Bug: Remote SSH Workspace becomes unusable after auto-continue limit is reached #67

Activity

1. Lexus2016 commented on Aug 24, 2026

   

   Lexus2016

   on Aug 24, 2026

   Owner

   More actions

   Thanks for the detailed report — the reproduction steps and the exact log lines made this findable.

   The auto-continue limit turned out to be a red herring. The real defect is a missed termination in the SSH transport, and it explains both symptoms at once.

   runSshSingle() awaits a promise that resolves only from the transport's onDone. Four paths in ClaudeSSH.send() ended a run without ever reaching it:

   1. A cleanly closed socket emits no 'error'. sshd's ClientAliveCountMax/LoginGraceTime drop, a NAT/firewall idle-reap on a long-running turn, and the client's own conn.end() all land on 'close' — and there was no conn.on('close') handler at all.
   2. The idle watchdog and the hard cap called conn.end() and nothing else. That only settles a run when a channel is open for ssh2 to tear down with it.
   3. The abort listener was registered inside the conn.exec callback, so a Stop pressed while the connection was still handshaking was silently ignored.
   4. conn.on('error') called onDone directly, bypassing the once-guard, so the socket close behind it fired it a second time.

   A pending await keeps the session's activeTasks entry alive, and nothing reaps it — the 15s orphan sweeper opens with if (activeTasks.has(sid)) continue. That is why the chat stopped accepting messages and why Restart Session did nothing: it was hitting "Task is still running" and returning early. Only a server restart cleared it.

   Fixed in #68:

   • Every way a remote run can end now goes through finish(), and the abort listener is registered before the connection is opened — so Stop always works, and a dropped link always ends the turn instead of hanging it.
   • Restart Session is now a recovery action, not a refusal. It aborts the live turn, waits for it to release the session, reaps it if it never does, and drops the cross-connection lock with it.
   • It is reachable without a working turn. The in-message button only rendered when the server managed to send a session_restart_available frame — and the run that needs recovering is precisely the one that never sends one. There is now a 🔄 Restart button in the session bar whenever the chat has a Claude session.
   • You do not lose the conversation. Restart clears claude_session_id, and the next message replays this chat's full history into the fresh remote session. That part already existed; it just was not reachable.
   • The failure message names the root cause. Instead of only "Agent did not complete after 3 auto-continues", it now says why — hitting the per-run turn limit (raise Max turns) reads differently from the SSH link failing, and those need different fixes from you.

   Two regression suites were added: one drives the SSH transport through a fake ssh2 client and asserts onDone fires exactly once across eight ways a run can end (the first scenario hangs on the old code), and one boots a real server against a claude that never exits and pins that a restart aborts and releases the session.

   Two of your suggestions I did not implement, so you know where things stand: there is no persistent SSH connection-status indicator, and no automatic re-creation of the remote agent. The connection is opened per run rather than held, so there is no long-lived state to display or revive — a failed turn now reports the SSH error inline instead. If you would still find a status indicator useful, a separate issue for it is welcome.

   Please reopen if you hit anything like this again after the next release.
2. Lexus2016

   closed this as completedin #68on Aug 25, 2026
3. malys commented on Aug 25, 2026

   

   malys

   on Aug 25, 2026

   Last edited by malys

   Author

   More actions

   7.9.0
   "⚠️ Agent did not complete after 3 auto-continues — it kept hitting the per-run turn limit (raise "Max turns" for this chat). The chat is still usable: send another message to continue, or use Restart Session to start a fresh remote session with this chat's history replayed."
   Restart session -> no effect

   ⏳ Auto-continuing (1/3) — resuming on remote...

   I'll read the catalog files to see what conditions and actions already exist.

   ⏳ Auto-continuing (2/3) — resuming on remote...

   ⏳ Auto-continuing (3/3) — resuming on remote...

   ⚠️ Agent did not complete after 3 auto-continues — it kept hitting the per-run turn limit (raise "Max turns" for this chat). The chat is still usable: send another message to continue, or use Restart Session to start a fresh remote session with this chat's history replayed.
   ✅ Done — 8 calls
   $0.643 · 4 in / 804 out · 3 turns · 4% ctx · 12.9s
   Done

   @Lexus2016 issue not solved
4. Its-My-Work

   added 2 commits that reference this issue on Aug 26, 2026

   fix: a remote SSH chat must never become permanently unusable (Lexus2…

   ...

   076f937

   fix: a turn that strands a background task must not report success

   ...

   3b2becc