Isolate analytics recorder failures

docs/ANALYTICS.md

@@ -8,7 +8,13 @@ Server-side recording of session volume, conversation content, tool usage, and f
 
 ## Why `better-sqlite3`, not `node:sqlite`
 
-Node has an experimental built-in SQLite module, but `better-sqlite3` was chosen instead: it's the mature, widely-used library with a stable synchronous API, which is what keeps every `AnalyticsRecorder` call in `session.ts` a plain fire-and-forget call rather than something that needs `await`/error-handling threaded through the session's existing synchronous methods.
+Node has an experimental built-in SQLite module, but `better-sqlite3` was chosen instead: it's the mature, widely-used library with a stable synchronous API, which keeps analytics writes compatible with the session's existing synchronous lifecycle methods.
+
+## Failure isolation
+
+Analytics is best-effort telemetry, not part of session correctness. Every `AnalyticsRecorder` call in `session.ts` runs through one non-throwing boundary: a full disk, locked database, or other recorder failure is reported through the server logger under the `analytics` category, then normal provider traffic, browser delivery, and teardown continue. Failure reporting is itself guarded so a broken logger cannot turn a telemetry failure back into a session failure.
+
+This boundary covers writes after the recorder has been constructed. Recorder construction still happens eagerly at process startup; an invalid/unwritable database path is a configuration or boot failure rather than a mid-session degradation case.
 
 ## The transcript-buffering gotcha
 

src/lib/server/core/session.spec.ts

@@ -112,6 +112,38 @@ class FakeAnalyticsRecorder implements AnalyticsRecorder {
 	}
 }
 
+type AnalyticsMethod =
+	'sessionStarted' | 'sessionEnded' | 'messageLogged' | 'toolCallLogged' | 'eventLogged';
+
+class ThrowingAnalyticsRecorder implements AnalyticsRecorder {
+	readonly calls: AnalyticsMethod[] = [];
+
+	private fail(method: AnalyticsMethod): never {
+		this.calls.push(method);
+		throw new Error(`Simulated ${method} storage failure.`);
+	}
+
+	sessionStarted(): void {
+		this.fail('sessionStarted');
+	}
+
+	sessionEnded(): void {
+		this.fail('sessionEnded');
+	}
+
+	messageLogged(): void {
+		this.fail('messageLogged');
+	}
+
+	toolCallLogged(): void {
+		this.fail('toolCallLogged');
+	}
+
+	eventLogged(): void {
+		this.fail('eventLogged');
+	}
+}
+
 function makeSocket() {
 	const socket = new FakeBrowserSocket();
 	return { socket, asWs: socket as unknown as WebSocket };
@@ -252,12 +284,103 @@ describe('RealtimeSession', () => {
 		recorder.close();
 	});
 
-	it('contains unexpected async message failures and leaves other sessions usable', async () => {
+	it('degrades gracefully when every analytics write fails', async () => {
+		const errorSpy = vi.spyOn(console, 'error').mockImplementation(() => {});
+		const provider = new FakeProvider();
+		const recorder = new ThrowingAnalyticsRecorder();
+		const { socket, asWs } = makeSocket();
+		const orchestrator = createRealtimeSession(
+			asWs,
+			provider,
+			DEFAULTS,
+			recorder,
+			TEST_TOOL_TIMEOUT_MS
+		);
+
+		socket.receive({ type: 'session.start' });
+		await vi.waitFor(() => expect(provider.sessions).toHaveLength(1));
+		await vi.waitFor(() =>
+			expect(socket.sent).toContainEqual(expect.objectContaining({ type: 'session.ready' }))
+		);
+
+		socket.receive({ type: 'input_text.send' });
+		await vi.waitFor(() =>
+			expect(socket.sent).toContainEqual(
+				expect.objectContaining({ type: 'error', code: 'invalid_client_event' })
+			)
+		);
+
+		socket.receive({ type: 'input_text.send', text: 'still works' });
+		await vi.waitFor(() =>
+			expect(provider.sessions[0].sent).toContainEqual({
+				type: 'input_text.send',
+				text: 'still works'
+			})
+		);
+
+		provider.sessions[0].emitToOrchestrator({
+			type: 'transcript.delta',
+			role: 'user',
+			text: 'voice input'
+		});
+		provider.sessions[0].emitToOrchestrator({
+			type: 'transcript.delta',
+			role: 'assistant',
+			text: 'voice output'
+		});
+		provider.sessions[0].emitToOrchestrator({ type: 'transcript.done', role: 'assistant' });
+		expect(socket.sent).toContainEqual({
+			type: 'transcript.delta',
+			role: 'user',
+			text: 'voice input'
+		});
+		expect(socket.sent).toContainEqual({ type: 'transcript.done', role: 'assistant' });
+
+		provider.sessions[0].emitToOrchestrator({
+			type: 'tool.call',
+			callId: 'call-1',
+			name: 'nonexistent_tool',
+			arguments: '{}'
+		});
+		await vi.waitFor(() =>
+			expect(provider.sessions[0].sent).toContainEqual({ type: 'response.trigger' })
+		);
+
+		provider.sessions[0].emitToOrchestrator({
+			type: 'session.closed',
+			reason: 'upstream_disconnected'
+		});
+		expect(socket.sent).toContainEqual({
+			type: 'session.closed',
+			reason: 'upstream_disconnected'
+		});
+		expect(orchestrator.getState()).toBe('closed');
+		expect(provider.sessions[0].closed).toBe(true);
+		expect(socket.closeCalls).toBe(1);
+		expect(socket.sent).not.toContainEqual(
+			expect.objectContaining({ type: 'error', code: 'internal_error' })
+		);
+		expect(recorder.calls).toEqual([
+			'sessionStarted',
+			'eventLogged',
+			'messageLogged',
+			'messageLogged',
+			'messageLogged',
+			'toolCallLogged',
+			'eventLogged',
+			'sessionEnded'
+		]);
+		expect(errorSpy).toHaveBeenCalledTimes(recorder.calls.length);
+		expect(errorSpy).toHaveBeenCalledWith(
+			'[analytics] Failed to record session start.',
+			expect.objectContaining({ message: expect.stringContaining('storage failure') })
+		);
+		errorSpy.mockRestore();
+	});
+
+	it('contains non-analytics async message failures and leaves other sessions usable', async () => {
 		const provider = new FakeProvider();
 		const recorder = new FakeAnalyticsRecorder();
-		recorder.messageLogged = () => {
-			throw new Error('analytics unavailable');
-		};
 		const affected = makeSocket();
 		const other = makeSocket();
 		createRealtimeSession(affected.asWs, provider, DEFAULTS, recorder, TEST_TOOL_TIMEOUT_MS);
@@ -267,8 +390,11 @@ describe('RealtimeSession', () => {
 		await vi.waitFor(() => expect(provider.sessions).toHaveLength(1));
 		other.socket.receive({ type: 'session.start' });
 		await vi.waitFor(() => expect(provider.sessions).toHaveLength(2));
-		affected.socket.receive({ type: 'input_text.send', text: 'hello' });
+		provider.sessions[0].send = () => {
+			throw new Error('provider send failed');
+		};
 
+		affected.socket.receive({ type: 'response.cancel' });
 		await vi.waitFor(() =>
 			expect(affected.socket.sent).toContainEqual({
 				type: 'error',

src/lib/server/core/session.ts

@@ -164,11 +164,13 @@ export class RealtimeSession {
 			// Typed text is always already a complete string — unlike the voice
 			// path, no buffering is needed.
 			this.logger.info('messages', 'User message', { role: 'user', text: event.text });
-			this.recorder.messageLogged({
+			this.recordAnalytics('message', () => {
-				sessionId: this.sessionId!,
+				this.recorder.messageLogged({
-				role: 'user',
+					sessionId: this.sessionId!,
-				text: event.text,
+					role: 'user',
-				loggedAt: Date.now()
+					text: event.text,
+					loggedAt: Date.now()
+				});
 			});
 		}
 		this.providerSession.send(event);
@@ -259,11 +261,13 @@ export class RealtimeSession {
 			role: 'system',
 			text: config.instructions
 		});
-		this.recorder.sessionStarted({
+		this.recordAnalytics('session start', () => {
-			sessionId,
+			this.recorder.sessionStarted({
-			model: config.model,
+				sessionId,
-			voice: config.voice,
+				model: config.model,
-			startedAt: Date.now()
+				voice: config.voice,
+				startedAt: Date.now()
+			});
 		});
 		// Recorded above before wiring onEvent, so there's no window where a
 		// provider event could reach handleProviderEvent before sessionId is set.
@@ -287,11 +291,13 @@ export class RealtimeSession {
 				// immediately rather than buffering. Rests on unverified live
 				// behavior; see docs/ANALYTICS.md.
 				this.logger.info('messages', 'User message', { role: 'user', text: event.text });
-				this.recorder.messageLogged({
+				this.recordAnalytics('message', () => {
-					sessionId: this.sessionId!,
+					this.recorder.messageLogged({
-					role: 'user',
+						sessionId: this.sessionId!,
-					text: event.text,
+						role: 'user',
-					loggedAt: Date.now()
+						text: event.text,
+						loggedAt: Date.now()
+					});
 				});
 			}
 			this.sendToBrowser(event);
@@ -301,11 +307,13 @@ export class RealtimeSession {
 			const text = this.assistantTranscriptBuffer;
 			this.assistantTranscriptBuffer = '';
 			this.logger.info('messages', 'Assistant message', { role: 'assistant', text });
-			this.recorder.messageLogged({
+			this.recordAnalytics('message', () => {
-				sessionId: this.sessionId!,
+				this.recorder.messageLogged({
-				role: 'assistant',
+					sessionId: this.sessionId!,
-				text,
+					role: 'assistant',
-				loggedAt: Date.now()
+					text,
+					loggedAt: Date.now()
+				});
 			});
 			this.sendToBrowser(event);
 			return;
@@ -391,14 +399,16 @@ export class RealtimeSession {
 				callId,
 				durationMs: Date.now() - meta.startedAt
 			});
-			this.recorder.toolCallLogged({
+			this.recordAnalytics('tool call', () => {
-				sessionId: this.sessionId!,
+				this.recorder.toolCallLogged({
-				callId,
+					sessionId: this.sessionId!,
-				name: meta.name,
+					callId,
-				arguments: meta.arguments,
+					name: meta.name,
-				output,
+					arguments: meta.arguments,
-				durationMs: Date.now() - meta.startedAt,
+					output,
-				loggedAt: Date.now()
+					durationMs: Date.now() - meta.startedAt,
+					loggedAt: Date.now()
+				});
 			});
 		}
 		this.outstandingToolCalls--;
@@ -423,7 +433,13 @@ export class RealtimeSession {
 			// Guarded: sessionId is still null if the browser disconnected
 			// mid-handshake, before start() ever reached 'active' — no sessions
 			// row exists yet to update in that case.
-			this.recorder.sessionEnded({ sessionId: this.sessionId, endedAt: Date.now(), reason });
+			this.recordAnalytics('session end', () => {
+				this.recorder.sessionEnded({
+					sessionId: this.sessionId!,
+					endedAt: Date.now(),
+					reason
+				});
+			});
 		}
 	}
 
@@ -432,17 +448,42 @@ export class RealtimeSession {
 			// Logged before the delivery guard below: the event "happened" from
 			// the session's point of view even if the browser socket is already
 			// dead by the time we try to deliver it.
-			this.recorder.eventLogged({
+			this.recordAnalytics('event', () => {
-				sessionId: this.sessionId,
+				this.recorder.eventLogged({
-				type: event.type,
+					sessionId: this.sessionId,
-				detail: JSON.stringify(event),
+					type: event.type,
-				loggedAt: Date.now()
+					detail: JSON.stringify(event),
+					loggedAt: Date.now()
+				});
 			});
 		}
 		if (this.browserSocket.readyState !== WebSocket.OPEN) return;
 		this.browserSocket.send(JSON.stringify(event));
 	}
 
+	/**
+	 * Analytics is best-effort telemetry, never part of session correctness.
+	 * A full disk, locked database, or broken recorder must not interrupt
+	 * provider traffic, browser delivery, or teardown.
+	 */
+	private recordAnalytics(operation: string, write: () => void): void {
+		try {
+			write();
+		} catch (error) {
+			let message = 'Unknown error';
+			try {
+				message = error instanceof Error ? error.message : String(error);
+			} catch {
+				// Even a hostile thrown value with a broken toString stays contained.
+			}
+			try {
+				this.logger.error('analytics', `Failed to record ${operation}.`, { message });
+			} catch {
+				// Telemetry failure reporting is best-effort too.
+			}
+		}
+	}
+
 	/**
 	 * Verifies a `debug.subscribe` request and replies with the result. In
 	 * dev, any request succeeds (frictionless — see docs/LOGGING.md); in