Review

Nice, focused fix — I traced the full flow and the core logic is correct and, importantly, loop-safe: restartAgent(id, false) sets resumed = false, so if the fresh start also exits non-zero the a().resumed guard is false and it falls through to markAgentExited rather than restarting again. Going straight from running → running (skipping the "exited" state) is the right call for a seamless recovery, and centralizing the signatures in RESUME_FAILURE_PATTERNS is a clean, extensible approach.

A few things worth addressing:

Important

• Add a unit test for isResumeArgsFailure. It's a pure function and src/lib/agent-args.test.ts already exists right beside it. Suggested cases: claude match, claude no-match, non-claude command (→ false), empty lastOutput (→ false), and a multi-line lastOutput where the pattern is on one of several lines.

Minor

• Manual "Resume" button now silently falls back too. The new onExit logic also fires when the user explicitly clicks Resume and it fails — the agent restarts fresh with no indication the resume failed. For the app-restart auto-resume case that silence is fine; just worth a conscious decision on whether an explicit Resume click should surface a brief notice.
• English-string matching is inherently fragile. No conversation found to continue will break if Claude rewrites/localizes the message, and the raw last_output carries ANSI escapes (only \r is stripped in pty.ts). The substring match still works as long as the codes bracket the phrase rather than split it — low risk for a short non-wrapping string — but it's a signature that'll need maintenance. Centralizing it makes that easy, so this is just a flag.
• code.exit_code !== 0 also matches null (signal kill). Harmless because isResumeArgsFailure must also match, but noting it.

Verdict: Ready to merge once the isResumeArgsFailure test is added.

Thank you very much!