fix(claude): actually stop the resume crash-loop ("No conversation found")

The earlier restarts<=1 guard never triggered: the bootstrap 'ready' ack fires
before the --resume failure, so it reset the restart budget every cycle → the
loop ran forever (276 restarts in the wild).

Now the adapter detects "No conversation found with session ID" on stderr and
emits a resume_failed event; ChatState sets a one-shot resumeBroken flag, and the
next restart comes up fresh (no --resume) instead of re-resuming the same doomed
id. The fresh session gets a new id that CAN be resumed on a later crash. The
desktop transcript is preserved throughout.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Author: gaoyu06

src/lib/backends/claude.test.ts

@@ -1143,5 +1143,9 @@ describe('claude adapter: restarts and robustness', () => {
 		// Tracing-formatted engine log lines are dropped as noise.
 		expect(adapter.translate({ __stderr: '2026-07-14T10:00:00.000Z DEBUG hitting cache' })).toEqual([]);
 		expect(adapter.translate({ __stderr: '2026-07-14T10:00:00Z ERROR boom' })).toEqual([]);
+		// A failed --resume signals resume_failed (breaks the crash-restart loop).
+		expect(
+			adapter.translate({ __stderr: 'No conversation found with session ID: e4f0405e-3919-4c10' })
+		).toEqual([{ type: 'resume_failed' }]);
 	});
 });

src/lib/backends/claude.ts

@@ -938,6 +938,11 @@ export function createClaudeAdapter(): EngineAdapter {
 				const line = raw.__stderr.replace(/\[[0-9;]*m/g, '').trim();
 				if (!line) return [];
 				if (/^\d{4}-\d{2}-\d{2}T\S+\s+(ERROR|WARN|INFO|DEBUG|TRACE)\b/.test(line)) return [];
+				// A --resume target the CLI can't find (session file gone / never
+				// persisted) makes it exit immediately. Signal it so the store stops
+				// re-resuming the same doomed id in an endless crash-restart loop
+				// (the bootstrap 'ready' otherwise keeps resetting the restart budget).
+				if (/No conversation found with session ID/i.test(line)) return [{ type: 'resume_failed' }];
 				return [{ type: 'info', message: `[claude] ${line}` }];
 			}
 			const msg = rec(raw);

src/lib/chat.svelte.ts

@@ -167,6 +167,10 @@ export class ChatState {
 	// Engine crash auto-restart bookkeeping (driven by the page on agent-exit).
 	restarts = 0;
 	restartWindowStart: number | null = null;
+	// Set when a claude --resume target isn't found: the next restart must NOT
+	// resume the same doomed id (it would crash-loop). One-shot — consumed by the
+	// store's restartSession, which then comes up fresh.
+	resumeBroken = false;
 	// Set while an intentional provider-switch restart is in flight, so the exit it
 	// causes isn't treated as a crash to auto-restart.
 	switching = false;
@@ -593,6 +597,11 @@ export class ChatState {
 			case 'retrying':
 				this.messages.push({ kind: 'system', text: `reconnecting… (attempt ${num(ev.attempt)})` });
 				break;
+			case 'resume_failed':
+				// The engine couldn't resume the session id — restart fresh instead of
+				// looping on the same doomed --resume.
+				this.resumeBroken = true;
+				break;
 			case 'compaction_progress':
 				this.compactionTokens = num(ev.output_tokens);
 				break;

src/lib/session.svelte.ts

@@ -217,11 +217,13 @@ export class SessionStore {
 		// claude resumes via the --resume spawn option (no /resume command in
 		// stream-json mode); codex resumes via the thread/resume RPC (thread id
 		// through SessionCtx); jucode resumes with the command after the handshake.
-		// A resume target that the engine can't find ("No conversation found …")
-		// makes it exit immediately — which would crash-loop forever re-resuming the
-		// same doomed id. So only resume on the first restart; if that also crashes,
-		// come back up fresh (the desktop transcript is preserved regardless).
-		const mayResume = sid && canResume && s.chat.restarts <= 1;
+		// A resume target the engine can't find ("No conversation found …") makes it
+		// exit immediately, which would crash-loop forever re-resuming the same
+		// doomed id (the bootstrap 'ready' keeps resetting the restart budget). When
+		// the adapter flagged a resume failure, come up fresh instead. One-shot: the
+		// fresh session gets a new id that CAN be resumed on a later crash.
+		const mayResume = sid && canResume && !s.chat.resumeBroken;
+		s.chat.resumeBroken = false;
 		const resumeViaSpawn = s.backendId === 'claude' && mayResume;
 		const resumeViaCtx = s.backendId === 'codex' && mayResume;
 		this.#spawn(

src/lib/session.test.ts

@@ -76,6 +76,23 @@ describe('SessionStore lifecycle', () => {
 		expect(snap[0].tabs).toEqual([{ sid: 'sid-0', title: 'kept', archived: true }]);
 	});
 
+	it('a flagged resume failure makes the next claude restart come up fresh', () => {
+		const store = new SessionStore();
+		const p = proj();
+		store.projects.push(p);
+		const id = store.addSession(p, undefined, 'claude');
+		const s = p.sessions.find((x) => x.id === id)!;
+		s.chat.sessionId = 'sid-x';
+		s.chat.messages.push({ kind: 'user', text: 'hi' }); // resumable
+		s.chat.resumeBroken = true;
+		vi.clearAllMocks();
+		store.restartSession(id);
+		// Spawned without a resume option, and the one-shot flag is consumed.
+		const call = (createSession as unknown as { mock: { calls: unknown[][] } }).mock.calls.at(-1);
+		expect((call?.[3] as { resume?: string } | undefined)?.resume).toBeUndefined();
+		expect(s.chat.resumeBroken).toBe(false);
+	});
+
 	it('removeProject tears down its sessions and clears a dangling activeId', () => {
 		const store = new SessionStore();
 		const p = proj();