feat(mcp): expose 6 MCP prompts + JMESPath schema parity (koala73#3892)

* feat(mcp): expose 6 MCP prompts + JMESPath schema parity

Turns on the MCP prompts capability with six concrete workflow templates
(country-briefing, energy-shock-watch, market-open-prep, conflict-pulse,
route-risk-check, freshness-audit). Each step pre-bakes a literal
JMESPath projection so the first execution lands on the right shape
without a discovery round-trip.

prompts/list + prompts/get are quota-exempt (per-minute limit still
applies) to mirror the describe_tool metadata posture — counting
template fetches against the 50/day Pro cap would discourage exploration.

`capabilities.prompts.listChanged: false` is advertised because the
stateless edge transport can't push notifications/prompts/list_changed
today.

Adds a parity test (tests/mcp-prompts.test.mjs) that compiles every
embedded JMESPath via the same parser used at runtime and asserts every
field identifier resolves to a property in the referenced tool's declared
outputSchema. A typo'd field path in a prompt OR a renamed field in a
tool's schema fails this test by name with both the prompt and the
broken path — verified by sabotaging both cases before commit.

SERVER_VERSION 1.7.0 -> 1.8.0; server-card prompts flag flipped in
lockstep. SERVER_INSTRUCTIONS mentions prompts/list as a discoverable
affordance.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* fix(mcp/prompts): strip empty-string optional args from rendered tool args

When a caller omits an optional prompt arg (e.g. country in
energy-shock-watch / conflict-pulse), the renderer used to emit
`arguments: {"country":""}` in the step block. The LLM then passed `""`
through to the tool literally. Today's referenced tools all truthy-check
filter args via argStr/argStrList (safe), but a future tool that guards
with `!== undefined` would try to filter by empty string instead of
serving the global view.

Strip top-level keys that resolved to '' so the no-filter call renders
as `arguments: {}` — unambiguous, matches the model's intent when the
arg is omitted. Required-arg absence already returned -32602 above, so
this only affects optional-arg absence.

Adds a regression test asserting the rendered block is `arguments: {}`
when an optional arg is omitted (and absent the literal `"country":""`).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: SebastienMelki

api/api-route-exceptions.json

@@ -114,6 +114,13 @@
       "owner": "@SebastienMelki",
       "removal_issue": null
     },
+    {
+      "path": "api/mcp/prompts/index.ts",
+      "category": "external-protocol",
+      "reason": "MCP PROMPT_REGISTRY + buildPromptResponse + prompts/list public-shape projection. Workflow templates surfaced via prompts/list + prompts/get JSON-RPC methods, shape dictated by the MCP spec.",
+      "owner": "@SebastienMelki",
+      "removal_issue": null
+    },
     {
       "path": "api/mcp-proxy.ts",
       "category": "external-protocol",

api/mcp.ts

@@ -64,10 +64,17 @@ export type {
 } from './mcp/types';
 export { compressDescription, utf8ByteLength } from './mcp/utils';
 
+export { buildPromptResponse, PROMPT_LIST_RESPONSE, PROMPT_REGISTRY } from './mcp/prompts/index';
+
 // Test-only escape hatch. Exposes the TOOL_REGISTRY by REFERENCE so mutations
 // inside `tests/mcp-tool-output-contracts.test.mjs` (which monkey-patches
 // `_execute` on individual RPC tools) propagate through the live binding.
+// PROMPT_REGISTRY follows the same live-binding contract so a test that
+// monkey-patches a prompt (e.g. the sabotage cases in mcp-prompts.test.mjs)
+// observes the same array the handler dispatches against.
+import { PROMPT_REGISTRY as __PROMPT_REGISTRY } from './mcp/prompts/index';
 import { TOOL_REGISTRY as __TOOL_REGISTRY } from './mcp/registry/index';
 export const __testing__ = {
   TOOL_REGISTRY: __TOOL_REGISTRY,
+  PROMPT_REGISTRY: __PROMPT_REGISTRY,
 };

api/mcp/constants.ts

@@ -125,6 +125,23 @@ export const SERVER_NAME = 'worldmonitor';
 //     unknown fields, and discovery on 2025-03-26 sessions should still benefit.
 //   - Purely additive on the wire — no input contract change. Bundle delta is
 //     documented in the v1.6.0 PR body.
+// Bumped 1.7.0 → 1.8.0 (2026-05-24) reflecting:
+//   - MCP `prompts` capability turned on. Six workflow templates exposed via
+//     prompts/list + prompts/get (country-briefing, energy-shock-watch,
+//     market-open-prep, conflict-pulse, route-risk-check, freshness-audit).
+//     Each template pre-bakes a literal JMESPath projection per step so the
+//     LLM doesn't have to discover the response shape on first execution.
+//   - Wire-visible additive capability — clients that ignore the new methods
+//     keep working; capable clients (Claude Desktop slash menu, MCP Inspector)
+//     surface the workflows in a discovery affordance.
+//   - prompts/list and prompts/get are quota-exempt (per-minute limit only)
+//     to mirror the describe_tool metadata posture — counting template
+//     fetches against the 50/day Pro cap would discourage exploration.
+//   - capabilities.prompts.listChanged = false advertised, because the
+//     stateless edge transport can't push notifications/prompts/list_changed
+//     today.
+//   - Server-card prompts capability flag flipped false → true in the same
+//     commit so external scanners see the wire and the card agree.
 // Bumped 1.6.0 → 1.7.0 (2026-05-23) reflecting:
 //   - De-blanket the `Tool.annotations` object. Previously buildPublicTool
 //     hard-coded `{ readOnlyHint: true, openWorldHint: true }` for every
@@ -144,7 +161,7 @@ export const SERVER_NAME = 'worldmonitor';
 //     hints keep working; new four-hint clients get a richer signal.
 // Keep aligned with public/.well-known/mcp/server-card.json::serverInfo.version
 // — discovery scanners cross-check both values.
-export const SERVER_VERSION = '1.7.0';
+export const SERVER_VERSION = '1.8.0';
 
 // MCP logging capability — valid severity levels per the 2025-03-26 spec
 // (RFC 5424 subset). Stateless HTTP transport: we ACK the level but do not
@@ -193,6 +210,8 @@ export const SERVER_INSTRUCTIONS = [
   `Limits: expression ≤ ${JMESPATH_MAX_EXPR_BYTES} bytes; projected payload ≤ ${JMESPATH_MAX_OUTPUT_BYTES} bytes. Failures return {_jmespath_error, original_keys} inside the normal result envelope. Bad expressions DO consume one daily quota unit on retry — original_keys is echoed so you can self-correct in one extra call.`,
   '',
   `tools/list returns COMPRESSED tool descriptions (first sentence, ≤${TOOL_DESCRIPTION_MAX_BYTES}B per tool). Call describe_tool({tool_name}) to get the full uncompressed definition for any tool you're considering — especially useful when the compressed entry is ambiguous about behaviour or argument semantics. describe_tool is metadata-only and is EXEMPT from the Pro daily quota (still counts toward the 60/min rate limit), so use it freely while exploring. describe_tool({tool_name: 'nonexistent'}) returns {error: 'unknown_tool', available: [...]} so you can self-correct.`,
+  '',
+  'Issue prompts/list to discover pre-built workflow templates (country-briefing, energy-shock-watch, market-open-prep, conflict-pulse, route-risk-check, freshness-audit). Each prompt pre-bakes a JMESPath projection per step so the first execution lands on the right shape. prompts/list + prompts/get are quota-exempt (per-minute limit only).',
 ].join('\n');
 
 // Country-code whitelist for get_consumer_prices. The consumer-prices seeder

api/mcp/handler.ts

@@ -14,6 +14,7 @@ import {
   SERVER_VERSION,
 } from './constants';
 import { dispatchToolsCall } from './dispatch';
+import { buildPromptResponse, PROMPT_LIST_RESPONSE } from './prompts/index';
 import { TOOL_LIST_BYTES, TOOL_LIST_RESPONSE } from './registry/index';
 import { rpcError, rpcOk } from './rpc';
 import { emitTelemetry, principalIdForLog } from './telemetry';
@@ -94,7 +95,11 @@ export async function mcpHandler(
       });
       return rpcOk(id, {
         protocolVersion: negotiatedVersion,
-        capabilities: { tools: {}, logging: {} },
+        // `prompts.listChanged: false` is the spec-correct value for our
+        // transport — the stateless edge route cannot push
+        // `notifications/prompts/list_changed`, so advertising `true` would
+        // be a lie. Same posture applies if/when `resources` lands later.
+        capabilities: { tools: {}, logging: {}, prompts: { listChanged: false } },
         serverInfo: { name: SERVER_NAME, version: SERVER_VERSION },
         instructions: SERVER_INSTRUCTIONS,
       }, { 'Mcp-Session-Id': sessionId, ...corsHeaders });
@@ -107,6 +112,22 @@ export async function mcpHandler(
       return rpcOk(id, { tools: TOOL_LIST_RESPONSE }, corsHeaders);
     case 'tools/call':
       return dispatchToolsCall(req, context, deps, body, corsHeaders, ctx);
+    // Prompts are metadata-class — they ship a workflow template, not data.
+    // Symmetric posture with `describe_tool`: quota-exempt (counting template
+    // fetches against the 50/day cap would discourage exploration, which
+    // defeats the prompt-discovery point), but the per-minute rate limit
+    // applied above still gates abusive loops.
+    case 'prompts/list':
+      return rpcOk(id, { prompts: PROMPT_LIST_RESPONSE }, corsHeaders);
+    case 'prompts/get': {
+      const params = body.params as { name?: unknown; arguments?: Record<string, unknown> } | null;
+      if (!params || typeof params.name !== 'string') {
+        return rpcError(id, -32602, 'Invalid params: missing prompt name');
+      }
+      const built = buildPromptResponse(params.name, params.arguments);
+      if (!built.ok) return rpcError(id, built.code, built.message);
+      return rpcOk(id, { description: built.description, messages: built.messages }, corsHeaders);
+    }
     case 'logging/setLevel': {
       const level = (body.params as { level?: string } | null)?.level;
       if (typeof level !== 'string' || !MCP_LOG_LEVELS.has(level)) {