Let agents idle for updates required by creation cancellation

docs/native-hosts.md

@@ -426,3 +426,13 @@ Docker command must return or time out before cleanup can run. Cancellation
 does not require the original creation to acknowledge completion. The agent
 records a removal tombstone before cleanup so a delayed creation command cannot
 recreate the instance, and resumes interrupted cleanup under the same operation ID.
+
+While cancellation requires a newer agent, the controller pauses all command
+delivery to that host, including background observations and command replays.
+Older agents hold their update lock across a whole command batch; continuing to
+feed them work could otherwise starve the updater. Already delivered work can
+finish and report normally, and heartbeats continue. The next automatic update
+check can install the release once the batch has drained. When a heartbeat
+advertises cancellation support, delivery resumes with cancellation cleanup
+first. This controller-side hold also recovers cancellations already waiting
+on older agents. Hosts without automatic updates need the current installer.

src/lib/server/hosts/index.ts

@@ -319,6 +319,13 @@ export function queueHostOperation(hostId: string, kind: string, payload: Record
 export function pendingHostOperations(hostId: string): HostOperation[] {
     const host = requireHost(hostId);
     if (host.mode === 'disabled' || !host.connected) return [];
+    // Older agents hold activity.lock across the entire delivered batch, including
+    // observations. Filtering only the unsupported cleanup can keep the updater
+    // busy indefinitely. Drain the old batch by withholding ALL further work;
+    // heartbeats and receipts still flow, and the new capability releases this hold.
+    // This must live in the controller so the fix also works before an agent update.
+    if (!host.inventory.capabilities?.includes('instance.cancel-create') && database().prepare(`SELECT 1 FROM ec_host_operations
+        WHERE host_id=? AND kind='instance.cancel-create' AND state IN ('pending','accepted','running','unknown') LIMIT 1`).get(hostId)) return [];
     // Accepted/running/unknown commands keep the same immutable ID. The agent
     // consults its durable journal before acknowledging or executing a replay.
     return database().transaction(() => {

src/lib/server/provisioning/native.ts

@@ -21,6 +21,7 @@ import { sha256Buffer } from '../fleet/files.js';
 import type { StagedFileChange } from '../fleet/staged-files.js';
 import type { LaunchPreview } from '../../deployment-plan.js';
 import type { ProvisioningAction } from '../../provisioning.js';
+import { hostAgentUpdatesEnabled } from '../../hosts.js';
 import { cancelCreation, finishCreationCancellation } from './cancel-creation.js';
 
 export interface NativeSelection { hostId: string; libraryId: string; image?: string; bindAddress?: string; publicAddress?: string; publicPort?: number }
@@ -282,8 +283,10 @@ async function processJob(job: Job, eligibility: (job: Job) => string | null) {
     }
     if (job.operation_id) {
         const operation = getHostOperation(job.operation_id);
-        if (operation?.kind === 'instance.cancel-create' && operation.state === 'pending' && !host.inventory.capabilities?.includes('instance.cancel-create')) {
+        if (operation?.kind === 'instance.cancel-create' && !['completed', 'failed'].includes(operation.state) && !host.inventory.capabilities?.includes('instance.cancel-create')) {
-            hold(job, 'Update the host agent to support creation cancellation. Creation is cancelled; cleanup and port release await the updated agent.'); return;
+            hold(job, hostAgentUpdatesEnabled(host.inventory)
+                ? 'Waiting for the host agent update. New command delivery is paused so its current batch can finish and the automatic updater can run. Cleanup resumes after the updated agent connects.'
+                : 'Update the host agent to support creation cancellation: rerun the current installer on this host. Automatic updates are not enabled. New command delivery is paused; cleanup and port release await the updated agent.'); return;
         }
         if (!operation || operation.state === 'unknown') { hold(job, 'Agent result is unknown. Reconciling the original operation; no replacement or repeated command will be sent.'); return; }
         if (operation.state === 'failed') {

tests/unit/native-hosts.test.ts

@@ -136,6 +136,43 @@ describe('native host readiness and reservations', () => {
 });
 
 describe('durable host operations', () => {
+    it('drains an old agent for a cancellation update while allowing receipts, heartbeats, and other hosts', async () => {
+        const api = await enrolled();
+        const original = api.queueHostOperation(api.host.id, 'instance.create', {}, { instanceId: 'cancelled-instance', generation: 1 });
+        const inspect = api.queueHostOperation(api.host.id, 'instance.inspect', {});
+        expect(api.pendingHostOperations(api.host.id).map(op => op.id)).toContain(original.id);
+        api.reportHostOperation(api.host.id, original.id, { state: 'running' });
+        api.reportHostOperation(api.host.id, inspect.id, { state: 'unknown' });
+        const cleanup = api.queueHostOperation(api.host.id, 'instance.cancel-create', { createOperationId: original.id }, { instanceId: 'cancelled-instance', generation: 1 });
+        const logs = api.queueHostOperation(api.host.id, 'logs.read', {});
+        const preparation = api.queueHostOperation(api.host.id, 'library.download', {});
+        const other = api.createHost({ name: 'Unaffected host' });
+        api.enrollHost(api.createEnrollment(other.id).token, inventory);
+        const otherWork = api.queueHostOperation(other.id, 'host.inspect', {});
+        const { db } = await import('../../src/lib/server/db/index.js');
+        const before = db().prepare('SELECT id,state,last_delivered_at FROM ec_host_operations WHERE host_id=? ORDER BY id').all(api.host.id);
+        for (let poll = 0; poll < 3; poll++) expect(api.pendingHostOperations(api.host.id)).toEqual([]);
+        expect(db().prepare('SELECT id,state,last_delivered_at FROM ec_host_operations WHERE host_id=? ORDER BY id').all(api.host.id)).toEqual(before);
+        expect(api.pendingHostOperations(other.id).map(op => op.id)).toEqual([otherWork.id]);
+        api.reportHostOperation(api.host.id, inspect.id, { state: 'completed', result: { healthy: true } });
+        api.heartbeat(api.host.id, inventory);
+        expect(api.pendingHostOperations(api.host.id)).toEqual([]);
+
+        // The controller hold must clear on capability, without waiting for cleanup
+        // or the original creation to finish (which would recreate the deadlock).
+        api.heartbeat(api.host.id, { ...inventory, capabilities: ['instance.cancel-create'] });
+        expect(api.pendingHostOperations(api.host.id).map(op => op.id)).toEqual([cleanup.id, logs.id, preparation.id]);
+        expect(api.getHostOperation(original.id)?.state).toBe('running');
+        expect(api.getHostOperation(cleanup.id)?.state).toBe('pending');
+    });
+
+    it.each(['completed', 'failed'] as const)('does not pause an old host for a terminal %s cancellation', async state => {
+        const api = await enrolled();
+        const cleanup = api.queueHostOperation(api.host.id, 'instance.cancel-create', { createOperationId: 'creation' });
+        api.reportHostOperation(api.host.id, cleanup.id, { state });
+        const work = api.queueHostOperation(api.host.id, 'host.inspect', {});
+        expect(api.pendingHostOperations(api.host.id).map(op => op.id)).toEqual([work.id]);
+    });
     it('rotates unresolved operation delivery so an old batch cannot starve new work', async () => {
         const api = await enrolled();
         const operations = Array.from({ length: 30 }, () => api.queueHostOperation(api.host.id, 'instance.inspect', {}));

tests/unit/native-provisioning.test.ts

@@ -142,6 +142,12 @@ describe('native instance provisioning', () => {
         expect(pendingHostOperations(hostId).filter(op => op.kind.startsWith('instance.'))).toEqual([]);
         await advance();
         expect((await detail(cleanup)).stage).toContain('Update the host agent');
+        expect((await detail(cleanup)).stage).toContain('Automatic updates are not enabled');
+        heartbeat(hostId, { ...inventory(), capabilities: ['agent.update'],
+            agentUpdate: { enabled: true, state: 'deferred', currentVersion: '1.2.2', targetVersion: '1.3.0' } });
+        await advance();
+        expect((await detail(cleanup)).stage).toContain('New command delivery is paused');
+        expect((await detail(cleanup)).stage).toContain('automatic updater can run');
         db().prepare('UPDATE ec_hosts SET last_heartbeat=? WHERE id=?').run('2000-01-01T00:00:00.000Z', hostId);
         await advance();
         expect((await detail(cleanup)).stage).toContain('unavailable');