diff --git a/src/providers/kubernetes.js b/src/providers/kubernetes.js
index b8550a5..81572b3 100644
--- a/src/providers/kubernetes.js
+++ b/src/providers/kubernetes.js
@@ -9,6 +9,7 @@ import { canonicalResourceRefs, createMetadata, deriveWorkspaceIdentity, labelHa
 import { createWorkspaceSecrets, getWorkspaceToken, rotateWorkspaceCredentials, selectGrantedCredentials } from '../auth.js';
 import { requireKubernetesEgress, validateGatewayImage, validateImage } from '../policy.js';
 import { grantedEgressPolicy } from '../egress-domains.js';
+import { parseAuthCanIList, permissionsNeedingProbe } from './rbac-listing.js';
 import { waitForHttpHealth } from '../health.js';
 import { KUBERNETES_TOKEN_FILE, KUBERNETES_TOKEN_MOUNT_PATH, PROVIDER_MODEL_AUTH_FILE, runtimeCommand, runtimeEnvironment } from '../runtime-command.js';
 import { cleanupTransaction, createTransaction } from '../lifecycle.js';
@@ -47,10 +48,16 @@ export function createKubernetesProvider({ policy, sourceDirectory }) {
       }
       throw new ProviderUnavailableError(`Kubernetes cluster is not reachable: ${cause instanceof Error ? cause.message : String(cause)}`, { provider, code: 'WORKSPACE_PROVIDER_CLUSTER_UNREACHABLE', cause });
     });
-    // Each check is a process spawn and a round trip, and there are two dozen of them:
-    // run sequentially they dominated the time to display readiness. Every denial is
+    // Every check is a process spawn, and two dozen of them cost seconds on Windows —
+    // creating the processes, not the round trips, which is why raising the concurrency
+    // did not help. One listing answers for all of them at the price of one, and whatever
+    // it does not clearly settle is still asked about directly below. Every denial is
     // collected rather than only the first, so one message names everything to request.
-    const denied = await mapWithConcurrency(requiredPermissions(policy), RBAC_PROBE_CONCURRENCY, async ([verb, resource]) => {
+    const listing = await kubectl(['auth', 'can-i', '--list', '-n', policy.kubernetes.namespace], { timeoutMs: 20_000 })
+      .then((result) => parseAuthCanIList(result.stdout))
+      .catch(() => []);
+    const unsettled = permissionsNeedingProbe(listing, requiredPermissions(policy));
+    const denied = await mapWithConcurrency(unsettled, RBAC_PROBE_CONCURRENCY, async ([verb, resource]) => {
       const { stdout } = await kubectl(['auth', 'can-i', verb, resource, '-n', policy.kubernetes.namespace], { timeoutMs: 20_000 });
       return stdout.trim() === 'yes' ? null : `${verb} ${resource}`;
     });
diff --git a/src/providers/rbac-listing.js b/src/providers/rbac-listing.js
new file mode 100644
index 0000000..711c4a1
--- /dev/null
+++ b/src/providers/rbac-listing.js
@@ -0,0 +1,51 @@
+/**
+ * Reading permissions from one `kubectl auth can-i --list` instead of one process per
+ * permission.
+ *
+ * There are twenty-six permissions to establish, and asking about each one separately
+ * costs about 3.6 seconds on Windows — process creation, not the cluster, and raising the
+ * concurrency does not help because the spawns are what saturates. The same answer as a
+ * single listing costs about 250 milliseconds. That difference is the readiness check
+ * feeling instant rather than looking stuck.
+ *
+ * The listing is a table meant for people, so parsing it is the fragile part. It is used
+ * only to *grant*: a permission the rules clearly cover is settled, and anything else —
+ * an unfamiliar row, a wildcard shape not handled here, a listing that failed outright —
+ * falls through to the explicit probe that was always there. A parse that understands
+ * less is slower, never wrong.
+ */
+
+/** One rule from the listing: which resources it names and which verbs it allows. */
+export function parseAuthCanIList(output) {
+  const rules = [];
+  for (const line of String(output ?? '').split(/\r?\n/)) {
+    // Three bracketed columns follow the resource: non-resource URLs, resource names,
+    // verbs. Anchoring on them avoids depending on the column widths, which shift with
+    // the longest resource name in the table.
+    const match = /^(\S*)\s+\[([^\]]*)\]\s+\[([^\]]*)\]\s+\[([^\]]*)\]\s*$/.exec(line);
+    if (!match) continue;
+    const [, resourceColumn, , resourceNames, verbColumn] = match;
+    // A rule limited to named objects does not grant the verb on the resource in general,
+    // and a row with no resource is about a URL path rather than an API resource.
+    if (!resourceColumn || resourceNames.trim()) continue;
+    const verbs = verbColumn.split(/\s+/).filter(Boolean);
+    if (verbs.length === 0) continue;
+    rules.push({ resources: resourceColumn.split(',').filter(Boolean), verbs });
+  }
+  return rules;
+}
+
+function ruleGrants(rule, verb, resource) {
+  if (!rule.verbs.includes('*') && !rule.verbs.includes(verb)) return false;
+  return rule.resources.some((candidate) => candidate === '*.*' || candidate === '*' || candidate === resource);
+}
+
+/** Whether the listing settles this permission. Unsure is reported as not granted. */
+export function listingGrants(rules, verb, resource) {
+  return rules.some((rule) => ruleGrants(rule, verb, resource));
+}
+
+/** The permissions the listing could not settle, which still need asking about directly. */
+export function permissionsNeedingProbe(rules, required) {
+  return required.filter(([verb, resource]) => !listingGrants(rules, verb, resource));
+}
diff --git a/src/providers/rbac-listing.test.js b/src/providers/rbac-listing.test.js
new file mode 100644
index 0000000..df1591a
--- /dev/null
+++ b/src/providers/rbac-listing.test.js
@@ -0,0 +1,81 @@
+import { describe, expect, it } from 'vitest';
+import { listingGrants, parseAuthCanIList, permissionsNeedingProbe } from './rbac-listing.js';
+
+/** Real output from an admin account, whose single wildcard rule covers everything. */
+const ADMIN = [
+  'Resources                                       Non-Resource URLs   Resource Names   Verbs',
+  '*.*                                             []                  []               [*]',
+  '                                                [*]                 []               [*]',
+  'selfsubjectreviews.authentication.k8s.io        []                  []               [create]',
+  '                                                [/healthz]          []               [get]',
+].join('\n');
+
+/** The shape a namespace-scoped role produces: named resources, explicit verb lists. */
+const SCOPED = [
+  'Resources                                    Non-Resource URLs   Resource Names   Verbs',
+  'pods                                         []                  []               [create get delete watch]',
+  'pods/exec                                    []                  []               [create]',
+  'secrets                                      []                  []               [create get delete update]',
+  'deployments.apps                             []                  []               [create get delete list]',
+  'services                                     []                  []               [create get delete]',
+].join('\n');
+
+const REQUIRED = [
+  ['create', 'pods'],
+  ['delete', 'pods'],
+  ['watch', 'pods'],
+  ['create', 'pods/exec'],
+  ['update', 'secrets'],
+  ['list', 'deployments.apps'],
+];
+
+describe('permissions read from a single listing', () => {
+  it('reads a wildcard rule as covering every permission asked about', () => {
+    const rules = parseAuthCanIList(ADMIN);
+    expect(permissionsNeedingProbe(rules, REQUIRED)).toEqual([]);
+  });
+
+  it('reads named resources and their verbs', () => {
+    const rules = parseAuthCanIList(SCOPED);
+    expect(listingGrants(rules, 'create', 'pods')).toBe(true);
+    expect(listingGrants(rules, 'watch', 'pods')).toBe(true);
+    expect(listingGrants(rules, 'create', 'pods/exec')).toBe(true);
+    expect(listingGrants(rules, 'update', 'secrets')).toBe(true);
+    expect(listingGrants(rules, 'list', 'deployments.apps')).toBe(true);
+  });
+
+  it('leaves anything the listing does not cover to be asked about directly', () => {
+    const rules = parseAuthCanIList(SCOPED);
+    // Absent from the listing entirely, and a verb the listed rule does not carry.
+    expect(permissionsNeedingProbe(rules, [
+      ['create', 'networkpolicies.networking.k8s.io'],
+      ['list', 'pods'],
+    ])).toEqual([
+      ['create', 'networkpolicies.networking.k8s.io'],
+      ['list', 'pods'],
+    ]);
+  });
+
+  it('ignores rules about URL paths rather than API resources', () => {
+    // `[/healthz] [] [get]` grants `get` on a path; reading it as a resource rule would
+    // hand out `get` on whatever happened to be asked about.
+    const rules = parseAuthCanIList(ADMIN.split('\n').filter((line) => !line.startsWith('*.*')).join('\n'));
+    expect(listingGrants(rules, 'get', 'pods')).toBe(false);
+  });
+
+  it('does not treat a rule limited to named objects as a rule about the resource', () => {
+    const listing = [
+      'Resources   Non-Resource URLs   Resource Names   Verbs',
+      'secrets     []                  [one-secret]     [get]',
+    ].join('\n');
+    expect(listingGrants(parseAuthCanIList(listing), 'get', 'secrets')).toBe(false);
+  });
+
+  it('treats an unreadable listing as settling nothing', () => {
+    // A failed call passes an empty listing, and an unfamiliar table must behave the same
+    // way: everything falls through to the probe rather than being assumed granted.
+    expect(permissionsNeedingProbe([], REQUIRED)).toEqual(REQUIRED);
+    expect(permissionsNeedingProbe(parseAuthCanIList('error: you must be logged in'), REQUIRED)).toEqual(REQUIRED);
+    expect(parseAuthCanIList(undefined)).toEqual([]);
+  });
+});