@agentv/core 4.27.0 → 4.28.0

package/dist/evaluation/validation/index.js

@@ -12,7 +12,7 @@ import {
   loadCasesFromFile,
   parseYamlValue,
   resolveFileReference
-} from "../../chunk-7LQI7772.js";
+} from "../../chunk-YFXMMBUG.js";
 
 // src/evaluation/validation/file-type.ts
 import { readFile } from "node:fs/promises";

package/dist/index.cjs

@@ -9059,7 +9059,7 @@ var init_llm_providers = __esm({
   "src/evaluation/providers/llm-providers.ts"() {
     "use strict";
     init_cjs_shims();
-    import_pi_ai = require("@mariozechner/pi-ai");
+    import_pi_ai = require("@earendil-works/pi-ai");
     (0, import_pi_ai.registerBuiltInApiProviders)();
     DEFAULT_SYSTEM_PROMPT = "You are a careful assistant. Follow all provided instructions and do not fabricate results.";
     OpenAIProvider = class {
@@ -9950,6 +9950,9 @@ var init_claude_cli = __esm({
           "--include-partial-messages",
           "--verbose"
         ];
+        if (this.config.bypassPermissions !== false) {
+          args.push("--dangerously-skip-permissions");
+        }
         if (this.config.model) {
           args.push("--model", this.config.model);
         }
@@ -14166,7 +14169,7 @@ async function promptInstall() {
   try {
     return await new Promise((resolve) => {
       rl.question(
-        "@mariozechner/pi-coding-agent is not installed. Install it now? (y/N) ",
+        "@earendil-works/pi-coding-agent is not installed. Install it now? (y/N) ",
         (answer) => resolve(answer.trim().toLowerCase() === "y")
       );
     });
@@ -14204,8 +14207,8 @@ function findAccessiblePath(paths) {
 async function tryImportLocalSdkModules() {
   try {
     [piCodingAgentModule, piAiModule] = await Promise.all([
-      import("@mariozechner/pi-coding-agent"),
-      import("@mariozechner/pi-ai")
+      import("@earendil-works/pi-coding-agent"),
+      import("@earendil-works/pi-ai")
     ]);
     return true;
   } catch {
@@ -14215,17 +14218,24 @@ async function tryImportLocalSdkModules() {
 async function tryImportManagedSdkModules() {
   const managedRoot = findManagedSdkInstallRoot();
   const piCodingAgentEntry = findAccessiblePath([
-    import_node_path25.default.join(managedRoot, "node_modules", "@mariozechner", "pi-coding-agent", "dist", "index.js")
+    import_node_path25.default.join(
+      managedRoot,
+      "node_modules",
+      "@earendil-works",
+      "pi-coding-agent",
+      "dist",
+      "index.js"
+    )
   ]);
   const piAiEntry = findAccessiblePath([
-    import_node_path25.default.join(managedRoot, "node_modules", "@mariozechner", "pi-ai", "dist", "index.js"),
+    import_node_path25.default.join(managedRoot, "node_modules", "@earendil-works", "pi-ai", "dist", "index.js"),
     import_node_path25.default.join(
       managedRoot,
       "node_modules",
-      "@mariozechner",
+      "@earendil-works",
       "pi-coding-agent",
       "node_modules",
-      "@mariozechner",
+      "@earendil-works",
       "pi-ai",
       "dist",
       "index.js"
@@ -14246,16 +14256,16 @@ async function tryImportGlobalSdkModules() {
   const globalNpmRoot = resolveGlobalNpmRoot();
   if (!globalNpmRoot) return false;
   const piCodingAgentEntry = findAccessiblePath([
-    buildGlobalModuleEntry("@mariozechner/pi-coding-agent", globalNpmRoot)
+    buildGlobalModuleEntry("@earendil-works/pi-coding-agent", globalNpmRoot)
   ]);
   const piAiEntry = findAccessiblePath([
-    buildGlobalModuleEntry("@mariozechner/pi-ai", globalNpmRoot),
+    buildGlobalModuleEntry("@earendil-works/pi-ai", globalNpmRoot),
     import_node_path25.default.join(
       globalNpmRoot,
-      "@mariozechner",
+      "@earendil-works",
       "pi-coding-agent",
       "node_modules",
-      "@mariozechner",
+      "@earendil-works",
       "pi-ai",
       "dist",
       "index.js"
@@ -14273,9 +14283,9 @@ async function tryImportGlobalSdkModules() {
   }
 }
 function installSdkModules(installDir) {
-  console.error(`Installing @mariozechner/pi-coding-agent into ${installDir} via npm...`);
+  console.error(`Installing @earendil-works/pi-coding-agent into ${installDir} via npm...`);
   (0, import_node_fs10.mkdirSync)(installDir, { recursive: true });
-  (0, import_node_child_process6.execSync)("npm install --no-save --no-package-lock @mariozechner/pi-coding-agent", {
+  (0, import_node_child_process6.execSync)("npm install --no-save --no-package-lock @earendil-works/pi-coding-agent", {
     cwd: installDir,
     stdio: "inherit"
   });
@@ -14292,7 +14302,7 @@ async function doLoadSdkModules() {
     }
   }
   throw new Error(
-    "pi-coding-agent SDK is not installed. Install it with:\n  npm install @mariozechner/pi-coding-agent"
+    "pi-coding-agent SDK is not installed. Install it with:\n  npm install @earendil-works/pi-coding-agent"
   );
 }
 async function loadSdkModules() {
@@ -15782,6 +15792,7 @@ function resolveClaudeConfig(target, env, _evalFilePath) {
   const systemPrompt = typeof systemPromptSource === "string" && systemPromptSource.trim().length > 0 ? systemPromptSource.trim() : void 0;
   const maxTurns = typeof target.max_turns === "number" ? target.max_turns : void 0;
   const maxBudgetUsd = typeof target.max_budget_usd === "number" ? target.max_budget_usd : void 0;
+  const bypassPermissions = target.bypass_permissions !== void 0 ? resolveOptionalBoolean(target.bypass_permissions) : void 0;
   return {
     executable,
     model,
@@ -15792,7 +15803,8 @@ function resolveClaudeConfig(target, env, _evalFilePath) {
     maxBudgetUsd,
     logDir,
     logFormat,
-    streamLog: streamLogResult.streamLog
+    streamLog: streamLogResult.streamLog,
+    bypassPermissions
   };
 }
 function resolveCcMirrorBinaryPath(variant) {
@@ -19086,6 +19098,9 @@ function gitEnv2() {
 function getSourceUrl(source) {
   return source.type === "git" ? source.url : source.path;
 }
+function isFullCommitSha(ref) {
+  return typeof ref === "string" && /^[0-9a-f]{40}$/i.test(ref);
+}
 async function git2(args, opts) {
   const { stdout } = await execFileAsync2("git", args, {
     cwd: opts?.cwd,
@@ -19208,8 +19223,10 @@ ${lines.join("\n")}`;
         }
         const ref = getRepoCheckoutRef(repo.checkout);
         const resolve = repo.checkout?.resolve ?? "remote";
+        const baseCommit = repo.checkout?.base_commit;
+        const shouldResolveLocally = resolve === "local" || repo.source.type === "git" && isFullCommitSha(baseCommit);
         let resolvedSha;
-        if (resolve === "remote" && repo.source.type === "git") {
+        if (!shouldResolveLocally && repo.source.type === "git") {
           const url = getSourceUrl(repo.source);
           try {
             const lsOutput = await this.runGit(["ls-remote", url, ref]);
@@ -24808,6 +24825,8 @@ __export(index_exports, {
   subscribeToCopilotSdkLogEntries: () => subscribeToCopilotSdkLogEntries,
   subscribeToPiLogEntries: () => subscribeToPiLogEntries,
   substituteVariables: () => substituteVariables,
+  syncBenchmark: () => syncBenchmark,
+  syncBenchmarks: () => syncBenchmarks,
   syncResultsRepo: () => syncResultsRepo,
   toCamelCaseDeep: () => toCamelCaseDeep,
   toSnakeCaseDeep: () => toSnakeCaseDeep,
@@ -25163,12 +25182,12 @@ var CONFIG_FILE_NAMES = [
   ".agentv/config.js"
 ];
 async function loadTsConfig(projectRoot) {
-  const { existsSync: existsSync10 } = await import("fs");
+  const { existsSync: existsSync11 } = await import("fs");
   const { pathToFileURL: pathToFileURL3 } = await import("url");
   const { join: join2 } = await import("path");
   for (const fileName of CONFIG_FILE_NAMES) {
     const filePath = join2(projectRoot, fileName);
-    if (!existsSync10(filePath)) {
+    if (!existsSync11(filePath)) {
       continue;
     }
     try {
@@ -25752,6 +25771,7 @@ init_cjs_shims();
 var import_node_fs20 = require("fs");
 var import_node_path55 = __toESM(require("path"), 1);
 var import_yaml2 = require("yaml");
+init_interpolation();
 init_yaml_loader();
 init_paths();
 function getBenchmarksRegistryPath() {
@@ -25763,22 +25783,33 @@ function fromYaml(raw) {
   if (typeof e.id !== "string" || typeof e.name !== "string" || typeof e.path !== "string") {
     return null;
   }
-  return {
+  const entry = {
     id: e.id,
     name: e.name,
     path: e.path,
     addedAt: typeof e.added_at === "string" ? e.added_at : "",
     lastOpenedAt: typeof e.last_opened_at === "string" ? e.last_opened_at : ""
   };
+  if (e.source && typeof e.source === "object") {
+    const s = e.source;
+    if (typeof s.url === "string" && typeof s.ref === "string") {
+      entry.source = { url: s.url, ref: s.ref };
+    }
+  }
+  return entry;
 }
 function toYaml(entry) {
-  return {
+  const yaml = {
     id: entry.id,
     name: entry.name,
     path: entry.path,
     added_at: entry.addedAt,
     last_opened_at: entry.lastOpenedAt
   };
+  if (entry.source) {
+    yaml.source = { url: entry.source.url, ref: entry.source.ref };
+  }
+  return yaml;
 }
 function loadBenchmarkRegistry() {
   const registryPath = getBenchmarksRegistryPath();
@@ -25791,7 +25822,8 @@ function loadBenchmarkRegistry() {
     if (!parsed || typeof parsed !== "object") {
       return { benchmarks: [] };
     }
-    const benchmarks = Array.isArray(parsed.benchmarks) ? parsed.benchmarks.map(fromYaml).filter((e) => e !== null) : [];
+    const env = process.env;
+    const benchmarks = Array.isArray(parsed.benchmarks) ? parsed.benchmarks.map((e) => fromYaml(interpolateEnv(e, env))).filter((e) => e !== null) : [];
     return { benchmarks };
   } catch {
     return { benchmarks: [] };
@@ -25890,6 +25922,35 @@ function discoverBenchmarks(rootDir, maxDepth = 2) {
   return results.sort();
 }
 
+// src/benchmark-sync.ts
+init_cjs_shims();
+var childProcess = __toESM(require("child_process"), 1);
+var import_node_fs21 = require("fs");
+async function syncBenchmark(entry) {
+  if (!entry.source) {
+    throw new Error(`Benchmark '${entry.id}' has no source defined`);
+  }
+  const { url, ref } = entry.source;
+  const dest = entry.path;
+  if ((0, import_node_fs21.existsSync)(`${dest}/.git`)) {
+    childProcess.execFileSync("git", ["-C", dest, "pull", "--ff-only"], { stdio: "inherit" });
+  } else {
+    childProcess.execFileSync(
+      "git",
+      ["clone", "--depth", "1", "--filter=blob:none", "--branch", ref, url, dest],
+      { stdio: "inherit" }
+    );
+  }
+}
+async function syncBenchmarks(entries) {
+  for (const entry of entries) {
+    if (!entry.source) continue;
+    console.log(`Syncing benchmark '${entry.id}' from ${entry.source.url}...`);
+    await syncBenchmark(entry);
+    console.log(`Benchmark '${entry.id}' synced.`);
+  }
+}
+
 // src/evaluation/baseline.ts
 init_cjs_shims();
 var STRIPPED_TOP_LEVEL_FIELDS = /* @__PURE__ */ new Set([
@@ -27379,6 +27440,8 @@ function createAgentKernel() {
   subscribeToCopilotSdkLogEntries,
   subscribeToPiLogEntries,
   substituteVariables,
+  syncBenchmark,
+  syncBenchmarks,
   syncResultsRepo,
   toCamelCaseDeep,
   toSnakeCaseDeep,