axiom 0.34.1 → 0.35.0

package/dist/bin.cjs

@@ -43,137 +43,13 @@ var init_cjs_shims = __esm({
   }
 });
 
-// src/bin.ts
-var bin_exports = {};
-__export(bin_exports, {
-  program: () => program
-});
-module.exports = __toCommonJS(bin_exports);
-init_cjs_shims();
-var import_commander3 = require("commander");
-
-// src/cli/commands/eval.command.ts
-init_cjs_shims();
-var import_commander = require("commander");
-var import_nanoid = require("nanoid");
-
-// src/evals/run-vitest.ts
-init_cjs_shims();
-
-// ../../node_modules/.pnpm/tinyrainbow@2.0.0/node_modules/tinyrainbow/dist/node.js
-init_cjs_shims();
-
-// ../../node_modules/.pnpm/tinyrainbow@2.0.0/node_modules/tinyrainbow/dist/chunk-BVHSVHOK.js
-init_cjs_shims();
-var f = {
-  reset: [0, 0],
-  bold: [1, 22, "\x1B[22m\x1B[1m"],
-  dim: [2, 22, "\x1B[22m\x1B[2m"],
-  italic: [3, 23],
-  underline: [4, 24],
-  inverse: [7, 27],
-  hidden: [8, 28],
-  strikethrough: [9, 29],
-  black: [30, 39],
-  red: [31, 39],
-  green: [32, 39],
-  yellow: [33, 39],
-  blue: [34, 39],
-  magenta: [35, 39],
-  cyan: [36, 39],
-  white: [37, 39],
-  gray: [90, 39],
-  bgBlack: [40, 49],
-  bgRed: [41, 49],
-  bgGreen: [42, 49],
-  bgYellow: [43, 49],
-  bgBlue: [44, 49],
-  bgMagenta: [45, 49],
-  bgCyan: [46, 49],
-  bgWhite: [47, 49],
-  blackBright: [90, 39],
-  redBright: [91, 39],
-  greenBright: [92, 39],
-  yellowBright: [93, 39],
-  blueBright: [94, 39],
-  magentaBright: [95, 39],
-  cyanBright: [96, 39],
-  whiteBright: [97, 39],
-  bgBlackBright: [100, 49],
-  bgRedBright: [101, 49],
-  bgGreenBright: [102, 49],
-  bgYellowBright: [103, 49],
-  bgBlueBright: [104, 49],
-  bgMagentaBright: [105, 49],
-  bgCyanBright: [106, 49],
-  bgWhiteBright: [107, 49]
-};
-var h = Object.entries(f);
-function a(n) {
-  return String(n);
-}
-a.open = "";
-a.close = "";
-function C(n = false) {
-  let e = typeof process != "undefined" ? process : void 0, i = (e == null ? void 0 : e.env) || {}, g = (e == null ? void 0 : e.argv) || [];
-  return !("NO_COLOR" in i || g.includes("--no-color")) && ("FORCE_COLOR" in i || g.includes("--color") || (e == null ? void 0 : e.platform) === "win32" || n && i.TERM !== "dumb" || "CI" in i) || typeof window != "undefined" && !!window.chrome;
-}
-function p(n = false) {
-  let e = C(n), i = (r2, t, c, o) => {
-    let l = "", s2 = 0;
-    do
-      l += r2.substring(s2, o) + c, s2 = o + t.length, o = r2.indexOf(t, s2);
-    while (~o);
-    return l + r2.substring(s2);
-  }, g = (r2, t, c = r2) => {
-    let o = (l) => {
-      let s2 = String(l), b = s2.indexOf(t, r2.length);
-      return ~b ? r2 + i(s2, t, c, b) + t : r2 + s2 + t;
-    };
-    return o.open = r2, o.close = t, o;
-  }, u2 = {
-    isColorSupported: e
-  }, d = (r2) => `\x1B[${r2}m`;
-  for (let [r2, t] of h)
-    u2[r2] = e ? g(
-      d(t[0]),
-      d(t[1]),
-      t[2]
-    ) : a;
-  return u2;
-}
-
-// ../../node_modules/.pnpm/tinyrainbow@2.0.0/node_modules/tinyrainbow/dist/node.js
-var import_tty = require("tty");
-var r = process.env.FORCE_TTY !== void 0 || (0, import_tty.isatty)(1);
-var u = p(r);
-
-// src/evals/run-vitest.ts
-var import_node_path = require("path");
-var import_node_fs = require("fs");
-var import_node_os = require("os");
-var import_node_path2 = __toESM(require("path"), 1);
-var import_vite_tsconfig_paths = __toESM(require("vite-tsconfig-paths"), 1);
-var import_node = require("vitest/node");
-
-// src/evals/reporter.ts
-init_cjs_shims();
-
-// src/evals/context/storage.ts
-init_cjs_shims();
-var import_api9 = require("@opentelemetry/api");
-
 // src/evals/context/manager.ts
-init_cjs_shims();
-var import_node_module = require("module");
-var CONTEXT_MANAGER_SYMBOL = Symbol.for("axiom.context_manager");
 function getGlobalContextManager() {
   return globalThis[CONTEXT_MANAGER_SYMBOL];
 }
 function setGlobalContextManager(manager) {
   globalThis[CONTEXT_MANAGER_SYMBOL] = manager;
 }
-var isNodeJS = typeof process !== "undefined" && !!process.versions?.node;
 function getContextManager() {
   const existing = getGlobalContextManager();
   if (existing) return existing;
@@ -229,27 +105,34 @@ function createAsyncHook(_name) {
     }
   };
 }
-
-// src/app-scope.ts
-init_cjs_shims();
+var import_node_module, CONTEXT_MANAGER_SYMBOL, isNodeJS;
+var init_manager = __esm({
+  "src/evals/context/manager.ts"() {
+    "use strict";
+    init_cjs_shims();
+    import_node_module = require("module");
+    CONTEXT_MANAGER_SYMBOL = Symbol.for("axiom.context_manager");
+    isNodeJS = typeof process !== "undefined" && !!process.versions?.node;
+  }
+});
 
 // src/evals/context/global-flags.ts
-init_cjs_shims();
-var GLOBAL_OVERRIDES_SYMBOL = Symbol.for("axiom.global_flag_overrides");
 function setRoot(val) {
   globalThis[GLOBAL_OVERRIDES_SYMBOL] = val;
 }
 function setGlobalFlagOverrides(overrides2) {
   setRoot(overrides2);
 }
-
-// src/validate-flags.ts
-init_cjs_shims();
-var import_zod4 = require("zod");
+var GLOBAL_OVERRIDES_SYMBOL;
+var init_global_flags = __esm({
+  "src/evals/context/global-flags.ts"() {
+    "use strict";
+    init_cjs_shims();
+    GLOBAL_OVERRIDES_SYMBOL = Symbol.for("axiom.global_flag_overrides");
+  }
+});
 
 // src/cli/utils/format-zod-errors.ts
-init_cjs_shims();
-var import_zod = require("zod");
 function formatZodErrors(error) {
   const issues = error.issues;
   const messages = [];
@@ -345,13 +228,16 @@ function generateExampleForIssue(issue, path3) {
   }
   return null;
 }
-
-// src/util/dot-path.ts
-init_cjs_shims();
-var import_zod2 = require("zod");
+var import_zod;
+var init_format_zod_errors = __esm({
+  "src/cli/utils/format-zod-errors.ts"() {
+    "use strict";
+    init_cjs_shims();
+    import_zod = require("zod");
+  }
+});
 
 // src/util/zod-internals.ts
-init_cjs_shims();
 function isZodV4Schema(schema) {
   if (!schema || typeof schema !== "object") return false;
   const s2 = schema;
@@ -381,20 +267,6 @@ function getDefRawType(def) {
   if (raw == null) return void 0;
   return typeof raw === "string" ? raw : String(raw);
 }
-var KNOWN_KINDS = /* @__PURE__ */ new Set([
-  "object",
-  "optional",
-  "default",
-  "nullable",
-  "readonly",
-  "prefault",
-  "nonoptional",
-  "catch",
-  "array",
-  "record",
-  "union",
-  "discriminatedunion"
-]);
 function getKind(schemaOrDef) {
   const def = schemaOrDef && typeof schemaOrDef === "object" && "type" in schemaOrDef ? schemaOrDef : getDef(schemaOrDef);
   const raw = getDefRawType(def);
@@ -429,16 +301,6 @@ function getDefaultValue(schema) {
   const def = getDef(schema);
   return def?.defaultValue;
 }
-var TRANSPARENT_WRAPPERS = [
-  "optional",
-  "nullable",
-  "default",
-  "readonly",
-  "prefault",
-  "nonoptional",
-  "catch"
-  // transparent for schema structure, but alters error behavior
-];
 function unwrapTransparent(schema) {
   let current = schema;
   for (let i = 0; i < 10; i++) {
@@ -454,6 +316,37 @@ function unwrapTransparent(schema) {
   }
   return current;
 }
+var KNOWN_KINDS, TRANSPARENT_WRAPPERS;
+var init_zod_internals = __esm({
+  "src/util/zod-internals.ts"() {
+    "use strict";
+    init_cjs_shims();
+    KNOWN_KINDS = /* @__PURE__ */ new Set([
+      "object",
+      "optional",
+      "default",
+      "nullable",
+      "readonly",
+      "prefault",
+      "nonoptional",
+      "catch",
+      "array",
+      "record",
+      "union",
+      "discriminatedunion"
+    ]);
+    TRANSPARENT_WRAPPERS = [
+      "optional",
+      "nullable",
+      "default",
+      "readonly",
+      "prefault",
+      "nonoptional",
+      "catch"
+      // transparent for schema structure, but alters error behavior
+    ];
+  }
+});
 
 // src/util/dot-path.ts
 function parsePath(path3) {
@@ -509,10 +402,17 @@ function isValidPath(schema, segments) {
   }
   return true;
 }
+var import_zod2;
+var init_dot_path = __esm({
+  "src/util/dot-path.ts"() {
+    "use strict";
+    init_cjs_shims();
+    import_zod2 = require("zod");
+    init_zod_internals();
+  }
+});
 
 // src/util/deep-partial-schema.ts
-init_cjs_shims();
-var import_zod3 = require("zod");
 function makeDeepPartial(schema) {
   const shape = schema.shape;
   const newShape = {};
@@ -562,180 +462,246 @@ function makeDeepPartialField(fieldSchema) {
   }
   return fieldSchema.optional();
 }
+var import_zod3;
+var init_deep_partial_schema = __esm({
+  "src/util/deep-partial-schema.ts"() {
+    "use strict";
+    init_cjs_shims();
+    import_zod3 = require("zod");
+    init_zod_internals();
+  }
+});
 
-// src/app-scope.ts
-var import_api8 = require("@opentelemetry/api");
-var import_zod5 = require("zod");
+// src/validate-flags.ts
+var import_zod4;
+var init_validate_flags = __esm({
+  "src/validate-flags.ts"() {
+    "use strict";
+    init_cjs_shims();
+    import_zod4 = require("zod");
+    init_global_flags();
+    init_format_zod_errors();
+    init_dot_path();
+    init_deep_partial_schema();
+    init_zod_internals();
+  }
+});
 
 // src/otel/utils/to-otel-attribute.ts
-init_cjs_shims();
-var import_api = require("@opentelemetry/api");
-
-// src/otel/index.ts
-init_cjs_shims();
-
-// src/otel/withSpan.ts
-init_cjs_shims();
-var import_api6 = require("@opentelemetry/api");
+var import_api;
+var init_to_otel_attribute = __esm({
+  "src/otel/utils/to-otel-attribute.ts"() {
+    "use strict";
+    init_cjs_shims();
+    import_api = require("@opentelemetry/api");
+  }
+});
 
 // src/otel/withSpanBaggageKey.ts
-init_cjs_shims();
-
-// src/otel/utils/wrapperUtils.ts
-init_cjs_shims();
-var import_api5 = require("@opentelemetry/api");
-
-// src/otel/semconv/attributes.ts
-init_cjs_shims();
-var import_semantic_conventions = require("@opentelemetry/semantic-conventions");
+var init_withSpanBaggageKey = __esm({
+  "src/otel/withSpanBaggageKey.ts"() {
+    "use strict";
+    init_cjs_shims();
+  }
+});
 
 // src/otel/semconv/eval_proposal.ts
-init_cjs_shims();
+var init_eval_proposal = __esm({
+  "src/otel/semconv/eval_proposal.ts"() {
+    "use strict";
+    init_cjs_shims();
+  }
+});
 
 // src/otel/semconv/attributes.ts
-var import_incubating = require("@opentelemetry/semantic-conventions/incubating");
+var import_semantic_conventions, import_incubating;
+var init_attributes = __esm({
+  "src/otel/semconv/attributes.ts"() {
+    "use strict";
+    init_cjs_shims();
+    import_semantic_conventions = require("@opentelemetry/semantic-conventions");
+    init_eval_proposal();
+    import_incubating = require("@opentelemetry/semantic-conventions/incubating");
+  }
+});
 
-// src/otel/startActiveSpan.ts
-init_cjs_shims();
-var import_api2 = require("@opentelemetry/api");
+// src/schema.ts
+var SCHEMA_VERSION, SCHEMA_BASE_URL, SCHEMA_URL;
+var init_schema = __esm({
+  "src/schema.ts"() {
+    "use strict";
+    init_cjs_shims();
+    SCHEMA_VERSION = "0.0.2";
+    SCHEMA_BASE_URL = "https://axiom.co/ai/schemas/";
+    SCHEMA_URL = `${SCHEMA_BASE_URL}${SCHEMA_VERSION}`;
+  }
+});
 
-// src/otel/initAxiomAI.ts
-init_cjs_shims();
-var import_api4 = require("@opentelemetry/api");
+// src/otel/startActiveSpan.ts
+var import_api2;
+var init_startActiveSpan = __esm({
+  "src/otel/startActiveSpan.ts"() {
+    "use strict";
+    init_cjs_shims();
+    import_api2 = require("@opentelemetry/api");
+  }
+});
 
 // package.json
-var package_default = {
-  name: "axiom",
-  version: "0.34.1",
-  type: "module",
-  author: "Axiom, Inc.",
-  contributors: [
-    "Islam Shehata <islam@axiom.co>",
-    "Chris Ehrlich <chris@axiom.co>",
-    "Gabriel de Andrade <gabriel@axiom.co>"
-  ],
-  scripts: {
-    dev: "tsup --watch",
-    build: "tsup && chmod +x dist/bin.js",
-    format: "prettier --write .",
-    "format:check": "prettier --check .",
-    lint: "eslint './**/*.{js,ts}'",
-    typecheck: "tsc --noEmit",
-    test: "vitest run",
-    "test:watch": "vitest --watch",
-    publint: "npx publint"
-  },
-  types: "./dist/index.d.ts",
-  main: "./dist/index.cjs",
-  module: "./dist/index.js",
-  bin: {
-    axiom: "./dist/bin.js"
-  },
-  exports: {
-    "./ai": {
-      import: {
-        types: "./dist/index.d.ts",
-        default: "./dist/index.js"
+var package_default;
+var init_package = __esm({
+  "package.json"() {
+    package_default = {
+      name: "axiom",
+      version: "0.35.0",
+      type: "module",
+      author: "Axiom, Inc.",
+      contributors: [
+        "Islam Shehata <islam@axiom.co>",
+        "Chris Ehrlich <chris@axiom.co>",
+        "Gabriel de Andrade <gabriel@axiom.co>"
+      ],
+      scripts: {
+        dev: "tsup --watch",
+        build: "tsup && chmod +x dist/bin.js",
+        format: "prettier --write .",
+        "format:check": "prettier --check .",
+        lint: "eslint './**/*.{js,ts}'",
+        typecheck: "tsc --noEmit",
+        test: "vitest run",
+        "test:watch": "vitest --watch",
+        publint: "npx publint"
       },
-      require: {
-        types: "./dist/index.d.cts",
-        default: "./dist/index.cjs"
-      }
-    },
-    "./ai/evals": {
-      import: {
-        types: "./dist/evals.d.ts",
-        default: "./dist/evals.js"
+      types: "./dist/index.d.ts",
+      main: "./dist/index.cjs",
+      module: "./dist/index.js",
+      bin: {
+        axiom: "./dist/bin.js"
       },
-      require: {
-        types: "./dist/evals.d.cts",
-        default: "./dist/evals.cjs"
-      }
-    },
-    "./ai/config": {
-      import: {
-        types: "./dist/config.d.ts",
-        default: "./dist/config.js"
+      exports: {
+        "./ai": {
+          import: {
+            types: "./dist/index.d.ts",
+            default: "./dist/index.js"
+          },
+          require: {
+            types: "./dist/index.d.cts",
+            default: "./dist/index.cjs"
+          }
+        },
+        "./ai/evals": {
+          import: {
+            types: "./dist/evals.d.ts",
+            default: "./dist/evals.js"
+          },
+          require: {
+            types: "./dist/evals.d.cts",
+            default: "./dist/evals.cjs"
+          }
+        },
+        "./ai/config": {
+          import: {
+            types: "./dist/config.d.ts",
+            default: "./dist/config.js"
+          },
+          require: {
+            types: "./dist/config.d.cts",
+            default: "./dist/config.cjs"
+          }
+        },
+        "./ai/feedback": {
+          import: {
+            types: "./dist/feedback.d.ts",
+            default: "./dist/feedback.js"
+          },
+          require: {
+            types: "./dist/feedback.d.cts",
+            default: "./dist/feedback.cjs"
+          }
+        }
       },
-      require: {
-        types: "./dist/config.d.cts",
-        default: "./dist/config.cjs"
-      }
-    }
-  },
-  keywords: [
-    "axiom",
-    "logging",
-    "ai",
-    "otel",
-    "opentelemetry"
-  ],
-  repository: {
-    type: "git",
-    url: "git+https://github.com/axiomhq/ai.git",
-    directory: "packages/ai"
-  },
-  license: "MIT",
-  dependencies: {
-    "@next/env": "^15.4.2",
-    "@opentelemetry/auto-instrumentations-node": "^0.60.1",
-    "@opentelemetry/context-async-hooks": "^2.0.1",
-    "@opentelemetry/exporter-trace-otlp-http": "^0.202.0",
-    "@opentelemetry/resources": "^2.0.1",
-    "@opentelemetry/sdk-trace-node": "^2.0.1",
-    "@opentelemetry/semantic-conventions": "^1.37.0",
-    "@sinclair/typebox": "^0.34.37",
-    c12: "^2.0.4",
-    commander: "^14.0.0",
-    defu: "^6.1.4",
-    handlebars: "^4.7.8",
-    nanoid: "^5.1.5",
-    open: "^10.1.0",
-    "vite-tsconfig-paths": "^5.1.4"
-  },
-  peerDependencies: {
-    "@opentelemetry/api": "^1.9.0",
-    zod: "^3.25.0 || ^4.0.0"
-  },
-  devDependencies: {
-    "@ai-sdk/anthropicv1": "npm:@ai-sdk/anthropic@^1.2.12",
-    "@ai-sdk/anthropicv2": "npm:@ai-sdk/anthropic@^2.0.44",
-    "@ai-sdk/openaiv1": "npm:@ai-sdk/openai@^1.3.24",
-    "@ai-sdk/openaiv2": "npm:@ai-sdk/openai@^2.0.67",
-    "@ai-sdk/providerv1": "npm:@ai-sdk/provider@^1.1.3",
-    "@ai-sdk/providerv2": "npm:@ai-sdk/provider@^2.0.0",
-    "@opentelemetry/api": "^1.9.0",
-    "@opentelemetry/core": "^2.0.1",
-    "@opentelemetry/sdk-trace-base": "^2.0.1",
-    "@opentelemetry/sdk-trace-node": "^2.0.1",
-    "@repo/eslint-config": "workspace:*",
-    "@types/node": "^22.15.29",
-    "@vitest/coverage-v8": "^4.0.0",
-    aiv4: "npm:ai@^4.3.19",
-    aiv5: "npm:ai@^5.0.93",
-    esbuild: "^0.25.8",
-    eslint: "catalog:",
-    msw: "^2.12.2",
-    prettier: "catalog:",
-    tinyrainbow: "^2.0.0",
-    tsup: "catalog:",
-    typescript: "catalog:",
-    vitest: "catalog:",
-    zod: "catalog:"
-  },
-  files: [
-    "dist"
-  ],
-  packageManager: "pnpm@10.16.1"
-};
+      keywords: [
+        "axiom",
+        "logging",
+        "ai",
+        "otel",
+        "opentelemetry"
+      ],
+      repository: {
+        type: "git",
+        url: "git+https://github.com/axiomhq/ai.git",
+        directory: "packages/ai"
+      },
+      license: "MIT",
+      dependencies: {
+        "@next/env": "^15.4.2",
+        "@opentelemetry/auto-instrumentations-node": "^0.60.1",
+        "@opentelemetry/context-async-hooks": "^2.0.1",
+        "@opentelemetry/exporter-trace-otlp-http": "^0.202.0",
+        "@opentelemetry/resources": "^2.0.1",
+        "@opentelemetry/sdk-trace-node": "^2.0.1",
+        "@opentelemetry/semantic-conventions": "^1.37.0",
+        "@sinclair/typebox": "^0.34.37",
+        c12: "^2.0.4",
+        commander: "^14.0.0",
+        defu: "^6.1.4",
+        handlebars: "^4.7.8",
+        nanoid: "^5.1.5",
+        open: "^10.1.0",
+        "vite-tsconfig-paths": "^5.1.4",
+        vitest: "^4.0.0"
+      },
+      peerDependencies: {
+        "@opentelemetry/api": "^1.9.0",
+        zod: "^3.25.0 || ^4.0.0"
+      },
+      devDependencies: {
+        "@ai-sdk/anthropicv1": "npm:@ai-sdk/anthropic@^1.2.12",
+        "@ai-sdk/anthropicv2": "npm:@ai-sdk/anthropic@^2.0.44",
+        "@ai-sdk/openaiv1": "npm:@ai-sdk/openai@^1.3.24",
+        "@ai-sdk/openaiv2": "npm:@ai-sdk/openai@^2.0.67",
+        "@ai-sdk/providerv1": "npm:@ai-sdk/provider@^1.1.3",
+        "@ai-sdk/providerv2": "npm:@ai-sdk/provider@^2.0.0",
+        "@opentelemetry/api": "^1.9.0",
+        "@opentelemetry/core": "^2.0.1",
+        "@opentelemetry/sdk-trace-base": "^2.0.1",
+        "@opentelemetry/sdk-trace-node": "^2.0.1",
+        "@repo/eslint-config": "workspace:*",
+        "@types/node": "^22.15.29",
+        "@vitest/coverage-v8": "^4.0.0",
+        aiv4: "npm:ai@^4.3.19",
+        aiv5: "npm:ai@^5.0.93",
+        esbuild: "^0.25.8",
+        eslint: "catalog:",
+        msw: "^2.12.2",
+        prettier: "catalog:",
+        tinyrainbow: "^2.0.0",
+        tsup: "catalog:",
+        typescript: "catalog:",
+        vitest: "catalog:",
+        zod: "catalog:"
+      },
+      files: [
+        "dist"
+      ],
+      packageManager: "pnpm@10.16.1"
+    };
+  }
+});
 
 // src/otel/utils/redaction.ts
-init_cjs_shims();
-var import_api3 = require("@opentelemetry/api");
-var AXIOM_AI_REDACTION_KEY = Symbol.for("__axiom_ai_redaction__");
+var import_api3, AXIOM_AI_REDACTION_KEY;
+var init_redaction = __esm({
+  "src/otel/utils/redaction.ts"() {
+    "use strict";
+    init_cjs_shims();
+    import_api3 = require("@opentelemetry/api");
+    init_withSpanBaggageKey();
+    AXIOM_AI_REDACTION_KEY = Symbol.for("__axiom_ai_redaction__");
+  }
+});
 
 // src/otel/initAxiomAI.ts
-var AXIOM_AI_SCOPE_KEY = Symbol.for("__axiom_ai_scope__");
 function extractTracerScope(tracer) {
   const tracerAny = tracer;
   const name = tracerAny._instrumentationScope?.name || tracerAny.instrumentationLibrary?.name || package_default.name;
@@ -758,51 +724,186 @@ function initAxiomAI(config) {
     globalThis[AXIOM_AI_REDACTION_KEY] = config.redactionPolicy;
   }
 }
+var import_api4, AXIOM_AI_SCOPE_KEY;
+var init_initAxiomAI = __esm({
+  "src/otel/initAxiomAI.ts"() {
+    "use strict";
+    init_cjs_shims();
+    import_api4 = require("@opentelemetry/api");
+    init_package();
+    init_redaction();
+    AXIOM_AI_SCOPE_KEY = Symbol.for("__axiom_ai_scope__");
+  }
+});
 
-// src/util/name-validation-runtime.ts
-init_cjs_shims();
+// src/otel/utils/wrapperUtils.ts
+var import_api5;
+var init_wrapperUtils = __esm({
+  "src/otel/utils/wrapperUtils.ts"() {
+    "use strict";
+    init_cjs_shims();
+    import_api5 = require("@opentelemetry/api");
+    init_attributes();
+    init_schema();
+    init_withSpanBaggageKey();
+    init_startActiveSpan();
+    init_initAxiomAI();
+  }
+});
 
-// src/otel/vercel.ts
-init_cjs_shims();
+// src/util/name-validation-runtime.ts
+var init_name_validation_runtime = __esm({
+  "src/util/name-validation-runtime.ts"() {
+    "use strict";
+    init_cjs_shims();
+  }
+});
 
-// src/otel/AxiomWrappedLanguageModelV1.ts
-init_cjs_shims();
+// src/otel/withSpan.ts
+var import_api6;
+var init_withSpan = __esm({
+  "src/otel/withSpan.ts"() {
+    "use strict";
+    init_cjs_shims();
+    import_api6 = require("@opentelemetry/api");
+    init_withSpanBaggageKey();
+    init_wrapperUtils();
+    init_name_validation_runtime();
+  }
+});
 
-// src/otel/middleware.ts
-init_cjs_shims();
-var import_api7 = require("@opentelemetry/api");
+// src/otel/utils/contentSanitizer.ts
+var init_contentSanitizer = __esm({
+  "src/otel/utils/contentSanitizer.ts"() {
+    "use strict";
+    init_cjs_shims();
+  }
+});
 
 // src/otel/completionUtils.ts
-init_cjs_shims();
-
-// src/otel/utils/contentSanitizer.ts
-init_cjs_shims();
+var init_completionUtils = __esm({
+  "src/otel/completionUtils.ts"() {
+    "use strict";
+    init_cjs_shims();
+    init_contentSanitizer();
+  }
+});
 
 // src/util/promptUtils.ts
-init_cjs_shims();
+var init_promptUtils = __esm({
+  "src/util/promptUtils.ts"() {
+    "use strict";
+    init_cjs_shims();
+  }
+});
 
 // src/otel/utils/normalized.ts
-init_cjs_shims();
+var init_normalized = __esm({
+  "src/otel/utils/normalized.ts"() {
+    "use strict";
+    init_cjs_shims();
+  }
+});
+
+// src/util/currentUnixTime.ts
+var init_currentUnixTime = __esm({
+  "src/util/currentUnixTime.ts"() {
+    "use strict";
+    init_cjs_shims();
+  }
+});
 
 // src/otel/streaming/aggregators.ts
-init_cjs_shims();
+var init_aggregators = __esm({
+  "src/otel/streaming/aggregators.ts"() {
+    "use strict";
+    init_cjs_shims();
+    init_currentUnixTime();
+  }
+});
 
-// src/util/currentUnixTime.ts
-init_cjs_shims();
+// src/otel/middleware.ts
+var import_api7;
+var init_middleware = __esm({
+  "src/otel/middleware.ts"() {
+    "use strict";
+    init_cjs_shims();
+    import_api7 = require("@opentelemetry/api");
+    init_attributes();
+    init_completionUtils();
+    init_promptUtils();
+    init_contentSanitizer();
+    init_wrapperUtils();
+    init_normalized();
+    init_aggregators();
+    init_redaction();
+  }
+});
+
+// src/otel/AxiomWrappedLanguageModelV1.ts
+var init_AxiomWrappedLanguageModelV1 = __esm({
+  "src/otel/AxiomWrappedLanguageModelV1.ts"() {
+    "use strict";
+    init_cjs_shims();
+    init_middleware();
+  }
+});
 
 // src/otel/AxiomWrappedLanguageModelV2.ts
-init_cjs_shims();
+var init_AxiomWrappedLanguageModelV2 = __esm({
+  "src/otel/AxiomWrappedLanguageModelV2.ts"() {
+    "use strict";
+    init_cjs_shims();
+    init_middleware();
+  }
+});
+
+// src/otel/vercel.ts
+var init_vercel = __esm({
+  "src/otel/vercel.ts"() {
+    "use strict";
+    init_cjs_shims();
+    init_AxiomWrappedLanguageModelV1();
+    init_AxiomWrappedLanguageModelV2();
+  }
+});
+
+// src/otel/index.ts
+var init_otel = __esm({
+  "src/otel/index.ts"() {
+    "use strict";
+    init_cjs_shims();
+    init_withSpan();
+    init_vercel();
+    init_attributes();
+  }
+});
+
+// src/app-scope.ts
+var import_api8, import_zod5;
+var init_app_scope = __esm({
+  "src/app-scope.ts"() {
+    "use strict";
+    init_cjs_shims();
+    init_global_flags();
+    init_storage();
+    init_validate_flags();
+    init_dot_path();
+    init_zod_internals();
+    import_api8 = require("@opentelemetry/api");
+    import_zod5 = require("zod");
+    init_to_otel_attribute();
+    init_otel();
+  }
+});
 
 // src/evals/context/storage.ts
-var CONFIG_SCOPE_SYMBOL = Symbol.for("axiom.eval.configScope");
-var AXIOM_CONFIG_SYMBOL = Symbol.for("axiom.eval.config");
 function getAxiomConfig() {
   return globalThis[AXIOM_CONFIG_SYMBOL];
 }
 function setAxiomConfig(config) {
   globalThis[AXIOM_CONFIG_SYMBOL] = config;
 }
-var EVAL_CONTEXT = createAsyncHook("eval-context");
 function getEvalContext() {
   const ctx = EVAL_CONTEXT.get();
   if (!ctx) {
@@ -842,628 +943,297 @@ function withEvalContext(options = {}, fn) {
     fn
   );
 }
-
-// src/evals/reporter.console-utils.ts
-init_cjs_shims();
-function formatPercentage(value) {
-  if (!Number.isFinite(value)) {
-    return "N/A";
-  }
-  return Number(value * 100).toFixed(2) + "%";
-}
-function formatDiff(current, baseline) {
-  if (!Number.isFinite(current) || !Number.isFinite(baseline)) {
-    return { text: "N/A", color: u.dim };
+var import_api9, CONFIG_SCOPE_SYMBOL, AXIOM_CONFIG_SYMBOL, EVAL_CONTEXT;
+var init_storage = __esm({
+  "src/evals/context/storage.ts"() {
+    "use strict";
+    init_cjs_shims();
+    import_api9 = require("@opentelemetry/api");
+    init_manager();
+    init_app_scope();
+    CONFIG_SCOPE_SYMBOL = Symbol.for("axiom.eval.configScope");
+    AXIOM_CONFIG_SYMBOL = Symbol.for("axiom.eval.config");
+    EVAL_CONTEXT = createAsyncHook("eval-context");
   }
-  const diff = current - baseline;
-  const diffText = (diff >= 0 ? "+" : "") + formatPercentage(diff);
-  const color = diff > 0 ? u.green : diff < 0 ? u.red : u.dim;
-  return { text: diffText, color };
-}
-function truncate(str, max) {
-  return str.length > max ? str.slice(0, max) + "\u2026" : str;
-}
-function stringify(value) {
-  try {
-    if (typeof value === "string") return value;
-    return JSON.stringify(value);
-  } catch {
-    return String(value);
+});
+
+// src/context.ts
+function overrideFlags(partial) {
+  const current = getEvalContext();
+  if (!current) {
+    if (process.env.NODE_ENV !== "test") {
+      console.warn("overrideFlags called outside of evaluation context");
+    }
+    return;
   }
-}
-function getCaseFingerprint(input, expected) {
-  const inputStr = typeof input === "string" ? input : JSON.stringify(input);
-  const expectedStr = typeof expected === "string" ? expected : JSON.stringify(expected);
-  return JSON.stringify({ input: inputStr, expected: expectedStr });
-}
-function printEvalNameAndFileName(testSuite, meta, logger = console.log) {
-  const cwd = process.cwd();
-  logger(
-    " ",
-    u.bgCyan(u.black(` ${testSuite.project.name} `)),
-    u.bgBlue(u.black(` ${meta.evaluation.name}-${meta.evaluation.version} `)),
-    u.dim(`(${testSuite.children.size} cases)`)
-  );
-  logger(" ", u.dim(testSuite.module.moduleId.replace(cwd, "")));
-}
-function printBaselineNameAndVersion(testMeta, logger = console.log) {
-  if (testMeta.evaluation.baseline) {
-    logger(
-      " ",
-      " baseline ",
-      u.bgMagenta(
-        u.black(` ${testMeta.evaluation.baseline.name}-${testMeta.evaluation.baseline.version} `)
-      )
-    );
-  } else {
-    logger(" ", u.bgWhite(u.blackBright(" baseline: ")), "none");
+  const overlayContext = {
+    ...current,
+    flags: { ...current.flags, ...partial },
+    // Merge for backwards compatibility
+    parent: current,
+    overrides: { ...partial }
+  };
+  const currentCtx = EVAL_CONTEXT.get();
+  if (currentCtx) {
+    Object.assign(currentCtx, overlayContext);
   }
-  logger("");
-}
-function printRuntimeFlags(testMeta, logger = console.log) {
-  if (testMeta.case.runtimeFlags && Object.keys(testMeta.case.runtimeFlags).length > 0) {
-    const entries = Object.entries(testMeta.case.runtimeFlags);
-    logger("   ", u.dim("runtime flags"));
-    for (const [k, v] of entries) {
-      switch (v.kind) {
-        case "replaced": {
-          const valText = truncate(stringify(v.value), 80);
-          const defText = truncate(stringify(v.default), 80);
-          logger("     ", `${k}: ${valText} (default: ${defText})`);
-          break;
-        }
-        case "introduced": {
-          const valText = truncate(stringify(v.value), 80);
-          logger("     ", `${k}: ${valText} (no default)`);
-          break;
-        }
-      }
-    }
+  for (const [key, value] of Object.entries(partial)) {
+    putOnSpan("flag", key, value);
   }
 }
-function printTestCaseCountStartDuration(testSuite, startTime, duration, logger = console.log) {
-  logger(" ");
-  logger(" ", u.dim("Cases"), testSuite.children.size);
-  logger(" ", u.dim("Start at"), new Date(startTime).toTimeString());
-  logger(" ", u.dim("Duration"), `${duration}s`);
+var init_context = __esm({
+  "src/context.ts"() {
+    "use strict";
+    init_cjs_shims();
+    init_storage();
+  }
+});
+
+// src/cli/utils/eval-context-runner.ts
+async function runEvalWithContext(overrides2, runFn) {
+  setGlobalFlagOverrides(overrides2);
+  return withEvalContext({ initialFlags: overrides2 }, async () => {
+    if (Object.keys(overrides2).length > 0) {
+      overrideFlags(overrides2);
+    }
+    return runFn();
+  });
 }
-function printTestCaseSuccessOrFailed(testMeta, ok, logger = console.log) {
-  const index = testMeta.case.index;
-  if (ok) {
-    logger(" ", u.yellow(` \u2714 case ${index}:`));
-  } else {
-    logger(" ", u.red(` \u2716 case ${index}: failed`));
-    for (const e of testMeta.case.errors ?? []) {
-      logger("", e.message);
+var init_eval_context_runner = __esm({
+  "src/cli/utils/eval-context-runner.ts"() {
+    "use strict";
+    init_cjs_shims();
+    init_context();
+    init_storage();
+    init_global_flags();
+  }
+});
+
+// src/cli/utils/parse-flag-overrides.ts
+function ensureNoSpaceSeparatedSyntax(flagName, value, nextToken, flagType) {
+  if (value === void 0 && nextToken !== void 0) {
+    if (flagType === "flag" && !nextToken.startsWith("-") && nextToken !== "true" && nextToken !== "false") {
+      console.error(`\u274C Invalid syntax: --flag.${flagName} ${nextToken}`);
+      console.error(`\u{1F4A1} Use: --flag.${flagName}=${nextToken}`);
+      process.exit(1);
+    } else if (flagType === "config" && !nextToken.startsWith("-")) {
+      console.error(`\u274C Invalid syntax: --flags-config ${nextToken}`);
+      console.error(`\u{1F4A1} Use: --flags-config=${nextToken}`);
+      process.exit(1);
     }
   }
 }
-function printTestCaseScores(testMeta, baselineCase, logger = console.log) {
-  const scores = testMeta.case.scores;
-  const keys = Object.keys(scores);
-  if (keys.length === 0) {
-    return;
+function collectFlagValidationErrors(overrides2, flagSchema) {
+  if (!flagSchema || Object.keys(overrides2).length === 0) {
+    return { success: true, errors: [] };
   }
-  const maxNameLength = Math.max(...keys.map((k) => k.length));
-  keys.forEach((k) => {
-    const scoreData = scores[k];
-    const hasError = scoreData.metadata?.error;
-    const v = scoreData.score ? scoreData.score : 0;
-    const rawCurrent = hasError ? "N/A" : formatPercentage(v);
-    const paddedCurrent = rawCurrent.padStart(7);
-    const coloredCurrent = hasError ? u.dim(paddedCurrent) : u.magentaBright(paddedCurrent);
-    const paddedName = k.padEnd(maxNameLength);
-    if (baselineCase?.scores[k]) {
-      const baselineScoreValue = baselineCase.scores[k].value;
-      const rawBaseline = formatPercentage(baselineScoreValue);
-      const paddedBaseline = rawBaseline.padStart(7);
-      const coloredBaseline = u.blueBright(paddedBaseline);
-      const { text: diffText, color: diffColor } = formatDiff(v, baselineScoreValue);
-      const paddedDiff = diffText.padStart(8);
-      logger(
-        `    ${paddedName}  ${coloredBaseline} \u2192 ${coloredCurrent}  ${hasError ? u.dim("(scorer not run)") : u.dim("(") + diffColor(paddedDiff) + u.dim(")")}`
-      );
-    } else {
-      logger(`    ${paddedName}  ${coloredCurrent} ${hasError ? u.dim("(scorer not run)") : ""}`);
+  assertZodV4(flagSchema, "flagSchema");
+  const schema = flagSchema;
+  const errors = [];
+  for (const dotPath of Object.keys(overrides2)) {
+    const segments = parsePath(dotPath);
+    if (!isValidPath(schema, segments)) {
+      errors.push({ type: "invalid_path", path: dotPath });
     }
-  });
+  }
+  if (errors.length > 0) {
+    return { success: false, errors };
+  }
+  const nestedObject = dotNotationToNested(overrides2);
+  const deepPartialSchema = makeDeepPartial(schema);
+  const result = deepPartialSchema.safeParse(nestedObject);
+  if (!result.success) {
+    errors.push({ type: "invalid_value", zodError: result.error });
+  }
+  return { success: errors.length === 0, errors };
 }
-function printOutOfScopeFlags(testMeta, logger = console.log) {
-  if (testMeta.case.outOfScopeFlags && testMeta.case.outOfScopeFlags.length > 0) {
-    const pickedFlagsText = testMeta.case.pickedFlags ? `(picked: ${testMeta.case.pickedFlags.map((f2) => `'${f2}'`).join(", ")})` : "(none)";
-    logger("   ", u.yellow(`\u26A0 Out-of-scope flags: ${pickedFlagsText}`));
-    testMeta.case.outOfScopeFlags.forEach((flag) => {
-      const timeStr = new Date(flag.accessedAt).toLocaleTimeString();
-      logger("     ", `${flag.flagPath} (at ${timeStr})`);
-      if (flag.stackTrace && flag.stackTrace.length > 0) {
-        flag.stackTrace.forEach((frame, i) => {
-          const prefix = i === flag.stackTrace.length - 1 ? " \u2514\u2500" : " \u251C\u2500";
-          logger("     ", u.dim(`${prefix} ${frame}`));
-        });
+function printFlagValidationErrorsAndExit(errors) {
+  console.error("\u274C Invalid CLI flags:");
+  for (const error of errors) {
+    if (error.type === "invalid_path") {
+      console.error(`  \u2022 flag '${error.path}': Invalid flag path`);
+    } else {
+      console.error(formatZodErrors(error.zodError));
+      const examples = generateFlagExamples(error.zodError);
+      if (examples.length > 0) {
+        console.error("\n\u{1F4A1} Valid examples:");
+        examples.forEach((example) => console.error(`  ${example}`));
       }
-    });
+    }
   }
+  process.exit(1);
 }
-function printCaseResult(test, baselineCasesByFingerprint, matchedIndices, logger = console.log) {
-  const ok = test.ok();
-  const testMeta = test.meta();
-  if (!testMeta?.case) {
-    return;
-  }
-  printTestCaseSuccessOrFailed(testMeta, ok, logger);
-  const fingerprint = getCaseFingerprint(testMeta.case.input, testMeta.case.expected);
-  const baselineCases = baselineCasesByFingerprint.get(fingerprint);
-  const baselineCase = baselineCases?.shift();
-  if (baselineCase) {
-    matchedIndices.add(baselineCase.index);
+function validateFlagOverrides(overrides2, flagSchema) {
+  const result = collectFlagValidationErrors(overrides2, flagSchema);
+  if (!result.success) {
+    printFlagValidationErrorsAndExit(result.errors);
   }
-  printTestCaseScores(testMeta, baselineCase, logger);
-  printRuntimeFlags(testMeta, logger);
-  printOutOfScopeFlags(testMeta, logger);
 }
-function printOrphanedBaselineCases(baseline, matchedIndices, logger = console.log) {
-  const orphanedCases = baseline.cases.filter((c) => !matchedIndices.has(c.index));
-  if (orphanedCases.length === 0) {
-    return;
+function coerceValue(raw) {
+  if (raw === "true") return true;
+  if (raw === "false") return false;
+  const num = Number(raw);
+  if (!Number.isNaN(num) && raw.trim() === num.toString()) {
+    return num;
   }
-  logger("");
-  logger(" ", u.yellow("Orphaned baseline cases:"));
-  for (const orphanedCase of orphanedCases) {
-    logger(
-      " ",
-      u.dim(
-        `case ${orphanedCase.index}: ${truncate(orphanedCase.input, 50)} (score: ${truncate(
-          JSON.stringify(orphanedCase.scores),
-          50
-        )})`
-      )
-    );
-    const keys = Object.keys(orphanedCase.scores);
-    if (keys.length > 0) {
-      const maxNameLength = Math.max(...keys.map((k) => k.length));
-      keys.forEach((k) => {
-        const scoreData = orphanedCase.scores[k];
-        const rawScore = formatPercentage(scoreData.value);
-        const paddedName = k.padEnd(maxNameLength);
-        const paddedScore = rawScore.padStart(7);
-        logger(`    ${paddedName}  ${u.blueBright(paddedScore)}`);
-      });
-    }
+  try {
+    return JSON.parse(raw);
+  } catch {
+    return raw;
   }
 }
-var reporterDate = (d) => {
-  const date = d.toISOString().slice(0, 10);
-  const hours = d.getUTCHours().toString().padStart(2, "0");
-  const minutes = d.getUTCMinutes().toString().padStart(2, "0");
-  return `${date}, ${hours}:${minutes} UTC`;
-};
-function printGlobalFlagOverrides(overrides2, defaults, logger = console.log) {
-  if (Object.keys(overrides2).length === 0) {
-    logger("");
-    logger(u.dim("Flag overrides: (none)"));
-    logger("");
-    return;
-  }
-  logger("");
-  logger("Flag overrides:");
-  for (const [key, value] of Object.entries(overrides2)) {
-    const defaultValue = defaults[key];
-    const valueStr = JSON.stringify(value);
-    const defaultStr = defaultValue !== void 0 ? JSON.stringify(defaultValue) : "none";
-    logger(`  \u2022 ${key}: ${valueStr} ${u.dim(`(default: ${defaultStr})`)}`);
-  }
-  logger("");
-}
-function printSuiteBox({
-  suite,
-  scorerAverages,
-  calculateBaselineScorerAverage: calculateBaselineScorerAverage2,
-  flagDiff,
-  logger = console.log
-}) {
-  const filename = suite.file.split("/").pop();
-  logger("\u250C\u2500");
-  logger(`\u2502  ${u.blue(suite.name)} ${u.gray(`(${filename})`)}`);
-  logger("\u251C\u2500");
-  const scorerNames = Object.keys(scorerAverages);
-  const maxNameLength = Math.max(...scorerNames.map((name) => name.length));
-  const allCasesErrored = (scorerName) => {
-    return suite.cases.every((caseData) => caseData.scores[scorerName]?.metadata?.error);
-  };
-  for (const scorerName of scorerNames) {
-    const avg = scorerAverages[scorerName];
-    const paddedName = scorerName.padEnd(maxNameLength);
-    const hasAllErrors = allCasesErrored(scorerName);
-    const baselineAvg = suite.baseline ? calculateBaselineScorerAverage2(suite.baseline, scorerName) : null;
-    if (baselineAvg !== null) {
-      const currentPercent = hasAllErrors ? u.dim("N/A") : formatPercentage(avg);
-      const baselinePercent = formatPercentage(baselineAvg);
-      const { text: diffText, color: diffColor } = formatDiff(avg, baselineAvg);
-      const paddedBaseline = baselinePercent.padStart(7);
-      const paddedCurrent = hasAllErrors ? currentPercent : currentPercent.padStart(7);
-      const diffDisplay = hasAllErrors ? u.dim("all cases failed") : diffColor(diffText.padStart(8));
-      logger(
-        `\u2502  ${paddedName}  ${u.blueBright(paddedBaseline)} \u2192 ${hasAllErrors ? paddedCurrent : u.magentaBright(paddedCurrent)}  (${diffDisplay})`
+function loadConfigFile(path3) {
+  const abs = (0, import_node_path.resolve)(process.cwd(), path3);
+  try {
+    const contents = (0, import_node_fs.readFileSync)(abs, "utf8");
+    const parsed = JSON.parse(contents);
+    if (typeof parsed !== "object" || Array.isArray(parsed) || parsed === null) {
+      console.error(
+        `\u274C Flags config must be a JSON object, got ${Array.isArray(parsed) ? "array" : typeof parsed}`
       );
-    } else {
-      const currentPercent = hasAllErrors ? u.red("N/A (all cases failed)") : formatPercentage(avg);
-      logger(`\u2502   \u2022 ${paddedName}  ${currentPercent}`);
+      process.exit(1);
     }
+    return parsed;
+  } catch (err) {
+    console.error(`\u274C Could not read or parse flags config "${path3}": ${err.message}`);
+    process.exit(1);
   }
-  logger("\u251C\u2500");
-  if (suite.baseline) {
-    const baselineTimestamp = suite.baseline.runAt ? reporterDate(new Date(suite.baseline.runAt)) : "unknown time";
-    logger(
-      `\u2502  Baseline: ${suite.baseline.name}-${suite.baseline.version} ${u.gray(`(${baselineTimestamp})`)}`
-    );
-  } else {
-    logger(`\u2502  Baseline: ${u.gray("(none)")}`);
-  }
-  const hasConfigChanges = flagDiff.length > 0;
-  logger("\u2502  Config changes:", hasConfigChanges ? "" : u.gray("(none)"));
-  if (hasConfigChanges) {
-    for (const { flag, current, baseline, default: defaultVal } of flagDiff) {
-      logger(`\u2502   \u2022 ${flag}: ${current ?? "<not set>"}`);
-      if (defaultVal !== void 0) {
-        logger(`\u2502       ${u.gray(`default: ${defaultVal}`)}`);
+}
+function extractOverrides(argv) {
+  const cleanedArgv2 = [];
+  const overrides2 = {};
+  let configPath = null;
+  let hasCliFlags = false;
+  let configPathCount = 0;
+  for (let i = 0; i < argv.length; i++) {
+    const token = argv[i];
+    const configMatch = token.match(CONFIG_RE);
+    const flagMatch = token.match(FLAG_RE);
+    if (configMatch) {
+      configPathCount++;
+      if (configPathCount > 1) {
+        console.error("\u274C Only one --flags-config can be supplied.");
+        process.exit(1);
       }
-      if (suite.baseline) {
-        logger(`\u2502       ${u.gray(`baseline: ${baseline ?? "<not set>"}`)}`);
+      const value = configMatch[1];
+      const nextToken = argv.length > i + 1 ? argv[i + 1] : void 0;
+      ensureNoSpaceSeparatedSyntax("flags-config", value, nextToken, "config");
+      if (!value) {
+        console.error("\u274C --flags-config requires a file path");
+        console.error("\u{1F4A1} Use: --flags-config=path/to/config.json");
+        process.exit(1);
       }
+      configPath = value;
+    } else if (flagMatch) {
+      hasCliFlags = true;
+      const key = flagMatch[1];
+      const value = flagMatch[2];
+      const nextToken = argv.length > i + 1 ? argv[i + 1] : void 0;
+      ensureNoSpaceSeparatedSyntax(key, value, nextToken, "flag");
+      const finalValue = value === void 0 ? "true" : value;
+      overrides2[key] = coerceValue(finalValue);
+    } else {
+      cleanedArgv2.push(token);
     }
   }
-  if (suite.outOfScopeFlags && suite.outOfScopeFlags.length > 0) {
-    const pickedFlagsText = suite.configFlags && suite.configFlags.length > 0 ? suite.configFlags.map((f2) => `'${f2}'`).join(", ") : "none";
-    logger("\u2502");
-    logger(`\u2502  ${u.yellow("\u26A0 Out-of-scope flags")} ${u.gray(`(picked: ${pickedFlagsText})`)}:`);
-    for (const flag of suite.outOfScopeFlags) {
-      const lastStackTraceFrame = flag.stackTrace[0];
-      const lastStackTraceFnName = lastStackTraceFrame.split(" ").shift();
-      const lastStackTraceFile = lastStackTraceFrame.split("/").pop()?.slice(0, -1);
-      logger(
-        `\u2502   \u2022 ${flag.flagPath} ${u.gray(`at ${lastStackTraceFnName} (${lastStackTraceFile})`)}`
-      );
-    }
+  if (configPath && hasCliFlags) {
+    console.error("\u274C Cannot use both --flags-config and --flag.* arguments together.");
+    console.error("Choose one approach:");
+    console.error("  \u2022 Config file: --flags-config=my-flags.json");
+    console.error("  \u2022 CLI flags: --flag.temperature=0.9 --flag.model=gpt-4o");
+    process.exit(1);
   }
-  logger("\u2514\u2500");
-}
-function calculateScorerAverages(suite) {
-  const scorerTotals = {};
-  for (const caseData of suite.cases) {
-    for (const [scorerName, score] of Object.entries(caseData.scores)) {
-      if (!scorerTotals[scorerName]) {
-        scorerTotals[scorerName] = { sum: 0, count: 0 };
-      }
-      if (!score.metadata?.error) {
-        scorerTotals[scorerName].sum += score.score || 0;
-        scorerTotals[scorerName].count += 1;
-      }
-    }
+  if (configPath) {
+    const configOverrides = loadConfigFile(configPath);
+    return { cleanedArgv: cleanedArgv2, overrides: configOverrides };
   }
-  const averages = {};
-  for (const [scorerName, totals] of Object.entries(scorerTotals)) {
-    averages[scorerName] = totals.count > 0 ? totals.sum / totals.count : 0;
+  return { cleanedArgv: cleanedArgv2, overrides: overrides2 };
+}
+var import_zod6, import_node_fs, import_node_path, FLAG_RE, CONFIG_RE;
+var init_parse_flag_overrides = __esm({
+  "src/cli/utils/parse-flag-overrides.ts"() {
+    "use strict";
+    init_cjs_shims();
+    import_zod6 = require("zod");
+    init_format_zod_errors();
+    import_node_fs = require("fs");
+    import_node_path = require("path");
+    init_dot_path();
+    init_deep_partial_schema();
+    init_zod_internals();
+    FLAG_RE = /^--flag\.([^=]+)(?:=(.*))?$/;
+    CONFIG_RE = /^--flags-config(?:=(.*))?$/;
   }
-  return averages;
+});
+
+// src/cli/utils/glob-utils.ts
+function isGlob(str) {
+  return /[*?[\]{}!]/.test(str);
 }
-function calculateBaselineScorerAverage(baseline, scorerName) {
-  const scores = [];
-  for (const caseData of baseline.cases) {
-    if (caseData.scores[scorerName]) {
-      scores.push(caseData.scores[scorerName].value);
-    }
+var init_glob_utils = __esm({
+  "src/cli/utils/glob-utils.ts"() {
+    "use strict";
+    init_cjs_shims();
   }
-  if (scores.length === 0) return null;
-  const sum = scores.reduce((acc, val) => acc + val, 0);
-  return sum / scores.length;
+});
+
+// src/util/errors.ts
+function getCircularReplacer() {
+  const seen = /* @__PURE__ */ new WeakSet();
+  return (_k, v) => {
+    if (typeof v === "object" && v !== null) {
+      if (seen.has(v)) return "[Circular]";
+      seen.add(v);
+    }
+    return v;
+  };
 }
-function calculateFlagDiff(suite) {
-  if (!suite.configFlags || suite.configFlags.length === 0) {
-    return [];
+function safeJson(x) {
+  try {
+    return JSON.stringify(x, getCircularReplacer());
+  } catch {
+    return String(x);
   }
-  const diffs = [];
-  const currentConfig = suite.flagConfig || {};
-  const baselineConfig = suite.baseline?.flagConfig || {};
-  const defaultConfig = suite.defaultFlagConfig || {};
-  const currentFlat = flattenObject(currentConfig);
-  const baselineFlat = flattenObject(baselineConfig);
-  const defaultFlat = flattenObject(defaultConfig);
-  const allKeys = /* @__PURE__ */ new Set([
-    ...Object.keys(currentFlat),
-    ...Object.keys(baselineFlat),
-    ...Object.keys(defaultFlat)
-  ]);
-  for (const key of allKeys) {
-    const isInScope = suite.configFlags.some((pattern) => key.startsWith(pattern));
-    if (!isInScope) continue;
-    const currentValue = currentFlat[key];
-    const baselineValue = baselineFlat[key];
-    const defaultValue = defaultFlat[key];
-    const currentStr = currentValue !== void 0 ? JSON.stringify(currentValue) : void 0;
-    const baselineStr = baselineValue !== void 0 ? JSON.stringify(baselineValue) : void 0;
-    const defaultStr = defaultValue !== void 0 ? JSON.stringify(defaultValue) : void 0;
-    const diffFromBaseline = suite.baseline && currentStr !== baselineStr;
-    const diffFromDefault = currentStr !== defaultStr;
-    if (diffFromBaseline || diffFromDefault) {
-      diffs.push({
-        flag: key,
-        current: currentStr,
-        baseline: suite.baseline ? baselineStr : void 0,
-        default: defaultStr
-      });
+}
+function errorToString(err) {
+  try {
+    if (typeof err === "string") return err;
+    if (err instanceof Error) {
+      return err.stack ?? err.message;
+    }
+    if (typeof err === "object" && err !== null) {
+      const msg = err.message;
+      const json = safeJson(err);
+      return msg ? `${msg} (${json})` : json;
     }
+    return String(err);
+  } catch {
+    return "[unserializable error]";
   }
-  return diffs;
 }
-function printFinalReport({
-  suiteData,
-  config,
-  registrationStatus,
-  isDebug,
-  logger = console.log
-}) {
-  logger("");
-  logger(u.bgBlue(u.white(" FINAL EVALUATION REPORT ")));
-  logger("");
-  for (const suite of suiteData) {
-    const scorerAverages = calculateScorerAverages(suite);
-    const flagDiff = calculateFlagDiff(suite);
-    printSuiteBox({ suite, scorerAverages, calculateBaselineScorerAverage, flagDiff, logger });
-    logger("");
+var AxiomCLIError;
+var init_errors = __esm({
+  "src/util/errors.ts"() {
+    "use strict";
+    init_cjs_shims();
+    AxiomCLIError = class extends Error {
+      constructor(message) {
+        super(message);
+        this.name = "AxiomCLIError";
+      }
+    };
   }
-  const runId = suiteData[0]?.runId;
-  const orgId = suiteData[0]?.orgId;
-  const anyRegistered = registrationStatus.some((s2) => s2.registered);
-  const anyFailed = registrationStatus.some((s2) => !s2.registered);
-  if (anyRegistered && orgId && config?.consoleEndpointUrl) {
-    if (suiteData.length === 1) {
-      const suite = suiteData[0];
-      const baselineParam = suite.baseline?.traceId ? `?baselineId=${suite.baseline.traceId}` : "";
-      logger("View eval result:");
-      logger(
-        `${config.consoleEndpointUrl}/${orgId}/ai-engineering/evaluations/${suite.name}/${suite.version}${baselineParam}`
-      );
-    } else {
-      logger("View full report:");
-      logger(`${config.consoleEndpointUrl}/${orgId}/ai-engineering/evaluations?runId=${runId}`);
-    }
-  } else if (isDebug) {
-    logger(u.dim("Results not uploaded to Axiom (debug mode)"));
-  } else {
-    logger("Results not available in Axiom UI (registration failed)");
-  }
-  if (anyFailed) {
-    logger("");
-    for (const status of registrationStatus) {
-      if (!status.registered) {
-        logger(u.yellow(`\u26A0\uFE0F  Warning: Failed to register "${status.name}" with Axiom`));
-        if (status.error) {
-          logger(u.dim(`   Error: ${status.error}`));
-        }
-        logger(u.dim(`   Results for this evaluation will not be available in the Axiom UI.`));
-      }
-    }
-  }
-}
-
-// src/config/resolver.ts
-init_cjs_shims();
-var buildConsoleUrl = (urlString) => {
-  const url = new URL(urlString);
-  return `${url.protocol}//app.${url.host.split("api.").at(-1)}`;
-};
-function resolveAxiomConnection(config, consoleUrlOverride) {
-  const consoleEndpointUrl = consoleUrlOverride ?? buildConsoleUrl(config.eval.url);
-  return {
-    url: config.eval.url,
-    consoleEndpointUrl,
-    token: config.eval.token,
-    dataset: config.eval.dataset,
-    orgId: config.eval.orgId
-  };
-}
-
-// src/evals/reporter.ts
-var AxiomReporter = class {
-  constructor() {
-    __publicField(this, "startTime", 0);
-    __publicField(this, "start", 0);
-    __publicField(this, "_endOfRunConfigEnd");
-    __publicField(this, "_suiteData", []);
-    __publicField(this, "_printedFlagOverrides", false);
-    __publicField(this, "_config");
-  }
-  onTestRunStart() {
-    this.start = performance.now();
-    this.startTime = (/* @__PURE__ */ new Date()).getTime();
-    const config = getAxiomConfig();
-    if (config) {
-      this._config = resolveAxiomConnection(config, getConsoleUrl());
-    }
-  }
-  async onTestSuiteReady(_testSuite) {
-    const meta = _testSuite.meta();
-    if (_testSuite.state() === "skipped" || !meta?.evaluation) {
-      return;
-    }
-    if (!this._printedFlagOverrides) {
-      const defaultsFromConfigEnd = meta.evaluation.configEnd?.flags ?? {};
-      const overridesFromConfigEnd = meta.evaluation.configEnd?.overrides ?? {};
-      if (Object.keys(overridesFromConfigEnd).length > 0) {
-        printGlobalFlagOverrides(overridesFromConfigEnd, defaultsFromConfigEnd);
-      }
-      this._printedFlagOverrides = true;
-    }
-    if (meta.evaluation.configEnd && !this._endOfRunConfigEnd) {
-      this._endOfRunConfigEnd = meta.evaluation.configEnd;
-    }
-  }
-  onTestCaseReady(test) {
-    const meta = test.meta();
-    if (!meta.case) return;
-  }
-  async onTestSuiteResult(testSuite) {
-    const meta = testSuite.meta();
-    if (testSuite.state() === "skipped" || !meta?.evaluation) {
-      return;
-    }
-    if (meta.evaluation.configEnd && !this._endOfRunConfigEnd) {
-      this._endOfRunConfigEnd = meta.evaluation.configEnd;
-    }
-    const durationSeconds = Number((performance.now() - this.start) / 1e3).toFixed(2);
-    const cases = [];
-    for (const test of testSuite.children) {
-      if (test.type !== "test") continue;
-      const testMeta = test.meta();
-      if (!testMeta?.case) continue;
-      cases.push({
-        index: testMeta.case.index,
-        scores: testMeta.case.scores,
-        outOfScopeFlags: testMeta.case.outOfScopeFlags,
-        errors: testMeta.case.errors,
-        runtimeFlags: testMeta.case.runtimeFlags
-      });
-    }
-    const cwd = process.cwd();
-    const relativePath = testSuite.module.moduleId.replace(cwd, "").replace(/^\//, "");
-    let suiteBaseline = meta.evaluation.baseline;
-    let flagConfig = meta.evaluation.flagConfig;
-    if (meta.evaluation.configEnd) {
-      const defaults = meta.evaluation.configEnd.flags ?? {};
-      const overrides2 = meta.evaluation.configEnd.overrides ?? {};
-      const defaultsFlat = flattenObject(defaults);
-      const overridesFlat = flattenObject(overrides2);
-      flagConfig = dotNotationToNested({ ...defaultsFlat, ...overridesFlat });
-    }
-    const defaultFlagConfig = meta.evaluation.configEnd?.flags;
-    this._suiteData.push({
-      version: meta.evaluation.version,
-      name: meta.evaluation.name,
-      file: relativePath,
-      duration: durationSeconds + "s",
-      baseline: suiteBaseline || null,
-      configFlags: meta.evaluation.configFlags,
-      flagConfig,
-      defaultFlagConfig,
-      runId: meta.evaluation.runId,
-      orgId: meta.evaluation.orgId,
-      cases,
-      outOfScopeFlags: meta.evaluation.outOfScopeFlags,
-      registrationStatus: meta.evaluation.registrationStatus
-    });
-    printEvalNameAndFileName(testSuite, meta);
-    printBaselineNameAndVersion(meta);
-    printTestCaseCountStartDuration(testSuite, this.startTime, durationSeconds);
-    const matchedBaselineIndices = /* @__PURE__ */ new Set();
-    const baselineCasesByFingerprint = /* @__PURE__ */ new Map();
-    if (suiteBaseline) {
-      for (const c of suiteBaseline.cases) {
-        const fp = getCaseFingerprint(c.input, c.expected);
-        const cases2 = baselineCasesByFingerprint.get(fp) || [];
-        cases2.push(c);
-        baselineCasesByFingerprint.set(fp, cases2);
-      }
-    }
-    for (const test of testSuite.children) {
-      if (test.type !== "test") continue;
-      printCaseResult(test, baselineCasesByFingerprint, matchedBaselineIndices);
-    }
-    if (suiteBaseline) {
-      printOrphanedBaselineCases(suiteBaseline, matchedBaselineIndices);
-    }
-    console.log("");
-  }
-  async onTestRunEnd(_testModules, _errors, _reason) {
-    const shouldClear = !process.env.CI && process.stdout.isTTY !== false;
-    if (shouldClear) {
-      process.stdout.write("\x1B[2J\x1B[0f");
-    }
-    const registrationStatus = this._suiteData.map((suite) => ({
-      name: suite.name,
-      registered: suite.registrationStatus?.status === "success",
-      error: suite.registrationStatus?.status === "failed" ? suite.registrationStatus.error : void 0
-    }));
-    const isDebug = process.env.AXIOM_DEBUG === "true";
-    printFinalReport({
-      suiteData: this._suiteData,
-      config: this._config,
-      registrationStatus,
-      isDebug
-    });
-  }
-};
-
-// src/evals/instrument.ts
-init_cjs_shims();
-var import_sdk_trace_node = require("@opentelemetry/sdk-trace-node");
-var import_resources = require("@opentelemetry/resources");
-var import_exporter_trace_otlp_http = require("@opentelemetry/exporter-trace-otlp-http");
-var import_api10 = require("@opentelemetry/api");
-
-// src/util/errors.ts
-init_cjs_shims();
-var AxiomCLIError = class extends Error {
-  constructor(message) {
-    super(message);
-    this.name = "AxiomCLIError";
-  }
-};
-function getCircularReplacer() {
-  const seen = /* @__PURE__ */ new WeakSet();
-  return (_k, v) => {
-    if (typeof v === "object" && v !== null) {
-      if (seen.has(v)) return "[Circular]";
-      seen.add(v);
-    }
-    return v;
-  };
-}
-function safeJson(x) {
-  try {
-    return JSON.stringify(x, getCircularReplacer());
-  } catch {
-    return String(x);
-  }
-}
-function errorToString(err) {
-  try {
-    if (typeof err === "string") return err;
-    if (err instanceof Error) {
-      return err.stack ?? err.message;
-    }
-    if (typeof err === "object" && err !== null) {
-      const msg = err.message;
-      const json = safeJson(err);
-      return msg ? `${msg} (${json})` : json;
-    }
-    return String(err);
-  } catch {
-    return "[unserializable error]";
-  }
-}
-
-// src/config/loader.ts
-init_cjs_shims();
-var import_c12 = require("c12");
-var import_defu = require("defu");
-
-// src/config/index.ts
-init_cjs_shims();
-var import_zod6 = require("zod");
-
-// src/cli/auth/index.ts
-init_cjs_shims();
+});
 
 // src/cli/auth/types.ts
-init_cjs_shims();
+var init_types = __esm({
+  "src/cli/auth/types.ts"() {
+    "use strict";
+    init_cjs_shims();
+  }
+});
 
 // src/cli/auth/config.ts
-init_cjs_shims();
-var import_fs = require("fs");
-var import_path = __toESM(require("path"), 1);
-var import_os = __toESM(require("os"), 1);
-var CONFIG_FILENAME = "config.json";
-var CONFIG_DIR_NAME = "axiom";
 function getConfigDir() {
   const platform = process.platform;
   const homeDir = import_os.default.homedir();
@@ -1510,64 +1280,81 @@ function getActiveProfile(config) {
   if (!profile) return null;
   return profile;
 }
+var import_fs, import_path, import_os, CONFIG_FILENAME, CONFIG_DIR_NAME;
+var init_config = __esm({
+  "src/cli/auth/config.ts"() {
+    "use strict";
+    init_cjs_shims();
+    import_fs = require("fs");
+    import_path = __toESM(require("path"), 1);
+    import_os = __toESM(require("os"), 1);
+    CONFIG_FILENAME = "config.json";
+    CONFIG_DIR_NAME = "axiom";
+  }
+});
 
 // src/cli/auth/oauth.ts
-init_cjs_shims();
-var import_crypto = require("crypto");
-var OAUTH_CLIENT_ID = "264d906a404efc209b027f6595e6b616";
-var OAUTH_AUTH_PATH = "/oauth/authorize";
-var OAUTH_TOKEN_PATH = "/oauth/token";
-var OAuth = class {
-  constructor(oauthBaseUrl) {
-    this.oauthBaseUrl = oauthBaseUrl;
-  }
-  static generateCodeVerifier() {
-    return (0, import_crypto.randomBytes)(32).toString("base64url");
-  }
-  static generateCodeChallenge(verifier) {
-    return (0, import_crypto.createHash)("sha256").update(verifier).digest("base64url");
-  }
-  static generateState() {
-    return (0, import_crypto.randomBytes)(16).toString("hex");
-  }
-  buildAuthUrl(params) {
-    const url = new URL(OAUTH_AUTH_PATH, this.oauthBaseUrl);
-    url.searchParams.set("client_id", OAUTH_CLIENT_ID);
-    url.searchParams.set("redirect_uri", params.redirectUri);
-    url.searchParams.set("response_type", "code");
-    url.searchParams.set("state", params.state);
-    url.searchParams.set("code_challenge", params.codeChallenge);
-    url.searchParams.set("code_challenge_method", "S256");
-    url.searchParams.set("scope", "*");
-    return url.toString();
-  }
-  async exchangeCodeForToken(params) {
-    const tokenUrl = new URL(OAUTH_TOKEN_PATH, this.oauthBaseUrl);
-    const body = new URLSearchParams({
-      grant_type: "authorization_code",
-      client_id: OAUTH_CLIENT_ID,
-      code: params.code,
-      redirect_uri: params.redirectUri,
-      code_verifier: params.codeVerifier
-    });
-    const response = await fetch(tokenUrl.toString(), {
-      method: "POST",
-      headers: {
-        "Content-Type": "application/x-www-form-urlencoded"
-      },
-      body: body.toString()
-    });
-    if (!response.ok) {
-      const errorText = await response.text();
-      throw new Error(`Token exchange failed: ${response.status} ${errorText}`);
-    }
-    const data = await response.json();
-    return data.access_token;
+var import_crypto, OAUTH_CLIENT_ID, OAUTH_AUTH_PATH, OAUTH_TOKEN_PATH, OAuth;
+var init_oauth = __esm({
+  "src/cli/auth/oauth.ts"() {
+    "use strict";
+    init_cjs_shims();
+    import_crypto = require("crypto");
+    OAUTH_CLIENT_ID = "264d906a404efc209b027f6595e6b616";
+    OAUTH_AUTH_PATH = "/oauth/authorize";
+    OAUTH_TOKEN_PATH = "/oauth/token";
+    OAuth = class {
+      constructor(oauthBaseUrl) {
+        this.oauthBaseUrl = oauthBaseUrl;
+      }
+      static generateCodeVerifier() {
+        return (0, import_crypto.randomBytes)(32).toString("base64url");
+      }
+      static generateCodeChallenge(verifier) {
+        return (0, import_crypto.createHash)("sha256").update(verifier).digest("base64url");
+      }
+      static generateState() {
+        return (0, import_crypto.randomBytes)(16).toString("hex");
+      }
+      buildAuthUrl(params) {
+        const url = new URL(OAUTH_AUTH_PATH, this.oauthBaseUrl);
+        url.searchParams.set("client_id", OAUTH_CLIENT_ID);
+        url.searchParams.set("redirect_uri", params.redirectUri);
+        url.searchParams.set("response_type", "code");
+        url.searchParams.set("state", params.state);
+        url.searchParams.set("code_challenge", params.codeChallenge);
+        url.searchParams.set("code_challenge_method", "S256");
+        url.searchParams.set("scope", "*");
+        return url.toString();
+      }
+      async exchangeCodeForToken(params) {
+        const tokenUrl = new URL(OAUTH_TOKEN_PATH, this.oauthBaseUrl);
+        const body = new URLSearchParams({
+          grant_type: "authorization_code",
+          client_id: OAUTH_CLIENT_ID,
+          code: params.code,
+          redirect_uri: params.redirectUri,
+          code_verifier: params.codeVerifier
+        });
+        const response = await fetch(tokenUrl.toString(), {
+          method: "POST",
+          headers: {
+            "Content-Type": "application/x-www-form-urlencoded"
+          },
+          body: body.toString()
+        });
+        if (!response.ok) {
+          const errorText = await response.text();
+          throw new Error(`Token exchange failed: ${response.status} ${errorText}`);
+        }
+        const data = await response.json();
+        return data.access_token;
+      }
+    };
   }
-};
+});
 
 // src/cli/auth/api.ts
-init_cjs_shims();
 async function fetchOrganizations(token, apiBaseUrl) {
   const response = await fetch(`${apiBaseUrl}/v2/orgs`, {
     headers: {
@@ -1593,10 +1380,15 @@ async function verifyToken(token, orgId, apiBaseUrl) {
   });
   return response.ok;
 }
+var init_api = __esm({
+  "src/cli/auth/api.ts"() {
+    "use strict";
+    init_cjs_shims();
+    init_errors();
+  }
+});
 
 // src/cli/auth/callback-server.ts
-init_cjs_shims();
-var import_http = __toESM(require("http"), 1);
 function escapeHtml(text) {
   const map = {
     "&": "&amp;",
@@ -1608,9 +1400,6 @@ function escapeHtml(text) {
   };
   return text.replace(/[&<>"']/g, (m2) => map[m2] || m2);
 }
-var SVG_LOGO = `<svg width="124" height="24" viewBox="0 0 124 24" fill="none" xmlns="http://www.w3.org/2000/svg" class="logo">
-<path d="M42.9919 16.8116H36.3696L35.5537 19.1572C35.4209 19.539 34.9714 19.8513 34.5548 19.8513H30.9779C30.5614 19.8513 30.3459 19.5449 30.499 19.1703L36.9816 3.31644C37.1346 2.9419 37.6009 2.63546 38.0174 2.63546H41.3718C41.7883 2.63546 42.2541 2.94207 42.4067 3.3168L48.8634 19.17C49.016 19.5447 48.8 19.8513 48.3835 19.8513H44.8067C44.3901 19.8513 43.9406 19.539 43.8078 19.1572L42.9919 16.8116ZM41.8232 13.4223L39.6807 7.18148L37.5383 13.4223H41.8232ZM64.1105 19.8513C63.694 19.8513 63.1767 19.5694 62.9611 19.2247L59.9029 14.3369L56.8447 19.2247C56.6291 19.5694 56.1119 19.8513 55.6953 19.8513H51.3392C50.9227 19.8513 50.7754 19.5801 51.0119 19.2486L56.8978 11.0013L51.5315 3.24414C51.2999 2.90937 51.4513 2.63546 51.8679 2.63546H55.89C56.3066 2.63546 56.8268 2.91563 57.0461 3.25807L59.9029 7.71959L62.7343 3.25955C62.9522 2.9163 63.4713 2.63546 63.8879 2.63546H67.938C68.3545 2.63546 68.5048 2.90868 68.272 3.24261L62.8801 10.9743L68.7935 19.2489C69.0303 19.5802 68.8832 19.8513 68.4666 19.8513H64.1105ZM76.525 19.119C76.525 19.5218 76.1841 19.8513 75.7675 19.8513H72.5522C72.1356 19.8513 71.7947 19.5218 71.7947 19.119V3.36771C71.7947 2.96498 72.1356 2.63546 72.5522 2.63546H75.7675C76.1841 2.63546 76.525 2.96498 76.525 3.36771V19.119ZM79.3736 11.1896C79.3736 6.18625 83.2688 2.15134 89.2511 2.15134C95.2334 2.15134 99.1289 6.18625 99.1289 11.1896C99.1289 16.2199 95.2334 20.2548 89.2511 20.2548C83.2688 20.2548 79.3736 16.2199 79.3736 11.1896ZM94.2873 11.1896C94.2873 8.58038 92.3953 6.21307 89.2511 6.21307C86.1349 6.21307 84.2149 8.58038 84.2149 11.1896C84.2149 13.7989 86.1349 16.2199 89.2511 16.2199C92.3675 16.2199 94.2873 13.7989 94.2873 11.1896ZM119.318 19.8513C118.902 19.8513 118.517 19.5245 118.463 19.1251L117.225 9.89847L113.65 19.1652C113.505 19.5426 113.045 19.8513 112.628 19.8513H111.055C110.638 19.8513 110.179 19.5423 110.034 19.1645L106.457 9.81766L105.218 19.125C105.165 19.5245 104.781 19.8513 104.364 19.8513H101.344C100.927 19.8513 100.633 19.5249 100.69 19.1259L102.932 3.3609C102.988 2.96191 103.376 2.63546 103.792 2.63546H106.59C107.007 2.63546 107.469 2.94351 107.617 3.32002L111.827 14.041L116.064 3.31949C116.213 2.94327 116.676 2.63546 117.092 2.63546H119.89C120.307 2.63546 120.694 2.96191 120.751 3.3609L122.993 19.1259C123.05 19.5249 122.755 19.8513 122.339 19.8513H119.318ZM23.9616 15.6531L18.8054 6.97021C18.5689 6.57115 17.9863 6.24465 17.5106 6.24465H14.2915C13.5433 6.24465 13.2365 5.73171 13.6097 5.1048L15.375 2.13986C15.5151 1.90455 15.5148 1.61487 15.3743 1.37981C15.2337 1.14476 14.9741 1 14.6933 1H10.2025C9.72681 1 9.14291 1.32577 8.90491 1.72395L0.17865 16.3217C-0.0594434 16.7199 -0.0595348 17.3715 0.178285 17.7698L2.42362 21.5297C2.79777 22.1561 3.41129 22.1569 3.78699 21.5313L5.54143 18.6103C5.91722 17.9847 6.53065 17.9854 6.9048 18.6119L8.49538 21.2754C8.7332 21.6737 9.317 21.9995 9.79273 21.9995H20.1698C20.6455 21.9995 21.2293 21.6737 21.4672 21.2754L23.959 17.1028C24.1968 16.7045 24.198 16.0521 23.9616 15.6531ZM16.9981 15.2352C17.3699 15.8629 17.0619 16.3765 16.3136 16.3765H8.24192C7.49372 16.3765 7.1876 15.864 7.56175 15.2375L11.6007 8.47417C11.9748 7.84772 12.5869 7.84774 12.9611 8.47421L16.9981 15.2352Z" fill="#121224"/>
-</svg>`;
 function renderCallbackPage(error) {
   const errorClass = error ? ' class="error"' : "";
   const errorMessage = error ? escapeHtml(error) : "";
@@ -1757,10 +1546,19 @@ async function waitForCallback(server, expectedState) {
     });
   });
 }
+var import_http, SVG_LOGO;
+var init_callback_server = __esm({
+  "src/cli/auth/callback-server.ts"() {
+    "use strict";
+    init_cjs_shims();
+    import_http = __toESM(require("http"), 1);
+    SVG_LOGO = `<svg width="124" height="24" viewBox="0 0 124 24" fill="none" xmlns="http://www.w3.org/2000/svg" class="logo">
+<path d="M42.9919 16.8116H36.3696L35.5537 19.1572C35.4209 19.539 34.9714 19.8513 34.5548 19.8513H30.9779C30.5614 19.8513 30.3459 19.5449 30.499 19.1703L36.9816 3.31644C37.1346 2.9419 37.6009 2.63546 38.0174 2.63546H41.3718C41.7883 2.63546 42.2541 2.94207 42.4067 3.3168L48.8634 19.17C49.016 19.5447 48.8 19.8513 48.3835 19.8513H44.8067C44.3901 19.8513 43.9406 19.539 43.8078 19.1572L42.9919 16.8116ZM41.8232 13.4223L39.6807 7.18148L37.5383 13.4223H41.8232ZM64.1105 19.8513C63.694 19.8513 63.1767 19.5694 62.9611 19.2247L59.9029 14.3369L56.8447 19.2247C56.6291 19.5694 56.1119 19.8513 55.6953 19.8513H51.3392C50.9227 19.8513 50.7754 19.5801 51.0119 19.2486L56.8978 11.0013L51.5315 3.24414C51.2999 2.90937 51.4513 2.63546 51.8679 2.63546H55.89C56.3066 2.63546 56.8268 2.91563 57.0461 3.25807L59.9029 7.71959L62.7343 3.25955C62.9522 2.9163 63.4713 2.63546 63.8879 2.63546H67.938C68.3545 2.63546 68.5048 2.90868 68.272 3.24261L62.8801 10.9743L68.7935 19.2489C69.0303 19.5802 68.8832 19.8513 68.4666 19.8513H64.1105ZM76.525 19.119C76.525 19.5218 76.1841 19.8513 75.7675 19.8513H72.5522C72.1356 19.8513 71.7947 19.5218 71.7947 19.119V3.36771C71.7947 2.96498 72.1356 2.63546 72.5522 2.63546H75.7675C76.1841 2.63546 76.525 2.96498 76.525 3.36771V19.119ZM79.3736 11.1896C79.3736 6.18625 83.2688 2.15134 89.2511 2.15134C95.2334 2.15134 99.1289 6.18625 99.1289 11.1896C99.1289 16.2199 95.2334 20.2548 89.2511 20.2548C83.2688 20.2548 79.3736 16.2199 79.3736 11.1896ZM94.2873 11.1896C94.2873 8.58038 92.3953 6.21307 89.2511 6.21307C86.1349 6.21307 84.2149 8.58038 84.2149 11.1896C84.2149 13.7989 86.1349 16.2199 89.2511 16.2199C92.3675 16.2199 94.2873 13.7989 94.2873 11.1896ZM119.318 19.8513C118.902 19.8513 118.517 19.5245 118.463 19.1251L117.225 9.89847L113.65 19.1652C113.505 19.5426 113.045 19.8513 112.628 19.8513H111.055C110.638 19.8513 110.179 19.5423 110.034 19.1645L106.457 9.81766L105.218 19.125C105.165 19.5245 104.781 19.8513 104.364 19.8513H101.344C100.927 19.8513 100.633 19.5249 100.69 19.1259L102.932 3.3609C102.988 2.96191 103.376 2.63546 103.792 2.63546H106.59C107.007 2.63546 107.469 2.94351 107.617 3.32002L111.827 14.041L116.064 3.31949C116.213 2.94327 116.676 2.63546 117.092 2.63546H119.89C120.307 2.63546 120.694 2.96191 120.751 3.3609L122.993 19.1259C123.05 19.5249 122.755 19.8513 122.339 19.8513H119.318ZM23.9616 15.6531L18.8054 6.97021C18.5689 6.57115 17.9863 6.24465 17.5106 6.24465H14.2915C13.5433 6.24465 13.2365 5.73171 13.6097 5.1048L15.375 2.13986C15.5151 1.90455 15.5148 1.61487 15.3743 1.37981C15.2337 1.14476 14.9741 1 14.6933 1H10.2025C9.72681 1 9.14291 1.32577 8.90491 1.72395L0.17865 16.3217C-0.0594434 16.7199 -0.0595348 17.3715 0.178285 17.7698L2.42362 21.5297C2.79777 22.1561 3.41129 22.1569 3.78699 21.5313L5.54143 18.6103C5.91722 17.9847 6.53065 17.9854 6.9048 18.6119L8.49538 21.2754C8.7332 21.6737 9.317 21.9995 9.79273 21.9995H20.1698C20.6455 21.9995 21.2293 21.6737 21.4672 21.2754L23.959 17.1028C24.1968 16.7045 24.198 16.0521 23.9616 15.6531ZM16.9981 15.2352C17.3699 15.8629 17.0619 16.3765 16.3136 16.3765H8.24192C7.49372 16.3765 7.1876 15.864 7.56175 15.2375L11.6007 8.47417C11.9748 7.84772 12.5869 7.84774 12.9611 8.47421L16.9981 15.2352Z" fill="#121224"/>
+</svg>`;
+  }
+});
 
 // src/cli/auth/global-auth.ts
-init_cjs_shims();
-var authContext = null;
 function getAuthContext() {
   return authContext;
 }
@@ -1776,9 +1574,31 @@ async function setupGlobalAuth() {
   }
   return authContext;
 }
+var authContext;
+var init_global_auth = __esm({
+  "src/cli/auth/global-auth.ts"() {
+    "use strict";
+    init_cjs_shims();
+    init_config();
+    authContext = null;
+  }
+});
+
+// src/cli/auth/index.ts
+var init_auth = __esm({
+  "src/cli/auth/index.ts"() {
+    "use strict";
+    init_cjs_shims();
+    init_types();
+    init_config();
+    init_oauth();
+    init_api();
+    init_callback_server();
+    init_global_auth();
+  }
+});
 
 // src/config/index.ts
-var DEFAULT_EVAL_INCLUDE = ["**/*.eval.{ts,js,mts,mjs,cts,cjs}"];
 function createPartialDefaults() {
   let token;
   let url;
@@ -1841,6 +1661,17 @@ function validateConfig(config) {
   }
   return config;
 }
+var import_zod7, DEFAULT_EVAL_INCLUDE;
+var init_config2 = __esm({
+  "src/config/index.ts"() {
+    "use strict";
+    init_cjs_shims();
+    import_zod7 = require("zod");
+    init_errors();
+    init_auth();
+    DEFAULT_EVAL_INCLUDE = ["**/*.eval.{ts,js,mts,mjs,cts,cjs}"];
+  }
+});
 
 // src/config/loader.ts
 function customMerger(target, source) {
@@ -1880,446 +1711,967 @@ async function loadConfig(cwd = process.cwd()) {
     throw new AxiomCLIError(`Failed to load config file: ${errorToString(error)}`);
   }
 }
+var import_c12, import_defu;
+var init_loader = __esm({
+  "src/config/loader.ts"() {
+    "use strict";
+    init_cjs_shims();
+    import_c12 = require("c12");
+    import_defu = require("defu");
+    init_config2();
+    init_errors();
+  }
+});
 
-// src/evals/instrument.ts
-var axiomProvider;
-var axiomTracer;
-var userProvider;
-var initializationPromise = null;
-var initialized = false;
-async function resolveInstrumentationHook(config) {
-  if (config.eval.instrumentation) {
-    return config.eval.instrumentation;
+// ../../node_modules/.pnpm/tinyrainbow@2.0.0/node_modules/tinyrainbow/dist/chunk-BVHSVHOK.js
+function a(n) {
+  return String(n);
+}
+function C(n = false) {
+  let e = typeof process != "undefined" ? process : void 0, i = (e == null ? void 0 : e.env) || {}, g = (e == null ? void 0 : e.argv) || [];
+  return !("NO_COLOR" in i || g.includes("--no-color")) && ("FORCE_COLOR" in i || g.includes("--color") || (e == null ? void 0 : e.platform) === "win32" || n && i.TERM !== "dumb" || "CI" in i) || typeof window != "undefined" && !!window.chrome;
+}
+function p(n = false) {
+  let e = C(n), i = (r2, t, c, o) => {
+    let l = "", s2 = 0;
+    do
+      l += r2.substring(s2, o) + c, s2 = o + t.length, o = r2.indexOf(t, s2);
+    while (~o);
+    return l + r2.substring(s2);
+  }, g = (r2, t, c = r2) => {
+    let o = (l) => {
+      let s2 = String(l), b = s2.indexOf(t, r2.length);
+      return ~b ? r2 + i(s2, t, c, b) + t : r2 + s2 + t;
+    };
+    return o.open = r2, o.close = t, o;
+  }, u2 = {
+    isColorSupported: e
+  }, d = (r2) => `\x1B[${r2}m`;
+  for (let [r2, t] of h)
+    u2[r2] = e ? g(
+      d(t[0]),
+      d(t[1]),
+      t[2]
+    ) : a;
+  return u2;
+}
+var f, h;
+var init_chunk_BVHSVHOK = __esm({
+  "../../node_modules/.pnpm/tinyrainbow@2.0.0/node_modules/tinyrainbow/dist/chunk-BVHSVHOK.js"() {
+    "use strict";
+    init_cjs_shims();
+    f = {
+      reset: [0, 0],
+      bold: [1, 22, "\x1B[22m\x1B[1m"],
+      dim: [2, 22, "\x1B[22m\x1B[2m"],
+      italic: [3, 23],
+      underline: [4, 24],
+      inverse: [7, 27],
+      hidden: [8, 28],
+      strikethrough: [9, 29],
+      black: [30, 39],
+      red: [31, 39],
+      green: [32, 39],
+      yellow: [33, 39],
+      blue: [34, 39],
+      magenta: [35, 39],
+      cyan: [36, 39],
+      white: [37, 39],
+      gray: [90, 39],
+      bgBlack: [40, 49],
+      bgRed: [41, 49],
+      bgGreen: [42, 49],
+      bgYellow: [43, 49],
+      bgBlue: [44, 49],
+      bgMagenta: [45, 49],
+      bgCyan: [46, 49],
+      bgWhite: [47, 49],
+      blackBright: [90, 39],
+      redBright: [91, 39],
+      greenBright: [92, 39],
+      yellowBright: [93, 39],
+      blueBright: [94, 39],
+      magentaBright: [95, 39],
+      cyanBright: [96, 39],
+      whiteBright: [97, 39],
+      bgBlackBright: [100, 49],
+      bgRedBright: [101, 49],
+      bgGreenBright: [102, 49],
+      bgYellowBright: [103, 49],
+      bgBlueBright: [104, 49],
+      bgMagentaBright: [105, 49],
+      bgCyanBright: [106, 49],
+      bgWhiteBright: [107, 49]
+    };
+    h = Object.entries(f);
+    a.open = "";
+    a.close = "";
   }
-  try {
-    const { config: loadedConfig } = await loadConfig(process.cwd());
-    return loadedConfig.eval.instrumentation ?? null;
-  } catch (error) {
-    throw new AxiomCLIError(
-      `Failed to reload instrumentation from config: ${errorToString(error)}`
-    );
+});
+
+// ../../node_modules/.pnpm/tinyrainbow@2.0.0/node_modules/tinyrainbow/dist/node.js
+var import_tty, r, u;
+var init_node = __esm({
+  "../../node_modules/.pnpm/tinyrainbow@2.0.0/node_modules/tinyrainbow/dist/node.js"() {
+    "use strict";
+    init_cjs_shims();
+    init_chunk_BVHSVHOK();
+    import_tty = require("tty");
+    r = process.env.FORCE_TTY !== void 0 || (0, import_tty.isatty)(1);
+    u = p(r);
+  }
+});
+
+// src/evals/reporter.console-utils.ts
+function formatPercentage(value) {
+  if (!Number.isFinite(value)) {
+    return "N/A";
   }
+  return Number(value * 100).toFixed(2) + "%";
 }
-async function runInstrumentationHook(hook, options) {
+function formatDiff(current, baseline) {
+  if (!Number.isFinite(current) || !Number.isFinite(baseline)) {
+    return { text: "N/A", color: u.dim };
+  }
+  const diff = current - baseline;
+  const diffText = (diff >= 0 ? "+" : "") + formatPercentage(diff);
+  const color = diff > 0 ? u.green : diff < 0 ? u.red : u.dim;
+  return { text: diffText, color };
+}
+function truncate(str, max) {
+  return str.length > max ? str.slice(0, max) + "\u2026" : str;
+}
+function stringify(value) {
   try {
-    return await hook(options);
-  } catch (error) {
-    throw new AxiomCLIError(`Failed to execute instrumentation hook: ${errorToString(error)}`);
+    if (typeof value === "string") return value;
+    return JSON.stringify(value);
+  } catch {
+    return String(value);
   }
 }
-function setupEvalProvider(connection) {
-  const headers = {
-    "X-Axiom-Dataset": connection.dataset,
-    ...connection.orgId ? { "X-AXIOM-ORG-ID": connection.orgId } : {}
-  };
-  if (connection.token) {
-    headers.Authorization = `Bearer ${connection.token}`;
+function getCaseFingerprint(input, expected) {
+  const inputStr = typeof input === "string" ? input : JSON.stringify(input);
+  const expectedStr = typeof expected === "string" ? expected : JSON.stringify(expected);
+  return JSON.stringify({ input: inputStr, expected: expectedStr });
+}
+function printEvalNameAndFileName(testSuite, meta, logger = console.log) {
+  const cwd = process.cwd();
+  logger(
+    " ",
+    u.bgCyan(u.black(` ${testSuite.project.name} `)),
+    u.bgBlue(u.black(` ${meta.evaluation.name}-${meta.evaluation.version} `)),
+    u.dim(`(${testSuite.children.size} cases)`)
+  );
+  logger(" ", u.dim(testSuite.module.moduleId.replace(cwd, "")));
+}
+function printBaselineNameAndVersion(testMeta, logger = console.log) {
+  if (testMeta.evaluation.baseline) {
+    logger(
+      " ",
+      " baseline ",
+      u.bgMagenta(
+        u.black(` ${testMeta.evaluation.baseline.name}-${testMeta.evaluation.baseline.version} `)
+      )
+    );
+  } else {
+    logger(" ", u.bgWhite(u.blackBright(" baseline: ")), "none");
   }
-  const collectorOptions = {
-    url: `${connection.url}/v1/traces`,
-    headers,
-    concurrencyLimit: 10
-  };
-  const exporter = new import_exporter_trace_otlp_http.OTLPTraceExporter(collectorOptions);
-  const processor = new import_sdk_trace_node.BatchSpanProcessor(exporter, {
-    maxQueueSize: 2048,
-    maxExportBatchSize: 512,
-    scheduledDelayMillis: 5e3,
-    exportTimeoutMillis: 3e4
-  });
-  axiomProvider = new import_sdk_trace_node.NodeTracerProvider({
-    resource: (0, import_resources.resourceFromAttributes)({
-      ["service.name"]: "axiom",
-      ["service.version"]: "0.34.1"
-    }),
-    spanProcessors: [processor]
+  logger("");
+}
+function printRuntimeFlags(testMeta, logger = console.log) {
+  if (testMeta.case.runtimeFlags && Object.keys(testMeta.case.runtimeFlags).length > 0) {
+    const entries = Object.entries(testMeta.case.runtimeFlags);
+    logger("   ", u.dim("runtime flags"));
+    for (const [k, v] of entries) {
+      switch (v.kind) {
+        case "replaced": {
+          const valText = truncate(stringify(v.value), 80);
+          const defText = truncate(stringify(v.default), 80);
+          logger("     ", `${k}: ${valText} (default: ${defText})`);
+          break;
+        }
+        case "introduced": {
+          const valText = truncate(stringify(v.value), 80);
+          logger("     ", `${k}: ${valText} (no default)`);
+          break;
+        }
+      }
+    }
+  }
+}
+function printTestCaseCountStartDuration(testSuite, startTime, duration, logger = console.log) {
+  logger(" ");
+  logger(" ", u.dim("Cases"), testSuite.children.size);
+  logger(" ", u.dim("Start at"), new Date(startTime).toTimeString());
+  logger(" ", u.dim("Duration"), `${duration}s`);
+}
+function printTestCaseSuccessOrFailed(testMeta, ok, logger = console.log) {
+  const index = testMeta.case.index;
+  if (ok) {
+    logger(" ", u.yellow(` \u2714 case ${index}:`));
+  } else {
+    logger(" ", u.red(` \u2716 case ${index}: failed`));
+    for (const e of testMeta.case.errors ?? []) {
+      logger("", e.message);
+    }
+  }
+}
+function printTestCaseScores(testMeta, baselineCase, logger = console.log) {
+  const scores = testMeta.case.scores;
+  const keys = Object.keys(scores);
+  if (keys.length === 0) {
+    return;
+  }
+  const maxNameLength = Math.max(...keys.map((k) => k.length));
+  keys.forEach((k) => {
+    const scoreData = scores[k];
+    const hasError = scoreData.metadata?.error;
+    const v = scoreData.score ? scoreData.score : 0;
+    const rawCurrent = hasError ? "N/A" : formatPercentage(v);
+    const paddedCurrent = rawCurrent.padStart(7);
+    const coloredCurrent = hasError ? u.dim(paddedCurrent) : u.magentaBright(paddedCurrent);
+    const paddedName = k.padEnd(maxNameLength);
+    if (baselineCase?.scores[k]) {
+      const baselineScoreValue = baselineCase.scores[k].value;
+      const rawBaseline = formatPercentage(baselineScoreValue);
+      const paddedBaseline = rawBaseline.padStart(7);
+      const coloredBaseline = u.blueBright(paddedBaseline);
+      const { text: diffText, color: diffColor } = formatDiff(v, baselineScoreValue);
+      const paddedDiff = diffText.padStart(8);
+      logger(
+        `    ${paddedName}  ${coloredBaseline} \u2192 ${coloredCurrent}  ${hasError ? u.dim("(scorer not run)") : u.dim("(") + diffColor(paddedDiff) + u.dim(")")}`
+      );
+    } else {
+      logger(`    ${paddedName}  ${coloredCurrent} ${hasError ? u.dim("(scorer not run)") : ""}`);
+    }
   });
-  axiomTracer = axiomProvider.getTracer("axiom", "0.34.1");
 }
-async function initInstrumentation(config) {
-  if (initialized) {
+function printOutOfScopeFlags(testMeta, logger = console.log) {
+  if (testMeta.case.outOfScopeFlags && testMeta.case.outOfScopeFlags.length > 0) {
+    const pickedFlagsText = testMeta.case.pickedFlags ? `(picked: ${testMeta.case.pickedFlags.map((f2) => `'${f2}'`).join(", ")})` : "(none)";
+    logger("   ", u.yellow(`\u26A0 Out-of-scope flags: ${pickedFlagsText}`));
+    testMeta.case.outOfScopeFlags.forEach((flag) => {
+      const timeStr = new Date(flag.accessedAt).toLocaleTimeString();
+      logger("     ", `${flag.flagPath} (at ${timeStr})`);
+      if (flag.stackTrace && flag.stackTrace.length > 0) {
+        flag.stackTrace.forEach((frame, i) => {
+          const prefix = i === flag.stackTrace.length - 1 ? " \u2514\u2500" : " \u251C\u2500";
+          logger("     ", u.dim(`${prefix} ${frame}`));
+        });
+      }
+    });
+  }
+}
+function printCaseResult(test, baselineCasesByFingerprint, matchedIndices, logger = console.log) {
+  const ok = test.ok();
+  const testMeta = test.meta();
+  if (!testMeta?.case) {
     return;
   }
-  if (initializationPromise) {
-    await initializationPromise;
+  printTestCaseSuccessOrFailed(testMeta, ok, logger);
+  const fingerprint = getCaseFingerprint(testMeta.case.input, testMeta.case.expected);
+  const baselineCases = baselineCasesByFingerprint.get(fingerprint);
+  const baselineCase = baselineCases?.shift();
+  if (baselineCase) {
+    matchedIndices.add(baselineCase.index);
+  }
+  printTestCaseScores(testMeta, baselineCase, logger);
+  printRuntimeFlags(testMeta, logger);
+  printOutOfScopeFlags(testMeta, logger);
+}
+function printOrphanedBaselineCases(baseline, matchedIndices, logger = console.log) {
+  const orphanedCases = baseline.cases.filter((c) => !matchedIndices.has(c.index));
+  if (orphanedCases.length === 0) {
     return;
   }
-  initializationPromise = (async () => {
-    if (!config.enabled) {
-      axiomTracer = import_api10.trace.getTracer("axiom", "0.34.1");
-      initialized = true;
-      return;
-    }
-    const connection = resolveAxiomConnection(config.config);
-    const hook = await resolveInstrumentationHook(config.config);
-    let hookResult = void 0;
-    if (hook) {
-      config.config.eval.instrumentation = hook;
-      hookResult = await runInstrumentationHook(hook, {
-        dataset: connection.dataset,
-        token: connection.token,
-        url: connection.url,
-        orgId: connection.orgId
+  logger("");
+  logger(" ", u.yellow("Orphaned baseline cases:"));
+  for (const orphanedCase of orphanedCases) {
+    logger(
+      " ",
+      u.dim(
+        `case ${orphanedCase.index}: ${truncate(orphanedCase.input, 50)} (score: ${truncate(
+          JSON.stringify(orphanedCase.scores),
+          50
+        )})`
+      )
+    );
+    const keys = Object.keys(orphanedCase.scores);
+    if (keys.length > 0) {
+      const maxNameLength = Math.max(...keys.map((k) => k.length));
+      keys.forEach((k) => {
+        const scoreData = orphanedCase.scores[k];
+        const rawScore = formatPercentage(scoreData.value);
+        const paddedName = k.padEnd(maxNameLength);
+        const paddedScore = rawScore.padStart(7);
+        logger(`    ${paddedName}  ${u.blueBright(paddedScore)}`);
       });
-      userProvider = hookResult?.provider ?? userProvider;
     }
-    setupEvalProvider(connection);
-    if (!hook) {
-      axiomProvider?.register();
-      if (axiomTracer) {
-        initAxiomAI({ tracer: axiomTracer });
+  }
+}
+function printGlobalFlagOverrides(overrides2, defaults, logger = console.log) {
+  if (Object.keys(overrides2).length === 0) {
+    logger("");
+    logger(u.dim("Flag overrides: (none)"));
+    logger("");
+    return;
+  }
+  logger("");
+  logger("Flag overrides:");
+  for (const [key, value] of Object.entries(overrides2)) {
+    const defaultValue = defaults[key];
+    const valueStr = JSON.stringify(value);
+    const defaultStr = defaultValue !== void 0 ? JSON.stringify(defaultValue) : "none";
+    logger(`  \u2022 ${key}: ${valueStr} ${u.dim(`(default: ${defaultStr})`)}`);
+  }
+  logger("");
+}
+function printSuiteBox({
+  suite,
+  scorerAverages,
+  calculateBaselineScorerAverage: calculateBaselineScorerAverage2,
+  flagDiff,
+  logger = console.log
+}) {
+  const filename = suite.file.split("/").pop();
+  logger("\u250C\u2500");
+  logger(`\u2502  ${u.blue(suite.name)} ${u.gray(`(${filename})`)}`);
+  logger("\u251C\u2500");
+  const scorerNames = Object.keys(scorerAverages);
+  const maxNameLength = Math.max(...scorerNames.map((name) => name.length));
+  const allCasesErrored = (scorerName) => {
+    return suite.cases.every((caseData) => caseData.scores[scorerName]?.metadata?.error);
+  };
+  for (const scorerName of scorerNames) {
+    const avg = scorerAverages[scorerName];
+    const paddedName = scorerName.padEnd(maxNameLength);
+    const hasAllErrors = allCasesErrored(scorerName);
+    const baselineAvg = suite.baseline ? calculateBaselineScorerAverage2(suite.baseline, scorerName) : null;
+    if (baselineAvg !== null) {
+      const currentPercent = hasAllErrors ? u.dim("N/A") : formatPercentage(avg);
+      const baselinePercent = formatPercentage(baselineAvg);
+      const { text: diffText, color: diffColor } = formatDiff(avg, baselineAvg);
+      const paddedBaseline = baselinePercent.padStart(7);
+      const paddedCurrent = hasAllErrors ? currentPercent : currentPercent.padStart(7);
+      const diffDisplay = hasAllErrors ? u.dim("all cases failed") : diffColor(diffText.padStart(8));
+      logger(
+        `\u2502  ${paddedName}  ${u.blueBright(paddedBaseline)} \u2192 ${hasAllErrors ? paddedCurrent : u.magentaBright(paddedCurrent)}  (${diffDisplay})`
+      );
+    } else {
+      const currentPercent = hasAllErrors ? u.red("N/A (all cases failed)") : formatPercentage(avg);
+      logger(`\u2502   \u2022 ${paddedName}  ${currentPercent}`);
+    }
+  }
+  logger("\u251C\u2500");
+  if (suite.baseline) {
+    const baselineTimestamp = suite.baseline.runAt ? reporterDate(new Date(suite.baseline.runAt)) : "unknown time";
+    logger(
+      `\u2502  Baseline: ${suite.baseline.name}-${suite.baseline.version} ${u.gray(`(${baselineTimestamp})`)}`
+    );
+  } else {
+    logger(`\u2502  Baseline: ${u.gray("(none)")}`);
+  }
+  const hasConfigChanges = flagDiff.length > 0;
+  logger("\u2502  Config changes:", hasConfigChanges ? "" : u.gray("(none)"));
+  if (hasConfigChanges) {
+    for (const { flag, current, baseline, default: defaultVal } of flagDiff) {
+      logger(`\u2502   \u2022 ${flag}: ${current ?? "<not set>"}`);
+      if (defaultVal !== void 0) {
+        logger(`\u2502       ${u.gray(`default: ${defaultVal}`)}`);
+      }
+      if (suite.baseline) {
+        logger(`\u2502       ${u.gray(`baseline: ${baseline ?? "<not set>"}`)}`);
       }
     }
-    initialized = true;
-  })();
-  try {
-    await initializationPromise;
-  } finally {
-    initializationPromise = null;
   }
+  if (suite.outOfScopeFlags && suite.outOfScopeFlags.length > 0) {
+    const pickedFlagsText = suite.configFlags && suite.configFlags.length > 0 ? suite.configFlags.map((f2) => `'${f2}'`).join(", ") : "none";
+    logger("\u2502");
+    logger(`\u2502  ${u.yellow("\u26A0 Out-of-scope flags")} ${u.gray(`(picked: ${pickedFlagsText})`)}:`);
+    for (const flag of suite.outOfScopeFlags) {
+      const lastStackTraceFrame = flag.stackTrace[0];
+      const lastStackTraceFnName = lastStackTraceFrame.split(" ").shift();
+      const lastStackTraceFile = lastStackTraceFrame.split("/").pop()?.slice(0, -1);
+      logger(
+        `\u2502   \u2022 ${flag.flagPath} ${u.gray(`at ${lastStackTraceFnName} (${lastStackTraceFile})`)}`
+      );
+    }
+  }
+  logger("\u2514\u2500");
 }
-var flush = async () => {
-  if (initializationPromise) {
-    await initializationPromise;
+function calculateScorerAverages(suite) {
+  const scorerTotals = {};
+  for (const caseData of suite.cases) {
+    for (const [scorerName, score] of Object.entries(caseData.scores)) {
+      if (!scorerTotals[scorerName]) {
+        scorerTotals[scorerName] = { sum: 0, count: 0 };
+      }
+      if (!score.metadata?.error) {
+        scorerTotals[scorerName].sum += score.score || 0;
+        scorerTotals[scorerName].count += 1;
+      }
+    }
   }
-  const tasks = [];
-  if (axiomProvider) {
-    tasks.push(axiomProvider.forceFlush());
+  const averages = {};
+  for (const [scorerName, totals] of Object.entries(scorerTotals)) {
+    averages[scorerName] = totals.count > 0 ? totals.sum / totals.count : 0;
   }
-  const candidateProviders = /* @__PURE__ */ new Set();
-  if (userProvider) {
-    candidateProviders.add(userProvider);
+  return averages;
+}
+function calculateBaselineScorerAverage(baseline, scorerName) {
+  const scores = [];
+  for (const caseData of baseline.cases) {
+    if (caseData.scores[scorerName]) {
+      scores.push(caseData.scores[scorerName].value);
+    }
   }
-  const globalProvider = import_api10.trace.getTracerProvider();
-  if (globalProvider) {
-    candidateProviders.add(globalProvider);
+  if (scores.length === 0) return null;
+  const sum = scores.reduce((acc, val) => acc + val, 0);
+  return sum / scores.length;
+}
+function calculateFlagDiff(suite) {
+  if (!suite.configFlags || suite.configFlags.length === 0) {
+    return [];
   }
-  for (const provider of candidateProviders) {
-    const flushFn = provider.forceFlush;
-    if (typeof flushFn === "function") {
-      tasks.push(
-        flushFn.call(provider).catch((error) => {
-          console.warn("[AxiomAI] Failed to flush tracer provider:", errorToString(error));
-        })
-      );
+  const diffs = [];
+  const currentConfig = suite.flagConfig || {};
+  const baselineConfig = suite.baseline?.flagConfig || {};
+  const defaultConfig = suite.defaultFlagConfig || {};
+  const currentFlat = flattenObject(currentConfig);
+  const baselineFlat = flattenObject(baselineConfig);
+  const defaultFlat = flattenObject(defaultConfig);
+  const allKeys = /* @__PURE__ */ new Set([
+    ...Object.keys(currentFlat),
+    ...Object.keys(baselineFlat),
+    ...Object.keys(defaultFlat)
+  ]);
+  for (const key of allKeys) {
+    const isInScope = suite.configFlags.some((pattern) => key.startsWith(pattern));
+    if (!isInScope) continue;
+    const currentValue = currentFlat[key];
+    const baselineValue = baselineFlat[key];
+    const defaultValue = defaultFlat[key];
+    const currentStr = currentValue !== void 0 ? JSON.stringify(currentValue) : void 0;
+    const baselineStr = baselineValue !== void 0 ? JSON.stringify(baselineValue) : void 0;
+    const defaultStr = defaultValue !== void 0 ? JSON.stringify(defaultValue) : void 0;
+    const diffFromBaseline = suite.baseline && currentStr !== baselineStr;
+    const diffFromDefault = currentStr !== defaultStr;
+    if (diffFromBaseline || diffFromDefault) {
+      diffs.push({
+        flag: key,
+        current: currentStr,
+        baseline: suite.baseline ? baselineStr : void 0,
+        default: defaultStr
+      });
     }
   }
-  if (tasks.length > 0) {
-    await Promise.all(tasks);
-  }
-};
-
-// src/evals/run-vitest.ts
-var printCollectedEvals = (result, rootDir) => {
-  if (!result.testModules || result.testModules.length === 0) {
-    console.log(u.yellow("\nNo evaluations found\n"));
-    return;
+  return diffs;
+}
+function printFinalReport({
+  suiteData,
+  config,
+  registrationStatus,
+  isDebug,
+  logger = console.log
+}) {
+  logger("");
+  logger(u.bgBlue(u.white(" FINAL EVALUATION REPORT ")));
+  logger("");
+  for (const suite of suiteData) {
+    const scorerAverages = calculateScorerAverages(suite);
+    const flagDiff = calculateFlagDiff(suite);
+    printSuiteBox({ suite, scorerAverages, calculateBaselineScorerAverage, flagDiff, logger });
+    logger("");
   }
-  console.log(u.bold("\nFound evaluations:\n"));
-  let totalEvals = 0;
-  let totalCases = 0;
-  for (const module2 of result.testModules) {
-    const relativePath = import_node_path2.default.relative(rootDir, module2.moduleId);
-    for (const suite of module2.children.suites()) {
-      totalEvals++;
-      const caseCount = suite.children.size;
-      totalCases += caseCount;
-      console.log(u.green(`\u2713 ${suite.name} (${caseCount} cases)`));
-      console.log(u.dim(`  ${relativePath}`));
-      console.log("");
+  const runId = suiteData[0]?.runId;
+  const orgId = suiteData[0]?.orgId;
+  const anyRegistered = registrationStatus.some((s2) => s2.registered);
+  const anyFailed = registrationStatus.some((s2) => !s2.registered);
+  if (anyRegistered && orgId && config?.consoleEndpointUrl) {
+    if (suiteData.length === 1) {
+      const suite = suiteData[0];
+      const baselineParam = suite.baseline?.traceId ? `?baselineId=${suite.baseline.traceId}` : "";
+      logger("View eval result:");
+      logger(
+        `${config.consoleEndpointUrl}/${orgId}/ai-engineering/evaluations/${suite.name}/${suite.version}${baselineParam}`
+      );
+    } else {
+      logger("View full report:");
+      logger(`${config.consoleEndpointUrl}/${orgId}/ai-engineering/evaluations?runId=${runId}`);
     }
+  } else if (isDebug) {
+    logger(u.dim("Results not uploaded to Axiom (debug mode)"));
+  } else {
+    logger("Results not available in Axiom UI (registration failed)");
   }
-  console.log(u.bold(`Total: ${totalEvals} evaluations, ${totalCases} test cases
-`));
-};
-var runVitest = async (dir, opts) => {
-  setAxiomConfig(opts.config);
-  await initInstrumentation({
-    enabled: !opts.debug && !opts.list,
-    config: opts.config
-  });
-  const providedConfig = {
-    ...opts.config,
-    eval: {
-      ...opts.config.eval,
-      // These can't be serialized, so we need to remove them
-      instrumentation: null,
-      flagSchema: null
-    }
-  };
-  if (opts.debug) {
-    console.log(u.bgWhite(u.blackBright(" Debug mode enabled ")));
-  }
-  const tmpDir = (0, import_node_path.join)((0, import_node_os.tmpdir)(), "axiom-eval", opts.runId);
-  (0, import_node_fs.mkdirSync)(tmpDir, { recursive: true });
-  const nameRegistryFile = (0, import_node_path.join)(tmpDir, "names.jsonl");
-  const abortFile = (0, import_node_path.join)(tmpDir, "abort.txt");
-  (0, import_node_fs.writeFileSync)(nameRegistryFile, "", "utf8");
-  if ((0, import_node_fs.existsSync)(abortFile)) {
-    (0, import_node_fs.unlinkSync)(abortFile);
-  }
-  process.env.AXIOM_NAME_REGISTRY_FILE = nameRegistryFile;
-  process.env.AXIOM_ABORT_FILE = abortFile;
-  if (opts.list) {
-    console.log(u.bgWhite(u.blackBright(" List mode ")));
-  }
-  const vi = await (0, import_node.createVitest)(
-    "test",
-    {
-      root: dir ? dir : process.cwd(),
-      mode: "test",
-      include: opts.include,
-      exclude: opts.exclude,
-      testNamePattern: opts.testNamePattern,
-      reporters: ["verbose", new AxiomReporter()],
-      environment: "node",
-      browser: void 0,
-      watch: opts.watch,
-      setupFiles: [],
-      // ignore user vitest.config.ts etc
-      name: "axiom:eval",
-      printConsoleTrace: true,
-      silent: false,
-      disableConsoleIntercept: true,
-      testTimeout: opts.config?.eval?.timeoutMs || 6e4,
-      globals: true,
-      runner: (0, import_node_path.resolve)(__dirname, "evals", "custom-runner.js"),
-      provide: {
-        baseline: opts.baseline,
-        debug: opts.debug,
-        list: opts.list,
-        overrides: opts.overrides,
-        axiomConfig: providedConfig,
-        runId: opts.runId,
-        consoleUrl: opts.consoleUrl
+  if (anyFailed) {
+    logger("");
+    for (const status of registrationStatus) {
+      if (!status.registered) {
+        logger(u.yellow(`\u26A0\uFE0F  Warning: Failed to register "${status.name}" with Axiom`));
+        if (status.error) {
+          logger(u.dim(`   Error: ${status.error}`));
+        }
+        logger(u.dim(`   Results for this evaluation will not be available in the Axiom UI.`));
       }
-    },
-    {
-      plugins: [(0, import_vite_tsconfig_paths.default)({ root: dir || process.cwd() })]
     }
-  );
-  if (opts.list) {
-    const result = await vi.collect();
-    printCollectedEvals(result, dir || process.cwd());
-    await vi.close();
-    process.exit(0);
-  }
-  await vi.start();
-  if ((0, import_node_fs.existsSync)(abortFile)) {
-    const message = (0, import_node_fs.readFileSync)(abortFile, "utf8");
-    console.error("\n" + message);
-    await vi.close();
-    process.exit(1);
   }
-  const dispose = (0, import_node.registerConsoleShortcuts)(vi, process.stdin, process.stdout);
-  if (!vi.shouldKeepServer()) {
-    dispose();
-    await flush();
-    await vi.close();
-    process.exit(0);
+}
+var reporterDate;
+var init_reporter_console_utils = __esm({
+  "src/evals/reporter.console-utils.ts"() {
+    "use strict";
+    init_cjs_shims();
+    init_node();
+    init_dot_path();
+    reporterDate = (d) => {
+      const date = d.toISOString().slice(0, 10);
+      const hours = d.getUTCHours().toString().padStart(2, "0");
+      const minutes = d.getUTCMinutes().toString().padStart(2, "0");
+      return `${date}, ${hours}:${minutes} UTC`;
+    };
   }
-  await flush();
-};
-
-// src/cli/commands/eval.command.ts
-var import_node_fs3 = require("fs");
-
-// src/cli/utils/eval-context-runner.ts
-init_cjs_shims();
+});
 
-// src/context.ts
-init_cjs_shims();
-function overrideFlags(partial) {
-  const current = getEvalContext();
-  if (!current) {
-    if (process.env.NODE_ENV !== "test") {
-      console.warn("overrideFlags called outside of evaluation context");
-    }
-    return;
-  }
-  const overlayContext = {
-    ...current,
-    flags: { ...current.flags, ...partial },
-    // Merge for backwards compatibility
-    parent: current,
-    overrides: { ...partial }
+// src/config/resolver.ts
+function resolveAxiomConnection(config, consoleUrlOverride) {
+  const consoleEndpointUrl = consoleUrlOverride ?? buildConsoleUrl(config.eval.url);
+  return {
+    url: config.eval.url,
+    consoleEndpointUrl,
+    token: config.eval.token,
+    dataset: config.eval.dataset,
+    orgId: config.eval.orgId
   };
-  const currentCtx = EVAL_CONTEXT.get();
-  if (currentCtx) {
-    Object.assign(currentCtx, overlayContext);
-  }
-  for (const [key, value] of Object.entries(partial)) {
-    putOnSpan("flag", key, value);
-  }
-}
-
-// src/cli/utils/eval-context-runner.ts
-async function runEvalWithContext(overrides2, runFn) {
-  setGlobalFlagOverrides(overrides2);
-  return withEvalContext({ initialFlags: overrides2 }, async () => {
-    if (Object.keys(overrides2).length > 0) {
-      overrideFlags(overrides2);
-    }
-    return runFn();
-  });
-}
-
-// src/cli/utils/parse-flag-overrides.ts
-init_cjs_shims();
-var import_zod7 = require("zod");
-var import_node_fs2 = require("fs");
-var import_node_path3 = require("path");
-var FLAG_RE = /^--flag\.([^=]+)(?:=(.*))?$/;
-var CONFIG_RE = /^--flags-config(?:=(.*))?$/;
-function ensureNoSpaceSeparatedSyntax(flagName, value, nextToken, flagType) {
-  if (value === void 0 && nextToken !== void 0) {
-    if (flagType === "flag" && !nextToken.startsWith("-") && nextToken !== "true" && nextToken !== "false") {
-      console.error(`\u274C Invalid syntax: --flag.${flagName} ${nextToken}`);
-      console.error(`\u{1F4A1} Use: --flag.${flagName}=${nextToken}`);
-      process.exit(1);
-    } else if (flagType === "config" && !nextToken.startsWith("-")) {
-      console.error(`\u274C Invalid syntax: --flags-config ${nextToken}`);
-      console.error(`\u{1F4A1} Use: --flags-config=${nextToken}`);
-      process.exit(1);
-    }
-  }
 }
-function collectFlagValidationErrors(overrides2, flagSchema) {
-  if (!flagSchema || Object.keys(overrides2).length === 0) {
-    return { success: true, errors: [] };
+var buildConsoleUrl;
+var init_resolver = __esm({
+  "src/config/resolver.ts"() {
+    "use strict";
+    init_cjs_shims();
+    buildConsoleUrl = (urlString) => {
+      const url = new URL(urlString);
+      return `${url.protocol}//app.${url.host.split("api.").at(-1)}`;
+    };
   }
-  assertZodV4(flagSchema, "flagSchema");
-  const schema = flagSchema;
-  const errors = [];
-  for (const dotPath of Object.keys(overrides2)) {
-    const segments = parsePath(dotPath);
-    if (!isValidPath(schema, segments)) {
-      errors.push({ type: "invalid_path", path: dotPath });
-    }
+});
+
+// src/evals/reporter.ts
+var AxiomReporter;
+var init_reporter = __esm({
+  "src/evals/reporter.ts"() {
+    "use strict";
+    init_cjs_shims();
+    init_storage();
+    init_reporter_console_utils();
+    init_resolver();
+    init_eval_command();
+    init_dot_path();
+    AxiomReporter = class {
+      constructor() {
+        __publicField(this, "startTime", 0);
+        __publicField(this, "start", 0);
+        __publicField(this, "_endOfRunConfigEnd");
+        __publicField(this, "_suiteData", []);
+        __publicField(this, "_printedFlagOverrides", false);
+        __publicField(this, "_config");
+      }
+      onTestRunStart() {
+        this.start = performance.now();
+        this.startTime = (/* @__PURE__ */ new Date()).getTime();
+        const config = getAxiomConfig();
+        if (config) {
+          this._config = resolveAxiomConnection(config, getConsoleUrl());
+        }
+      }
+      async onTestSuiteReady(_testSuite) {
+        const meta = _testSuite.meta();
+        if (_testSuite.state() === "skipped" || !meta?.evaluation) {
+          return;
+        }
+        if (!this._printedFlagOverrides) {
+          const defaultsFromConfigEnd = meta.evaluation.configEnd?.flags ?? {};
+          const overridesFromConfigEnd = meta.evaluation.configEnd?.overrides ?? {};
+          if (Object.keys(overridesFromConfigEnd).length > 0) {
+            printGlobalFlagOverrides(overridesFromConfigEnd, defaultsFromConfigEnd);
+          }
+          this._printedFlagOverrides = true;
+        }
+        if (meta.evaluation.configEnd && !this._endOfRunConfigEnd) {
+          this._endOfRunConfigEnd = meta.evaluation.configEnd;
+        }
+      }
+      onTestCaseReady(test) {
+        const meta = test.meta();
+        if (!meta.case) return;
+      }
+      async onTestSuiteResult(testSuite) {
+        const meta = testSuite.meta();
+        if (testSuite.state() === "skipped" || !meta?.evaluation) {
+          return;
+        }
+        if (meta.evaluation.configEnd && !this._endOfRunConfigEnd) {
+          this._endOfRunConfigEnd = meta.evaluation.configEnd;
+        }
+        const durationSeconds = Number((performance.now() - this.start) / 1e3).toFixed(2);
+        const cases = [];
+        for (const test of testSuite.children) {
+          if (test.type !== "test") continue;
+          const testMeta = test.meta();
+          if (!testMeta?.case) continue;
+          cases.push({
+            index: testMeta.case.index,
+            scores: testMeta.case.scores,
+            outOfScopeFlags: testMeta.case.outOfScopeFlags,
+            errors: testMeta.case.errors,
+            runtimeFlags: testMeta.case.runtimeFlags
+          });
+        }
+        const cwd = process.cwd();
+        const relativePath = testSuite.module.moduleId.replace(cwd, "").replace(/^\//, "");
+        let suiteBaseline = meta.evaluation.baseline;
+        let flagConfig = meta.evaluation.flagConfig;
+        if (meta.evaluation.configEnd) {
+          const defaults = meta.evaluation.configEnd.flags ?? {};
+          const overrides2 = meta.evaluation.configEnd.overrides ?? {};
+          const defaultsFlat = flattenObject(defaults);
+          const overridesFlat = flattenObject(overrides2);
+          flagConfig = dotNotationToNested({ ...defaultsFlat, ...overridesFlat });
+        }
+        const defaultFlagConfig = meta.evaluation.configEnd?.flags;
+        this._suiteData.push({
+          version: meta.evaluation.version,
+          name: meta.evaluation.name,
+          file: relativePath,
+          duration: durationSeconds + "s",
+          baseline: suiteBaseline || null,
+          configFlags: meta.evaluation.configFlags,
+          flagConfig,
+          defaultFlagConfig,
+          runId: meta.evaluation.runId,
+          orgId: meta.evaluation.orgId,
+          cases,
+          outOfScopeFlags: meta.evaluation.outOfScopeFlags,
+          registrationStatus: meta.evaluation.registrationStatus
+        });
+        printEvalNameAndFileName(testSuite, meta);
+        printBaselineNameAndVersion(meta);
+        printTestCaseCountStartDuration(testSuite, this.startTime, durationSeconds);
+        const matchedBaselineIndices = /* @__PURE__ */ new Set();
+        const baselineCasesByFingerprint = /* @__PURE__ */ new Map();
+        if (suiteBaseline) {
+          for (const c of suiteBaseline.cases) {
+            const fp = getCaseFingerprint(c.input, c.expected);
+            const cases2 = baselineCasesByFingerprint.get(fp) || [];
+            cases2.push(c);
+            baselineCasesByFingerprint.set(fp, cases2);
+          }
+        }
+        for (const test of testSuite.children) {
+          if (test.type !== "test") continue;
+          printCaseResult(test, baselineCasesByFingerprint, matchedBaselineIndices);
+        }
+        if (suiteBaseline) {
+          printOrphanedBaselineCases(suiteBaseline, matchedBaselineIndices);
+        }
+        console.log("");
+      }
+      async onTestRunEnd(_testModules, _errors, _reason) {
+        const shouldClear = !process.env.CI && process.stdout.isTTY !== false;
+        if (shouldClear) {
+          process.stdout.write("\x1B[2J\x1B[0f");
+        }
+        const registrationStatus = this._suiteData.map((suite) => ({
+          name: suite.name,
+          registered: suite.registrationStatus?.status === "success",
+          error: suite.registrationStatus?.status === "failed" ? suite.registrationStatus.error : void 0
+        }));
+        const isDebug = process.env.AXIOM_DEBUG === "true";
+        printFinalReport({
+          suiteData: this._suiteData,
+          config: this._config,
+          registrationStatus,
+          isDebug
+        });
+      }
+    };
   }
-  if (errors.length > 0) {
-    return { success: false, errors };
+});
+
+// src/evals/instrument.ts
+async function resolveInstrumentationHook(config) {
+  if (config.eval.instrumentation) {
+    return config.eval.instrumentation;
   }
-  const nestedObject = dotNotationToNested(overrides2);
-  const deepPartialSchema = makeDeepPartial(schema);
-  const result = deepPartialSchema.safeParse(nestedObject);
-  if (!result.success) {
-    errors.push({ type: "invalid_value", zodError: result.error });
+  try {
+    const { config: loadedConfig } = await loadConfig(process.cwd());
+    return loadedConfig.eval.instrumentation ?? null;
+  } catch (error) {
+    throw new AxiomCLIError(
+      `Failed to reload instrumentation from config: ${errorToString(error)}`
+    );
   }
-  return { success: errors.length === 0, errors };
 }
-function printFlagValidationErrorsAndExit(errors) {
-  console.error("\u274C Invalid CLI flags:");
-  for (const error of errors) {
-    if (error.type === "invalid_path") {
-      console.error(`  \u2022 flag '${error.path}': Invalid flag path`);
-    } else {
-      console.error(formatZodErrors(error.zodError));
-      const examples = generateFlagExamples(error.zodError);
-      if (examples.length > 0) {
-        console.error("\n\u{1F4A1} Valid examples:");
-        examples.forEach((example) => console.error(`  ${example}`));
-      }
-    }
+async function runInstrumentationHook(hook, options) {
+  try {
+    return await hook(options);
+  } catch (error) {
+    throw new AxiomCLIError(`Failed to execute instrumentation hook: ${errorToString(error)}`);
   }
-  process.exit(1);
 }
-function validateFlagOverrides(overrides2, flagSchema) {
-  const result = collectFlagValidationErrors(overrides2, flagSchema);
-  if (!result.success) {
-    printFlagValidationErrorsAndExit(result.errors);
+function setupEvalProvider(connection) {
+  const headers = {
+    "X-Axiom-Dataset": connection.dataset,
+    ...connection.orgId ? { "X-AXIOM-ORG-ID": connection.orgId } : {}
+  };
+  if (connection.token) {
+    headers.Authorization = `Bearer ${connection.token}`;
   }
+  const collectorOptions = {
+    url: `${connection.url}/v1/traces`,
+    headers,
+    concurrencyLimit: 10
+  };
+  const exporter = new import_exporter_trace_otlp_http.OTLPTraceExporter(collectorOptions);
+  const processor = new import_sdk_trace_node.BatchSpanProcessor(exporter, {
+    maxQueueSize: 2048,
+    maxExportBatchSize: 512,
+    scheduledDelayMillis: 5e3,
+    exportTimeoutMillis: 3e4
+  });
+  axiomProvider = new import_sdk_trace_node.NodeTracerProvider({
+    resource: (0, import_resources.resourceFromAttributes)({
+      ["service.name"]: "axiom",
+      ["service.version"]: "0.35.0"
+    }),
+    spanProcessors: [processor]
+  });
+  axiomTracer = axiomProvider.getTracer("axiom", "0.35.0");
 }
-function coerceValue(raw) {
-  if (raw === "true") return true;
-  if (raw === "false") return false;
-  const num = Number(raw);
-  if (!Number.isNaN(num) && raw.trim() === num.toString()) {
-    return num;
+async function initInstrumentation(config) {
+  if (initialized) {
+    return;
   }
-  try {
-    return JSON.parse(raw);
-  } catch {
-    return raw;
+  if (initializationPromise) {
+    await initializationPromise;
+    return;
   }
-}
-function loadConfigFile(path3) {
-  const abs = (0, import_node_path3.resolve)(process.cwd(), path3);
-  try {
-    const contents = (0, import_node_fs2.readFileSync)(abs, "utf8");
-    const parsed = JSON.parse(contents);
-    if (typeof parsed !== "object" || Array.isArray(parsed) || parsed === null) {
-      console.error(
-        `\u274C Flags config must be a JSON object, got ${Array.isArray(parsed) ? "array" : typeof parsed}`
-      );
-      process.exit(1);
+  initializationPromise = (async () => {
+    if (!config.enabled) {
+      axiomTracer = import_api10.trace.getTracer("axiom", "0.35.0");
+      initialized = true;
+      return;
     }
-    return parsed;
-  } catch (err) {
-    console.error(`\u274C Could not read or parse flags config "${path3}": ${err.message}`);
-    process.exit(1);
+    const connection = resolveAxiomConnection(config.config);
+    const hook = await resolveInstrumentationHook(config.config);
+    let hookResult = void 0;
+    if (hook) {
+      config.config.eval.instrumentation = hook;
+      hookResult = await runInstrumentationHook(hook, {
+        dataset: connection.dataset,
+        token: connection.token,
+        url: connection.url,
+        orgId: connection.orgId
+      });
+      userProvider = hookResult?.provider ?? userProvider;
+    }
+    setupEvalProvider(connection);
+    if (!hook) {
+      axiomProvider?.register();
+      if (axiomTracer) {
+        initAxiomAI({ tracer: axiomTracer });
+      }
+    }
+    initialized = true;
+  })();
+  try {
+    await initializationPromise;
+  } finally {
+    initializationPromise = null;
   }
 }
-function extractOverrides(argv) {
-  const cleanedArgv2 = [];
-  const overrides2 = {};
-  let configPath = null;
-  let hasCliFlags = false;
-  let configPathCount = 0;
-  for (let i = 0; i < argv.length; i++) {
-    const token = argv[i];
-    const configMatch = token.match(CONFIG_RE);
-    const flagMatch = token.match(FLAG_RE);
-    if (configMatch) {
-      configPathCount++;
-      if (configPathCount > 1) {
-        console.error("\u274C Only one --flags-config can be supplied.");
-        process.exit(1);
+var import_sdk_trace_node, import_resources, import_exporter_trace_otlp_http, import_api10, axiomProvider, axiomTracer, userProvider, initializationPromise, initialized, flush;
+var init_instrument = __esm({
+  "src/evals/instrument.ts"() {
+    "use strict";
+    init_cjs_shims();
+    import_sdk_trace_node = require("@opentelemetry/sdk-trace-node");
+    import_resources = require("@opentelemetry/resources");
+    import_exporter_trace_otlp_http = require("@opentelemetry/exporter-trace-otlp-http");
+    import_api10 = require("@opentelemetry/api");
+    init_initAxiomAI();
+    init_resolver();
+    init_errors();
+    init_loader();
+    initializationPromise = null;
+    initialized = false;
+    flush = async () => {
+      if (initializationPromise) {
+        await initializationPromise;
       }
-      const value = configMatch[1];
-      const nextToken = argv.length > i + 1 ? argv[i + 1] : void 0;
-      ensureNoSpaceSeparatedSyntax("flags-config", value, nextToken, "config");
-      if (!value) {
-        console.error("\u274C --flags-config requires a file path");
-        console.error("\u{1F4A1} Use: --flags-config=path/to/config.json");
-        process.exit(1);
+      const tasks = [];
+      if (axiomProvider) {
+        tasks.push(axiomProvider.forceFlush());
       }
-      configPath = value;
-    } else if (flagMatch) {
-      hasCliFlags = true;
-      const key = flagMatch[1];
-      const value = flagMatch[2];
-      const nextToken = argv.length > i + 1 ? argv[i + 1] : void 0;
-      ensureNoSpaceSeparatedSyntax(key, value, nextToken, "flag");
-      const finalValue = value === void 0 ? "true" : value;
-      overrides2[key] = coerceValue(finalValue);
-    } else {
-      cleanedArgv2.push(token);
-    }
-  }
-  if (configPath && hasCliFlags) {
-    console.error("\u274C Cannot use both --flags-config and --flag.* arguments together.");
-    console.error("Choose one approach:");
-    console.error("  \u2022 Config file: --flags-config=my-flags.json");
-    console.error("  \u2022 CLI flags: --flag.temperature=0.9 --flag.model=gpt-4o");
-    process.exit(1);
-  }
-  if (configPath) {
-    const configOverrides = loadConfigFile(configPath);
-    return { cleanedArgv: cleanedArgv2, overrides: configOverrides };
+      const candidateProviders = /* @__PURE__ */ new Set();
+      if (userProvider) {
+        candidateProviders.add(userProvider);
+      }
+      const globalProvider = import_api10.trace.getTracerProvider();
+      if (globalProvider) {
+        candidateProviders.add(globalProvider);
+      }
+      for (const provider of candidateProviders) {
+        const flushFn = provider.forceFlush;
+        if (typeof flushFn === "function") {
+          tasks.push(
+            flushFn.call(provider).catch((error) => {
+              console.warn("[AxiomAI] Failed to flush tracer provider:", errorToString(error));
+            })
+          );
+        }
+      }
+      if (tasks.length > 0) {
+        await Promise.all(tasks);
+      }
+    };
   }
-  return { cleanedArgv: cleanedArgv2, overrides: overrides2 };
-}
+});
 
-// src/cli/utils/glob-utils.ts
-init_cjs_shims();
-function isGlob(str) {
-  return /[*?[\]{}!]/.test(str);
-}
+// src/evals/run-vitest.ts
+var run_vitest_exports = {};
+__export(run_vitest_exports, {
+  runVitest: () => runVitest
+});
+var import_node_path2, import_node_fs2, import_node_os, import_node_path3, import_vite_tsconfig_paths, import_node, printCollectedEvals, runVitest;
+var init_run_vitest = __esm({
+  "src/evals/run-vitest.ts"() {
+    "use strict";
+    init_cjs_shims();
+    init_node();
+    import_node_path2 = require("path");
+    import_node_fs2 = require("fs");
+    import_node_os = require("os");
+    import_node_path3 = __toESM(require("path"), 1);
+    import_vite_tsconfig_paths = __toESM(require("vite-tsconfig-paths"), 1);
+    import_node = require("vitest/node");
+    init_reporter();
+    init_instrument();
+    init_storage();
+    printCollectedEvals = (result, rootDir) => {
+      if (!result.testModules || result.testModules.length === 0) {
+        console.log(u.yellow("\nNo evaluations found\n"));
+        return;
+      }
+      console.log(u.bold("\nFound evaluations:\n"));
+      let totalEvals = 0;
+      let totalCases = 0;
+      for (const module2 of result.testModules) {
+        const relativePath = import_node_path3.default.relative(rootDir, module2.moduleId);
+        for (const suite of module2.children.suites()) {
+          totalEvals++;
+          const caseCount = suite.children.size;
+          totalCases += caseCount;
+          console.log(u.green(`\u2713 ${suite.name} (${caseCount} cases)`));
+          console.log(u.dim(`  ${relativePath}`));
+          console.log("");
+        }
+      }
+      console.log(u.bold(`Total: ${totalEvals} evaluations, ${totalCases} test cases
+`));
+    };
+    runVitest = async (dir, opts) => {
+      setAxiomConfig(opts.config);
+      await initInstrumentation({
+        enabled: !opts.debug && !opts.list,
+        config: opts.config
+      });
+      const providedConfig = {
+        ...opts.config,
+        eval: {
+          ...opts.config.eval,
+          // These can't be serialized, so we need to remove them
+          instrumentation: null,
+          flagSchema: null
+        }
+      };
+      if (opts.debug) {
+        console.log(u.bgWhite(u.blackBright(" Debug mode enabled ")));
+      }
+      const tmpDir = (0, import_node_path2.join)((0, import_node_os.tmpdir)(), "axiom-eval", opts.runId);
+      (0, import_node_fs2.mkdirSync)(tmpDir, { recursive: true });
+      const nameRegistryFile = (0, import_node_path2.join)(tmpDir, "names.jsonl");
+      const abortFile = (0, import_node_path2.join)(tmpDir, "abort.txt");
+      (0, import_node_fs2.writeFileSync)(nameRegistryFile, "", "utf8");
+      if ((0, import_node_fs2.existsSync)(abortFile)) {
+        (0, import_node_fs2.unlinkSync)(abortFile);
+      }
+      process.env.AXIOM_NAME_REGISTRY_FILE = nameRegistryFile;
+      process.env.AXIOM_ABORT_FILE = abortFile;
+      if (opts.list) {
+        console.log(u.bgWhite(u.blackBright(" List mode ")));
+      }
+      const vi = await (0, import_node.createVitest)(
+        "test",
+        {
+          root: dir ? dir : process.cwd(),
+          mode: "test",
+          include: opts.include,
+          exclude: opts.exclude,
+          testNamePattern: opts.testNamePattern,
+          reporters: ["verbose", new AxiomReporter()],
+          environment: "node",
+          browser: void 0,
+          watch: opts.watch,
+          setupFiles: [],
+          // ignore user vitest.config.ts etc
+          name: "axiom:eval",
+          printConsoleTrace: true,
+          silent: false,
+          disableConsoleIntercept: true,
+          testTimeout: opts.config?.eval?.timeoutMs || 6e4,
+          globals: true,
+          runner: (0, import_node_path2.resolve)(__dirname, "evals", "custom-runner.js"),
+          provide: {
+            baseline: opts.baseline,
+            debug: opts.debug,
+            list: opts.list,
+            overrides: opts.overrides,
+            axiomConfig: providedConfig,
+            runId: opts.runId,
+            consoleUrl: opts.consoleUrl
+          }
+        },
+        {
+          plugins: [(0, import_vite_tsconfig_paths.default)({ root: dir || process.cwd() })]
+        }
+      );
+      if (opts.list) {
+        const result = await vi.collect();
+        printCollectedEvals(result, dir || process.cwd());
+        await vi.close();
+        process.exit(0);
+      }
+      await vi.start();
+      if ((0, import_node_fs2.existsSync)(abortFile)) {
+        const message = (0, import_node_fs2.readFileSync)(abortFile, "utf8");
+        console.error("\n" + message);
+        await vi.close();
+        process.exit(1);
+      }
+      const dispose = (0, import_node.registerConsoleShortcuts)(vi, process.stdin, process.stdout);
+      if (!vi.shouldKeepServer()) {
+        dispose();
+        await flush();
+        await vi.close();
+        process.exit(0);
+      }
+      await flush();
+    };
+  }
+});
 
 // src/cli/commands/eval.command.ts
-var createRunId = (0, import_nanoid.customAlphabet)("0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ", 10);
-var consoleUrl;
 function getConsoleUrl() {
   return consoleUrl;
 }
@@ -2344,93 +2696,136 @@ function getDefaultOrgId(value) {
   const authContext2 = getAuthContext();
   return authContext2?.orgId ?? process.env.AXIOM_ORG_ID;
 }
-var loadEvalCommand = (program2, flagOverrides = {}) => {
-  return program2.addCommand(
-    new import_commander.Command("eval").description("run evals locally").addArgument(
-      new import_commander.Argument("[target]", "file, directory, glob pattern, or eval name").default(
-        ".",
-        "any *.eval.ts file in current directory"
-      )
-    ).option("-w, --watch true", "keep server running and watch for changes", false).option("-t, --token <TOKEN>", "axiom token", getDefaultToken).option("-d, --dataset <DATASET>", "axiom dataset name", process.env.AXIOM_DATASET).option("-u, --url <AXIOM URL>", "axiom url", getDefaultUrl).option("-o, --org-id <ORG ID>", "axiom organization id", getDefaultOrgId).option("-b, --baseline <BASELINE ID>", "id of baseline evaluation to compare against").option("--debug", "run locally without any network operations", false).option("--list", "list evaluations and test cases without running them", false).addOption(new import_commander.Option("-c, --console-url <URL>", "console url override").hideHelp()).action(async (target, options) => {
-      try {
-        if (options.debug) {
-          process.env.AXIOM_DEBUG = "true";
-        }
-        let include = [];
-        let exclude;
-        let testNamePattern;
-        const isGlobPattern = isGlob(target);
-        const { config: loadedConfig } = await loadConfig(".");
-        validateFlagOverrides(flagOverrides, loadedConfig.eval.flagSchema);
-        const config = {
-          ...loadedConfig,
-          eval: {
-            ...loadedConfig.eval,
-            ...options.token && { token: options.token },
-            ...options.url && { url: options.url },
-            ...options.dataset && { dataset: options.dataset },
-            ...options.orgId && { orgId: options.orgId }
-          }
-        };
-        if (isGlobPattern) {
-          include = [target];
-        } else {
+var import_commander, import_nanoid, import_node_fs3, createRunId, consoleUrl, loadEvalCommand;
+var init_eval_command = __esm({
+  "src/cli/commands/eval.command.ts"() {
+    "use strict";
+    init_cjs_shims();
+    import_commander = require("commander");
+    import_nanoid = require("nanoid");
+    import_node_fs3 = require("fs");
+    init_eval_context_runner();
+    init_parse_flag_overrides();
+    init_glob_utils();
+    init_loader();
+    init_errors();
+    init_global_auth();
+    init_node();
+    createRunId = (0, import_nanoid.customAlphabet)("0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ", 10);
+    loadEvalCommand = (program2, flagOverrides = {}) => {
+      return program2.addCommand(
+        new import_commander.Command("eval").description("run evals locally").addArgument(
+          new import_commander.Argument("[target]", "file, directory, glob pattern, or eval name").default(
+            ".",
+            "any *.eval.ts file in current directory"
+          )
+        ).option("-w, --watch true", "keep server running and watch for changes", false).option("-t, --token <TOKEN>", "axiom token", getDefaultToken).option("-d, --dataset <DATASET>", "axiom dataset name", process.env.AXIOM_DATASET).option("-u, --url <AXIOM URL>", "axiom url", getDefaultUrl).option("-o, --org-id <ORG ID>", "axiom organization id", getDefaultOrgId).option("-b, --baseline <BASELINE ID>", "id of baseline evaluation to compare against").option("--debug", "run locally without any network operations", false).option("--list", "list evaluations and test cases without running them", false).addOption(new import_commander.Option("-c, --console-url <URL>", "console url override").hideHelp()).action(async (target, options) => {
           try {
-            const stat = (0, import_node_fs3.lstatSync)(target);
-            if (stat.isDirectory()) {
-              include = config?.eval?.include || [];
-            } else {
+            if (options.debug) {
+              process.env.AXIOM_DEBUG = "true";
+            }
+            let include = [];
+            let exclude;
+            let testNamePattern;
+            const isGlobPattern = isGlob(target);
+            const { config: loadedConfig } = await loadConfig(".");
+            validateFlagOverrides(flagOverrides, loadedConfig.eval.flagSchema);
+            const config = {
+              ...loadedConfig,
+              eval: {
+                ...loadedConfig.eval,
+                ...options.token && { token: options.token },
+                ...options.url && { url: options.url },
+                ...options.dataset && { dataset: options.dataset },
+                ...options.orgId && { orgId: options.orgId }
+              }
+            };
+            if (isGlobPattern) {
               include = [target];
+            } else {
+              try {
+                const stat = (0, import_node_fs3.lstatSync)(target);
+                if (stat.isDirectory()) {
+                  include = config?.eval?.include || [];
+                } else {
+                  include = [target];
+                }
+              } catch {
+                testNamePattern = new RegExp(target, "i");
+                include = config?.eval?.include || [];
+              }
             }
-          } catch {
-            testNamePattern = new RegExp(target, "i");
-            include = config?.eval?.include || [];
-          }
-        }
-        exclude = config?.eval?.exclude;
-        if (!config?.eval?.instrumentation) {
-          console.warn(
-            u.yellow(
-              "\u26A0 App instrumentation (`eval.instrumentation` in `axiom.config.ts`) not configured. Using default provider."
-            )
-          );
-          console.log("");
-        }
-        const runId = createRunId();
-        consoleUrl = options.consoleUrl;
-        await runEvalWithContext(flagOverrides, async () => {
-          return runVitest(".", {
-            watch: options.watch,
-            baseline: options.baseline,
-            include,
-            exclude,
-            testNamePattern,
-            debug: options.debug,
-            list: options.list,
-            overrides: flagOverrides,
-            config,
-            runId,
-            consoleUrl: options.consoleUrl
-          });
-        });
-      } catch (error) {
-        if (error instanceof AxiomCLIError) {
-          console.error(`
+            exclude = config?.eval?.exclude;
+            if (!config?.eval?.instrumentation) {
+              console.warn(
+                u.yellow(
+                  "\u26A0 App instrumentation (`eval.instrumentation` in `axiom.config.ts`) not configured. Using default provider."
+                )
+              );
+              console.log("");
+            }
+            const runId = createRunId();
+            consoleUrl = options.consoleUrl;
+            let runVitestModule;
+            try {
+              runVitestModule = await Promise.resolve().then(() => (init_run_vitest(), run_vitest_exports));
+            } catch (err) {
+              if (err && typeof err === "object" && "code" in err && (err.code === "ERR_MODULE_NOT_FOUND" || err.code === "MODULE_NOT_FOUND")) {
+                throw new AxiomCLIError("Failed to load vitest.");
+              }
+              throw err;
+            }
+            const { runVitest: runVitest2 } = runVitestModule;
+            await runEvalWithContext(flagOverrides, async () => {
+              return runVitest2(".", {
+                watch: options.watch,
+                baseline: options.baseline,
+                include,
+                exclude,
+                testNamePattern,
+                debug: options.debug,
+                list: options.list,
+                overrides: flagOverrides,
+                config,
+                runId,
+                consoleUrl: options.consoleUrl
+              });
+            });
+          } catch (error) {
+            if (error instanceof AxiomCLIError) {
+              console.error(`
 \u274C ${error.message}
 `);
-          process.exit(1);
-        }
-        throw error;
-      }
-    })
-  );
-};
+              process.exit(1);
+            }
+            throw error;
+          }
+        })
+      );
+    };
+  }
+});
+
+// src/bin.ts
+var bin_exports = {};
+__export(bin_exports, {
+  program: () => program
+});
+module.exports = __toCommonJS(bin_exports);
+init_cjs_shims();
+var import_commander3 = require("commander");
+init_eval_command();
 
 // src/cli/commands/auth.command.ts
 init_cjs_shims();
 
 // src/cli/commands/auth-login.command.ts
 init_cjs_shims();
+init_oauth();
+init_callback_server();
+init_config();
+init_api();
+init_errors();
 var BASE_HOSTNAME = "axiom.co";
 var getApiUrl = (hostname) => {
   return `https://api.${hostname}`;
@@ -2583,6 +2978,8 @@ function loadAuthLoginCommand(auth, root) {
 
 // src/cli/commands/auth-logout.command.ts
 init_cjs_shims();
+init_config();
+init_errors();
 async function logoutCommand(alias) {
   const config = await loadGlobalConfig();
   const profileToRemove = alias || config.active_profile;
@@ -2628,6 +3025,9 @@ function loadAuthLogoutCommand(auth, root) {
 
 // src/cli/commands/auth-status.command.ts
 init_cjs_shims();
+init_config();
+init_api();
+init_errors();
 async function statusCommand() {
   const config = await loadGlobalConfig();
   if (Object.keys(config.profiles).length === 0) {
@@ -2693,6 +3093,8 @@ function loadAuthStatusCommand(auth, program2) {
 
 // src/cli/commands/auth-switch.command.ts
 init_cjs_shims();
+init_config();
+init_errors();
 async function promptSelect2(message, choices) {
   console.log(`
 ${message}`);
@@ -2788,6 +3190,8 @@ function loadAuthCommand(program2) {
 }
 
 // src/bin.ts
+init_parse_flag_overrides();
+init_global_auth();
 var import_env = __toESM(require("@next/env"), 1);
 
 // src/cli/commands/version.command.ts
@@ -2796,7 +3200,7 @@ var import_commander2 = require("commander");
 var loadVersionCommand = (program2) => {
   return program2.addCommand(
     new import_commander2.Command("version").description("cli version").action(() => {
-      console.log("0.34.1");
+      console.log("0.35.0");
     })
   );
 };
@@ -2806,7 +3210,7 @@ var { loadEnvConfig } = import_env.default;
 loadEnvConfig(process.cwd());
 var { cleanedArgv, overrides } = extractOverrides(process.argv.slice(2));
 var program = new import_commander3.Command();
-program.name("axiom").description("Axiom's CLI to manage your objects and run evals").version("0.34.1");
+program.name("axiom").description("Axiom's CLI to manage your objects and run evals").version("0.35.0");
 program.hook("preAction", async (_, actionCommand) => {
   const commandName = actionCommand.name();
   const parentCommand = actionCommand.parent;