agentv 4.31.0-next.1 → 4.31.1

package/dist/{chunk-BCZHBAUK.js → chunk-ZFTDIK4V.js}

@@ -16,7 +16,7 @@ import {
   toSnakeCaseDeep,
   writeArtifactsFromResults,
   writeInitialBenchmarkArtifact
-} from "./chunk-BMSSZSND.js";
+} from "./chunk-IJPWTVDU.js";
 import {
   ResponseCache,
   RunBudgetTracker,
@@ -24,12 +24,15 @@ import {
   directPushResults,
   directorySizeBytes,
   getResultsRepoStatus,
+  listGitRuns,
   loadTsConfig,
+  materializeGitRun,
+  normalizeResultsConfig,
   resolveResultsRepoRunsDir,
   shouldEnableCache,
   shouldSkipCacheForTemperature,
   syncResultsRepo
-} from "./chunk-FLSABQ33.js";
+} from "./chunk-LOYPSIE7.js";
 import {
   CLI_PLACEHOLDERS,
   COMMON_TARGET_SETTINGS,
@@ -61,7 +64,7 @@ import {
   subscribeToCopilotSdkLogEntries,
   subscribeToPiLogEntries,
   toCamelCaseDeep
-} from "./chunk-2Z2V7RJO.js";
+} from "./chunk-377GONL7.js";
 
 // src/commands/eval/shared.ts
 import { constants } from "node:fs";
@@ -177,7 +180,7 @@ async function findRepoRoot(start) {
 // package.json
 var package_default = {
   name: "agentv",
-  version: "4.31.0-next.1",
+  version: "4.31.1",
   description: "CLI entry point for AgentV",
   type: "module",
   repository: {
@@ -285,7 +288,7 @@ async function discoverTargetsFile(options) {
 }
 
 // src/commands/eval/run-eval.ts
-import { constants as constants4, existsSync as existsSync4, mkdirSync } from "node:fs";
+import { constants as constants4, existsSync as existsSync5, mkdirSync } from "node:fs";
 import { access as access5, readFile as readFile8 } from "node:fs/promises";
 import path15 from "node:path";
 import { pathToFileURL } from "node:url";
@@ -514,6 +517,7 @@ ${ANSI_GREEN}Update complete.${ANSI_RESET}`);
 }
 
 // src/commands/results/remote.ts
+import { existsSync as existsSync3 } from "node:fs";
 import path5 from "node:path";
 
 // src/commands/inspect/utils.ts
@@ -1074,19 +1078,30 @@ function formatScore(score) {
 }
 
 // src/commands/results/remote.ts
+var gitRunsCache = /* @__PURE__ */ new Map();
+var GIT_RUNS_CACHE_TTL_MS = 6e4;
+function cachedListGitRuns(repoDir) {
+  const now = Date.now();
+  const cached = gitRunsCache.get(repoDir);
+  if (cached && cached.expiresAt > now) {
+    return cached.data;
+  }
+  const promise = listGitRuns(repoDir);
+  gitRunsCache.set(repoDir, { data: promise, expiresAt: now + GIT_RUNS_CACHE_TTL_MS });
+  promise.catch(() => {
+  }).finally(() => {
+    const entry = gitRunsCache.get(repoDir);
+    if (entry && entry.expiresAt <= Date.now()) {
+      gitRunsCache.delete(repoDir);
+    }
+  });
+  return promise;
+}
 var REMOTE_RUN_PREFIX = "remote::";
 var SIZE_WARNING_BYTES = 10 * 1024 * 1024;
 function getStatusMessage(error) {
   return error instanceof Error ? error.message : String(error);
 }
-function normalizeResultsConfig(config) {
-  return {
-    repo: config.repo,
-    path: config.path,
-    auto_push: config.auto_push === true,
-    branch_prefix: config.branch_prefix?.trim() || "eval-results"
-  };
-}
 function getRelativeRunPath(cwd, runDir) {
   const relative = path5.relative(path5.join(cwd, ".agentv", "results", "runs"), runDir);
   if (!relative.startsWith("..") && !path5.isAbsolute(relative)) {
@@ -1124,7 +1139,14 @@ function encodeRemoteRunId(filename) {
 async function getRemoteResultsStatus(cwd) {
   const config = await loadNormalizedResultsConfig(cwd);
   const status = getResultsRepoStatus(config);
-  const runCount = config && status.available ? listResultFilesFromRunsDir(resolveResultsRepoRunsDir(config)).length : 0;
+  let runCount = 0;
+  if (config && status.available) {
+    try {
+      runCount = (await cachedListGitRuns(config.path)).length;
+    } catch {
+      runCount = listResultFilesFromRunsDir(resolveResultsRepoRunsDir(config)).length;
+    }
+  }
   return {
     ...status,
     run_count: runCount
@@ -1165,14 +1187,43 @@ async function listMergedResultFiles(cwd, limit) {
       remote_status: remoteStatus
     };
   }
-  const remoteRuns = listResultFilesFromRunsDir(resolveResultsRepoRunsDir(config)).map(
-    (meta) => ({
-      ...meta,
-      filename: encodeRemoteRunId(meta.filename),
-      raw_filename: meta.filename,
-      source: "remote"
-    })
-  );
+  let remoteRuns = [];
+  if (config.mode === "github") {
+    try {
+      const gitRuns = await cachedListGitRuns(config.path);
+      remoteRuns = gitRuns.map((r) => ({
+        filename: encodeRemoteRunId(r.run_id),
+        raw_filename: r.run_id,
+        source: "remote",
+        path: path5.join(config.path, r.manifest_path),
+        displayName: r.display_name,
+        timestamp: r.timestamp,
+        testCount: r.test_count,
+        passRate: r.pass_rate || 0,
+        avgScore: r.avg_score || 0,
+        sizeBytes: r.size_bytes || 0
+      }));
+    } catch (error) {
+      console.error("git-native listing failed, falling back", error);
+      remoteRuns = listResultFilesFromRunsDir(resolveResultsRepoRunsDir(config)).map(
+        (meta) => ({
+          ...meta,
+          filename: encodeRemoteRunId(meta.filename),
+          raw_filename: meta.filename,
+          source: "remote"
+        })
+      );
+    }
+  } else {
+    remoteRuns = listResultFilesFromRunsDir(resolveResultsRepoRunsDir(config)).map(
+      (meta) => ({
+        ...meta,
+        filename: encodeRemoteRunId(meta.filename),
+        raw_filename: meta.filename,
+        source: "remote"
+      })
+    );
+  }
   const merged = [...localRuns, ...remoteRuns].sort(
     (a, b) => b.timestamp.localeCompare(a.timestamp)
   );
@@ -1185,6 +1236,21 @@ async function findRunById(cwd, runId) {
   const { runs } = await listMergedResultFiles(cwd);
   return runs.find((run) => run.filename === runId);
 }
+async function ensureRemoteRunAvailable(cwd, meta) {
+  if (meta.source !== "remote" || existsSync3(meta.path)) {
+    return;
+  }
+  const config = await loadNormalizedResultsConfig(cwd);
+  if (!config) {
+    throw new Error("Remote results are not configured");
+  }
+  const relativeManifestPath = path5.relative(config.path, meta.path).split(path5.sep).join("/");
+  if (relativeManifestPath.length === 0 || relativeManifestPath === meta.path || relativeManifestPath.startsWith("../")) {
+    throw new Error(`Remote manifest path is outside the results repo clone: ${meta.path}`);
+  }
+  const relativeRunPath = path5.posix.relative("runs", path5.posix.dirname(relativeManifestPath));
+  await materializeGitRun(config.path, relativeRunPath);
+}
 async function maybeAutoExportRunArtifacts(payload) {
   const config = await loadNormalizedResultsConfig(payload.cwd);
   if (!config?.auto_push) {
@@ -2474,7 +2540,7 @@ async function loadNonErrorResults(jsonlPath) {
 }
 
 // src/commands/eval/run-cache.ts
-import { existsSync as existsSync3 } from "node:fs";
+import { existsSync as existsSync4 } from "node:fs";
 import { mkdir as mkdir6, readFile, writeFile as writeFile5 } from "node:fs/promises";
 import path13 from "node:path";
 var CACHE_FILENAME = "cache.json";
@@ -2498,7 +2564,7 @@ async function loadRunCache(cwd) {
 async function resolveCachedRunDir(cwd) {
   const cache = await loadRunCache(cwd);
   if (!cache?.lastRunDir) return void 0;
-  if (!existsSync3(cache.lastRunDir)) return void 0;
+  if (!existsSync4(cache.lastRunDir)) return void 0;
   return cache.lastRunDir;
 }
 async function saveRunCache(cwd, resultPath) {
@@ -4205,22 +4271,43 @@ async function validateConfigFile(filePath) {
         });
       } else {
         const resultsRecord = results;
-        if (typeof resultsRecord.repo !== "string" || resultsRecord.repo.trim().length === 0) {
+        if (resultsRecord.mode !== "github") {
           errors.push({
             severity: "error",
             filePath,
-            location: "results.repo",
-            message: "Field 'results.repo' must be a non-empty string"
+            location: "results.mode",
+            message: "Field 'results.mode' must be 'github'"
           });
         }
-        if (typeof resultsRecord.path !== "string" || resultsRecord.path.trim().length === 0) {
+        if (typeof resultsRecord.repo !== "string" || resultsRecord.repo.trim().length === 0) {
           errors.push({
             severity: "error",
             filePath,
-            location: "results.path",
-            message: "Field 'results.path' must be a non-empty string"
+            location: "results.repo",
+            message: "Field 'results.repo' must be a non-empty string"
           });
         }
+        if (resultsRecord.path !== void 0) {
+          if (typeof resultsRecord.path !== "string" || resultsRecord.path.trim().length === 0) {
+            errors.push({
+              severity: "error",
+              filePath,
+              location: "results.path",
+              message: "Field 'results.path' must be a non-empty string"
+            });
+          } else {
+            const p = resultsRecord.path.trim();
+            const isFilesystemPath = p.startsWith("/") || p.startsWith("~/") || p.startsWith("~\\") || p === "~" || /^[A-Za-z]:[/\\]/.test(p);
+            if (!isFilesystemPath) {
+              errors.push({
+                severity: "error",
+                filePath,
+                location: "results.path",
+                message: `'results.path' must be an absolute or home-relative filesystem path (e.g., ~/data/agentv-results). Found: '${p}'. Remove 'path' to use the default.`
+              });
+            }
+          }
+        }
         if (resultsRecord.auto_push !== void 0 && typeof resultsRecord.auto_push !== "boolean") {
           errors.push({
             severity: "error",
@@ -5385,7 +5472,7 @@ async function runEvalCommand(input) {
     const explicitResumeDir = options.outputDir ?? options.artifacts;
     if (explicitResumeDir) {
       const resumeIndexPath = path15.join(path15.resolve(explicitResumeDir), "index.jsonl");
-      if (existsSync4(resumeIndexPath)) {
+      if (existsSync5(resumeIndexPath)) {
         const content = await readFile8(resumeIndexPath, "utf8");
         const existingResults = parseJsonlResults(content);
         resumeSkipKeys = /* @__PURE__ */ new Set();
@@ -5468,7 +5555,7 @@ async function runEvalCommand(input) {
   const useFileExport = !!options.otelFile;
   if (options.exportOtel || useFileExport) {
     try {
-      const { OtelTraceExporter, OTEL_BACKEND_PRESETS } = await import("./dist-2E6ZNN32.js");
+      const { OtelTraceExporter, OTEL_BACKEND_PRESETS } = await import("./dist-CRYAFKLS.js");
       let endpoint = process.env.OTEL_EXPORTER_OTLP_ENDPOINT;
       let headers = {};
       if (options.otelBackend) {
@@ -5667,7 +5754,7 @@ async function runEvalCommand(input) {
   const activeTestFiles = resolvedTestFiles.filter((f) => fileMetadata.has(f));
   let transcriptProviderFactory;
   if (options.transcript) {
-    const { TranscriptProvider } = await import("./dist-2E6ZNN32.js");
+    const { TranscriptProvider } = await import("./dist-CRYAFKLS.js");
     const transcriptProvider = await TranscriptProvider.fromFile(options.transcript);
     const totalTests = [...fileMetadata.values()].reduce(
       (sum, meta) => sum + meta.testCases.length,
@@ -5844,7 +5931,7 @@ async function runEvalCommand(input) {
     if (usesDefaultArtifactWorkspace && allResults.length > 0) {
       const evalFile = activeTestFiles.length === 1 ? activeTestFiles[0] : "";
       if (isResumeAppend) {
-        const { writePerTestArtifacts } = await import("./artifact-writer-K4B2S7OE.js");
+        const { writePerTestArtifacts } = await import("./artifact-writer-FZ5RUHWC.js");
         await writePerTestArtifacts(allResults, runDir, {
           experiment: normalizeExperimentName(options.experiment)
         });
@@ -6050,6 +6137,7 @@ export {
   syncRemoteResults,
   listMergedResultFiles,
   findRunById,
+  ensureRemoteRunAvailable,
   maybeAutoExportRunArtifacts,
   resolveRunCacheFile,
   loadRunCache,
@@ -6068,4 +6156,4 @@ export {
   getCategories,
   filterByCategory
 };
-//# sourceMappingURL=chunk-BCZHBAUK.js.map
+//# sourceMappingURL=chunk-ZFTDIK4V.js.map