archal 0.9.9 → 0.9.10

package/dist/index.cjs

@@ -31869,7 +31869,6 @@ var init_scenario_schemas = __esm({
       timeout: external_exports3.number().default(120),
       runs: external_exports3.number().default(5),
       evaluatorModel: external_exports3.string().optional(),
-      difficulty: external_exports3.enum(["easy", "medium", "hard"]).optional(),
       tags: external_exports3.array(external_exports3.string()).default([])
     });
   }
@@ -58333,8 +58332,7 @@ function loadArchalFile(explicitPath, ...searchFrom) {
     );
   }
   if (typeof c["agentModel"] === "string") r.agentModel = c["agentModel"].trim() || void 0;
-  if (typeof c["harness"] === "string") r.harness = c["harness"].trim() || void 0;
-  if (typeof c["model"] === "string") r.model = c["model"].trim() || void 0;
+  if (typeof c["evaluatorModel"] === "string") r.evaluatorModel = c["evaluatorModel"].trim() || void 0;
   if (typeof c["runs"] === "number" && Number.isInteger(c["runs"]) && c["runs"] > 0) r.runs = c["runs"];
   if (typeof c["timeout"] === "number" && c["timeout"] > 0) r.timeout = c["timeout"];
   return { config: r, configDir: (0, import_node_path28.dirname)(filePath) };
@@ -60424,21 +60422,12 @@ function parseConfigSection(configText) {
         result.evaluatorModel = value;
         break;
       }
-      case "difficulty": {
-        const normalized = value.toLowerCase();
-        if (normalized === "easy" || normalized === "medium" || normalized === "hard") {
-          result.difficulty = normalized;
-        } else {
-          warn(`Invalid difficulty "${value}" in scenario config \u2014 ignoring. Must be "easy", "medium", or "hard".`);
-        }
-        break;
-      }
       case "tags": {
         result.tags = value.split(",").map((tag) => tag.trim()).filter(Boolean);
         break;
       }
       default: {
-        const knownKeys = ["twins", "timeout", "seed", "runs", "evaluator", "evaluator-model", "evaluatormodel", "model", "difficulty", "tags"];
+        const knownKeys = ["twins", "timeout", "seed", "runs", "evaluator", "evaluator-model", "evaluatormodel", "model", "tags"];
         const close = knownKeys.find((k) => k.startsWith(key));
         const hint = close ? ` Did you mean "${close}"?` : ` Known keys: ${knownKeys.join(", ")}.`;
         warn(`Unknown config key "${key}" in scenario \u2014 ignored.${hint}`);
@@ -60502,8 +60491,7 @@ function parseScenarioMarkdown(markdown, sourcePath) {
     timeout: parsedConfig.timeout ?? 180,
     runs: parsedConfig.runs ?? 1,
     evaluatorModel: parsedConfig.evaluatorModel,
-    tags: parsedConfig.tags ?? [],
-    difficulty: parsedConfig.difficulty
+    tags: parsedConfig.tags ?? []
   };
   debug("Parsed scenario", {
     title: sections.title,
@@ -60813,7 +60801,7 @@ function traceEntriesForTool(trace, toolName) {
 // src/runner/reporter-files.ts
 function writeLastRunLog(report) {
   try {
-    const dir = (0, import_node_path8.join)(process.cwd(), ".archal");
+    const dir = (0, import_node_path8.join)(process.cwd(), ".archal", "cache");
     if (!(0, import_node_fs10.existsSync)(dir)) {
       (0, import_node_fs10.mkdirSync)(dir, { recursive: true });
     }
@@ -60839,7 +60827,7 @@ function writeLastRunLog(report) {
 }
 function saveRunTrace(report) {
   try {
-    const dir = (0, import_node_path8.join)(process.cwd(), ".archal", "runs");
+    const dir = (0, import_node_path8.join)(process.cwd(), ".archal", "cache", "runs");
     (0, import_node_fs10.mkdirSync)(dir, { recursive: true });
     const ts = (/* @__PURE__ */ new Date()).toISOString().replace(/[:.]/g, "-").slice(0, 19);
     const slug2 = (report.scenarioTitle ?? "untitled").toLowerCase().replace(/[^a-z0-9]+/g, "-").replace(/^-|-$/g, "").slice(0, 60);
@@ -63404,6 +63392,55 @@ function buildCriterionMaps(scenario) {
   }
   return { criterionDescriptions, criterionTypes };
 }
+function summarizeExistingState(snapshot) {
+  if (!snapshot) return "(existing state)";
+  const nonEmpty = Object.entries(snapshot).filter(([, entities]) => Array.isArray(entities) && entities.length > 0).map(([collection, entities]) => ({ collection, count: entities.length })).sort((a, b) => b.count - a.count).slice(0, 2);
+  if (nonEmpty.length === 0) return "(existing state)";
+  return nonEmpty.map(({ count, collection }) => `${count} ${collection}`).join(", ");
+}
+function emitSkipReapplyBanner(declaredSeed, stateSummary) {
+  const body = `Seed not re-applied: scenario declares seed "${declaredSeed}" but reusing existing
+   session state from an earlier \`archal twin start\` (has: ${stateSummary}).
+   Pass --fresh-seed to reset and re-apply, or run \`archal twin stop\` and retry.`;
+  warn(body);
+}
+var RESET_STATE_TIMEOUT_MS = 3e4;
+var RESET_STATE_RETRIES = 6;
+var RESET_STATE_BACKOFF_MS = [1500, 2500, 3500, 5e3, 6e3, 7e3];
+async function resetLiveTwinState(twinApiUrl, bearerToken, adminAuth) {
+  const baseUrl = twinApiUrl.replace(/\/(mcp|api)\/?$/, "");
+  const url2 = `${baseUrl}/state`;
+  const headers = buildTwinHttpHeaders(bearerToken, adminAuth);
+  let response;
+  try {
+    response = await fetchWithRetry2(
+      url2,
+      { method: "DELETE", headers },
+      {
+        maxRetries: RESET_STATE_RETRIES,
+        timeoutMs: RESET_STATE_TIMEOUT_MS,
+        backoffMs: RESET_STATE_BACKOFF_MS,
+        label: "twin-state-reset"
+      }
+    );
+  } catch (error49) {
+    const detail2 = error49 instanceof Error ? error49.message : String(error49);
+    return { ok: false, status: 0, detail: detail2 };
+  }
+  if (response.ok) {
+    return { ok: true };
+  }
+  let detail;
+  try {
+    const raw = await response.text();
+    const parsed = raw.trim() ? JSON.parse(raw) : {};
+    const msg = parsed["error"] ?? parsed["message"];
+    detail = typeof msg === "string" ? msg : raw.slice(0, 200) || void 0;
+  } catch {
+    detail = void 0;
+  }
+  return { ok: false, status: response.status, detail };
+}
 async function hydrateStaticSeedSelections(seedSelections, seedDescription, options) {
   const autoSelections = generateSeedSelections(
     seedSelections.map((selection) => selection.twinName),
@@ -63463,6 +63500,43 @@ async function hydrateStaticSeedSelections(seedSelections, seedDescription, opti
           continue;
         }
         if (options.isReusedEnvSession && twinHasExistingState(selection.twinName)) {
+          if (options.freshSeed) {
+            const twinUrl2 = options.cloudTwinUrls[selection.twinName];
+            if (!twinUrl2) {
+              throw new Error(
+                `--fresh-seed requested but hosted session is missing the API base URL for twin "${selection.twinName}". Re-provision the hosted session before retrying.`
+              );
+            }
+            info(
+              `--fresh-seed: resetting twin "${selection.twinName}" and re-applying seed "${selection.seedName}".`
+            );
+            const resetResult = await resetLiveTwinState(
+              twinUrl2,
+              options.apiBearerToken,
+              adminAuth
+            );
+            if (!resetResult.ok) {
+              throw new Error(
+                `--fresh-seed: failed to reset state on twin "${selection.twinName}" (HTTP ${resetResult.status}${resetResult.detail ? `: ${resetResult.detail}` : ""}). Stop the twin session with \`archal twin stop\` and retry.`
+              );
+            }
+            const loadResult2 = await loadNamedSeedIntoLiveTwin(
+              twinUrl2,
+              selection.seedName,
+              options.apiBearerToken,
+              adminAuth
+            );
+            if (!loadResult2.ok) {
+              throw new Error(
+                `--fresh-seed: could not load seed "${selection.seedName}" into twin "${selection.twinName}" (HTTP ${loadResult2.status}${loadResult2.detail ? `: ${loadResult2.detail}` : ""}). Re-provision the hosted session with the requested seed, or point --seed at a seed this twin exposes.`
+              );
+            }
+            continue;
+          }
+          const stateSummary = summarizeExistingState(
+            probedSnapshot ? normalizeSeedState(probedSnapshot[selection.twinName]) : null
+          );
+          emitSkipReapplyBanner(selection.seedName, stateSummary);
           info(
             `Using existing state on reused session; scenario seed '${selection.seedName}' not re-applied.`
           );
@@ -64311,40 +64385,27 @@ function resolveMarkdownPromptOrder(markdownFiles) {
 // src/runner/harness-paths.ts
 var import_node_fs24 = require("fs");
 var import_node_path22 = require("path");
-var MANIFEST_FILE = "archal-harness.json";
-var EXPLICIT_HARNESS_ENTRY_CANDIDATES = [
-  "harness.ts",
-  "harness.js",
-  "harness.mjs",
-  "harness.cjs",
-  "archal-harness.ts",
-  "archal-harness.js",
-  "archal-harness.mjs",
-  "archal-harness.cjs"
-];
-var PROJECT_ARCHAL_HARNESS_FILENAMES = /* @__PURE__ */ new Set([
+var MANIFEST_FILE = "harness.json";
+var HARNESS_ENTRY_CANDIDATES = [
   "harness.ts",
   "harness.js",
   "harness.mjs",
   "harness.cjs"
-]);
-var PROJECT_ARCHAL_HARNESS_ENTRY_CANDIDATES = [
-  "harness.ts",
-  "harness.js",
-  "harness.mjs",
-  "harness.cjs",
-  "archal-harness.ts",
-  "archal-harness.js",
-  "archal-harness.mjs",
-  "archal-harness.cjs"
 ];
+var PROJECT_ARCHAL_HARNESS_FILENAMES = new Set(HARNESS_ENTRY_CANDIDATES);
 var HARNESS_DISCOVERY_BOUNDARY_FILES = [
   ".git",
-  ".archal.json"
+  ".archal"
 ];
-function trimToUndefined(value) {
-  const trimmed = value?.trim();
-  return trimmed ? trimmed : void 0;
+var PROJECT_ROOT_MARKERS = [
+  "package.json",
+  ".archal.json",
+  ".git"
+];
+function isProperAncestor(candidate, target) {
+  if (candidate === target) return false;
+  const withSep = candidate.endsWith(import_node_path22.sep) ? candidate : candidate + import_node_path22.sep;
+  return target.startsWith(withSep);
 }
 function isProjectArchalHarnessFile(entryPath) {
   return (0, import_node_path22.basename)((0, import_node_path22.dirname)(entryPath)) === ".archal" && PROJECT_ARCHAL_HARNESS_FILENAMES.has((0, import_node_path22.basename)(entryPath).toLowerCase());
@@ -64356,6 +64417,28 @@ function resolveHarnessExecutionRootFromPath(entryPath) {
   }
   return entryDir;
 }
+function resolveHarnessProjectRoot(startPath, fallback, stopAt) {
+  let current;
+  try {
+    const stats = (0, import_node_fs24.statSync)(startPath);
+    current = stats.isDirectory() ? startPath : (0, import_node_path22.dirname)(startPath);
+  } catch {
+    current = (0, import_node_path22.dirname)(startPath);
+  }
+  const effectiveFallback = fallback ?? current;
+  while (true) {
+    for (const marker of PROJECT_ROOT_MARKERS) {
+      if ((0, import_node_fs24.existsSync)((0, import_node_path22.resolve)(current, marker))) {
+        return current;
+      }
+    }
+    const parent = (0, import_node_path22.dirname)(current);
+    if (parent === current || stopAt !== void 0 && isProperAncestor(current, stopAt)) {
+      return effectiveFallback;
+    }
+    current = parent;
+  }
+}
 function normalizeHarnessDiscoveryStart(startPath) {
   const trimmed = startPath.trim();
   if (!trimmed) {
@@ -64400,20 +64483,11 @@ function discoverRepoLocalHarness(...startPaths) {
     let current = startDir;
     while (!visited.has(current)) {
       visited.add(current);
-      for (const candidate of PROJECT_ARCHAL_HARNESS_ENTRY_CANDIDATES) {
+      for (const candidate of HARNESS_ENTRY_CANDIDATES) {
         if ((0, import_node_fs24.existsSync)((0, import_node_path22.resolve)(current, candidate))) {
-          return {
-            harnessDir: current,
-            reason: candidate.startsWith("archal-") ? "archal-harness" : "harness"
-          };
+          return { harnessDir: current };
         }
       }
-      if (trimToUndefined(readPackageJson3(current)?.archal?.harness)) {
-        return {
-          harnessDir: current,
-          reason: "package.json#archal.harness"
-        };
-      }
       const reachedBoundary = HARNESS_DISCOVERY_BOUNDARY_FILES.some((file2) => (0, import_node_fs24.existsSync)((0, import_node_path22.resolve)(current, file2)));
       const parent = (0, import_node_path22.dirname)(current);
       if (reachedBoundary || parent === current) {
@@ -64439,29 +64513,11 @@ function resolveHarnessLocation(harnessInput, visited = /* @__PURE__ */ new Set(
     return {
       executionRoot,
       manifestPath: (0, import_node_path22.resolve)(executionRoot, MANIFEST_FILE),
-      entryPath: resolvedInput
-    };
-  }
-  const packageJson = readPackageJson3(resolvedInput);
-  const configuredHarnessPath = trimToUndefined(packageJson?.archal?.harness);
-  if (configuredHarnessPath) {
-    const configuredAbsolutePath = (0, import_node_path22.resolve)(resolvedInput, configuredHarnessPath);
-    if (!(0, import_node_fs24.existsSync)(configuredAbsolutePath)) {
-      throw new Error(
-        `package.json archal.harness points to a missing path: ${configuredAbsolutePath}`
-      );
-    }
-    const configuredStats = (0, import_node_fs24.statSync)(configuredAbsolutePath);
-    if (configuredStats.isDirectory()) {
-      return resolveHarnessLocation(configuredAbsolutePath, visited);
-    }
-    return {
-      executionRoot: resolvedInput,
-      manifestPath: (0, import_node_path22.resolve)(resolvedInput, MANIFEST_FILE),
-      entryPath: configuredAbsolutePath
+      entryPath: resolvedInput,
+      projectRoot: resolveHarnessProjectRoot(resolvedInput, executionRoot)
     };
   }
-  for (const candidate of EXPLICIT_HARNESS_ENTRY_CANDIDATES) {
+  for (const candidate of HARNESS_ENTRY_CANDIDATES) {
     const candidatePath = (0, import_node_path22.resolve)(resolvedInput, candidate);
     if (!(0, import_node_fs24.existsSync)(candidatePath)) {
       continue;
@@ -64469,12 +64525,14 @@ function resolveHarnessLocation(harnessInput, visited = /* @__PURE__ */ new Set(
     return {
       executionRoot: resolvedInput,
       manifestPath: (0, import_node_path22.resolve)(resolvedInput, MANIFEST_FILE),
-      entryPath: candidatePath
+      entryPath: candidatePath,
+      projectRoot: resolvedInput
     };
   }
   return {
     executionRoot: resolvedInput,
-    manifestPath: (0, import_node_path22.resolve)(resolvedInput, MANIFEST_FILE)
+    manifestPath: (0, import_node_path22.resolve)(resolvedInput, MANIFEST_FILE),
+    projectRoot: resolvedInput
   };
 }
 
@@ -64496,14 +64554,14 @@ var DIRECT_FILE_RUNNERS = {
   ".mjs": { command: "node", args: [] },
   ".cjs": { command: "node", args: [] }
 };
-function inferLocalCommandFromEntryPath(executionRoot, entryPath) {
+function inferLocalCommandFromEntryPath(spawnCwd, entryPath) {
   const runner = DIRECT_FILE_RUNNERS[(0, import_node_path23.extname)(entryPath).toLowerCase()];
   if (!runner) {
     throw new Error(
       `Unsupported harness entrypoint extension for ${entryPath}. Use .ts, .js, .mjs, or .cjs.`
     );
   }
-  const relativeEntryPath = (0, import_node_path23.relative)(executionRoot, entryPath) || entryPath;
+  const relativeEntryPath = (0, import_node_path23.relative)(spawnCwd, entryPath) || entryPath;
   return {
     command: runner.command,
     args: [...runner.args, relativeEntryPath]
@@ -64520,7 +64578,7 @@ function parseHarnessManifest(manifestPath) {
     );
   }
 }
-function trimToUndefined2(value) {
+function trimToUndefined(value) {
   const trimmed = value?.trim();
   return trimmed ? trimmed : void 0;
 }
@@ -64530,10 +64588,11 @@ function buildResolvedLocalHarness(location, explicitModel, manifest) {
     command: manifest.local.command,
     args: manifest.local.args,
     env: manifest.local.env
-  } : location.entryPath ? inferLocalCommandFromEntryPath(location.executionRoot, location.entryPath) : void 0;
-  const model = explicitModel ?? trimToUndefined2(manifest?.defaultModel);
+  } : location.entryPath ? inferLocalCommandFromEntryPath(location.projectRoot, location.entryPath) : void 0;
+  const model = explicitModel ?? trimToUndefined(manifest?.defaultModel);
   return {
     harnessDir: location.executionRoot,
+    projectRoot: location.projectRoot,
     manifestPath: location.manifestPath,
     manifest,
     model,
@@ -64543,7 +64602,7 @@ function buildResolvedLocalHarness(location, explicitModel, manifest) {
 }
 function resolveLocalHarness(harnessDirInput, explicitModel) {
   const location = resolveHarnessLocation(harnessDirInput);
-  const explicit = trimToUndefined2(explicitModel);
+  const explicit = trimToUndefined(explicitModel);
   if (!(0, import_node_fs25.existsSync)(location.manifestPath)) {
     if (location.entryPath) {
       return buildResolvedLocalHarness(location, explicit);
@@ -64677,7 +64736,7 @@ function resolveEngineConfiguration(scenario, options, timeoutSeconds, projectCo
         model: rawEngineModel?.trim() || void 0,
         command: "docker",
         args: [],
-        cwd: harnessDir,
+        cwd: resolveHarnessProjectRoot(harnessDir),
         promptContext: void 0
       };
     } else {
@@ -64705,8 +64764,12 @@ function resolveEngineConfiguration(scenario, options, timeoutSeconds, projectCo
         model: resolvedHarness.model,
         command: commandConfig.command,
         args: commandConfig.args,
-        env: commandConfig.env,
-        cwd: resolvedHarness.harnessDir,
+        env: {
+          ...commandConfig.env,
+          ARCHAL_PROJECT_ROOT: resolvedHarness.projectRoot,
+          ARCHAL_HARNESS_ENTRY_DIR: resolvedHarness.harnessDir
+        },
+        cwd: resolvedHarness.projectRoot,
         promptContext: resolvedHarness.promptContext
       };
     }
@@ -65107,7 +65170,7 @@ function parseBatchJudgeResponse(text, criteriaIds) {
       if (results.size > 0) {
         const missing = criteriaIds.filter((id) => !results.has(id));
         if (missing.length > 0) {
-          warn(`Batch parse: ${missing.length} criteria missing from response (will fall back to fail/0.3): ${missing.join(", ")}`);
+          warn(`Batch parse: ${missing.length} criteria missing from response (sequential fallback will be attempted, #2551): ${missing.join(", ")}`);
         }
         return results;
       }
@@ -66219,6 +66282,28 @@ function resolveJudgeProvider(config2) {
   }
   return { provider, apiKey, canUseManagedRouting };
 }
+function accumulateTokenUsageSummaries(primary, secondary) {
+  if (!primary && !secondary) return null;
+  const a = primary ?? { inputTokens: 0, outputTokens: 0 };
+  const b = secondary ?? { inputTokens: 0, outputTokens: 0 };
+  const callCount = typeof a.llmCallCount === "number" || typeof b.llmCallCount === "number" ? (a.llmCallCount ?? 0) + (b.llmCallCount ?? 0) : void 0;
+  const combined = {
+    inputTokens: a.inputTokens + b.inputTokens,
+    outputTokens: a.outputTokens + b.outputTokens,
+    ...callCount !== void 0 ? { llmCallCount: callCount } : {}
+  };
+  return combined;
+}
+async function evaluateMissingCriteriaSequentially(missingCriteria, context, config2) {
+  const results = /* @__PURE__ */ new Map();
+  let tokenUsage = null;
+  for (const criterion of missingCriteria) {
+    const singleResult = await evaluateWithLlm(criterion, context, config2);
+    results.set(criterion.id, singleResult.evaluation);
+    tokenUsage = accumulateTokenUsageSummaries(tokenUsage, singleResult.tokenUsage);
+  }
+  return { results, tokenUsage };
+}
 async function evaluateBatchWithLlm(criteria, context, config2) {
   const resolved = resolveJudgeProvider(config2);
   if (!resolved) {
@@ -66242,7 +66327,8 @@ async function evaluateBatchWithLlm(criteria, context, config2) {
     provider,
     traceLength: String(context.trace.length)
   });
-  const maxTokens = Math.min(256 + criteria.length * 512, 4096);
+  const BATCH_MAX_TOKENS_CEILING = 32768;
+  const maxTokens = Math.min(256 + criteria.length * 512, BATCH_MAX_TOKENS_CEILING);
   try {
     const response = await callLlmWithUsage({
       provider,
@@ -66269,6 +66355,17 @@ async function evaluateBatchWithLlm(criteria, context, config2) {
     const text = response.text;
     const criteriaIds = criteria.map((c) => c.id);
     const parsed = parseBatchJudgeResponse(text, criteriaIds);
+    const missingCriteria = criteria.filter((c) => !parsed.has(c.id));
+    let fallbackResults = /* @__PURE__ */ new Map();
+    let fallbackTokenUsage = null;
+    if (missingCriteria.length > 0) {
+      warn(
+        `Batch evaluation missing ${String(missingCriteria.length)}/${String(criteria.length)} criteria \u2014 falling back to sequential single-criterion evaluation (#2551)`
+      );
+      const fallback = await evaluateMissingCriteriaSequentially(missingCriteria, context, config2);
+      fallbackResults = fallback.results;
+      fallbackTokenUsage = fallback.tokenUsage;
+    }
     return {
       evaluations: criteria.map((c) => {
         const result = parsed.get(c.id);
@@ -66285,15 +66382,24 @@ async function evaluateBatchWithLlm(criteria, context, config2) {
             explanation: result.explanation
           };
         }
-        warn(`Batch response missing result for criterion "${c.id}" \u2014 defaulting to fail`);
+        const fallback = fallbackResults.get(c.id);
+        if (fallback) {
+          debug("LLM judge batch fallback (sequential)", {
+            criterion: c.id,
+            status: fallback.status,
+            confidence: fallback.confidence.toFixed(2)
+          });
+          return fallback;
+        }
+        warn(`Batch + sequential fallback both missing result for criterion "${c.id}" \u2014 defaulting to fail`);
         return {
           criterionId: c.id,
           status: "fail",
           confidence: 0.3,
-          explanation: "Criterion not found in batch evaluation response"
+          explanation: "Criterion not found in batch or sequential evaluation response"
         };
       }),
-      tokenUsage: response.usage
+      tokenUsage: accumulateTokenUsageSummaries(response.usage, fallbackTokenUsage)
     };
   } catch (err) {
     const message = errorMessage(err);
@@ -68749,7 +68855,7 @@ async function assertLocalHarnessBootable(harnessInput, explicitModel, extraEnv)
       command: harness.localCommand.command,
       args: harness.localCommand.args,
       timeoutMs,
-      cwd: harness.harnessDir,
+      cwd: harness.projectRoot,
       env: {
         ...harness.localCommand.env,
         ...extraEnv,
@@ -68762,19 +68868,21 @@ async function assertLocalHarnessBootable(harnessInput, explicitModel, extraEnv)
         ARCHAL_REST_CONFIG: restConfigPath,
         ARCHAL_METRICS_FILE: metricsFilePath,
         ARCHAL_AGENT_TRACE_FILE: traceFilePath,
-        ARCHAL_PREFLIGHT: "1"
+        ARCHAL_PREFLIGHT: "1",
+        ARCHAL_PROJECT_ROOT: harness.projectRoot,
+        ARCHAL_HARNESS_ENTRY_DIR: harness.harnessDir
       }
     });
     if (result.timedOut) {
       throw new Error(formatLocalHarnessFailure(
         `Harness boot timed out after ${Math.round(timeoutMs / 1e3)}s.`,
-        { cwd: harness.harnessDir, phase: "preflight" }
+        { cwd: harness.projectRoot, phase: "preflight" }
       ));
     }
     if (result.exitCode !== 0) {
       const detail = trimSnippet(result.stderr) || trimSnippet(result.stdout) || `Harness exited with code ${result.exitCode}.`;
       throw new Error(formatLocalHarnessFailure(detail, {
-        cwd: harness.harnessDir,
+        cwd: harness.projectRoot,
         phase: "preflight"
       }));
     }
@@ -68783,7 +68891,7 @@ async function assertLocalHarnessBootable(harnessInput, explicitModel, extraEnv)
       throw err;
     }
     throw new Error(formatLocalHarnessFailure(errorMessage(err), {
-      cwd: harness.harnessDir,
+      cwd: harness.projectRoot,
       phase: "preflight"
     }));
   } finally {
@@ -71576,11 +71684,11 @@ function resolveEngineMode(opts) {
   }
   if (opts.task && opts.task.trim()) {
     throw new CliUsageError(
-      'No agent configured.\n  --task still needs a runnable agent path.\n  Archal looked for: --harness, repo-local harness discovery, or .archal.json with "agent".\n  Next step: archal harness check ./.archal/harness.ts\n  Example:   archal run --task "List recent issues" --harness ./.archal/harness.ts --twin github\n  Or add:    { "agent": "npx tsx ./.archal/harness.ts", "twins": ["github"] }'
+      'No agent configured.\n  --task still needs a runnable agent path.\n  Archal looked for: --harness, repo-local harness discovery, or .archal.json with "agent".\n  Example:   archal run --task "List recent issues" --harness ./.archal/harness.ts --twin github\n  Or add:    { "agent": "npx tsx ./.archal/harness.ts", "twins": ["github"] }'
     );
   }
   throw new CliUsageError(
-    'No agent configured.\n  Archal could not find a runnable agent path.\n  Archal looked for: --harness, repo-local harness discovery, or .archal.json with "agent".\n  Next step: archal harness check ./.archal/harness.ts\n  Example:   archal run scenario.md --harness ./.archal/harness.ts\n  Or add:    { "agent": "npx tsx ./.archal/harness.ts", "twins": ["github"] }'
+    'No agent configured.\n  Archal could not find a runnable agent path.\n  Archal looked for: --harness, repo-local harness discovery, or .archal.json with "agent".\n  Example:   archal run scenario.md --harness ./.archal/harness.ts\n  Or add:    { "agent": "npx tsx ./.archal/harness.ts", "twins": ["github"] }'
   );
 }
 function resolveRequestedEngineModel(opts) {
@@ -71839,19 +71947,17 @@ function resolveHarnessAndEngine(opts, timeout, resolved) {
         if (opts.task && opts.task.trim()) {
           validationError(
             "No agent configured.",
-            '--task still needs a runnable agent path.\n  Archal looked for: --harness, repo-local harness discovery, or .archal.json with "agent".\n  Next step: archal harness check ./.archal/harness.ts\n  Example:   archal run --task "List recent issues" --harness ./.archal/harness.ts --twin github\n  Or add:    { "agent": "npx tsx ./.archal/harness.ts", "twins": ["github"] }'
+            '--task still needs a runnable agent path.\n  Archal looked for: --harness, repo-local harness discovery, or .archal.json with "agent".\n  Example:   archal run --task "List recent issues" --harness ./.archal/harness.ts --twin github\n  Or add:    { "agent": "npx tsx ./.archal/harness.ts", "twins": ["github"] }'
           );
         }
         validationError(
           "No agent configured.",
-          'Archal could not find a runnable agent path.\n  Archal looked for: --harness, repo-local harness discovery, or .archal.json with "agent".\n  Next step: archal harness check ./.archal/harness.ts\n  Example:   archal run scenario.md --harness ./.archal/harness.ts\n  Or add:    { "agent": "npx tsx ./.archal/harness.ts", "twins": ["github"] }'
+          'Archal could not find a runnable agent path.\n  Archal looked for: --harness, repo-local harness discovery, or .archal.json with "agent".\n  Example:   archal run scenario.md --harness ./.archal/harness.ts\n  Or add:    { "agent": "npx tsx ./.archal/harness.ts", "twins": ["github"] }'
         );
       }
       inferredRepoHarness = true;
       opts.harnessDir = inferredHarness.harnessDir;
-      info(
-        `Using repo-local harness from ${inferredHarness.harnessDir} (${inferredHarness.reason}).`
-      );
+      info(`Using repo-local harness from ${inferredHarness.harnessDir}.`);
     }
   }
   assertDockerAndSandboxAreCompatible(dockerExplicitlyRequested, opts.sandbox);
@@ -71924,6 +72030,21 @@ function resolveHarnessAndEngine(opts, timeout, resolved) {
   if (opts.sandbox) {
     return engine;
   }
+  if (engine.mode === "local" && opts.proxy === void 0) {
+    opts.proxy = true;
+    info("TLS proxy enabled by default for local harness runs. Use --no-proxy to disable.");
+    const existingProxy = process.env["HTTPS_PROXY"]?.trim() || process.env["https_proxy"]?.trim();
+    const existingCa = process.env["NODE_EXTRA_CA_CERTS"]?.trim();
+    if (existingProxy || existingCa) {
+      const collisions = [
+        existingProxy ? "HTTPS_PROXY" : null,
+        existingCa ? "NODE_EXTRA_CA_CERTS" : null
+      ].filter(Boolean).join(" / ");
+      warn(
+        `Auto-enabled --proxy will override your shell's ${collisions} for the agent process. If you rely on a corporate CA or custom proxy, pass --no-proxy and wire twin base URLs via ARCHAL_<TWIN>_BASE_URL env vars instead.`
+      );
+    }
+  }
   if (engine.mode === "local" && !process.env["ARCHAL_ENGINE_API_KEY"]) {
     const requestedModel = firstNonEmpty(
       engine.model,
@@ -72260,9 +72381,9 @@ function resolveRunConfigWithSources(scenarioPath, opts, sources, preloaded) {
         detail: "ARCHAL_MODEL"
       },
       {
-        value: toTrimmedString(archalConfig?.config.model),
+        value: toTrimmedString(archalConfig?.config.evaluatorModel),
         source: "project",
-        detail: ".archal.json:model"
+        detail: ".archal.json:evaluatorModel"
       },
       {
         value: toTrimmedString(dashboardConfig["runEvaluatorModel"]),
@@ -72293,11 +72414,6 @@ function resolveRunConfigWithSources(scenarioPath, opts, sources, preloaded) {
         source: "env",
         detail: "ARCHAL_DEFAULT_HARNESS"
       },
-      {
-        value: archalConfig?.config.harness ? (0, import_node_path38.resolve)(archalConfig.configDir, archalConfig.config.harness) : void 0,
-        source: "project",
-        detail: ".archal.json:harness"
-      },
       {
         value: toTrimmedString(dashboardConfig["runHarness"]),
         source: "project",
@@ -72583,9 +72699,6 @@ function applyResolvedRunConfigToOptions(opts, resolved) {
   if (resolved.model.source === "project") {
     opts.model = resolved.model.value;
   }
-  if (resolved.harness.value && resolved.harness.source === "project" && resolved.harness.detail === ".archal.json:harness" && !opts.harness) {
-    opts.harness = resolved.harness.value;
-  }
   const hasExplicitLocalHarness = (() => {
     if (firstNonEmpty(opts.harnessDir, process.env["ARCHAL_HARNESS_DIR"])) {
       return true;
@@ -78088,6 +78201,7 @@ async function executeRunForScenario(scenarioArg, opts, command, configDefaults,
           hostedSessionId: ctx.backendSessionId,
           isReusedEnvSession: ctx.isReusedEnvSession,
           keepState: opts.keepState,
+          freshSeed: opts.freshSeed,
           allowEmptyState: opts.allowEmptyState,
           noSeedCache: !opts.seedCache,
           staticSeed: hasStaticSeed,
@@ -78191,8 +78305,33 @@ function deriveTitle(task) {
   const lastSpace = truncated.lastIndexOf(" ");
   return lastSpace > 20 ? truncated.slice(0, lastSpace) : truncated;
 }
+var LEADING_CONTEXT_PATTERNS = [
+  /^using\s+[^,]+?,\s+/,
+  /^(?:please|kindly)\s+/,
+  /^(?:can|could|would|will)\s+you\s+(?:please\s+)?/,
+  /^(?:i\s+want\s+(?:you\s+)?to|i\s+(?:would\s+)?(?:'d\s+)?like\s+(?:you\s+)?to|i\s+need\s+(?:you\s+)?to|i'?d\s+like\s+(?:you\s+)?to)\s+/,
+  /^help\s+me\s+(?:please\s+)?/,
+  /^for\s+me,\s+/
+];
+function stripLeadingContext(input) {
+  let cur = input;
+  for (let i = 0; i < 4; i++) {
+    let changed = false;
+    for (const pattern of LEADING_CONTEXT_PATTERNS) {
+      const next = cur.replace(pattern, "");
+      if (next !== cur) {
+        cur = next;
+        changed = true;
+      }
+    }
+    if (!changed) break;
+  }
+  return cur;
+}
 function isReadOnlyTask(task) {
-  const normalized = task.trim().toLowerCase().replace(/\s+/g, " ").replace(/^(please|can you|could you|would you|help me)\s+/, "");
+  const normalized = stripLeadingContext(
+    task.trim().toLowerCase().replace(/\s+/g, " ")
+  );
   const writePrefixes = [
     "mark ",
     "star ",
@@ -78298,10 +78437,10 @@ function isReadOnlyTask(task) {
   return true;
 }
 function generateTaskScenario(options) {
-  const { task, twins } = options;
+  const { task, twins, readOnlyOverride } = options;
   const title = deriveTitle(task);
   const twinsConfig = twins.join(", ");
-  const readOnly = isReadOnlyTask(task);
+  const readOnly = typeof readOnlyOverride === "boolean" ? readOnlyOverride : isReadOnlyTask(task);
   const criteria = [];
   if (readOnly) {
     criteria.push(
@@ -78474,7 +78613,11 @@ async function resolveRunCommandScenarios(scenarioArg, opts, command) {
     if (resolvedTwins.length === 0) {
       throw new CliUsageError("--task requires --twin (could not auto-detect).");
     }
-    taskScenarioPath = generateTaskScenario({ task: opts.task, twins: resolvedTwins });
+    taskScenarioPath = generateTaskScenario({
+      task: opts.task,
+      twins: resolvedTwins,
+      readOnlyOverride: opts.readOnly === true ? true : void 0
+    });
     scenariosToRun.push(taskScenarioPath);
     if (!opts.seed && !opts.seedOverrides && !opts.fileSeedPaths) {
       opts.seed = "empty";
@@ -78516,6 +78659,7 @@ function applyRunHelpVisibility(cmd, showAdvancedHelp) {
     "--config",
     "--task",
     "--twin",
+    "--read-only",
     "--runs",
     "--timeout",
     "--output",
@@ -78526,6 +78670,7 @@ function applyRunHelpVisibility(cmd, showAdvancedHelp) {
     "--verbose",
     "--reuse-session",
     "--keep-state",
+    "--fresh-seed",
     "--help"
   ]);
   for (const option of cmd.options) {
@@ -78552,12 +78697,17 @@ function createRunCommand() {
       return acc;
     },
     []
+  ).addOption(
+    new Option(
+      "--read-only",
+      "Force the generated --task scenario to use the read-only (single-criterion) scaffold, bypassing the heuristic classifier. Only meaningful alongside --task."
+    )
   ).option("-n, --runs <count>", "Number of runs", String(configDefaults.runs)).option("-t, --timeout <seconds>", "Timeout per run in seconds", String(configDefaults.timeout)).addOption(new Option("-m, --model <model>", "Evaluator model").hideHelp()).addOption(
     new Option("-o, --output <format>", "Output format: terminal, json").default("terminal")
   ).addOption(new Option("--seed <name-or-path>", "Seed name or file path (.json / .md)")).addOption(new Option("--tag <tag>", "Only run if scenario has this tag")).addOption(new Option("-q, --quiet", "Suppress non-error output")).addOption(new Option("-v, --verbose", "Enable debug logging")).addOption(
     new Option(
       "--harness <path>",
-      "Path to a runnable headless harness (e.g. ./.archal/harness.ts). Omit this flag and Archal discovers a repo-local harness by walking up from cwd for a top-level harness.{ts,js,mjs,cjs}, archal-harness.{ts,js,mjs,cjs}, or package.json#archal.harness. You can also set `agent` in .archal.json."
+      "Path to a runnable headless harness (e.g. ./.archal/harness.ts). Omit this flag and Archal discovers a repo-local harness by walking up from cwd for a top-level harness.{ts,js,mjs,cjs}. You can also set `agent` in .archal.json."
     )
   ).addOption(
     new Option(
@@ -78570,18 +78720,28 @@ function createRunCommand() {
       "Skip the scenario's named-seed re-apply on a reused session. Use after sideloading twin state with 'archal twin seed --file'. Alias: --no-reseed."
     )
   ).addOption(new Option("--no-reseed").hideHelp()).addOption(
+    new Option(
+      "--fresh-seed",
+      "Force re-apply of the scenario's named seed on a reused session, wiping any existing twin state first. Opposite of --keep-state. Use when an earlier `archal twin start` left state around and you want the scenario's declared seed applied cleanly. Mutually exclusive with --keep-state."
+    )
+  ).addOption(
     new Option(
       "--allow-empty-state",
       "Allow the run to proceed when the reused-session state probe fails. Rare debug-only; default off \u2014 a probe failure normally fails the run to protect sideloaded state."
     ).hideHelp()
-  ).addOption(new Option("--rate-limit <count>").hideHelp()).addOption(new Option("--pass-threshold <score>").default("0").hideHelp()).addOption(new Option("--api-key <key>").hideHelp()).addOption(new Option("--engine-endpoint <url>").hideHelp()).addOption(new Option("--engine-token <token>").hideHelp()).addOption(new Option("--agent-model <model>").hideHelp()).addOption(new Option("--engine-twin-urls <path>").hideHelp()).addOption(new Option("--engine-timeout <seconds>").hideHelp()).addOption(new Option("--dockerfile <path>").hideHelp()).addOption(new Option("--api-base-urls <path>").hideHelp()).addOption(new Option("--api-proxy-url <url>").hideHelp()).addOption(new Option("--preflight-only").hideHelp()).addOption(new Option("--seed-cache").hideHelp()).addOption(new Option("--clear-seed-cache").hideHelp()).addOption(new Option("--replay-seed <path>").hideHelp()).addOption(new Option("--save-seed <path>").hideHelp()).addOption(new Option("--no-failure-analysis").hideHelp()).addOption(new Option("--allow-ambiguous-seed").hideHelp()).addOption(new Option("--strict-seed").hideHelp()).addOption(new Option("--run-project-id <id>").hideHelp()).addOption(new Option("--run-project-name <name>").hideHelp()).addOption(new Option("--run-project-description <text>").hideHelp()).addOption(new Option("--sandbox").hideHelp()).addOption(new Option("--openclaw-home <dir>").hideHelp()).addOption(new Option("--workspace <dir>").hideHelp()).addOption(new Option("--openclaw-config <path>").hideHelp()).addOption(new Option("--openclaw-version <version>").hideHelp()).addOption(new Option("--openclaw-eval-mode <mode>").hideHelp()).addOption(new Option("--docker").hideHelp()).addOption(new Option("--no-docker").hideHelp()).addOption(new Option("--proxy", "Route agent HTTP traffic through TLS proxy to twins")).addOption(new Option("--advanced").hideHelp()).addHelpText(
+  ).addOption(new Option("--rate-limit <count>").hideHelp()).addOption(new Option("--pass-threshold <score>").default("0").hideHelp()).addOption(new Option("--api-key <key>").hideHelp()).addOption(new Option("--engine-endpoint <url>").hideHelp()).addOption(new Option("--engine-token <token>").hideHelp()).addOption(new Option("--agent-model <model>").hideHelp()).addOption(new Option("--engine-twin-urls <path>").hideHelp()).addOption(new Option("--engine-timeout <seconds>").hideHelp()).addOption(new Option("--dockerfile <path>").hideHelp()).addOption(new Option("--api-base-urls <path>").hideHelp()).addOption(new Option("--api-proxy-url <url>").hideHelp()).addOption(new Option("--preflight-only").hideHelp()).addOption(new Option("--seed-cache").hideHelp()).addOption(new Option("--clear-seed-cache").hideHelp()).addOption(new Option("--replay-seed <path>").hideHelp()).addOption(new Option("--save-seed <path>").hideHelp()).addOption(new Option("--no-failure-analysis").hideHelp()).addOption(new Option("--allow-ambiguous-seed").hideHelp()).addOption(new Option("--strict-seed").hideHelp()).addOption(new Option("--run-project-id <id>").hideHelp()).addOption(new Option("--run-project-name <name>").hideHelp()).addOption(new Option("--run-project-description <text>").hideHelp()).addOption(new Option("--sandbox").hideHelp()).addOption(new Option("--openclaw-home <dir>").hideHelp()).addOption(new Option("--workspace <dir>").hideHelp()).addOption(new Option("--openclaw-config <path>").hideHelp()).addOption(new Option("--openclaw-version <version>").hideHelp()).addOption(new Option("--openclaw-eval-mode <mode>").hideHelp()).addOption(new Option("--docker").hideHelp()).addOption(new Option("--no-docker").hideHelp()).addOption(new Option("--proxy", "Force the TLS proxy on (on by default for local harness runs).").hideHelp()).addOption(new Option("--no-proxy", "Disable the TLS proxy for local harness runs.").hideHelp()).addOption(new Option("--advanced").hideHelp()).addHelpText(
     "after",
-    '\nExamples:\n  archal run                                    # uses .archal.json in current directory\n  archal run --config path/to/.archal.json      # use specific config\n  archal run scenario.md                        # run a single scenario directly\n  archal run scenario.md --runs 5               # run 5 times\n  archal harness check ./.archal/harness.ts     # verify a repo-local headless harness\n  archal run --task "List recent issues" --harness ./.archal/harness.ts --twin github\n                                                # inline task with an explicit harness path\n  archal twin start github                      # then in the same shell:\n  archal run scenario.md --reuse-session        # runs against the active twin session\n  archal run scenario.md --reuse-session ses-abc # targets a specific session id\n\n.archal.json (one valid project-default setup path):\n  {\n    "title": "my-agent-tests",\n    "agent": "npx tsx ./.archal/harness.ts",\n    "twins": ["github"],\n    "agentModel": "gemini-2.0-flash"\n  }\n\nFor full reference: https://docs.archal.ai/cli/run\n'
+    '\nExamples:\n  archal run                                    # uses .archal.json in current directory\n  archal run --config path/to/.archal.json      # use specific config\n  archal run scenario.md                        # run a single scenario directly\n  archal run scenario.md --runs 5               # run 5 times\n  archal run --task "List recent issues" --harness ./.archal/harness.ts --twin github\n                                                # inline task with an explicit harness path\n  archal twin start github                      # then in the same shell:\n  archal run scenario.md --reuse-session        # runs against the active twin session\n  archal run scenario.md --reuse-session ses-abc # targets a specific session id\n\n.archal.json (one valid project-default setup path):\n  {\n    "title": "my-agent-tests",\n    "agent": "npx tsx ./.archal/harness.ts",\n    "twins": ["github"],\n    "agentModel": "gemini-2.0-flash"\n  }\n\nFor full reference: https://docs.archal.ai/cli/run\n'
   ).action(async (scenarioArg, opts, command) => {
     const reseedOpt = opts.reseed;
     if (reseedOpt === false) {
       opts.keepState = true;
     }
+    if (opts.freshSeed && opts.keepState) {
+      command.error(
+        "error: --fresh-seed and --keep-state are mutually exclusive. Pick one: --fresh-seed to wipe and re-apply the scenario's seed, or --keep-state to preserve the existing twin state."
+      );
+    }
     const {
       archalFile,
       configuredAgentConfig,
@@ -79279,7 +79439,6 @@ function isRunnableScenarioCandidate(filePath) {
 }
 function toLocalScenarioRows(opts) {
   const tagFilter = opts.tag?.toLowerCase();
-  const difficultyFilter = opts.difficulty?.toLowerCase();
   const rows = [];
   const localResolution = findLocalScenariosDir();
   const localDir = localResolution.dir;
@@ -79299,7 +79458,6 @@ function toLocalScenarioRows(opts) {
         const scenarioTags = scenario.config.tags.map((t) => t.toLowerCase());
         if (!scenarioTags.includes(tagFilter)) continue;
       }
-      if (difficultyFilter && (scenario.config.difficulty ?? "") !== difficultyFilter) continue;
       rows.push({
         scenario: scenario.title,
         slug: (0, import_node_path49.relative)(localDir, file2).replace(/\\/gu, "/").replace(/\.md$/u, "") || (0, import_node_path49.basename)(file2, ".md"),
@@ -79313,13 +79471,9 @@ function toLocalScenarioRows(opts) {
 }
 function matchesScenarioFilters(scenario, opts) {
   const tagFilter = opts.tag?.toLowerCase();
-  const difficultyFilter = opts.difficulty?.toLowerCase();
   if (tagFilter && !scenario.tags.some((tag) => tag.toLowerCase() === tagFilter)) {
     return false;
   }
-  if (difficultyFilter && (scenario.difficulty ?? "") !== difficultyFilter) {
-    return false;
-  }
   return true;
 }
 async function listScenarios(opts) {
@@ -79385,7 +79539,7 @@ Found ${rows.length} scenario(s)`);
 }
 function createScenarioCommand() {
   const scenario = new Command("scenario").description("Browse and manage scenarios");
-  scenario.command("list").description("List runnable scenarios").option("--tag <tag>", "Filter scenarios by tag").option("--difficulty <level>", "Filter by difficulty (easy, medium, hard)").option("--json", "Output as JSON").action(async (opts) => {
+  scenario.command("list").description("List runnable scenarios").option("--tag <tag>", "Filter scenarios by tag").option("--json", "Output as JSON").action(async (opts) => {
     await listScenarios(opts);
   });
   return scenario;
@@ -80440,6 +80594,21 @@ function sortSessionsByCreatedAt(sessions) {
 function isActiveTwinSession(session) {
   return session.status !== "ended" && session.status !== "expired" && session.status !== "completed";
 }
+async function probeSavedTwinSession(token, saved) {
+  const result = await getSessionStatus(token, saved.sessionId, saved.controlPlaneBaseUrl);
+  if (!result.ok) {
+    if (result.status === 404) {
+      clearSession();
+      return { state: "stale", reason: "not found" };
+    }
+    throw new CliRuntimeError(`Server error fetching session status: ${result.error}`);
+  }
+  if (!isActiveTwinSession(result.data)) {
+    clearSession();
+    return { state: "stale", reason: result.data.status };
+  }
+  return { state: "alive", response: result.data };
+}
 var WORKER_HEALTH_TIMEOUT_BASE_MS = 6e4;
 var WORKER_HEALTH_TIMEOUT_PER_TWIN_MS = 2e4;
 var WORKER_HEALTH_TIMEOUT_MAX_MS = 18e4;
@@ -80759,14 +80928,21 @@ function createTwinCommand() {
     });
   });
   command.command("start").argument("[twins...]", "Twins to start").option("--all", "Start all available twins").option("--seed <seeds...>", "Seeds per twin (for example: stripe:small-business github:enterprise-repo)").option("--setup <description>", "Describe desired state in natural language and seed the first twin after startup").option("--setup-file <path>", "Read setup description from a file and seed the first twin after startup").option("--seed-file <path>", "Load a JSON or markdown seed file after the first twin is ready").option("--name <name>", "Custom session name shown on the dashboard (defaults to twin names)").option("--ttl-seconds <seconds>", "Requested twin lifetime in seconds (capped server-side)").option("--fresh", "Bypass idempotency replay \u2014 always create a new session even if request params match a recent one").description("Start a persistent hosted twin session with hosted twin APIs").action(async (twins, opts) => {
+    const token = requireToken();
     const existing = loadSession();
     if (existing) {
-      throw new CliRuntimeError(
-        `Active session found: ${existing.sessionId}
+      const probe = await probeSavedTwinSession(token, existing);
+      if (probe.state === "alive") {
+        throw new CliRuntimeError(
+          `Active session found: ${existing.sessionId}
 Run 'archal twin stop' first, or 'archal twin status' to inspect it.`
+        );
+      }
+      process.stderr.write(
+        `${DIM6}[twin] Clearing stale local session ${existing.sessionId} (${probe.reason}); starting a new one.${RESET6}
+`
       );
     }
-    const token = requireToken();
     const requestedTwins = await resolveRequestedTwins(token, twins, opts.all);
     if (requestedTwins.length === 0) {
       if (opts.all) {
@@ -80843,25 +81019,15 @@ Run 'archal twin stop' first, or 'archal twin status' to inspect it.`
 `);
       return;
     }
-    const result = await getSessionStatus(token, saved.sessionId, saved.controlPlaneBaseUrl);
-    if (!result.ok) {
-      if (result.status === 404) {
-        process.stderr.write(`${DIM6}Session ${saved.sessionId} is no longer active (not found).${RESET6}
-`);
-        clearSession();
-        return;
-      }
-      throw new CliRuntimeError(`Server error fetching session status: ${result.error}`);
-    }
-    const response = result.data;
-    if (response.status === "ended" || response.status === "expired" || response.status === "completed") {
+    const probe = await probeSavedTwinSession(token, saved);
+    if (probe.state === "stale") {
       process.stderr.write(
-        `${DIM6}Session ${saved.sessionId} is no longer active (${response.status}).${RESET6}
+        `${DIM6}Session ${saved.sessionId} is no longer active (${probe.reason}).${RESET6}
 `
       );
-      clearSession();
       return;
     }
+    const response = probe.response;
     if (opts.json) {
       process.stdout.write(`${JSON.stringify({ ...response, sessionId: saved.sessionId }, null, 2)}
 `);

package/dist/seed/dynamic-generator.cjs

@@ -26994,7 +26994,6 @@ var scenarioConfigSchema = external_exports3.object({
   timeout: external_exports3.number().default(120),
   runs: external_exports3.number().default(5),
   evaluatorModel: external_exports3.string().optional(),
-  difficulty: external_exports3.enum(["easy", "medium", "hard"]).optional(),
   tags: external_exports3.array(external_exports3.string()).default([])
 });
 

package/dist/vitest/{chunk-YV6BH6DO.js → chunk-KTMNDJFB.js}

@@ -27194,7 +27194,6 @@ var scenarioConfigSchema = external_exports3.object({
   timeout: external_exports3.number().default(120),
   runs: external_exports3.number().default(5),
   evaluatorModel: external_exports3.string().optional(),
-  difficulty: external_exports3.enum(["easy", "medium", "hard"]).optional(),
   tags: external_exports3.array(external_exports3.string()).default([])
 });
 

package/dist/vitest/{chunk-RKYS44AS.js → chunk-L6HSMJ3F.js}

@@ -18,7 +18,7 @@ import {
   requestedSeedsMatchSession,
   sleep,
   trimEnv
-} from "./chunk-YV6BH6DO.js";
+} from "./chunk-KTMNDJFB.js";
 
 // src/runtime-module-resolution.ts
 import { dirname, extname, resolve } from "path";

package/dist/vitest/index.cjs

@@ -30086,7 +30086,6 @@ var scenarioConfigSchema = external_exports3.object({
   timeout: external_exports3.number().default(120),
   runs: external_exports3.number().default(5),
   evaluatorModel: external_exports3.string().optional(),
-  difficulty: external_exports3.enum(["easy", "medium", "hard"]).optional(),
   tags: external_exports3.array(external_exports3.string()).default([])
 });
 

package/dist/vitest/index.js

@@ -8,11 +8,11 @@ import {
   readArchalVitestConfig,
   resetArchalTwins,
   resolveRuntimeModule
-} from "./chunk-RKYS44AS.js";
+} from "./chunk-L6HSMJ3F.js";
 import {
   encodeConfig,
   getSessionIdFilePath
-} from "./chunk-YV6BH6DO.js";
+} from "./chunk-KTMNDJFB.js";
 import {
   __commonJS,
   __toESM

package/dist/vitest/runtime/hosted-session-reaper.cjs

@@ -27434,7 +27434,6 @@ var scenarioConfigSchema = external_exports3.object({
   timeout: external_exports3.number().default(120),
   runs: external_exports3.number().default(5),
   evaluatorModel: external_exports3.string().optional(),
-  difficulty: external_exports3.enum(["easy", "medium", "hard"]).optional(),
   tags: external_exports3.array(external_exports3.string()).default([])
 });
 

package/dist/vitest/runtime/hosted-session-reaper.js

@@ -4,7 +4,7 @@ import {
   createHostedAuthLease,
   parsePositiveInteger,
   runHostedSessionReaper
-} from "../chunk-YV6BH6DO.js";
+} from "../chunk-KTMNDJFB.js";
 import "../chunk-YJICENME.js";
 
 // src/runtime/hosted-session-reaper.ts

package/dist/vitest/runtime/setup-files.js

@@ -1,7 +1,7 @@
 import {
   bootstrapArchalVitestRouting
-} from "../chunk-RKYS44AS.js";
-import "../chunk-YV6BH6DO.js";
+} from "../chunk-L6HSMJ3F.js";
+import "../chunk-KTMNDJFB.js";
 import "../chunk-YJICENME.js";
 
 // src/runtime/setup-files.ts

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "archal",
-  "version": "0.9.9",
+  "version": "0.9.10",
   "description": "Test your agents & integrations against digital twins",
   "type": "module",
   "main": "dist/index.cjs",