agentv 4.3.3 → 4.4.0

package/dist/{chunk-5DDVNHOS.js → chunk-YORCRL4G.js}

@@ -24,27 +24,32 @@ import {
   validateFileReferences,
   validateTargetsFile,
   writeArtifactsFromResults
-} from "./chunk-EW4COQU2.js";
+} from "./chunk-BAYNXTX6.js";
 import {
   DEFAULT_CATEGORY,
   PASS_THRESHOLD,
+  addProject,
   createBuiltinRegistry,
   deriveCategory,
+  discoverProjects,
   executeScript,
   getAgentvHome,
   getOutputFilenames,
+  getProject,
   getWorkspacePoolRoot,
   isAgentSkillsFormat,
+  loadProjectRegistry,
   loadTestById,
   loadTestSuite,
   loadTests,
   normalizeLineEndings,
   parseAgentSkillsEvals,
+  removeProject,
   toCamelCaseDeep,
   toSnakeCaseDeep as toSnakeCaseDeep2,
   transpileEvalYamlFile,
   trimBaselineResult
-} from "./chunk-HMOXP7T5.js";
+} from "./chunk-63NDZ6UC.js";
 import {
   __commonJS,
   __esm,
@@ -4218,7 +4223,7 @@ var evalRunCommand = command({
   },
   handler: async (args) => {
     if (args.evalPaths.length === 0 && process.stdin.isTTY) {
-      const { launchInteractiveWizard } = await import("./interactive-6BO4RY6U.js");
+      const { launchInteractiveWizard } = await import("./interactive-YNSOO2BS.js");
       await launchInteractiveWizard();
       return;
     }
@@ -6255,7 +6260,16 @@ function loadStudioConfig(agentvDir) {
   if (!parsed || typeof parsed !== "object") {
     return { ...DEFAULTS };
   }
-  const threshold = typeof parsed.pass_threshold === "number" ? parsed.pass_threshold : DEFAULTS.pass_threshold;
+  const studio = parsed.studio;
+  let threshold = DEFAULTS.pass_threshold;
+  if (studio && typeof studio === "object" && !Array.isArray(studio)) {
+    const studioThreshold = studio.pass_threshold;
+    if (typeof studioThreshold === "number") {
+      threshold = studioThreshold;
+    }
+  } else if (typeof parsed.pass_threshold === "number") {
+    threshold = parsed.pass_threshold;
+  }
   return {
     pass_threshold: Math.min(1, Math.max(0, threshold))
   };
@@ -6265,7 +6279,18 @@ function saveStudioConfig(agentvDir, config) {
     mkdirSync2(agentvDir, { recursive: true });
   }
   const configPath = path9.join(agentvDir, "config.yaml");
-  const yamlStr = stringifyYaml2(config);
+  let existing = {};
+  if (existsSync7(configPath)) {
+    const raw = readFileSync8(configPath, "utf-8");
+    const parsed = parseYaml(raw);
+    if (parsed && typeof parsed === "object" && !Array.isArray(parsed)) {
+      existing = parsed;
+    }
+  }
+  const { pass_threshold: _, ...rest } = existing;
+  existing = rest;
+  existing.studio = { ...config };
+  const yamlStr = stringifyYaml2(existing);
   writeFileSync3(configPath, yamlStr, "utf-8");
 }
 
@@ -6289,11 +6314,366 @@ function writeFeedback(cwd, data) {
   writeFileSync4(feedbackPath(cwd), `${JSON.stringify(data, null, 2)}
 `, "utf8");
 }
+function buildFileTree(dirPath, relativeTo) {
+  if (!existsSync8(dirPath) || !statSync4(dirPath).isDirectory()) {
+    return [];
+  }
+  const entries2 = readdirSync3(dirPath, { withFileTypes: true });
+  return entries2.sort((a, b) => {
+    if (a.isDirectory() !== b.isDirectory()) return a.isDirectory() ? -1 : 1;
+    return a.name.localeCompare(b.name);
+  }).map((entry) => {
+    const fullPath = path10.join(dirPath, entry.name);
+    const relPath = path10.relative(relativeTo, fullPath);
+    if (entry.isDirectory()) {
+      return {
+        name: entry.name,
+        path: relPath,
+        type: "dir",
+        children: buildFileTree(fullPath, relativeTo)
+      };
+    }
+    return { name: entry.name, path: relPath, type: "file" };
+  });
+}
+function inferLanguage(filePath) {
+  const ext = path10.extname(filePath).toLowerCase();
+  const langMap = {
+    ".json": "json",
+    ".jsonl": "json",
+    ".ts": "typescript",
+    ".tsx": "typescript",
+    ".js": "javascript",
+    ".jsx": "javascript",
+    ".md": "markdown",
+    ".yaml": "yaml",
+    ".yml": "yaml",
+    ".log": "plaintext",
+    ".txt": "plaintext",
+    ".py": "python",
+    ".sh": "shell",
+    ".bash": "shell",
+    ".css": "css",
+    ".html": "html",
+    ".xml": "xml",
+    ".svg": "xml",
+    ".toml": "toml",
+    ".diff": "diff",
+    ".patch": "diff"
+  };
+  return langMap[ext] ?? "plaintext";
+}
+function stripHeavyFields(results) {
+  return results.map((r) => {
+    const { requests, trace, ...rest } = r;
+    const toolCalls = trace?.toolCalls && Object.keys(trace.toolCalls).length > 0 ? trace.toolCalls : void 0;
+    const graderDurationMs = (r.scores ?? []).reduce((sum, s) => sum + (s.durationMs ?? 0), 0);
+    return {
+      ...rest,
+      ...toolCalls && { _toolCalls: toolCalls },
+      ...graderDurationMs > 0 && { _graderDurationMs: graderDurationMs }
+    };
+  });
+}
+function handleRuns(c3, { searchDir }) {
+  const metas = listResultFiles(searchDir);
+  return c3.json({
+    runs: metas.map((m) => {
+      let target;
+      let experiment;
+      try {
+        const records = loadLightweightResults(m.path);
+        if (records.length > 0) {
+          target = records[0].target;
+          experiment = records[0].experiment;
+        }
+      } catch {
+      }
+      return {
+        filename: m.filename,
+        path: m.path,
+        timestamp: m.timestamp,
+        test_count: m.testCount,
+        pass_rate: m.passRate,
+        avg_score: m.avgScore,
+        size_bytes: m.sizeBytes,
+        ...target && { target },
+        ...experiment && { experiment }
+      };
+    })
+  });
+}
+function handleRunDetail(c3, { searchDir }) {
+  const filename = c3.req.param("filename");
+  const meta = listResultFiles(searchDir).find((m) => m.filename === filename);
+  if (!meta) return c3.json({ error: "Run not found" }, 404);
+  try {
+    const loaded = patchTestIds(loadManifestResults(meta.path));
+    return c3.json({ results: stripHeavyFields(loaded), source: meta.filename });
+  } catch {
+    return c3.json({ error: "Failed to load run" }, 500);
+  }
+}
+function handleRunDatasets(c3, { searchDir, agentvDir }) {
+  const filename = c3.req.param("filename");
+  const meta = listResultFiles(searchDir).find((m) => m.filename === filename);
+  if (!meta) return c3.json({ error: "Run not found" }, 404);
+  try {
+    const loaded = patchTestIds(loadManifestResults(meta.path));
+    const { pass_threshold } = loadStudioConfig(agentvDir);
+    const datasetMap = /* @__PURE__ */ new Map();
+    for (const r of loaded) {
+      const ds = r.dataset ?? r.target ?? "default";
+      const entry = datasetMap.get(ds) ?? { total: 0, passed: 0, scoreSum: 0 };
+      entry.total++;
+      if (r.score >= pass_threshold) entry.passed++;
+      entry.scoreSum += r.score;
+      datasetMap.set(ds, entry);
+    }
+    const datasets = [...datasetMap.entries()].map(([name, entry]) => ({
+      name,
+      total: entry.total,
+      passed: entry.passed,
+      failed: entry.total - entry.passed,
+      avg_score: entry.total > 0 ? entry.scoreSum / entry.total : 0
+    }));
+    return c3.json({ datasets });
+  } catch {
+    return c3.json({ error: "Failed to load datasets" }, 500);
+  }
+}
+function handleRunCategories(c3, { searchDir, agentvDir }) {
+  const filename = c3.req.param("filename");
+  const meta = listResultFiles(searchDir).find((m) => m.filename === filename);
+  if (!meta) return c3.json({ error: "Run not found" }, 404);
+  try {
+    const loaded = patchTestIds(loadManifestResults(meta.path));
+    const { pass_threshold } = loadStudioConfig(agentvDir);
+    const categoryMap = /* @__PURE__ */ new Map();
+    for (const r of loaded) {
+      const cat = r.category ?? DEFAULT_CATEGORY;
+      const entry = categoryMap.get(cat) ?? {
+        total: 0,
+        passed: 0,
+        scoreSum: 0,
+        datasets: /* @__PURE__ */ new Set()
+      };
+      entry.total++;
+      if (r.score >= pass_threshold) entry.passed++;
+      entry.scoreSum += r.score;
+      entry.datasets.add(r.dataset ?? r.target ?? "default");
+      categoryMap.set(cat, entry);
+    }
+    const categories = [...categoryMap.entries()].map(([name, entry]) => ({
+      name,
+      total: entry.total,
+      passed: entry.passed,
+      failed: entry.total - entry.passed,
+      avg_score: entry.total > 0 ? entry.scoreSum / entry.total : 0,
+      dataset_count: entry.datasets.size
+    }));
+    return c3.json({ categories });
+  } catch {
+    return c3.json({ error: "Failed to load categories" }, 500);
+  }
+}
+function handleCategoryDatasets(c3, { searchDir, agentvDir }) {
+  const filename = c3.req.param("filename");
+  const category = decodeURIComponent(c3.req.param("category") ?? "");
+  const meta = listResultFiles(searchDir).find((m) => m.filename === filename);
+  if (!meta) return c3.json({ error: "Run not found" }, 404);
+  try {
+    const loaded = patchTestIds(loadManifestResults(meta.path));
+    const { pass_threshold } = loadStudioConfig(agentvDir);
+    const filtered = loaded.filter((r) => (r.category ?? DEFAULT_CATEGORY) === category);
+    const datasetMap = /* @__PURE__ */ new Map();
+    for (const r of filtered) {
+      const ds = r.dataset ?? r.target ?? "default";
+      const entry = datasetMap.get(ds) ?? { total: 0, passed: 0, scoreSum: 0 };
+      entry.total++;
+      if (r.score >= pass_threshold) entry.passed++;
+      entry.scoreSum += r.score;
+      datasetMap.set(ds, entry);
+    }
+    const datasets = [...datasetMap.entries()].map(([name, entry]) => ({
+      name,
+      total: entry.total,
+      passed: entry.passed,
+      failed: entry.total - entry.passed,
+      avg_score: entry.total > 0 ? entry.scoreSum / entry.total : 0
+    }));
+    return c3.json({ datasets });
+  } catch {
+    return c3.json({ error: "Failed to load datasets" }, 500);
+  }
+}
+function handleEvalDetail(c3, { searchDir }) {
+  const filename = c3.req.param("filename");
+  const evalId = c3.req.param("evalId");
+  const meta = listResultFiles(searchDir).find((m) => m.filename === filename);
+  if (!meta) return c3.json({ error: "Run not found" }, 404);
+  try {
+    const loaded = patchTestIds(loadManifestResults(meta.path));
+    const result = loaded.find((r) => r.testId === evalId);
+    if (!result) return c3.json({ error: "Eval not found" }, 404);
+    return c3.json({ eval: result });
+  } catch {
+    return c3.json({ error: "Failed to load eval" }, 500);
+  }
+}
+function handleEvalFiles(c3, { searchDir }) {
+  const filename = c3.req.param("filename");
+  const evalId = c3.req.param("evalId");
+  const meta = listResultFiles(searchDir).find((m) => m.filename === filename);
+  if (!meta) return c3.json({ error: "Run not found" }, 404);
+  try {
+    const content = readFileSync9(meta.path, "utf8");
+    const records = parseResultManifest(content);
+    const record = records.find((r) => (r.test_id ?? r.eval_id) === evalId);
+    if (!record) return c3.json({ error: "Eval not found" }, 404);
+    const baseDir = path10.dirname(meta.path);
+    const knownPaths = [
+      record.grading_path,
+      record.timing_path,
+      record.input_path,
+      record.output_path,
+      record.response_path
+    ].filter((p) => !!p);
+    if (knownPaths.length === 0) return c3.json({ files: [] });
+    const artifactDirs = knownPaths.map((p) => path10.dirname(p));
+    let commonDir = artifactDirs[0];
+    for (const dir of artifactDirs) {
+      while (!dir.startsWith(commonDir)) {
+        commonDir = path10.dirname(commonDir);
+      }
+    }
+    const artifactAbsDir = path10.join(baseDir, commonDir);
+    const files = buildFileTree(artifactAbsDir, baseDir);
+    return c3.json({ files });
+  } catch {
+    return c3.json({ error: "Failed to load file tree" }, 500);
+  }
+}
+function handleEvalFileContent(c3, { searchDir }) {
+  const filename = c3.req.param("filename");
+  const evalId = c3.req.param("evalId");
+  const meta = listResultFiles(searchDir).find((m) => m.filename === filename);
+  if (!meta) return c3.json({ error: "Run not found" }, 404);
+  const marker = `/runs/${filename}/evals/${evalId}/files/`;
+  const markerIdx = c3.req.path.indexOf(marker);
+  const filePath = markerIdx >= 0 ? c3.req.path.slice(markerIdx + marker.length) : "";
+  if (!filePath) return c3.json({ error: "No file path specified" }, 400);
+  const baseDir = path10.dirname(meta.path);
+  const absolutePath = path10.resolve(baseDir, filePath);
+  if (!absolutePath.startsWith(path10.resolve(baseDir) + path10.sep) && absolutePath !== path10.resolve(baseDir)) {
+    return c3.json({ error: "Path traversal not allowed" }, 403);
+  }
+  if (!existsSync8(absolutePath) || !statSync4(absolutePath).isFile()) {
+    return c3.json({ error: "File not found" }, 404);
+  }
+  try {
+    const fileContent = readFileSync9(absolutePath, "utf8");
+    const language = inferLanguage(absolutePath);
+    return c3.json({ content: fileContent, language });
+  } catch {
+    return c3.json({ error: "Failed to read file" }, 500);
+  }
+}
+function handleExperiments(c3, { searchDir, agentvDir }) {
+  const metas = listResultFiles(searchDir);
+  const { pass_threshold } = loadStudioConfig(agentvDir);
+  const experimentMap = /* @__PURE__ */ new Map();
+  for (const m of metas) {
+    try {
+      const records = loadLightweightResults(m.path);
+      for (const r of records) {
+        const experiment = r.experiment ?? "default";
+        const entry = experimentMap.get(experiment) ?? {
+          targets: /* @__PURE__ */ new Set(),
+          runFilenames: /* @__PURE__ */ new Set(),
+          evalCount: 0,
+          passedCount: 0,
+          lastTimestamp: ""
+        };
+        entry.runFilenames.add(m.filename);
+        if (r.target) entry.targets.add(r.target);
+        entry.evalCount++;
+        if (r.score >= pass_threshold) entry.passedCount++;
+        if (r.timestamp && r.timestamp > entry.lastTimestamp) {
+          entry.lastTimestamp = r.timestamp;
+        }
+        experimentMap.set(experiment, entry);
+      }
+    } catch {
+    }
+  }
+  const experiments = [...experimentMap.entries()].map(([name, entry]) => ({
+    name,
+    run_count: entry.runFilenames.size,
+    target_count: entry.targets.size,
+    eval_count: entry.evalCount,
+    passed_count: entry.passedCount,
+    pass_rate: entry.evalCount > 0 ? entry.passedCount / entry.evalCount : 0,
+    last_run: entry.lastTimestamp || null
+  }));
+  return c3.json({ experiments });
+}
+function handleTargets(c3, { searchDir, agentvDir }) {
+  const metas = listResultFiles(searchDir);
+  const { pass_threshold } = loadStudioConfig(agentvDir);
+  const targetMap = /* @__PURE__ */ new Map();
+  for (const m of metas) {
+    try {
+      const records = loadLightweightResults(m.path);
+      for (const r of records) {
+        const target = r.target ?? "default";
+        const entry = targetMap.get(target) ?? {
+          experiments: /* @__PURE__ */ new Set(),
+          runFilenames: /* @__PURE__ */ new Set(),
+          evalCount: 0,
+          passedCount: 0
+        };
+        entry.runFilenames.add(m.filename);
+        if (r.experiment) entry.experiments.add(r.experiment);
+        entry.evalCount++;
+        if (r.score >= pass_threshold) entry.passedCount++;
+        targetMap.set(target, entry);
+      }
+    } catch {
+    }
+  }
+  const targets = [...targetMap.entries()].map(([name, entry]) => ({
+    name,
+    run_count: entry.runFilenames.size,
+    experiment_count: entry.experiments.size,
+    eval_count: entry.evalCount,
+    passed_count: entry.passedCount,
+    pass_rate: entry.evalCount > 0 ? entry.passedCount / entry.evalCount : 0
+  }));
+  return c3.json({ targets });
+}
+function handleConfig(c3, { agentvDir }) {
+  return c3.json(loadStudioConfig(agentvDir));
+}
+function handleFeedbackRead(c3, { searchDir }) {
+  const resultsDir = path10.join(searchDir, ".agentv", "results");
+  return c3.json(readFeedback(existsSync8(resultsDir) ? resultsDir : searchDir));
+}
 function createApp(results, resultDir, cwd, sourceFile, options) {
   const searchDir = cwd ?? resultDir;
   const agentvDir = path10.join(searchDir, ".agentv");
+  const defaultCtx = { searchDir, agentvDir };
   const app2 = new Hono();
-  app2.get("/api/config", (c3) => c3.json(loadStudioConfig(agentvDir)));
+  function withProject(c3, handler) {
+    const project = getProject(c3.req.param("projectId") ?? "");
+    if (!project || !existsSync8(project.path)) {
+      return c3.json({ error: "Project not found" }, 404);
+    }
+    return handler(c3, {
+      searchDir: project.path,
+      agentvDir: path10.join(project.path, ".agentv")
+    });
+  }
   app2.post("/api/config", async (c3) => {
     try {
       const body = await c3.req.json();
@@ -6308,60 +6688,100 @@ function createApp(results, resultDir, cwd, sourceFile, options) {
       return c3.json({ error: "Failed to save config" }, 500);
     }
   });
-  const studioDistPath = options?.studioDir ?? resolveStudioDistDir();
-  if (!studioDistPath || !existsSync8(path10.join(studioDistPath, "index.html"))) {
-    throw new Error('Studio dist not found. Run "bun run build" in apps/studio/ to build the SPA.');
+  function projectEntryToWire(entry) {
+    return {
+      id: entry.id,
+      name: entry.name,
+      path: entry.path,
+      added_at: entry.addedAt,
+      last_opened_at: entry.lastOpenedAt
+    };
   }
-  app2.get("/", (c3) => {
-    const indexPath = path10.join(studioDistPath, "index.html");
-    if (existsSync8(indexPath)) {
-      return c3.html(readFileSync9(indexPath, "utf8"));
-    }
-    return c3.notFound();
-  });
-  app2.get("/api/runs", (c3) => {
-    const metas = listResultFiles(searchDir);
-    return c3.json({
-      runs: metas.map((m) => {
-        let target;
-        let experiment;
-        try {
-          const records = loadLightweightResults(m.path);
-          if (records.length > 0) {
-            target = records[0].target;
-            experiment = records[0].experiment;
-          }
-        } catch {
+  app2.get("/api/projects", (c3) => {
+    const registry = loadProjectRegistry();
+    const projects = registry.projects.map((p) => {
+      let runCount = 0;
+      let passRate = 0;
+      let lastRun = null;
+      try {
+        const metas = listResultFiles(p.path);
+        runCount = metas.length;
+        if (metas.length > 0) {
+          const totalPassRate = metas.reduce((sum, m) => sum + m.passRate, 0);
+          passRate = totalPassRate / metas.length;
+          lastRun = metas[0].timestamp;
         }
-        return {
-          filename: m.filename,
-          path: m.path,
-          timestamp: m.timestamp,
-          test_count: m.testCount,
-          pass_rate: m.passRate,
-          avg_score: m.avgScore,
-          size_bytes: m.sizeBytes,
-          ...target && { target },
-          ...experiment && { experiment }
-        };
-      })
+      } catch {
+      }
+      return {
+        ...projectEntryToWire(p),
+        run_count: runCount,
+        pass_rate: passRate,
+        last_run: lastRun
+      };
     });
+    return c3.json({ projects });
   });
-  app2.get("/api/runs/:filename", (c3) => {
-    const filename = c3.req.param("filename");
-    const metas = listResultFiles(searchDir);
-    const meta = metas.find((m) => m.filename === filename);
-    if (!meta) {
-      return c3.json({ error: "Run not found" }, 404);
+  app2.post("/api/projects", async (c3) => {
+    try {
+      const body = await c3.req.json();
+      if (!body.path) return c3.json({ error: "Missing path" }, 400);
+      const entry = addProject(body.path);
+      return c3.json(projectEntryToWire(entry), 201);
+    } catch (err2) {
+      return c3.json({ error: err2.message }, 400);
+    }
+  });
+  app2.delete("/api/projects/:projectId", (c3) => {
+    const removed = removeProject(c3.req.param("projectId") ?? "");
+    if (!removed) return c3.json({ error: "Project not found" }, 404);
+    return c3.json({ ok: true });
+  });
+  app2.get("/api/projects/:projectId/summary", (c3) => {
+    const project = getProject(c3.req.param("projectId") ?? "");
+    if (!project) return c3.json({ error: "Project not found" }, 404);
+    try {
+      const metas = listResultFiles(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 c3.json({
+        id: project.id,
+        name: project.name,
+        path: project.path,
+        run_count: runCount,
+        pass_rate: passRate,
+        last_run: lastRun
+      });
+    } catch {
+      return c3.json({ error: "Failed to read project" }, 500);
     }
+  });
+  app2.post("/api/projects/discover", async (c3) => {
     try {
-      const loaded = patchTestIds(loadManifestResults(meta.path));
-      const lightResults = stripHeavyFields(loaded);
-      return c3.json({ results: lightResults, source: meta.filename });
+      const body = await c3.req.json();
+      if (!body.path) return c3.json({ error: "Missing path" }, 400);
+      const discovered = discoverProjects(body.path);
+      const registered = discovered.map((p) => projectEntryToWire(addProject(p)));
+      return c3.json({ discovered: registered });
     } catch (err2) {
-      return c3.json({ error: "Failed to load run" }, 500);
+      return c3.json({ error: err2.message }, 400);
     }
   });
+  app2.get("/api/config", (c3) => handleConfig(c3, defaultCtx));
+  app2.get("/api/runs", (c3) => handleRuns(c3, defaultCtx));
+  app2.get("/api/runs/:filename", (c3) => handleRunDetail(c3, defaultCtx));
+  app2.get("/api/runs/:filename/datasets", (c3) => handleRunDatasets(c3, defaultCtx));
+  app2.get("/api/runs/:filename/categories", (c3) => handleRunCategories(c3, defaultCtx));
+  app2.get(
+    "/api/runs/:filename/categories/:category/datasets",
+    (c3) => handleCategoryDatasets(c3, defaultCtx)
+  );
+  app2.get("/api/runs/:filename/evals/:evalId", (c3) => handleEvalDetail(c3, defaultCtx));
+  app2.get("/api/runs/:filename/evals/:evalId/files", (c3) => handleEvalFiles(c3, defaultCtx));
+  app2.get("/api/runs/:filename/evals/:evalId/files/*", (c3) => handleEvalFileContent(c3, defaultCtx));
+  app2.get("/api/experiments", (c3) => handleExperiments(c3, defaultCtx));
+  app2.get("/api/targets", (c3) => handleTargets(c3, defaultCtx));
   app2.get("/api/feedback", (c3) => {
     const data = readFeedback(resultDir);
     return c3.json(data);
@@ -6404,127 +6824,6 @@ function createApp(results, resultDir, cwd, sourceFile, options) {
     writeFeedback(resultDir, existing);
     return c3.json(existing);
   });
-  app2.get("/api/runs/:filename/datasets", (c3) => {
-    const filename = c3.req.param("filename");
-    const metas = listResultFiles(searchDir);
-    const meta = metas.find((m) => m.filename === filename);
-    if (!meta) {
-      return c3.json({ error: "Run not found" }, 404);
-    }
-    try {
-      const loaded = patchTestIds(loadManifestResults(meta.path));
-      const { pass_threshold } = loadStudioConfig(agentvDir);
-      const datasetMap = /* @__PURE__ */ new Map();
-      for (const r of loaded) {
-        const ds = r.dataset ?? r.target ?? "default";
-        const entry = datasetMap.get(ds) ?? { total: 0, passed: 0, scoreSum: 0 };
-        entry.total++;
-        if (r.score >= pass_threshold) entry.passed++;
-        entry.scoreSum += r.score;
-        datasetMap.set(ds, entry);
-      }
-      const datasets = [...datasetMap.entries()].map(([name, entry]) => ({
-        name,
-        total: entry.total,
-        passed: entry.passed,
-        failed: entry.total - entry.passed,
-        avg_score: entry.total > 0 ? entry.scoreSum / entry.total : 0
-      }));
-      return c3.json({ datasets });
-    } catch {
-      return c3.json({ error: "Failed to load datasets" }, 500);
-    }
-  });
-  app2.get("/api/runs/:filename/categories", (c3) => {
-    const filename = c3.req.param("filename");
-    const metas = listResultFiles(searchDir);
-    const meta = metas.find((m) => m.filename === filename);
-    if (!meta) {
-      return c3.json({ error: "Run not found" }, 404);
-    }
-    try {
-      const loaded = patchTestIds(loadManifestResults(meta.path));
-      const { pass_threshold } = loadStudioConfig(agentvDir);
-      const categoryMap = /* @__PURE__ */ new Map();
-      for (const r of loaded) {
-        const cat = r.category ?? DEFAULT_CATEGORY;
-        const entry = categoryMap.get(cat) ?? {
-          total: 0,
-          passed: 0,
-          scoreSum: 0,
-          datasets: /* @__PURE__ */ new Set()
-        };
-        entry.total++;
-        if (r.score >= pass_threshold) entry.passed++;
-        entry.scoreSum += r.score;
-        entry.datasets.add(r.dataset ?? r.target ?? "default");
-        categoryMap.set(cat, entry);
-      }
-      const categories = [...categoryMap.entries()].map(([name, entry]) => ({
-        name,
-        total: entry.total,
-        passed: entry.passed,
-        failed: entry.total - entry.passed,
-        avg_score: entry.total > 0 ? entry.scoreSum / entry.total : 0,
-        dataset_count: entry.datasets.size
-      }));
-      return c3.json({ categories });
-    } catch {
-      return c3.json({ error: "Failed to load categories" }, 500);
-    }
-  });
-  app2.get("/api/runs/:filename/categories/:category/datasets", (c3) => {
-    const filename = c3.req.param("filename");
-    const category = decodeURIComponent(c3.req.param("category"));
-    const metas = listResultFiles(searchDir);
-    const meta = metas.find((m) => m.filename === filename);
-    if (!meta) {
-      return c3.json({ error: "Run not found" }, 404);
-    }
-    try {
-      const loaded = patchTestIds(loadManifestResults(meta.path));
-      const { pass_threshold } = loadStudioConfig(agentvDir);
-      const filtered = loaded.filter((r) => (r.category ?? DEFAULT_CATEGORY) === category);
-      const datasetMap = /* @__PURE__ */ new Map();
-      for (const r of filtered) {
-        const ds = r.dataset ?? r.target ?? "default";
-        const entry = datasetMap.get(ds) ?? { total: 0, passed: 0, scoreSum: 0 };
-        entry.total++;
-        if (r.score >= pass_threshold) entry.passed++;
-        entry.scoreSum += r.score;
-        datasetMap.set(ds, entry);
-      }
-      const datasets = [...datasetMap.entries()].map(([name, entry]) => ({
-        name,
-        total: entry.total,
-        passed: entry.passed,
-        failed: entry.total - entry.passed,
-        avg_score: entry.total > 0 ? entry.scoreSum / entry.total : 0
-      }));
-      return c3.json({ datasets });
-    } catch {
-      return c3.json({ error: "Failed to load datasets" }, 500);
-    }
-  });
-  app2.get("/api/runs/:filename/evals/:evalId", (c3) => {
-    const filename = c3.req.param("filename");
-    const evalId = c3.req.param("evalId");
-    const metas = listResultFiles(searchDir);
-    const meta = metas.find((m) => m.filename === filename);
-    if (!meta) {
-      return c3.json({ error: "Run not found" }, 404);
-    }
-    try {
-      const loaded = patchTestIds(loadManifestResults(meta.path));
-      const result = loaded.find((r) => r.testId === evalId);
-      if (!result) {
-        return c3.json({ error: "Eval not found" }, 404);
-      }
-      return c3.json({ eval: result });
-    } catch {
-      return c3.json({ error: "Failed to load eval" }, 500);
-    }
-  });
   app2.get("/api/index", (c3) => {
     const metas = listResultFiles(searchDir);
     const entries2 = metas.map((m) => {
@@ -6545,204 +6844,49 @@ function createApp(results, resultDir, cwd, sourceFile, options) {
     });
     return c3.json({ entries: entries2 });
   });
-  function buildFileTree(dirPath, relativeTo) {
-    if (!existsSync8(dirPath) || !statSync4(dirPath).isDirectory()) {
-      return [];
-    }
-    const entries2 = readdirSync3(dirPath, { withFileTypes: true });
-    return entries2.sort((a, b) => {
-      if (a.isDirectory() !== b.isDirectory()) return a.isDirectory() ? -1 : 1;
-      return a.name.localeCompare(b.name);
-    }).map((entry) => {
-      const fullPath = path10.join(dirPath, entry.name);
-      const relPath = path10.relative(relativeTo, fullPath);
-      if (entry.isDirectory()) {
-        return {
-          name: entry.name,
-          path: relPath,
-          type: "dir",
-          children: buildFileTree(fullPath, relativeTo)
-        };
-      }
-      return { name: entry.name, path: relPath, type: "file" };
-    });
-  }
-  function inferLanguage(filePath) {
-    const ext = path10.extname(filePath).toLowerCase();
-    const langMap = {
-      ".json": "json",
-      ".jsonl": "json",
-      ".ts": "typescript",
-      ".tsx": "typescript",
-      ".js": "javascript",
-      ".jsx": "javascript",
-      ".md": "markdown",
-      ".yaml": "yaml",
-      ".yml": "yaml",
-      ".log": "plaintext",
-      ".txt": "plaintext",
-      ".py": "python",
-      ".sh": "shell",
-      ".bash": "shell",
-      ".css": "css",
-      ".html": "html",
-      ".xml": "xml",
-      ".svg": "xml",
-      ".toml": "toml",
-      ".diff": "diff",
-      ".patch": "diff"
-    };
-    return langMap[ext] ?? "plaintext";
+  app2.get("/api/projects/:projectId/config", (c3) => withProject(c3, handleConfig));
+  app2.get("/api/projects/:projectId/runs", (c3) => withProject(c3, handleRuns));
+  app2.get("/api/projects/:projectId/runs/:filename", (c3) => withProject(c3, handleRunDetail));
+  app2.get(
+    "/api/projects/:projectId/runs/:filename/datasets",
+    (c3) => withProject(c3, handleRunDatasets)
+  );
+  app2.get(
+    "/api/projects/:projectId/runs/:filename/categories",
+    (c3) => withProject(c3, handleRunCategories)
+  );
+  app2.get(
+    "/api/projects/:projectId/runs/:filename/categories/:category/datasets",
+    (c3) => withProject(c3, handleCategoryDatasets)
+  );
+  app2.get(
+    "/api/projects/:projectId/runs/:filename/evals/:evalId",
+    (c3) => withProject(c3, handleEvalDetail)
+  );
+  app2.get(
+    "/api/projects/:projectId/runs/:filename/evals/:evalId/files",
+    (c3) => withProject(c3, handleEvalFiles)
+  );
+  app2.get(
+    "/api/projects/:projectId/runs/:filename/evals/:evalId/files/*",
+    (c3) => withProject(c3, handleEvalFileContent)
+  );
+  app2.get("/api/projects/:projectId/experiments", (c3) => withProject(c3, handleExperiments));
+  app2.get("/api/projects/:projectId/targets", (c3) => withProject(c3, handleTargets));
+  app2.get("/api/projects/:projectId/feedback", (c3) => withProject(c3, handleFeedbackRead));
+  const studioDistPath = options?.studioDir ?? resolveStudioDistDir();
+  if (!studioDistPath || !existsSync8(path10.join(studioDistPath, "index.html"))) {
+    throw new Error('Studio dist not found. Run "bun run build" in apps/studio/ to build the SPA.');
   }
-  app2.get("/api/runs/:filename/evals/:evalId/files", (c3) => {
-    const filename = c3.req.param("filename");
-    const evalId = c3.req.param("evalId");
-    const metas = listResultFiles(searchDir);
-    const meta = metas.find((m) => m.filename === filename);
-    if (!meta) {
-      return c3.json({ error: "Run not found" }, 404);
-    }
-    try {
-      const content = readFileSync9(meta.path, "utf8");
-      const records = parseResultManifest(content);
-      const record = records.find((r) => (r.test_id ?? r.eval_id) === evalId);
-      if (!record) {
-        return c3.json({ error: "Eval not found" }, 404);
-      }
-      const baseDir = path10.dirname(meta.path);
-      const knownPaths = [
-        record.grading_path,
-        record.timing_path,
-        record.input_path,
-        record.output_path,
-        record.response_path
-      ].filter((p) => !!p);
-      if (knownPaths.length === 0) {
-        return c3.json({ files: [] });
-      }
-      const artifactDirs = knownPaths.map((p) => path10.dirname(p));
-      let commonDir = artifactDirs[0];
-      for (const dir of artifactDirs) {
-        while (!dir.startsWith(commonDir)) {
-          commonDir = path10.dirname(commonDir);
-        }
-      }
-      const artifactAbsDir = path10.join(baseDir, commonDir);
-      const files = buildFileTree(artifactAbsDir, baseDir);
-      return c3.json({ files });
-    } catch {
-      return c3.json({ error: "Failed to load file tree" }, 500);
-    }
-  });
-  app2.get("/api/runs/:filename/evals/:evalId/files/*", (c3) => {
-    const filename = c3.req.param("filename");
-    const evalId = c3.req.param("evalId");
-    const metas = listResultFiles(searchDir);
-    const meta = metas.find((m) => m.filename === filename);
-    if (!meta) {
-      return c3.json({ error: "Run not found" }, 404);
-    }
-    const requestPath = c3.req.path;
-    const prefix = `/api/runs/${filename}/evals/${evalId}/files/`;
-    const filePath = requestPath.slice(prefix.length);
-    if (!filePath) {
-      return c3.json({ error: "No file path specified" }, 400);
-    }
-    const baseDir = path10.dirname(meta.path);
-    const absolutePath = path10.resolve(baseDir, filePath);
-    if (!absolutePath.startsWith(path10.resolve(baseDir) + path10.sep) && absolutePath !== path10.resolve(baseDir)) {
-      return c3.json({ error: "Path traversal not allowed" }, 403);
-    }
-    if (!existsSync8(absolutePath) || !statSync4(absolutePath).isFile()) {
-      return c3.json({ error: "File not found" }, 404);
-    }
-    try {
-      const fileContent = readFileSync9(absolutePath, "utf8");
-      const language = inferLanguage(absolutePath);
-      return c3.json({ content: fileContent, language });
-    } catch {
-      return c3.json({ error: "Failed to read file" }, 500);
-    }
-  });
-  app2.get("/api/experiments", (c3) => {
-    const metas = listResultFiles(searchDir);
-    const { pass_threshold } = loadStudioConfig(agentvDir);
-    const experimentMap = /* @__PURE__ */ new Map();
-    for (const m of metas) {
-      try {
-        const records = loadLightweightResults(m.path);
-        for (const r of records) {
-          const experiment = r.experiment ?? "default";
-          const entry = experimentMap.get(experiment) ?? {
-            targets: /* @__PURE__ */ new Set(),
-            runFilenames: /* @__PURE__ */ new Set(),
-            evalCount: 0,
-            passedCount: 0,
-            lastTimestamp: ""
-          };
-          entry.runFilenames.add(m.filename);
-          if (r.target) entry.targets.add(r.target);
-          entry.evalCount++;
-          if (r.score >= pass_threshold) entry.passedCount++;
-          if (r.timestamp && r.timestamp > entry.lastTimestamp) {
-            entry.lastTimestamp = r.timestamp;
-          }
-          experimentMap.set(experiment, entry);
-        }
-      } catch {
-      }
-    }
-    const experiments = [...experimentMap.entries()].map(([name, entry]) => ({
-      name,
-      run_count: entry.runFilenames.size,
-      target_count: entry.targets.size,
-      eval_count: entry.evalCount,
-      passed_count: entry.passedCount,
-      pass_rate: entry.evalCount > 0 ? entry.passedCount / entry.evalCount : 0,
-      last_run: entry.lastTimestamp || null
-    }));
-    return c3.json({ experiments });
-  });
-  app2.get("/api/targets", (c3) => {
-    const metas = listResultFiles(searchDir);
-    const { pass_threshold } = loadStudioConfig(agentvDir);
-    const targetMap = /* @__PURE__ */ new Map();
-    for (const m of metas) {
-      try {
-        const records = loadLightweightResults(m.path);
-        for (const r of records) {
-          const target = r.target ?? "default";
-          const entry = targetMap.get(target) ?? {
-            experiments: /* @__PURE__ */ new Set(),
-            runFilenames: /* @__PURE__ */ new Set(),
-            evalCount: 0,
-            passedCount: 0
-          };
-          entry.runFilenames.add(m.filename);
-          if (r.experiment) entry.experiments.add(r.experiment);
-          entry.evalCount++;
-          if (r.score >= pass_threshold) entry.passedCount++;
-          targetMap.set(target, entry);
-        }
-      } catch {
-      }
-    }
-    const targets = [...targetMap.entries()].map(([name, entry]) => ({
-      name,
-      run_count: entry.runFilenames.size,
-      experiment_count: entry.experiments.size,
-      eval_count: entry.evalCount,
-      passed_count: entry.passedCount,
-      pass_rate: entry.evalCount > 0 ? entry.passedCount / entry.evalCount : 0
-    }));
-    return c3.json({ targets });
+  app2.get("/", (c3) => {
+    const indexPath = path10.join(studioDistPath, "index.html");
+    if (existsSync8(indexPath)) return c3.html(readFileSync9(indexPath, "utf8"));
+    return c3.notFound();
   });
   app2.get("/assets/*", (c3) => {
     const assetPath = c3.req.path;
     const filePath = path10.join(studioDistPath, assetPath);
-    if (!existsSync8(filePath)) {
-      return c3.notFound();
-    }
+    if (!existsSync8(filePath)) return c3.notFound();
     const content = readFileSync9(filePath);
     const ext = path10.extname(filePath);
     const mimeTypes = {
@@ -6764,13 +6908,9 @@ function createApp(results, resultDir, cwd, sourceFile, options) {
     });
   });
   app2.get("*", (c3) => {
-    if (c3.req.path.startsWith("/api/")) {
-      return c3.json({ error: "Not found" }, 404);
-    }
+    if (c3.req.path.startsWith("/api/")) return c3.json({ error: "Not found" }, 404);
     const indexPath = path10.join(studioDistPath, "index.html");
-    if (existsSync8(indexPath)) {
-      return c3.html(readFileSync9(indexPath, "utf8"));
-    }
+    if (existsSync8(indexPath)) return c3.html(readFileSync9(indexPath, "utf8"));
     return c3.notFound();
   });
   return app2;
@@ -6794,18 +6934,6 @@ function resolveStudioDistDir() {
   }
   return void 0;
 }
-function stripHeavyFields(results) {
-  return results.map((r) => {
-    const { requests, trace, ...rest } = r;
-    const toolCalls = trace?.toolCalls && Object.keys(trace.toolCalls).length > 0 ? trace.toolCalls : void 0;
-    const graderDurationMs = (r.scores ?? []).reduce((sum, s) => sum + (s.durationMs ?? 0), 0);
-    return {
-      ...rest,
-      ...toolCalls && { _toolCalls: toolCalls },
-      ...graderDurationMs > 0 && { _graderDurationMs: graderDurationMs }
-    };
-  });
-}
 var resultsServeCommand = command({
   name: "studio",
   description: "Start AgentV Studio \u2014 a local dashboard for reviewing evaluation results",
@@ -6826,11 +6954,66 @@ var resultsServeCommand = command({
       long: "dir",
       short: "d",
       description: "Working directory (default: current directory)"
+    }),
+    multi: flag({
+      long: "multi",
+      description: "Launch in multi-project dashboard mode"
+    }),
+    add: option({
+      type: optional(string),
+      long: "add",
+      description: "Register a project by path"
+    }),
+    remove: option({
+      type: optional(string),
+      long: "remove",
+      description: "Unregister a project by ID"
+    }),
+    discover: option({
+      type: optional(string),
+      long: "discover",
+      description: "Scan a directory tree for repos with .agentv/"
     })
   },
-  handler: async ({ source, port, dir }) => {
+  handler: async ({ source, port, dir, multi, add, remove, discover }) => {
     const cwd = dir ?? process.cwd();
     const listenPort = port ?? (process.env.PORT ? Number(process.env.PORT) : 3117);
+    if (add) {
+      try {
+        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);
+      }
+      return;
+    }
+    if (remove) {
+      const removed = removeProject(remove);
+      if (removed) {
+        console.log(`Unregistered project: ${remove}`);
+      } else {
+        console.error(`Project not found: ${remove}`);
+        process.exit(1);
+      }
+      return;
+    }
+    if (discover) {
+      const discovered = discoverProjects(discover);
+      if (discovered.length === 0) {
+        console.log(`No projects with .agentv/ found under ${discover}`);
+        return;
+      }
+      for (const p of discovered) {
+        const entry = addProject(p);
+        console.log(`Registered: ${entry.name} (${entry.id}) at ${entry.path}`);
+      }
+      console.log(`
+Discovered ${discovered.length} project(s).`);
+      return;
+    }
+    const registry = loadProjectRegistry();
+    const isMultiProject = multi || registry.projects.length > 0;
     try {
       let results = [];
       let sourceFile;
@@ -6858,16 +7041,16 @@ var resultsServeCommand = command({
       }
       const resultDir = sourceFile ? path10.dirname(path10.resolve(sourceFile)) : cwd;
       const app2 = createApp(results, resultDir, cwd, sourceFile);
-      if (results.length > 0 && sourceFile) {
+      if (isMultiProject) {
+        console.log(`Multi-project mode: ${registry.projects.length} project(s) registered`);
+      } else if (results.length > 0 && sourceFile) {
         console.log(`Serving ${results.length} result(s) from ${sourceFile}`);
       } else {
         console.log("No results found. Dashboard will show an empty state.");
         console.log("Run an evaluation to see results: agentv eval <eval-file>");
       }
       console.log(`Dashboard: http://localhost:${listenPort}`);
-      console.log(`Feedback API: http://localhost:${listenPort}/api/feedback`);
-      console.log(`Result picker API: http://localhost:${listenPort}/api/runs`);
-      console.log(`Feedback file: ${feedbackPath(resultDir)}`);
+      console.log(`Projects API: http://localhost:${listenPort}/api/projects`);
       console.log("Press Ctrl+C to stop");
       const { serve: startServer } = await import("@hono/node-server");
       startServer({
@@ -8042,7 +8225,7 @@ function isYamlFile(filePath) {
 }
 
 // src/commands/validate/index.ts
-async function runValidateCommand(paths) {
+async function runValidateCommand(paths, maxWarnings) {
   if (paths.length === 0) {
     console.error("Error: No paths specified. Usage: agentv validate <paths...>");
     process.exit(1);
@@ -8053,6 +8236,18 @@ async function runValidateCommand(paths) {
   if (summary.invalidFiles > 0) {
     process.exit(1);
   }
+  if (maxWarnings !== void 0) {
+    const warningCount = summary.results.reduce(
+      (count, r) => count + r.errors.filter((e) => e.severity === "warning").length,
+      0
+    );
+    if (warningCount > maxWarnings) {
+      console.error(
+        `Found ${warningCount} warning${warningCount === 1 ? "" : "s"} (max allowed: ${maxWarnings})`
+      );
+      process.exit(1);
+    }
+  }
 }
 var validateCommand = command({
   name: "validate",
@@ -8062,11 +8257,16 @@ var validateCommand = command({
       type: string,
       displayName: "paths",
       description: "Files or directories to validate"
+    }),
+    maxWarnings: option({
+      type: optional(number),
+      long: "max-warnings",
+      description: "Maximum number of warnings allowed before failing (e.g., --max-warnings 0)"
     })
   },
-  handler: async ({ paths }) => {
+  handler: async ({ paths, maxWarnings }) => {
     try {
-      await runValidateCommand(paths);
+      await runValidateCommand(paths, maxWarnings);
     } catch (error) {
       console.error(`Error: ${error.message}`);
       process.exit(1);
@@ -8414,4 +8614,4 @@ export {
   preprocessArgv,
   runCli
 };
-//# sourceMappingURL=chunk-5DDVNHOS.js.map
+//# sourceMappingURL=chunk-YORCRL4G.js.map