test: strengthen MCP directory preflight

packages/mcp/DIRECTORY_PREFLIGHT.md

@@ -6,9 +6,27 @@ ChatGPT directory.
 - [ ] Run `npm run test --workspace @terminal49/mcp -- --run`,
       `npm run build --workspace @terminal49/mcp`, and
       `npm run lint --workspace @terminal49/mcp`.
+- [ ] Confirm every tool has a non-empty `title`, `readOnlyHint`, and
+      `destructiveHint`. Confirm `track_container` is the only write tool and
+      no tool is destructive. CI enforces this in
+      [`annotations.test.ts`](./src/annotations.test.ts).
+- [ ] Confirm tool names remain at most 64 characters, list `page_size` stays
+      bounded to 1–25, and the advertised `include` choices/defaults remain
+      lean. CI locks these contracts in
+      [`protocol-compat.test.ts`](./src/protocol-compat.test.ts).
 - [ ] Compare the live `tools/list` response with the tool reference in
       `docs/mcp/home.mdx`. Confirm parameter names, required fields, defaults,
       limits, and descriptions match.
+- [ ] Confirm the live connector still uses HTTPS, OAuth, and Streamable HTTP
+      at `https://mcp.terminal49.com`. CI checks the locked submission metadata
+      and exercises the protocol surface over Streamable HTTP in the annotation
+      and protocol-compatibility tests.
+- [ ] Confirm the tool surface stays task-specific with no catch-all request
+      tool such as `api_request`.
+- [ ] Confirm tools and server instructions request only their declared
+      parameters—not chat history, conversation context, or user memory. CI
+      rejects conversation- or memory-shaped arguments and instructions in
+      [`protocol-compat.test.ts`](./src/protocol-compat.test.ts).
 - [ ] Inspect successful tool results and server/tool instructions. They must
       not contain `_agent_steering`, `_response_contract`,
       `presentation_guidance`, `suggested_follow_ups`, or `suggested_tools`.
@@ -18,7 +36,13 @@ ChatGPT directory.
 - [ ] Confirm `track_container.number` is required and `get_container`
       `transport_events` returns only event counts plus the latest event. Use
       `get_container_transport_events` for full history.
-- [ ] Run `mcpdemo` against `https://mcp.terminal49.com` using Anthropic's
+- [ ] Recheck the [privacy policy](https://terminal49.com/privacy) and MCP data
+      practices: tools access the authenticated private account, use only tool
+      arguments, and only `track_container` creates data.
+- [ ] Populate and run `mcpdemo` with at least three working prompts. Exercise
+      every tool in MCP Inspector and a Claude custom connector; use the
+      [live eval](./eval/README.md) as the automated baseline.
+- [ ] Review the results against Anthropic's
       current [submission requirements](https://claude.com/docs/connectors/building/submission)
       and [Software Directory Policy](https://support.claude.com/en/articles/13145358-anthropic-software-directory-policy).
       Resolve every failure before resubmitting.

packages/mcp/src/annotations.test.ts

@@ -118,6 +118,19 @@ describe('MCP tool annotations', () => {
     expect(annotations?.idempotentHint).toBe(false);
   });
 
+  it('advertises track_container as the only write and no destructive tools', () => {
+    const tools = getRegisteredTools();
+    const writeTools = Object.entries(tools)
+      .filter(([, tool]) => tool.annotations?.readOnlyHint === false)
+      .map(([name]) => name);
+    const destructiveTools = Object.entries(tools)
+      .filter(([, tool]) => tool.annotations?.destructiveHint === true)
+      .map(([name]) => name);
+
+    expect(writeTools).toEqual(['track_container']);
+    expect(destructiveTools).toEqual([]);
+  });
+
   it('marks every tool as private-account-only and non-destructive', () => {
     const tools = getRegisteredTools();
 
@@ -150,6 +163,13 @@ describe('MCP tool annotations', () => {
       expect(tool.title, `${name}.title`).toEqual(expect.any(String));
       expect(tool.title?.trim().length, `${name}.title`).toBeGreaterThan(0);
       expect(tool.annotations, name).toBeDefined();
+      expect(tool.annotations?.readOnlyHint, `${name}.readOnlyHint`).toBeTypeOf(
+        'boolean',
+      );
+      expect(
+        tool.annotations?.destructiveHint,
+        `${name}.destructiveHint`,
+      ).toBeTypeOf('boolean');
     }
   });
 

packages/mcp/src/protocol-compat.test.ts

@@ -25,7 +25,10 @@ const openConnections: Array<{
 
 type AdvertisedProperty = {
   type?: string;
+  default?: unknown;
   maxLength?: number;
+  maxItems?: number;
+  exclusiveMinimum?: number;
   maximum?: number;
   description?: string;
   items?: { enum?: string[] };
@@ -173,11 +176,35 @@ describe('MCP protocol compatibility', () => {
 
     expect(tools).toHaveLength(10);
     for (const tool of tools) {
+      expect(tool.name.length, tool.name).toBeLessThanOrEqual(64);
       expect(tool.inputSchema.properties, tool.name).not.toHaveProperty(
         'intent',
       );
+      for (const propertyName of Object.keys(
+        tool.inputSchema.properties ?? {},
+      )) {
+        expect(propertyName, `${tool.name}.${propertyName}`).not.toMatch(
+          /(?:chat|conversation|history|memory|messages?)/i,
+        );
+      }
     }
 
+    const advertisedInstructions = [
+      TERMINAL49_SERVER_INSTRUCTIONS,
+      ...tools.flatMap((tool) => [
+        tool.description ?? '',
+        ...Object.values(
+          (tool.inputSchema as AdvertisedInputSchema).properties ?? {},
+        ).map((property) => property.description ?? ''),
+      ]),
+    ].join('\n');
+    expect(advertisedInstructions).not.toMatch(
+      /\b(?:send|provide|share|upload|attach)\b.{0,40}\b(?:chat|conversation) history\b/i,
+    );
+    expect(advertisedInstructions).not.toMatch(
+      /\b(?:send|provide|share|upload|attach)\b.{0,40}\buser memory\b/i,
+    );
+
     expect(
       toolSchemas.get('search_container')?.properties?.query,
     ).toMatchObject({
@@ -195,11 +222,47 @@ describe('MCP protocol compatibility', () => {
       'list_containers',
       'list_tracking_requests',
     ]) {
-      expect(toolSchemas.get(name)?.properties?.page_size?.maximum, name).toBe(
-        25,
-      );
+      expect(
+        toolSchemas.get(name)?.properties?.page,
+        `${name}.page`,
+      ).toMatchObject({
+        type: 'integer',
+        exclusiveMinimum: 0,
+      });
+      expect(
+        toolSchemas.get(name)?.properties?.page_size,
+        `${name}.page_size`,
+      ).toMatchObject({
+        default: 25,
+        type: 'integer',
+        exclusiveMinimum: 0,
+        maximum: 25,
+      });
     }
 
+    expect(toolSchemas.get('get_container')?.properties?.include).toMatchObject(
+      {
+        default: ['shipment'],
+        items: {
+          enum: ['shipment', 'pod_terminal', 'transport_events'],
+        },
+      },
+    );
+    expect(
+      toolSchemas.get('list_containers')?.properties?.include,
+    ).toMatchObject({
+      items: {
+        enum: ['shipment', 'pod_terminal'],
+      },
+      maxItems: 2,
+    });
+    expect(
+      toolSchemas.get('get_shipment_details')?.properties?.include_containers,
+    ).toMatchObject({ default: true });
+    expect(
+      toolSchemas.get('list_shipments')?.properties?.include_containers,
+    ).toMatchObject({ default: false });
+
     const trackingRequestSchema = toolSchemas.get('list_tracking_requests');
     expect(trackingRequestSchema?.properties).not.toHaveProperty('filters');
     expect(trackingRequestSchema?.properties).not.toHaveProperty(