fix: keep the turn open while a command runs in the background

[KobeVrancken]

A session that backgrounds a command shows in Paseo as idle while it goes on working. Found on a live session of mine this evening, and it is the one launch kind the subagent hold does not cover.

Stacked on #22, which is stacked on #21, #20 and #19, so this branch carries all four and the diff against main includes them. The commit to read here is the last one.

What was happening

A background command is dispatched exactly as an asynchronous agent is — the tool answers the moment it starts, Claude goes idle, and a <task-notification> wakes Claude when it ends — but only agents were counted, so the Stop hook in between ended the turn. Paseo reads a session as busy from its open turn and from nothing else, and nothing can reopen an ended one.

From the session shipping remi-plus#731, all times local:

time	event
21:32:02	async agent launched; 21:32:23 Stop → turn held
21:37:42	Bash(run_in_background) "Run codex reviewer" started
21:40:39	the last agent reports → nothing outstanding that the adapter counts
21:41:12	Stop → finishTurn(end_turn), Paseo goes idle
21:44:54	the review reports, waking Claude
21:45 → 21:56	Claude launches three more agents and keeps working, with no turn open

transcript-translator.ts said as much in as many words: "A notification for a background command rather than an agent names no card here, and is left alone."

What changed

• A background command now holds the turn the way an agent does. Claude records the id its report will name in backgroundTaskId beside the launching tool result, and the notification carries it as the <task-id>, so the launch and the report are joined the way an agent's are.
• It has no card: its tool call is the launch, which finished, and everything it does goes to a file the adapter only learns the name of when the report arrives. All that is kept for one is whether a report is still owed.
• Its bound is a flat thirty minutes from the moment the turn was held rather than a silence, because a running command shows nothing to measure silence against; starting or reporting another restarts it. Of the 91 background commands that reported across this box's transcripts the median took 4 minutes and the longest genuine wait was a 21-minute code review; the ones that ran for hours were servers and poll loops, which never report at all and are what the bound is there to let go of.
• Each kind of work waited on is bounded on its own, and the turn ends only once every one of them has run out — a command's thirty minutes are not cut short by the agents beside it having gone quiet first.
• Giving up abandons the command as it does an agent, so the next turn does not hold for a poll interval and give up on it again.
• TaskNotification.agentId is now taskId, which is what the tag is called and what it holds: an agent's id or a command's.

Not covered

• A command Claude kills with KillShell is not distinguished. If no notification follows one, the turn waits its bound out rather than ending at the kill — no transcript on this box has a KillShell in it to read the shape from.
• Thirty minutes is a bound, not a promise: a command that runs longer than that and then reports lands outside the turn, which is today's behaviour for every background command.

Verification

pnpm typecheck and pnpm test at the root: 127 adapter tests, all green, five of them new — the hold across a command's report, the bound and the abandon after it, a command outliving the agents beside it, and the translator's counting, settling and replay.

🤖 Generated with Claude Code

https://claude.ai/code/session_016CbhLTXNiW7pV8oEhz76gp