agentHost: re-enable retryable errors (#333260)

* Revert "agentHost: temporarily disable retryable errors"

Reverts #332839 now that the retry issue is fixed.

(Written by Copilot)

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* agentHost: preserve non-resumable error handling

Keep subagent errors non-resumable when their parent mapping is missing, and retain dispatcher coverage for rejecting non-resumable retry requests.

(Written by Copilot)

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

---------

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Author: roblourens

src/vs/platform/agentHost/node/agentSideEffects.ts

@@ -1593,7 +1593,7 @@ export class AgentSideEffects extends Disposable {
 						type: ActionType.ChatError,
 						turnId: action.turnId,
 						duration: execution.duration + execution.stopWatch.elapsed(),
-						part: createErrorResponsePart(failure.error),
+						part: createErrorResponsePart(failure.error, true),
 					});
 					this._completeTurn(channel, action.turnId, 'error', failure);
 					this._toolCallTracker.clearSession(channel);

src/vs/platform/agentHost/node/copilot/copilotAgentSession.ts

@@ -1594,7 +1594,7 @@ export class CopilotAgentSession extends Disposable {
 			type: ActionType.ChatError,
 			turnId: turn.id,
 			duration: turn.duration,
-			part: createErrorResponsePart(error),
+			part: createErrorResponsePart(error, true),
 		});
 		this._clearActiveTurn();
 		return turn.id;
@@ -2644,7 +2644,7 @@ export class CopilotAgentSession extends Disposable {
 			part: createErrorResponsePart({
 				errorType: 'developmentRecoverableError',
 				message: localize('copilotAgent.developmentRecoverableError', "Injected recoverable development error ({0}/{1}).", attempt, totalFailures),
-			}),
+			}, true),
 		});
 		this._clearActiveTurn();
 	}
@@ -5003,7 +5003,7 @@ export class CopilotAgentSession extends Disposable {
 				type: ActionType.ChatError,
 				turnId: this._turnId,
 				duration: turn?.duration ?? 0,
-				part: createErrorResponsePart(buildChatErrorInfoFromCopilotSdkFields(e.data)),
+				part: createErrorResponsePart(buildChatErrorInfoFromCopilotSdkFields(e.data), !e.agentId && turn !== undefined),
 			}, parentToolCallId);
 			if (!parentToolCallId) {
 				this._clearActiveTurn();

src/vs/platform/agentHost/node/copilot/mapSessionEvents.ts

@@ -672,7 +672,7 @@ export async function mapSessionEvents(
 					rootRequestActive = false;
 					parentTurnState = TurnState.Error;
 					parentTurnTerminated = true;
-					parentBuilder.responseParts.push(createErrorResponsePart(buildChatErrorInfoFromCopilotSdkFields(e.data)));
+					parentBuilder.responseParts.push(createErrorResponsePart(buildChatErrorInfoFromCopilotSdkFields(e.data), true));
 					parentBuilder.waitingStartedAt = currentEventTimestamp;
 					touch(parentBuilder);
 				}
@@ -784,7 +784,7 @@ export async function mapSessionEvents(
 	}
 
 	if (options && !(options instanceof URI) && options.interruptedTurnError && parentBuilder && rootRequestActive && parentTurnState !== TurnState.Error) {
-		parentBuilder.responseParts.push(createErrorResponsePart(options.interruptedTurnError));
+		parentBuilder.responseParts.push(createErrorResponsePart(options.interruptedTurnError, true));
 		parentTurnState = TurnState.Error;
 	}
 	flushParent();

src/vs/platform/agentHost/test/node/agentService.test.ts

@@ -791,7 +791,7 @@ suite('AgentService (node dispatcher)', () => {
 				});
 			});
 
-			test('finalizes the same turn with a non-resumable error when continuation fails immediately', async () => {
+			test('finalizes the same turn with another resumable error when continuation fails immediately', async () => {
 				const { chat } = await createErroredTurn();
 				copilotAgent.chats.resumeTurn = async () => {
 					throw new Error('continuation failed');
@@ -814,7 +814,7 @@ suite('AgentService (node dispatcher)', () => {
 					state: TurnState.Error,
 					errors: [
 						createErrorResponsePart({ errorType: 'requestFailed', message: 'failed' }, true),
-						createErrorResponsePart({ errorType: 'sendFailed', message: 'Error: continuation failed' }),
+						createErrorResponsePart({ errorType: 'sendFailed', message: 'Error: continuation failed' }, true),
 					],
 					durationAtLeastInitial: true,
 				});

src/vs/platform/agentHost/test/node/copilotAgentSession.test.ts

@@ -1464,6 +1464,7 @@ suite('CopilotAgentSession', () => {
 				errorType: 'executionInterrupted',
 				message: 'The agent was interrupted before this request finished.',
 			},
+			resumable: true,
 		});
 	});
 
@@ -5704,7 +5705,7 @@ Use the attached image as context.
 			});
 		});
 
-		test('the development $error-ui path emits an error even with attachments', async () => {
+		test('the development $error-ui path emits a resumable error even with attachments', async () => {
 			const { session, mockSession, signals } = await createAgentSession(disposables);
 
 			await session.send('$error-ui', [{
@@ -5730,6 +5731,7 @@ Use the attached image as context.
 							errorType: 'developmentRecoverableError',
 							message: 'Injected recoverable development error (1/1).',
 						},
+						resumable: true,
 					},
 				}],
 			});

src/vs/platform/agentHost/test/node/e2e/KNOWN_ISSUES.md

@@ -776,25 +776,18 @@ Use the affected provider command with `--grep "<exact test title>"` and tempora
     --grep "accepted steering followed by abort"
   ```
 
-### Retryable Copilot errors are temporarily disabled
+### Mid-turn host shutdown recovery is record-only
 
-Copilot errors currently end a turn without offering an in-place retry. The retry protocol remains implemented, but the host intentionally omits the `resumable` marker from live, restored, and repeated errors until the feature is re-enabled.
+A user can lose the Agent Host process while a model response is still streaming. Reopening the session should restore the unfinished request as a resumable error, and retrying should continue that same turn without adding another user message.
 
-- Tests:
-  - `resumes a failed turn in place`
-  - `resumes the same turn after repeated failures`
-  - `restores and resumes a turn interrupted by host shutdown`
-- Scope: Copilot on all platforms and execution modes.
-- Expected when enabled: a failed turn is marked resumable, and retrying continues the same turn without adding another user message. An unfinished request restored after host shutdown has the same behavior.
-- Observed: Copilot errors intentionally omit the `resumable` marker, so clients cannot request an in-place retry.
-- Gate: all three scenarios are unconditionally skipped. The host-shutdown scenario additionally requires direct `AGENT_HOST_REPLAY_RECORD=1` mode because replay has no active streaming window to terminate.
-- Run after removing the temporary gate:
+- Test: `restores and resumes a turn interrupted by host shutdown`.
+- Scope: deterministic replay for Copilot.
+- Expected: the host dies after streaming starts but before any terminal turn action; restoration synthesizes a resumable `executionInterrupted` error, and a zero-message continuation completes the same turn.
+- Observed: replay serves the full recorded response immediately, leaving no active streaming window in which to kill the host before turn completion.
+- Gate: direct `AGENT_HOST_REPLAY_RECORD=1` mode only.
+- Run:
 
   ```bash
-  ./scripts/test-integration.sh --run \
-    src/vs/platform/agentHost/test/node/e2e/providers/copilotAgentHostE2E.integrationTest.ts \
-    --grep "resumes a failed turn in place|resumes the same turn after repeated failures"
-
   AGENT_HOST_REPLAY_RECORD=1 ./scripts/test-integration.sh --run \
     src/vs/platform/agentHost/test/node/e2e/providers/copilotAgentHostE2E.integrationTest.ts \
     --grep "restores and resumes a turn interrupted by host shutdown"

src/vs/platform/agentHost/test/node/e2e/providers/copilotAgentHostE2E.integrationTest.ts

@@ -181,7 +181,7 @@ suite('Agent Host E2E — Copilot (Copilot-specific)', function () {
 			90_000,
 		);
 		const liveErrorPart = (getActionEnvelope(liveNotification).action as ChatErrorAction).part;
-		assert.strictEqual(liveErrorPart.resumable, undefined);
+		assert.strictEqual(liveErrorPart.resumable, true);
 
 		client = await lease.restart();
 		client.setWorkingDirectory(workingDirectory);
@@ -201,12 +201,11 @@ suite('Agent Host E2E — Copilot (Copilot-specific)', function () {
 		}, {
 			state: TurnState.Error,
 			error: liveErrorPart.error,
-			resumable: undefined,
+			resumable: true,
 		});
 	});
 
-	// Retryable errors are temporarily disabled.
-	test.skip('resumes a failed turn in place', async function () {
+	test('resumes a failed turn in place', async function () {
 		this.timeout(180_000);
 		const workingDirectory = await mkdtemp(join(tmpdir(), 'copilot-failed-turn-resume-'));
 		tempDirs.push(workingDirectory);
@@ -345,7 +344,7 @@ suite('Agent Host E2E — Copilot (Copilot-specific)', function () {
 		}
 	});
 
-	test.skip('resumes the same turn after repeated failures', async function () {
+	test('resumes the same turn after repeated failures', async function () {
 		this.timeout(180_000);
 		const workingDirectory = await mkdtemp(join(tmpdir(), 'copilot-repeated-failed-turn-resume-'));
 		tempDirs.push(workingDirectory);
@@ -426,8 +425,8 @@ suite('Agent Host E2E — Copilot (Copilot-specific)', function () {
 		});
 	});
 
-	// Retryable errors are temporarily disabled.
-	test.skip('restores and resumes a turn interrupted by host shutdown', async function () {
+	// Replay serves the full recorded response immediately, so it has no active streaming window to terminate.
+	(RECORD_ONLY ? test : test.skip)('restores and resumes a turn interrupted by host shutdown', async function () {
 		this.timeout(240_000);
 		const workingDirectory = await mkdtemp(join(tmpdir(), 'copilot-host-shutdown-resume-'));
 		tempDirs.push(workingDirectory);

src/vs/platform/agentHost/test/node/mapSessionEvents.test.ts

@@ -146,7 +146,7 @@ suite('mapSessionEvents — history replay', () => {
 		});
 	});
 
-	test('restores an unfinished request as an error on the same turn', async () => {
+	test('restores an unfinished request as a resumable error on the same turn', async () => {
 		const events: ISessionEvent[] = [
 			{ type: 'user.message', id: 'interrupted-turn', data: { interactionId: 'm1', content: 'Keep working' } },
 			{ type: 'assistant.turn_start', data: { turnId: 'sdk-turn' } },
@@ -171,6 +171,7 @@ suite('mapSessionEvents — history replay', () => {
 			errorPart: {
 				kind: ResponsePartKind.Error,
 				error: interruptedTurnError,
+				resumable: true,
 			},
 		});
 	});
@@ -232,11 +233,11 @@ suite('mapSessionEvents — history replay', () => {
 				{ kind: ResponsePartKind.Markdown, content: 'Second segment' },
 				{ kind: ResponsePartKind.Error },
 			],
-			resumable: undefined,
+			resumable: true,
 		});
 	});
 
-	test('keeps an error terminal when a later notification starts another turn', async () => {
+	test('keeps a resumable error terminal when a later notification starts another turn', async () => {
 		const events: ISessionEvent[] = [
 			{ type: 'user.message', id: 'failed-turn', timestamp: '2026-08-11T00:00:00.000Z', data: { interactionId: 'm1', content: 'Start the background agent' } },
 			{ type: 'assistant.turn_start', timestamp: '2026-08-11T00:00:00.100Z', data: { turnId: 'sdk-turn-1' } },
@@ -273,10 +274,10 @@ suite('mapSessionEvents — history replay', () => {
 			state: TurnState.Complete,
 			parts: [{ kind: ResponsePartKind.Markdown, content: 'The background agent finished.' }],
 		}]);
-		assert.strictEqual(getErrorResponsePart(turns[0])?.resumable, undefined);
+		assert.strictEqual(getErrorResponsePart(turns[0])?.resumable, true);
 	});
 
-	test('keeps an error as the final part when a late tool completion arrives', async () => {
+	test('keeps a resumable error as the final part when a late tool completion arrives', async () => {
 		const events: ISessionEvent[] = [
 			{ type: 'user.message', id: 'failed-turn', data: { interactionId: 'm1', content: 'Run a command' } },
 			{ type: 'assistant.turn_start', data: { turnId: 'sdk-turn-1' } },
@@ -294,7 +295,7 @@ suite('mapSessionEvents — history replay', () => {
 		}, {
 			state: TurnState.Error,
 			parts: [{ kind: ResponsePartKind.Error }],
-			resumable: undefined,
+			resumable: true,
 		});
 	});