khoj.el kills the Emacs server (and other buffer-less processes) when the khoj server is unreachable

[dustinfarris]

Describe the bug

When the khoj server is unreachable, khoj.el's failed-request handler kills unrelated Emacs processes — including the emacsclient server listener, whose socket file is then removed by server-sentinel. The practical symptom: ~60 seconds after starting Emacs (when the auto-index timer first fires), emacsclient stops working with can't find socket; have you started the server?, and every hourly re-index kills any manually restarted server again. Buffer-less processes other than the server listener (e.g. live emacsclient connections) are killed too.

Root cause

khoj--delete-open-network-connections-to-server (khoj.el:1118, v2.0.0-beta.28) iterates the whole process-list and matches each process by the name of its buffer:

(defun khoj--delete-open-network-connections-to-server ()
  "Delete all network connections to khoj server."
  (dolist (proc (process-list))
    (let ((proc-buf (buffer-name (process-buffer proc)))
          (khoj-network-proc-buf (string-join (split-string khoj-server-url "://") " ")))
      (when (string-match (format "%s" khoj-network-proc-buf) proc-buf)
        (ignore-errors (delete-process proc))))))

Two problems combine:

1. Many processes have no buffer (process-buffer → nil), and (buffer-name nil) returns the name of the current buffer — it does not error.
2. This function is called from the url-retrieve failure callback in khoj--send-index-update-request (khoj.el:447), where the current buffer is the failed url-http buffer, e.g. *http localhost:42110* — which matches the "http localhost:42110" pattern.

So every buffer-less process in the session matches and gets delete-processd. That includes the server-start listener process; deleting it flips its status to closed, and server-sentinel then deletes the server socket file.

The deletions also cascade: deleting one pending url-http process fires its sentinel, which invokes the failure callback and khoj--delete-open-network-connections-to-server again, recursively.

To Reproduce

1. Configure khoj.el with khoj-auto-index enabled (the default) and a khoj-server-url pointing at a server that is not running.
2. Start Emacs with server-start (or a daemon), verify emacsclient --eval t works.
3. Wait for the first auto-index run (60s), or trigger M-x khoj--server-index-files manually.
4. emacsclient now fails with can't find socket; lsof -U -a -p <emacs-pid> shows the listening socket fd is gone, and the socket file has been removed.

Backtrace captured with advice on delete-process at the moment the listener dies:

delete-process(#<process server>)          ; ← the Emacs server listener
khoj--delete-open-network-connections-to-server()
#f(compiled-function (status) ...)((:error (error connection-failed "deleted\n" :host "localhost" :service 42110)))
url-http-activate-callback()
url-http-async-sentinel(#<process localhost> "deleted\n")

Suggested fix

Only consider processes that actually own a live buffer:

(defun khoj--delete-open-network-connections-to-server ()
  "Delete all network connections to khoj server."
  (let ((khoj-network-proc-buf (string-join (split-string khoj-server-url "://") " ")))
    (dolist (proc (process-list))
      (let ((buf (process-buffer proc)))
        (when (and buf
                   (buffer-live-p buf)
                   (string-match-p khoj-network-proc-buf (buffer-name buf)))
          (ignore-errors (delete-process proc)))))))

Optionally, khoj--server-index-files could also skip the index run entirely (with a message) when a quick /api/health check fails, instead of building and sending the full multipart body only to fail per batch.

Platform

• Server: Docker (server not running is the trigger)
• Client: Emacs 30.2, khoj.el 2.0.0-beta.28 (macOS)

If applicable, share the file type you were trying to index

org-mode (irrelevant to the bug — any failed index request triggers it)

[tianmind-studio]

I'm working on the focused fix suggested here: only inspect processes with live buffers, match the Khoj URL buffer name literally, and add an ERT regression test covering the matching connection, a bufferless process, and unrelated/lookalike buffers. I'll link a draft PR once the test suite passes.