Regression (3.2.4): double Ctrl-C on an agent no longer restarts it fresh — classified as deliberate quit

[waleedkadous]

Problem

Since PR #1244 (shipped in v3.2.4), killing an agent with Ctrl-C Ctrl-C no longer restarts it. The intended UX for double Ctrl-C is restart fresh — it's how users recover a wedged agent. Instead the session now ends cleanly and stays dead, requiring a manual respawn.

Root cause

PR #1244 (bugfix #1241) added exit classification so a deliberate quit (typing exit / /quit) stops auto-restarting the agent. The classifier is:

// shellper-protocol.ts:128
export function isDeliberateExit(exit) {
  if (exit.code !== 0) return false;
  return !exit.signal; // (null/undefined/''/'0')
}

The Claude CLI traps SIGINT and exits code 0 on a double Ctrl-C — indistinguishable, at the exit-status level, from a typed quit. So the double-Ctrl-C path is now classified deliberate → no restart. The #1241 test matrix covered typed quits and crash exits but not the double-Ctrl-C recovery gesture.

Expected behavior

• Typing exit / /quit in the agent → session ends cleanly, no restart (the shellper/launch loops: auto-restart should only trigger on unnatural exits — a deliberate quit should end cleanly #1241 fix — keep this).
• Double Ctrl-C → agent restarts fresh (pre-3.2.4 behavior — restore this).
• Crashes / signals / nonzero exits → auto-restart (unchanged).

Implementation hint (not prescriptive)

Exit status alone can't discriminate the two gestures. One candidate: the shellper/PTY layer sees the raw input stream — recent ^C (0x03) bytes in the input immediately preceding the exit could mark the exit as Ctrl-C-driven and therefore restartable. Another: check whether the Claude CLI distinguishes these cases in any observable way (exit code, stderr, tty state) before inventing input-sniffing. Verify against the real CLI, not assumptions.

Notes

Regression is live in published v3.2.4 — fix likely warrants a fast-follow patch release.

[waleedkadous]

On it! Working on a fix now.

[waleedkadous]

Spec correction from the architect (human decision, 2026-07-27)

The first framing of this issue was wrong. The problem is not discriminating ^C^C from a typed quit — it's that a clean harness exit kills the shellper/session at all.

Corrected expected behavior:

• Any clean exit of the harness (double Ctrl-C, /quit, exit — no discrimination needed): the shellper and terminal session survive, and the harness (e.g. claude) is rerun in the same PTY with recovery disabled — no --resume, fresh conversation.
• Unnatural exits (crash, signal, nonzero code): keep the existing restart-with-recovery behavior (--resume of the persisted session id) — that is what recovery is for.
• Truly ending a session happens only via explicit kill (afx / UI / DELETE /api/terminals/:id), never via an in-terminal gesture.

Consequence for PR #1266: the ^C-byte detection machinery is no longer needed and should be dropped; the fix becomes 'clean exit → fresh rerun without recovery args'. Simpler than what was built.