@dalmia/calibrate-mcp 0.0.43 → 0.0.44

package/bin/mcp-server.js

@@ -51946,9 +51946,9 @@ var init_config = __esm(() => {
   SDK_METADATA = {
     language: "typescript",
     openapiDocVersion: "0.1.0",
-    sdkVersion: "0.0.43",
+    sdkVersion: "0.0.44",
     genVersion: "2.915.1",
-    userAgent: "speakeasy-sdk/mcp-typescript 0.0.43 2.915.1 0.1.0 @dalmia/calibrate-mcp"
+    userAgent: "speakeasy-sdk/mcp-typescript 0.0.44 2.915.1 0.1.0 @dalmia/calibrate-mcp"
   };
 });
 
@@ -55723,6 +55723,17 @@ var init_modelrunsummary = __esm(() => {
   }).describe("Flat summary for one model in a benchmark run-LIST item. The full results\nfor each case of a model live on the benchmark detail endpoint\n(`GET /agent-tests/benchmark/{task_id}`), not here.");
 });
 
+// src/models/runlistevaluator.ts
+var RunListEvaluator$zodSchema;
+var init_runlistevaluator = __esm(() => {
+  init_zod();
+  RunListEvaluator$zodSchema = object({
+    name: string2().describe("Name of the evaluator"),
+    uuid: string2().nullable().optional().describe("ID of the evaluator. Absent for the `Tool call` entry, which is not an evaluator in the library")
+  }).describe(`One evaluator on a run-LIST row: just enough to name it and open it.
+The rubric, prompt and version live on the run-DETAIL endpoint.`);
+});
+
 // src/models/testruncasesummary.ts
 var TestRunCaseSummary$zodSchema;
 var init_testruncasesummary = __esm(() => {
@@ -55742,6 +55753,7 @@ var AgentTestRunListItemType$zodSchema, AgentTestRunListItem$zodSchema;
 var init_agenttestrunlistitem = __esm(() => {
   init_zod();
   init_modelrunsummary();
+  init_runlistevaluator();
   init_taskstatus();
   init_testruncasesummary();
   AgentTestRunListItemType$zodSchema = _enum([
@@ -55752,7 +55764,7 @@ var init_agenttestrunlistitem = __esm(() => {
     cost: record(string2(), any()).nullable().optional().describe("Aggregated cost as `{mean, min, max, count}` (USD)"),
     created_at: string2().describe("When the run was created (ISO 8601 UTC)"),
     error: boolean2().default(false).describe("True if the run failed"),
-    evaluators: array(string2()).optional().describe("Names of the evaluators that judged this run, deduplicated and in display order. `Tool call` is appended when any test in the run was a tool-call test. Empty when the run had no evaluators"),
+    evaluators: array(RunListEvaluator$zodSchema).optional().describe("The evaluators that judged this run, deduplicated and in display order. A `Tool call` entry is appended when any test in the run was a tool-call test. That entry has no `uuid`, because it is not an evaluator in the library. Empty when the run had no evaluators"),
     failed: int().nullable().optional().describe("Number of test cases that failed"),
     is_public: boolean2().default(false).describe("Whether the run is shared publicly"),
     latency_ms: record(string2(), any()).nullable().optional().describe("Aggregated latency in milliseconds, as `{p50, p95, p99, count}`"),
@@ -60384,7 +60396,7 @@ hits its trace limit.
 function createMCPServer(deps) {
   const server = new McpServer({
     name: "CalibrateMcp",
-    version: "0.0.43"
+    version: "0.0.44"
   });
   const getClient = deps.getSDK || (() => new CalibrateMcpCore({
     security: deps.security,
@@ -61650,7 +61662,7 @@ http_headers = { "api-key-auth" = "YOUR_API_KEY_AUTH" }`;
         <h1>Instructions</h1>
         <p>One-click installation for Claude Desktop users</p>
         <div class="instruction-item">
-          <a href="https://github.com/dalmia/calibrate-mcp/releases/download/v0.0.43/mcp-server.mcpb" download="mcp-server.mcpb" class="action-button header-action" style="display: inline-flex; margin-bottom: 16px;">
+          <a href="https://github.com/dalmia/calibrate-mcp/releases/download/v0.0.44/mcp-server.mcpb" download="mcp-server.mcpb" class="action-button header-action" style="display: inline-flex; margin-bottom: 16px;">
             \uD83D\uDCE5 Download MCP Bundle
           </a>
         </div>
@@ -64537,7 +64549,7 @@ var routes = buildRouteMap({
 var app = buildApplication(routes, {
   name: "mcp",
   versionInfo: {
-    currentVersion: "0.0.43"
+    currentVersion: "0.0.44"
   }
 });
 run(app, process4.argv.slice(2), buildContext(process4));
@@ -64545,5 +64557,5 @@ export {
   app
 };
 
-//# debugId=AF946BF2D9C57E1564756E2164756E21
+//# debugId=D9742F01AE30E23564756E2164756E21
 //# sourceMappingURL=mcp-server.js.map