agentv 4.28.0 → 4.29.0-next.1

package/dist/{chunk-2WS3BEPV.js → chunk-GLJVO5PK.js}

@@ -45,7 +45,7 @@ import {
   validateFileReferences,
   validateTargetsFile,
   validateWorkspacePaths
-} from "./chunk-NL5H3TIC.js";
+} from "./chunk-FEDIWLKK.js";
 import {
   RESULT_INDEX_FILENAME,
   aggregateRunDir,
@@ -53,27 +53,27 @@ import {
   resolveRunManifestPath,
   toSnakeCaseDeep as toSnakeCaseDeep2,
   writeArtifactsFromResults
-} from "./chunk-7T6AF75O.js";
+} from "./chunk-OS67VZUO.js";
 import {
   DEFAULT_CATEGORY,
-  addBenchmark,
+  addProject,
   deriveCategory,
   discoverClaudeSessions,
   discoverCodexSessions,
-  getBenchmark,
   getOutputFilenames,
-  loadBenchmarkRegistry,
+  getProject,
+  loadProjectRegistry,
   parseClaudeSession,
   parseCodexSession,
   readTranscriptFile,
-  removeBenchmark,
+  removeProject,
   runBeforeSessionHook,
   scanRepoDeps,
-  syncBenchmarks,
+  syncProjects,
   toTranscriptJsonLines,
   transpileEvalYamlFile,
   trimBaselineResult
-} from "./chunk-WZVOY2W2.js";
+} from "./chunk-VZMGBDJD.js";
 import {
   DEFAULT_THRESHOLD,
   createBuiltinRegistry,
@@ -4053,7 +4053,7 @@ var evalRunCommand = command({
   },
   handler: async (args) => {
     if (args.evalPaths.length === 0 && process.stdin.isTTY) {
-      const { launchInteractiveWizard } = await import("./interactive-7R2K3CBY.js");
+      const { launchInteractiveWizard } = await import("./interactive-5LEM6ITD.js");
       await launchInteractiveWizard();
       return;
     }
@@ -4121,7 +4121,7 @@ var evalRunCommand = command({
 // src/commands/eval/index.ts
 var evalCommand = subcommands({
   name: "eval",
-  description: "Evaluation commands",
+  description: "Evaluation commands. Shorthand: `agentv eval <eval-paths...>` aliases `agentv eval run <eval-paths...>`.",
   cmds: {
     run: evalRunCommand,
     assert: evalAssertCommand,
@@ -9338,7 +9338,7 @@ import { Hono } from "hono";
 
 // src/commands/results/eval-runner.ts
 import { execFileSync as execFileSync2, spawn } from "node:child_process";
-import { existsSync as existsSync12 } from "node:fs";
+import { createWriteStream, existsSync as existsSync12, mkdirSync as mkdirSync3 } from "node:fs";
 import path17 from "node:path";
 import { fileURLToPath as fileURLToPath2 } from "node:url";
 var activeRuns = /* @__PURE__ */ new Map();
@@ -9365,6 +9365,14 @@ function getActiveRunTarget(indexJsonlPath) {
   }
   return void 0;
 }
+function getActiveRunStatus(indexJsonlPath) {
+  for (const run2 of activeRuns.values()) {
+    if (run2.outputDir && path17.join(run2.outputDir, "index.jsonl") === indexJsonlPath) {
+      return run2.status;
+    }
+  }
+  return void 0;
+}
 async function discoverTargetsInProject(cwd) {
   const repoRoot = await findRepoRoot(cwd) ?? cwd;
   let targetsFilePath;
@@ -9477,6 +9485,17 @@ function isCommandAvailable(cmd) {
     return false;
   }
 }
+function openConsoleLogStream(outputDir) {
+  try {
+    mkdirSync3(outputDir, { recursive: true });
+    const stream = createWriteStream(path17.join(outputDir, "console.log"), { flags: "w" });
+    stream.on("error", () => {
+    });
+    return stream;
+  } catch {
+    return void 0;
+  }
+}
 function registerEvalRoutes(app2, getCwd, options) {
   const readOnly = options?.readOnly === true;
   app2.get("/api/eval/discover", async (c4) => {
@@ -9553,13 +9572,16 @@ function registerEvalRoutes(app2, getCwd, options) {
       });
       run2.process = child;
       run2.status = "running";
+      const logStream = openConsoleLogStream(outputDir);
       child.stdout?.on("data", (chunk) => {
+        logStream?.write(chunk);
         run2.stdout += chunk.toString();
         if (run2.stdout.length > 1e5) {
           run2.stdout = run2.stdout.slice(-8e4);
         }
       });
       child.stderr?.on("data", (chunk) => {
+        logStream?.write(chunk);
         run2.stderr += chunk.toString();
         if (run2.stderr.length > 1e5) {
           run2.stderr = run2.stderr.slice(-8e4);
@@ -9570,6 +9592,7 @@ function registerEvalRoutes(app2, getCwd, options) {
         run2.status = code === 0 ? "finished" : "failed";
         run2.finishedAt = (/* @__PURE__ */ new Date()).toISOString();
         run2.process = void 0;
+        logStream?.end();
         pruneFinishedRuns();
       });
       child.on("error", (err2) => {
@@ -9578,6 +9601,10 @@ function registerEvalRoutes(app2, getCwd, options) {
 Process error: ${err2.message}`;
         run2.finishedAt = (/* @__PURE__ */ new Date()).toISOString();
         run2.process = void 0;
+        logStream?.write(`
+Process error: ${err2.message}
+`);
+        logStream?.end();
       });
       return c4.json(
         {
@@ -9649,7 +9676,7 @@ Process error: ${err2.message}`;
     const args = buildCliArgs(body);
     return c4.json({ command: buildCliPreview(args) });
   });
-  app2.get("/api/benchmarks/:benchmarkId/eval/discover", async (c4) => {
+  app2.get("/api/projects/:projectId/eval/discover", async (c4) => {
     const cwd = getCwd(c4);
     try {
       const files = await discoverEvalFiles(cwd);
@@ -9664,7 +9691,7 @@ Process error: ${err2.message}`;
       return c4.json({ error: err2.message, eval_files: [] }, 500);
     }
   });
-  app2.get("/api/benchmarks/:benchmarkId/eval/targets", async (c4) => {
+  app2.get("/api/projects/:projectId/eval/targets", async (c4) => {
     const cwd = getCwd(c4);
     try {
       const names = await discoverTargetsInProject(cwd);
@@ -9673,7 +9700,7 @@ Process error: ${err2.message}`;
       return c4.json({ error: err2.message, targets: [] }, 500);
     }
   });
-  app2.post("/api/benchmarks/:benchmarkId/eval/run", async (c4) => {
+  app2.post("/api/projects/:projectId/eval/run", async (c4) => {
     if (readOnly) {
       return c4.json({ error: "Studio is running in read-only mode" }, 403);
     }
@@ -9722,11 +9749,14 @@ Process error: ${err2.message}`;
       });
       run2.process = child;
       run2.status = "running";
+      const logStream = openConsoleLogStream(outputDir);
       child.stdout?.on("data", (chunk) => {
+        logStream?.write(chunk);
         run2.stdout += chunk.toString();
         if (run2.stdout.length > 1e5) run2.stdout = run2.stdout.slice(-8e4);
       });
       child.stderr?.on("data", (chunk) => {
+        logStream?.write(chunk);
         run2.stderr += chunk.toString();
         if (run2.stderr.length > 1e5) run2.stderr = run2.stderr.slice(-8e4);
       });
@@ -9735,6 +9765,7 @@ Process error: ${err2.message}`;
         run2.status = code === 0 ? "finished" : "failed";
         run2.finishedAt = (/* @__PURE__ */ new Date()).toISOString();
         run2.process = void 0;
+        logStream?.end();
         pruneFinishedRuns();
       });
       child.on("error", (err2) => {
@@ -9743,6 +9774,10 @@ Process error: ${err2.message}`;
 Process error: ${err2.message}`;
         run2.finishedAt = (/* @__PURE__ */ new Date()).toISOString();
         run2.process = void 0;
+        logStream?.write(`
+Process error: ${err2.message}
+`);
+        logStream?.end();
       });
       return c4.json({ id: runId, status: run2.status, command: command2 }, 202);
     } catch (err2) {
@@ -9752,7 +9787,7 @@ Process error: ${err2.message}`;
       return c4.json({ error: err2.message }, 500);
     }
   });
-  app2.post("/api/benchmarks/:benchmarkId/eval/run/:id/stop", (c4) => {
+  app2.post("/api/projects/:projectId/eval/run/:id/stop", (c4) => {
     if (readOnly) {
       return c4.json({ error: "Studio is running in read-only mode" }, 403);
     }
@@ -9769,7 +9804,7 @@ Process error: ${err2.message}`;
     }
     return c4.json({ stopped: true, status: run2.status });
   });
-  app2.get("/api/benchmarks/:benchmarkId/eval/status/:id", (c4) => {
+  app2.get("/api/projects/:projectId/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);
@@ -9784,7 +9819,7 @@ Process error: ${err2.message}`;
       stderr: run2.stderr.slice(-5e3)
     });
   });
-  app2.get("/api/benchmarks/:benchmarkId/eval/runs", (c4) => {
+  app2.get("/api/projects/:projectId/eval/runs", (c4) => {
     const runs = [...activeRuns.values()].map((r) => ({
       id: r.id,
       status: r.status,
@@ -9797,7 +9832,7 @@ Process error: ${err2.message}`;
     runs.sort((a, b) => b.started_at.localeCompare(a.started_at));
     return c4.json({ runs });
   });
-  app2.post("/api/benchmarks/:benchmarkId/eval/preview", async (c4) => {
+  app2.post("/api/projects/:projectId/eval/preview", async (c4) => {
     let body;
     try {
       body = await c4.req.json();
@@ -9887,7 +9922,7 @@ function normalizeTags(tags) {
 }
 
 // src/commands/results/studio-config.ts
-import { existsSync as existsSync14, mkdirSync as mkdirSync3, readFileSync as readFileSync11, writeFileSync as writeFileSync5 } from "node:fs";
+import { existsSync as existsSync14, mkdirSync as mkdirSync4, readFileSync as readFileSync11, writeFileSync as writeFileSync5 } from "node:fs";
 import path19 from "node:path";
 import { stringify as stringifyYaml2 } from "yaml";
 var DEFAULTS = {
@@ -9921,7 +9956,7 @@ function loadStudioConfig(agentvDir) {
 }
 function saveStudioConfig(agentvDir, config) {
   if (!existsSync14(agentvDir)) {
-    mkdirSync3(agentvDir, { recursive: true });
+    mkdirSync4(agentvDir, { recursive: true });
   }
   const configPath = path19.join(agentvDir, "config.yaml");
   let existing = {};
@@ -9977,14 +10012,11 @@ Serving most recent: ${metas[0].path}
   }
   return metas[0].path;
 }
-function resolveDashboardMode(benchmarkCount, options) {
+function resolveDashboardMode(projectCount, options) {
   if (options.single === true) {
-    return { isMultiBenchmark: false, showMultiWarning: options.multi === true };
+    return { projectDashboard: false };
   }
-  if (options.multi === true) {
-    return { isMultiBenchmark: true, showMultiWarning: true };
-  }
-  return { isMultiBenchmark: benchmarkCount > 1, showMultiWarning: false };
+  return { projectDashboard: projectCount > 1 };
 }
 function feedbackPath(resultDir) {
   return path20.join(resultDir, "feedback.json");
@@ -10096,6 +10128,7 @@ async function handleRuns(c4, { searchDir, agentvDir }) {
         }
       } catch {
       }
+      const liveStatus = getActiveRunStatus(m.path);
       const tagsEntry = readRunTags(m.path);
       return {
         filename: m.filename,
@@ -10109,11 +10142,30 @@ async function handleRuns(c4, { searchDir, agentvDir }) {
         source: m.source,
         ...target && { target },
         ...experiment && { experiment },
-        ...tagsEntry && { tags: tagsEntry.tags }
+        ...tagsEntry && { tags: tagsEntry.tags },
+        ...liveStatus && { status: liveStatus }
       };
     })
   });
 }
+async function handleRunLog(c4, { searchDir }) {
+  const filename = c4.req.param("filename") ?? "";
+  const meta = await findRunById(searchDir, filename);
+  if (!meta) return c4.json({ error: "Run not found" }, 404);
+  if (meta.source === "remote") {
+    return c4.json({ error: "Console log is not available for remote runs" }, 404);
+  }
+  const logPath = path20.join(path20.dirname(meta.path), "console.log");
+  if (!existsSync15(logPath)) {
+    return c4.json({ error: "Console log not found for this run" }, 404);
+  }
+  try {
+    const content = readFileSync12(logPath, "utf8");
+    return c4.text(content);
+  } catch {
+    return c4.json({ error: "Failed to read console log" }, 500);
+  }
+}
 async function handleRunDetail(c4, { searchDir }) {
   const filename = c4.req.param("filename") ?? "";
   const meta = await findRunById(searchDir, filename);
@@ -10523,8 +10575,8 @@ function handleConfig(c4, { agentvDir, searchDir }, options) {
   return c4.json({
     ...loadStudioConfig(agentvDir),
     read_only: options?.readOnly === true,
-    benchmark_name: path20.basename(searchDir),
-    multi_benchmark_dashboard: options?.multiBenchmarkDashboard === true
+    project_name: path20.basename(searchDir),
+    project_dashboard: options?.projectDashboard === true
   });
 }
 function handleFeedbackRead(c4, { searchDir }) {
@@ -10581,14 +10633,14 @@ function createApp(results, resultDir, cwd, sourceFile, options) {
   const defaultCtx = { searchDir, agentvDir };
   const readOnly = options?.readOnly === true;
   const app2 = new Hono();
-  function withBenchmark(c4, handler) {
-    const benchmark = getBenchmark(c4.req.param("benchmarkId") ?? "");
-    if (!benchmark || !existsSync15(benchmark.path)) {
-      return c4.json({ error: "Benchmark not found" }, 404);
+  function withProject(c4, handler) {
+    const project = getProject(c4.req.param("projectId") ?? "");
+    if (!project || !existsSync15(project.path)) {
+      return c4.json({ error: "Project not found" }, 404);
     }
     return handler(c4, {
-      searchDir: benchmark.path,
-      agentvDir: path20.join(benchmark.path, ".agentv")
+      searchDir: project.path,
+      agentvDir: path20.join(project.path, ".agentv")
     });
   }
   app2.post("/api/config", async (c4) => {
@@ -10608,7 +10660,7 @@ function createApp(results, resultDir, cwd, sourceFile, options) {
       return c4.json({ error: "Failed to save config" }, 500);
     }
   });
-  function benchmarkEntryToWire(entry) {
+  function projectEntryToWire(entry) {
     return {
       id: entry.id,
       name: entry.name,
@@ -10617,10 +10669,10 @@ function createApp(results, resultDir, cwd, sourceFile, options) {
       last_opened_at: entry.lastOpenedAt
     };
   }
-  app2.get("/api/benchmarks", async (c4) => {
-    const registry = loadBenchmarkRegistry();
-    const benchmarks = await Promise.all(
-      registry.benchmarks.map(async (p) => {
+  app2.get("/api/projects", async (c4) => {
+    const registry = loadProjectRegistry();
+    const projects = await Promise.all(
+      registry.projects.map(async (p) => {
         let runCount = 0;
         let passRate = 0;
         let lastRun = null;
@@ -10635,52 +10687,52 @@ function createApp(results, resultDir, cwd, sourceFile, options) {
         } catch {
         }
         return {
-          ...benchmarkEntryToWire(p),
+          ...projectEntryToWire(p),
           run_count: runCount,
           pass_rate: passRate,
           last_run: lastRun
         };
       })
     );
-    return c4.json({ benchmarks });
+    return c4.json({ projects });
   });
-  app2.post("/api/benchmarks", async (c4) => {
+  app2.post("/api/projects", 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 = addBenchmark(body.path);
-      return c4.json(benchmarkEntryToWire(entry), 201);
+      const entry = addProject(body.path);
+      return c4.json(projectEntryToWire(entry), 201);
     } catch (err2) {
       return c4.json({ error: err2.message }, 400);
     }
   });
-  app2.get("/api/benchmarks/:benchmarkId/summary", async (c4) => {
-    const benchmark = getBenchmark(c4.req.param("benchmarkId") ?? "");
-    if (!benchmark) return c4.json({ error: "Benchmark not found" }, 404);
+  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);
     try {
-      const { runs: metas } = await listMergedResultFiles(benchmark.path);
+      const { runs: metas } = await listMergedResultFiles(project.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: benchmark.id,
-        name: benchmark.name,
-        path: benchmark.path,
+        id: project.id,
+        name: project.name,
+        path: project.path,
         run_count: runCount,
         pass_rate: passRate,
         last_run: lastRun
       });
     } catch {
-      return c4.json({ error: "Failed to read benchmark" }, 500);
+      return c4.json({ error: "Failed to read project" }, 500);
     }
   });
-  app2.get("/api/benchmarks/all-runs", async (c4) => {
-    const registry = loadBenchmarkRegistry();
+  app2.get("/api/projects/all-runs", async (c4) => {
+    const registry = loadProjectRegistry();
     const allRuns = [];
-    for (const p of registry.benchmarks) {
+    for (const p of registry.projects) {
       try {
         const { runs: metas } = await listMergedResultFiles(p.path);
         for (const m of metas) {
@@ -10706,8 +10758,8 @@ function createApp(results, resultDir, cwd, sourceFile, options) {
             source: m.source,
             ...target && { target },
             ...experiment && { experiment },
-            benchmark_id: p.id,
-            benchmark_name: p.name
+            project_id: p.id,
+            project_name: p.name
           });
         }
       } catch {
@@ -10716,19 +10768,19 @@ function createApp(results, resultDir, cwd, sourceFile, options) {
     allRuns.sort((a, b) => b.timestamp.localeCompare(a.timestamp));
     return c4.json({ runs: allRuns });
   });
-  app2.delete("/api/benchmarks/:benchmarkId", (c4) => {
+  app2.delete("/api/projects/:projectId", (c4) => {
     if (readOnly) {
       return c4.json({ error: "Studio is running in read-only mode" }, 403);
     }
-    const removed = removeBenchmark(c4.req.param("benchmarkId") ?? "");
-    if (!removed) return c4.json({ error: "Benchmark not found" }, 404);
+    const removed = removeProject(c4.req.param("projectId") ?? "");
+    if (!removed) return c4.json({ error: "Project not found" }, 404);
     return c4.json({ ok: true });
   });
   app2.get(
     "/api/config",
     (c4) => handleConfig(c4, defaultCtx, {
       readOnly,
-      multiBenchmarkDashboard: options?.multiBenchmarkDashboard
+      projectDashboard: options?.projectDashboard
     })
   );
   app2.get("/api/remote/status", async (c4) => c4.json(await getRemoteResultsStatus(searchDir)));
@@ -10747,6 +10799,7 @@ function createApp(results, resultDir, cwd, sourceFile, options) {
     return handleRunTagsDelete(c4, defaultCtx);
   });
   app2.get("/api/runs/:filename", (c4) => handleRunDetail(c4, defaultCtx));
+  app2.get("/api/runs/:filename/log", (c4) => handleRunLog(c4, defaultCtx));
   app2.get("/api/runs/:filename/suites", (c4) => handleRunSuites(c4, defaultCtx));
   app2.get("/api/runs/:filename/categories", (c4) => handleRunCategories(c4, defaultCtx));
   app2.get(
@@ -10826,75 +10879,73 @@ function createApp(results, resultDir, cwd, sourceFile, options) {
     return c4.json({ entries: entries2 });
   });
   app2.get(
-    "/api/benchmarks/:benchmarkId/config",
-    (c4) => withBenchmark(
+    "/api/projects/:projectId/config",
+    (c4) => withProject(
       c4,
       (ctx, dataCtx) => handleConfig(ctx, dataCtx, {
         readOnly,
-        multiBenchmarkDashboard: options?.multiBenchmarkDashboard
+        projectDashboard: options?.projectDashboard
       })
     )
   );
   app2.get(
-    "/api/benchmarks/:benchmarkId/remote/status",
-    (c4) => withBenchmark(
+    "/api/projects/:projectId/remote/status",
+    (c4) => withProject(
       c4,
       async (ctx, dataCtx) => ctx.json(await getRemoteResultsStatus(dataCtx.searchDir))
     )
   );
   app2.post(
-    "/api/benchmarks/:benchmarkId/remote/sync",
-    (c4) => withBenchmark(c4, async (ctx, dataCtx) => ctx.json(await syncRemoteResults(dataCtx.searchDir)))
+    "/api/projects/:projectId/remote/sync",
+    (c4) => withProject(c4, async (ctx, dataCtx) => ctx.json(await syncRemoteResults(dataCtx.searchDir)))
   );
-  app2.get("/api/benchmarks/:benchmarkId/runs", (c4) => withBenchmark(c4, handleRuns));
-  app2.put("/api/benchmarks/:benchmarkId/runs/:filename/tags", (c4) => {
+  app2.get("/api/projects/:projectId/runs", (c4) => withProject(c4, handleRuns));
+  app2.put("/api/projects/:projectId/runs/:filename/tags", (c4) => {
     if (readOnly) {
       return c4.json({ error: "Studio is running in read-only mode" }, 403);
     }
-    return withBenchmark(c4, handleRunTagsPut);
+    return withProject(c4, handleRunTagsPut);
   });
-  app2.delete("/api/benchmarks/:benchmarkId/runs/:filename/tags", (c4) => {
+  app2.delete("/api/projects/:projectId/runs/:filename/tags", (c4) => {
     if (readOnly) {
       return c4.json({ error: "Studio is running in read-only mode" }, 403);
     }
-    return withBenchmark(c4, handleRunTagsDelete);
+    return withProject(c4, handleRunTagsDelete);
   });
-  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/:filename", (c4) => withProject(c4, handleRunDetail));
+  app2.get("/api/projects/:projectId/runs/:filename/log", (c4) => withProject(c4, handleRunLog));
+  app2.get("/api/projects/:projectId/runs/:filename/suites", (c4) => withProject(c4, handleRunSuites));
   app2.get(
-    "/api/benchmarks/:benchmarkId/runs/:filename/categories",
-    (c4) => withBenchmark(c4, handleRunCategories)
+    "/api/projects/:projectId/runs/:filename/categories",
+    (c4) => withProject(c4, handleRunCategories)
   );
   app2.get(
-    "/api/benchmarks/:benchmarkId/runs/:filename/categories/:category/suites",
-    (c4) => withBenchmark(c4, handleCategorySuites)
+    "/api/projects/:projectId/runs/:filename/categories/:category/suites",
+    (c4) => withProject(c4, handleCategorySuites)
   );
   app2.get(
-    "/api/benchmarks/:benchmarkId/runs/:filename/evals/:evalId",
-    (c4) => withBenchmark(c4, handleEvalDetail)
+    "/api/projects/:projectId/runs/:filename/evals/:evalId",
+    (c4) => withProject(c4, handleEvalDetail)
   );
   app2.get(
-    "/api/benchmarks/:benchmarkId/runs/:filename/evals/:evalId/files",
-    (c4) => withBenchmark(c4, handleEvalFiles)
+    "/api/projects/:projectId/runs/:filename/evals/:evalId/files",
+    (c4) => withProject(c4, handleEvalFiles)
   );
   app2.get(
-    "/api/benchmarks/:benchmarkId/runs/:filename/evals/:evalId/files/*",
-    (c4) => withBenchmark(c4, handleEvalFileContent)
+    "/api/projects/:projectId/runs/:filename/evals/:evalId/files/*",
+    (c4) => withProject(c4, handleEvalFileContent)
   );
-  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));
+  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));
   registerEvalRoutes(
     app2,
     (c4) => {
-      const benchmarkId = c4.req.param("benchmarkId");
-      if (benchmarkId) {
-        const benchmark = getBenchmark(benchmarkId);
-        if (benchmark) return benchmark.path;
+      const projectId = c4.req.param("projectId");
+      if (projectId) {
+        const project = getProject(projectId);
+        if (project) return project.path;
       }
       return searchDir;
     },
@@ -10981,36 +11032,32 @@ var resultsServeCommand = command({
       short: "d",
       description: "Working directory (default: current directory)"
     }),
-    multi: flag({
-      long: "multi",
-      description: "Launch in multi-benchmark dashboard mode (deprecated; use auto-detect or --single)"
-    }),
     single: flag({
       long: "single",
-      description: "Force single-benchmark dashboard mode"
+      description: "Force single-project dashboard mode"
     }),
     add: option({
       type: optional(string),
       long: "add",
-      description: "Register a benchmark by path"
+      description: "Register a project by path"
     }),
     remove: option({
       type: optional(string),
       long: "remove",
-      description: "Unregister a benchmark by ID"
+      description: "Unregister a project by ID"
     }),
     readOnly: flag({
       long: "read-only",
       description: "Disable write operations and launch Studio in read-only leaderboard mode"
     })
   },
-  handler: async ({ source, port, dir, multi, single, add, remove, readOnly }) => {
+  handler: async ({ source, port, dir, single, add, remove, readOnly }) => {
     const cwd = dir ?? process.cwd();
     const listenPort = port ?? (process.env.PORT ? Number(process.env.PORT) : 3117);
     if (add) {
       try {
-        const entry = addBenchmark(add);
-        console.log(`Registered benchmark: ${entry.name} (${entry.id}) at ${entry.path}`);
+        const entry = addProject(add);
+        console.log(`Registered project: ${entry.name} (${entry.id}) at ${entry.path}`);
       } catch (err2) {
         console.error(`Error: ${err2.message}`);
         process.exit(1);
@@ -11018,11 +11065,11 @@ var resultsServeCommand = command({
       return;
     }
     if (remove) {
-      const removed = removeBenchmark(remove);
+      const removed = removeProject(remove);
       if (removed) {
-        console.log(`Unregistered benchmark: ${remove}`);
+        console.log(`Unregistered project: ${remove}`);
       } else {
-        console.error(`Benchmark not found: ${remove}`);
+        console.error(`Project not found: ${remove}`);
         process.exit(1);
       }
       return;
@@ -11032,12 +11079,9 @@ var resultsServeCommand = command({
     if (yamlConfig?.required_version) {
       await enforceRequiredVersion(yamlConfig.required_version);
     }
-    const registry = loadBenchmarkRegistry();
-    const { isMultiBenchmark, showMultiWarning } = resolveDashboardMode(
-      registry.benchmarks.length,
-      { multi, single }
-    );
-    await syncBenchmarks(registry.benchmarks);
+    const registry = loadProjectRegistry();
+    const { projectDashboard } = resolveDashboardMode(registry.projects.length, { single });
+    await syncProjects(registry.projects);
     try {
       let results = [];
       let sourceFile;
@@ -11061,15 +11105,10 @@ var resultsServeCommand = command({
       const resultDir = sourceFile ? path20.dirname(path20.resolve(sourceFile)) : cwd;
       const app2 = createApp(results, resultDir, cwd, sourceFile, {
         readOnly,
-        multiBenchmarkDashboard: isMultiBenchmark
+        projectDashboard
       });
-      if (showMultiWarning) {
-        console.warn(
-          "Warning: --multi is deprecated. Studio now auto-detects multi-benchmark mode when multiple benchmarks are registered. Use --single to force the single-benchmark view."
-        );
-      }
-      if (isMultiBenchmark) {
-        console.log(`Multi-benchmark mode: ${registry.benchmarks.length} benchmark(s) registered`);
+      if (projectDashboard) {
+        console.log(`Project dashboard: ${registry.projects.length} project(s) registered`);
       } else if (results.length > 0 && sourceFile) {
         console.log(`Serving ${results.length} result(s) from ${sourceFile}`);
       } else {
@@ -11077,7 +11116,7 @@ var resultsServeCommand = command({
         console.log("Run an evaluation to see results: agentv eval <eval-file>");
       }
       console.log(`Dashboard: http://localhost:${listenPort}`);
-      console.log(`Benchmarks API: http://localhost:${listenPort}/api/benchmarks`);
+      console.log(`Projects API: http://localhost:${listenPort}/api/projects`);
       console.log("Press Ctrl+C to stop");
       const { serve: startServer } = await import("@hono/node-server");
       startServer({
@@ -12549,4 +12588,4 @@ export {
   preprocessArgv,
   runCli
 };
-//# sourceMappingURL=chunk-2WS3BEPV.js.map
+//# sourceMappingURL=chunk-GLJVO5PK.js.map