[a-min-7]

ghostel-exit-functions runs before the buffer is reaped, which is the right place for a client to react to its own child dying — but on the native-PTY path the child's exit status never reached it. The native reaper writes the status as a bare number into the event pipe; ghostel--events-filter deletes the pipe without publishing it, so the sentinel surfaces "finished" and an rc=0 exit and an rc=1 exit are indistinguishable. This is the API gap behind herdr.el's takeover prompt, which currently re-reads the event stream through :around advice on ghostel--events-filter.

This records the status before the deletion and passes it to ghostel-exit-functions as a third argument, as agreed in #744.

New contract: (BUFFER EVENT STATUS).

• STATUS is an integer, or nil when it cannot be determined.
• A child killed by a signal reports 128 plus the signal number on both backends. process-exit-status alone reports the bare signal number on the Emacs-PTY path, so the sentinel normalizes it; the native reaper already used 128 plus the signal number.
• Handlers that declare two arguments keep receiving (BUFFER EVENT): the runner checks func-arity per handler, so existing hooks (ghostel-eshell--visual-exit included) do not change.

Measured against a real refusal on herdr 0.9.1, native-PTY path, module built from this branch:

case	EVENT	STATUS
second herdr agent attach to a held pane (refused; stderr: terminal … already has an attached client; retry with --takeover)	"finished\n"	1
clean herdr client exit	"finished\n"	0
an attached client detaching normally	"finished\n"	0

EVENT is identical in all three; STATUS is the distinguishing value.

This fixes the status half of #744 only; the stderr half remains open. The child's stderr still does not render into the VT — re-measured empty at hook time in the same run. The takeover prompt no longer needs it now that the status is authoritative, but it is a separate gap and is not addressed here.

Version floor for consumers. This adds a third hook argument, so a consumer that gates on the status should require a ghostel release containing this change. herdr.el's fork keeps its :around workaround on ghostel--events-filter until such a release exists; its Package-Requires then needs the corresponding ghostel minimum.

Test: ghostel-test-exec-exit-hook-receives-status runs rc=0, rc=1 and SIGTERM cases on both local PTY backends, pins the three-argument contract, and asserts that a two-argument handler is still called. It fails against the pre-change code, where the three-argument handler never receives a status.

Refs #744.

[dakra]

@a-min-7 thanks. Would you also be ok with PR #748 ?
I'm not strictly against a numeric 3rd argument (which I personally find cleaner too),
but Emacs process sentinels, compilation-finish-functions, and vterm-exit-functions all do the buffer+event args only and you have to compare the strings in the event argument.
PR #748 is therefore a lot smaller than this here (9 added lines vs 48).
wdyt?

[a-min-7]

Fair enough on the convention — (BUFFER EVENT) is what sentinels and compilation-finish-functions do, and keeping the hook surface unchanged is the smaller change. I ran #748 on both backends with herdr 0.9.1 (current stable is 0.9.5): a refused attach via its nested-herdr guard (rc=1) gives "exited abnormally with code 1\n" and a clean exit gives "finished\n" on both, so the advice on ghostel--events-filter in herdr.el can go, and the split-status case reassembles correctly.

One thing before merging: the newline framing makes the Elisp require the new module. With a pre-#748 module the filter holds the bare number forever and never calls delete-process, so the sentinel never runs — no exit hook fires, and the buffer still looks live indefinitely (measured). MELPA users can get the new Elisp against the released 0.58.0 module, and the minimum-version check will not reject it. Could you either bump ghostel--minimum-module-version in the same change, or make the status path tolerate the un-terminated form?

One documentation point: "same strings on both backends" holds for normal exits only. SIGTERM gives "terminated: 15\n" on the Emacs PTY and "exited abnormally with code 143\n" on native (measured); SIGKILL/SIGINT follow from the same 128+N path (137/130), where Emacs says "killed: 9\n" / "interrupt: 2\n". Still distinguishable, but a consumer written against the Emacs vocabulary silently misses — worth a docstring or commit-message note, and a signal-case test would be ideal. A CHANGELOG entry would also match the recent commits.

Minor, though the newline framing makes it load-bearing: ghostel--event-buf is buffer-local but filters run in the caller's buffer, so holding the status there assumes the same buffer across reads; process-put on the pipe would be sturdier.

If #748 goes in with those addressed, I'll close #746 — the two-argument hook plus the synthesized strings already covers what prompted it (herdr telling a refused attach from a clean exit without reaching into internals), and the convention argument wins.

Environment:

Emacs    GNU Emacs 31.1 (build prefix /Users/amonecke/opt/emacs-31.1; invoked via /Users/amonecke/.local/bin/emacs)
Host     macOS 26.7.1, arm64, MacOSX27.0 SDK
Ghostel  19ff12b (PR head)
Module   Zig 0.16.0; zig build --prefix . -Dtarget=aarch64-macos.13.0 -Doptimize=ReleaseFast -Dcpu=baseline
         (the default native target fails identically on the base branch against this SDK; not #748)
Herdr    0.9.1 at test time

[dakra]

OK, it's merged and I also triggered a new release.
Closing this PR then.

[a-min-7]

Thanks — that was quick. Merged plus a release is more than I expected today; the two-argument hook is the smaller and more idiomatic change, and I'm happy to have lost that argument.

I'll close out my branch here.