Skip to content

Remote runtime does not recover after macOS lock or sleep over Tailscale #9092

Closed

Description

@matheustimbo

Summary

A macOS Orca desktop paired with an Orca runtime on Ubuntu over Tailscale does not recover after the Mac is locked or sleeps. When the Mac returns, the terminal and control sockets are reset, but the UI remains broken until Orca is fully closed and reopened.

Reproduction

  1. Run Orca on an Ubuntu host and expose the pairing endpoint on its Tailscale address.
  2. Pair and control that runtime from Orca on macOS.
  3. Lock the MacBook long enough for the WebSocket connections to become half-open.
  4. Unlock the MacBook and return to an active remote terminal.

Actual behavior

The terminal shows one or both of these errors and does not recover:

  • Remote Orca runtime stopped responding; the stream connection was reset.
  • Remote Orca runtime stopped responding; resetting the control connection.

Closing and reopening Orca is required.

Expected behavior

Remote subscriptions should remain alive across an arbitrary offline or sleep interval, retry with bounded backoff after wake, and restore the terminal without showing a fatal PTY error for transient network unavailability.

Root cause

  • Shared-control reconnects stop after seven attempts and permanently close passive subscriptions after roughly 31 seconds plus jitter.
  • The dedicated terminal stream attempts one immediate resubscribe. If Tailscale or Wi-Fi is still restoring the route, that attempt fails permanently and surfaces a recoverable transport error as a fatal terminal error.
  • Pending attempts were not generation-guarded across detach and reattach, so a late failure could also rearm stale retries.

Fix direction

Keep passive shared-control subscriptions retrying at a capped 15-second delay until explicitly closed; retry terminal resubscription with the same capped backoff; suppress only known transient transport errors; and invalidate stale async reconnect work across terminal lifecycle changes.

Activity

  1. e-jung commented on Jul 17, 2026

    @e-jung

    Corroborating this from another real deployment (@ericjung):

    • macOS Orca client paired to an Ubuntu Linux arm64 headless orca serve runtime over Tailscale
    • server is Orca 1.4.144
    • after leaving the Mac asleep/locked and returning, the client repeatedly shows Remote Orca runtime stopped responding; the stream connection was reset.
    • the terminal does not recover on its own; closing and reopening Orca is the current workaround

    A read-only source/runtime audit matches the mechanism described here: the 25-second inbound-silence watchdog can fire on the first timer tick after macOS resumes, passive shared-control recovery ends after the finite seven-attempt schedule, and terminal resubscription can lose its one recovery attempt while the control lane is still coming back.

    This also reproduces the broader recovery race in #8180 with a Linux host rather than a Windows host. No auth tokens, device records, or private runtime logs are included here. I can collect sanitized macOS pmset timing plus Orca client reconnect logs on the next occurrence if that would help validate a fix.

  2. fitz04 commented on Jul 27, 2026

    @fitz04

    Repro on macOS↔macOS over Tailscale, v1.4.158 — freeze during active use (no sleep/lock involved)

    • Client: Orca 1.4.158, MacBook (Apple Silicon, macOS 15). Server: headless orca serve 1.4.158 on a Mac mini M4, paired over Tailscale (ws://100.x.x.x:6767).
    • Terminal (Claude Code TUI) froze twice within ~90 min while actively typing — last frame stays on screen, input silently dropped. The MacBook was awake and in use, so this reproduces beyond the sleep/lock trigger (possibly App Nap tick stall or a Tailscale path change).
    • Server side after the freeze: agent processes still alive (orphaned), and the daemon log shows a serve-* session create→kill churn right after the freeze (23:24:47–23:29:01 UTC, several sessions killed within seconds of creation). The PTY session was recreated — orca terminal read returns only the shell command echo while the GUI still renders the stale TUI frame.
    • Client side: no crash, renderer heap ~125MB, no disconnect events traced. Minutes later the CLI could list/read remote terminals fine, so the transport itself had recovered — only the app's subscription never did.
    • Workaround confirmed: full app quit/relaunch; recovered agent sessions from their transcript files on the server.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

bugSomething isn't workingorca-remote-serverRemote Orca Server / orca serve / paired remote runtime

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions