dsh-superagent-prototype: two defects with fixes (service process on Node 24; late GitHub mentions lost)

[mikel38487]

I installed deepseek-ai/dsh-superagent-prototype following its SETUP.md and ran it against a real repository (Ubuntu 26.04 under WSL2, Node.js 24.22.2, dsh 0.2.1-rc.5, GitHub CLI 2.103). Two defects showed up in the first hour. I fixed both, added regression tests and verified them; the prototype repository has pull requests and issues disabled, so I am posting the report here.

1. The service script cannot find a running service on Node.js 24

Node 24 names its main thread MainThread, and ps -o comm= / ucomm= reports that name for the process. processes() in scripts/dsh-superagent matched it against node and dsh, so status printed pids: [] and exited 1 while the service was listening and its tmux session was alive. start then waited out its 60-second health check and reported a timeout, and stop found nothing to stop.

$ ps -p 6651 -o pid=,comm=,ucomm=
   6651 MainThread      MainThread
$ ps -p 6651 -o pid=,args=
   6651 node /usr/local/bin/dsh --profile superagent --no-open
$ python3 scripts/dsh-superagent status
{ "pids": [], ... }   # exit 1

Fix: take the process name from the first word of its command line instead of the comm column. The ps format stays the same, so tmux, sleep and dsh plugin … are still rejected. A regression test uses a fixture whose comm is MainThread.

2. A comment can be lost for good when GitHub reports the mention late

pollNotifications read a notified thread from max(thread cursor, notifications window), where the window starts at the previous round start minus five seconds. intakeComments drops comments older than that, and readNotified then moved the thread's cursor to the new round start — past the dropped comment. The comment is never read again, while the notification stays unread in the account's inbox because the since query no longer returns it ("only show results that were last updated after the given time").

Reproduced twice on a live repository:

• two mentions in one issue never reached the gate;
• the thread cursor advanced 19:42:59 → 19:47:05 with no trigger in between;
• gh api notifications still listed the thread with updated_at 19:42:07, older than the window being read.

Fix: a thread read before starts at its own read position; a thread the Source never read starts an hour before the notifications window. Re-reading is harmless — trigger ids are deterministic and Intake.receive drops duplicates. Two regression tests cover both cases.

Verification

pnpm typecheck is clean, the full pnpm vitest run passes (59 files, 1200 tests), pnpm build succeeds, and the patch applies to a fresh clone of main. Both new tests fail without the patch. Two independent reviewers looked at the change; their findings (a weaker first version of fix 2, an existing false positive in serves() for one-off dsh --profile … commands, and a ps parsing caveat with spaces in the binary path) are documented with the branch.

Minor documentation gaps

SETUP.md and docs/006-deployment.md do not say that the bundle is POSIX-only: worker-bin is #!/bin/sh shims and the service script uses ps, tmux and fcntl, so native Windows cannot work — WSL was needed. Also gh 2.48+ is required, but the probe only prints the version; Ubuntu 26.04's apt package is 2.46, which fails later in setup.

Where the code is

The changes are on branches of a fork, ready to become pull requests if the repository is opened for contributions:

• pr/service-command — https://github.com/mikel38487/dsh-superagent-prototype/tree/pr/service-command
• pr/notified-thread — https://github.com/mikel38487/dsh-superagent-prototype/tree/pr/notified-thread

I can attach the patch or a short reproduction script for the second defect on request.

P.S. The author address in the prototype's package.json (team@dsh.dev) does not deliver: dsh.dev has an SOA record but no MX, A or AAAA.