A -p session that outlives its launcher cannot be stopped except by shutdown --force

[jasoncarreira]

Summary

A session started with -p keeps running after its launcher process exits — which is the documented
print-and-exit behaviour — but it is then unreachable by prime-agent stop under any identifier I could find.
The only way I found to end it is prime-agent shutdown --force, which stops every agent and the daemon.

A session created through the daemon protocol ({"type":"create"}) does not have this problem: it appears in
prime-agent list and stop works. So this looks specific to client-owned sessions whose owning launcher has
gone away, rather than to autonomous or headless runs in general.

Environment. prime-agent 0.7.2, daemon protocol 7 (protocol-7-schema-16-1bcb9e7f1a49), build 83a0f9f-dirty, Node 24.11.1, macOS 15 (darwin 25.5.0).

Reproduction

nohup prime-agent --autonomous \
  --autonomous-max-turns 300 --autonomous-max-continuations 200 --autonomous-timeout-ms 900000 \
  -ne -ns -e /path/to/an/extension --cwd /some/repo \
  -p "Count from 1 to 300, one number per turn. Never stop early." < /dev/null &

The launcher exits almost immediately (expected for -p). The session keeps working — I confirmed its transcript
under ~/.prime/agent/sessions/<uuid>.jsonl growing by 8,125 bytes in 10 seconds.

Then:

command	result
prime-agent list	session absent; only an unrelated older session listed
prime-agent list --all	present, lifecycle: live, id: 01a01195-bc7c-705a-884f-600495c5df07
prime-agent stop 01a01195-bc7c-705a-884f-600495c5df07	Error: Unknown active session — and the transcript kept growing (+4,098 bytes)
stop <transcript uuid>	Error: Unknown active session
stop prime-agent:<uuid>.jsonl	Error: Unknown active session
prime-agent shutdown --force	stopped it (and the daemon)

Expected

Either stop accepts the identifier list --all reports for a live session, or list/stop explain that the
session is client-owned and offer a way to reach it. Ending one session should not require stopping every agent.

Contrast that may localise it

# daemon-created session — behaves correctly
DaemonClient.request({type: "create"})  -> activeSessionId d7190de5f08a
prime-agent list                        -> shows d7190de5f08a, lifecycle live
prime-agent stop d7190de5f08a           -> ok

The protocol has list { includeClientOwned?: boolean }, complete_owned_session and promote_owned_session
commands, which suggests client-owned sessions are a known category that the list/stop CLI surface may simply
not cover.

Why it matters to us

We drive an autonomous build tool from these sessions and need supervision to be able to stop a run it started.
Not being able to stop one specific session without killing the daemon is the blocker; everything else about the
autonomous surface has worked well.

Not verified

Whether attach/send are affected the same way, and whether this is specific to -ne (I reproduced only with
-ne -ns -e, because that is the shape our use case requires).