crewly 1.20.294 → 1.20.296

package/dist/backend/backend/src/index.js

@@ -77,7 +77,7 @@ import { parseSlackThreadKey } from './services/slack/slack-thread-key.js';
 import { LIVENESS_MONITOR_CONSTANTS, INPUT_CIRCUIT_CONSTANTS, INPUT_BLOCKED_RETRY_CONSTANTS } from './constants.js';
 import { InputBlockedRetryService } from './services/messaging/input-blocked-retry.service.js';
 import { driveCapabilities } from './services/drive/drive-agent.service.js';
-import { SUB_AGENT_QUEUE_CONSTANTS, CHAT_CONTEXT_CONSTANTS, SAFE_RESTART, AUTO_UPDATE_CONSTANTS, PROCESS_EXIT_CODES, CLAUDE_STARTUP_CONSTANTS, WEB_CONSTANTS, TICKET_CONSTANTS, UNASSIGNED_ROUTE_CONSTANTS, CLOUD_TALK_CONSTANTS, STANDING_ANSWERS_CONSTANTS, TICKET_AUTOPILOT_CONSTANTS, TICKET_HYGIENE_CONSTANTS, EXPERIMENT_CONSTANTS, WORK_ITEM_DESTINATION_CONSTANTS, CODEX_USAGE_SYNC_CONSTANTS, ANTIGRAVITY_USAGE_SYNC_CONSTANTS, OWNER_AUTH_CONSTANTS, CREWLY_APPS_CONSTANTS, SLACK_AGENT_DM_CONSTANTS, BRIEFING_CONSTANTS, DRIVE_CONSTANTS } from './constants.js';
+import { SUB_AGENT_QUEUE_CONSTANTS, CHAT_CONTEXT_CONSTANTS, SAFE_RESTART, AUTO_UPDATE_CONSTANTS, PROCESS_EXIT_CODES, CLAUDE_STARTUP_CONSTANTS, WEB_CONSTANTS, TICKET_CONSTANTS, UNASSIGNED_ROUTE_CONSTANTS, CLOUD_TALK_CONSTANTS, STANDING_ANSWERS_CONSTANTS, TICKET_AUTOPILOT_CONSTANTS, TICKET_HYGIENE_CONSTANTS, EXPERIMENT_CONSTANTS, WORK_ITEM_DESTINATION_CONSTANTS, CODEX_USAGE_SYNC_CONSTANTS, ANTIGRAVITY_USAGE_SYNC_CONSTANTS, OWNER_AUTH_CONSTANTS, CREWLY_APPS_CONSTANTS, SLACK_AGENT_DM_CONSTANTS, BRIEFING_CONSTANTS, DRIVE_CONSTANTS, POOL_PRUNE_CONSTANTS } from './constants.js';
 import { randomUUID } from 'crypto';
 import { PtyActivityTrackerService } from './services/agent/pty-activity-tracker.service.js';
 import { InFlightTurnTracker } from './services/restart/in-flight-turn-tracker.service.js';
@@ -88,6 +88,7 @@ import { initialDecider, nextDecider } from './services/task-pool/untargeted-rou
 import { activeAcceptance, formatTicketMarker, formatTicketNumber } from './types/v2/ticket.types.js';
 import { createWorkItem, TERMINAL_WORK_ITEM_STATUSES } from './types/v2/work-item.types.js';
 import { runPoolArchiveMigration } from './services/task-pool/pool-archive-migration.js';
+import { runPoolPrune } from './services/task-pool/pool-prune.js';
 import { createPtyTurnProbe } from './services/restart/turn-probe.js';
 import { RestartDrainService, resolveRestartDrainMs, resolveBackgroundDrainMs, } from './services/restart/restart-drain.service.js';
 import { interruptedTurnsPath, loadInterruptedTurns, saveInterruptedTurns, writeInterruptedTurns, resumeInterruptedTurns, planCommitmentNotes, } from './services/restart/interrupted-turns.js';
@@ -185,6 +186,7 @@ import { TaskHistorySubscriber } from './services/memory/task-history.subscriber
 import { TeamHealthWatchdogService, LiveTeamHealthDataProvider, loadTeamHealthConfig, setTeamHealthWatchdogSingleton, getTeamHealthWatchdogSingleton, } from './services/team-health/index.js';
 import { traceHarness, traceTurnActivity } from './services/trace/trace-recorder.js';
 import { inputCircuitStats, setInputCircuitOpenListener } from './services/session/input-circuit-breaker.js';
+import { looksLikeShellPrompt } from './utils/screen-excerpt.js';
 import { harnessPastesSinceOutsideInput, shownMarkers } from './services/session/input-ledger.js';
 // ESM __dirname equivalent using import.meta.url
 const __filename = fileURLToPath(import.meta.url);
@@ -764,13 +766,25 @@ export class CrewlyServer {
         // Ticket loop §4: one-time archive of old WorkItems (marker file makes
         // it run once). Through the pool's own storage cache, so no later
         // flush can write the archived items back. Never deletes.
+        const poolPruneLogger = LoggerService.getInstance().createComponentLogger('PoolPrune');
+        const pruneOnce = () => runPoolPrune(TaskPoolService.getInstance().getStorage(), { logger: poolPruneLogger }).catch((pruneErr) => {
+            this.logger.warn('Task pool prune failed (pool left as is; retried next interval)', {
+                error: pruneErr instanceof Error ? pruneErr.message : String(pruneErr),
+            });
+        });
         void runPoolArchiveMigration(TaskPoolService.getInstance().getStorage(), {
             logger: LoggerService.getInstance().createComponentLogger('PoolArchive'),
-        }).catch((archiveErr) => {
+        })
+            .catch((archiveErr) => {
             this.logger.warn('Task pool one-time archive failed (pool left as is; retried next boot)', {
                 error: archiveErr instanceof Error ? archiveErr.message : String(archiveErr),
             });
-        });
+        })
+            // Then keep it small: finished items and ended claims move to the
+            // archive, once now and every few hours (crewly#1108).
+            .then(pruneOnce);
+        const poolPruneTimer = setInterval(() => void pruneOnce(), POOL_PRUNE_CONSTANTS.INTERVAL_MS);
+        poolPruneTimer.unref?.();
         // Atlas 2026-05-23 fix: wire the agent-liveness gate so claimFromPool /
         // claimSpecificItem refuse to put a WI into `running` when the requesting
         // agent's session is dead. delegate-task's "self-heal fix #1" used to
@@ -2417,7 +2431,20 @@ export class CrewlyServer {
                     activateAgent: async (agentSession) => {
                         const { activateAgentBySession } = await import('./controllers/team/team.controller.js');
                         const res = await activateAgentBySession(this.apiController, agentSession);
-                        return res.success;
+                        // A start waiting for a running-agent slot is deferred, not failed.
+                        return res.success ? true : res.deferred ? 'deferred' : false;
+                    },
+                    // The message waits on the agent's persisted queue for a deferred start.
+                    holdForStart: (agentSession, prompt) => {
+                        SubAgentMessageQueue.getInstance().enqueue(agentSession, prompt);
+                        // Nothing else starts the agent when a slot frees: retry until it does.
+                        void import('./services/messaging/deferred-start-retry.js').then(async ({ scheduleDeferredStartRetry }) => {
+                            const { activateAgentBySession } = await import('./controllers/team/team.controller.js');
+                            await scheduleDeferredStartRetry(agentSession, {
+                                activate: (name) => activateAgentBySession(this.apiController, name),
+                                hasPending: (name) => SubAgentMessageQueue.getInstance().hasPending(name),
+                            });
+                        });
                     },
                     // Every owner message that reached an agent is watched until
                     // it is answered (specs/2026-09-30-owner-message-guarantee.md).
@@ -6370,6 +6397,17 @@ export class CrewlyServer {
             };
             // Every delivery refused for a while: automatic redelivery is slowed
             // to a probe now and then, and the owner told once (crewly#1028).
+            // What the unreadable screen shows (redacted rows), so the owner can
+            // tell a dialog from a bare shell without opening the terminal.
+            const screenNote = (sessionName, state) => {
+                if (state !== 'unknown')
+                    return '';
+                const excerpt = SessionCommandHelper.unreadableExcerpt(sessionName);
+                if (!excerpt)
+                    return '';
+                const hint = looksLikeShellPrompt(excerpt) ? ' It looks like the runtime exited to a shell prompt.' : '';
+                return `${hint}\nScreen (secrets masked):\n${excerpt.split('\n').map((l) => `> ${l}`).join('\n')}`;
+            };
             setInputCircuitOpenListener((info) => {
                 traceHarness('guard.block', {
                     session: info.sessionName,
@@ -6422,7 +6460,7 @@ export class CrewlyServer {
                             ? 'its input box cannot be read (a dialog or an unfamiliar screen)'
                             : `its input box holds ${notice.inputLength} characters that Crewly did not write`;
                         const maxMin = Math.round(INPUT_CIRCUIT_CONSTANTS.PROBE_MAX_MS / 60000);
-                        tell(notice.sessionName, `Needs you: ${notice.sessionName} has not received any message for ${minutes} min: ${what}, so Crewly refused to type over it (${notice.refusals} attempts). Crewly has stopped retrying every few seconds and now tries again only every few minutes (at most every ${maxMin} min). To fix it, open ${notice.sessionName}'s terminal and clear the input box (Ctrl+U), or press Enter if that text is yours. Delivery resumes on its own after that.`, notice.message || undefined);
+                        tell(notice.sessionName, `Needs you: ${notice.sessionName} has not received any message for ${minutes} min: ${what}, so Crewly refused to type over it (${notice.refusals} attempts). Crewly has stopped retrying every few seconds and now tries again only every few minutes (at most every ${maxMin} min). To fix it, open ${notice.sessionName}'s terminal and clear the input box (Ctrl+U), or press Enter if that text is yours. Delivery resumes on its own after that.${screenNote(notice.sessionName, notice.blockedState)}`, notice.message || undefined);
                         return;
                     }
                     if (notice.state === 'busy') {
@@ -6432,7 +6470,7 @@ export class CrewlyServer {
                     const what = notice.state === 'unknown'
                         ? 'its input box cannot be read (a dialog or an unfamiliar screen)'
                         : `its input box holds ${notice.inputLength} characters of text not written by Crewly`;
-                    tell(notice.sessionName, `Needs you: messages to ${notice.sessionName} are waiting because ${what}. Crewly will not type over it. Tried ${notice.refusals} times over ${minutes} min; it keeps retrying. Clear the agent's input (or answer its screen) to let them through.`, notice.message);
+                    tell(notice.sessionName, `Needs you: messages to ${notice.sessionName} are waiting because ${what}. Crewly will not type over it. Tried ${notice.refusals} times over ${minutes} min; it keeps retrying. Clear the agent's input (or answer its screen) to let them through.${screenNote(notice.sessionName, notice.state)}`, notice.message);
                 },
             });
             // A WorkItem brief held by the restart drain gets its hand-over