thinkwork-cli 0.8.2 → 0.9.0

package/dist/cli.js

@@ -68,6 +68,32 @@ function printKeyValue(pairs) {
     console.log(`  ${label}${v ?? chalk.dim("\u2014")}`);
   }
 }
+function printTable(rows, columns) {
+  if (jsonMode) return;
+  if (rows.length === 0) {
+    console.log(chalk.dim("  (no results)"));
+    return;
+  }
+  const widths = columns.map((c) => {
+    const header2 = c.header.length;
+    const maxRow = Math.max(
+      ...rows.map((r) => (c.format ? c.format(r[c.key]) : String(r[c.key] ?? "")).length)
+    );
+    return Math.max(header2, maxRow);
+  });
+  const header = columns.map((c, i) => chalk.bold(c.header.padEnd(widths[i]))).join("  ");
+  console.log(`  ${header}`);
+  console.log(
+    "  " + widths.map((w) => chalk.dim("\u2500".repeat(w))).join("  ")
+  );
+  for (const row of rows) {
+    const line = columns.map((c, i) => {
+      const v = c.format ? c.format(row[c.key]) : String(row[c.key] ?? "");
+      return v.padEnd(widths[i]);
+    }).join("  ");
+    console.log(`  ${line}`);
+  }
+}
 function logStderr(message) {
   process.stderr.write(message + "\n");
 }
@@ -312,6 +338,18 @@ function requireTty(label) {
     process.exit(1);
   }
 }
+async function promptOrExit(fn) {
+  try {
+    return await fn();
+  } catch (err) {
+    if (isCancellation(err)) {
+      console.log("");
+      console.log("  Cancelled.");
+      process.exit(0);
+    }
+    throw err;
+  }
+}
 
 // src/lib/resolve-stage.ts
 async function resolveStage(opts = {}) {
@@ -2049,9 +2087,9 @@ function registerInitCommand(program2) {
         printError("Stage name is required. Pass -s <name> or re-run in an interactive terminal.");
         process.exit(1);
       }
-      const { input: input3 } = await import("@inquirer/prompts");
+      const { input: input5 } = await import("@inquirer/prompts");
       try {
-        stage = await input3({
+        stage = await input5({
           message: "Stage name (e.g. dev, staging, prod):",
           validate: (v) => validateStage(v).error ?? true
         });
@@ -2912,13 +2950,13 @@ Examples:
       --auth-type tenant_api_key --api-key sk-abc -s dev -t acme
 
   # OAuth connector (users connect from the mobile app)
-  $ thinkwork mcp add lastmile --url https://mcp-dev.lastmile-tei.com/crm \\
+  $ thinkwork mcp add lastmile --url https://dev-mcp.lastmile-tei.com/crm \\
       --auth-type per_user_oauth --oauth-provider lastmile -s dev -t acme
 `
   ).action(
     async (nameArg, opts) => {
       try {
-        const { input: input3 } = await import("@inquirer/prompts");
+        const { input: input5 } = await import("@inquirer/prompts");
         const { stage, api, tenant } = await resolveMcpContext(opts);
         let name = nameArg;
         if (!name) {
@@ -2926,7 +2964,7 @@ Examples:
             printError("Name is required. Pass it as a positional arg.");
             process.exit(1);
           }
-          name = await input3({ message: "Server name:" });
+          name = await input5({ message: "Server name:" });
         }
         let url = opts.url;
         if (!url) {
@@ -2934,7 +2972,7 @@ Examples:
             printError("--url is required. Pass it as a flag.");
             process.exit(1);
           }
-          url = await input3({
+          url = await input5({
             message: "MCP server URL:",
             validate: (v) => v.startsWith("http://") || v.startsWith("https://") ? true : "URL must start with http:// or https://"
           });
@@ -3094,7 +3132,7 @@ Examples:
   ).option("-s, --stage <name>", "Deployment stage").option("-t, --tenant <slug>", "Tenant slug").option("--agent <id>", "Agent ID").action(
     async (mcpServerArg, opts) => {
       try {
-        const { input: input3 } = await import("@inquirer/prompts");
+        const { input: input5 } = await import("@inquirer/prompts");
         const { api, tenant } = await resolveMcpContext(opts);
         const server = await resolveServer(mcpServerArg, api, tenant.slug);
         let agent = opts.agent;
@@ -3103,7 +3141,7 @@ Examples:
             printError("--agent is required. Pass it as a flag.");
             process.exit(1);
           }
-          agent = await input3({ message: "Agent ID:" });
+          agent = await input5({ message: "Agent ID:" });
         }
         const result = await apiFetch(
           api.apiUrl,
@@ -3126,7 +3164,7 @@ Examples:
   ).option("-s, --stage <name>", "Deployment stage").option("-t, --tenant <slug>", "Tenant slug").option("--agent <id>", "Agent ID").action(
     async (mcpServerArg, opts) => {
       try {
-        const { input: input3 } = await import("@inquirer/prompts");
+        const { input: input5 } = await import("@inquirer/prompts");
         const { api, tenant } = await resolveMcpContext(opts);
         const server = await resolveServer(mcpServerArg, api, tenant.slug);
         let agent = opts.agent;
@@ -3135,7 +3173,7 @@ Examples:
             printError("--agent is required. Pass it as a flag.");
             process.exit(1);
           }
-          agent = await input3({ message: "Agent ID:" });
+          agent = await input5({ message: "Agent ID:" });
         }
         await apiFetch(
           api.apiUrl,
@@ -3866,6 +3904,24 @@ async function getGqlClient(opts) {
     tenantSlug: auth.tenantSlug
   };
 }
+async function gqlQuery(client, doc, variables) {
+  const res = await client.query(doc, variables).toPromise();
+  return unwrap(res);
+}
+async function gqlMutate(client, doc, variables) {
+  const res = await client.mutation(doc, variables).toPromise();
+  return unwrap(res);
+}
+function unwrap(res) {
+  if (res.error) {
+    const msg = res.error.graphQLErrors.map((e) => e.message).filter(Boolean).join("; ") || res.error.networkError?.message || "GraphQL request failed";
+    throw new Error(msg);
+  }
+  if (!res.data) {
+    throw new Error("GraphQL request returned no data.");
+  }
+  return res.data;
+}
 
 // src/commands/me.ts
 var ME_QUERY = gql`
@@ -4612,6 +4668,1865 @@ Examples:
   ).action(() => notYetImplemented("dashboard", 5));
 }
 
+// src/commands/eval/run.ts
+import { select as select8, checkbox, confirm as confirm2, input as input3 } from "@inquirer/prompts";
+import ora2 from "ora";
+
+// src/gql/graphql.ts
+var CliEvalRunsDocument = { "kind": "Document", "definitions": [{ "kind": "OperationDefinition", "operation": "query", "name": { "kind": "Name", "value": "CliEvalRuns" }, "variableDefinitions": [{ "kind": "VariableDefinition", "variable": { "kind": "Variable", "name": { "kind": "Name", "value": "tenantId" } }, "type": { "kind": "NonNullType", "type": { "kind": "NamedType", "name": { "kind": "Name", "value": "ID" } } } }, { "kind": "VariableDefinition", "variable": { "kind": "Variable", "name": { "kind": "Name", "value": "agentId" } }, "type": { "kind": "NamedType", "name": { "kind": "Name", "value": "ID" } } }, { "kind": "VariableDefinition", "variable": { "kind": "Variable", "name": { "kind": "Name", "value": "limit" } }, "type": { "kind": "NamedType", "name": { "kind": "Name", "value": "Int" } } }, { "kind": "VariableDefinition", "variable": { "kind": "Variable", "name": { "kind": "Name", "value": "offset" } }, "type": { "kind": "NamedType", "name": { "kind": "Name", "value": "Int" } } }], "selectionSet": { "kind": "SelectionSet", "selections": [{ "kind": "Field", "name": { "kind": "Name", "value": "evalRuns" }, "arguments": [{ "kind": "Argument", "name": { "kind": "Name", "value": "tenantId" }, "value": { "kind": "Variable", "name": { "kind": "Name", "value": "tenantId" } } }, { "kind": "Argument", "name": { "kind": "Name", "value": "agentId" }, "value": { "kind": "Variable", "name": { "kind": "Name", "value": "agentId" } } }, { "kind": "Argument", "name": { "kind": "Name", "value": "limit" }, "value": { "kind": "Variable", "name": { "kind": "Name", "value": "limit" } } }, { "kind": "Argument", "name": { "kind": "Name", "value": "offset" }, "value": { "kind": "Variable", "name": { "kind": "Name", "value": "offset" } } }], "selectionSet": { "kind": "SelectionSet", "selections": [{ "kind": "Field", "name": { "kind": "Name", "value": "totalCount" } }, { "kind": "Field", "name": { "kind": "Name", "value": "items" }, "selectionSet": { "kind": "SelectionSet", "selections": [{ "kind": "Field", "name": { "kind": "Name", "value": "id" } }, { "kind": "Field", "name": { "kind": "Name", "value": "status" } }, { "kind": "Field", "name": { "kind": "Name", "value": "model" } }, { "kind": "Field", "name": { "kind": "Name", "value": "categories" } }, { "kind": "Field", "name": { "kind": "Name", "value": "agentId" } }, { "kind": "Field", "name": { "kind": "Name", "value": "agentName" } }, { "kind": "Field", "name": { "kind": "Name", "value": "agentTemplateId" } }, { "kind": "Field", "name": { "kind": "Name", "value": "agentTemplateName" } }, { "kind": "Field", "name": { "kind": "Name", "value": "totalTests" } }, { "kind": "Field", "name": { "kind": "Name", "value": "passed" } }, { "kind": "Field", "name": { "kind": "Name", "value": "failed" } }, { "kind": "Field", "name": { "kind": "Name", "value": "passRate" } }, { "kind": "Field", "name": { "kind": "Name", "value": "regression" } }, { "kind": "Field", "name": { "kind": "Name", "value": "costUsd" } }, { "kind": "Field", "name": { "kind": "Name", "value": "errorMessage" } }, { "kind": "Field", "name": { "kind": "Name", "value": "startedAt" } }, { "kind": "Field", "name": { "kind": "Name", "value": "completedAt" } }, { "kind": "Field", "name": { "kind": "Name", "value": "createdAt" } }] } }] } }] } }] };
+var CliEvalRunDocument = { "kind": "Document", "definitions": [{ "kind": "OperationDefinition", "operation": "query", "name": { "kind": "Name", "value": "CliEvalRun" }, "variableDefinitions": [{ "kind": "VariableDefinition", "variable": { "kind": "Variable", "name": { "kind": "Name", "value": "id" } }, "type": { "kind": "NonNullType", "type": { "kind": "NamedType", "name": { "kind": "Name", "value": "ID" } } } }], "selectionSet": { "kind": "SelectionSet", "selections": [{ "kind": "Field", "name": { "kind": "Name", "value": "evalRun" }, "arguments": [{ "kind": "Argument", "name": { "kind": "Name", "value": "id" }, "value": { "kind": "Variable", "name": { "kind": "Name", "value": "id" } } }], "selectionSet": { "kind": "SelectionSet", "selections": [{ "kind": "Field", "name": { "kind": "Name", "value": "id" } }, { "kind": "Field", "name": { "kind": "Name", "value": "status" } }, { "kind": "Field", "name": { "kind": "Name", "value": "model" } }, { "kind": "Field", "name": { "kind": "Name", "value": "categories" } }, { "kind": "Field", "name": { "kind": "Name", "value": "agentId" } }, { "kind": "Field", "name": { "kind": "Name", "value": "agentName" } }, { "kind": "Field", "name": { "kind": "Name", "value": "agentTemplateId" } }, { "kind": "Field", "name": { "kind": "Name", "value": "agentTemplateName" } }, { "kind": "Field", "name": { "kind": "Name", "value": "totalTests" } }, { "kind": "Field", "name": { "kind": "Name", "value": "passed" } }, { "kind": "Field", "name": { "kind": "Name", "value": "failed" } }, { "kind": "Field", "name": { "kind": "Name", "value": "passRate" } }, { "kind": "Field", "name": { "kind": "Name", "value": "regression" } }, { "kind": "Field", "name": { "kind": "Name", "value": "costUsd" } }, { "kind": "Field", "name": { "kind": "Name", "value": "errorMessage" } }, { "kind": "Field", "name": { "kind": "Name", "value": "startedAt" } }, { "kind": "Field", "name": { "kind": "Name", "value": "completedAt" } }, { "kind": "Field", "name": { "kind": "Name", "value": "createdAt" } }] } }] } }] };
+var CliEvalRunResultsDocument = { "kind": "Document", "definitions": [{ "kind": "OperationDefinition", "operation": "query", "name": { "kind": "Name", "value": "CliEvalRunResults" }, "variableDefinitions": [{ "kind": "VariableDefinition", "variable": { "kind": "Variable", "name": { "kind": "Name", "value": "runId" } }, "type": { "kind": "NonNullType", "type": { "kind": "NamedType", "name": { "kind": "Name", "value": "ID" } } } }], "selectionSet": { "kind": "SelectionSet", "selections": [{ "kind": "Field", "name": { "kind": "Name", "value": "evalRunResults" }, "arguments": [{ "kind": "Argument", "name": { "kind": "Name", "value": "runId" }, "value": { "kind": "Variable", "name": { "kind": "Name", "value": "runId" } } }], "selectionSet": { "kind": "SelectionSet", "selections": [{ "kind": "Field", "name": { "kind": "Name", "value": "id" } }, { "kind": "Field", "name": { "kind": "Name", "value": "testCaseId" } }, { "kind": "Field", "name": { "kind": "Name", "value": "testCaseName" } }, { "kind": "Field", "name": { "kind": "Name", "value": "category" } }, { "kind": "Field", "name": { "kind": "Name", "value": "status" } }, { "kind": "Field", "name": { "kind": "Name", "value": "score" } }, { "kind": "Field", "name": { "kind": "Name", "value": "durationMs" } }, { "kind": "Field", "name": { "kind": "Name", "value": "agentSessionId" } }, { "kind": "Field", "name": { "kind": "Name", "value": "input" } }, { "kind": "Field", "name": { "kind": "Name", "value": "expected" } }, { "kind": "Field", "name": { "kind": "Name", "value": "actualOutput" } }, { "kind": "Field", "name": { "kind": "Name", "value": "evaluatorResults" } }, { "kind": "Field", "name": { "kind": "Name", "value": "assertions" } }, { "kind": "Field", "name": { "kind": "Name", "value": "errorMessage" } }, { "kind": "Field", "name": { "kind": "Name", "value": "createdAt" } }] } }] } }] };
+var CliEvalTestCasesDocument = { "kind": "Document", "definitions": [{ "kind": "OperationDefinition", "operation": "query", "name": { "kind": "Name", "value": "CliEvalTestCases" }, "variableDefinitions": [{ "kind": "VariableDefinition", "variable": { "kind": "Variable", "name": { "kind": "Name", "value": "tenantId" } }, "type": { "kind": "NonNullType", "type": { "kind": "NamedType", "name": { "kind": "Name", "value": "ID" } } } }, { "kind": "VariableDefinition", "variable": { "kind": "Variable", "name": { "kind": "Name", "value": "category" } }, "type": { "kind": "NamedType", "name": { "kind": "Name", "value": "String" } } }, { "kind": "VariableDefinition", "variable": { "kind": "Variable", "name": { "kind": "Name", "value": "search" } }, "type": { "kind": "NamedType", "name": { "kind": "Name", "value": "String" } } }], "selectionSet": { "kind": "SelectionSet", "selections": [{ "kind": "Field", "name": { "kind": "Name", "value": "evalTestCases" }, "arguments": [{ "kind": "Argument", "name": { "kind": "Name", "value": "tenantId" }, "value": { "kind": "Variable", "name": { "kind": "Name", "value": "tenantId" } } }, { "kind": "Argument", "name": { "kind": "Name", "value": "category" }, "value": { "kind": "Variable", "name": { "kind": "Name", "value": "category" } } }, { "kind": "Argument", "name": { "kind": "Name", "value": "search" }, "value": { "kind": "Variable", "name": { "kind": "Name", "value": "search" } } }], "selectionSet": { "kind": "SelectionSet", "selections": [{ "kind": "Field", "name": { "kind": "Name", "value": "id" } }, { "kind": "Field", "name": { "kind": "Name", "value": "name" } }, { "kind": "Field", "name": { "kind": "Name", "value": "category" } }, { "kind": "Field", "name": { "kind": "Name", "value": "query" } }, { "kind": "Field", "name": { "kind": "Name", "value": "systemPrompt" } }, { "kind": "Field", "name": { "kind": "Name", "value": "agentTemplateId" } }, { "kind": "Field", "name": { "kind": "Name", "value": "agentTemplateName" } }, { "kind": "Field", "name": { "kind": "Name", "value": "agentcoreEvaluatorIds" } }, { "kind": "Field", "name": { "kind": "Name", "value": "tags" } }, { "kind": "Field", "name": { "kind": "Name", "value": "enabled" } }, { "kind": "Field", "name": { "kind": "Name", "value": "source" } }, { "kind": "Field", "name": { "kind": "Name", "value": "createdAt" } }, { "kind": "Field", "name": { "kind": "Name", "value": "updatedAt" } }] } }] } }] };
+var CliEvalTestCaseDocument = { "kind": "Document", "definitions": [{ "kind": "OperationDefinition", "operation": "query", "name": { "kind": "Name", "value": "CliEvalTestCase" }, "variableDefinitions": [{ "kind": "VariableDefinition", "variable": { "kind": "Variable", "name": { "kind": "Name", "value": "id" } }, "type": { "kind": "NonNullType", "type": { "kind": "NamedType", "name": { "kind": "Name", "value": "ID" } } } }], "selectionSet": { "kind": "SelectionSet", "selections": [{ "kind": "Field", "name": { "kind": "Name", "value": "evalTestCase" }, "arguments": [{ "kind": "Argument", "name": { "kind": "Name", "value": "id" }, "value": { "kind": "Variable", "name": { "kind": "Name", "value": "id" } } }], "selectionSet": { "kind": "SelectionSet", "selections": [{ "kind": "Field", "name": { "kind": "Name", "value": "id" } }, { "kind": "Field", "name": { "kind": "Name", "value": "tenantId" } }, { "kind": "Field", "name": { "kind": "Name", "value": "name" } }, { "kind": "Field", "name": { "kind": "Name", "value": "category" } }, { "kind": "Field", "name": { "kind": "Name", "value": "query" } }, { "kind": "Field", "name": { "kind": "Name", "value": "systemPrompt" } }, { "kind": "Field", "name": { "kind": "Name", "value": "agentTemplateId" } }, { "kind": "Field", "name": { "kind": "Name", "value": "agentTemplateName" } }, { "kind": "Field", "name": { "kind": "Name", "value": "assertions" } }, { "kind": "Field", "name": { "kind": "Name", "value": "agentcoreEvaluatorIds" } }, { "kind": "Field", "name": { "kind": "Name", "value": "tags" } }, { "kind": "Field", "name": { "kind": "Name", "value": "enabled" } }, { "kind": "Field", "name": { "kind": "Name", "value": "source" } }, { "kind": "Field", "name": { "kind": "Name", "value": "createdAt" } }, { "kind": "Field", "name": { "kind": "Name", "value": "updatedAt" } }] } }] } }] };
+var CliAgentTemplatesForEvalDocument = { "kind": "Document", "definitions": [{ "kind": "OperationDefinition", "operation": "query", "name": { "kind": "Name", "value": "CliAgentTemplatesForEval" }, "variableDefinitions": [{ "kind": "VariableDefinition", "variable": { "kind": "Variable", "name": { "kind": "Name", "value": "tenantId" } }, "type": { "kind": "NonNullType", "type": { "kind": "NamedType", "name": { "kind": "Name", "value": "ID" } } } }], "selectionSet": { "kind": "SelectionSet", "selections": [{ "kind": "Field", "name": { "kind": "Name", "value": "agentTemplates" }, "arguments": [{ "kind": "Argument", "name": { "kind": "Name", "value": "tenantId" }, "value": { "kind": "Variable", "name": { "kind": "Name", "value": "tenantId" } } }], "selectionSet": { "kind": "SelectionSet", "selections": [{ "kind": "Field", "name": { "kind": "Name", "value": "id" } }, { "kind": "Field", "name": { "kind": "Name", "value": "name" } }, { "kind": "Field", "name": { "kind": "Name", "value": "slug" } }, { "kind": "Field", "name": { "kind": "Name", "value": "model" } }, { "kind": "Field", "name": { "kind": "Name", "value": "isPublished" } }] } }] } }] };
+var CliTenantBySlugDocument = { "kind": "Document", "definitions": [{ "kind": "OperationDefinition", "operation": "query", "name": { "kind": "Name", "value": "CliTenantBySlug" }, "variableDefinitions": [{ "kind": "VariableDefinition", "variable": { "kind": "Variable", "name": { "kind": "Name", "value": "slug" } }, "type": { "kind": "NonNullType", "type": { "kind": "NamedType", "name": { "kind": "Name", "value": "String" } } } }], "selectionSet": { "kind": "SelectionSet", "selections": [{ "kind": "Field", "name": { "kind": "Name", "value": "tenantBySlug" }, "arguments": [{ "kind": "Argument", "name": { "kind": "Name", "value": "slug" }, "value": { "kind": "Variable", "name": { "kind": "Name", "value": "slug" } } }], "selectionSet": { "kind": "SelectionSet", "selections": [{ "kind": "Field", "name": { "kind": "Name", "value": "id" } }, { "kind": "Field", "name": { "kind": "Name", "value": "slug" } }, { "kind": "Field", "name": { "kind": "Name", "value": "name" } }] } }] } }] };
+var CliStartEvalRunDocument = { "kind": "Document", "definitions": [{ "kind": "OperationDefinition", "operation": "mutation", "name": { "kind": "Name", "value": "CliStartEvalRun" }, "variableDefinitions": [{ "kind": "VariableDefinition", "variable": { "kind": "Variable", "name": { "kind": "Name", "value": "tenantId" } }, "type": { "kind": "NonNullType", "type": { "kind": "NamedType", "name": { "kind": "Name", "value": "ID" } } } }, { "kind": "VariableDefinition", "variable": { "kind": "Variable", "name": { "kind": "Name", "value": "input" } }, "type": { "kind": "NonNullType", "type": { "kind": "NamedType", "name": { "kind": "Name", "value": "StartEvalRunInput" } } } }], "selectionSet": { "kind": "SelectionSet", "selections": [{ "kind": "Field", "name": { "kind": "Name", "value": "startEvalRun" }, "arguments": [{ "kind": "Argument", "name": { "kind": "Name", "value": "tenantId" }, "value": { "kind": "Variable", "name": { "kind": "Name", "value": "tenantId" } } }, { "kind": "Argument", "name": { "kind": "Name", "value": "input" }, "value": { "kind": "Variable", "name": { "kind": "Name", "value": "input" } } }], "selectionSet": { "kind": "SelectionSet", "selections": [{ "kind": "Field", "name": { "kind": "Name", "value": "id" } }, { "kind": "Field", "name": { "kind": "Name", "value": "status" } }, { "kind": "Field", "name": { "kind": "Name", "value": "model" } }, { "kind": "Field", "name": { "kind": "Name", "value": "categories" } }, { "kind": "Field", "name": { "kind": "Name", "value": "agentTemplateId" } }, { "kind": "Field", "name": { "kind": "Name", "value": "agentTemplateName" } }, { "kind": "Field", "name": { "kind": "Name", "value": "totalTests" } }, { "kind": "Field", "name": { "kind": "Name", "value": "createdAt" } }] } }] } }] };
+var CliCancelEvalRunDocument = { "kind": "Document", "definitions": [{ "kind": "OperationDefinition", "operation": "mutation", "name": { "kind": "Name", "value": "CliCancelEvalRun" }, "variableDefinitions": [{ "kind": "VariableDefinition", "variable": { "kind": "Variable", "name": { "kind": "Name", "value": "id" } }, "type": { "kind": "NonNullType", "type": { "kind": "NamedType", "name": { "kind": "Name", "value": "ID" } } } }], "selectionSet": { "kind": "SelectionSet", "selections": [{ "kind": "Field", "name": { "kind": "Name", "value": "cancelEvalRun" }, "arguments": [{ "kind": "Argument", "name": { "kind": "Name", "value": "id" }, "value": { "kind": "Variable", "name": { "kind": "Name", "value": "id" } } }], "selectionSet": { "kind": "SelectionSet", "selections": [{ "kind": "Field", "name": { "kind": "Name", "value": "id" } }, { "kind": "Field", "name": { "kind": "Name", "value": "status" } }, { "kind": "Field", "name": { "kind": "Name", "value": "completedAt" } }] } }] } }] };
+var CliDeleteEvalRunDocument = { "kind": "Document", "definitions": [{ "kind": "OperationDefinition", "operation": "mutation", "name": { "kind": "Name", "value": "CliDeleteEvalRun" }, "variableDefinitions": [{ "kind": "VariableDefinition", "variable": { "kind": "Variable", "name": { "kind": "Name", "value": "id" } }, "type": { "kind": "NonNullType", "type": { "kind": "NamedType", "name": { "kind": "Name", "value": "ID" } } } }], "selectionSet": { "kind": "SelectionSet", "selections": [{ "kind": "Field", "name": { "kind": "Name", "value": "deleteEvalRun" }, "arguments": [{ "kind": "Argument", "name": { "kind": "Name", "value": "id" }, "value": { "kind": "Variable", "name": { "kind": "Name", "value": "id" } } }] }] } }] };
+var CliCreateEvalTestCaseDocument = { "kind": "Document", "definitions": [{ "kind": "OperationDefinition", "operation": "mutation", "name": { "kind": "Name", "value": "CliCreateEvalTestCase" }, "variableDefinitions": [{ "kind": "VariableDefinition", "variable": { "kind": "Variable", "name": { "kind": "Name", "value": "tenantId" } }, "type": { "kind": "NonNullType", "type": { "kind": "NamedType", "name": { "kind": "Name", "value": "ID" } } } }, { "kind": "VariableDefinition", "variable": { "kind": "Variable", "name": { "kind": "Name", "value": "input" } }, "type": { "kind": "NonNullType", "type": { "kind": "NamedType", "name": { "kind": "Name", "value": "CreateEvalTestCaseInput" } } } }], "selectionSet": { "kind": "SelectionSet", "selections": [{ "kind": "Field", "name": { "kind": "Name", "value": "createEvalTestCase" }, "arguments": [{ "kind": "Argument", "name": { "kind": "Name", "value": "tenantId" }, "value": { "kind": "Variable", "name": { "kind": "Name", "value": "tenantId" } } }, { "kind": "Argument", "name": { "kind": "Name", "value": "input" }, "value": { "kind": "Variable", "name": { "kind": "Name", "value": "input" } } }], "selectionSet": { "kind": "SelectionSet", "selections": [{ "kind": "Field", "name": { "kind": "Name", "value": "id" } }, { "kind": "Field", "name": { "kind": "Name", "value": "name" } }, { "kind": "Field", "name": { "kind": "Name", "value": "category" } }] } }] } }] };
+var CliUpdateEvalTestCaseDocument = { "kind": "Document", "definitions": [{ "kind": "OperationDefinition", "operation": "mutation", "name": { "kind": "Name", "value": "CliUpdateEvalTestCase" }, "variableDefinitions": [{ "kind": "VariableDefinition", "variable": { "kind": "Variable", "name": { "kind": "Name", "value": "id" } }, "type": { "kind": "NonNullType", "type": { "kind": "NamedType", "name": { "kind": "Name", "value": "ID" } } } }, { "kind": "VariableDefinition", "variable": { "kind": "Variable", "name": { "kind": "Name", "value": "input" } }, "type": { "kind": "NonNullType", "type": { "kind": "NamedType", "name": { "kind": "Name", "value": "UpdateEvalTestCaseInput" } } } }], "selectionSet": { "kind": "SelectionSet", "selections": [{ "kind": "Field", "name": { "kind": "Name", "value": "updateEvalTestCase" }, "arguments": [{ "kind": "Argument", "name": { "kind": "Name", "value": "id" }, "value": { "kind": "Variable", "name": { "kind": "Name", "value": "id" } } }, { "kind": "Argument", "name": { "kind": "Name", "value": "input" }, "value": { "kind": "Variable", "name": { "kind": "Name", "value": "input" } } }], "selectionSet": { "kind": "SelectionSet", "selections": [{ "kind": "Field", "name": { "kind": "Name", "value": "id" } }, { "kind": "Field", "name": { "kind": "Name", "value": "name" } }, { "kind": "Field", "name": { "kind": "Name", "value": "category" } }, { "kind": "Field", "name": { "kind": "Name", "value": "enabled" } }] } }] } }] };
+var CliDeleteEvalTestCaseDocument = { "kind": "Document", "definitions": [{ "kind": "OperationDefinition", "operation": "mutation", "name": { "kind": "Name", "value": "CliDeleteEvalTestCase" }, "variableDefinitions": [{ "kind": "VariableDefinition", "variable": { "kind": "Variable", "name": { "kind": "Name", "value": "id" } }, "type": { "kind": "NonNullType", "type": { "kind": "NamedType", "name": { "kind": "Name", "value": "ID" } } } }], "selectionSet": { "kind": "SelectionSet", "selections": [{ "kind": "Field", "name": { "kind": "Name", "value": "deleteEvalTestCase" }, "arguments": [{ "kind": "Argument", "name": { "kind": "Name", "value": "id" }, "value": { "kind": "Variable", "name": { "kind": "Name", "value": "id" } } }] }] } }] };
+var CliSeedEvalTestCasesDocument = { "kind": "Document", "definitions": [{ "kind": "OperationDefinition", "operation": "mutation", "name": { "kind": "Name", "value": "CliSeedEvalTestCases" }, "variableDefinitions": [{ "kind": "VariableDefinition", "variable": { "kind": "Variable", "name": { "kind": "Name", "value": "tenantId" } }, "type": { "kind": "NonNullType", "type": { "kind": "NamedType", "name": { "kind": "Name", "value": "ID" } } } }, { "kind": "VariableDefinition", "variable": { "kind": "Variable", "name": { "kind": "Name", "value": "categories" } }, "type": { "kind": "ListType", "type": { "kind": "NonNullType", "type": { "kind": "NamedType", "name": { "kind": "Name", "value": "String" } } } } }], "selectionSet": { "kind": "SelectionSet", "selections": [{ "kind": "Field", "name": { "kind": "Name", "value": "seedEvalTestCases" }, "arguments": [{ "kind": "Argument", "name": { "kind": "Name", "value": "tenantId" }, "value": { "kind": "Variable", "name": { "kind": "Name", "value": "tenantId" } } }, { "kind": "Argument", "name": { "kind": "Name", "value": "categories" }, "value": { "kind": "Variable", "name": { "kind": "Name", "value": "categories" } } }] }] } }] };
+var CliMeDocument = { "kind": "Document", "definitions": [{ "kind": "OperationDefinition", "operation": "query", "name": { "kind": "Name", "value": "CliMe" }, "selectionSet": { "kind": "SelectionSet", "selections": [{ "kind": "Field", "name": { "kind": "Name", "value": "me" }, "selectionSet": { "kind": "SelectionSet", "selections": [{ "kind": "Field", "name": { "kind": "Name", "value": "id" } }, { "kind": "Field", "name": { "kind": "Name", "value": "email" } }, { "kind": "Field", "name": { "kind": "Name", "value": "name" } }, { "kind": "Field", "name": { "kind": "Name", "value": "tenantId" } }] } }] } }] };
+var CliWikiTenantBySlugDocument = { "kind": "Document", "definitions": [{ "kind": "OperationDefinition", "operation": "query", "name": { "kind": "Name", "value": "CliWikiTenantBySlug" }, "variableDefinitions": [{ "kind": "VariableDefinition", "variable": { "kind": "Variable", "name": { "kind": "Name", "value": "slug" } }, "type": { "kind": "NonNullType", "type": { "kind": "NamedType", "name": { "kind": "Name", "value": "String" } } } }], "selectionSet": { "kind": "SelectionSet", "selections": [{ "kind": "Field", "name": { "kind": "Name", "value": "tenantBySlug" }, "arguments": [{ "kind": "Argument", "name": { "kind": "Name", "value": "slug" }, "value": { "kind": "Variable", "name": { "kind": "Name", "value": "slug" } } }], "selectionSet": { "kind": "SelectionSet", "selections": [{ "kind": "Field", "name": { "kind": "Name", "value": "id" } }, { "kind": "Field", "name": { "kind": "Name", "value": "slug" } }, { "kind": "Field", "name": { "kind": "Name", "value": "name" } }] } }] } }] };
+var CliAllTenantAgentsForWikiDocument = { "kind": "Document", "definitions": [{ "kind": "OperationDefinition", "operation": "query", "name": { "kind": "Name", "value": "CliAllTenantAgentsForWiki" }, "variableDefinitions": [{ "kind": "VariableDefinition", "variable": { "kind": "Variable", "name": { "kind": "Name", "value": "tenantId" } }, "type": { "kind": "NonNullType", "type": { "kind": "NamedType", "name": { "kind": "Name", "value": "ID" } } } }], "selectionSet": { "kind": "SelectionSet", "selections": [{ "kind": "Field", "name": { "kind": "Name", "value": "allTenantAgents" }, "arguments": [{ "kind": "Argument", "name": { "kind": "Name", "value": "tenantId" }, "value": { "kind": "Variable", "name": { "kind": "Name", "value": "tenantId" } } }, { "kind": "Argument", "name": { "kind": "Name", "value": "includeSystem" }, "value": { "kind": "BooleanValue", "value": false } }, { "kind": "Argument", "name": { "kind": "Name", "value": "includeSubAgents" }, "value": { "kind": "BooleanValue", "value": false } }], "selectionSet": { "kind": "SelectionSet", "selections": [{ "kind": "Field", "name": { "kind": "Name", "value": "id" } }, { "kind": "Field", "name": { "kind": "Name", "value": "name" } }, { "kind": "Field", "name": { "kind": "Name", "value": "slug" } }, { "kind": "Field", "name": { "kind": "Name", "value": "type" } }, { "kind": "Field", "name": { "kind": "Name", "value": "status" } }] } }] } }] };
+var CliCompileWikiNowDocument = { "kind": "Document", "definitions": [{ "kind": "OperationDefinition", "operation": "mutation", "name": { "kind": "Name", "value": "CliCompileWikiNow" }, "variableDefinitions": [{ "kind": "VariableDefinition", "variable": { "kind": "Variable", "name": { "kind": "Name", "value": "tenantId" } }, "type": { "kind": "NonNullType", "type": { "kind": "NamedType", "name": { "kind": "Name", "value": "ID" } } } }, { "kind": "VariableDefinition", "variable": { "kind": "Variable", "name": { "kind": "Name", "value": "ownerId" } }, "type": { "kind": "NonNullType", "type": { "kind": "NamedType", "name": { "kind": "Name", "value": "ID" } } } }, { "kind": "VariableDefinition", "variable": { "kind": "Variable", "name": { "kind": "Name", "value": "modelId" } }, "type": { "kind": "NamedType", "name": { "kind": "Name", "value": "String" } } }], "selectionSet": { "kind": "SelectionSet", "selections": [{ "kind": "Field", "name": { "kind": "Name", "value": "compileWikiNow" }, "arguments": [{ "kind": "Argument", "name": { "kind": "Name", "value": "tenantId" }, "value": { "kind": "Variable", "name": { "kind": "Name", "value": "tenantId" } } }, { "kind": "Argument", "name": { "kind": "Name", "value": "ownerId" }, "value": { "kind": "Variable", "name": { "kind": "Name", "value": "ownerId" } } }, { "kind": "Argument", "name": { "kind": "Name", "value": "modelId" }, "value": { "kind": "Variable", "name": { "kind": "Name", "value": "modelId" } } }], "selectionSet": { "kind": "SelectionSet", "selections": [{ "kind": "Field", "name": { "kind": "Name", "value": "id" } }, { "kind": "Field", "name": { "kind": "Name", "value": "tenantId" } }, { "kind": "Field", "name": { "kind": "Name", "value": "ownerId" } }, { "kind": "Field", "name": { "kind": "Name", "value": "status" } }, { "kind": "Field", "name": { "kind": "Name", "value": "trigger" } }, { "kind": "Field", "name": { "kind": "Name", "value": "dedupeKey" } }, { "kind": "Field", "name": { "kind": "Name", "value": "attempt" } }, { "kind": "Field", "name": { "kind": "Name", "value": "createdAt" } }] } }] } }] };
+var CliResetWikiCursorDocument = { "kind": "Document", "definitions": [{ "kind": "OperationDefinition", "operation": "mutation", "name": { "kind": "Name", "value": "CliResetWikiCursor" }, "variableDefinitions": [{ "kind": "VariableDefinition", "variable": { "kind": "Variable", "name": { "kind": "Name", "value": "tenantId" } }, "type": { "kind": "NonNullType", "type": { "kind": "NamedType", "name": { "kind": "Name", "value": "ID" } } } }, { "kind": "VariableDefinition", "variable": { "kind": "Variable", "name": { "kind": "Name", "value": "ownerId" } }, "type": { "kind": "NonNullType", "type": { "kind": "NamedType", "name": { "kind": "Name", "value": "ID" } } } }, { "kind": "VariableDefinition", "variable": { "kind": "Variable", "name": { "kind": "Name", "value": "force" } }, "type": { "kind": "NamedType", "name": { "kind": "Name", "value": "Boolean" } } }], "selectionSet": { "kind": "SelectionSet", "selections": [{ "kind": "Field", "name": { "kind": "Name", "value": "resetWikiCursor" }, "arguments": [{ "kind": "Argument", "name": { "kind": "Name", "value": "tenantId" }, "value": { "kind": "Variable", "name": { "kind": "Name", "value": "tenantId" } } }, { "kind": "Argument", "name": { "kind": "Name", "value": "ownerId" }, "value": { "kind": "Variable", "name": { "kind": "Name", "value": "ownerId" } } }, { "kind": "Argument", "name": { "kind": "Name", "value": "force" }, "value": { "kind": "Variable", "name": { "kind": "Name", "value": "force" } } }], "selectionSet": { "kind": "SelectionSet", "selections": [{ "kind": "Field", "name": { "kind": "Name", "value": "tenantId" } }, { "kind": "Field", "name": { "kind": "Name", "value": "ownerId" } }, { "kind": "Field", "name": { "kind": "Name", "value": "cursorCleared" } }, { "kind": "Field", "name": { "kind": "Name", "value": "pagesArchived" } }] } }] } }] };
+var CliWikiCompileJobsDocument = { "kind": "Document", "definitions": [{ "kind": "OperationDefinition", "operation": "query", "name": { "kind": "Name", "value": "CliWikiCompileJobs" }, "variableDefinitions": [{ "kind": "VariableDefinition", "variable": { "kind": "Variable", "name": { "kind": "Name", "value": "tenantId" } }, "type": { "kind": "NonNullType", "type": { "kind": "NamedType", "name": { "kind": "Name", "value": "ID" } } } }, { "kind": "VariableDefinition", "variable": { "kind": "Variable", "name": { "kind": "Name", "value": "ownerId" } }, "type": { "kind": "NamedType", "name": { "kind": "Name", "value": "ID" } } }, { "kind": "VariableDefinition", "variable": { "kind": "Variable", "name": { "kind": "Name", "value": "limit" } }, "type": { "kind": "NamedType", "name": { "kind": "Name", "value": "Int" } } }], "selectionSet": { "kind": "SelectionSet", "selections": [{ "kind": "Field", "name": { "kind": "Name", "value": "wikiCompileJobs" }, "arguments": [{ "kind": "Argument", "name": { "kind": "Name", "value": "tenantId" }, "value": { "kind": "Variable", "name": { "kind": "Name", "value": "tenantId" } } }, { "kind": "Argument", "name": { "kind": "Name", "value": "ownerId" }, "value": { "kind": "Variable", "name": { "kind": "Name", "value": "ownerId" } } }, { "kind": "Argument", "name": { "kind": "Name", "value": "limit" }, "value": { "kind": "Variable", "name": { "kind": "Name", "value": "limit" } } }], "selectionSet": { "kind": "SelectionSet", "selections": [{ "kind": "Field", "name": { "kind": "Name", "value": "id" } }, { "kind": "Field", "name": { "kind": "Name", "value": "tenantId" } }, { "kind": "Field", "name": { "kind": "Name", "value": "ownerId" } }, { "kind": "Field", "name": { "kind": "Name", "value": "status" } }, { "kind": "Field", "name": { "kind": "Name", "value": "trigger" } }, { "kind": "Field", "name": { "kind": "Name", "value": "dedupeKey" } }, { "kind": "Field", "name": { "kind": "Name", "value": "attempt" } }, { "kind": "Field", "name": { "kind": "Name", "value": "claimedAt" } }, { "kind": "Field", "name": { "kind": "Name", "value": "startedAt" } }, { "kind": "Field", "name": { "kind": "Name", "value": "finishedAt" } }, { "kind": "Field", "name": { "kind": "Name", "value": "error" } }, { "kind": "Field", "name": { "kind": "Name", "value": "metrics" } }, { "kind": "Field", "name": { "kind": "Name", "value": "createdAt" } }] } }] } }] };
+
+// src/gql/gql.ts
+var documents = {
+  "\n  query CliEvalRuns($tenantId: ID!, $agentId: ID, $limit: Int, $offset: Int) {\n    evalRuns(tenantId: $tenantId, agentId: $agentId, limit: $limit, offset: $offset) {\n      totalCount\n      items {\n        id\n        status\n        model\n        categories\n        agentId\n        agentName\n        agentTemplateId\n        agentTemplateName\n        totalTests\n        passed\n        failed\n        passRate\n        regression\n        costUsd\n        errorMessage\n        startedAt\n        completedAt\n        createdAt\n      }\n    }\n  }\n": CliEvalRunsDocument,
+  "\n  query CliEvalRun($id: ID!) {\n    evalRun(id: $id) {\n      id\n      status\n      model\n      categories\n      agentId\n      agentName\n      agentTemplateId\n      agentTemplateName\n      totalTests\n      passed\n      failed\n      passRate\n      regression\n      costUsd\n      errorMessage\n      startedAt\n      completedAt\n      createdAt\n    }\n  }\n": CliEvalRunDocument,
+  "\n  query CliEvalRunResults($runId: ID!) {\n    evalRunResults(runId: $runId) {\n      id\n      testCaseId\n      testCaseName\n      category\n      status\n      score\n      durationMs\n      agentSessionId\n      input\n      expected\n      actualOutput\n      evaluatorResults\n      assertions\n      errorMessage\n      createdAt\n    }\n  }\n": CliEvalRunResultsDocument,
+  "\n  query CliEvalTestCases($tenantId: ID!, $category: String, $search: String) {\n    evalTestCases(tenantId: $tenantId, category: $category, search: $search) {\n      id\n      name\n      category\n      query\n      systemPrompt\n      agentTemplateId\n      agentTemplateName\n      agentcoreEvaluatorIds\n      tags\n      enabled\n      source\n      createdAt\n      updatedAt\n    }\n  }\n": CliEvalTestCasesDocument,
+  "\n  query CliEvalTestCase($id: ID!) {\n    evalTestCase(id: $id) {\n      id\n      tenantId\n      name\n      category\n      query\n      systemPrompt\n      agentTemplateId\n      agentTemplateName\n      assertions\n      agentcoreEvaluatorIds\n      tags\n      enabled\n      source\n      createdAt\n      updatedAt\n    }\n  }\n": CliEvalTestCaseDocument,
+  "\n  query CliAgentTemplatesForEval($tenantId: ID!) {\n    agentTemplates(tenantId: $tenantId) {\n      id\n      name\n      slug\n      model\n      isPublished\n    }\n  }\n": CliAgentTemplatesForEvalDocument,
+  "\n  query CliTenantBySlug($slug: String!) {\n    tenantBySlug(slug: $slug) {\n      id\n      slug\n      name\n    }\n  }\n": CliTenantBySlugDocument,
+  "\n  mutation CliStartEvalRun($tenantId: ID!, $input: StartEvalRunInput!) {\n    startEvalRun(tenantId: $tenantId, input: $input) {\n      id\n      status\n      model\n      categories\n      agentTemplateId\n      agentTemplateName\n      totalTests\n      createdAt\n    }\n  }\n": CliStartEvalRunDocument,
+  "\n  mutation CliCancelEvalRun($id: ID!) {\n    cancelEvalRun(id: $id) {\n      id\n      status\n      completedAt\n    }\n  }\n": CliCancelEvalRunDocument,
+  "\n  mutation CliDeleteEvalRun($id: ID!) {\n    deleteEvalRun(id: $id)\n  }\n": CliDeleteEvalRunDocument,
+  "\n  mutation CliCreateEvalTestCase($tenantId: ID!, $input: CreateEvalTestCaseInput!) {\n    createEvalTestCase(tenantId: $tenantId, input: $input) {\n      id\n      name\n      category\n    }\n  }\n": CliCreateEvalTestCaseDocument,
+  "\n  mutation CliUpdateEvalTestCase($id: ID!, $input: UpdateEvalTestCaseInput!) {\n    updateEvalTestCase(id: $id, input: $input) {\n      id\n      name\n      category\n      enabled\n    }\n  }\n": CliUpdateEvalTestCaseDocument,
+  "\n  mutation CliDeleteEvalTestCase($id: ID!) {\n    deleteEvalTestCase(id: $id)\n  }\n": CliDeleteEvalTestCaseDocument,
+  "\n  mutation CliSeedEvalTestCases($tenantId: ID!, $categories: [String!]) {\n    seedEvalTestCases(tenantId: $tenantId, categories: $categories)\n  }\n": CliSeedEvalTestCasesDocument,
+  "\n  query CliMe {\n    me {\n      id\n      email\n      name\n      tenantId\n    }\n  }\n": CliMeDocument,
+  "\n  query CliWikiTenantBySlug($slug: String!) {\n    tenantBySlug(slug: $slug) {\n      id\n      slug\n      name\n    }\n  }\n": CliWikiTenantBySlugDocument,
+  "\n  query CliAllTenantAgentsForWiki($tenantId: ID!) {\n    allTenantAgents(tenantId: $tenantId, includeSystem: false, includeSubAgents: false) {\n      id\n      name\n      slug\n      type\n      status\n    }\n  }\n": CliAllTenantAgentsForWikiDocument,
+  "\n  mutation CliCompileWikiNow($tenantId: ID!, $ownerId: ID!, $modelId: String) {\n    compileWikiNow(tenantId: $tenantId, ownerId: $ownerId, modelId: $modelId) {\n      id\n      tenantId\n      ownerId\n      status\n      trigger\n      dedupeKey\n      attempt\n      createdAt\n    }\n  }\n": CliCompileWikiNowDocument,
+  "\n  mutation CliResetWikiCursor($tenantId: ID!, $ownerId: ID!, $force: Boolean) {\n    resetWikiCursor(tenantId: $tenantId, ownerId: $ownerId, force: $force) {\n      tenantId\n      ownerId\n      cursorCleared\n      pagesArchived\n    }\n  }\n": CliResetWikiCursorDocument,
+  "\n  query CliWikiCompileJobs($tenantId: ID!, $ownerId: ID, $limit: Int) {\n    wikiCompileJobs(tenantId: $tenantId, ownerId: $ownerId, limit: $limit) {\n      id\n      tenantId\n      ownerId\n      status\n      trigger\n      dedupeKey\n      attempt\n      claimedAt\n      startedAt\n      finishedAt\n      error\n      metrics\n      createdAt\n    }\n  }\n": CliWikiCompileJobsDocument
+};
+function graphql(source) {
+  return documents[source] ?? {};
+}
+
+// src/commands/eval/gql.ts
+var EvalRunsDoc = graphql(`
+  query CliEvalRuns($tenantId: ID!, $agentId: ID, $limit: Int, $offset: Int) {
+    evalRuns(tenantId: $tenantId, agentId: $agentId, limit: $limit, offset: $offset) {
+      totalCount
+      items {
+        id
+        status
+        model
+        categories
+        agentId
+        agentName
+        agentTemplateId
+        agentTemplateName
+        totalTests
+        passed
+        failed
+        passRate
+        regression
+        costUsd
+        errorMessage
+        startedAt
+        completedAt
+        createdAt
+      }
+    }
+  }
+`);
+var EvalRunDoc = graphql(`
+  query CliEvalRun($id: ID!) {
+    evalRun(id: $id) {
+      id
+      status
+      model
+      categories
+      agentId
+      agentName
+      agentTemplateId
+      agentTemplateName
+      totalTests
+      passed
+      failed
+      passRate
+      regression
+      costUsd
+      errorMessage
+      startedAt
+      completedAt
+      createdAt
+    }
+  }
+`);
+var EvalRunResultsDoc = graphql(`
+  query CliEvalRunResults($runId: ID!) {
+    evalRunResults(runId: $runId) {
+      id
+      testCaseId
+      testCaseName
+      category
+      status
+      score
+      durationMs
+      agentSessionId
+      input
+      expected
+      actualOutput
+      evaluatorResults
+      assertions
+      errorMessage
+      createdAt
+    }
+  }
+`);
+var EvalTestCasesDoc = graphql(`
+  query CliEvalTestCases($tenantId: ID!, $category: String, $search: String) {
+    evalTestCases(tenantId: $tenantId, category: $category, search: $search) {
+      id
+      name
+      category
+      query
+      systemPrompt
+      agentTemplateId
+      agentTemplateName
+      agentcoreEvaluatorIds
+      tags
+      enabled
+      source
+      createdAt
+      updatedAt
+    }
+  }
+`);
+var EvalTestCaseDoc = graphql(`
+  query CliEvalTestCase($id: ID!) {
+    evalTestCase(id: $id) {
+      id
+      tenantId
+      name
+      category
+      query
+      systemPrompt
+      agentTemplateId
+      agentTemplateName
+      assertions
+      agentcoreEvaluatorIds
+      tags
+      enabled
+      source
+      createdAt
+      updatedAt
+    }
+  }
+`);
+var AgentTemplatesForEvalDoc = graphql(`
+  query CliAgentTemplatesForEval($tenantId: ID!) {
+    agentTemplates(tenantId: $tenantId) {
+      id
+      name
+      slug
+      model
+      isPublished
+    }
+  }
+`);
+var TenantBySlugDoc = graphql(`
+  query CliTenantBySlug($slug: String!) {
+    tenantBySlug(slug: $slug) {
+      id
+      slug
+      name
+    }
+  }
+`);
+var StartEvalRunDoc = graphql(`
+  mutation CliStartEvalRun($tenantId: ID!, $input: StartEvalRunInput!) {
+    startEvalRun(tenantId: $tenantId, input: $input) {
+      id
+      status
+      model
+      categories
+      agentTemplateId
+      agentTemplateName
+      totalTests
+      createdAt
+    }
+  }
+`);
+var CancelEvalRunDoc = graphql(`
+  mutation CliCancelEvalRun($id: ID!) {
+    cancelEvalRun(id: $id) {
+      id
+      status
+      completedAt
+    }
+  }
+`);
+var DeleteEvalRunDoc = graphql(`
+  mutation CliDeleteEvalRun($id: ID!) {
+    deleteEvalRun(id: $id)
+  }
+`);
+var CreateEvalTestCaseDoc = graphql(`
+  mutation CliCreateEvalTestCase($tenantId: ID!, $input: CreateEvalTestCaseInput!) {
+    createEvalTestCase(tenantId: $tenantId, input: $input) {
+      id
+      name
+      category
+    }
+  }
+`);
+var UpdateEvalTestCaseDoc = graphql(`
+  mutation CliUpdateEvalTestCase($id: ID!, $input: UpdateEvalTestCaseInput!) {
+    updateEvalTestCase(id: $id, input: $input) {
+      id
+      name
+      category
+      enabled
+    }
+  }
+`);
+var DeleteEvalTestCaseDoc = graphql(`
+  mutation CliDeleteEvalTestCase($id: ID!) {
+    deleteEvalTestCase(id: $id)
+  }
+`);
+var SeedEvalTestCasesDoc = graphql(`
+  mutation CliSeedEvalTestCases($tenantId: ID!, $categories: [String!]) {
+    seedEvalTestCases(tenantId: $tenantId, categories: $categories)
+  }
+`);
+
+// src/commands/eval/helpers.ts
+async function resolveEvalContext(opts) {
+  const region = opts.region ?? "us-east-1";
+  const stage = await resolveStage({ flag: opts.stage, region });
+  const session = loadStageSession(stage);
+  const { client, tenantSlug: ctxTenantSlug } = await getGqlClient({ stage, region });
+  const flagOrEnv = opts.tenant ?? process.env.THINKWORK_TENANT;
+  if (flagOrEnv) {
+    if (session?.tenantSlug === flagOrEnv && session.tenantId) {
+      return { stage, region, client, tenantId: session.tenantId, tenantSlug: flagOrEnv };
+    }
+    const data = await gqlQuery(client, TenantBySlugDoc, { slug: flagOrEnv });
+    if (!data.tenantBySlug) {
+      printError(`Tenant "${flagOrEnv}" not found.`);
+      process.exit(1);
+    }
+    return {
+      stage,
+      region,
+      client,
+      tenantId: data.tenantBySlug.id,
+      tenantSlug: data.tenantBySlug.slug
+    };
+  }
+  if (session?.tenantId && session.tenantSlug) {
+    return {
+      stage,
+      region,
+      client,
+      tenantId: session.tenantId,
+      tenantSlug: session.tenantSlug
+    };
+  }
+  if (ctxTenantSlug) {
+    const data = await gqlQuery(client, TenantBySlugDoc, { slug: ctxTenantSlug });
+    if (data.tenantBySlug) {
+      return {
+        stage,
+        region,
+        client,
+        tenantId: data.tenantBySlug.id,
+        tenantSlug: data.tenantBySlug.slug
+      };
+    }
+  }
+  printError(
+    `No tenant resolved for stage "${stage}". Pass --tenant <slug>, set THINKWORK_TENANT, or run \`thinkwork login --stage ${stage}\`.`
+  );
+  process.exit(1);
+}
+function fmtIso(iso) {
+  if (!iso) return "\u2014";
+  const d = new Date(iso);
+  if (Number.isNaN(d.getTime())) return iso;
+  return d.toISOString().replace(/\.\d{3}Z$/, "Z");
+}
+function fmtPercent(value) {
+  if (value == null) return "\u2014";
+  return `${(value * 100).toFixed(1)}%`;
+}
+function fmtUsd(value) {
+  if (value == null) return "\u2014";
+  return `$${value.toFixed(4)}`;
+}
+function isTerminalStatus(status) {
+  return status === "completed" || status === "failed" || status === "cancelled";
+}
+
+// src/commands/eval/run.ts
+async function runEvalRun(opts) {
+  const ctx = await resolveEvalContext(opts);
+  const interactive = isInteractive();
+  let agentTemplateId = opts.agentTemplate ?? null;
+  let categories = opts.category ?? null;
+  let testCaseIds = opts.testCase ?? null;
+  const scopeSatisfied = testCaseIds && testCaseIds.length > 0 || categories && categories.length > 0 || opts.all === true;
+  if (!agentTemplateId || !scopeSatisfied) {
+    if (!interactive) {
+      const missing = [];
+      if (!agentTemplateId) missing.push("--agent-template");
+      if (!scopeSatisfied) missing.push("one of --all | --category | --test-case");
+      printError(
+        `Missing required flag(s) in non-interactive session: ${missing.join(", ")}.`
+      );
+      process.exit(1);
+    }
+  }
+  if (!agentTemplateId) {
+    const data = await gqlQuery(ctx.client, AgentTemplatesForEvalDoc, {
+      tenantId: ctx.tenantId
+    });
+    const templates = data.agentTemplates ?? [];
+    if (templates.length === 0) {
+      printError("No agent templates defined for this tenant. Create one first.");
+      process.exit(1);
+    }
+    requireTty("Agent template");
+    agentTemplateId = await promptOrExit(
+      () => select8({
+        message: "Agent template to run against?",
+        choices: templates.map((t) => ({
+          name: `${t.name}${t.model ? `  (${t.model})` : ""}${t.isPublished ? "" : "  [draft]"}`,
+          value: t.id
+        })),
+        loop: false
+      })
+    );
+  }
+  if (!scopeSatisfied) {
+    const scope = await promptOrExit(
+      () => select8({
+        message: "How should we pick test cases?",
+        choices: [
+          { name: "All enabled test cases", value: "all" },
+          { name: "Filter by category", value: "category" },
+          { name: "Pick specific test cases", value: "specific" }
+        ],
+        loop: false
+      })
+    );
+    if (scope === "all") {
+      categories = null;
+      testCaseIds = null;
+      opts.all = true;
+    } else if (scope === "category") {
+      const tcData = await gqlQuery(ctx.client, EvalTestCasesDoc, {
+        tenantId: ctx.tenantId
+      });
+      const distinctCategories = Array.from(
+        new Set((tcData.evalTestCases ?? []).map((tc) => tc.category))
+      ).sort();
+      if (distinctCategories.length === 0) {
+        printError(
+          "No test cases exist for this tenant yet. Run `thinkwork eval seed` to load the starter pack."
+        );
+        process.exit(1);
+      }
+      const picked = await promptOrExit(
+        () => checkbox({
+          message: "Which categories? (space to toggle, enter to confirm)",
+          choices: distinctCategories.map((c) => ({ name: c, value: c })),
+          required: true,
+          loop: false
+        })
+      );
+      categories = picked;
+    } else {
+      const tcData = await gqlQuery(ctx.client, EvalTestCasesDoc, {
+        tenantId: ctx.tenantId
+      });
+      const options = (tcData.evalTestCases ?? []).filter((tc) => tc.enabled);
+      if (options.length === 0) {
+        printError("No enabled test cases to pick from.");
+        process.exit(1);
+      }
+      const picked = await promptOrExit(
+        () => checkbox({
+          message: "Which test cases?",
+          choices: options.map((tc) => ({
+            name: `${tc.name}  (${tc.category})`,
+            value: tc.id
+          })),
+          required: true,
+          loop: false
+        })
+      );
+      testCaseIds = picked;
+    }
+  }
+  if (!opts.model && interactive) {
+    const entered = await promptOrExit(
+      () => input3({
+        message: "Model override? (blank for template default)",
+        default: ""
+      })
+    );
+    if (entered.trim()) opts.model = entered.trim();
+  }
+  if (interactive && !isJsonMode()) {
+    const summaryLines = [
+      ["Stage", ctx.stage],
+      ["Tenant", ctx.tenantSlug],
+      ["Agent template", agentTemplateId]
+    ];
+    if (opts.model) summaryLines.push(["Model", opts.model]);
+    if (categories && categories.length) summaryLines.push(["Categories", categories.join(", ")]);
+    if (testCaseIds && testCaseIds.length)
+      summaryLines.push(["Test cases", `${testCaseIds.length} picked`]);
+    if (opts.all && !categories?.length && !testCaseIds?.length)
+      summaryLines.push(["Scope", "all enabled test cases"]);
+    printKeyValue(summaryLines);
+    const proceed = await promptOrExit(
+      () => confirm2({ message: "Start run?", default: true })
+    );
+    if (!proceed) {
+      logStderr("Cancelled.");
+      process.exit(0);
+    }
+  }
+  const mutRes = await gqlMutate(ctx.client, StartEvalRunDoc, {
+    tenantId: ctx.tenantId,
+    input: {
+      agentTemplateId,
+      agentId: opts.agent ?? null,
+      model: opts.model ?? null,
+      categories: categories ?? null,
+      testCaseIds: testCaseIds ?? null
+    }
+  });
+  const run2 = mutRes.startEvalRun;
+  if (isJsonMode()) {
+    printJson({ runId: run2.id, status: run2.status, model: run2.model, categories: run2.categories });
+  } else {
+    printSuccess(`Started eval run ${run2.id} (status: ${run2.status}).`);
+  }
+  if (!opts.watch) return;
+  const timeoutSec = Number.parseInt(opts.timeout ?? "900", 10);
+  await pollUntilTerminal(ctx.client, run2.id, 3, timeoutSec);
+}
+async function pollUntilTerminal(client, runId, intervalSec, timeoutSec) {
+  const deadline = Date.now() + timeoutSec * 1e3;
+  const spinner = isJsonMode() ? null : ora2({ text: "Waiting for run to complete\u2026" }).start();
+  try {
+    while (Date.now() < deadline) {
+      const data = await gqlQuery(client, EvalRunDoc, { id: runId });
+      const run2 = data.evalRun;
+      if (!run2) {
+        if (spinner) spinner.fail("Run disappeared from the database.");
+        process.exit(1);
+      }
+      if (spinner) {
+        spinner.text = `status=${run2.status}  ${run2.passed}/${run2.totalTests} passed  (${fmtPercent(run2.passRate)})`;
+      }
+      if (isTerminalStatus(run2.status)) {
+        if (spinner) {
+          if (run2.status === "completed") spinner.succeed(`completed \u2014 ${run2.passed}/${run2.totalTests} (${fmtPercent(run2.passRate)})`);
+          else if (run2.status === "failed") spinner.fail(`failed \u2014 ${run2.errorMessage ?? "unknown error"}`);
+          else spinner.warn("cancelled");
+        }
+        if (isJsonMode()) {
+          printJson({
+            runId: run2.id,
+            status: run2.status,
+            passed: run2.passed,
+            failed: run2.failed,
+            totalTests: run2.totalTests,
+            passRate: run2.passRate,
+            errorMessage: run2.errorMessage
+          });
+        }
+        if (run2.status === "completed") process.exit(0);
+        process.exit(1);
+      }
+      await new Promise((r) => setTimeout(r, intervalSec * 1e3));
+    }
+    if (spinner) spinner.warn(`timeout after ${timeoutSec}s`);
+    process.exit(2);
+  } catch (err) {
+    if (spinner) spinner.fail(err instanceof Error ? err.message : String(err));
+    throw err;
+  }
+}
+
+// src/commands/eval/list.ts
+async function runEvalList(opts) {
+  const ctx = await resolveEvalContext(opts);
+  const data = await gqlQuery(ctx.client, EvalRunsDoc, {
+    tenantId: ctx.tenantId,
+    agentId: opts.agent ?? null,
+    limit: Number.parseInt(opts.limit ?? "25", 10),
+    offset: Number.parseInt(opts.offset ?? "0", 10)
+  });
+  const rows = (data.evalRuns.items ?? []).map((r) => ({
+    id: r.id,
+    status: r.status,
+    template: r.agentTemplateName ?? r.agentTemplateId ?? "\u2014",
+    categories: (r.categories ?? []).join(", ") || "\u2014",
+    tests: `${r.passed}/${r.totalTests}`,
+    passRate: fmtPercent(r.passRate),
+    cost: fmtUsd(r.costUsd),
+    started: fmtIso(r.startedAt)
+  }));
+  if (isJsonMode()) {
+    printJson({ totalCount: data.evalRuns.totalCount, items: data.evalRuns.items });
+    return;
+  }
+  printTable(rows, [
+    { key: "id", header: "RUN ID" },
+    { key: "status", header: "STATUS" },
+    { key: "template", header: "TEMPLATE" },
+    { key: "categories", header: "CATEGORIES" },
+    { key: "tests", header: "PASS/TOTAL" },
+    { key: "passRate", header: "PASS RATE" },
+    { key: "cost", header: "COST" },
+    { key: "started", header: "STARTED" }
+  ]);
+}
+
+// src/commands/eval/get.ts
+async function runEvalGet(runId, opts) {
+  const ctx = await resolveEvalContext(opts);
+  const data = await gqlQuery(ctx.client, EvalRunDoc, { id: runId });
+  if (!data.evalRun) {
+    printError(`Run ${runId} not found.`);
+    process.exit(1);
+  }
+  const run2 = data.evalRun;
+  const results = opts.results === false ? [] : (await gqlQuery(ctx.client, EvalRunResultsDoc, { runId })).evalRunResults ?? [];
+  if (isJsonMode()) {
+    printJson({ run: run2, results });
+    return;
+  }
+  printKeyValue([
+    ["Run ID", run2.id],
+    ["Status", run2.status],
+    ["Agent template", run2.agentTemplateName ?? run2.agentTemplateId ?? "\u2014"],
+    ["Agent", run2.agentName ?? run2.agentId ?? "\u2014"],
+    ["Model", run2.model ?? "\u2014"],
+    ["Categories", (run2.categories ?? []).join(", ") || "\u2014"],
+    ["Pass/Total", `${run2.passed}/${run2.totalTests}`],
+    ["Pass rate", fmtPercent(run2.passRate)],
+    ["Regression", run2.regression ? "YES" : "no"],
+    ["Cost", fmtUsd(run2.costUsd)],
+    ["Error", run2.errorMessage ?? "\u2014"],
+    ["Started", fmtIso(run2.startedAt)],
+    ["Completed", fmtIso(run2.completedAt)]
+  ]);
+  if (opts.results !== false && results.length > 0) {
+    console.log("");
+    const rows = results.map((r) => ({
+      name: r.testCaseName ?? "\u2014",
+      category: r.category ?? "\u2014",
+      status: r.status,
+      score: r.score == null ? "\u2014" : r.score.toFixed(3),
+      duration: r.durationMs == null ? "\u2014" : `${r.durationMs}ms`
+    }));
+    printTable(rows, [
+      { key: "name", header: "TEST CASE" },
+      { key: "category", header: "CATEGORY" },
+      { key: "status", header: "STATUS" },
+      { key: "score", header: "SCORE" },
+      { key: "duration", header: "DURATION" }
+    ]);
+  }
+}
+
+// src/commands/eval/watch.ts
+import ora3 from "ora";
+async function runEvalWatch(runId, opts) {
+  const ctx = await resolveEvalContext(opts);
+  const intervalSec = Number.parseInt(opts.interval ?? "3", 10);
+  const timeoutSec = Number.parseInt(opts.timeout ?? "900", 10);
+  const deadline = Date.now() + timeoutSec * 1e3;
+  const spinner = isJsonMode() ? null : ora3({ text: `Watching run ${runId}\u2026` }).start();
+  try {
+    while (Date.now() < deadline) {
+      const data = await gqlQuery(ctx.client, EvalRunDoc, { id: runId });
+      const run2 = data.evalRun;
+      if (!run2) {
+        if (spinner) spinner.fail(`Run ${runId} not found.`);
+        process.exit(1);
+      }
+      if (spinner) {
+        spinner.text = `status=${run2.status}  ${run2.passed}/${run2.totalTests} passed  (${fmtPercent(run2.passRate)})`;
+      }
+      if (isTerminalStatus(run2.status)) {
+        if (spinner) {
+          if (run2.status === "completed")
+            spinner.succeed(`completed \u2014 ${run2.passed}/${run2.totalTests} (${fmtPercent(run2.passRate)})`);
+          else if (run2.status === "failed") spinner.fail(`failed \u2014 ${run2.errorMessage ?? "unknown error"}`);
+          else spinner.warn("cancelled");
+        }
+        if (isJsonMode()) {
+          printJson({
+            runId: run2.id,
+            status: run2.status,
+            passed: run2.passed,
+            failed: run2.failed,
+            totalTests: run2.totalTests,
+            passRate: run2.passRate,
+            errorMessage: run2.errorMessage
+          });
+        }
+        process.exit(run2.status === "completed" ? 0 : 1);
+      }
+      await new Promise((r) => setTimeout(r, intervalSec * 1e3));
+    }
+    if (spinner) spinner.warn(`timeout after ${timeoutSec}s`);
+    process.exit(2);
+  } catch (err) {
+    if (spinner) spinner.fail(err instanceof Error ? err.message : String(err));
+    throw err;
+  }
+}
+
+// src/commands/eval/cancel.ts
+async function runEvalCancel(runId, opts) {
+  const ctx = await resolveEvalContext(opts);
+  const data = await gqlMutate(ctx.client, CancelEvalRunDoc, { id: runId });
+  if (isJsonMode()) {
+    printJson({ runId: data.cancelEvalRun.id, status: data.cancelEvalRun.status });
+    return;
+  }
+  printSuccess(`Cancelled run ${data.cancelEvalRun.id} (status: ${data.cancelEvalRun.status}).`);
+}
+
+// src/commands/eval/delete.ts
+import { confirm as confirm3 } from "@inquirer/prompts";
+async function runEvalDelete(runId, opts) {
+  const ctx = await resolveEvalContext(opts);
+  if (!opts.yes) {
+    if (!isInteractive()) {
+      printError("Refusing to delete without --yes in a non-interactive session.");
+      process.exit(1);
+    }
+    requireTty("Confirmation");
+    const go = await promptOrExit(
+      () => confirm3({
+        message: `Permanently delete run ${runId} and its results?`,
+        default: false
+      })
+    );
+    if (!go) {
+      logStderr("Cancelled.");
+      process.exit(0);
+    }
+  }
+  const data = await gqlMutate(ctx.client, DeleteEvalRunDoc, { id: runId });
+  if (isJsonMode()) {
+    printJson({ runId, deleted: data.deleteEvalRun });
+    return;
+  }
+  if (data.deleteEvalRun) printSuccess(`Deleted run ${runId}.`);
+  else printError(`Server reported not-deleted for ${runId}.`);
+}
+
+// src/commands/eval/categories.ts
+async function runEvalCategories(opts) {
+  const ctx = await resolveEvalContext(opts);
+  const data = await gqlQuery(ctx.client, EvalTestCasesDoc, { tenantId: ctx.tenantId });
+  const counts = /* @__PURE__ */ new Map();
+  for (const tc of data.evalTestCases ?? []) {
+    const entry = counts.get(tc.category) ?? { total: 0, enabled: 0 };
+    entry.total += 1;
+    if (tc.enabled) entry.enabled += 1;
+    counts.set(tc.category, entry);
+  }
+  const rows = Array.from(counts.entries()).sort(([a], [b]) => a.localeCompare(b)).map(([category, { total, enabled }]) => ({
+    category,
+    enabled: String(enabled),
+    total: String(total)
+  }));
+  if (isJsonMode()) {
+    printJson(rows);
+    return;
+  }
+  printTable(rows, [
+    { key: "category", header: "CATEGORY" },
+    { key: "enabled", header: "ENABLED" },
+    { key: "total", header: "TOTAL" }
+  ]);
+}
+
+// src/commands/eval/seed.ts
+async function runEvalSeed(opts) {
+  const ctx = await resolveEvalContext(opts);
+  const data = await gqlMutate(ctx.client, SeedEvalTestCasesDoc, {
+    tenantId: ctx.tenantId,
+    categories: opts.category && opts.category.length > 0 ? opts.category : null
+  });
+  if (isJsonMode()) {
+    printJson({ inserted: data.seedEvalTestCases });
+    return;
+  }
+  printSuccess(`Seeded ${data.seedEvalTestCases} new test case(s). (Duplicates were skipped.)`);
+}
+
+// src/commands/eval/test-case/list.ts
+async function runEvalTestCaseList(opts) {
+  const ctx = await resolveEvalContext(opts);
+  const data = await gqlQuery(ctx.client, EvalTestCasesDoc, {
+    tenantId: ctx.tenantId,
+    category: opts.category ?? null,
+    search: opts.search ?? null
+  });
+  const rows = (data.evalTestCases ?? []).map((tc) => ({
+    id: tc.id,
+    name: tc.name,
+    category: tc.category,
+    template: tc.agentTemplateName ?? "\u2014",
+    evaluators: (tc.agentcoreEvaluatorIds ?? []).join(", ") || "\u2014",
+    enabled: tc.enabled ? "yes" : "no",
+    updated: fmtIso(tc.updatedAt)
+  }));
+  if (isJsonMode()) {
+    printJson(data.evalTestCases ?? []);
+    return;
+  }
+  printTable(rows, [
+    { key: "id", header: "ID" },
+    { key: "name", header: "NAME" },
+    { key: "category", header: "CATEGORY" },
+    { key: "template", header: "TEMPLATE" },
+    { key: "evaluators", header: "EVALUATORS" },
+    { key: "enabled", header: "ENABLED" },
+    { key: "updated", header: "UPDATED" }
+  ]);
+}
+
+// src/commands/eval/test-case/get.ts
+async function runEvalTestCaseGet(id, opts) {
+  const ctx = await resolveEvalContext(opts);
+  const data = await gqlQuery(ctx.client, EvalTestCaseDoc, { id });
+  if (!data.evalTestCase) {
+    printError(`Test case ${id} not found.`);
+    process.exit(1);
+  }
+  const tc = data.evalTestCase;
+  if (isJsonMode()) {
+    printJson(tc);
+    return;
+  }
+  printKeyValue([
+    ["ID", tc.id],
+    ["Name", tc.name],
+    ["Category", tc.category],
+    ["Agent template", tc.agentTemplateName ?? tc.agentTemplateId ?? "\u2014"],
+    ["Source", tc.source],
+    ["Enabled", tc.enabled ? "yes" : "no"],
+    ["Evaluators", (tc.agentcoreEvaluatorIds ?? []).join(", ") || "\u2014"],
+    ["Tags", (tc.tags ?? []).join(", ") || "\u2014"],
+    ["Created", fmtIso(tc.createdAt)],
+    ["Updated", fmtIso(tc.updatedAt)]
+  ]);
+  console.log("");
+  console.log("  QUERY");
+  console.log("  \u2500\u2500\u2500\u2500\u2500");
+  console.log(`  ${tc.query.split("\n").join("\n  ")}`);
+  if (tc.systemPrompt) {
+    console.log("");
+    console.log("  SYSTEM PROMPT");
+    console.log("  \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500");
+    console.log(`  ${tc.systemPrompt.split("\n").join("\n  ")}`);
+  }
+  if (tc.assertions) {
+    console.log("");
+    console.log("  ASSERTIONS");
+    console.log("  \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500");
+    console.log(`  ${tc.assertions}`);
+  }
+}
+
+// src/commands/eval/test-case/create.ts
+import { readFileSync as readFileSync6 } from "fs";
+import { input as input4, select as select9, checkbox as checkbox2 } from "@inquirer/prompts";
+var DEFAULT_EVALUATORS = [
+  "Builtin.Helpfulness",
+  "Builtin.Correctness",
+  "Builtin.Faithfulness",
+  "Builtin.ToolSelectionAccuracy",
+  "Builtin.ToolParameterAccuracy",
+  "Builtin.GoalSuccessRate"
+];
+async function runEvalTestCaseCreate(opts) {
+  const ctx = await resolveEvalContext(opts);
+  const interactive = isInteractive();
+  let name = opts.name;
+  let category = opts.category;
+  let query = opts.query;
+  let evaluators = opts.evaluator;
+  let agentTemplateId = opts.agentTemplate ?? null;
+  if (!name || !category || !query) {
+    if (!interactive) {
+      const missing = [];
+      if (!name) missing.push("--name");
+      if (!category) missing.push("--category");
+      if (!query) missing.push("--query");
+      printError(`Missing required flag(s): ${missing.join(", ")}.`);
+      process.exit(1);
+    }
+  }
+  if (!name) {
+    requireTty("Name");
+    name = await promptOrExit(
+      () => input4({ message: "Test case name?", validate: (v) => v.trim().length > 0 || "Required" })
+    );
+  }
+  if (!category) {
+    category = await promptOrExit(
+      () => input4({ message: "Category (free-form label)?", validate: (v) => v.trim().length > 0 || "Required" })
+    );
+  }
+  if (!query) {
+    query = await promptOrExit(
+      () => input4({ message: "Query the agent under test will receive?", validate: (v) => v.trim().length > 0 || "Required" })
+    );
+  }
+  if (interactive && agentTemplateId === null) {
+    const tpls = await gqlQuery(ctx.client, AgentTemplatesForEvalDoc, { tenantId: ctx.tenantId });
+    const templates = tpls.agentTemplates ?? [];
+    if (templates.length > 0) {
+      const choice = await promptOrExit(
+        () => select9({
+          message: "Pin to an agent template? (Enter for none)",
+          choices: [
+            { name: "\u2014 none \u2014 (runner picks)", value: "" },
+            ...templates.map((t) => ({ name: `${t.name}${t.model ? `  (${t.model})` : ""}`, value: t.id }))
+          ],
+          loop: false
+        })
+      );
+      agentTemplateId = choice === "" ? null : choice;
+    }
+  }
+  if (interactive && (!evaluators || evaluators.length === 0)) {
+    const picked = await promptOrExit(
+      () => checkbox2({
+        message: "Evaluators to run for this test case?",
+        choices: DEFAULT_EVALUATORS.map((e) => ({ name: e, value: e, checked: e === "Builtin.Helpfulness" })),
+        loop: false
+      })
+    );
+    evaluators = picked;
+  }
+  let assertions = null;
+  if (opts.assertionsFile) {
+    const parsed = JSON.parse(readFileSync6(opts.assertionsFile, "utf8"));
+    if (!Array.isArray(parsed)) {
+      printError(`--assertions-file must contain a JSON array.`);
+      process.exit(1);
+    }
+    assertions = parsed;
+  }
+  const mutation = await gqlMutate(ctx.client, CreateEvalTestCaseDoc, {
+    tenantId: ctx.tenantId,
+    input: {
+      name,
+      category,
+      query,
+      systemPrompt: opts.systemPrompt ?? null,
+      agentTemplateId,
+      agentcoreEvaluatorIds: evaluators && evaluators.length > 0 ? evaluators : null,
+      tags: opts.tag && opts.tag.length > 0 ? opts.tag : null,
+      enabled: opts.enabled ?? true,
+      assertions
+    }
+  });
+  if (isJsonMode()) {
+    printJson(mutation.createEvalTestCase);
+    return;
+  }
+  printSuccess(
+    `Created test case ${mutation.createEvalTestCase.id} "${mutation.createEvalTestCase.name}" (${mutation.createEvalTestCase.category}).`
+  );
+}
+
+// src/commands/eval/test-case/update.ts
+import { readFileSync as readFileSync7 } from "fs";
+async function runEvalTestCaseUpdate(id, opts) {
+  const ctx = await resolveEvalContext(opts);
+  const input5 = {};
+  if (opts.name !== void 0) input5.name = opts.name;
+  if (opts.category !== void 0) input5.category = opts.category;
+  if (opts.query !== void 0) input5.query = opts.query;
+  if (opts.systemPrompt !== void 0) input5.systemPrompt = opts.systemPrompt;
+  if (opts.agentTemplate !== void 0) input5.agentTemplateId = opts.agentTemplate;
+  if (opts.evaluator !== void 0) input5.agentcoreEvaluatorIds = opts.evaluator;
+  if (opts.tag !== void 0) input5.tags = opts.tag;
+  if (opts.enabled !== void 0) input5.enabled = opts.enabled;
+  if (opts.assertionsFile) {
+    const parsed = JSON.parse(readFileSync7(opts.assertionsFile, "utf8"));
+    if (!Array.isArray(parsed)) {
+      printError(`--assertions-file must contain a JSON array.`);
+      process.exit(1);
+    }
+    input5.assertions = parsed;
+  }
+  if (Object.keys(input5).length === 0) {
+    printError("No fields to update. Pass at least one --<field>.");
+    process.exit(1);
+  }
+  const res = await gqlMutate(ctx.client, UpdateEvalTestCaseDoc, { id, input: input5 });
+  if (isJsonMode()) {
+    printJson(res.updateEvalTestCase);
+    return;
+  }
+  printSuccess(`Updated test case ${res.updateEvalTestCase.id}.`);
+}
+
+// src/commands/eval/test-case/delete.ts
+import { confirm as confirm4 } from "@inquirer/prompts";
+async function runEvalTestCaseDelete(id, opts) {
+  const ctx = await resolveEvalContext(opts);
+  if (!opts.yes) {
+    if (!isInteractive()) {
+      printError("Refusing to delete without --yes in a non-interactive session.");
+      process.exit(1);
+    }
+    requireTty("Confirmation");
+    const go = await promptOrExit(
+      () => confirm4({ message: `Permanently delete test case ${id}?`, default: false })
+    );
+    if (!go) {
+      logStderr("Cancelled.");
+      process.exit(0);
+    }
+  }
+  const res = await gqlMutate(ctx.client, DeleteEvalTestCaseDoc, { id });
+  if (isJsonMode()) {
+    printJson({ id, deleted: res.deleteEvalTestCase });
+    return;
+  }
+  if (res.deleteEvalTestCase) printSuccess(`Deleted test case ${id}.`);
+  else printError(`Server reported not-deleted for ${id}.`);
+}
+
+// src/commands/eval.ts
+function registerEvalCommand(program2) {
+  const evals = program2.command("eval").alias("evals").description(
+    "Run evaluations against your agents and manage eval test cases. Integrates with the Evaluations Studio in the admin UI."
+  );
+  evals.command("run").description(
+    "Start an evaluation run. Prompts for missing values in a TTY; fails fast in non-interactive sessions."
+  ).option("-s, --stage <name>", "Deployment stage").option("-t, --tenant <slug>", "Tenant slug").option("-r, --region <region>", "AWS region", "us-east-1").option("--agent-template <id>", "Run-level agent template ID").option("--agent <id>", "Optional agent under test").option("--model <id>", "Optional model override").option("--category <name...>", "Only run these categories (repeatable)").option("--test-case <id...>", "Only run these specific test case IDs (repeatable)").option("--all", "Run all enabled test cases for the tenant").option("--watch", "Block and poll until the run reaches a terminal status").option("--timeout <seconds>", "Max wait seconds for --watch (default 900)", "900").addHelpText(
+    "after",
+    `
+Examples:
+  # Fire and return \u2014 prints the runId; view results in the admin UI
+  $ thinkwork eval run --agent-template tpl-abc --category tool-safety
+
+  # Pick categories + test cases interactively
+  $ thinkwork eval run
+
+  # Block until done
+  $ thinkwork eval run --agent-template tpl-abc --all --watch --timeout 1800
+`
+  ).action(runEvalRun);
+  evals.command("list").alias("ls").description("List recent eval runs for the tenant.").option("-s, --stage <name>", "Deployment stage").option("-t, --tenant <slug>", "Tenant slug").option("-r, --region <region>", "AWS region", "us-east-1").option("--agent <id>", "Filter by agent under test").option("--limit <n>", "Max rows (default 25)", "25").option("--offset <n>", "Skip N rows", "0").action(runEvalList);
+  evals.command("get <runId>").description("Show one eval run with its per-test-case results.").option("-s, --stage <name>", "Deployment stage").option("-t, --tenant <slug>", "Tenant slug").option("-r, --region <region>", "AWS region", "us-east-1").option("--results", "Also fetch per-test-case results (default: true)", true).option("--no-results", "Skip fetching per-test-case results").action(runEvalGet);
+  evals.command("watch <runId>").description("Poll an eval run until it reaches a terminal status.").option("-s, --stage <name>", "Deployment stage").option("-t, --tenant <slug>", "Tenant slug").option("-r, --region <region>", "AWS region", "us-east-1").option("--interval <seconds>", "Poll interval (default 3)", "3").option("--timeout <seconds>", "Max wait seconds (default 900)", "900").action(runEvalWatch);
+  evals.command("cancel <runId>").description("Cancel a running or pending eval run.").option("-s, --stage <name>", "Deployment stage").option("-t, --tenant <slug>", "Tenant slug").option("-r, --region <region>", "AWS region", "us-east-1").action(runEvalCancel);
+  evals.command("delete <runId>").description("Delete an eval run and its results. Requires confirmation unless --yes.").option("-s, --stage <name>", "Deployment stage").option("-t, --tenant <slug>", "Tenant slug").option("-r, --region <region>", "AWS region", "us-east-1").option("-y, --yes", "Skip the confirmation prompt").action(runEvalDelete);
+  evals.command("categories").description("List distinct categories present across the tenant's test cases.").option("-s, --stage <name>", "Deployment stage").option("-t, --tenant <slug>", "Tenant slug").option("-r, --region <region>", "AWS region", "us-east-1").action(runEvalCategories);
+  evals.command("seed").description(
+    "Idempotently seed the maniflow starter pack (96 test cases across 9 categories). Safe to re-run."
+  ).option("-s, --stage <name>", "Deployment stage").option("-t, --tenant <slug>", "Tenant slug").option("-r, --region <region>", "AWS region", "us-east-1").option("--category <name...>", "Only seed these categories (repeatable)").action(runEvalSeed);
+  const tc = evals.command("test-case").alias("test-cases").description("Manage individual eval test cases (CRUD).");
+  tc.command("list").alias("ls").description("List test cases, optionally filtered by category or search.").option("-s, --stage <name>", "Deployment stage").option("-t, --tenant <slug>", "Tenant slug").option("-r, --region <region>", "AWS region", "us-east-1").option("--category <name>", "Filter by a single category").option("--search <q>", "Substring match on test case name").action(runEvalTestCaseList);
+  tc.command("get <id>").description("Show a single test case.").option("-s, --stage <name>", "Deployment stage").option("-t, --tenant <slug>", "Tenant slug").option("-r, --region <region>", "AWS region", "us-east-1").action(runEvalTestCaseGet);
+  tc.command("create").description("Create a new test case. Prompts for missing values in a TTY.").option("-s, --stage <name>", "Deployment stage").option("-t, --tenant <slug>", "Tenant slug").option("-r, --region <region>", "AWS region", "us-east-1").option("--name <text>", "Human-readable name").option("--category <name>", "Category label (e.g. tool-safety, red-team)").option("--query <text>", "The user-facing query this agent will receive").option("--system-prompt <text>", "Optional system-prompt override").option("--agent-template <id>", "Pin to a specific agent template").option("--evaluator <id...>", "AgentCore evaluator IDs (repeatable)").option("--tag <name...>", "Tags (repeatable)").option("--enabled", "Mark enabled (default)", true).option("--no-enabled", "Mark disabled").option("--assertions-file <path>", "JSON file containing an array of assertions").action(runEvalTestCaseCreate);
+  tc.command("update <id>").description("Update a test case. Only supplied fields are changed.").option("-s, --stage <name>", "Deployment stage").option("-t, --tenant <slug>", "Tenant slug").option("-r, --region <region>", "AWS region", "us-east-1").option("--name <text>").option("--category <name>").option("--query <text>").option("--system-prompt <text>").option("--agent-template <id>").option("--evaluator <id...>", "Replace AgentCore evaluator IDs (repeatable)").option("--tag <name...>", "Replace tags (repeatable)").option("--enabled", "Mark enabled").option("--no-enabled", "Mark disabled").option("--assertions-file <path>", "JSON file containing an array of assertions (replaces all)").action(runEvalTestCaseUpdate);
+  tc.command("delete <id>").description("Delete a test case. Requires confirmation unless --yes.").option("-s, --stage <name>", "Deployment stage").option("-t, --tenant <slug>", "Tenant slug").option("-r, --region <region>", "AWS region", "us-east-1").option("-y, --yes", "Skip the confirmation prompt").action(runEvalTestCaseDelete);
+}
+
+// src/commands/wiki/compile.ts
+import ora4 from "ora";
+
+// src/commands/wiki/gql.ts
+var TenantBySlugDoc2 = graphql(`
+  query CliWikiTenantBySlug($slug: String!) {
+    tenantBySlug(slug: $slug) {
+      id
+      slug
+      name
+    }
+  }
+`);
+var AllTenantAgentsForWikiDoc = graphql(`
+  query CliAllTenantAgentsForWiki($tenantId: ID!) {
+    allTenantAgents(tenantId: $tenantId, includeSystem: false, includeSubAgents: false) {
+      id
+      name
+      slug
+      type
+      status
+    }
+  }
+`);
+var CompileWikiNowDoc = graphql(`
+  mutation CliCompileWikiNow($tenantId: ID!, $ownerId: ID!, $modelId: String) {
+    compileWikiNow(tenantId: $tenantId, ownerId: $ownerId, modelId: $modelId) {
+      id
+      tenantId
+      ownerId
+      status
+      trigger
+      dedupeKey
+      attempt
+      createdAt
+    }
+  }
+`);
+var ResetWikiCursorDoc = graphql(`
+  mutation CliResetWikiCursor($tenantId: ID!, $ownerId: ID!, $force: Boolean) {
+    resetWikiCursor(tenantId: $tenantId, ownerId: $ownerId, force: $force) {
+      tenantId
+      ownerId
+      cursorCleared
+      pagesArchived
+    }
+  }
+`);
+var WikiCompileJobsDoc = graphql(`
+  query CliWikiCompileJobs($tenantId: ID!, $ownerId: ID, $limit: Int) {
+    wikiCompileJobs(tenantId: $tenantId, ownerId: $ownerId, limit: $limit) {
+      id
+      tenantId
+      ownerId
+      status
+      trigger
+      dedupeKey
+      attempt
+      claimedAt
+      startedAt
+      finishedAt
+      error
+      metrics
+      createdAt
+    }
+  }
+`);
+
+// src/commands/wiki/helpers.ts
+import { select as select10 } from "@inquirer/prompts";
+async function resolveWikiContext(opts) {
+  const region = opts.region ?? "us-east-1";
+  const stage = await resolveStage({ flag: opts.stage, region });
+  const session = loadStageSession(stage);
+  const { client, tenantSlug: ctxTenantSlug } = await getGqlClient({
+    stage,
+    region
+  });
+  const flagOrEnv = opts.tenant ?? process.env.THINKWORK_TENANT;
+  if (flagOrEnv) {
+    if (session?.tenantSlug === flagOrEnv && session.tenantId) {
+      return {
+        stage,
+        region,
+        client,
+        tenantId: session.tenantId,
+        tenantSlug: flagOrEnv
+      };
+    }
+    const data = await gqlQuery(client, TenantBySlugDoc2, { slug: flagOrEnv });
+    if (!data.tenantBySlug) {
+      printError(`Tenant "${flagOrEnv}" not found.`);
+      process.exit(1);
+    }
+    return {
+      stage,
+      region,
+      client,
+      tenantId: data.tenantBySlug.id,
+      tenantSlug: data.tenantBySlug.slug
+    };
+  }
+  if (session?.tenantId && session.tenantSlug) {
+    return {
+      stage,
+      region,
+      client,
+      tenantId: session.tenantId,
+      tenantSlug: session.tenantSlug
+    };
+  }
+  if (ctxTenantSlug) {
+    const data = await gqlQuery(client, TenantBySlugDoc2, {
+      slug: ctxTenantSlug
+    });
+    if (data.tenantBySlug) {
+      return {
+        stage,
+        region,
+        client,
+        tenantId: data.tenantBySlug.id,
+        tenantSlug: data.tenantBySlug.slug
+      };
+    }
+  }
+  printError(
+    `No tenant resolved for stage "${stage}". Pass --tenant <slug>, set THINKWORK_TENANT, or run \`thinkwork login --stage ${stage}\`.`
+  );
+  process.exit(1);
+}
+async function resolveAgentScope(ctx, opts, config = {}) {
+  const allowAll = config.allowAll ?? true;
+  const needList = opts.agent != null && !isUuid(opts.agent) || opts.all === true || opts.agent == null && !opts.all;
+  let agents = [];
+  const loadAgents = async () => {
+    const data = await gqlQuery(ctx.client, AllTenantAgentsForWikiDoc, {
+      tenantId: ctx.tenantId
+    });
+    return data.allTenantAgents ?? [];
+  };
+  if (opts.agent) {
+    if (isUuid(opts.agent)) {
+      return {
+        mode: "single",
+        agentId: opts.agent,
+        agentLabel: opts.agent
+      };
+    }
+    agents = await loadAgents();
+    const needle = opts.agent.toLowerCase();
+    const matches = agents.filter(
+      (a) => [a.name, a.slug].some(
+        (v) => v != null && String(v).toLowerCase() === needle
+      )
+    );
+    if (matches.length === 0) {
+      printError(
+        `Agent "${opts.agent}" not found. Pass the UUID or a matching name/slug.`
+      );
+      process.exit(1);
+    }
+    if (matches.length > 1) {
+      printError(
+        `"${opts.agent}" matches ${matches.length} agents. Pass the UUID instead \u2014 candidates: ${matches.map((a) => a.id).join(", ")}`
+      );
+      process.exit(1);
+    }
+    return {
+      mode: "single",
+      agentId: matches[0].id,
+      agentLabel: matches[0].name ?? matches[0].id
+    };
+  }
+  if (opts.all) {
+    if (!allowAll) {
+      printError(
+        "--all is not supported for this command. Rebuild one agent at a time."
+      );
+      process.exit(1);
+    }
+    if (!needList) agents = await loadAgents();
+    else if (agents.length === 0) agents = await loadAgents();
+    return {
+      mode: "all",
+      agentIds: agents.map((a) => a.id),
+      agentLabels: Object.fromEntries(
+        agents.map((a) => [a.id, a.name ?? a.id])
+      )
+    };
+  }
+  if (!isInteractive()) {
+    requireTty(allowAll ? "Agent (or --all)" : "Agent");
+    throw new Error("unreachable");
+  }
+  agents = await loadAgents();
+  if (agents.length === 0) {
+    printError("No agents found for this tenant.");
+    process.exit(1);
+  }
+  const choices = [];
+  if (allowAll) {
+    choices.push({ name: "All agents (fan out)", value: "__all__" });
+  }
+  for (const a of agents) {
+    const label = a.name ?? a.id;
+    const slugPart = a.slug ? `  (${a.slug})` : "";
+    choices.push({ name: `${label}${slugPart}  [${a.id}]`, value: a.id });
+  }
+  const pick = await promptOrExit(
+    () => select10({
+      message: "Which agent?",
+      choices,
+      loop: false
+    })
+  );
+  if (pick === "__all__") {
+    return {
+      mode: "all",
+      agentIds: agents.map((a) => a.id),
+      agentLabels: Object.fromEntries(
+        agents.map((a) => [a.id, a.name ?? a.id])
+      )
+    };
+  }
+  const picked = agents.find((a) => a.id === pick);
+  return {
+    mode: "single",
+    agentId: picked.id,
+    agentLabel: picked.name ?? picked.id
+  };
+}
+function classifyMutationError(err) {
+  const message = err?.message ?? String(err);
+  const forbidden = /Admin-only|Access denied|tenant mismatch|outside tenant/i.test(message);
+  return { forbidden, message };
+}
+function printForbiddenHint(tenantSlug) {
+  printError(
+    `Admin access to tenant "${tenantSlug}" is required for wiki operations. Ask your tenant owner to promote your membership or use an admin API key.`
+  );
+}
+function isTerminalCompileStatus(status) {
+  return status === "succeeded" || status === "failed" || status === "cancelled" || status === "skipped";
+}
+function shortJobId(id) {
+  return id.slice(0, 8);
+}
+
+// src/commands/wiki/compile.ts
+async function runWikiCompile(opts) {
+  const ctx = await resolveWikiContext(opts);
+  const scope = await resolveAgentScope(ctx, opts, { allowAll: true });
+  const targets = scope.mode === "single" ? [{ id: scope.agentId, label: scope.agentLabel }] : scope.agentIds.map((id) => ({
+    id,
+    label: scope.agentLabels[id] ?? id
+  }));
+  if (targets.length === 0) {
+    if (isJsonMode()) {
+      printJson({
+        ok: true,
+        scope: { tenantId: ctx.tenantId, tenantSlug: ctx.tenantSlug, agentIds: [] },
+        jobs: [],
+        errors: []
+      });
+    } else {
+      printWarning("No agents found for this tenant \u2014 nothing to compile.");
+    }
+    return;
+  }
+  const jobs = [];
+  const errors = [];
+  let forbiddenHit = false;
+  for (const target of targets) {
+    const spinner = isJsonMode() || scope.mode === "all" ? null : ora4({
+      text: `Enqueuing compile for ${target.label}\u2026`,
+      prefixText: "  "
+    }).start();
+    try {
+      const data = await gqlMutate(ctx.client, CompileWikiNowDoc, {
+        tenantId: ctx.tenantId,
+        ownerId: target.id,
+        modelId: opts.model ?? null
+      });
+      const job = data.compileWikiNow;
+      jobs.push({
+        agentId: target.id,
+        agentLabel: target.label,
+        jobId: job.id,
+        status: job.status,
+        error: null
+      });
+      if (spinner) {
+        spinner.succeed(
+          `${target.label}  \u2192  job=${shortJobId(job.id)} (${job.status})`
+        );
+      } else if (!isJsonMode()) {
+        console.log(
+          `  \u2713 ${target.label}  \u2192  job=${shortJobId(job.id)} (${job.status})`
+        );
+      }
+    } catch (err) {
+      const classified = classifyMutationError(err);
+      errors.push({ agentId: target.id, message: classified.message });
+      jobs.push({
+        agentId: target.id,
+        agentLabel: target.label,
+        jobId: null,
+        status: null,
+        error: classified.message
+      });
+      if (spinner) spinner.fail(`${target.label}  \u2192  ${classified.message}`);
+      else if (!isJsonMode())
+        console.log(`  \u2717 ${target.label}  \u2192  ${classified.message}`);
+      if (classified.forbidden) {
+        forbiddenHit = true;
+        if (scope.mode === "all") {
+          break;
+        }
+      }
+    }
+  }
+  const anyFailed = errors.length > 0;
+  const ok = !anyFailed;
+  if (isJsonMode()) {
+    printJson({
+      ok,
+      scope: {
+        tenantId: ctx.tenantId,
+        tenantSlug: ctx.tenantSlug,
+        mode: scope.mode,
+        agentIds: targets.map((t) => t.id)
+      },
+      model: opts.model ?? null,
+      jobs,
+      errors
+    });
+  } else if (scope.mode === "all") {
+    console.log("");
+    printKeyValue([
+      ["Tenant", ctx.tenantSlug],
+      ["Agents queued", `${jobs.filter((j) => j.jobId).length} / ${targets.length}`],
+      ["Failures", String(errors.length)],
+      ["Model override", opts.model ?? "(default)"]
+    ]);
+  }
+  if (forbiddenHit) {
+    printForbiddenHint(ctx.tenantSlug);
+    process.exit(2);
+  }
+  if (anyFailed) {
+    process.exit(1);
+  }
+  if (!isJsonMode() && jobs.length === 1) {
+    printSuccess(`Compile enqueued for ${jobs[0].agentLabel}.`);
+    console.log(
+      `  Use \`thinkwork wiki status --tenant ${ctx.tenantSlug} --agent ${jobs[0].agentId} --watch\` to follow the job.`
+    );
+  }
+  if (opts.watch && scope.mode === "single" && jobs.length === 1 && jobs[0].jobId) {
+    await watchSingleJob(ctx, {
+      agentId: jobs[0].agentId,
+      jobId: jobs[0].jobId,
+      agentLabel: jobs[0].agentLabel
+    });
+  } else if (opts.watch && scope.mode === "all") {
+    printWarning(
+      "--watch is ignored for --all. Use `thinkwork wiki status --tenant " + ctx.tenantSlug + " --watch` instead."
+    );
+  }
+}
+async function watchSingleJob(ctx, target) {
+  const spinner = isJsonMode() ? null : ora4({
+    text: `Watching job ${shortJobId(target.jobId)} for ${target.agentLabel}\u2026`,
+    prefixText: "  "
+  }).start();
+  const intervalMs = 3e3;
+  const deadline = Date.now() + 15 * 60 * 1e3;
+  try {
+    while (Date.now() < deadline) {
+      const data = await gqlQuery(ctx.client, WikiCompileJobsDoc, {
+        tenantId: ctx.tenantId,
+        ownerId: target.agentId,
+        limit: 5
+      });
+      const job = data.wikiCompileJobs.find((j) => j.id === target.jobId);
+      if (!job) {
+        if (spinner) spinner.warn("job not visible yet \u2014 polling\u2026");
+      } else {
+        if (spinner) spinner.text = `status=${job.status}  attempt=${job.attempt}`;
+        if (isTerminalCompileStatus(job.status)) {
+          if (spinner) {
+            if (job.status === "succeeded") spinner.succeed(`succeeded`);
+            else if (job.status === "skipped") spinner.info("skipped");
+            else spinner.fail(`${job.status}${job.error ? ` \u2014 ${job.error}` : ""}`);
+          }
+          if (isJsonMode()) {
+            printJson({
+              ok: job.status === "succeeded",
+              jobId: job.id,
+              status: job.status,
+              error: job.error,
+              metrics: job.metrics
+            });
+          }
+          process.exit(job.status === "succeeded" ? 0 : 1);
+        }
+      }
+      await new Promise((r) => setTimeout(r, intervalMs));
+    }
+    if (spinner) spinner.warn("watch timeout \u2014 job still in progress.");
+    process.exit(2);
+  } catch (err) {
+    if (spinner) spinner.fail(err instanceof Error ? err.message : String(err));
+    printError("Watch failed. The compile job itself may still complete.");
+    process.exit(1);
+  }
+}
+
+// src/commands/wiki/rebuild.ts
+import { confirm as confirm5 } from "@inquirer/prompts";
+import ora5 from "ora";
+async function runWikiRebuild(opts) {
+  if (opts.all) {
+    printError(
+      "--all is not supported for rebuild. Rebuild one agent at a time to avoid mass-archiving pages across the tenant."
+    );
+    process.exit(1);
+  }
+  const ctx = await resolveWikiContext(opts);
+  const scope = await resolveAgentScope(ctx, opts, { allowAll: false });
+  if (scope.mode !== "single") {
+    printError("rebuild requires a single agent.");
+    process.exit(1);
+  }
+  const { agentId, agentLabel } = scope;
+  const skipConfirm = opts.yes === true || isJsonMode();
+  if (!skipConfirm) {
+    if (!isInteractive()) {
+      requireTty("Rebuild confirmation (--yes)");
+    }
+    const ok = await promptOrExit(
+      () => confirm5({
+        message: `Rebuild wiki for ${agentLabel}? This archives every active page in the scope and recompiles from scratch.`,
+        default: false
+      })
+    );
+    if (!ok) {
+      if (!isJsonMode()) console.log("  Cancelled.");
+      process.exit(0);
+    }
+  }
+  const resetSpinner = isJsonMode() ? null : ora5({
+    text: `Archiving active pages for ${agentLabel}\u2026`,
+    prefixText: "  "
+  }).start();
+  let pagesArchived = 0;
+  try {
+    const data = await gqlMutate(ctx.client, ResetWikiCursorDoc, {
+      tenantId: ctx.tenantId,
+      ownerId: agentId,
+      force: true
+    });
+    pagesArchived = data.resetWikiCursor.pagesArchived;
+    if (resetSpinner)
+      resetSpinner.succeed(
+        `${pagesArchived} page${pagesArchived === 1 ? "" : "s"} archived, cursor cleared.`
+      );
+  } catch (err) {
+    const classified = classifyMutationError(err);
+    if (resetSpinner) resetSpinner.fail(`Reset failed: ${classified.message}`);
+    const result2 = {
+      ok: false,
+      scope: {
+        tenantId: ctx.tenantId,
+        tenantSlug: ctx.tenantSlug,
+        agentId
+      },
+      pagesArchived: null,
+      jobId: null,
+      error: classified.message
+    };
+    if (isJsonMode()) printJson(result2);
+    if (classified.forbidden) {
+      printForbiddenHint(ctx.tenantSlug);
+      process.exit(2);
+    }
+    process.exit(1);
+  }
+  const compileSpinner = isJsonMode() ? null : ora5({
+    text: `Enqueuing fresh compile for ${agentLabel}\u2026`,
+    prefixText: "  "
+  }).start();
+  let jobId = null;
+  try {
+    const data = await gqlMutate(ctx.client, CompileWikiNowDoc, {
+      tenantId: ctx.tenantId,
+      ownerId: agentId,
+      modelId: opts.model ?? null
+    });
+    jobId = data.compileWikiNow.id;
+    if (compileSpinner)
+      compileSpinner.succeed(
+        `Compile enqueued \u2014 job=${shortJobId(jobId)}  status=${data.compileWikiNow.status}`
+      );
+  } catch (err) {
+    const classified = classifyMutationError(err);
+    if (compileSpinner)
+      compileSpinner.fail(`Compile enqueue failed: ${classified.message}`);
+    const result2 = {
+      ok: false,
+      scope: {
+        tenantId: ctx.tenantId,
+        tenantSlug: ctx.tenantSlug,
+        agentId
+      },
+      pagesArchived,
+      jobId: null,
+      error: classified.message
+    };
+    if (isJsonMode()) printJson(result2);
+    else {
+      printWarning(
+        `Reset succeeded (${pagesArchived} page${pagesArchived === 1 ? "" : "s"} archived) but compile enqueue failed. Retry with:
+  thinkwork wiki compile --tenant ${ctx.tenantSlug} --agent ${agentId}`
+      );
+    }
+    if (classified.forbidden) {
+      printForbiddenHint(ctx.tenantSlug);
+      process.exit(2);
+    }
+    process.exit(1);
+  }
+  const result = {
+    ok: true,
+    scope: {
+      tenantId: ctx.tenantId,
+      tenantSlug: ctx.tenantSlug,
+      agentId
+    },
+    pagesArchived,
+    jobId,
+    error: null
+  };
+  if (isJsonMode()) {
+    printJson(result);
+  } else {
+    console.log("");
+    printKeyValue([
+      ["Tenant", ctx.tenantSlug],
+      ["Agent", agentLabel],
+      ["Pages archived", String(pagesArchived)],
+      ["Compile job", jobId ?? "\u2014"],
+      ["Model override", opts.model ?? "(default)"]
+    ]);
+    printSuccess(`Rebuild enqueued for ${agentLabel}.`);
+  }
+  if (opts.watch && jobId) {
+    await watchRebuildJob(ctx, { agentId, jobId, agentLabel });
+  }
+}
+async function watchRebuildJob(ctx, target) {
+  const spinner = isJsonMode() ? null : ora5({
+    text: `Watching rebuild job ${shortJobId(target.jobId)}\u2026`,
+    prefixText: "  "
+  }).start();
+  const intervalMs = 3e3;
+  const deadline = Date.now() + 15 * 60 * 1e3;
+  try {
+    while (Date.now() < deadline) {
+      const data = await gqlQuery(ctx.client, WikiCompileJobsDoc, {
+        tenantId: ctx.tenantId,
+        ownerId: target.agentId,
+        limit: 5
+      });
+      const job = data.wikiCompileJobs.find((j) => j.id === target.jobId);
+      if (job) {
+        if (spinner) spinner.text = `status=${job.status}  attempt=${job.attempt}`;
+        if (isTerminalCompileStatus(job.status)) {
+          if (spinner) {
+            if (job.status === "succeeded") spinner.succeed("rebuild succeeded");
+            else if (job.status === "skipped") spinner.info("rebuild skipped");
+            else spinner.fail(`${job.status}${job.error ? ` \u2014 ${job.error}` : ""}`);
+          }
+          process.exit(job.status === "succeeded" ? 0 : 1);
+        }
+      }
+      await new Promise((r) => setTimeout(r, intervalMs));
+    }
+    if (spinner) spinner.warn("watch timeout \u2014 rebuild still in progress.");
+    process.exit(2);
+  } catch (err) {
+    if (spinner) spinner.fail(err instanceof Error ? err.message : String(err));
+    process.exit(1);
+  }
+}
+
+// src/commands/wiki/status.ts
+import ora6 from "ora";
+var DEFAULT_WATCH_INTERVAL_MS = 3e3;
+async function runWikiStatus(opts) {
+  const ctx = await resolveWikiContext(opts);
+  const ownerId = opts.agent ?? null;
+  const limit = toInt(opts.limit, 10);
+  const timeoutSec = toInt(opts.timeout, 900);
+  let agentNameById = null;
+  const resolveAgentName = async () => {
+    if (agentNameById) return agentNameById;
+    const data = await gqlQuery(ctx.client, AllTenantAgentsForWikiDoc, {
+      tenantId: ctx.tenantId
+    });
+    agentNameById = Object.fromEntries(
+      (data.allTenantAgents ?? []).map((a) => [a.id, a.name ?? a.id])
+    );
+    return agentNameById;
+  };
+  let jobs;
+  try {
+    jobs = await fetchJobs(ctx, { ownerId, limit });
+  } catch (err) {
+    const classified = classifyMutationError(err);
+    printError(classified.message);
+    if (classified.forbidden) {
+      printForbiddenHint(ctx.tenantSlug);
+      process.exit(2);
+    }
+    process.exit(1);
+  }
+  if (!opts.watch) {
+    await renderJobs(jobs, {
+      tenantSlug: ctx.tenantSlug,
+      scope: ownerId ? { agentId: ownerId } : { tenantWide: true },
+      resolveAgentName
+    });
+    process.exit(0);
+  }
+  if (jobs.length === 0) {
+    printWarning(
+      `No compile jobs yet for this scope. --watch exits immediately; re-run once activity starts.`
+    );
+    if (isJsonMode()) {
+      printJson({
+        ok: true,
+        scope: { tenantId: ctx.tenantId, ownerId },
+        jobs: []
+      });
+    }
+    process.exit(0);
+  }
+  const latestId = jobs[0].id;
+  const spinner = isJsonMode() ? null : ora6({
+    text: `Watching job ${shortJobId(latestId)}\u2026`,
+    prefixText: "  "
+  }).start();
+  const deadline = Date.now() + timeoutSec * 1e3;
+  try {
+    while (Date.now() < deadline) {
+      const next = await fetchJobs(ctx, { ownerId, limit: 5 });
+      const latest = next.find((j) => j.id === latestId) ?? next[0];
+      if (latest) {
+        if (spinner)
+          spinner.text = `status=${latest.status}  attempt=${latest.attempt}`;
+        if (isTerminalCompileStatus(latest.status)) {
+          if (spinner) {
+            if (latest.status === "succeeded") spinner.succeed("succeeded");
+            else if (latest.status === "skipped") spinner.info("skipped");
+            else
+              spinner.fail(
+                `${latest.status}${latest.error ? ` \u2014 ${latest.error}` : ""}`
+              );
+          }
+          if (isJsonMode()) {
+            printJson({
+              ok: latest.status === "succeeded",
+              scope: { tenantId: ctx.tenantId, ownerId },
+              job: latest
+            });
+          }
+          process.exit(latest.status === "succeeded" ? 0 : 2);
+        }
+      }
+      await new Promise((r) => setTimeout(r, DEFAULT_WATCH_INTERVAL_MS));
+    }
+    if (spinner) spinner.warn(`watch timeout after ${timeoutSec}s.`);
+    process.exit(2);
+  } catch (err) {
+    if (spinner) spinner.fail(err instanceof Error ? err.message : String(err));
+    process.exit(1);
+  }
+}
+async function fetchJobs(ctx, args) {
+  const data = await gqlQuery(ctx.client, WikiCompileJobsDoc, {
+    tenantId: ctx.tenantId,
+    ownerId: args.ownerId,
+    limit: args.limit
+  });
+  return data.wikiCompileJobs;
+}
+async function renderJobs(jobs, args) {
+  if (isJsonMode()) {
+    printJson({
+      ok: true,
+      scope: "agentId" in args.scope ? { agentId: args.scope.agentId } : { tenantWide: true },
+      jobs
+    });
+    return;
+  }
+  if (jobs.length === 0) {
+    const label = "agentId" in args.scope ? `agent ${args.scope.agentId}` : `tenant ${args.tenantSlug}`;
+    console.log(`  No recent compile jobs for ${label}.`);
+    return;
+  }
+  const names = "tenantWide" in args.scope ? await args.resolveAgentName() : {};
+  const rows = jobs.map((j) => ({
+    id: j.id.slice(0, 8),
+    agent: "tenantWide" in args.scope ? names[j.ownerId] ?? j.ownerId.slice(0, 8) : "\u2014",
+    status: j.status,
+    trigger: j.trigger,
+    attempt: String(j.attempt),
+    duration: fmtDuration(j.startedAt, j.finishedAt),
+    records: extractMetric(j.metrics, "records_read"),
+    pages: extractMetric(j.metrics, "pages_upserted"),
+    started: fmtIso2(j.startedAt ?? j.createdAt)
+  }));
+  const columns = [
+    { key: "id", header: "JOB" }
+  ];
+  if ("tenantWide" in args.scope) columns.push({ key: "agent", header: "AGENT" });
+  columns.push(
+    { key: "status", header: "STATUS" },
+    { key: "trigger", header: "TRIGGER" },
+    { key: "attempt", header: "TRY" },
+    { key: "duration", header: "DUR" },
+    { key: "records", header: "RECS" },
+    { key: "pages", header: "PAGES" },
+    { key: "started", header: "STARTED" }
+  );
+  printTable(rows, columns);
+}
+function toInt(v, fallback) {
+  if (v == null || v === "") return fallback;
+  const n = typeof v === "number" ? v : Number.parseInt(v, 10);
+  return Number.isFinite(n) && n > 0 ? n : fallback;
+}
+function fmtIso2(iso) {
+  if (!iso) return "\u2014";
+  const d = new Date(iso);
+  if (Number.isNaN(d.getTime())) return iso;
+  return d.toISOString().replace(/\.\d{3}Z$/, "Z");
+}
+function fmtDuration(startedAt, finishedAt) {
+  if (!startedAt) return "\u2014";
+  const start = new Date(startedAt).getTime();
+  const end = finishedAt ? new Date(finishedAt).getTime() : Date.now();
+  if (!Number.isFinite(start) || !Number.isFinite(end) || end < start) return "\u2014";
+  const sec = Math.round((end - start) / 1e3);
+  if (sec < 60) return `${sec}s`;
+  const m = Math.floor(sec / 60);
+  const s = sec % 60;
+  return `${m}m${s.toString().padStart(2, "0")}s`;
+}
+function extractMetric(metrics, key) {
+  if (!metrics || typeof metrics !== "object") return "\u2014";
+  const v = metrics[key];
+  if (v == null) return "\u2014";
+  if (typeof v === "number") return String(v);
+  return String(v);
+}
+
+// src/commands/wiki.ts
+function registerWikiCommand(program2) {
+  const wiki = program2.command("wiki").description(
+    "Compile and rebuild agent wiki pages (Compounding Memory). Admin-only."
+  );
+  wiki.command("compile").description(
+    "Enqueue a wiki compile for a single agent (--agent) or all tenant agents (--all)."
+  ).option("-s, --stage <name>", "Deployment stage").option("-t, --tenant <slug>", "Tenant slug").option("--agent <id>", "Agent ID, slug, or name. Bypasses the picker.").option("--all", "Fan out to every non-system agent in the tenant.").option(
+    "--model <id>",
+    "Bedrock model ID override for this run. Defaults to server BEDROCK_MODEL_ID."
+  ).option(
+    "--watch",
+    "After enqueue, poll wiki_compile_jobs until the job reaches a terminal state (single-agent only)."
+  ).addHelpText(
+    "after",
+    `
+Examples:
+  # Interactive: pick the agent (or "All agents") from a list
+  $ thinkwork wiki compile
+
+  # Single agent, scripted
+  $ thinkwork wiki compile --tenant acme --agent agt-xyz --json
+
+  # Fan-out across every agent in the tenant
+  $ thinkwork wiki compile --tenant acme --all
+
+  # Spike a different Bedrock model for one run
+  $ thinkwork wiki compile --tenant acme --agent agt-xyz \\
+      --model anthropic.claude-sonnet-4-6-v1:0
+`
+  ).action(async (opts, cmd) => {
+    const parent = cmd.parent?.parent;
+    await runWikiCompile({
+      ...opts,
+      stage: opts.stage ?? parent?.opts().stage,
+      json: parent?.opts().json === true
+    });
+  });
+  wiki.command("rebuild").description(
+    "Destructive: archive an agent's active wiki pages and enqueue a fresh compile. Single-agent only."
+  ).option("-s, --stage <name>", "Deployment stage").option("-t, --tenant <slug>", "Tenant slug").option("--agent <id>", "Agent ID, slug, or name. Bypasses the picker.").option(
+    "--model <id>",
+    "Bedrock model ID override for the post-reset compile."
+  ).option("-y, --yes", "Skip the confirmation prompt.").option(
+    "--watch",
+    "After enqueue, poll wiki_compile_jobs until the job reaches a terminal state."
+  ).addHelpText(
+    "after",
+    `
+Examples:
+  # Interactive confirm
+  $ thinkwork wiki rebuild --tenant acme --agent agt-xyz
+
+  # Scripted (no prompt)
+  $ thinkwork wiki rebuild --tenant acme --agent agt-xyz --yes --json
+`
+  ).action(async (opts, cmd) => {
+    const parent = cmd.parent?.parent;
+    await runWikiRebuild({
+      ...opts,
+      stage: opts.stage ?? parent?.opts().stage,
+      json: parent?.opts().json === true
+    });
+  });
+  wiki.command("status").description(
+    "Show recent compile jobs for a tenant (optionally filtered to one agent)."
+  ).option("-s, --stage <name>", "Deployment stage").option("-t, --tenant <slug>", "Tenant slug").option(
+    "--agent <id>",
+    "Restrict to a single agent. Omit for tenant-wide recent activity."
+  ).option("-n, --limit <n>", "Max jobs to return.", "10").option("--watch", "Poll until the most-recent job reaches a terminal state.").option(
+    "--timeout <seconds>",
+    "Max seconds to watch (default 900 = 15m).",
+    "900"
+  ).addHelpText(
+    "after",
+    `
+Examples:
+  # Tenant-wide (admin)
+  $ thinkwork wiki status --tenant acme
+
+  # Single agent, watching until a job settles
+  $ thinkwork wiki status --tenant acme --agent agt-xyz --watch
+`
+  ).action(async (opts, cmd) => {
+    const parent = cmd.parent?.parent;
+    await runWikiStatus({
+      ...opts,
+      stage: opts.stage ?? parent?.opts().stage,
+      json: parent?.opts().json === true
+    });
+  });
+}
+
 // src/cli.ts
 var program = new Command();
 program.name("thinkwork").description(
@@ -4676,6 +6591,8 @@ registerBudgetCommand(program);
 registerPerformanceCommand(program);
 registerTraceCommand(program);
 registerDashboardCommand(program);
+registerEvalCommand(program);
+registerWikiCommand(program);
 for (const cmd of program.commands) {
   if (!cmd.options.some((o) => o.long === "--json")) {
     cmd.option("--json", "Emit machine-readable JSON on stdout.");