agentv 4.11.2 → 4.12.0

package/dist/{chunk-7TJ2PON3.js → chunk-VVWPD4CN.js}

@@ -42,25 +42,25 @@ import {
   validateFileReferences,
   validateTargetsFile,
   writeArtifactsFromResults
-} from "./chunk-MIP46NEN.js";
+} from "./chunk-4MEGL2E3.js";
 import {
   DEFAULT_CATEGORY,
   DEFAULT_THRESHOLD,
-  addProject,
+  addBenchmark,
   createBuiltinRegistry,
   deriveCategory,
+  discoverBenchmarks,
   discoverClaudeSessions,
   discoverCodexSessions,
   discoverCopilotSessions,
-  discoverProjects,
   executeScript,
   getAgentvHome,
+  getBenchmark,
   getOutputFilenames,
-  getProject,
   getWorkspacePoolRoot,
   isAgentSkillsFormat,
   listTargetNames,
-  loadProjectRegistry,
+  loadBenchmarkRegistry,
   loadTestSuite,
   normalizeLineEndings,
   parseAgentSkillsEvals,
@@ -69,14 +69,14 @@ import {
   parseCopilotEvents,
   readTargetDefinitions,
   readTranscriptFile,
-  removeProject,
+  removeBenchmark,
   scanRepoDeps,
   toCamelCaseDeep,
   toSnakeCaseDeep as toSnakeCaseDeep2,
   toTranscriptJsonLine,
   transpileEvalYamlFile,
   trimBaselineResult
-} from "./chunk-FQGY6QXQ.js";
+} from "./chunk-CXAO4VPP.js";
 import {
   __commonJS,
   __require,
@@ -3912,7 +3912,7 @@ var evalRunCommand = command({
   },
   handler: async (args) => {
     if (args.evalPaths.length === 0 && process.stdin.isTTY) {
-      const { launchInteractiveWizard } = await import("./interactive-LRW3X5OF.js");
+      const { launchInteractiveWizard } = await import("./interactive-SP2LWOQX.js");
       await launchInteractiveWizard();
       return;
     }
@@ -7252,13 +7252,14 @@ var resultsCommand = subcommands({
 // src/commands/results/serve.ts
 import { existsSync as existsSync12, readFileSync as readFileSync10, readdirSync as readdirSync4, statSync as statSync5, writeFileSync as writeFileSync4 } from "node:fs";
 import path16 from "node:path";
-import { fileURLToPath as fileURLToPath2 } from "node:url";
+import { fileURLToPath as fileURLToPath3 } from "node:url";
 import { Hono } from "hono";
 
 // src/commands/results/eval-runner.ts
-import { spawn } from "node:child_process";
+import { execFileSync, spawn } from "node:child_process";
 import { existsSync as existsSync10 } from "node:fs";
 import path14 from "node:path";
+import { fileURLToPath as fileURLToPath2 } from "node:url";
 var activeRuns = /* @__PURE__ */ new Map();
 function generateRunId() {
   const now = /* @__PURE__ */ new Date();
@@ -7342,16 +7343,27 @@ function resolveCliPath(cwd) {
   ];
   for (const c4 of candidates) {
     if (existsSync10(c4)) {
-      return { bunPath: "bun", cliPath: c4 };
+      return { binPath: "bun", args: [c4] };
     }
   }
-  const currentDir = typeof __dirname !== "undefined" ? __dirname : path14.dirname(new URL(import.meta.url).pathname);
+  const currentDir = typeof __dirname !== "undefined" ? __dirname : path14.dirname(fileURLToPath2(import.meta.url));
   const fromSrc = path14.resolve(currentDir, "../../../cli.ts");
   const fromDist = path14.resolve(currentDir, "../../cli.js");
-  if (existsSync10(fromSrc)) return { bunPath: "bun", cliPath: fromSrc };
-  if (existsSync10(fromDist)) return { bunPath: "bun", cliPath: fromDist };
+  if (existsSync10(fromSrc)) return { binPath: "bun", args: [fromSrc] };
+  if (existsSync10(fromDist)) return { binPath: "bun", args: [fromDist] };
+  if (isCommandAvailable("agentv")) {
+    return { binPath: "agentv", args: [] };
+  }
   return void 0;
 }
+function isCommandAvailable(cmd) {
+  try {
+    execFileSync(process.platform === "win32" ? "where" : "which", [cmd], { stdio: "ignore" });
+    return true;
+  } catch {
+    return false;
+  }
+}
 function registerEvalRoutes(app2, getCwd, options) {
   const readOnly = options?.readOnly === true;
   app2.get("/api/eval/discover", async (c4) => {
@@ -7409,7 +7421,7 @@ function registerEvalRoutes(app2, getCwd, options) {
     };
     activeRuns.set(runId, run2);
     try {
-      const child = spawn(cliPaths.bunPath, [cliPaths.cliPath, ...args], {
+      const child = spawn(cliPaths.binPath, [...cliPaths.args, ...args], {
         cwd,
         stdio: ["ignore", "pipe", "pipe"],
         env: { ...process.env }
@@ -7494,7 +7506,7 @@ Process error: ${err2.message}`;
     const args = buildCliArgs(body);
     return c4.json({ command: buildCliPreview(args) });
   });
-  app2.get("/api/projects/:projectId/eval/discover", async (c4) => {
+  app2.get("/api/benchmarks/:benchmarkId/eval/discover", async (c4) => {
     const cwd = getCwd(c4);
     try {
       const files = await discoverEvalFiles(cwd);
@@ -7509,7 +7521,7 @@ Process error: ${err2.message}`;
       return c4.json({ error: err2.message, eval_files: [] }, 500);
     }
   });
-  app2.get("/api/projects/:projectId/eval/targets", async (c4) => {
+  app2.get("/api/benchmarks/:benchmarkId/eval/targets", async (c4) => {
     const cwd = getCwd(c4);
     try {
       const names = await discoverTargetsInProject(cwd);
@@ -7518,7 +7530,7 @@ Process error: ${err2.message}`;
       return c4.json({ error: err2.message, targets: [] }, 500);
     }
   });
-  app2.post("/api/projects/:projectId/eval/run", async (c4) => {
+  app2.post("/api/benchmarks/:benchmarkId/eval/run", async (c4) => {
     const cwd = getCwd(c4);
     let body;
     try {
@@ -7546,7 +7558,7 @@ Process error: ${err2.message}`;
     };
     activeRuns.set(runId, run2);
     try {
-      const child = spawn(cliPaths.bunPath, [cliPaths.cliPath, ...args], {
+      const child = spawn(cliPaths.binPath, [...cliPaths.args, ...args], {
         cwd,
         stdio: ["ignore", "pipe", "pipe"],
         env: { ...process.env }
@@ -7583,7 +7595,7 @@ Process error: ${err2.message}`;
       return c4.json({ error: err2.message }, 500);
     }
   });
-  app2.get("/api/projects/:projectId/eval/status/:id", (c4) => {
+  app2.get("/api/benchmarks/:benchmarkId/eval/status/:id", (c4) => {
     const id = c4.req.param("id");
     const run2 = activeRuns.get(id ?? "");
     if (!run2) return c4.json({ error: "Run not found" }, 404);
@@ -7598,7 +7610,7 @@ Process error: ${err2.message}`;
       stderr: run2.stderr.slice(-5e3)
     });
   });
-  app2.get("/api/projects/:projectId/eval/runs", (c4) => {
+  app2.get("/api/benchmarks/:benchmarkId/eval/runs", (c4) => {
     const runs = [...activeRuns.values()].map((r) => ({
       id: r.id,
       status: r.status,
@@ -7610,7 +7622,7 @@ Process error: ${err2.message}`;
     runs.sort((a, b) => b.started_at.localeCompare(a.started_at));
     return c4.json({ runs });
   });
-  app2.post("/api/projects/:projectId/eval/preview", async (c4) => {
+  app2.post("/api/benchmarks/:benchmarkId/eval/preview", async (c4) => {
     let body;
     try {
       body = await c4.req.json();
@@ -8164,14 +8176,14 @@ function createApp(results, resultDir, cwd, sourceFile, options) {
   const defaultCtx = { searchDir, agentvDir };
   const readOnly = options?.readOnly === true;
   const app2 = new Hono();
-  function withProject(c4, handler) {
-    const project = getProject(c4.req.param("projectId") ?? "");
-    if (!project || !existsSync12(project.path)) {
+  function withBenchmark(c4, handler) {
+    const benchmark = getBenchmark(c4.req.param("benchmarkId") ?? "");
+    if (!benchmark || !existsSync12(benchmark.path)) {
       return c4.json({ error: "Project not found" }, 404);
     }
     return handler(c4, {
-      searchDir: project.path,
-      agentvDir: path16.join(project.path, ".agentv")
+      searchDir: benchmark.path,
+      agentvDir: path16.join(benchmark.path, ".agentv")
     });
   }
   app2.post("/api/config", async (c4) => {
@@ -8191,7 +8203,7 @@ function createApp(results, resultDir, cwd, sourceFile, options) {
       return c4.json({ error: "Failed to save config" }, 500);
     }
   });
-  function projectEntryToWire(entry) {
+  function benchmarkEntryToWire(entry) {
     return {
       id: entry.id,
       name: entry.name,
@@ -8200,10 +8212,10 @@ function createApp(results, resultDir, cwd, sourceFile, options) {
       last_opened_at: entry.lastOpenedAt
     };
   }
-  app2.get("/api/projects", async (c4) => {
-    const registry = loadProjectRegistry();
-    const projects = await Promise.all(
-      registry.projects.map(async (p) => {
+  app2.get("/api/benchmarks", async (c4) => {
+    const registry = loadBenchmarkRegistry();
+    const benchmarks = await Promise.all(
+      registry.benchmarks.map(async (p) => {
         let runCount = 0;
         let passRate = 0;
         let lastRun = null;
@@ -8218,48 +8230,48 @@ function createApp(results, resultDir, cwd, sourceFile, options) {
         } catch {
         }
         return {
-          ...projectEntryToWire(p),
+          ...benchmarkEntryToWire(p),
           run_count: runCount,
           pass_rate: passRate,
           last_run: lastRun
         };
       })
     );
-    return c4.json({ projects });
+    return c4.json({ projects: benchmarks });
   });
-  app2.post("/api/projects", async (c4) => {
+  app2.post("/api/benchmarks", async (c4) => {
     if (readOnly) {
       return c4.json({ error: "Studio is running in read-only mode" }, 403);
     }
     try {
       const body = await c4.req.json();
       if (!body.path) return c4.json({ error: "Missing path" }, 400);
-      const entry = addProject(body.path);
-      return c4.json(projectEntryToWire(entry), 201);
+      const entry = addBenchmark(body.path);
+      return c4.json(benchmarkEntryToWire(entry), 201);
     } catch (err2) {
       return c4.json({ error: err2.message }, 400);
     }
   });
-  app2.delete("/api/projects/:projectId", (c4) => {
+  app2.delete("/api/benchmarks/:benchmarkId", (c4) => {
     if (readOnly) {
       return c4.json({ error: "Studio is running in read-only mode" }, 403);
     }
-    const removed = removeProject(c4.req.param("projectId") ?? "");
+    const removed = removeBenchmark(c4.req.param("benchmarkId") ?? "");
     if (!removed) return c4.json({ error: "Project not found" }, 404);
     return c4.json({ ok: true });
   });
-  app2.get("/api/projects/:projectId/summary", async (c4) => {
-    const project = getProject(c4.req.param("projectId") ?? "");
-    if (!project) return c4.json({ error: "Project not found" }, 404);
+  app2.get("/api/benchmarks/:benchmarkId/summary", async (c4) => {
+    const benchmark = getBenchmark(c4.req.param("benchmarkId") ?? "");
+    if (!benchmark) return c4.json({ error: "Project not found" }, 404);
     try {
-      const { runs: metas } = await listMergedResultFiles(project.path);
+      const { runs: metas } = await listMergedResultFiles(benchmark.path);
       const runCount = metas.length;
       const passRate = runCount > 0 ? metas.reduce((s, m) => s + m.passRate, 0) / runCount : 0;
       const lastRun = metas.length > 0 ? metas[0].timestamp : null;
       return c4.json({
-        id: project.id,
-        name: project.name,
-        path: project.path,
+        id: benchmark.id,
+        name: benchmark.name,
+        path: benchmark.path,
         run_count: runCount,
         pass_rate: passRate,
         last_run: lastRun
@@ -8268,24 +8280,24 @@ function createApp(results, resultDir, cwd, sourceFile, options) {
       return c4.json({ error: "Failed to read project" }, 500);
     }
   });
-  app2.post("/api/projects/discover", async (c4) => {
+  app2.post("/api/benchmarks/discover", async (c4) => {
     if (readOnly) {
       return c4.json({ error: "Studio is running in read-only mode" }, 403);
     }
     try {
       const body = await c4.req.json();
       if (!body.path) return c4.json({ error: "Missing path" }, 400);
-      const discovered = discoverProjects(body.path);
-      const registered = discovered.map((p) => projectEntryToWire(addProject(p)));
+      const discovered = discoverBenchmarks(body.path);
+      const registered = discovered.map((p) => benchmarkEntryToWire(addBenchmark(p)));
       return c4.json({ discovered: registered });
     } catch (err2) {
       return c4.json({ error: err2.message }, 400);
     }
   });
-  app2.get("/api/projects/all-runs", async (c4) => {
-    const registry = loadProjectRegistry();
+  app2.get("/api/benchmarks/all-runs", async (c4) => {
+    const registry = loadBenchmarkRegistry();
     const allRuns = [];
-    for (const p of registry.projects) {
+    for (const p of registry.benchmarks) {
       try {
         const { runs: metas } = await listMergedResultFiles(p.path);
         for (const m of metas) {
@@ -8411,8 +8423,8 @@ function createApp(results, resultDir, cwd, sourceFile, options) {
     return c4.json({ entries: entries2 });
   });
   app2.get(
-    "/api/projects/:projectId/config",
-    (c4) => withProject(
+    "/api/benchmarks/:benchmarkId/config",
+    (c4) => withBenchmark(
       c4,
       (ctx, dataCtx) => handleConfig(ctx, dataCtx, {
         readOnly,
@@ -8421,50 +8433,53 @@ function createApp(results, resultDir, cwd, sourceFile, options) {
     )
   );
   app2.get(
-    "/api/projects/:projectId/remote/status",
-    (c4) => withProject(
+    "/api/benchmarks/:benchmarkId/remote/status",
+    (c4) => withBenchmark(
       c4,
       async (ctx, dataCtx) => ctx.json(await getRemoteResultsStatus(dataCtx.searchDir))
     )
   );
   app2.post(
-    "/api/projects/:projectId/remote/sync",
-    (c4) => withProject(c4, async (ctx, dataCtx) => ctx.json(await syncRemoteResults(dataCtx.searchDir)))
+    "/api/benchmarks/:benchmarkId/remote/sync",
+    (c4) => withBenchmark(c4, async (ctx, dataCtx) => ctx.json(await syncRemoteResults(dataCtx.searchDir)))
+  );
+  app2.get("/api/benchmarks/:benchmarkId/runs", (c4) => withBenchmark(c4, handleRuns));
+  app2.get("/api/benchmarks/:benchmarkId/runs/:filename", (c4) => withBenchmark(c4, handleRunDetail));
+  app2.get(
+    "/api/benchmarks/:benchmarkId/runs/:filename/suites",
+    (c4) => withBenchmark(c4, handleRunSuites)
   );
-  app2.get("/api/projects/:projectId/runs", (c4) => withProject(c4, handleRuns));
-  app2.get("/api/projects/:projectId/runs/:filename", (c4) => withProject(c4, handleRunDetail));
-  app2.get("/api/projects/:projectId/runs/:filename/suites", (c4) => withProject(c4, handleRunSuites));
   app2.get(
-    "/api/projects/:projectId/runs/:filename/categories",
-    (c4) => withProject(c4, handleRunCategories)
+    "/api/benchmarks/:benchmarkId/runs/:filename/categories",
+    (c4) => withBenchmark(c4, handleRunCategories)
   );
   app2.get(
-    "/api/projects/:projectId/runs/:filename/categories/:category/suites",
-    (c4) => withProject(c4, handleCategorySuites)
+    "/api/benchmarks/:benchmarkId/runs/:filename/categories/:category/suites",
+    (c4) => withBenchmark(c4, handleCategorySuites)
   );
   app2.get(
-    "/api/projects/:projectId/runs/:filename/evals/:evalId",
-    (c4) => withProject(c4, handleEvalDetail)
+    "/api/benchmarks/:benchmarkId/runs/:filename/evals/:evalId",
+    (c4) => withBenchmark(c4, handleEvalDetail)
   );
   app2.get(
-    "/api/projects/:projectId/runs/:filename/evals/:evalId/files",
-    (c4) => withProject(c4, handleEvalFiles)
+    "/api/benchmarks/:benchmarkId/runs/:filename/evals/:evalId/files",
+    (c4) => withBenchmark(c4, handleEvalFiles)
   );
   app2.get(
-    "/api/projects/:projectId/runs/:filename/evals/:evalId/files/*",
-    (c4) => withProject(c4, handleEvalFileContent)
+    "/api/benchmarks/:benchmarkId/runs/:filename/evals/:evalId/files/*",
+    (c4) => withBenchmark(c4, handleEvalFileContent)
   );
-  app2.get("/api/projects/:projectId/experiments", (c4) => withProject(c4, handleExperiments));
-  app2.get("/api/projects/:projectId/compare", (c4) => withProject(c4, handleCompare));
-  app2.get("/api/projects/:projectId/targets", (c4) => withProject(c4, handleTargets));
-  app2.get("/api/projects/:projectId/feedback", (c4) => withProject(c4, handleFeedbackRead));
+  app2.get("/api/benchmarks/:benchmarkId/experiments", (c4) => withBenchmark(c4, handleExperiments));
+  app2.get("/api/benchmarks/:benchmarkId/compare", (c4) => withBenchmark(c4, handleCompare));
+  app2.get("/api/benchmarks/:benchmarkId/targets", (c4) => withBenchmark(c4, handleTargets));
+  app2.get("/api/benchmarks/:benchmarkId/feedback", (c4) => withBenchmark(c4, handleFeedbackRead));
   registerEvalRoutes(
     app2,
     (c4) => {
-      const projectId = c4.req.param("projectId");
-      if (projectId) {
-        const project = getProject(projectId);
-        if (project) return project.path;
+      const benchmarkId = c4.req.param("benchmarkId");
+      if (benchmarkId) {
+        const benchmark = getBenchmark(benchmarkId);
+        if (benchmark) return benchmark.path;
       }
       return searchDir;
     },
@@ -8512,7 +8527,7 @@ function createApp(results, resultDir, cwd, sourceFile, options) {
   return app2;
 }
 function resolveStudioDistDir() {
-  const currentDir = typeof __dirname !== "undefined" ? __dirname : path16.dirname(fileURLToPath2(import.meta.url));
+  const currentDir = typeof __dirname !== "undefined" ? __dirname : path16.dirname(fileURLToPath3(import.meta.url));
   const candidates = [
     // From src/commands/results/ → sibling apps/studio/dist
     path16.resolve(currentDir, "../../../../studio/dist"),
@@ -8584,7 +8599,7 @@ var resultsServeCommand = command({
     const listenPort = port ?? (process.env.PORT ? Number(process.env.PORT) : 3117);
     if (add) {
       try {
-        const entry = addProject(add);
+        const entry = addBenchmark(add);
         console.log(`Registered project: ${entry.name} (${entry.id}) at ${entry.path}`);
       } catch (err2) {
         console.error(`Error: ${err2.message}`);
@@ -8593,7 +8608,7 @@ var resultsServeCommand = command({
       return;
     }
     if (remove) {
-      const removed = removeProject(remove);
+      const removed = removeBenchmark(remove);
       if (removed) {
         console.log(`Unregistered project: ${remove}`);
       } else {
@@ -8603,21 +8618,21 @@ var resultsServeCommand = command({
       return;
     }
     if (discover) {
-      const discovered = discoverProjects(discover);
+      const discovered = discoverBenchmarks(discover);
       if (discovered.length === 0) {
         console.log(`No projects with .agentv/ found under ${discover}`);
         return;
       }
       for (const p of discovered) {
-        const entry = addProject(p);
+        const entry = addBenchmark(p);
         console.log(`Registered: ${entry.name} (${entry.id}) at ${entry.path}`);
       }
       console.log(`
 Discovered ${discovered.length} project(s).`);
       return;
     }
-    const registry = loadProjectRegistry();
-    const { isMultiProject, showMultiWarning } = resolveDashboardMode(registry.projects.length, {
+    const registry = loadBenchmarkRegistry();
+    const { isMultiProject, showMultiWarning } = resolveDashboardMode(registry.benchmarks.length, {
       multi,
       single
     });
@@ -8652,7 +8667,7 @@ Discovered ${discovered.length} project(s).`);
         );
       }
       if (isMultiProject) {
-        console.log(`Multi-project mode: ${registry.projects.length} project(s) registered`);
+        console.log(`Multi-project mode: ${registry.benchmarks.length} project(s) registered`);
       } else if (results.length > 0 && sourceFile) {
         console.log(`Serving ${results.length} result(s) from ${sourceFile}`);
       } else {
@@ -8660,7 +8675,7 @@ Discovered ${discovered.length} project(s).`);
         console.log("Run an evaluation to see results: agentv eval <eval-file>");
       }
       console.log(`Dashboard: http://localhost:${listenPort}`);
-      console.log(`Projects API: http://localhost:${listenPort}/api/projects`);
+      console.log(`Benchmarks API: http://localhost:${listenPort}/api/benchmarks`);
       console.log("Press Ctrl+C to stop");
       const { serve: startServer } = await import("@hono/node-server");
       startServer({
@@ -9835,4 +9850,4 @@ export {
   preprocessArgv,
   runCli
 };
-//# sourceMappingURL=chunk-7TJ2PON3.js.map
+//# sourceMappingURL=chunk-VVWPD4CN.js.map