@agent-relay/factory 0.1.48 → 0.1.50

package/dist/orchestrator/factory.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"factory.d.ts","sourceRoot":"","sources":["../../src/orchestrator/factory.ts"],"names":[],"mappings":"AAIA,OAAO,EAAuB,KAAK,aAAa,EAAE,MAAM,kBAAkB,CAAA;AAI1E,OAAO,KAAK,EAMV,WAAW,EASX,WAAW,EAKZ,MAAM,UAAU,CAAA;AAiCjB,OAAO,KAAK,EACV,cAAc,EACd,OAAO,EACP,mBAAmB,EACnB,YAAY,EACZ,aAAa,EACb,mBAAmB,EAEnB,qBAAqB,EACrB,oBAAoB,EACpB,mBAAmB,EAGnB,QAAQ,EACR,eAAe,EACf,WAAW,EAGX,cAAc,EAEf,MAAM,UAAU,CAAA;AA4BjB,KAAK,YAAY,GAAG,cAAc,GAAG,YAAY,GAAG,YAAY,GAAG,oBAAoB,GAAG,OAAO,CAAA;AACjG,KAAK,QAAQ,GAAG,CAAC,OAAO,EAAE,mBAAmB,KAAK,IAAI,CAAA;AAoFtD,KAAK,iBAAiB,GAAG;IACvB,KAAK,EAAE,MAAM,CAAA;IACb,QAAQ,EAAE,MAAM,CAAA;IAChB,IAAI,EAAE,MAAM,CAAA;IACZ,MAAM,EAAE,MAAM,CAAA;IACd,KAAK,EAAE,MAAM,CAAA;IACb,IAAI,EAAE,MAAM,CAAA;IACZ,GAAG,EAAE,MAAM,CAAA;IACX,KAAK,EAAE,MAAM,CAAA;IACb,MAAM,EAAE,MAAM,EAAE,CAAA;IAChB,MAAM,CAAC,EAAE,MAAM,CAAA;IACf,IAAI,EAAE,MAAM,CAAA;IACZ,GAAG,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;CAC7B,CAAA;AAgGD,eAAO,MAAM,mCAAmC,iDAAiD,CAAA;AACjG,eAAO,MAAM,kCAAkC,gDAAgD,CAAA;AAa/F,wBAAgB,aAAa,CAAC,MAAM,EAAE,aAAa,EAAE,KAAK,EAAE,YAAY,GAAG,OAAO,CAEjF;AAED,qBAAa,WAAY,YAAW,OAAO;;gBAiK7B,MAAM,EAAE,aAAa,EAAE,KAAK,EAAE,YAAY;IAkIhD,uBAAuB,CAAC,KAAK,EAAE,QAAQ,GAAG,OAAO,CAAC,IAAI,CAAC;IAiEvD,KAAK,CAAC,IAAI,GAAE,mBAAwB,GAAG,OAAO,CAAC,IAAI,CAAC;IAsHpD,IAAI,IAAI,OAAO,CAAC,IAAI,CAAC;IAkJrB,OAAO,IAAI,OAAO,CAAC,IAAI,CAAC;IA8mBxB,OAAO,CAAC,IAAI,GAAE;QAAE,MAAM,CAAC,EAAE,OAAO,CAAA;KAAO,GAAG,OAAO,CAAC,eAAe,CAAC;IAkoBlE,OAAO,CAAC,IAAI,GAAE,qBAA0B,GAAG,OAAO,CAAC,eAAe,EAAE,CAAC;IAyDrE,WAAW,CAAC,KAAK,EAAE,WAAW,GAAG,OAAO,CAAC,cAAc,CAAC;IAOxD,QAAQ,CAAC,QAAQ,EAAE,cAAc,EAAE,IAAI,GAAE;QAAE,MAAM,CAAC,EAAE,OAAO,CAAC;QAAC,eAAe,CAAC,EAAE,OAAO,CAAA;KAAO,GAAG,OAAO,CAAC,cAAc,CAAC;IAsT7H,MAAM,IAAI,aAAa;IAiBvB,EAAE,CAAC,KAAK,EAAE,YAAY,EAAE,QAAQ,EAAE,QAAQ,GAAG,MAAM,IAAI;CAolRxD;AAED,wBAAgB,gBAAgB,CAAC,IAAI,EAAE,MAAM,EAAE,OAAO,EAAE,OAAO,GAAG,WAAW,CAiC5E;AAyBD,wBAAsB,oCAAoC,CACxD,KAAK,EAAE,IAAI,CAAC,WAAW,EAAE,UAAU,CAAC,EACpC,IAAI,EAAE,MAAM,GACX,OAAO,CAAC,WAAW,CAAC,CAetB;AAED,wBAAgB,gBAAgB,CAAC,IAAI,EAAE,MAAM,EAAE,OAAO,EAAE,OAAO,GAAG,iBAAiB,CA6BlF;AAED,wBAAgB,uBAAuB,CAAC,IAAI,EAAE,MAAM,EAAE,OAAO,EAAE,OAAO,GAAG,WAAW,CAEnF;AAoDD,wBAAsB,wBAAwB,CAC5C,IAAI,SAAsC,GACzC,OAAO,CAAC,oBAAoB,GAAG,SAAS,CAAC,CAM3C;AAED,wBAAgB,wBAAwB,CACtC,SAAS,EAAE,oBAAoB,GAAG,SAAS,EAC3C,IAAI,GAAE;IAAE,KAAK,CAAC,EAAE,MAAM,CAAC;IAAC,OAAO,CAAC,EAAE,MAAM,CAAA;CAAO,GAC9C,mBAAmB,CAgBrB;AAED,wBAAgB,iBAAiB,CAAC,KAAK,EAAE,WAAW,GAAG,OAAO,CAO7D;AAED,wBAAgB,mBAAmB,CAAC,KAAK,EAAE,WAAW,GAAG,OAAO,CAiB/D;AAglBD,eAAO,MAAM,oBAAoB,GAAI,MAAM,MAAM,KAAG;IAAE,KAAK,EAAE,MAAM,CAAC;IAAC,IAAI,EAAE,MAAM,CAAC;IAAC,MAAM,EAAE,MAAM,CAAC;IAAC,IAAI,CAAC,EAAE,MAAM,CAAA;CAAE,GAAG,SAwBpH,CAAA;AA0ZD,eAAO,MAAM,2BAA2B,GAAI,QAAQ,aAAa,KAAG,MAAM,EAItE,CAAA;AA2mBJ,eAAO,MAAM,WAAW,GAAI,MAAM,MAAM,KAAG,MAC4B,CAAA;AAkUvE,eAAO,MAAM,eAAe,GAAI,OAAO,WAAW,KAAG,MAAM,GAAG,SAG7D,CAAA"}
+{"version":3,"file":"factory.d.ts","sourceRoot":"","sources":["../../src/orchestrator/factory.ts"],"names":[],"mappings":"AAIA,OAAO,EAAuB,KAAK,aAAa,EAAE,MAAM,kBAAkB,CAAA;AAI1E,OAAO,KAAK,EAMV,WAAW,EASX,WAAW,EAKZ,MAAM,UAAU,CAAA;AAiCjB,OAAO,KAAK,EACV,cAAc,EACd,OAAO,EACP,mBAAmB,EACnB,YAAY,EACZ,aAAa,EACb,mBAAmB,EAEnB,qBAAqB,EACrB,oBAAoB,EACpB,mBAAmB,EAGnB,QAAQ,EACR,eAAe,EACf,WAAW,EAGX,cAAc,EAEf,MAAM,UAAU,CAAA;AA4BjB,KAAK,YAAY,GAAG,cAAc,GAAG,YAAY,GAAG,YAAY,GAAG,oBAAoB,GAAG,OAAO,CAAA;AACjG,KAAK,QAAQ,GAAG,CAAC,OAAO,EAAE,mBAAmB,KAAK,IAAI,CAAA;AAoFtD,KAAK,iBAAiB,GAAG;IACvB,KAAK,EAAE,MAAM,CAAA;IACb,QAAQ,EAAE,MAAM,CAAA;IAChB,IAAI,EAAE,MAAM,CAAA;IACZ,MAAM,EAAE,MAAM,CAAA;IACd,KAAK,EAAE,MAAM,CAAA;IACb,IAAI,EAAE,MAAM,CAAA;IACZ,GAAG,EAAE,MAAM,CAAA;IACX,KAAK,EAAE,MAAM,CAAA;IACb,MAAM,EAAE,MAAM,EAAE,CAAA;IAChB,MAAM,CAAC,EAAE,MAAM,CAAA;IACf,IAAI,EAAE,MAAM,CAAA;IACZ,GAAG,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;CAC7B,CAAA;AAiGD,eAAO,MAAM,mCAAmC,iDAAiD,CAAA;AACjG,eAAO,MAAM,kCAAkC,gDAAgD,CAAA;AAa/F,wBAAgB,aAAa,CAAC,MAAM,EAAE,aAAa,EAAE,KAAK,EAAE,YAAY,GAAG,OAAO,CAEjF;AAED,qBAAa,WAAY,YAAW,OAAO;;gBAoK7B,MAAM,EAAE,aAAa,EAAE,KAAK,EAAE,YAAY;IAkIhD,uBAAuB,CAAC,KAAK,EAAE,QAAQ,GAAG,OAAO,CAAC,IAAI,CAAC;IAiEvD,KAAK,CAAC,IAAI,GAAE,mBAAwB,GAAG,OAAO,CAAC,IAAI,CAAC;IAsHpD,IAAI,IAAI,OAAO,CAAC,IAAI,CAAC;IAmJrB,OAAO,IAAI,OAAO,CAAC,IAAI,CAAC;IA8mBxB,OAAO,CAAC,IAAI,GAAE;QAAE,MAAM,CAAC,EAAE,OAAO,CAAA;KAAO,GAAG,OAAO,CAAC,eAAe,CAAC;IAkoBlE,OAAO,CAAC,IAAI,GAAE,qBAA0B,GAAG,OAAO,CAAC,eAAe,EAAE,CAAC;IAyDrE,WAAW,CAAC,KAAK,EAAE,WAAW,GAAG,OAAO,CAAC,cAAc,CAAC;IAOxD,QAAQ,CAAC,QAAQ,EAAE,cAAc,EAAE,IAAI,GAAE;QAAE,MAAM,CAAC,EAAE,OAAO,CAAC;QAAC,eAAe,CAAC,EAAE,OAAO,CAAA;KAAO,GAAG,OAAO,CAAC,cAAc,CAAC;IAsT7H,MAAM,IAAI,aAAa;IAiBvB,EAAE,CAAC,KAAK,EAAE,YAAY,EAAE,QAAQ,EAAE,QAAQ,GAAG,MAAM,IAAI;CAsqRxD;AAED,wBAAgB,gBAAgB,CAAC,IAAI,EAAE,MAAM,EAAE,OAAO,EAAE,OAAO,GAAG,WAAW,CAiC5E;AAyBD,wBAAsB,oCAAoC,CACxD,KAAK,EAAE,IAAI,CAAC,WAAW,EAAE,UAAU,CAAC,EACpC,IAAI,EAAE,MAAM,GACX,OAAO,CAAC,WAAW,CAAC,CAetB;AAED,wBAAgB,gBAAgB,CAAC,IAAI,EAAE,MAAM,EAAE,OAAO,EAAE,OAAO,GAAG,iBAAiB,CA6BlF;AAED,wBAAgB,uBAAuB,CAAC,IAAI,EAAE,MAAM,EAAE,OAAO,EAAE,OAAO,GAAG,WAAW,CAEnF;AAoDD,wBAAsB,wBAAwB,CAC5C,IAAI,SAAsC,GACzC,OAAO,CAAC,oBAAoB,GAAG,SAAS,CAAC,CAM3C;AAED,wBAAgB,wBAAwB,CACtC,SAAS,EAAE,oBAAoB,GAAG,SAAS,EAC3C,IAAI,GAAE;IAAE,KAAK,CAAC,EAAE,MAAM,CAAC;IAAC,OAAO,CAAC,EAAE,MAAM,CAAA;CAAO,GAC9C,mBAAmB,CAgBrB;AAED,wBAAgB,iBAAiB,CAAC,KAAK,EAAE,WAAW,GAAG,OAAO,CAO7D;AAED,wBAAgB,mBAAmB,CAAC,KAAK,EAAE,WAAW,GAAG,OAAO,CAiB/D;AAglBD,eAAO,MAAM,oBAAoB,GAAI,MAAM,MAAM,KAAG;IAAE,KAAK,EAAE,MAAM,CAAC;IAAC,IAAI,EAAE,MAAM,CAAC;IAAC,MAAM,EAAE,MAAM,CAAC;IAAC,IAAI,CAAC,EAAE,MAAM,CAAA;CAAE,GAAG,SAwBpH,CAAA;AA0ZD,eAAO,MAAM,2BAA2B,GAAI,QAAQ,aAAa,KAAG,MAAM,EAItE,CAAA;AA2mBJ,eAAO,MAAM,WAAW,GAAI,MAAM,MAAM,KAAG,MAC4B,CAAA;AAkUvE,eAAO,MAAM,eAAe,GAAI,OAAO,WAAW,KAAG,MAAM,GAAG,SAG7D,CAAA"}

package/dist/orchestrator/factory.js

@@ -90,6 +90,7 @@ const DISPATCH_LIFECYCLE_LEASE_MS = 5 * 60_000;
 const DISPATCH_LIFECYCLE_RENEW_MS = 60_000;
 const DISPATCH_LIFECYCLE_RETRY_MS = 1_000;
 const STARTUP_AGENT_EXIT_DRAIN_TIMEOUT_MS = 30_000;
+const RECONCILED_AGENT_EXIT_CONCURRENCY = 4;
 const SLACK_EVENT_WATERMARK_CACHE_MS = 60_000;
 const SLACK_CONVERSATION_TURN_LEASE_MS = 60_000;
 const SLACK_CONVERSATION_TURN_RETRY_MS = 1_000;
@@ -207,6 +208,7 @@ export class FactoryLoop {
     #dispatchTerminalWaiters = new Map();
     #dispatchLifecycleRetryTimers = new Map();
     #dispatchLifecycleDrives = new Set();
+    #abandonedDispatchReasons = new Map();
     #dispatchLifecycleCapacityWaitLogged = new Set();
     #dispatchLifecycleOwnershipWaitLogged = new Set();
     #localReleaseCheckpoints = new Map();
@@ -246,6 +248,8 @@ export class FactoryLoop {
     #completionSweepActive = false;
     #completionInFlight = new Set();
     #agentExitsInFlight = new Map();
+    #reconciledAgentExitsActive = 0;
+    #reconciledAgentExitWaiters = [];
     #agentLifecycleSignalsInFlight = new Map();
     #startupAgentAdoptionActive = false;
     #startupRosterExitSignals;
@@ -608,6 +612,7 @@ export class FactoryLoop {
         for (const timer of this.#dispatchLifecycleRetryTimers.values())
             clearTimeout(timer);
         this.#dispatchLifecycleRetryTimers.clear();
+        this.#abandonedDispatchReasons.clear();
         this.#dispatchLifecycleOwnershipWaitLogged.clear();
         if (this.#completionSweepTimer)
             clearTimeout(this.#completionSweepTimer);
@@ -2341,7 +2346,13 @@ export class FactoryLoop {
         const previous = this.#agentExitsInFlight.get(name) ?? Promise.resolve();
         const handling = previous
             .catch(() => undefined)
-            .then(async () => await this.#handleAgentExit(name, reason))
+            .then(async () => {
+            if (reason === 'reconciled-missing') {
+                await this.#withReconciledAgentExitSlot(async () => this.#handleAgentExit(name, reason));
+                return;
+            }
+            await this.#handleAgentExit(name, reason);
+        })
             .catch((error) => this.#error(error))
             .finally(() => {
             if (this.#agentExitsInFlight.get(name) === handling) {
@@ -2376,7 +2387,7 @@ export class FactoryLoop {
                     this.#scheduleDispatchLifecycleRetry(inFlightRecordFromLifecycle(claim.lifecycle));
                     continue;
                 }
-                this.#dispatchLifecycleEpochs.set(key, claim.lease.epoch);
+                this.#dispatchLifecycleEpochs.set(claim.key ?? key, claim.lease.epoch);
                 if (claim.lifecycle.phase === 'waiting-for-human')
                     continue;
                 const durableRecord = inFlightRecordFromLifecycle(claim.lifecycle);
@@ -2448,7 +2459,12 @@ export class FactoryLoop {
                 // the hydrated durable session is no longer usable. Classify it as a
                 // reconciled miss so remote implementers recover a PR or restart
                 // instead of waiting forever for branch replication after a crash.
-                for (const name of signalled)
+                const rolePriority = (name) => {
+                    const role = batch.getIssueByAgent(name)?.agents.get(name)?.spec.role;
+                    return role === 'implementer' ? 0 : role === 'babysitter' ? 1 : 2;
+                };
+                const orderedSignals = [...signalled].sort((left, right) => rolePriority(left) - rolePriority(right) || left.localeCompare(right));
+                for (const name of orderedSignals)
                     this.#queueAgentExit(name, 'reconciled-missing');
                 if (synthesized.length > 0) {
                     this.#increment('startupRosterMissingExitsSynthesized', synthesized.length);
@@ -2566,6 +2582,20 @@ export class FactoryLoop {
             clearTimeout(timer);
         return completed;
     }
+    async #withReconciledAgentExitSlot(run) {
+        if (this.#reconciledAgentExitsActive >= RECONCILED_AGENT_EXIT_CONCURRENCY) {
+            this.#increment('reconciledAgentExitBackpressure');
+            await new Promise((resolve) => this.#reconciledAgentExitWaiters.push(resolve));
+        }
+        this.#reconciledAgentExitsActive += 1;
+        try {
+            await run();
+        }
+        finally {
+            this.#reconciledAgentExitsActive -= 1;
+            this.#reconciledAgentExitWaiters.shift()?.();
+        }
+    }
     #scheduleDispatchLifecycleRenewal() {
         if (this.#dispatchLifecycleRenewTimer || this.#dispatchLifecycleEpochs.size === 0)
             return;
@@ -2620,7 +2650,7 @@ export class FactoryLoop {
                 : `dispatch lifecycle is owned by ${claim.lifecycle.lease?.owner ?? 'another publisher'}`;
             throw new Error(`Refusing to dispatch ${decision.issue.key}: ${reason}`);
         }
-        this.#dispatchLifecycleEpochs.set(key, claim.lease.epoch);
+        this.#dispatchLifecycleEpochs.set(claim.key ?? key, claim.lease.epoch);
         this.#scheduleDispatchLifecycleRenewal();
         if (claim.created) {
             await this.#reportLifecycle(claim.lifecycle, 'run.started');
@@ -2845,7 +2875,7 @@ export class FactoryLoop {
             if (!claim.acquired || !claim.lease) {
                 throw new DispatchLifecycleOwnedElsewhereError(claim.lifecycle.lease?.leaseUntilMs);
             }
-            this.#dispatchLifecycleEpochs.set(key, claim.lease.epoch);
+            this.#dispatchLifecycleEpochs.set(claim.key ?? key, claim.lease.epoch);
             this.#scheduleDispatchLifecycleRenewal();
             lifecycle = claim.lifecycle;
             acquiredNow = true;
@@ -2893,6 +2923,11 @@ export class FactoryLoop {
             return;
         if (acquiredNow && this.#config.babysitter.enabled)
             await this.#restoreBabysitterOwnership();
+        const abandonedReason = this.#abandonedDispatchReasons.get(key);
+        if (abandonedReason !== undefined) {
+            await this.#abandonStuckDispatch(record, abandonedReason);
+            return;
+        }
         if (acquiredNow && lifecycle.phase === 'running') {
             if (this.#fleet.hydrateTracked) {
                 this.#fleet.hydrateTracked(lifecycle.agents.map((agent) => ({
@@ -4567,6 +4602,7 @@ export class FactoryLoop {
         if (isCompletionReason(reason)) {
             if (exiting?.spec.role === 'implementer' && await this.#issueHasCompletionPr(record, {
                 openOnly: this.#config.babysitter.enabled,
+                preferExactBranch: tracingReconciledExit,
             }, exiting)) {
                 if (this.#config.babysitter.enabled)
                     await this.#ensureBabysitterForIssue(record);
@@ -4624,6 +4660,7 @@ export class FactoryLoop {
             const hasCompletionPr = tracked.spec.role === 'implementer'
                 ? await this.#issueHasCompletionPr(record, {
                     openOnly: this.#config.babysitter.enabled,
+                    preferExactBranch: tracingReconciledExit,
                 }, tracked)
                 : false;
             if (tracingReconciledExit) {
@@ -4729,30 +4766,43 @@ export class FactoryLoop {
                     return;
                 }
             }
-            let recovered = false;
-            if (tracked.sessionRef) {
-                const resumeKey = `${issueKey(record.issue)}:${name}:${tracked.sessionRef}`;
-                if (await this.#state.isResumed(this.#workspaceId, resumeKey)) {
-                    // Already resumed once and STILL exiting with no completion PR — the
-                    // agent isn't making progress. Conclude the dispatch so a human
-                    // notices AND the reviewer waiting on the implementer's DM is torn
-                    // down, instead of leaving the issue silently in-flight forever with
-                    // a live reviewer stalling the owned-broker dispose-wait (#67).
-                    if (tracked.spec.role === 'implementer') {
-                        await this.#concludeTerminalImplementer(record, name, 'stalled-no-pr');
-                    }
-                    return;
+            // Bound recovery by the durable logical agent, not by the session ref.
+            // Harnesses may return a fresh session ref from every successful resume,
+            // and no-sessionRef workers use respawn instead. Keying only resumptions
+            // by the changing ref (and not recording respawns at all) let an agent
+            // that exited immediately consume a batch slot forever while Factory
+            // continuously recreated it. One successful recovery is enough; a
+            // subsequent no-PR implementer exit is terminal and frees the slot.
+            const recoveryLifecycle = await this.#state.getDispatchLifecycle(this.#workspaceId, issueKey(record.issue));
+            const recoveryRunIdentity = recoveryLifecycle?.runId
+                ?? tracked.spec.branch
+                ?? batch.invocationIdFor(record.issue, tracked.spec);
+            const recoveryKey = `${recoveryRunIdentity}:${tracked.spec.name}`;
+            if (await this.#state.isResumed(this.#workspaceId, recoveryKey)) {
+                if (tracked.spec.role === 'implementer') {
+                    await this.#concludeTerminalImplementer(record, name, 'stalled-no-pr');
                 }
-                const existing = this.#resumeInFlight.get(resumeKey);
-                if (existing) {
+                return;
+            }
+            const existing = this.#resumeInFlight.get(recoveryKey);
+            if (existing) {
+                try {
                     await existing;
-                    return;
                 }
+                catch {
+                    // The initiating recovery handler owns classification and logging.
+                    // Followers only coalesce onto its lifetime and must not re-report
+                    // the same failure.
+                }
+                return;
+            }
+            let recovered = false;
+            if (tracked.sessionRef) {
                 const resume = this.#resumeTrackedAgent(record, name, tracked);
-                this.#resumeInFlight.set(resumeKey, resume);
+                this.#resumeInFlight.set(recoveryKey, resume);
                 try {
                     await resume;
-                    await this.#state.markResumed(this.#workspaceId, resumeKey);
+                    await this.#state.markResumed(this.#workspaceId, recoveryKey);
                     recovered = true;
                 }
                 catch (error) {
@@ -4760,11 +4810,11 @@ export class FactoryLoop {
                         // The broker never released this agent's name on exit
                         // (relay#1116-family), so re-registering collides with the stuck
                         // name. Retrying just re-collides forever. Treat it as terminal for
-                        // this name: record the resume key so subsequent exit events
+                        // this name: record the recovery key so subsequent exit events
                         // short-circuit, count it, and warn once. The external reaper / a
                         // broker restart reclaims the leaked name.
                         this.#fleet.markAgentTerminal?.(name, 'resume-already-exists');
-                        await this.#state.markResumed(this.#workspaceId, resumeKey);
+                        await this.#state.markResumed(this.#workspaceId, recoveryKey);
                         this.#increment('resumeNameCollisions');
                         this.#logger.warn?.('[factory] resume skipped: broker still holds agent name (relay#1116); not retrying', {
                             issue: record.issue.key,
@@ -4783,12 +4833,12 @@ export class FactoryLoop {
                     }
                 }
                 finally {
-                    this.#resumeInFlight.delete(resumeKey);
+                    this.#resumeInFlight.delete(recoveryKey);
                 }
             }
             else {
                 const invocationId = `${batch.invocationIdFor(record.issue, tracked.spec)}:restart:${this.#clock.now()}`;
-                try {
+                const respawn = (async () => {
                     await this.#prepareAgentWorktree(record, tracked.spec);
                     const result = await this.#fleet.spawn({
                         name: tracked.spec.name,
@@ -4804,6 +4854,11 @@ export class FactoryLoop {
                         channel: tracked.spec.channel,
                     });
                     batch.recordSpawn(record, tracked.spec, invocationId, result);
+                })();
+                this.#resumeInFlight.set(recoveryKey, respawn);
+                try {
+                    await respawn;
+                    await this.#state.markResumed(this.#workspaceId, recoveryKey);
                     recovered = true;
                 }
                 catch (error) {
@@ -4817,6 +4872,7 @@ export class FactoryLoop {
                     // the dispatch so a reviewer blocked on the implementer's DM does not
                     // hang the owned-broker dispose-wait (#67).
                     this.#fleet.markAgentTerminal?.(name, 'respawn-already-exists');
+                    await this.#state.markResumed(this.#workspaceId, recoveryKey);
                     this.#increment('resumeNameCollisions');
                     this.#logger.warn?.('[factory] respawn skipped: broker still holds agent name (relay#1116); not retrying', {
                         issue: record.issue.key,
@@ -4826,6 +4882,9 @@ export class FactoryLoop {
                         await this.#concludeTerminalImplementer(record, name, 'respawn-already-exists');
                     }
                 }
+                finally {
+                    this.#resumeInFlight.delete(recoveryKey);
+                }
             }
             if (recovered && tracked.spec.role === 'implementer') {
                 await this.#writeInFlightRegistry();
@@ -5302,6 +5361,8 @@ export class FactoryLoop {
     // the release-driven exit event so it cannot re-trigger a resume before the
     // record leaves the batch.
     async #abandonStuckDispatch(record, reason) {
+        const key = issueKey(record.issue);
+        this.#abandonedDispatchReasons.set(key, reason);
         const agents = [...record.agents];
         for (const [agentName, tracked] of agents) {
             if (tracked.spec.role === 'implementer')
@@ -5334,6 +5395,22 @@ export class FactoryLoop {
             await this.#writeInFlightRegistry();
             return;
         }
+        // Batch completion alone only frees the process-local slot. Durable
+        // capacity is computed from lifecycle phases, so leaving this row in
+        // `publishing`/`running` makes every queued issue wait forever even though
+        // all agents and worktrees are already gone. Fence the terminal phase
+        // before promoting the next issue.
+        if (!await this.#saveDispatchLifecycle(record, 'abandoned', undefined, reason)) {
+            this.#increment('abandonedDispatchReleaseRetries');
+            // #saveDispatchLifecycle already schedules the generic durable retry. The
+            // pending reason makes that retry re-acquire ownership and return here,
+            // rather than merely restoring the old running lifecycle and leaking the
+            // slot. The abandoned-specific scheduler remains the non-durable fallback.
+            this.#scheduleAbandonedDispatchRetry(record, reason);
+            await this.#writeInFlightRegistry();
+            return;
+        }
+        this.#abandonedDispatchReasons.delete(key);
         await this.#recordDispatchTerminal(record.issue);
         const next = (await this.#batch()).complete(record.issue);
         await this.#drainReadyClarificationWake();
@@ -5367,7 +5444,12 @@ export class FactoryLoop {
             if (!issue) {
                 return false;
             }
-            if (implementer?.spec.branch && record.decision.implementers.length > 1) {
+            // During startup recovery, the implementer's deterministic branch is the
+            // strongest and cheapest lookup. Avoid scanning mounted PR metadata across
+            // every configured repository. Normal event-driven exits keep using the
+            // webhook-fed mount so they do not introduce a GitHub API dependency.
+            if (implementer?.spec.branch &&
+                (opts.preferExactBranch || record.decision.implementers.length > 1)) {
                 const sourceOwner = record.issue.path
                     ? githubIssuePathParts(record.issue.path)?.owner
                     : undefined;