@dalmia/calibrate-mcp 0.0.62 → 0.0.63

package/esm/landing-page.js

@@ -915,7 +915,7 @@ 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/v0.0.62/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/v0.0.63/mcp-server.mcpb" download="mcp-server.mcpb" class="action-button header-action" style="display: inline-flex; margin-bottom: 16px;">
             📥 Download MCP Bundle
           </a>
         </div>

package/esm/lib/config.d.ts

@@ -35,8 +35,8 @@ export declare function serverURLFromOptions(options: SDKOptions): URL | null;
 export declare const SDK_METADATA: {
     readonly language: "typescript";
     readonly openapiDocVersion: "0.1.0";
-    readonly sdkVersion: "0.0.62";
+    readonly sdkVersion: "0.0.63";
     readonly genVersion: "2.915.1";
-    readonly userAgent: "speakeasy-sdk/mcp-typescript 0.0.62 2.915.1 0.1.0 @dalmia/calibrate-mcp";
+    readonly userAgent: "speakeasy-sdk/mcp-typescript 0.0.63 2.915.1 0.1.0 @dalmia/calibrate-mcp";
 };
 //# sourceMappingURL=config.d.ts.map

package/esm/lib/config.js

@@ -27,8 +27,8 @@ export function serverURLFromOptions(options) {
 export const SDK_METADATA = {
     language: "typescript",
     openapiDocVersion: "0.1.0",
-    sdkVersion: "0.0.62",
+    sdkVersion: "0.0.63",
     genVersion: "2.915.1",
-    userAgent: "speakeasy-sdk/mcp-typescript 0.0.62 2.915.1 0.1.0 @dalmia/calibrate-mcp",
+    userAgent: "speakeasy-sdk/mcp-typescript 0.0.63 2.915.1 0.1.0 @dalmia/calibrate-mcp",
 };
 //# sourceMappingURL=config.js.map

package/esm/mcp-server/mcp-server.js

@@ -18,7 +18,7 @@ const routes = buildRouteMap({
 export const app = buildApplication(routes, {
     name: "mcp",
     versionInfo: {
-        currentVersion: "0.0.62",
+        currentVersion: "0.0.63",
     },
 });
 run(app, process.argv.slice(2), buildContext(process));

package/esm/mcp-server/server.js

@@ -51,7 +51,7 @@ import { tool$tracesCreate } from "./tools/tracesCreate.js";
 export function createMCPServer(deps) {
     const server = new McpServer({
         name: "CalibrateMcp",
-        version: "0.0.62",
+        version: "0.0.63",
     });
     const getClient = deps.getSDK || (() => new CalibrateMcpCore({
         security: deps.security,

package/esm/models/agentcreate.js

@@ -33,7 +33,7 @@ export const AgentCreateInteractionType$zodSchema = z.enum([
     "general",
 ]).describe("What the agent expects in the request body:\n\n- `conversation`: a normal back-and-forth agent, answers within an ongoing conversation. Receives `{\"messages\": [...]}`\n- `general`: a one-shot agent, takes a single plain input and produces a single plain output, no conversation. Receives `{\"input\": \"...\"}`");
 export const AgentCreate$zodSchema = z.object({
-    config: z.record(z.string(), z.any()).nullable().optional().describe("Agent behavioral config. The keys depend on `type`.\n\n**`type=agent`**, built inside Calibrate:\n- `system_prompt`: the agent's instructions\n- `llm.model`: `provider/model`, e.g. `openai/gpt-4.1` or `google/gemini-2.5-flash`\n- `stt.provider`: `deepgram`, `openai`, `cartesia`, `elevenlabs`, `google`, `sarvam`, or `smallest`\n- `tts.provider`: `cartesia`, `openai`, `google`, `elevenlabs`, `sarvam`, or `smallest`\n- `settings.agent_speaks_first`, `settings.max_assistant_turns`\n- `system_tools.end_call`: let the agent end the call\n- `data_extraction_fields`: `[{name, type, description, required}]`\n\n```json\n{\n  \"system_prompt\": \"You are a helpful support agent.\",\n  \"llm\": {\"model\": \"openai/gpt-4.1\"},\n  \"stt\": {\"provider\": \"deepgram\"},\n  \"tts\": {\"provider\": \"elevenlabs\"},\n  \"settings\": {\"agent_speaks_first\": true, \"max_assistant_turns\": 50}\n}\n```\n\n**`type=connection`**, your own HTTP endpoint:\n- `agent_url`: public HTTP(S) endpoint your agent is called at\n- `agent_headers`: headers sent on each request, e.g. auth\n- `benchmark_provider`: `openrouter` by default. Other values: `openai`, `google`, `anthropic`, `meta-llama`, `mistralai`, `deepseek`, `x-ai`, `cohere`, `qwen`, or `ai21`\n\n```json\n{\n  \"agent_url\": \"https://api.example.com/agent\",\n  \"agent_headers\": {\"Authorization\": \"Bearer <token>\"},\n  \"benchmark_provider\": \"openrouter\"\n}\n```\n\nEvery request Calibrate makes to your endpoint carries the header\n`X-Calibrate-Eval: 1`. Read it to tell a test run from a real user, for example\nto tag the trace you send back or to skip sending one.\n\nFor `type=agent`, omitted keys inherit managed defaults. Omit `config` entirely to use all defaults. For `type=connection`, `config` is stored as-is and must contain `agent_url`"),
+    config: z.record(z.string(), z.any()).nullable().optional().describe("Agent behavioral config. The keys depend on `type`.\n\n**`type=agent`**, built inside Calibrate:\n- `system_prompt`: the agent's instructions\n- `llm.model`: `provider/model`, e.g. `openai/gpt-4.1` or `google/gemini-2.5-flash`\n- `stt.provider`: `deepgram`, `openai`, `cartesia`, `elevenlabs`, `google`, `sarvam`, or `smallest`\n- `tts.provider`: `cartesia`, `openai`, `google`, `elevenlabs`, `sarvam`, or `smallest`\n- `settings.agent_speaks_first`, `settings.max_assistant_turns`\n- `system_tools.end_call`: let the agent end the call\n- `data_extraction_fields`: `[{name, type, description, required}]`\n\n```json\n{\n  \"system_prompt\": \"You are a helpful support agent.\",\n  \"llm\": {\"model\": \"openai/gpt-4.1\"},\n  \"stt\": {\"provider\": \"deepgram\"},\n  \"tts\": {\"provider\": \"elevenlabs\"},\n  \"settings\": {\"agent_speaks_first\": true, \"max_assistant_turns\": 50}\n}\n```\n\n**`type=connection`**, your own HTTP endpoint:\n- `agent_url`: public HTTP(S) endpoint your agent is called at\n- `agent_headers`: headers sent on each request, e.g. auth\n- `benchmark_provider`: `openrouter` by default. Other values: `openai`, `google`, `anthropic`, `meta-llama`, `mistralai`, `deepseek`, `x-ai`, `cohere`, `qwen`, or `ai21`\n\n```json\n{\n  \"agent_url\": \"https://api.example.com/agent\",\n  \"agent_headers\": {\"Authorization\": \"Bearer <token>\"},\n  \"benchmark_provider\": \"openrouter\"\n}\n```\n\n**Either type**:\n- `traces.scoring.enabled`: whether the traces you send for this agent are scored by its linked evaluators. On unless you set it to `false`. Turning it back on needs at least one linked evaluator that can score traces\n\n```json\n{\"traces\": {\"scoring\": {\"enabled\": false}}}\n```\n\nEvery request Calibrate makes to your endpoint carries the header\n`X-Calibrate-Eval: 1`. Read it to tell a test run from a real user, for example\nto tag the trace you send back or to skip sending one.\n\nFor `type=agent`, omitted keys inherit managed defaults. Omit `config` entirely to use all defaults. For `type=connection`, `config` is stored as-is and must contain `agent_url`"),
     interaction_type: AgentCreateInteractionType$zodSchema.default("conversation")
         .describe("What the agent expects in the request body:\n\n- `conversation`: a normal back-and-forth agent, answers within an ongoing conversation. Receives `{\"messages\": [...]}`\n- `general`: a one-shot agent, takes a single plain input and produces a single plain output, no conversation. Receives `{\"input\": \"...\"}`"),
     name: z.string().describe("Agent name, unique within the workspace"),

package/esm/models/agentcreate.js.map

@@ -1 +1 @@
-{"version":3,"file":"agentcreate.js","sourceRoot":"","sources":["../../src/models/agentcreate.ts"],"names":[],"mappings":"AAAA;;GAEG;AAEH,OAAO,KAAK,CAAC,MAAM,KAAK,CAAC;AAGzB;;;;;GAKG;AACH,MAAM,CAAC,MAAM,eAAe,GAAG;IAC7B,KAAK,EAAE,OAAO;IACd,UAAU,EAAE,YAAY;CAChB,CAAC;AASX,MAAM,CAAC,MAAM,yBAAyB,GAAG,CAAC,CAAC,IAAI,CAAC;IAC9C,OAAO;IACP,YAAY;CACb,CAAC,CAAC,QAAQ,CACT,+FAA+F,CAChG,CAAC;AAEF;;;;;;;GAOG;AACH,MAAM,CAAC,MAAM,0BAA0B,GAAG;IACxC,YAAY,EAAE,cAAc;IAC5B,OAAO,EAAE,SAAS;CACV,CAAC;AAaX,MAAM,CAAC,MAAM,oCAAoC,GAAG,CAAC,CAAC,IAAI,CAAC;IACzD,cAAc;IACd,SAAS;CACV,CAAC,CAAC,QAAQ,CACT,0TAA0T,CAC3T,CAAC;AASF,MAAM,CAAC,MAAM,qBAAqB,GAA2B,CAAC,CAAC,MAAM,CAAC;IACpE,MAAM,EAAE,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,MAAM,EAAE,EAAE,CAAC,CAAC,GAAG,EAAE,CAAC,CAAC,QAAQ,EAAE,CAAC,QAAQ,EAAE,CAAC,QAAQ,CAClE,kxDAAkxD,CACnxD;IACD,gBAAgB,EAAE,oCAAoC,CAAC,OAAO,CAAC,cAAc,CAAC;SAC3E,QAAQ,CACP,0TAA0T,CAC3T;IACH,IAAI,EAAE,CAAC,CAAC,MAAM,EAAE,CAAC,QAAQ,CAAC,yCAAyC,CAAC;IACpE,IAAI,EAAE,yBAAyB,CAAC,OAAO,CAAC,OAAO,CAAC,CAAC,QAAQ,CACvD,+FAA+F,CAChG;CACF,CAAC,CAAC"}
+{"version":3,"file":"agentcreate.js","sourceRoot":"","sources":["../../src/models/agentcreate.ts"],"names":[],"mappings":"AAAA;;GAEG;AAEH,OAAO,KAAK,CAAC,MAAM,KAAK,CAAC;AAGzB;;;;;GAKG;AACH,MAAM,CAAC,MAAM,eAAe,GAAG;IAC7B,KAAK,EAAE,OAAO;IACd,UAAU,EAAE,YAAY;CAChB,CAAC;AASX,MAAM,CAAC,MAAM,yBAAyB,GAAG,CAAC,CAAC,IAAI,CAAC;IAC9C,OAAO;IACP,YAAY;CACb,CAAC,CAAC,QAAQ,CACT,+FAA+F,CAChG,CAAC;AAEF;;;;;;;GAOG;AACH,MAAM,CAAC,MAAM,0BAA0B,GAAG;IACxC,YAAY,EAAE,cAAc;IAC5B,OAAO,EAAE,SAAS;CACV,CAAC;AAaX,MAAM,CAAC,MAAM,oCAAoC,GAAG,CAAC,CAAC,IAAI,CAAC;IACzD,cAAc;IACd,SAAS;CACV,CAAC,CAAC,QAAQ,CACT,0TAA0T,CAC3T,CAAC;AASF,MAAM,CAAC,MAAM,qBAAqB,GAA2B,CAAC,CAAC,MAAM,CAAC;IACpE,MAAM,EAAE,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,MAAM,EAAE,EAAE,CAAC,CAAC,GAAG,EAAE,CAAC,CAAC,QAAQ,EAAE,CAAC,QAAQ,EAAE,CAAC,QAAQ,CAClE,okEAAokE,CACrkE;IACD,gBAAgB,EAAE,oCAAoC,CAAC,OAAO,CAAC,cAAc,CAAC;SAC3E,QAAQ,CACP,0TAA0T,CAC3T;IACH,IAAI,EAAE,CAAC,CAAC,MAAM,EAAE,CAAC,QAAQ,CAAC,yCAAyC,CAAC;IACpE,IAAI,EAAE,yBAAyB,CAAC,OAAO,CAAC,OAAO,CAAC,CAAC,QAAQ,CACvD,+FAA+F,CAChG;CACF,CAAC,CAAC"}

package/esm/models/agentupdate.js

@@ -3,7 +3,7 @@
  */
 import * as z from "zod";
 export const AgentUpdate$zodSchema = z.object({
-    config: z.record(z.string(), z.any()).nullable().optional().describe("Agent behavioral config. The keys depend on `type`.\n\n**`type=agent`**, built inside Calibrate:\n- `system_prompt`: the agent's instructions\n- `llm.model`: `provider/model`, e.g. `openai/gpt-4.1` or `google/gemini-2.5-flash`\n- `stt.provider`: `deepgram`, `openai`, `cartesia`, `elevenlabs`, `google`, `sarvam`, or `smallest`\n- `tts.provider`: `cartesia`, `openai`, `google`, `elevenlabs`, `sarvam`, or `smallest`\n- `settings.agent_speaks_first`, `settings.max_assistant_turns`\n- `system_tools.end_call`: let the agent end the call\n- `data_extraction_fields`: `[{name, type, description, required}]`\n\n```json\n{\n  \"system_prompt\": \"You are a helpful support agent.\",\n  \"llm\": {\"model\": \"openai/gpt-4.1\"},\n  \"stt\": {\"provider\": \"deepgram\"},\n  \"tts\": {\"provider\": \"elevenlabs\"},\n  \"settings\": {\"agent_speaks_first\": true, \"max_assistant_turns\": 50}\n}\n```\n\n**`type=connection`**, your own HTTP endpoint:\n- `agent_url`: public HTTP(S) endpoint your agent is called at\n- `agent_headers`: headers sent on each request, e.g. auth\n- `benchmark_provider`: `openrouter` by default. Other values: `openai`, `google`, `anthropic`, `meta-llama`, `mistralai`, `deepseek`, `x-ai`, `cohere`, `qwen`, or `ai21`\n\n```json\n{\n  \"agent_url\": \"https://api.example.com/agent\",\n  \"agent_headers\": {\"Authorization\": \"Bearer <token>\"},\n  \"benchmark_provider\": \"openrouter\"\n}\n```\n\nEvery request Calibrate makes to your endpoint carries the header\n`X-Calibrate-Eval: 1`. Read it to tell a test run from a real user, for example\nto tag the trace you send back or to skip sending one.\n\nReplaces the stored config. Omit to leave unchanged\n\nFor `type=connection`, changing `agent_url` or `agent_headers` resets the connection and benchmark verification flags"),
+    config: z.record(z.string(), z.any()).nullable().optional().describe("Agent behavioral config. The keys depend on `type`.\n\n**`type=agent`**, built inside Calibrate:\n- `system_prompt`: the agent's instructions\n- `llm.model`: `provider/model`, e.g. `openai/gpt-4.1` or `google/gemini-2.5-flash`\n- `stt.provider`: `deepgram`, `openai`, `cartesia`, `elevenlabs`, `google`, `sarvam`, or `smallest`\n- `tts.provider`: `cartesia`, `openai`, `google`, `elevenlabs`, `sarvam`, or `smallest`\n- `settings.agent_speaks_first`, `settings.max_assistant_turns`\n- `system_tools.end_call`: let the agent end the call\n- `data_extraction_fields`: `[{name, type, description, required}]`\n\n```json\n{\n  \"system_prompt\": \"You are a helpful support agent.\",\n  \"llm\": {\"model\": \"openai/gpt-4.1\"},\n  \"stt\": {\"provider\": \"deepgram\"},\n  \"tts\": {\"provider\": \"elevenlabs\"},\n  \"settings\": {\"agent_speaks_first\": true, \"max_assistant_turns\": 50}\n}\n```\n\n**`type=connection`**, your own HTTP endpoint:\n- `agent_url`: public HTTP(S) endpoint your agent is called at\n- `agent_headers`: headers sent on each request, e.g. auth\n- `benchmark_provider`: `openrouter` by default. Other values: `openai`, `google`, `anthropic`, `meta-llama`, `mistralai`, `deepseek`, `x-ai`, `cohere`, `qwen`, or `ai21`\n\n```json\n{\n  \"agent_url\": \"https://api.example.com/agent\",\n  \"agent_headers\": {\"Authorization\": \"Bearer <token>\"},\n  \"benchmark_provider\": \"openrouter\"\n}\n```\n\n**Either type**:\n- `traces.scoring.enabled`: whether the traces you send for this agent are scored by its linked evaluators. On unless you set it to `false`. Turning it back on needs at least one linked evaluator that can score traces\n\n```json\n{\"traces\": {\"scoring\": {\"enabled\": false}}}\n```\n\nEvery request Calibrate makes to your endpoint carries the header\n`X-Calibrate-Eval: 1`. Read it to tell a test run from a real user, for example\nto tag the trace you send back or to skip sending one.\n\nReplaces the stored config. Omit to leave unchanged\n\nFor `type=connection`, changing `agent_url` or `agent_headers` resets the connection and benchmark verification flags"),
     name: z.string().nullable().optional().describe("New agent name. Omit to leave the name unchanged"),
 });
 //# sourceMappingURL=agentupdate.js.map

package/esm/models/agentupdate.js.map

@@ -1 +1 @@
-{"version":3,"file":"agentupdate.js","sourceRoot":"","sources":["../../src/models/agentupdate.ts"],"names":[],"mappings":"AAAA;;GAEG;AAEH,OAAO,KAAK,CAAC,MAAM,KAAK,CAAC;AAOzB,MAAM,CAAC,MAAM,qBAAqB,GAA2B,CAAC,CAAC,MAAM,CAAC;IACpE,MAAM,EAAE,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,MAAM,EAAE,EAAE,CAAC,CAAC,GAAG,EAAE,CAAC,CAAC,QAAQ,EAAE,CAAC,QAAQ,EAAE,CAAC,QAAQ,CAClE,6wDAA6wD,CAC9wD;IACD,IAAI,EAAE,CAAC,CAAC,MAAM,EAAE,CAAC,QAAQ,EAAE,CAAC,QAAQ,EAAE,CAAC,QAAQ,CAC7C,kDAAkD,CACnD;CACF,CAAC,CAAC"}
+{"version":3,"file":"agentupdate.js","sourceRoot":"","sources":["../../src/models/agentupdate.ts"],"names":[],"mappings":"AAAA;;GAEG;AAEH,OAAO,KAAK,CAAC,MAAM,KAAK,CAAC;AAOzB,MAAM,CAAC,MAAM,qBAAqB,GAA2B,CAAC,CAAC,MAAM,CAAC;IACpE,MAAM,EAAE,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,MAAM,EAAE,EAAE,CAAC,CAAC,GAAG,EAAE,CAAC,CAAC,QAAQ,EAAE,CAAC,QAAQ,EAAE,CAAC,QAAQ,CAClE,+jEAA+jE,CAChkE;IACD,IAAI,EAAE,CAAC,CAAC,MAAM,EAAE,CAAC,QAAQ,EAAE,CAAC,QAAQ,EAAE,CAAC,QAAQ,CAC7C,kDAAkD,CACnD;CACF,CAAC,CAAC"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@dalmia/calibrate-mcp",
-  "version": "0.0.62",
+  "version": "0.0.63",
   "author": "Speakeasy",
   "type": "module",
   "sideEffects": false,

package/src/landing-page.ts

@@ -930,7 +930,7 @@ 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/v0.0.62/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/v0.0.63/mcp-server.mcpb" download="mcp-server.mcpb" class="action-button header-action" style="display: inline-flex; margin-bottom: 16px;">
             📥 Download MCP Bundle
           </a>
         </div>

package/src/lib/config.ts

@@ -65,8 +65,8 @@ export function serverURLFromOptions(options: SDKOptions): URL | null {
 export const SDK_METADATA = {
   language: "typescript",
   openapiDocVersion: "0.1.0",
-  sdkVersion: "0.0.62",
+  sdkVersion: "0.0.63",
   genVersion: "2.915.1",
   userAgent:
-    "speakeasy-sdk/mcp-typescript 0.0.62 2.915.1 0.1.0 @dalmia/calibrate-mcp",
+    "speakeasy-sdk/mcp-typescript 0.0.63 2.915.1 0.1.0 @dalmia/calibrate-mcp",
 } as const;

package/src/mcp-server/mcp-server.ts

@@ -21,7 +21,7 @@ const routes = buildRouteMap({
 export const app = buildApplication(routes, {
   name: "mcp",
   versionInfo: {
-    currentVersion: "0.0.62",
+    currentVersion: "0.0.63",
   },
 });
 

package/src/mcp-server/server.ts

@@ -73,7 +73,7 @@ export function createMCPServer(deps: {
 }) {
   const server = new McpServer({
     name: "CalibrateMcp",
-    version: "0.0.62",
+    version: "0.0.63",
   });
 
   const getClient = deps.getSDK || (() =>

package/src/models/agentcreate.ts

@@ -70,7 +70,7 @@ export type AgentCreate = {
 
 export const AgentCreate$zodSchema: z.ZodType<AgentCreate> = z.object({
   config: z.record(z.string(), z.any()).nullable().optional().describe(
-    "Agent behavioral config. The keys depend on `type`.\n\n**`type=agent`**, built inside Calibrate:\n- `system_prompt`: the agent's instructions\n- `llm.model`: `provider/model`, e.g. `openai/gpt-4.1` or `google/gemini-2.5-flash`\n- `stt.provider`: `deepgram`, `openai`, `cartesia`, `elevenlabs`, `google`, `sarvam`, or `smallest`\n- `tts.provider`: `cartesia`, `openai`, `google`, `elevenlabs`, `sarvam`, or `smallest`\n- `settings.agent_speaks_first`, `settings.max_assistant_turns`\n- `system_tools.end_call`: let the agent end the call\n- `data_extraction_fields`: `[{name, type, description, required}]`\n\n```json\n{\n  \"system_prompt\": \"You are a helpful support agent.\",\n  \"llm\": {\"model\": \"openai/gpt-4.1\"},\n  \"stt\": {\"provider\": \"deepgram\"},\n  \"tts\": {\"provider\": \"elevenlabs\"},\n  \"settings\": {\"agent_speaks_first\": true, \"max_assistant_turns\": 50}\n}\n```\n\n**`type=connection`**, your own HTTP endpoint:\n- `agent_url`: public HTTP(S) endpoint your agent is called at\n- `agent_headers`: headers sent on each request, e.g. auth\n- `benchmark_provider`: `openrouter` by default. Other values: `openai`, `google`, `anthropic`, `meta-llama`, `mistralai`, `deepseek`, `x-ai`, `cohere`, `qwen`, or `ai21`\n\n```json\n{\n  \"agent_url\": \"https://api.example.com/agent\",\n  \"agent_headers\": {\"Authorization\": \"Bearer <token>\"},\n  \"benchmark_provider\": \"openrouter\"\n}\n```\n\nEvery request Calibrate makes to your endpoint carries the header\n`X-Calibrate-Eval: 1`. Read it to tell a test run from a real user, for example\nto tag the trace you send back or to skip sending one.\n\nFor `type=agent`, omitted keys inherit managed defaults. Omit `config` entirely to use all defaults. For `type=connection`, `config` is stored as-is and must contain `agent_url`",
+    "Agent behavioral config. The keys depend on `type`.\n\n**`type=agent`**, built inside Calibrate:\n- `system_prompt`: the agent's instructions\n- `llm.model`: `provider/model`, e.g. `openai/gpt-4.1` or `google/gemini-2.5-flash`\n- `stt.provider`: `deepgram`, `openai`, `cartesia`, `elevenlabs`, `google`, `sarvam`, or `smallest`\n- `tts.provider`: `cartesia`, `openai`, `google`, `elevenlabs`, `sarvam`, or `smallest`\n- `settings.agent_speaks_first`, `settings.max_assistant_turns`\n- `system_tools.end_call`: let the agent end the call\n- `data_extraction_fields`: `[{name, type, description, required}]`\n\n```json\n{\n  \"system_prompt\": \"You are a helpful support agent.\",\n  \"llm\": {\"model\": \"openai/gpt-4.1\"},\n  \"stt\": {\"provider\": \"deepgram\"},\n  \"tts\": {\"provider\": \"elevenlabs\"},\n  \"settings\": {\"agent_speaks_first\": true, \"max_assistant_turns\": 50}\n}\n```\n\n**`type=connection`**, your own HTTP endpoint:\n- `agent_url`: public HTTP(S) endpoint your agent is called at\n- `agent_headers`: headers sent on each request, e.g. auth\n- `benchmark_provider`: `openrouter` by default. Other values: `openai`, `google`, `anthropic`, `meta-llama`, `mistralai`, `deepseek`, `x-ai`, `cohere`, `qwen`, or `ai21`\n\n```json\n{\n  \"agent_url\": \"https://api.example.com/agent\",\n  \"agent_headers\": {\"Authorization\": \"Bearer <token>\"},\n  \"benchmark_provider\": \"openrouter\"\n}\n```\n\n**Either type**:\n- `traces.scoring.enabled`: whether the traces you send for this agent are scored by its linked evaluators. On unless you set it to `false`. Turning it back on needs at least one linked evaluator that can score traces\n\n```json\n{\"traces\": {\"scoring\": {\"enabled\": false}}}\n```\n\nEvery request Calibrate makes to your endpoint carries the header\n`X-Calibrate-Eval: 1`. Read it to tell a test run from a real user, for example\nto tag the trace you send back or to skip sending one.\n\nFor `type=agent`, omitted keys inherit managed defaults. Omit `config` entirely to use all defaults. For `type=connection`, `config` is stored as-is and must contain `agent_url`",
   ),
   interaction_type: AgentCreateInteractionType$zodSchema.default("conversation")
     .describe(

package/src/models/agentupdate.ts

@@ -11,7 +11,7 @@ export type AgentUpdate = {
 
 export const AgentUpdate$zodSchema: z.ZodType<AgentUpdate> = z.object({
   config: z.record(z.string(), z.any()).nullable().optional().describe(
-    "Agent behavioral config. The keys depend on `type`.\n\n**`type=agent`**, built inside Calibrate:\n- `system_prompt`: the agent's instructions\n- `llm.model`: `provider/model`, e.g. `openai/gpt-4.1` or `google/gemini-2.5-flash`\n- `stt.provider`: `deepgram`, `openai`, `cartesia`, `elevenlabs`, `google`, `sarvam`, or `smallest`\n- `tts.provider`: `cartesia`, `openai`, `google`, `elevenlabs`, `sarvam`, or `smallest`\n- `settings.agent_speaks_first`, `settings.max_assistant_turns`\n- `system_tools.end_call`: let the agent end the call\n- `data_extraction_fields`: `[{name, type, description, required}]`\n\n```json\n{\n  \"system_prompt\": \"You are a helpful support agent.\",\n  \"llm\": {\"model\": \"openai/gpt-4.1\"},\n  \"stt\": {\"provider\": \"deepgram\"},\n  \"tts\": {\"provider\": \"elevenlabs\"},\n  \"settings\": {\"agent_speaks_first\": true, \"max_assistant_turns\": 50}\n}\n```\n\n**`type=connection`**, your own HTTP endpoint:\n- `agent_url`: public HTTP(S) endpoint your agent is called at\n- `agent_headers`: headers sent on each request, e.g. auth\n- `benchmark_provider`: `openrouter` by default. Other values: `openai`, `google`, `anthropic`, `meta-llama`, `mistralai`, `deepseek`, `x-ai`, `cohere`, `qwen`, or `ai21`\n\n```json\n{\n  \"agent_url\": \"https://api.example.com/agent\",\n  \"agent_headers\": {\"Authorization\": \"Bearer <token>\"},\n  \"benchmark_provider\": \"openrouter\"\n}\n```\n\nEvery request Calibrate makes to your endpoint carries the header\n`X-Calibrate-Eval: 1`. Read it to tell a test run from a real user, for example\nto tag the trace you send back or to skip sending one.\n\nReplaces the stored config. Omit to leave unchanged\n\nFor `type=connection`, changing `agent_url` or `agent_headers` resets the connection and benchmark verification flags",
+    "Agent behavioral config. The keys depend on `type`.\n\n**`type=agent`**, built inside Calibrate:\n- `system_prompt`: the agent's instructions\n- `llm.model`: `provider/model`, e.g. `openai/gpt-4.1` or `google/gemini-2.5-flash`\n- `stt.provider`: `deepgram`, `openai`, `cartesia`, `elevenlabs`, `google`, `sarvam`, or `smallest`\n- `tts.provider`: `cartesia`, `openai`, `google`, `elevenlabs`, `sarvam`, or `smallest`\n- `settings.agent_speaks_first`, `settings.max_assistant_turns`\n- `system_tools.end_call`: let the agent end the call\n- `data_extraction_fields`: `[{name, type, description, required}]`\n\n```json\n{\n  \"system_prompt\": \"You are a helpful support agent.\",\n  \"llm\": {\"model\": \"openai/gpt-4.1\"},\n  \"stt\": {\"provider\": \"deepgram\"},\n  \"tts\": {\"provider\": \"elevenlabs\"},\n  \"settings\": {\"agent_speaks_first\": true, \"max_assistant_turns\": 50}\n}\n```\n\n**`type=connection`**, your own HTTP endpoint:\n- `agent_url`: public HTTP(S) endpoint your agent is called at\n- `agent_headers`: headers sent on each request, e.g. auth\n- `benchmark_provider`: `openrouter` by default. Other values: `openai`, `google`, `anthropic`, `meta-llama`, `mistralai`, `deepseek`, `x-ai`, `cohere`, `qwen`, or `ai21`\n\n```json\n{\n  \"agent_url\": \"https://api.example.com/agent\",\n  \"agent_headers\": {\"Authorization\": \"Bearer <token>\"},\n  \"benchmark_provider\": \"openrouter\"\n}\n```\n\n**Either type**:\n- `traces.scoring.enabled`: whether the traces you send for this agent are scored by its linked evaluators. On unless you set it to `false`. Turning it back on needs at least one linked evaluator that can score traces\n\n```json\n{\"traces\": {\"scoring\": {\"enabled\": false}}}\n```\n\nEvery request Calibrate makes to your endpoint carries the header\n`X-Calibrate-Eval: 1`. Read it to tell a test run from a real user, for example\nto tag the trace you send back or to skip sending one.\n\nReplaces the stored config. Omit to leave unchanged\n\nFor `type=connection`, changing `agent_url` or `agent_headers` resets the connection and benchmark verification flags",
   ),
   name: z.string().nullable().optional().describe(
     "New agent name. Omit to leave the name unchanged",