatic check-in from Observatory: a short status card on the running work, please…". Each answer is a fresh status card in the thread, so the latest reply is always the tracker.
  • It stops on its own. The first reply whose verdict no longer says "running" (✅, 🔴, or a 🟠 about something else) ends it. It also stops after 24 check-ins (12 hours). A thread that said "running" more than 3 hours ago, before the app saw it, is left alone, so a launch doesn't ping every old thread at once.
  • Shown and switchable on the verdict line in the Read view: ● Checking in · next in 27m, which you click to stop. Any other 🟠 verdict offers Check in every 30m to turn it on by hand. The choice is per thread and kept across launches.
  • No interruptions. A check-in waits while the thread has a draft in its reply box or a queued prompt, never goes to leads, and wakes an asleep thread without selecting it: followUp gets a background option that attaches behind the current view, so what you're looking at doesn't change.
  • Sandbox: the fake agent answers running… with that verdict, and VITE_CHECK_IN_MS=60000 shortens the wait in a dev build.

Seen working (PC sandbox)

Before (main) After
before after

Clicked off, it offers to turn back on: off

With VITE_CHECK_IN_MS=60000, the check-ins reached the agent once a minute. The fake agent never changes its verdict, so they kept coming. This shot has the first, longer wording of the prompt: sent

Checks

npm run build && npm test (625 pass, including the new tests/checkin.test.mjs); cargo fmt --check, clippy -D warnings and cargo test pass.

Not tried on a real Claude thread yet. Each check-in is a normal agent turn, so a 4-hour job costs about 8 short turns.

🤖 Generated with Claude Code

@yobazy
yobazy merged commit 4b55e53 into main Nov 12, 2026
1 check passed
@yobazy
yobazy deleted the moment-2026-11-12t20-10-54-642z branch November 12, 2026 20:53
yobazy added a commit that referenced this pull request Nov 13, 2026
…nning elsewhere finishes

"Needs you: Nothing until it finishes" with a verdict saying the run is
in progress counted as asking you something, so a run started on the PC
showed in Decisions as if you hadn't started it. Uses the same running
verdict check-ins (#218) watch for. "Nothing yet, but…" still asks.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
yobazy added a commit that referenced this pull request Nov 13, 2026
… you (#221)

From a marked moment (2026-11-13 07:37): a thread that had started a run
on the PC still showed in Decisions with its old "Paste this to start
it" card, as if nothing had been sent.

Two causes, two commits:

1. **Decisions missed the new reply.** It re-reads a thread only when
its status changes. The turn that started the run (a "! cmd", then a
quick reply) never moved the status, so the card from before stayed.
Decisions now re-reads its listed threads, and any that moved in the
last six hours, when it opens and every minute while it's open. A
replayed moment's thread (`scripts/drive replay`) now reads in Decisions
too.
2. **The new reply counted as a decision anyway.** Its Needs you was
"Nothing until it finishes. I'm watching the run…" with the verdict "🟠
Keep open: round four is running". A Needs you that starts with
"Nothing"/"None" (with no "but"/"unless" in that sentence) no longer
counts while the verdict says something is running, the same check the
30-minute check-ins (#218) use. "Nothing yet, but confirm the API key"
and "Nothing until you've tried it" still ask.

The thread stays in Follow up, and check-ins keep asking it for status
every 30 minutes.

## Before / after
Browser mock with the moment's own transcript stood in for a thread, in
Decisions:

| Before | After |
|---|---|
|
![before](https://raw.githubusercontent.com/yobazy/observatory/pr-shots/moment-2026-11-13-decisions/before.png)
|
![after](https://raw.githubusercontent.com/yobazy/observatory/pr-shots/moment-2026-11-13-decisions/after.png)
|

Also checked in the mock: with only the transcript up to the
start-command reply, the thread shows (correct); the next reply landing
with no status change leaves it showing; `recheck()` takes it off.

The PC sandbox couldn't show it: its fake threads sit as Working, so
none reach Decisions.

## Checks
- `npm run build && npm test` pass. One run hit the existing timing test
"answers read in linear time" (over 200 ms while the Mac was busy); it
passed on three reruns.
- No Rust changes, so cargo was not run.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant