Triage
Confirmed on current main (ed809f7ad). This is a real desktop SSH bug. Not already fixed. Stable v0.0.42 (719a76ca1, the build in this report) already launches ssh the same way and already abandons connection setup after 15 seconds. The 0.0.43 nightlies do not handle the Tailscale check banner either. Updating will not change this.
No open issue or PR covers it. #13636 and #11053 are about editor deep links when a host has Tailscale SSH and no local sshd. #10897 only changes the wording of the in-app password / verification-code dialog. Neither path sees this banner.
What you reported. An SSH environment whose Tailscale ACL uses "action": "check" sits on Reconnecting: <host> did not respond during connection setup. The host is up, and a normal interactive ssh prints:
| 1 | # Tailscale SSH requires an additional check. |
| 2 | # To authenticate, visit: https://login.tailscale.com/a/xxxxxxxx |
T3 never shows or opens that URL. The workaround in the issue is right: run ssh yourself, open the link, approve it, and the next T3 attempt connects. Removing the check rule also avoids it.
The Windows browser popup is not T3. There is no login.tailscale.com handling anywhere in this repo, and no Windows-only branch that opens a browser for SSH stderr. Tailscale's own client does that. When check mode trips, the local tailscaled is told to open the URL (control says to open URL …; opening… in tailscale/tailscale#7589). The Windows GUI actually opens it. On Linux, especially tailscaled under systemd on a Wayland session (Omarchy / Hyprland), that open often never reaches a browser (tailscale/tailscale#8551). A normal terminal still shows the banner. T3's ssh is not a terminal, and T3 never reads the banner, so on Linux there is no second chance.
What the code does
1. The row text is the generic setup timeout, not the SSH banner.
The Connections subtitle is Reconnecting: ${connection.error}:
That error is produced when connection establishment loses a race with a 15 second sleep:
SSH prepare runs inside that attempt (resolver → ssh.prepare → ensureSshEnvironment). Tailscale holds the session open until the link is approved, which is longer than 15 seconds, so this timeout always wins. The supervisor then retries forever. The more specific Could not prepare the SSH environment: … error never gets a chance to surface.
Losing that race interrupts the attempt. The child-process spawner terminates ssh when the owning effect is interrupted, so the held session is killed before anyone can finish the check.
2. ssh is non-interactive, and stderr is kept until the process exits.
Desktop SSH, when no password is cached, uses BatchMode=yes because the in-app password prompt service is available:
The command is spawned with stdin closed immediately. The tunnel also passes -n -N. There is no PTY.
stderr is only accumulated for the failure reported after ssh exits:
The check banner is written while ssh is still running. Nothing scans it. The 15 second interrupt throws that buffer away, so the URL never becomes the connection error either. The in-app password dialog does not apply: it only runs after ssh exits with a permission-denied style failure (isSshAuthFailure). A Tailscale check is not that. It is a server banner while the session waits.
The tunnel's own readiness timeout is 20 seconds (SSH_READY_TIMEOUT_MS), which is already longer than the supervisor's 15 seconds, so it does not matter here.
3. Approving the check is enough for a later attempt.
Tailscale remembers a successful check for the rule's checkPeriod. The next ssh proceeds without a new prompt. That is why the manual workaround unsticks T3, and why Windows looks like it works: Tailscale opens the browser itself, the user approves, the first T3 attempt may still die at 15 seconds, and the retry connects. On Linux the approval never happens, so every retry hits the same wall.
Fix
T3 should watch ssh stderr while the process is alive, detect https://login.tailscale.com/…, and open it with the desktop external-open path (shell.openExternal / xdg-open). The same URL needs to show in the connection status, because opening a browser can fail on this session. The 15 second setup timeout should not kill ssh once that banner is seen; either hold the attempt until the check finishes or the Tailscale session gives up, or surface the URL and keep retrying only after the user has had a chance to approve. Opening the browser and still tearing the session down at 15 seconds recreates the stuck loop whenever approval takes longer than that.
v0.0.42 already has both the 15 second race and the "read stderr only after exit" behavior, so this is not a regression in a later build.