@@ -13,7 +13,13 @@ const { pages } = vi.hoisted(() => {
         workflows: [{ id: 'wf-1', name: 'Lead Scorer', folderPath: '/Sales', isDeployed: true }],
         nextCursorKeys: null,
       }),
-      skills: page({ skills: [{ name: 'research' }], hasMore: false }),
+      skills: page({
+        skills: [
+          { id: 'builtin-research', name: 'research', description: 'Research a topic.' },
+          { id: 'skill-1', name: 'invoice-formatting', description: '' },
+        ],
+        hasMore: false,
+      }),
       customTools: page({ tools: [{ id: 'ct-1', title: 'Lookup' }] }),
       credentials: page({
         credentials: [
@@ -67,14 +73,18 @@ mocks.secrets.execute.mockImplementation(async () => ({
 }))
 
 describe('buildWorkspaceInventory', () => {
-  it('reads every world under the caller principal, by name and id, and names the capped worlds', async () => {
+  it('reads every world under the caller principal, by name and id, skills with their descriptions, and names the capped worlds', async () => {
     const principal = { kind: 'session', userId: 'user-1', sessionId: 's-1' } as const
     const inventory = await buildWorkspaceInventory(principal, 'ws-1')
     expect(inventory.workflows).toEqual([
       { id: 'wf-1', name: 'Lead Scorer', folder: '/Sales', deployed: true },
     ])
     expect(inventory.tables).toEqual([{ id: 'tbl_1', name: 'leads' }])
     expect(inventory.files).toEqual([{ path: 'files/Reports/q3.md', size: 12 }])
+    expect(inventory.skills).toEqual([
+      { id: 'builtin-research', name: 'research', description: 'Research a topic.' },
+      { id: 'skill-1', name: 'invoice-formatting' },
+    ])
     expect(inventory.credentials).toEqual([
       { id: 'cred-1', name: 'Slack bot', provider: 'slack', type: 'service_account' },
     ])
 
@@ -20,7 +20,8 @@ const FILE_LIMIT = 150
 
 /**
  * What exists in the workspace, by name and id, packed with every chat request so the
- * agent orients without a round of listings per world. Same use cases, same principal,
+ * agent orients without a round of listings per world. Skills also carry their description:
+ * the agent picks one from that catalog and loads its body with `skills get <id>`. Same use cases, same principal,
  * same authorization as the v2 listings the agent would otherwise call one by one.
  *
  * Each world is read independently: a world that fails to list is left empty and
@@ -118,7 +119,11 @@ export async function buildWorkspaceInventory(
         input: { workspaceId, sortBy: 'name', sortOrder: 'asc', limit: WORLD_LIMIT, offset: 0 },
       })
       cut('skills', page.hasMore)
-      return page.skills.map((s) => ({ name: s.name }))
+      return page.skills.map((s) => ({
+        id: s.id,
+        name: s.name,
+        ...(s.description ? { description: s.description } : {}),
+      }))
     }, []),
     world('customTools', async () => {
       const page = await listWorkspaceCustomToolsUseCase.execute({
 
@@ -72,7 +72,9 @@ const WorkspaceInventorySchema = z.object({
   tables: z.array(InventoryNamed),
   knowledgeBases: z.array(InventoryNamed),
   files: z.array(z.object({ path: z.string(), size: z.number().optional() })),
-  skills: z.array(z.object({ name: z.string() })),
+  skills: z.array(
+    z.object({ id: z.string().optional(), name: z.string(), description: z.string().optional() }),
+  ),
   customTools: z.array(z.object({ id: z.string(), title: z.string() })),
   mcpServers: z.array(InventoryNamed),
   credentials: z.array(
@@ -352,8 +354,9 @@ export interface ChatRequest extends StreamResponseReceipt {
   /**
    * What exists in the workspace, by name and id, so the agent orients without a round
    * of listings per world (an orientation turn on dev spent nine tool rounds and ~20K
-   * tokens learning this). Names and ids only — never a tree, never contents — and each
-   * world capped; a capped world is named in `truncated` so the agent lists it itself.
+   * tokens learning this). Names, ids and skill descriptions — never a tree, never
+   * contents — and each world capped; a capped world is named in `truncated` so the agent
+   * lists it itself.
    * Rendered as a request-local trailing message, never into the transcript.
    */
   inventory?: WorkspaceInventory | undefined;
@@ -366,7 +369,11 @@ export interface WorkspaceInventory {
   knowledgeBases: { id: string; name: string }[];
   /** `files/...` paths as the CLI prints them. */
   files: { path: string; size?: number | undefined }[];
-  skills: { name: string }[];
+  /**
+   * The skill catalog: the agent loads a matching skill with `skills get <id>`. `id` and
+   * `description` are optional so either side can ship first.
+   */
+  skills: { id?: string | undefined; name: string; description?: string | undefined }[];
   customTools: { id: string; title: string }[];
   mcpServers: { id: string; name: string }[];
   credentials: { id: string; name: string; provider?: string | undefined; type?: string | undefined }[];