@borgee/agents-host 0.2.33 → 0.2.35

package/dist/providers/codex/cli-client.js

@@ -6,8 +6,18 @@ import spawn from 'cross-spawn';
 import { PROTOCOL_VERSION, client, methods, ndJsonStream, } from '@agentclientprotocol/sdk';
 import { HostLogger, summarizeChildStderr, summarizeError } from '../../debug.js';
 import { assertCodexProjectDocumentSize, buildCodexProjectDocument } from './project-doc.js';
+import { IDLE_BACKEND_SHUTDOWN_DISABLED_MS, IdleBackendShutdownScheduler, } from '../idle-backend-shutdown.js';
 const SESSION_TAINTED_ERRORS = new WeakSet();
 const DEFAULT_IDLE_SESSION_TTL_MS = 2 * 24 * 60 * 60 * 1000;
+const IDLE_BACKEND_STOPPED_MESSAGE = 'Codex ACP backend stopped after idle timeout';
+/**
+ * A fatal teardown has callers waiting on it, so it force-kills quickly. An
+ * idle recycle has nobody waiting, and the CLIs unlink their on-disk session
+ * locks during SIGTERM cleanup — a lock left behind by SIGKILL makes the next
+ * restore believe the session is still in use (github/copilot-cli#3255). So the
+ * idle path trades latency nobody observes for a clean exit.
+ */
+const DEFAULT_IDLE_SHUTDOWN_GRACE_PERIOD_MS = 5_000;
 const DEFAULT_SHUTDOWN_GRACE_PERIOD_MS = 250;
 const DEFAULT_SHUTDOWN_FORCE_KILL_WAIT_MS = 250;
 const DEFAULT_SESSION_CAPABILITIES = {
@@ -30,6 +40,8 @@ const DEFAULT_RUNTIME = {
     protocolVersion: PROTOCOL_VERSION,
     cwd: process.cwd(),
     idleSessionTtlMs: DEFAULT_IDLE_SESSION_TTL_MS,
+    idleBackendShutdownMs: IDLE_BACKEND_SHUTDOWN_DISABLED_MS,
+    idleShutdownGracePeriodMs: DEFAULT_IDLE_SHUTDOWN_GRACE_PERIOD_MS,
     shutdownGracePeriodMs: DEFAULT_SHUTDOWN_GRACE_PERIOD_MS,
     shutdownForceKillWaitMs: DEFAULT_SHUTDOWN_FORCE_KILL_WAIT_MS,
     readFile: async (path, options) => {
@@ -286,6 +298,14 @@ export class CodexCliClient {
     logger;
     runtime;
     channels = new Map();
+    /**
+     * The provider session currently bound to channelId, or undefined before a
+     * turn has opened one. Surfaced so agents-host can record which session
+     * worked a task without the agent being asked to report its own id.
+     */
+    sessionIdForChannel(channelId) {
+        return this.channels.get(channelId)?.session?.sessionId;
+    }
     persistedSessions = new Map();
     closingSessions = new WeakSet();
     pendingSessionStarts = new Set();
@@ -298,7 +318,13 @@ export class CodexCliClient {
     startPromise;
     shutdownPromise;
     childExitPromise;
-    resolveChildExit;
+    /**
+     * Bumped on every spawn so the `exit` / `connection.closed` handlers of a
+     * backend we deliberately recycled stay silent instead of latching a fatal
+     * error onto the client that already owns its successor.
+     */
+    backendGeneration = 0;
+    idleShutdownPromise;
     fatalError = null;
     disposing = false;
     backendClosed = false;
@@ -306,6 +332,7 @@ export class CodexCliClient {
     sessionStoreLoadPromise = null;
     sessionStoreWriteQueue = Promise.resolve();
     sessionCapabilities = DEFAULT_SESSION_CAPABILITIES;
+    idleBackendShutdown;
     constructor(command, args = [], runtimeOverrides = {}, sessionStore, resolveSessionStoreAgentId = () => undefined, logger = new HostLogger()) {
         this.command = command;
         this.args = args;
@@ -313,6 +340,16 @@ export class CodexCliClient {
         this.resolveSessionStoreAgentId = resolveSessionStoreAgentId;
         this.logger = logger;
         this.runtime = { ...DEFAULT_RUNTIME, ...runtimeOverrides };
+        this.idleBackendShutdown = new IdleBackendShutdownScheduler({
+            idleShutdownMs: this.runtime.idleBackendShutdownMs,
+            isIdle: () => this.isBackendIdle(),
+            shutdown: () => this.shutdownIdleBackend(),
+            onShutdownFailed: (error) => {
+                this.logger.error('failed to stop the idle Codex ACP backend', {
+                    error: summarizeError(error),
+                });
+            },
+        });
         this.fatalPromise = new Promise((_, reject) => {
             this.rejectFatalPromise = reject;
         });
@@ -334,6 +371,7 @@ export class CodexCliClient {
         const sessionRouting = resolveSessionRouting(preparedTurn);
         const state = this.getOrCreateChannelState(sessionRouting.key, preparedTurn.channelId, sessionRouting.persistence);
         this.clearIdleTimer(state);
+        this.idleBackendShutdown.cancel();
         const turn = createDeferredTurn(preparedTurn.channelId, preparedTurn.prompt, preparedTurn.promptContext, sessionRouting.persistence, options);
         state.queue.push(turn);
         this.processChannelQueue(sessionRouting.key, state);
@@ -342,12 +380,17 @@ export class CodexCliClient {
     async dispose() {
         const error = new Error('Codex ACP backend stopped');
         this.disposing = true;
+        this.idleBackendShutdown.cancel();
         this.logger?.debug('stopping Codex ACP backend');
         const closed = this.connection?.closed ?? Promise.resolve();
         this.failAll(error);
         let thrown;
         try {
-            await Promise.all([closed, this.shutdownPromise ?? Promise.resolve()]);
+            await Promise.all([
+                closed,
+                this.idleShutdownPromise ?? Promise.resolve(),
+                this.shutdownPromise ?? Promise.resolve(),
+            ]);
         }
         catch (disposeError) {
             thrown = disposeError;
@@ -368,6 +411,12 @@ export class CodexCliClient {
         if (this.fatalError) {
             throw this.fatalError;
         }
+        if (this.idleShutdownPromise) {
+            await this.idleShutdownPromise;
+            if (this.fatalError) {
+                throw this.fatalError;
+            }
+        }
         if (!this.startPromise) {
             this.startPromise = this.startBackend();
         }
@@ -380,6 +429,8 @@ export class CodexCliClient {
         }
     }
     async startBackend() {
+        const generation = ++this.backendGeneration;
+        const ownsCurrentBackend = () => (this.backendGeneration === generation && !this.disposing);
         const launch = resolveAdapterLaunch(this.command, this.args);
         this.logger?.debug('starting Codex ACP process', {
             command: launch.command,
@@ -396,13 +447,20 @@ export class CodexCliClient {
         });
         child.stderr.resume();
         this.logger?.debug('spawned Codex ACP process', { pid: child.pid });
+        let resolveChildExit;
+        this.childExitPromise = new Promise((resolve) => {
+            resolveChildExit = resolve;
+        });
         child.once('error', (error) => {
             this.logger?.debugError('Codex ACP process error', error);
+            if (!ownsCurrentBackend()) {
+                return;
+            }
             this.failAll(new Error(`Codex ACP process error: ${normalizeError(error).message}`));
         });
         child.once('exit', (code, signal) => {
-            this.resolveChildExit?.();
-            if (this.disposing) {
+            resolveChildExit();
+            if (!ownsCurrentBackend()) {
                 this.logger?.debug('Codex ACP process exited during shutdown', { code, signal });
                 return;
             }
@@ -410,9 +468,6 @@ export class CodexCliClient {
             const suffix = signal ? `signal ${signal}` : `code ${String(code ?? 'unknown')}`;
             this.failAll(new Error(`Codex ACP process exited unexpectedly (${suffix})`));
         });
-        this.childExitPromise = new Promise((resolveChildExit) => {
-            this.resolveChildExit = resolveChildExit;
-        });
         const output = Writable.toWeb(child.stdin);
         const input = Readable.toWeb(child.stdout);
         const stream = this.runtime.ndJsonStream(output, input);
@@ -422,7 +477,7 @@ export class CodexCliClient {
         const connection = app.connect(stream);
         this.connection = connection;
         void connection.closed.then(() => {
-            if (!this.disposing) {
+            if (ownsCurrentBackend()) {
                 this.logger?.debugError('Codex ACP connection closed unexpectedly');
                 this.failAll(new Error('Codex ACP connection closed unexpectedly'));
             }
@@ -512,6 +567,7 @@ export class CodexCliClient {
             finally {
                 state.processing = false;
                 this.reconcileIdleChannelState(channelId, state);
+                this.idleBackendShutdown.reconcile();
             }
         })();
     }
@@ -856,6 +912,7 @@ export class CodexCliClient {
         }
         this.fatalError = error;
         this.rejectFatalPromise(error);
+        this.idleBackendShutdown.cancel();
         for (const [channelId, state] of this.channels.entries()) {
             this.clearIdleTimer(state);
             state.activeTurn?.reject(error);
@@ -1073,6 +1130,66 @@ export class CodexCliClient {
             return;
         }
         this.backendClosed = true;
+        await this.closeBackendProcess(error, this.runtime.shutdownGracePeriodMs);
+    }
+    /**
+     * Backend teardown without the fatal latch: the client stays usable and
+     * `ensureStarted()` spawns a replacement process on the next turn. Only safe
+     * while `isBackendIdle()` holds, because every channel's live ACP session is
+     * dropped here and no turn is ever replayed onto its successor.
+     */
+    async shutdownIdleBackend() {
+        // Detach the running backend before any signal reaches it, so its own exit
+        // and connection-closed handlers read this teardown as deliberate instead
+        // of latching a fatal error on a client that stays usable.
+        this.backendGeneration += 1;
+        const error = new Error(IDLE_BACKEND_STOPPED_MESSAGE);
+        const sessions = [];
+        for (const [channelId, state] of [...this.channels.entries()]) {
+            this.clearIdleTimer(state);
+            if (state.session) {
+                sessions.push(state.session);
+            }
+            state.session = undefined;
+            state.sessionCwd = undefined;
+            state.sessionAdditionalDirectories = undefined;
+            state.sessionVisibilityKey = undefined;
+            this.channels.delete(channelId);
+        }
+        this.logger?.debug('stopping idle Codex ACP backend', {
+            channels: sessions.length,
+            idleShutdownMs: this.runtime.idleBackendShutdownMs,
+        });
+        for (const session of sessions) {
+            this.closeSession(session);
+        }
+        const idleShutdownPromise = this.closeBackendProcess(error, this.runtime.idleShutdownGracePeriodMs).finally(() => {
+            this.startPromise = undefined;
+            this.childExitPromise = undefined;
+            this.sessionCapabilities = DEFAULT_SESSION_CAPABILITIES;
+            this.idleShutdownPromise = undefined;
+        });
+        this.idleShutdownPromise = idleShutdownPromise;
+        await idleShutdownPromise;
+    }
+    isBackendIdle() {
+        if (this.disposing || this.fatalError || this.backendClosed || this.idleShutdownPromise) {
+            return false;
+        }
+        if (!this.child) {
+            return false;
+        }
+        if (this.pendingSessionStarts.size > 0) {
+            return false;
+        }
+        for (const state of this.channels.values()) {
+            if (state.activeTurn || state.queue.length > 0 || state.sessionPromise || state.processing) {
+                return false;
+            }
+        }
+        return true;
+    }
+    async closeBackendProcess(error, gracePeriodMs) {
         const connection = this.connection;
         const child = this.child;
         await this.waitForPendingSessionStarts();
@@ -1084,9 +1201,9 @@ export class CodexCliClient {
             return;
         }
         const childExitPromise = this.childExitPromise ?? Promise.resolve();
-        this.logger?.debug('sending SIGTERM to Codex ACP process');
+        this.logger?.debug('sending SIGTERM to Codex ACP process', { gracePeriodMs });
         child.kill('SIGTERM');
-        const exitedAfterTerm = await this.waitForChildExit(childExitPromise, this.runtime.shutdownGracePeriodMs);
+        const exitedAfterTerm = await this.waitForChildExit(childExitPromise, gracePeriodMs);
         if (exitedAfterTerm) {
             this.logger?.debug('Codex ACP process exited after SIGTERM');
             this.child = undefined;

package/dist/providers/codex/project-doc.js

@@ -2,6 +2,7 @@ import { basename } from 'node:path';
 import { buildAttentionSummaryLines } from '../../context/attention.js';
 import { buildCollaborationCapabilityDeclarationSummaryLines, buildMissedCollaborationDiagnosticSummaryLines, } from '../../context/collaboration-capabilities-diagnostics.js';
 import { buildCollaborationOutcomeSummaryLines } from '../../context/collaboration-outcome.js';
+import { MAIN_SESSION_DELEGATION_LINES } from '../../context/main-session-delegation.js';
 import { buildTaskThreadCollaborationSummaryLines } from '../../context/task-thread-collaboration.js';
 const PROJECT_DOC_MAX_BYTES = 32 * 1024;
 function buildSkillRuntimeLines(context) {
@@ -37,6 +38,8 @@ export function buildCodexProjectDocument(context) {
         '',
         'This workspace belongs to one Borgee channel session hosted by agents-host.',
         'Keep visible replies concise, do not claim actions you did not perform, and prefer the per-turn prompt when it provides newer turn-local details.',
+        '',
+        ...MAIN_SESSION_DELEGATION_LINES,
         ...(context?.channelContextPayloadPath
             ? [
                 '',

package/dist/providers/copilot/adapter.js

@@ -11,7 +11,9 @@ export class CopilotProviderAdapter {
         const text = await this.cli.generateReply(preparedTurn, {
             onProgress: createAwaitingUserProgressHandler(options?.onProgress),
         });
-        return parseProviderReply(text);
+        // Read the session AFTER the turn: a first turn on a channel opens the
+        // session as it runs, so before this point there is nothing to report.
+        return { ...parseProviderReply(text), sessionId: this.cli.sessionIdForChannel(input.channelId) };
     }
     async dispose() {
         await this.cli.dispose();

package/dist/providers/copilot/cli-client.d.ts

@@ -13,6 +13,8 @@ interface CopilotAcpRuntime {
     protocolVersion: typeof PROTOCOL_VERSION;
     cwd: string;
     idleSessionTtlMs: number;
+    idleBackendShutdownMs: number;
+    idleShutdownGracePeriodMs: number;
     shutdownGracePeriodMs: number;
     shutdownForceKillWaitMs: number;
 }
@@ -37,6 +39,12 @@ export declare class CopilotCliClient {
     private readonly permissionPolicy;
     private readonly runtime;
     private readonly channels;
+    /**
+     * The provider session currently bound to channelId, or undefined before a
+     * turn has opened one. Surfaced so agents-host can record which session
+     * worked a task without the agent being asked to report its own id.
+     */
+    sessionIdForChannel(channelId: string): string | undefined;
     private readonly persistedSessions;
     private readonly closingSessions;
     private readonly pendingSessionStarts;
@@ -49,7 +57,13 @@ export declare class CopilotCliClient {
     private startPromise?;
     private shutdownPromise?;
     private childExitPromise?;
-    private resolveChildExit?;
+    /**
+     * Bumped on every spawn so the `exit` / `connection.closed` handlers of a
+     * backend we deliberately recycled stay silent instead of latching a fatal
+     * error onto the client that already owns its successor.
+     */
+    private backendGeneration;
+    private idleShutdownPromise?;
     private fatalError;
     private disposing;
     private backendClosed;
@@ -57,6 +71,7 @@ export declare class CopilotCliClient {
     private sessionStoreLoadPromise;
     private sessionStoreWriteQueue;
     private sessionCapabilities;
+    private readonly idleBackendShutdown;
     constructor(command: string, _ignoredArgs?: string[], runtimeOverrides?: Partial<CopilotAcpRuntime>, sessionStore?: CopilotChannelSessionStore | undefined, resolveSessionStoreAgentId?: () => string | undefined, logger?: DebugLogger, permissionPolicy?: CopilotPermissionPolicyOptions);
     generateReply(turn: PreparedProviderTurnInput, options?: ProviderGenerateOptions): Promise<string>;
     generateReply(channelId: string, prompt: string, options?: ProviderGenerateOptions): Promise<string>;
@@ -92,6 +107,15 @@ export declare class CopilotCliClient {
     private trackSessionStoreOperation;
     private evictIdleChannel;
     private shutdownBackend;
+    /**
+     * Backend teardown without the fatal latch: the client stays usable and
+     * `ensureStarted()` spawns a replacement process on the next turn. Only safe
+     * while `isBackendIdle()` holds, because every channel's live ACP session is
+     * dropped here and no turn is ever replayed onto its successor.
+     */
+    private shutdownIdleBackend;
+    private isBackendIdle;
+    private closeBackendProcess;
     private waitForPendingSessionStarts;
     private waitForPendingSessionCloses;
     private waitForChildExit;

package/dist/providers/copilot/cli-client.js

@@ -3,8 +3,18 @@ import spawn from 'cross-spawn';
 import { PROTOCOL_VERSION, client, methods, ndJsonStream, } from '@agentclientprotocol/sdk';
 import { HostLogger, summarizeChildStderr, summarizeError } from '../../debug.js';
 import { resolveCopilotPermissionResponse } from '../../policy/copilot-permission.js';
+import { IDLE_BACKEND_SHUTDOWN_DISABLED_MS, IdleBackendShutdownScheduler, } from '../idle-backend-shutdown.js';
 const SESSION_TAINTED_ERRORS = new WeakSet();
 const DEFAULT_IDLE_SESSION_TTL_MS = 2 * 24 * 60 * 60 * 1000;
+const IDLE_BACKEND_STOPPED_MESSAGE = 'Copilot ACP backend stopped after idle timeout';
+/**
+ * A fatal teardown has callers waiting on it, so it force-kills quickly. An
+ * idle recycle has nobody waiting, and the CLIs unlink their on-disk session
+ * locks during SIGTERM cleanup — a lock left behind by SIGKILL makes the next
+ * restore believe the session is still in use (github/copilot-cli#3255). So the
+ * idle path trades latency nobody observes for a clean exit.
+ */
+const DEFAULT_IDLE_SHUTDOWN_GRACE_PERIOD_MS = 5_000;
 const DEFAULT_SHUTDOWN_GRACE_PERIOD_MS = 250;
 const DEFAULT_SHUTDOWN_FORCE_KILL_WAIT_MS = 250;
 const QUEUED_TURN_DROPPED_MESSAGE = 'Copilot ACP session was reset after a failed turn; queued turns were dropped instead of replaying them on a fresh session';
@@ -16,6 +26,8 @@ const DEFAULT_RUNTIME = {
     protocolVersion: PROTOCOL_VERSION,
     cwd: process.cwd(),
     idleSessionTtlMs: DEFAULT_IDLE_SESSION_TTL_MS,
+    idleBackendShutdownMs: IDLE_BACKEND_SHUTDOWN_DISABLED_MS,
+    idleShutdownGracePeriodMs: DEFAULT_IDLE_SHUTDOWN_GRACE_PERIOD_MS,
     shutdownGracePeriodMs: DEFAULT_SHUTDOWN_GRACE_PERIOD_MS,
     shutdownForceKillWaitMs: DEFAULT_SHUTDOWN_FORCE_KILL_WAIT_MS,
 };
@@ -206,6 +218,14 @@ export class CopilotCliClient {
     permissionPolicy;
     runtime;
     channels = new Map();
+    /**
+     * The provider session currently bound to channelId, or undefined before a
+     * turn has opened one. Surfaced so agents-host can record which session
+     * worked a task without the agent being asked to report its own id.
+     */
+    sessionIdForChannel(channelId) {
+        return this.channels.get(channelId)?.session?.sessionId;
+    }
     persistedSessions = new Map();
     closingSessions = new WeakSet();
     pendingSessionStarts = new Set();
@@ -218,7 +238,13 @@ export class CopilotCliClient {
     startPromise;
     shutdownPromise;
     childExitPromise;
-    resolveChildExit;
+    /**
+     * Bumped on every spawn so the `exit` / `connection.closed` handlers of a
+     * backend we deliberately recycled stay silent instead of latching a fatal
+     * error onto the client that already owns its successor.
+     */
+    backendGeneration = 0;
+    idleShutdownPromise;
     fatalError = null;
     disposing = false;
     backendClosed = false;
@@ -226,6 +252,7 @@ export class CopilotCliClient {
     sessionStoreLoadPromise = null;
     sessionStoreWriteQueue = Promise.resolve();
     sessionCapabilities = DEFAULT_SESSION_CAPABILITIES;
+    idleBackendShutdown;
     constructor(command, _ignoredArgs = [], runtimeOverrides = {}, sessionStore, resolveSessionStoreAgentId = () => undefined, logger = new HostLogger(), permissionPolicy = {}) {
         this.command = command;
         this.sessionStore = sessionStore;
@@ -233,6 +260,16 @@ export class CopilotCliClient {
         this.logger = logger;
         this.permissionPolicy = permissionPolicy;
         this.runtime = { ...DEFAULT_RUNTIME, ...runtimeOverrides };
+        this.idleBackendShutdown = new IdleBackendShutdownScheduler({
+            idleShutdownMs: this.runtime.idleBackendShutdownMs,
+            isIdle: () => this.isBackendIdle(),
+            shutdown: () => this.shutdownIdleBackend(),
+            onShutdownFailed: (error) => {
+                this.logger.error('failed to stop the idle Copilot ACP backend', {
+                    error: summarizeError(error),
+                });
+            },
+        });
         this.fatalPromise = new Promise((_, reject) => {
             this.rejectFatalPromise = reject;
         });
@@ -254,6 +291,7 @@ export class CopilotCliClient {
         const sessionRouting = resolveSessionRouting(preparedTurn);
         const state = this.getOrCreateChannelState(sessionRouting.key, preparedTurn.channelId, sessionRouting.persistence);
         this.clearIdleTimer(state);
+        this.idleBackendShutdown.cancel();
         const turn = createDeferredTurn(preparedTurn.channelId, preparedTurn.prompt, sessionRouting.persistence, preparedTurn.promptContext, options);
         state.queue.push(turn);
         this.processChannelQueue(sessionRouting.key, state);
@@ -262,12 +300,17 @@ export class CopilotCliClient {
     async dispose() {
         const error = new Error('Copilot ACP backend stopped');
         this.disposing = true;
+        this.idleBackendShutdown.cancel();
         this.logger?.debug('stopping Copilot ACP backend');
         const closed = this.connection?.closed ?? Promise.resolve();
         this.failAll(error);
         let thrown;
         try {
-            await Promise.all([closed, this.shutdownPromise ?? Promise.resolve()]);
+            await Promise.all([
+                closed,
+                this.idleShutdownPromise ?? Promise.resolve(),
+                this.shutdownPromise ?? Promise.resolve(),
+            ]);
         }
         catch (disposeError) {
             thrown = disposeError;
@@ -288,6 +331,12 @@ export class CopilotCliClient {
         if (this.fatalError) {
             throw this.fatalError;
         }
+        if (this.idleShutdownPromise) {
+            await this.idleShutdownPromise;
+            if (this.fatalError) {
+                throw this.fatalError;
+            }
+        }
         if (!this.startPromise) {
             this.startPromise = this.startBackend();
         }
@@ -300,6 +349,8 @@ export class CopilotCliClient {
         }
     }
     async startBackend() {
+        const generation = ++this.backendGeneration;
+        const ownsCurrentBackend = () => (this.backendGeneration === generation && !this.disposing);
         this.logger?.debug('starting Copilot ACP process', {
             command: this.command,
             args: ['--acp'],
@@ -315,13 +366,20 @@ export class CopilotCliClient {
         });
         child.stderr.resume();
         this.logger?.debug('spawned Copilot ACP process', { pid: child.pid });
+        let resolveChildExit;
+        this.childExitPromise = new Promise((resolve) => {
+            resolveChildExit = resolve;
+        });
         child.once('error', (error) => {
             this.logger?.debugError('Copilot ACP process error', error);
+            if (!ownsCurrentBackend()) {
+                return;
+            }
             this.failAll(new Error(`Copilot ACP process error: ${normalizeError(error).message}`));
         });
         child.once('exit', (code, signal) => {
-            this.resolveChildExit?.();
-            if (this.disposing) {
+            resolveChildExit();
+            if (!ownsCurrentBackend()) {
                 this.logger?.debug('Copilot ACP process exited during shutdown', { code, signal });
                 return;
             }
@@ -329,9 +387,6 @@ export class CopilotCliClient {
             const suffix = signal ? `signal ${signal}` : `code ${String(code ?? 'unknown')}`;
             this.failAll(new Error(`Copilot ACP process exited unexpectedly (${suffix})`));
         });
-        this.childExitPromise = new Promise((resolve) => {
-            this.resolveChildExit = resolve;
-        });
         const output = Writable.toWeb(child.stdin);
         const input = Readable.toWeb(child.stdout);
         const stream = this.runtime.ndJsonStream(output, input);
@@ -341,7 +396,7 @@ export class CopilotCliClient {
         const connection = app.connect(stream);
         this.connection = connection;
         void connection.closed.then(() => {
-            if (!this.disposing) {
+            if (ownsCurrentBackend()) {
                 this.logger?.debugError('Copilot ACP connection closed unexpectedly');
                 this.failAll(new Error('Copilot ACP connection closed unexpectedly'));
             }
@@ -428,6 +483,7 @@ export class CopilotCliClient {
             finally {
                 state.processing = false;
                 this.reconcileIdleChannelState(channelId, state);
+                this.idleBackendShutdown.reconcile();
             }
         })();
     }
@@ -643,6 +699,7 @@ export class CopilotCliClient {
         }
         this.fatalError = error;
         this.rejectFatalPromise(error);
+        this.idleBackendShutdown.cancel();
         for (const [channelId, state] of this.channels.entries()) {
             this.clearIdleTimer(state);
             state.activeTurn?.reject(error);
@@ -884,6 +941,64 @@ export class CopilotCliClient {
             return;
         }
         this.backendClosed = true;
+        await this.closeBackendProcess(error, this.runtime.shutdownGracePeriodMs);
+    }
+    /**
+     * Backend teardown without the fatal latch: the client stays usable and
+     * `ensureStarted()` spawns a replacement process on the next turn. Only safe
+     * while `isBackendIdle()` holds, because every channel's live ACP session is
+     * dropped here and no turn is ever replayed onto its successor.
+     */
+    async shutdownIdleBackend() {
+        // Detach the running backend before any signal reaches it, so its own exit
+        // and connection-closed handlers read this teardown as deliberate instead
+        // of latching a fatal error on a client that stays usable.
+        this.backendGeneration += 1;
+        const error = new Error(IDLE_BACKEND_STOPPED_MESSAGE);
+        const sessions = [];
+        for (const [channelId, state] of [...this.channels.entries()]) {
+            this.clearIdleTimer(state);
+            if (state.session) {
+                sessions.push(state.session);
+            }
+            state.session = undefined;
+            state.sessionCwd = undefined;
+            this.channels.delete(channelId);
+        }
+        this.logger?.debug('stopping idle Copilot ACP backend', {
+            channels: sessions.length,
+            idleShutdownMs: this.runtime.idleBackendShutdownMs,
+        });
+        for (const session of sessions) {
+            this.closeSession(session);
+        }
+        const idleShutdownPromise = this.closeBackendProcess(error, this.runtime.idleShutdownGracePeriodMs).finally(() => {
+            this.startPromise = undefined;
+            this.childExitPromise = undefined;
+            this.sessionCapabilities = DEFAULT_SESSION_CAPABILITIES;
+            this.idleShutdownPromise = undefined;
+        });
+        this.idleShutdownPromise = idleShutdownPromise;
+        await idleShutdownPromise;
+    }
+    isBackendIdle() {
+        if (this.disposing || this.fatalError || this.backendClosed || this.idleShutdownPromise) {
+            return false;
+        }
+        if (!this.child) {
+            return false;
+        }
+        if (this.pendingSessionStarts.size > 0) {
+            return false;
+        }
+        for (const state of this.channels.values()) {
+            if (state.activeTurn || state.queue.length > 0 || state.sessionPromise || state.processing) {
+                return false;
+            }
+        }
+        return true;
+    }
+    async closeBackendProcess(error, gracePeriodMs) {
         const connection = this.connection;
         const child = this.child;
         await this.waitForPendingSessionStarts();
@@ -895,9 +1010,9 @@ export class CopilotCliClient {
             return;
         }
         const childExitPromise = this.childExitPromise ?? Promise.resolve();
-        this.logger?.debug('sending SIGTERM to Copilot ACP process');
+        this.logger?.debug('sending SIGTERM to Copilot ACP process', { gracePeriodMs });
         child.kill('SIGTERM');
-        const exitedAfterTerm = await this.waitForChildExit(childExitPromise, this.runtime.shutdownGracePeriodMs);
+        const exitedAfterTerm = await this.waitForChildExit(childExitPromise, gracePeriodMs);
         if (exitedAfterTerm) {
             this.logger?.debug('Copilot ACP process exited after SIGTERM');
             this.child = undefined;