fix(restart): keep watching a CLI that is slow to quit, and hand over the resume line

[eneskirca]

Fixes #899.

What happened. "Restart agent and shell" on a long-running Claude node (a cron session open for 19 h, 28 MB transcript) reported:

Restart failed: the pane did not return to a shell in time, so the CLI was not relaunched. Nothing was killed — check the pane.

The pane then showed Claude's own farewell and a clean shell prompt. So the CLI did quit, a moment after performExitPhase stopped watching at RESTART_EXIT_TIMEOUT_MS (6 s). Nothing was left to send the resume line, and the node was left at a bare shell with no agent in it.

Why the exit poll, not the detection, is the likely failure. Once claude quits, #{pane_current_command} becomes the shell, and isShellCommand accepts it on the very next poll. The pane query is a local tmux display that takes milliseconds. What remains is a CLI that took longer than 6 s to quit. Not measured: how long that exit really took, and why a large session is slower to quit.

Change

• performExitPhase takes an optional lateExitMs. After the base window it keeps polling, but only while the pane still reads. A pane it cannot see ends the run exactly as before. A query that wedges inside the base window still lapses at the base deadline. When the option is absent, behavior is byte-identical.
• Both user-asked restart paths use RESTART_LATE_EXIT_MS (60 s):
  • performRestartResume, by default.
  • The restart-shell branch in TerminalNode, which is the action from the report.
• Not changed: the Eco sweep and Pause keep the bare 6 s. The sweep is serialized across the canvas and must not stall on one node.
• The exit-timeout notice (exitTimeoutNotice) names the window actually waited. It also carries the bare resume line (claude --resume <id>) for a CLI that quits after even the late window. A custom agent with no resume line gets "Check the pane." instead of a guessed command.

Tests (agent-restart.test.ts)

• New cases:
  • A late quit after the base window is resumed.
  • A CLI that never quits is still reported, and only after the late window runs out.
  • An unreadable pane does not spend the late window.
  • performExitPhase has no late window unless the caller asks for one.
  • Two cases for the notice text.
• Three existing "never quits" cases now advance past the late window as well.
• Mutations:
  • Dropping the performRestartResume default turns 2 cases red.
  • Dropping the pane === null stop turns 2 cases red, including the existing wedged-query case.
• npm run typecheck is clean. Full suite on Linux: 12886 passed, 53 skipped.

Trade-off. The node stays held by guardConcurrentRestart for the whole wait, which is deliberate: a background watcher would release it, and a second /exit or restart could then collide with a late resume. The cost is that a node whose CLI never quits holds a bulk restart for up to 66 s instead of 6 s. Bulk restart only picks idle nodes, and those normally quit in well under a second.

Not covered

• The restart-shell branch's use of the late window has no test of its own. It is one argument at a component call site.
• Eco and Pause can hit the same late quit. That is left for a separate change.
• The resume delivery's first write still carries no kill-line. A user who types into the shell during the longer wait can splice into the resume line. The risk is not new, but the window is now longer.
• No device run. A Mac check against a large, long-running session is owed.

Surfaces

• Desktop and Server Edition: both run the same renderer code, so both are affected.
• Relay tabs use the same closure.
• Mobile: not applicable. It has no restart action.

🤖 Generated with Claude Code