agentv 4.25.6-next.1 → 4.26.0-next.1

package/dist/{chunk-77RRTBK2.js → chunk-4QP2SFRS.js}

@@ -45,7 +45,7 @@ import {
   validateFileReferences,
   validateTargetsFile,
   validateWorkspacePaths
-} from "./chunk-5RGZVQIE.js";
+} from "./chunk-63LUDTLO.js";
 import {
   RESULT_INDEX_FILENAME,
   aggregateRunDir,
@@ -3997,7 +3997,7 @@ var evalRunCommand = command({
   },
   handler: async (args) => {
     if (args.evalPaths.length === 0 && process.stdin.isTTY) {
-      const { launchInteractiveWizard } = await import("./interactive-G5U6DHNL.js");
+      const { launchInteractiveWizard } = await import("./interactive-CCOGDYRK.js");
       await launchInteractiveWizard();
       return;
     }
@@ -9241,6 +9241,18 @@ async function discoverTargetsInProject(cwd) {
     return [];
   }
 }
+function validateResumeOptions(req) {
+  const modes = [];
+  if (req.resume) modes.push("resume");
+  if (req.rerun_failed) modes.push("rerun_failed");
+  if (req.retry_errors?.trim()) {
+    modes.push("retry_errors");
+  }
+  if (modes.length > 1) {
+    return `resume, rerun_failed, and retry_errors are mutually exclusive (got: ${modes.join(", ")})`;
+  }
+  return void 0;
+}
 function buildCliArgs(req) {
   const args = ["eval"];
   if (req.suite_filter?.trim()) {
@@ -9269,6 +9281,18 @@ function buildCliArgs(req) {
   if (req.dry_run) {
     args.push("--dry-run");
   }
+  if (req.output?.trim()) {
+    args.push("--output", req.output.trim());
+  }
+  if (req.resume) {
+    args.push("--resume");
+  }
+  if (req.rerun_failed) {
+    args.push("--rerun-failed");
+  }
+  if (req.retry_errors?.trim()) {
+    args.push("--retry-errors", req.retry_errors.trim());
+  }
   return args;
 }
 function buildCliPreview(args) {
@@ -9285,8 +9309,8 @@ function resolveCliPath(cwd) {
     }
   }
   const currentDir = typeof __dirname !== "undefined" ? __dirname : path16.dirname(fileURLToPath2(import.meta.url));
-  const fromSrc = path16.resolve(currentDir, "../../../cli.ts");
-  const fromDist = path16.resolve(currentDir, "../../cli.js");
+  const fromSrc = path16.resolve(currentDir, "../../cli.ts");
+  const fromDist = path16.resolve(currentDir, "cli.js");
   if (existsSync11(fromSrc)) return { binPath: "bun", args: [fromSrc] };
   if (existsSync11(fromDist)) return { binPath: "bun", args: [fromDist] };
   if (isCommandAvailable("agentv")) {
@@ -9342,6 +9366,10 @@ function registerEvalRoutes(app2, getCwd, options) {
     if (!body.suite_filter?.trim() && (!body.test_ids || body.test_ids.length === 0)) {
       return c4.json({ error: "Provide suite_filter or test_ids" }, 400);
     }
+    const resumeError = validateResumeOptions(body);
+    if (resumeError) {
+      return c4.json({ error: resumeError }, 400);
+    }
     const cliPaths = resolveCliPath(cwd);
     if (!cliPaths) {
       return c4.json({ error: "Cannot locate agentv CLI entry point" }, 500);
@@ -9471,6 +9499,9 @@ Process error: ${err2.message}`;
     }
   });
   app2.post("/api/benchmarks/:benchmarkId/eval/run", async (c4) => {
+    if (readOnly) {
+      return c4.json({ error: "Studio is running in read-only mode" }, 403);
+    }
     const cwd = getCwd(c4);
     let body;
     try {
@@ -9481,6 +9512,10 @@ Process error: ${err2.message}`;
     if (!body.suite_filter?.trim() && (!body.test_ids || body.test_ids.length === 0)) {
       return c4.json({ error: "Provide suite_filter or test_ids" }, 400);
     }
+    const resumeError = validateResumeOptions(body);
+    if (resumeError) {
+      return c4.json({ error: resumeError }, 400);
+    }
     const cliPaths = resolveCliPath(cwd);
     if (!cliPaths) {
       return c4.json({ error: "Cannot locate agentv CLI entry point" }, 500);
@@ -9884,15 +9919,35 @@ async function handleRunDetail(c4, { searchDir }) {
   if (!meta) return c4.json({ error: "Run not found" }, 404);
   try {
     const loaded = loadManifestResults(meta.path);
+    const resumeMeta = meta.source === "local" ? deriveResumeMeta(searchDir, meta.path) : {};
     return c4.json({
       results: stripHeavyFields(loaded),
       source: meta.source,
-      source_label: meta.displayName
+      source_label: meta.displayName,
+      ...resumeMeta
     });
   } catch {
     return c4.json({ error: "Failed to load run" }, 500);
   }
 }
+function deriveResumeMeta(cwd, manifestPath) {
+  const out = {};
+  const runDir = path19.dirname(manifestPath);
+  const relative3 = path19.relative(cwd, runDir);
+  out.run_dir = relative3 !== "" && !relative3.startsWith("..") ? relative3 : runDir;
+  try {
+    const benchmarkPath = path19.join(runDir, "benchmark.json");
+    if (existsSync14(benchmarkPath)) {
+      const parsed = JSON.parse(readFileSync12(benchmarkPath, "utf8"));
+      const evalFile = parsed.metadata?.eval_file;
+      if (typeof evalFile === "string" && evalFile.trim()) {
+        out.suite_filter = evalFile.trim();
+      }
+    }
+  } catch {
+  }
+  return out;
+}
 async function handleRunSuites(c4, { searchDir, agentvDir }) {
   const filename = c4.req.param("filename") ?? "";
   const meta = await findRunById(searchDir, filename);
@@ -11991,4 +12046,4 @@ export {
   preprocessArgv,
   runCli
 };
-//# sourceMappingURL=chunk-77RRTBK2.js.map
+//# sourceMappingURL=chunk-4QP2SFRS.js.map