feat(cursor): mid-turn Steer via concurrent ACP session/prompt (#888) (#1609)

* feat(shared): steer capability gates and live steered signal schemas

- STEERING_SUPPORTED_FLAVORS / isSteeringSupportedForSession gate which
  agents can deliver queued messages into the active turn (pi, codex,
  cursor ACP; legacy stream-json cursor excluded)
- AgentState.steeringActive, DecryptedMessage.steered and
  messages-consumed  live signal (never persisted by the hub)

* feat(cli): queue reservations and steered messages-consumed option

- MessageQueue2 gains takeByLocalId/restoreReservation/
  beginReservationDispatch/commitReservation so an async steer can reserve
  a queued row without racing the main loop's turn/start drain
- emitMessagesConsumed accepts steered: true to mark mid-turn delivery

* feat(codex): mid-turn steer via app-server turn/steer (#888)

- CodexAppServerClient.steerTurn + TurnSteerParams/Response types
- CodexRemoteLauncher registers the steer-queued-message RPC handler:
  reserves the queued row, validates it against the active turn (no
  control commands, matching mode hash), injects via turn/steer with an
  epoch guard that invalidates in-flight steers on abort/cleanup
- steeringActive agent state tracks the active-turn window
- hub syncEngine gate opens to codex; messages-consumed relays steered

* feat(web): Steered badge and steer gating for codex sessions

- HappyUserMessage shows a ↳ Steered badge fed by the live
  messages-consumed steered signal, preserved across server echoes and
  refetches (mergeMessages carries the optimistic marker)
- SessionChat gates canSteer via isSteeringSupportedForSession instead of
  the pi-only check
- clearStaleQueuedStatus normalizes a queued status on an invoked message
- fix(web): drop duplicate showSessionSummaryInChat in markdown test
  (upstream typecheck breakage)

* feat(acp): split request dispatch from completion and add soft steer

- AcpStdioTransport.sendRequestWithDispatch separates stdin-accepted
  dispatch from the JSON-RPC response, keeping sendRequest behavior
  unchanged
- AcpSdkBackend tracks concurrent session/prompt requests with an
  activePromptRequests counter (main prompt + soft steers); response
  completion stays pending until every concurrent prompt settles
- beginSoftSteerPrompt kicks off a concurrent session/prompt (Cursor GUI
  Send semantics — no cancel, no handler swap) returning {dispatched,
  completed}; softSteerPrompt awaits the full response for direct callers

* feat(cursor): mid-turn soft steer via concurrent session/prompt (#888)

- CursorAcpRemoteLauncher registers the steer-queued-message RPC handler:
  reserves the queued row, rejects control commands and mode mismatches,
  then soft-injects via beginSoftSteerPrompt without canceling the
  in-flight turn
- Acks the hub once stdin accepts the inject (not on turn completion) to
  stay inside the 30s RPC window; the launcher stays busy until the
  concurrent prompt settles so handlers are not swapped mid-inject
- steeringActive agent state mirrors the active-turn window; abort and
  cleanup reset it and invalidate pending steers
- Legacy stream-json Cursor sessions register a steer handler that
  reports unsupported

* fix(codex,shared): address bot findings on steer gate and ambiguous turn/steer

- STEERING_SUPPORTED_FLAVORS / isSteeringSupportedForSession advertise
  codex and pi only; cursor joins when its soft-steer handler lands (#1609)
- turn/steer now splits dispatch (stdin accepted) from completion (turn
  finished): the hub RPC acks once dispatch succeeds — never on the
  concurrent turn's completion, which can exceed the 30s RPC window
- queue row commits only after the turn settles; a rejected/aborted steer
  restores the row so the message still delivers via turn/start, and a
  dispatched steer is never restored (no duplicate delivery)
- steer carries clientUserMessageId (echoed as userMessage.clientId) so
  ambiguous transport failures can reconcile the thread later
- client tests cover dispatch/complete split and stdin-write failure

* feat(shared): advertise cursor in the steer gate now that its handler lands

Cursor ACP sessions pass the web and hub steer gates; legacy stream-json
cursor sessions stay excluded.

* fix(codex): reconcile dispatched steers before restoring; align error copy

- A dispatched turn/steer whose completion fails (disconnect / protocol
  error) is now reconciled via thread/read by clientUserMessageId before
  the queued row is restored — the instruction is only re-delivered by
  turn/start when the thread never received it
- Reconcile targets the pinned steer thread, not whichever turn is
  current when completion fails
- syncEngine unsupported-flavor error now matches the capability gate
  (Pi and Codex only until the cursor handler lands)
- launcher tests cover steer success (ack on dispatch), reconcile-accepted
  and reconcile-rejected outcomes

* fix(codex): consume the row at dispatch; drop background reconcile

- The hub RPC acks and the queue row is consumed as soon as stdin accepts
  turn/steer; completion is background-only logging. A dispatched steer is
  never restored, so the same localId cannot be re-delivered via turn/start
  after the caller was told the steer succeeded
- Dispatch failure (stdin write error) still restores the row and reports
  failure
- steer.completed rejection is always handled (no unhandled rejection on
  the dispatch-failure path)
- tests updated: completion failure after dispatch keeps the row consumed;
  dispatch failure restores it

* fix(cursor): consume the row at dispatch; keep waiters for prompt gating

- The hub RPC acks and the queue row is consumed as soon as stdin accepts
  the concurrent session/prompt; completion is background-only. A
  dispatched steer is never restored (no duplicate via the next prompt)
- softSteerWaiters are registered before awaiting dispatch so the main
  loop's finally cannot start the next prompt mid-inject; they still gate
  prompt handover on completion
- tests updated: post-dispatch ACP rejection keeps the row consumed

* fix(cursor,hub): never hang teardown on unresolved soft steer; align diagnostics

- Prompt-finally waits for soft-steer completion only when not exiting;
  the outer finally no longer waits at all — cleanup() disconnects the
  ACP transport, which rejects pending requests and settles the waiters
- syncEngine gate diagnostics and JSDoc name all supported flavors
  (Pi, Codex, Cursor ACP)
- regression test: Switch with an unresolved soft-steer completion still
  reaches teardown

* fix(codex): distinguish definite rejection from indeterminate completion

- Transport-level failures (timeout, abort, disconnect, spawn, protocol)
  carry an indeterminate marker; explicit JSON-RPC error responses do not
- After a dispatched steer, turn completion resolves → commit + consumed;
  a definite app-server rejection restores the row (instruction was never
  accepted, so turn/start cannot duplicate it); an indeterminate outcome
  leaves the row reserved so it can never be delivered twice
- Completion handling registers before awaiting dispatch so the
  dispatch-failure path cannot leak an unhandled rejection
- client/launcher tests cover explicit rejection (restore), indeterminate
  outcome (row stays reserved) and dispatch failure

* fix(codex): reconcile indeterminate steers instead of a permanent reservation

- After an indeterminate completion (disconnect/protocol), reconcile the
  thread by clientUserMessageId immediately: accepted → commit + consumed,
  provably rejected → restore, still unreadable → keep the reservation and
  retry from the main-loop top on later passes (post-reconnect)
- A row never sits in dispatching forever: the hub cannot stamp it invoked
  while the instruction may never have been accepted
- tests: indeterminate keeps reserved while thread unreadable; accepted
  reconciliation consumes; rejected path restores

* fix(cursor): abort drops soft-steer waiters so the next prompt never blocks

- Ordinary Abort (shouldExit false) now clears softSteerWaiters: the
  prompt finally cannot wait forever on a soft steer whose completion is
  unbounded; the ACP cancel rejects in-flight requests, and cleanup()
  settles leftovers on session end
- regression test: unresolved soft-steer completion after Abort no longer
  blocks the next prompt

* fix(codex): accept all thread item shapes; retry reconcile; ack through abort

- Reconcile matcher accepts userMessage/user_message with clientId/
  client_id, matching the shapes the thread parser supports — an accepted
  steer can no longer be misclassified as rejected
- A pending reconciliation schedules a wakeLoop retry, so a temporary
  app-server outage cannot strand the reservation behind waitForTurnOrRecovery
- The success-path ACK no longer checks the steer epoch: the hub already
  reported steered on dispatch, so commit + messages-consumed must reach
  it even when an abort resets the queue in between

* fix(cursor,acp): abort force-settles soft-steer bookkeeping

- AcpSdkBackend.abortSoftSteers() drops the concurrent-prompt counter and
  notifies response-complete so the next turn's waitForResponseComplete()
  cannot block on a soft steer that will never settle after abort
- handleAbort calls it before clearing the waiters; the main prompt's own
  finishPromptRequest stays guarded by Math.max(0, ...)
- unit tests cover counter release and no-op when idle

* fix(codex): reinit reconnected app-server; keep reconcile retries alive

- thread/read after a disconnect auto-connects a fresh app-server, which
  must be initialized before any request — reconcile now ensures
  connect + initialize (isConnected getter added to the client)
- every still-unknown loop-top reconciliation schedules the next retry,
  so recovery without external traffic is eventually observed
- launcher mock gains isConnected

* test(acp): match finishPromptRequest epoch signature in whitebox test

* fix(codex): timer-driven reconciliation; init tracking; abort-safe ACK

- Reconciliation runs on a self-rescheduling 1s timer independent of the
  main loop (wakes it too), so idle loops and waitForTurnOrRecovery still
  observe app-server recovery; abort clears nothing implicitly — the ACK
  path commits and consumes even when the reservation was cancelled
- Absence of a durable client id is ambiguous: unmatched reads stay
  'unknown' and keep retrying instead of restoring the row
- CodexAppServerClient tracks initialized state (reset on disconnect/exit)
  so ensureAppServerInitialized re-initializes a fresh process before
  thread/read; initialize failures leave the flag false for the next retry
- tests: accepted reconciliation via scheduled timer, indeterminate
  keeps reserved, explicit rejection restores

* fix(codex): bind reconciliation to the launcher lifecycle

- runSteerReconciliation clears any armed retry timer on entry and never
  installs a second one, so loop-top and timer-driven passes cannot
  multiply
- shuttingDown is set when the main loop ends: timers are cleared and the
  pending map is dropped, so an unresolved steer can never respawn an
  app-server after cleanup (remote-to-local switch included)

* fix(cursor): abort releases an in-progress soft-steer wait

- The prompt-finally wait races Promise.allSettled against the abort
  signal: an Abort that clears the waiters now also releases a wait that
  already started, so the launcher always reaches the next queued prompt

* fix(codex): report steered only after app-server acceptance

- The handler now awaits steer.completed (the inject-acceptance response):
  an explicit JSON-RPC rejection surfaces as failed and restores the row
  for the normal turn/start path instead of a false steered
- Transport failure after dispatch reports 'Steer outcome is being
  reconciled' and keeps the row reserved while the timer-driven thread
  reconciliation runs
- dispatch-failure path also swallows the paired completion rejection

* fix(cursor,acp): commit on ACP acceptance; distinguish transport failures

- AcpStdioTransport marks transport-level failures (timeout, closed,
  stdin write) as indeterminate; explicit JSON-RPC error responses are not
- The steer handler commits + consumes on completion (ACP acceptance) and
  restores the row on an explicit rejection; an indeterminate transport
  failure keeps the row reserved so a delivered instruction is never
  re-sent, and the ACK reaches the hub even when abort reset the queue
- launcher/transport tests updated for the three outcomes

* fix(steer): tri-state cancel, clear-safe reservations, bounded acceptance wait

- MessageQueue2.cancelByLocalId returns 'in-flight' for a dispatching
  steer reservation: the hub neither deletes the row nor stamps invoked_at
  (new CancelMessageResponse 'busy' status; web restores the optimistic
  row); pushIsolateAndClear and reset/close share cancelReservations so
  /clear-style commands cannot have a rejected steer resurrect a discarded
  prompt
- turn/steer acceptance wait bounded at 25s (< hub 30s RPC timeout): a
  lost response is indeterminate and funnels into thread reconciliation
  instead of stranding the reservation
- tests updated for the tri-state cancel contract

* test(cursor): match tri-state cancel contract for dispatching steers

* fix(cursor): drop duplicate promptInFlight declaration after upstream merge

* fix(codex,web): busy-aware edit flow; bound reconciliation reads

- QueuedMessagesBar edit flow treats a 'busy' cancel as unsuccessful: it
  never prefills the composer when the row is inside an async steer, so a
  second client cannot send a duplicate
- reconcileSteerByClientId bounds thread/read with a 5s timeout so a
  connected-but-silent app-server cannot hold the reservation in-flight
  indefinitely

* fix(steer): inFlight-dominated cancel acks; bounded reconciliation

- hub cancel-queued-message acks check inFlight before removed: a stale
  duplicate socket reporting removed can no longer delete the durable row
  while another socket is dispatching the steer
- reconciliation entries expire after 60s and mark delivered: after the
  rejection window, a dispatched steer that the app-server never proved
  (client ids dropped on restart) is committed instead of polling
  thread/read forever
- pre-dispatch failures (abort before write included) never enter
  reconciliation — they restore the row and report failure

* fix(cursor,acp,web): indeterminate close marks, steer gating precision

- AcpStdioTransport.rejectAllPending marks close/protocol failures
  indeterminate, so an accepted-but-close-interrupted soft steer restores
  nothing (no duplicate delivery)
- the abort race in the soft-steer wait removes its listener in finally
  (no accumulation across repeated waits)
- SessionChat gates the Steer button on agentState.steeringActive for
  codex/cursor instead of the queued-grace thinking flag, so Steer is not
  exposed before the launcher can accept it
- codex pre-dispatch abort never enters reconciliation (merged from #1606)

* fix(steer): persist indeterminate outcomes without replay

* fix(cursor): hold ambiguous steers for explicit resolution

* fix(steer): make ambiguous delivery restart-safe

* fix(cursor): make ambiguous delivery restart-safe

* fix(steer): recover crash-held rows and preserve retry dedup

* fix(steer): ack retries and bound stdin dispatch

* fix(cursor): reject steers when prompt generation changes

* fix(steer): reconcile indeterminate dispatches and serialize retries

* fix(cursor): preserve soft-steer reservations across abort

* fix(codex): classify stdin callback failures as indeterminate

* fix(steer): recheck indeterminate cancels after ACK

* fix(steer): close retry and abort races

* fix(cursor): hold ambiguous dispatch failures

* fix(steer): serialize live retries and abort admission

* fix(steer): distinguish live dispatching from unknown

* fix(cursor): bound ACP dispatch acknowledgements

* fix(steer): keep ACK failures held and reconcile busy cancel

* fix(cursor): preserve state when dispatch ACK is uncertain

* fix(steer): distinguish held cancel from removal

* fix(store): combine schema v24 migrations

* fix(cursor): distinguish held cancel from removal

* fix(store): reserve schema v25 for steer delivery state

* fix(cursor): suppress late ACP updates after abort

* fix(steer): keep held cancel state and notify requeue

* fix(cursor): isolate late updates after abort

* fix(steer): release explicitly cancelled unknown reservations

* test(cursor): cover explicit held cancellation

* fix(codex): reject cancelled reservations before native steer

* fix(cursor): reject cancelled reservations before ACP steer

* fix(codex): make reservation restore atomic with state

* fix(cursor): make reservation restore atomic with state

* fix(codex): terminate abandoned transport writes

* fix(cursor): hard-stop abandoned writes and update native queue state

* fix(steer): own abandoned app-server lifecycle and consume races

* fix(cursor): isolate aborts and add native retry resolution

* fix(codex): confirm dispatch and recover abandoned turns

* test(codex): mock abandoned transport callback

* fix(native): reconcile retry responses

* fix(codex): clear visible turn state on transport loss

* fix(steer): claim retries and cover native delivery state

* fix(native): resync busy cancel outcomes

* fix(native): preserve indeterminate state on Android hydration

* fix(steer): make retry claims single-winner

* fix(cursor): hold restore failures for explicit resolution

* fix(steer): serialize concurrent retry claims

* fix(socket): tolerate missing steer-state ACK callbacks

* fix(native): serialize retry operations

* docs(web): document unknown steer delivery and retry controls

* fix(steer): handle retry failures and abort-before-connect

* fix(cursor): drain foreground prompt after soft-steer abort

* fix(steer): reinitialize after transport loss and finish iOS retry errors

* fix(steer): preserve indeterminate rows across reconnect gaps

* test(web): mock indeterminate queued recovery state

* fix(steer): recover consumed ACK tombstones

* fix(steer): expose consumed cancel tombstones

* fix(cursor): drain soft steers before handler replacement

* fix(cursor): preserve buffered output on abort
This commit is contained in:
SSU-WEI HUANG
2026-08-20 08:53:41 +08:00
committed by GitHub
parent 1f5602ad6a
commit 0f7a3da68b
15 changed files with 1138 additions and 57 deletions
+238 -3
View File
@@ -34,7 +34,9 @@ import { buildCursorModelsSeedPayload, seedCursorModelsCache } from '@/modules/c
import { readSharedCursorModelsCache } from '@/modules/common/cursorModelsSharedCache';
import type { AcpSdkBackend } from '@/agent/backends/acp';
import type { AcpStderrError } from '@/agent/backends/acp/AcpStdioTransport';
import { isAcpIndeterminateError } from '@/agent/backends/acp/AcpStdioTransport';
import { registerAcpSessionTitleSync } from '@/agent/acpSessionTitle';
import { RPC_METHODS } from '@hapi/protocol/rpcMethods';
import {
cursorHapiMcpServerId,
installCursorMcpOverlay,
@@ -50,6 +52,8 @@ import {
stripRetryableCursorError
} from './cursorAutoRetry';
const CURSOR_ABORT_DRAIN_TIMEOUT_MS = 5_000;
class CursorAcpRemoteLauncher extends RemoteLauncherBase {
private readonly session: CursorSession;
private backend: ReturnType<typeof createCursorAcpBackend> | null = null;
@@ -63,6 +67,11 @@ class CursorAcpRemoteLauncher extends RemoteLauncherBase {
private defaultBackendModel: string | null = null;
private unregisterModelApplyHandler: (() => void) | null = null;
private modelApplySeq = 0;
private activePromptModeHash: string | null = null;
/** True while a backend.prompt turn is in flight. */
private promptInFlight = false;
/** Concurrent soft-steer session/prompt RPCs still running after kickoff. */
private softSteerWaiters: Promise<void>[] = [];
/** True when ACP process was spawned with `--auto-review`. */
private spawnedWithAutoReview = false;
/** Avoid re-queueing `/auto-review` on every mid-session mode sync. */
@@ -72,7 +81,6 @@ class CursorAcpRemoteLauncher extends RemoteLauncherBase {
private pendingRetryableFromStderr = false;
private pendingInlineRetryableError = false;
private attemptProducedToolActivity = false;
private promptInFlight = false;
private userAbortRequested = false;
constructor(session: CursorSession) {
@@ -94,6 +102,7 @@ class CursorAcpRemoteLauncher extends RemoteLauncherBase {
protected async runMainLoop(): Promise<void> {
const session = this.session;
const messageBuffer = this.messageBuffer;
session.client.updateAgentState?.((state) => ({ ...state, steeringActive: false }));
const { server: happyServer, mcpServers } = await buildHapiMcpBridge(session.client, {
enableChangeTitle: false,
@@ -414,13 +423,165 @@ class CursorAcpRemoteLauncher extends RemoteLauncherBase {
onSwitch: () => this.handleSwitchRequest()
});
// Soft steer = Cursor GUI "Send" (next-opportune / soft inject): fire a
// concurrent session/prompt without canceling the in-flight turn. Abort
// remains the hard stop path (GUI "Stop & send").
session.client.rpcHandlerManager.registerHandler(
RPC_METHODS.SteerQueuedMessage,
async (payload: unknown) => {
const localId = typeof (payload as { localId?: unknown } | null)?.localId === 'string'
? (payload as { localId: string }).localId
: '';
if (!localId) {
return { steered: false, error: 'Missing localId' };
}
const backend = this.backend;
const acpSessionId = this.acpSessionId;
if (!this.promptInFlight || !acpSessionId || !backend) {
return { steered: false, error: 'No active steerable turn' };
}
const targetPromptGeneration = backend.getPromptGeneration();
const taken = session.queue.takeByLocalId(localId);
if (!taken) {
return { steered: false, error: 'Message not in queue' };
}
const isControlCommand = Boolean(taken.item.isolate)
|| parseCursorSpecialCommand(taken.item.message).type !== null;
if (isControlCommand) {
session.queue.restoreReservation(taken);
return { steered: false, error: 'Control commands cannot be steered' };
}
if (this.activePromptModeHash !== taken.item.modeHash) {
session.queue.restoreReservation(taken);
return { steered: false, error: 'Queued message mode differs from the active turn' };
}
// Ack the hub once the soft-steer request is kicked off — not when
// the concurrent session/prompt finishes. ACP treats that response as
// turn completion, which can exceed the hub's 30s Socket.IO RPC timeout
// and report a false failure after the inject already started.
// Keep the launcher busy until that background prompt settles so we
// do not emit ready / start the next backend.prompt() while it runs.
if (!session.queue.beginReservationDispatch(taken)) {
return { steered: false, error: 'Steer cancelled' };
}
const dispatchStatePersisted = await session.client.setSteerDeliveryState([localId], 'dispatching');
if (!dispatchStatePersisted) {
session.queue.markReservationIndeterminate(taken);
session.client.emitSteerIndeterminate([localId]);
return { steered: false, error: 'Steer state is indeterminate' };
}
const restoreQueuedReservation = async (): Promise<boolean> => {
if (!taken.originIndeterminate) {
const persisted = await session.client.setSteerDeliveryState([localId], 'queued');
if (!persisted) {
session.queue.markReservationIndeterminate(taken);
session.client.emitSteerIndeterminate([localId]);
return false;
}
}
if (taken.state !== 'dispatching' || !session.queue.restoreReservation(taken)) {
session.client.emitSteerIndeterminate([localId]);
return false;
}
return true;
};
if (taken.state !== 'dispatching') {
session.client.emitSteerIndeterminate([localId]);
return { steered: false, error: 'Steer cancelled' };
}
if (!this.promptInFlight
|| this.backend !== backend
|| this.acpSessionId !== acpSessionId
|| backend.getPromptGeneration() !== targetPromptGeneration) {
await restoreQueuedReservation();
return { steered: false, error: 'Active turn changed' };
}
let steer: { dispatched: Promise<void>; completed: Promise<void> };
try {
steer = backend.beginSoftSteerPrompt(acpSessionId, [{
type: 'text',
text: taken.item.message
}]);
} catch (error) {
if (isAcpIndeterminateError(error)) {
if (session.queue.markReservationIndeterminate(taken)) {
session.client.emitSteerIndeterminate([localId]);
}
logger.debug('[cursor-acp] soft-steer dispatch outcome unknown', error);
return { steered: false, error: 'Steer outcome is being reconciled' };
}
logger.debug('[cursor-acp] soft-steer failed to start', error);
await restoreQueuedReservation();
return { steered: false, error: 'Failed to soft-steer into active turn' };
}
// Completion still gates the next prompt (handler swap safety);
// register the waiter before awaiting dispatch so the main loop's
// finally cannot slip a prompt in between.
const steerDone = Promise.all([steer.dispatched, steer.completed]).then(() => {}, (error) => {
logger.debug('[cursor-acp] soft-steer completion failed after dispatch', error);
});
this.softSteerWaiters.push(steerDone);
const removeWaiter = () => {
this.softSteerWaiters = this.softSteerWaiters.filter((p) => p !== steerDone);
};
void steerDone.then(removeWaiter);
try {
await steer.dispatched;
} catch (error) {
if (isAcpIndeterminateError(error)) {
if (session.queue.markReservationIndeterminate(taken)) {
session.client.emitSteerIndeterminate([localId]);
}
logger.debug('[cursor-acp] soft-steer dispatch outcome unknown', error);
return { steered: false, error: 'Steer outcome is being reconciled' };
}
await restoreQueuedReservation();
logger.debug('[cursor-acp] soft-steer failed to start', error);
return { steered: false, error: 'Failed to soft-steer into active turn' };
}
// The RPC acks once stdin accepted the inject. The queue row is
// committed only when the concurrent prompt settles: an explicit
// JSON-RPC rejection means ACP never accepted the instruction
// (restore it for the next prompt), while a transport failure
// (abort/disconnect) keeps the row reserved — never re-delivered.
void steer.completed.then(() => {
// Completion means ACP accepted the inject: the ACK must
// reach the hub even when an abort reset the queue and
// cancelled the reservation in between.
session.queue.commitReservation(taken);
messageBuffer.addMessage(taken.item.message, 'user');
session.client.emitMessagesConsumed([localId], { steered: true });
}, (error) => {
if (isAcpIndeterminateError(error)) {
// Do not leave the reservation dispatching forever. Hold
// it outside automatic replay and persist the ambiguous
// outcome; a later explicit Steer retries this same row.
if (session.queue.markReservationIndeterminate(taken)) {
session.client.emitSteerIndeterminate([localId]);
}
logger.debug('[cursor-acp] soft-steer outcome unknown after dispatch; row held for explicit resolution', error);
return;
}
void restoreQueuedReservation().then((restored) => {
if (restored) {
logger.debug('[cursor-acp] soft-steer rejected by ACP; row restored', error);
}
});
});
return { steered: true };
}
);
const sendReady = () => {
session.sendSessionEvent({ type: 'ready' });
};
try {
while (!this.shouldExit) {
const waitSignal = this.abortController.signal;
const batch = await session.queue.waitForMessagesAndGetAsString(waitSignal);
if (!batch) {
if (waitSignal.aborted && !this.shouldExit) {
continue;
@@ -463,6 +624,9 @@ class CursorAcpRemoteLauncher extends RemoteLauncherBase {
}];
session.onThinkingChange(true);
this.promptInFlight = true;
session.client.updateAgentState?.((state) => ({ ...state, steeringActive: true }));
this.activePromptModeHash = batch.hash;
try {
this.promptInFlight = true;
@@ -508,6 +672,31 @@ class CursorAcpRemoteLauncher extends RemoteLauncherBase {
}
} finally {
this.promptInFlight = false;
session.client.updateAgentState?.((state) => ({ ...state, steeringActive: false }));
// Soft-steers share the ACP session; wait for them before ready /
// the next prompt so message handlers are not swapped mid-inject.
// An Abort (which clears the waiters) must release this wait too:
// race the settle against the abort signal so the launcher never
// blocks on a soft steer whose completion is unbounded.
if (this.softSteerWaiters.length > 0 && !this.shouldExit) {
const waitSignal = this.abortController.signal;
let releaseWait!: () => void;
const abortListener = () => releaseWait();
if (!waitSignal.aborted) {
waitSignal.addEventListener('abort', abortListener, { once: true });
}
try {
await Promise.race([
Promise.allSettled([...this.softSteerWaiters]),
new Promise<void>((resolve) => { releaseWait = resolve; })
]);
} finally {
// Repeated waits must not accumulate abort listeners.
waitSignal.removeEventListener('abort', abortListener);
}
this.softSteerWaiters = [];
}
this.activePromptModeHash = null;
this.pendingRetryableError = null;
this.pendingRetryableFromStderr = false;
this.pendingInlineRetryableError = false;
@@ -520,6 +709,11 @@ class CursorAcpRemoteLauncher extends RemoteLauncherBase {
}
}
}
} finally {
// No wait here: Exit/Switch must reach cleanup() promptly; it
// disconnects the ACP transport, rejecting pending soft-steer
// requests and settling any waiters.
}
}
protected async cleanup(): Promise<void> {
@@ -530,6 +724,13 @@ class CursorAcpRemoteLauncher extends RemoteLauncherBase {
try {
this.clearAbortHandlers(this.session.client.rpcHandlerManager);
this.session.client.rpcHandlerManager.registerHandler(RPC_METHODS.SteerQueuedMessage, async () => ({
steered: false,
error: 'Session ending'
}));
this.promptInFlight = false;
this.session.client.updateAgentState?.((state) => ({ ...state, steeringActive: false }));
this.softSteerWaiters = [];
this.unregisterModelApplyHandler?.();
this.unregisterModelApplyHandler = null;
@@ -887,13 +1088,47 @@ class CursorAcpRemoteLauncher extends RemoteLauncherBase {
private async handleAbort(): Promise<void> {
this.userAbortRequested = true;
const backend = this.backend;
const sessionId = this.session.sessionId;
const sessionId = this.acpSessionId ?? this.session.sessionId;
if (backend && sessionId) {
const pendingSoftSteers = [...this.softSteerWaiters];
await backend.cancelPrompt(sessionId);
// Drop soft-steer bookkeeping first; retain the foreground prompt
// count and wait for both boundaries before a new handler is used.
backend.abortSoftSteers();
if (!this.shouldExit) {
let timeout: ReturnType<typeof setTimeout> | null = null;
const drained = await Promise.race([
Promise.all([
backend.waitForResponseComplete(),
Promise.allSettled(pendingSoftSteers)
]).then(() => true),
new Promise<boolean>((resolve) => {
timeout = setTimeout(() => resolve(false), CURSOR_ABORT_DRAIN_TIMEOUT_MS);
timeout.unref?.();
})
]);
if (timeout) clearTimeout(timeout);
if (!drained) {
// An ACP request that ignores cancel cannot safely share a
// handler with the next prompt. End the launcher instead
// of allowing late updates to cross the turn boundary.
logger.warn('[cursor-acp] abort drain timed out; ending session to isolate late ACP updates');
this.shouldExit = true;
}
}
}
await this.permissionAdapter?.cancelAll('User aborted');
await this.extensionAdapter?.cancelAll('User aborted');
this.session.queue.reset();
// A soft steer may settle after Abort; preserve its reservation until
// the completion callback records accepted or indeterminate.
this.session.queue.reset({ preserveDispatchingReservations: true });
this.promptInFlight = false;
// Abort is the hard-stop path: drop soft-steer waiters so the prompt
// finally cannot block the next prompt on a soft steer whose completion
// is unbounded and may never settle. Soft counters were already reset
// above; only the foreground prompt was drained before continuing.
this.softSteerWaiters = [];
this.session.client.updateAgentState?.((state) => ({ ...state, steeringActive: false }));
this.session.onThinkingChange(false);
this.abortController.abort();
this.abortController = new AbortController();