agentv 4.5.2 → 4.6.0

package/dist/{chunk-HF5UGZSZ.js → chunk-U2LSJ6Y4.js}

@@ -24,7 +24,7 @@ import {
   validateFileReferences,
   validateTargetsFile,
   writeArtifactsFromResults
-} from "./chunk-OIBYQMCK.js";
+} from "./chunk-5GZJIXTY.js";
 import {
   DEFAULT_CATEGORY,
   PASS_THRESHOLD,
@@ -50,13 +50,10 @@ import {
   toSnakeCaseDeep as toSnakeCaseDeep2,
   transpileEvalYamlFile,
   trimBaselineResult
-} from "./chunk-7DRAXDVC.js";
+} from "./chunk-KQQTEWZF.js";
 import {
   __commonJS,
-  __esm,
-  __export,
   __require,
-  __toCommonJS,
   __toESM
 } from "./chunk-5H446C7X.js";
 
@@ -281,7 +278,7 @@ var require_ms = __commonJS({
 var require_common = __commonJS({
   "../../node_modules/.bun/debug@4.4.3/node_modules/debug/src/common.js"(exports, module) {
     "use strict";
-    function setup(env3) {
+    function setup(env2) {
       createDebug.debug = createDebug;
       createDebug.default = createDebug;
       createDebug.coerce = coerce;
@@ -290,8 +287,8 @@ var require_common = __commonJS({
       createDebug.enabled = enabled;
       createDebug.humanize = require_ms();
       createDebug.destroy = destroy;
-      Object.keys(env3).forEach((key) => {
-        createDebug[key] = env3[key];
+      Object.keys(env2).forEach((key) => {
+        createDebug[key] = env2[key];
       });
       createDebug.names = [];
       createDebug.skips = [];
@@ -626,153 +623,118 @@ var require_browser = __commonJS({
   }
 });
 
-// ../../node_modules/.bun/supports-color@10.2.2/node_modules/supports-color/index.js
-var supports_color_exports = {};
-__export(supports_color_exports, {
-  createSupportsColor: () => createSupportsColor2,
-  default: () => supports_color_default2
-});
-import process3 from "node:process";
-import os2 from "node:os";
-import tty2 from "node:tty";
-function hasFlag2(flag2, argv = globalThis.Deno ? globalThis.Deno.args : process3.argv) {
-  const prefix = flag2.startsWith("-") ? "" : flag2.length === 1 ? "-" : "--";
-  const position = argv.indexOf(prefix + flag2);
-  const terminatorPosition = argv.indexOf("--");
-  return position !== -1 && (terminatorPosition === -1 || position < terminatorPosition);
-}
-function envForceColor2() {
-  if (!("FORCE_COLOR" in env2)) {
-    return;
-  }
-  if (env2.FORCE_COLOR === "true") {
-    return 1;
-  }
-  if (env2.FORCE_COLOR === "false") {
-    return 0;
-  }
-  if (env2.FORCE_COLOR.length === 0) {
-    return 1;
-  }
-  const level = Math.min(Number.parseInt(env2.FORCE_COLOR, 10), 3);
-  if (![0, 1, 2, 3].includes(level)) {
-    return;
-  }
-  return level;
-}
-function translateLevel2(level) {
-  if (level === 0) {
-    return false;
-  }
-  return {
-    level,
-    hasBasic: true,
-    has256: level >= 2,
-    has16m: level >= 3
-  };
-}
-function _supportsColor2(haveStream, { streamIsTTY, sniffFlags = true } = {}) {
-  const noFlagForceColor = envForceColor2();
-  if (noFlagForceColor !== void 0) {
-    flagForceColor2 = noFlagForceColor;
-  }
-  const forceColor = sniffFlags ? flagForceColor2 : noFlagForceColor;
-  if (forceColor === 0) {
-    return 0;
-  }
-  if (sniffFlags) {
-    if (hasFlag2("color=16m") || hasFlag2("color=full") || hasFlag2("color=truecolor")) {
-      return 3;
-    }
-    if (hasFlag2("color=256")) {
-      return 2;
-    }
-  }
-  if ("TF_BUILD" in env2 && "AGENT_NAME" in env2) {
-    return 1;
-  }
-  if (haveStream && !streamIsTTY && forceColor === void 0) {
-    return 0;
-  }
-  const min = forceColor || 0;
-  if (env2.TERM === "dumb") {
-    return min;
+// ../../node_modules/.bun/has-flag@4.0.0/node_modules/has-flag/index.js
+var require_has_flag = __commonJS({
+  "../../node_modules/.bun/has-flag@4.0.0/node_modules/has-flag/index.js"(exports, module) {
+    "use strict";
+    module.exports = (flag2, argv = process.argv) => {
+      const prefix = flag2.startsWith("-") ? "" : flag2.length === 1 ? "-" : "--";
+      const position = argv.indexOf(prefix + flag2);
+      const terminatorPosition = argv.indexOf("--");
+      return position !== -1 && (terminatorPosition === -1 || position < terminatorPosition);
+    };
   }
-  if (process3.platform === "win32") {
-    const osRelease = os2.release().split(".");
-    if (Number(osRelease[0]) >= 10 && Number(osRelease[2]) >= 10586) {
-      return Number(osRelease[2]) >= 14931 ? 3 : 2;
+});
+
+// ../../node_modules/.bun/supports-color@7.2.0/node_modules/supports-color/index.js
+var require_supports_color = __commonJS({
+  "../../node_modules/.bun/supports-color@7.2.0/node_modules/supports-color/index.js"(exports, module) {
+    "use strict";
+    var os2 = __require("os");
+    var tty2 = __require("tty");
+    var hasFlag2 = require_has_flag();
+    var { env: env2 } = process;
+    var forceColor;
+    if (hasFlag2("no-color") || hasFlag2("no-colors") || hasFlag2("color=false") || hasFlag2("color=never")) {
+      forceColor = 0;
+    } else if (hasFlag2("color") || hasFlag2("colors") || hasFlag2("color=true") || hasFlag2("color=always")) {
+      forceColor = 1;
     }
-    return 1;
-  }
-  if ("CI" in env2) {
-    if (["GITHUB_ACTIONS", "GITEA_ACTIONS", "CIRCLECI"].some((key) => key in env2)) {
-      return 3;
+    if ("FORCE_COLOR" in env2) {
+      if (env2.FORCE_COLOR === "true") {
+        forceColor = 1;
+      } else if (env2.FORCE_COLOR === "false") {
+        forceColor = 0;
+      } else {
+        forceColor = env2.FORCE_COLOR.length === 0 ? 1 : Math.min(parseInt(env2.FORCE_COLOR, 10), 3);
+      }
     }
-    if (["TRAVIS", "APPVEYOR", "GITLAB_CI", "BUILDKITE", "DRONE"].some((sign) => sign in env2) || env2.CI_NAME === "codeship") {
-      return 1;
+    function translateLevel2(level) {
+      if (level === 0) {
+        return false;
+      }
+      return {
+        level,
+        hasBasic: true,
+        has256: level >= 2,
+        has16m: level >= 3
+      };
     }
-    return min;
-  }
-  if ("TEAMCITY_VERSION" in env2) {
-    return /^(9\.(0*[1-9]\d*)\.|\d{2,}\.)/.test(env2.TEAMCITY_VERSION) ? 1 : 0;
-  }
-  if (env2.COLORTERM === "truecolor") {
-    return 3;
-  }
-  if (env2.TERM === "xterm-kitty") {
-    return 3;
-  }
-  if (env2.TERM === "xterm-ghostty") {
-    return 3;
-  }
-  if (env2.TERM === "wezterm") {
-    return 3;
-  }
-  if ("TERM_PROGRAM" in env2) {
-    const version = Number.parseInt((env2.TERM_PROGRAM_VERSION || "").split(".")[0], 10);
-    switch (env2.TERM_PROGRAM) {
-      case "iTerm.app": {
-        return version >= 3 ? 3 : 2;
+    function supportsColor2(haveStream, streamIsTTY) {
+      if (forceColor === 0) {
+        return 0;
       }
-      case "Apple_Terminal": {
+      if (hasFlag2("color=16m") || hasFlag2("color=full") || hasFlag2("color=truecolor")) {
+        return 3;
+      }
+      if (hasFlag2("color=256")) {
+        return 2;
+      }
+      if (haveStream && !streamIsTTY && forceColor === void 0) {
+        return 0;
+      }
+      const min = forceColor || 0;
+      if (env2.TERM === "dumb") {
+        return min;
+      }
+      if (process.platform === "win32") {
+        const osRelease = os2.release().split(".");
+        if (Number(osRelease[0]) >= 10 && Number(osRelease[2]) >= 10586) {
+          return Number(osRelease[2]) >= 14931 ? 3 : 2;
+        }
+        return 1;
+      }
+      if ("CI" in env2) {
+        if (["TRAVIS", "CIRCLECI", "APPVEYOR", "GITLAB_CI", "GITHUB_ACTIONS", "BUILDKITE"].some((sign) => sign in env2) || env2.CI_NAME === "codeship") {
+          return 1;
+        }
+        return min;
+      }
+      if ("TEAMCITY_VERSION" in env2) {
+        return /^(9\.(0*[1-9]\d*)\.|\d{2,}\.)/.test(env2.TEAMCITY_VERSION) ? 1 : 0;
+      }
+      if (env2.COLORTERM === "truecolor") {
+        return 3;
+      }
+      if ("TERM_PROGRAM" in env2) {
+        const version = parseInt((env2.TERM_PROGRAM_VERSION || "").split(".")[0], 10);
+        switch (env2.TERM_PROGRAM) {
+          case "iTerm.app":
+            return version >= 3 ? 3 : 2;
+          case "Apple_Terminal":
+            return 2;
+        }
+      }
+      if (/-256(color)?$/i.test(env2.TERM)) {
         return 2;
       }
+      if (/^screen|^xterm|^vt100|^vt220|^rxvt|color|ansi|cygwin|linux/i.test(env2.TERM)) {
+        return 1;
+      }
+      if ("COLORTERM" in env2) {
+        return 1;
+      }
+      return min;
     }
-  }
-  if (/-256(color)?$/i.test(env2.TERM)) {
-    return 2;
-  }
-  if (/^screen|^xterm|^vt100|^vt220|^rxvt|color|ansi|cygwin|linux/i.test(env2.TERM)) {
-    return 1;
-  }
-  if ("COLORTERM" in env2) {
-    return 1;
-  }
-  return min;
-}
-function createSupportsColor2(stream, options = {}) {
-  const level = _supportsColor2(stream, {
-    streamIsTTY: stream && stream.isTTY,
-    ...options
-  });
-  return translateLevel2(level);
-}
-var env2, flagForceColor2, supportsColor2, supports_color_default2;
-var init_supports_color = __esm({
-  "../../node_modules/.bun/supports-color@10.2.2/node_modules/supports-color/index.js"() {
-    "use strict";
-    ({ env: env2 } = process3);
-    if (hasFlag2("no-color") || hasFlag2("no-colors") || hasFlag2("color=false") || hasFlag2("color=never")) {
-      flagForceColor2 = 0;
-    } else if (hasFlag2("color") || hasFlag2("colors") || hasFlag2("color=true") || hasFlag2("color=always")) {
-      flagForceColor2 = 1;
+    function getSupportLevel(stream) {
+      const level = supportsColor2(stream, stream && stream.isTTY);
+      return translateLevel2(level);
     }
-    supportsColor2 = {
-      stdout: createSupportsColor2({ isTTY: tty2.isatty(1) }),
-      stderr: createSupportsColor2({ isTTY: tty2.isatty(2) })
+    module.exports = {
+      supportsColor: getSupportLevel,
+      stdout: translateLevel2(supportsColor2(true, tty2.isatty(1))),
+      stderr: translateLevel2(supportsColor2(true, tty2.isatty(2)))
     };
-    supports_color_default2 = supportsColor2;
   }
 });
 
@@ -780,7 +742,7 @@ var init_supports_color = __esm({
 var require_node = __commonJS({
   "../../node_modules/.bun/debug@4.4.3/node_modules/debug/src/node.js"(exports, module) {
     "use strict";
-    var tty3 = __require("tty");
+    var tty2 = __require("tty");
     var util = __require("util");
     exports.init = init;
     exports.log = log;
@@ -795,8 +757,8 @@ var require_node = __commonJS({
     );
     exports.colors = [6, 2, 3, 4, 5, 1];
     try {
-      const supportsColor3 = (init_supports_color(), __toCommonJS(supports_color_exports));
-      if (supportsColor3 && (supportsColor3.stderr || supportsColor3).level >= 2) {
+      const supportsColor2 = require_supports_color();
+      if (supportsColor2 && (supportsColor2.stderr || supportsColor2).level >= 2) {
         exports.colors = [
           20,
           21,
@@ -898,7 +860,7 @@ var require_node = __commonJS({
       return obj;
     }, {});
     function useColors() {
-      return "colors" in exports.inspectOpts ? Boolean(exports.inspectOpts.colors) : tty3.isatty(process.stderr.fd);
+      return "colors" in exports.inspectOpts ? Boolean(exports.inspectOpts.colors) : tty2.isatty(process.stderr.fd);
     }
     function formatArgs(args) {
       const { namespace: name, useColors: useColors2 } = this;
@@ -1630,8 +1592,8 @@ function fullFlag(config) {
       }
       const defaults = [];
       if (config.env) {
-        const env3 = process.env[config.env] === void 0 ? "" : `=${source_default.italic(process.env[config.env])}`;
-        defaults.push(`env: ${config.env}${env3}`);
+        const env2 = process.env[config.env] === void 0 ? "" : `=${source_default.italic(process.env[config.env])}`;
+        defaults.push(`env: ${config.env}${env2}`);
       }
       if (config.defaultValue) {
         try {
@@ -2248,8 +2210,8 @@ function fullOption(config) {
       usage += ` <${displayName}>`;
       const defaults = [];
       if (config.env) {
-        const env3 = process.env[config.env] === void 0 ? "" : `=${source_default.italic(process.env[config.env])}`;
-        defaults.push(`env: ${config.env}${env3}`);
+        const env2 = process.env[config.env] === void 0 ? "" : `=${source_default.italic(process.env[config.env])}`;
+        defaults.push(`env: ${config.env}${env2}`);
       }
       if (config.defaultValue) {
         try {
@@ -3998,11 +3960,21 @@ var evalRunCommand = command({
       type: optional(number),
       long: "threshold",
       description: "Per-test score threshold (0-1, default 0.8). Exit 1 if any test scores below this value"
+    }),
+    tag: multioption({
+      type: array(string),
+      long: "tag",
+      description: "Only run eval files that have this tag (repeatable, AND logic)"
+    }),
+    excludeTag: multioption({
+      type: array(string),
+      long: "exclude-tag",
+      description: "Skip eval files that have this tag (repeatable, file skipped if any match)"
     })
   },
   handler: async (args) => {
     if (args.evalPaths.length === 0 && process.stdin.isTTY) {
-      const { launchInteractiveWizard } = await import("./interactive-BOJUYBJS.js");
+      const { launchInteractiveWizard } = await import("./interactive-6D3ULOMN.js");
       await launchInteractiveWizard();
       return;
     }
@@ -4039,7 +4011,9 @@ var evalRunCommand = command({
       graderTarget: args.graderTarget,
       model: args.model,
       outputMessages: args.outputMessages,
-      threshold: args.threshold
+      threshold: args.threshold,
+      tag: args.tag,
+      excludeTag: args.excludeTag
     };
     const result = await runEvalCommand({ testFiles: resolvedPaths, rawOptions });
     if (result?.thresholdFailed) {
@@ -4886,15 +4860,15 @@ function loadEnvFile(dir) {
   while (true) {
     const candidate = join4(current, ".env");
     if (existsSync3(candidate)) {
-      const env3 = {};
+      const env2 = {};
       for (const line of readFileSync4(candidate, "utf8").split("\n")) {
         const trimmed = line.trim();
         if (!trimmed || trimmed.startsWith("#")) continue;
         const eqIdx = trimmed.indexOf("=");
         if (eqIdx === -1) continue;
-        env3[trimmed.slice(0, eqIdx).trim()] = trimmed.slice(eqIdx + 1).trim();
+        env2[trimmed.slice(0, eqIdx).trim()] = trimmed.slice(eqIdx + 1).trim();
       }
-      return env3;
+      return env2;
     }
     const parent = dirname2(current);
     if (parent === current) break;
@@ -8569,4 +8543,4 @@ export {
   preprocessArgv,
   runCli
 };
-//# sourceMappingURL=chunk-HF5UGZSZ.js.map
+//# sourceMappingURL=chunk-U2LSJ6Y4.js.map