@nathapp/nax 0.70.0-canary.5 → 0.70.0-canary.7

package/dist/nax.js

@@ -16828,7 +16828,8 @@ var init_schemas_execution = __esm(() => {
     escalation: exports_external.object({
       enabled: exports_external.boolean(),
       tierOrder: exports_external.array(TierConfigSchema).min(1, { message: "tierOrder must have at least one tier" }),
-      escalateEntireBatch: exports_external.boolean().optional()
+      escalateEntireBatch: exports_external.boolean().optional(),
+      resetMode: exports_external.enum(["initial", "last"]).default("initial")
     })
   });
   RectificationConfigSchema = exports_external.object({
@@ -16872,7 +16873,7 @@ var init_schemas_execution = __esm(() => {
   }).superRefine((value, ctx) => {
     if (value.mode !== "provision" && value.setupCommand !== null) {
       ctx.addIssue({
-        code: exports_external.ZodIssueCode.custom,
+        code: "custom",
         path: ["setupCommand"],
         message: "execution.worktreeDependencies.setupCommand requires mode 'provision'"
       });
@@ -17354,7 +17355,8 @@ var init_schemas3 = __esm(() => {
           { tier: "balanced", attempts: 3 },
           { tier: "powerful", attempts: 2 }
         ],
-        escalateEntireBatch: true
+        escalateEntireBatch: true,
+        resetMode: "initial"
       }
     }),
     routing: RoutingConfigSchema.default({
@@ -27956,7 +27958,8 @@ var init_static_rules = __esm(() => {
 
 // src/prd/types.ts
 function getContextFiles(story) {
-  const files = story.contextFiles ?? story.relevantFiles ?? [];
+  const legacyFiles = story.relevantFiles;
+  const files = story.contextFiles ?? legacyFiles ?? [];
   return files.map((f) => typeof f === "string" ? f : f.path);
 }
 function getExpectedFiles(story) {
@@ -28468,16 +28471,28 @@ function markStoryFailed(prd, storyId, failureCategory, failureStage, statusWrit
     }
   }
 }
-function resetFailedStoriesToPending(prd, resetRef = false, storyIsolation) {
+function resetFailedStoriesToPending(prd, opts = {}) {
+  const { resetRef = false, storyIsolation, resetMode = "initial" } = opts;
   const reset = [];
   for (const story of prd.userStories) {
-    if (story.status === "failed") {
-      story.status = "pending";
-      if (resetRef || storyIsolation === "worktree") {
-        story.storyGitRef = undefined;
-      }
-      reset.push(story);
+    if (story.status !== "failed")
+      continue;
+    story.status = "pending";
+    story.attempts = 0;
+    if (resetMode === "initial" && story.routing) {
+      story.routing = {
+        ...story.routing,
+        ...story.routing.initialModelTier !== undefined && {
+          modelTier: story.routing.initialModelTier
+        },
+        agent: story.routing.initialAgent
+      };
+      story.escalations = [];
+    }
+    if (resetRef || storyIsolation === "worktree") {
+      story.storyGitRef = undefined;
     }
+    reset.push(story);
   }
   return reset;
 }
@@ -29675,10 +29690,12 @@ function formatPriorFailures(failures) {
     return "";
   }
   const parts = [];
-  parts.push(`## Prior Failures (Structured Context)
-`);
   for (const failure of failures) {
     parts.push(`### Attempt ${failure.attempt} \u2014 ${failure.modelTier}`);
+    if (failure.agent) {
+      const profilePart = failure.agentProfileId ? ` (profile: ${failure.agentProfileId})` : "";
+      parts.push(`**Agent:** ${failure.agent}${profilePart}`);
+    }
     parts.push(`**Stage:** ${failure.stage}`);
     parts.push(`**Summary:** ${failure.summary}`);
     if (failure.testFailures && failure.testFailures.length > 0) {
@@ -29784,10 +29801,14 @@ function formatContextAsMarkdown(built) {
     byType.set(element.type, existing);
   }
   renderSection(sections, byType, "progress", `## Progress
+`, renderSimple);
+  renderSection(sections, byType, "prior-failures", `## Prior Failures (Structured Context)
 `, renderSimple);
   renderErrorSection(sections, byType);
   renderSection(sections, byType, "test-coverage", "", renderSimple);
   renderSection(sections, byType, "story", `## Current Story
+`, renderSimple);
+  renderSection(sections, byType, "planning-analysis", `## Planning Analysis
 `, renderSimple);
   renderSection(sections, byType, "dependency", `## Dependency Stories
 `, renderSimple);
@@ -36411,7 +36432,7 @@ var init_acceptance_refine = __esm(() => {
     jsonMode: true,
     config: acceptanceConfigSelector,
     retry: { preset: "transient-network", maxAttempts: 2, baseDelayMs: 0 },
-    model: (_input, ctx) => ctx.config.acceptance.generateModel ?? ctx.config.acceptance.model,
+    model: (_input, ctx) => ctx.config.acceptance.model,
     timeoutMs: (_input, ctx) => ctx.config.acceptance.timeoutMs,
     build(input, _ctx) {
       const prompt = new AcceptancePromptBuilder().buildRefinementPrompt(input.criteria, input.codebaseContext, {
@@ -52512,6 +52533,7 @@ async function collectStoryMetrics(ctx, storyStartTime) {
     startedAt: storyStartTime,
     completedAt: new Date().toISOString(),
     fullSuiteGatePassed,
+    ...ctx.fullSuiteGateFailingFiles && ctx.fullSuiteGateFailingFiles.length > 0 ? { failingTestFiles: ctx.fullSuiteGateFailingFiles } : {},
     runtimeCrashes: ctx.storyRuntimeCrashes ?? 0,
     tokens: agentResult?.tokenUsage ? new TokenUsage({
       inputTokens: agentResult.tokenUsage.inputTokens,
@@ -55044,6 +55066,15 @@ function extractPhaseFindings(output) {
   const success2 = "success" in record2 ? record2.success === true : ("passed" in record2) ? record2.passed === true : findings.length === 0;
   return success2 ? [] : findings;
 }
+function gateFailureKeys(gateOutput) {
+  const keys = new Set;
+  for (const f of extractPhaseFindings(gateOutput)) {
+    if (f.source !== "test-runner")
+      continue;
+    keys.add(`${f.file ?? ""}::${f.rule ?? ""}`);
+  }
+  return keys;
+}
 function shouldSkipPhaseForRectification(phase, state, phaseOutputs) {
   if (phase.kind !== "full-suite-gate")
     return false;
@@ -55433,6 +55464,8 @@ class ExecutionPlan {
         break;
       }
     }
+    const gateName = this.state.fullSuiteGate?.slot.op.name;
+    const preRectGateFailureKeys = gateName ? gateFailureKeys(phaseOutputs[gateName]) : new Set;
     const rectResult = await runRectification(this.ctx, this.state, phaseCosts, phaseOutputs);
     if (this.state.rectification && (!rectResult.rectificationExhausted || rectResult.liteScopeIncomplete)) {
       let resumeRectifyUsed = false;
@@ -55526,9 +55559,12 @@ class ExecutionPlan {
       });
     }
     const verifierName = this.state.verifier?.slot.op.name;
-    const gateName = this.state.fullSuiteGate?.slot.op.name;
-    const verifierPassedSsot = verifierName !== undefined && phaseExplicitlyPassed(phaseOutputs[verifierName]);
-    if (verifierPassedSsot && gateName !== undefined && !phasePassed(gateName, phaseOutputs[gateName], this.ctx.storyId)) {
+    const verifierExplicitlyPassed = verifierName !== undefined && phaseExplicitlyPassed(phaseOutputs[verifierName]);
+    const gateRegressedDuringRect = gateName !== undefined && [...gateFailureKeys(phaseOutputs[gateName])].some((k) => !preRectGateFailureKeys.has(k));
+    const verifierPassedSsot = verifierExplicitlyPassed && !gateRegressedDuringRect;
+    if (verifierExplicitlyPassed && gateRegressedDuringRect) {
+      logger?.warn("story-orchestrator", "Gate regressed during rectification after verifier passed \u2014 verifier verdict is stale, failing story", { storyId: this.ctx.storyId, packageDir: this.ctx.packageDir });
+    } else if (verifierPassedSsot && gateName !== undefined && !phasePassed(gateName, phaseOutputs[gateName], this.ctx.storyId)) {
       logger?.warn("story-orchestrator", "Full-suite gate failed but verifier judged story OK \u2014 treating gate failures as unrelated regressions", { storyId: this.ctx.storyId, packageDir: this.ctx.packageDir });
     }
     const success2 = Object.entries(phaseOutputs).every(([name, output]) => {
@@ -55566,7 +55602,8 @@ class ExecutionPlan {
       totalCostUsd,
       durationMs,
       phaseOutputs,
-      ...rectResult
+      ...rectResult,
+      gateRegressedDuringRect
     };
   }
 }
@@ -56217,7 +56254,7 @@ function extractPauseReason(phaseOutputs) {
   }
   return;
 }
-function deriveTddFailureCategory(phaseOutputs, unfixedFindings) {
+function deriveTddFailureCategory(phaseOutputs, unfixedFindings, gateRegressedDuringRect) {
   const testWriterOutput = phaseOutputs[testWriterOp.name];
   if (testWriterOutput?.success === false) {
     return "session-failure";
@@ -56233,7 +56270,7 @@ function deriveTddFailureCategory(phaseOutputs, unfixedFindings) {
     }
     return "tests-failing";
   }
-  const verifierPassed = verifierOutput?.success === true;
+  const verifierPassed = verifierOutput?.success === true && !gateRegressedDuringRect;
   if (!verifierPassed && unfixedFindings && unfixedFindings.length > 0) {
     const rectOutput = phaseOutputs.rectification;
     if (rectOutput?.exitReason && EXHAUSTED_EXIT_REASONS.has(rectOutput.exitReason) && unfixedFindings.some((f) => f.source === "test-runner")) {
@@ -56282,6 +56319,11 @@ async function applyPostRunInspection(ctx, planResult, opts) {
   if (fullSuiteGateOutput?.passed) {
     ctx.fullSuiteGatePassed = true;
   }
+  const gateFailingFiles = [
+    ...new Set((fullSuiteGateOutput?.findings ?? []).map((f) => f.file).filter((f) => !!f))
+  ];
+  if (gateFailingFiles.length > 0)
+    ctx.fullSuiteGateFailingFiles = gateFailingFiles;
   ctx.selfVerification = parseSelfVerificationMarker(agentResult.output ?? "", ctx.workdir);
   const selfVerificationFailed = ctx.selfVerification.lint === "fail" || ctx.selfVerification.typecheck === "fail";
   if (ctx.config.context?.v2?.enabled && ctx.sessionScratchDir) {
@@ -56348,7 +56390,7 @@ async function applyPostRunInspection(ctx, planResult, opts) {
     }
   }
   const pauseReason = extractPauseReason(planResult.phaseOutputs);
-  const failureCategory = isTdd && !planResult.success ? deriveTddFailureCategory(planResult.phaseOutputs, planResult.unfixedFindings) : undefined;
+  const failureCategory = isTdd && !planResult.success ? deriveTddFailureCategory(planResult.phaseOutputs, planResult.unfixedFindings, planResult.gateRegressedDuringRect) : undefined;
   if (isTdd && !planResult.success && !failureCategory) {
     const phaseSignals = {};
     for (const [name, output] of Object.entries(planResult.phaseOutputs)) {
@@ -57203,6 +57245,7 @@ var init_routing2 = __esm(() => {
       const neverEscalated = !hasEscalationRecords;
       const initialAgent = ctx.story.routing?.initialAgent ?? (neverEscalated ? routing.agent : undefined);
       const initialProfileId = ctx.story.routing?.initialProfileId ?? (neverEscalated ? ctx.story.routing?.agentProfileId : undefined);
+      const initialModelTier = ctx.story.routing?.initialModelTier ?? (neverEscalated ? routing.modelTier : undefined);
       ctx.story.routing = {
         ...ctx.story.routing ?? {},
         complexity: routing.complexity,
@@ -57212,7 +57255,8 @@ var init_routing2 = __esm(() => {
         modelTier: routing.modelTier,
         ...routing.agent !== undefined && { agent: routing.agent },
         ...initialAgent !== undefined && { initialAgent },
-        ...initialProfileId !== undefined && { initialProfileId }
+        ...initialProfileId !== undefined && { initialProfileId },
+        ...initialModelTier !== undefined && { initialModelTier }
       };
       if (ctx.prdPath) {
         await _routingDeps.savePRD(ctx.prd, ctx.prdPath);
@@ -60425,7 +60469,7 @@ var package_default;
 var init_package = __esm(() => {
   package_default = {
     name: "@nathapp/nax",
-    version: "0.70.0-canary.5",
+    version: "0.70.0-canary.7",
     description: "AI Coding Agent Orchestrator \u2014 loops until done",
     type: "module",
     bin: {
@@ -60520,8 +60564,8 @@ var init_version = __esm(() => {
   NAX_VERSION = package_default.version;
   NAX_COMMIT = (() => {
     try {
-      if (/^[0-9a-f]{6,10}$/.test("e8c2ab46"))
-        return "e8c2ab46";
+      if (/^[0-9a-f]{6,10}$/.test("08567b02"))
+        return "08567b02";
     } catch {}
     try {
       const result = Bun.spawnSync(["git", "rev-parse", "--short", "HEAD"], {
@@ -61475,6 +61519,15 @@ async function findResponsibleStory(testFile, workdir, passedStories) {
   }
   return;
 }
+function findResponsibleStoryByTransition(testFile, snapshots) {
+  const ordered = [...snapshots].sort((a, b) => a.completedAt.localeCompare(b.completedAt) || a.storyId.localeCompare(b.storyId));
+  for (const snap of ordered) {
+    if (snap.failingTestFiles?.includes(testFile)) {
+      return snap.storyId;
+    }
+  }
+  return;
+}
 async function runDeferredRegression(options) {
   const logger = getSafeLogger();
   const { config: config2, prd, workdir, runtime } = options;
@@ -61622,8 +61675,21 @@ async function runDeferredRegression(options) {
     }
   } else {
     const testFilesArray = Array.from(testFilesInFailures);
+    const snapshots = options.storyMetrics ?? [];
+    const passedById = new Map(passedStories.map((s) => [s.id, s]));
     for (const testFile of testFilesArray) {
-      const responsibleStory = await findResponsibleStory(testFile, workdir, passedStories);
+      let responsibleStory;
+      const transitionId = findResponsibleStoryByTransition(testFile, snapshots);
+      if (transitionId && passedById.has(transitionId)) {
+        responsibleStory = passedById.get(transitionId);
+        logger?.info("regression", "Mapped test to story via gate transition", {
+          storyId: transitionId,
+          testFile
+        });
+      }
+      if (!responsibleStory) {
+        responsibleStory = await findResponsibleStory(testFile, workdir, passedStories);
+      }
       if (responsibleStory) {
         affectedStories.add(responsibleStory.id);
         affectedStoriesObjs.set(responsibleStory.id, responsibleStory);
@@ -61808,7 +61874,12 @@ async function handleRunCompletion(options) {
       config: config2,
       prd,
       workdir,
-      runtime: options.runtime
+      runtime: options.runtime,
+      storyMetrics: options.isSequential === false ? undefined : allStoryMetrics.map((m) => ({
+        storyId: m.storyId,
+        completedAt: m.completedAt,
+        failingTestFiles: m.failingTestFiles
+      }))
     });
     const lastRunAt = new Date().toISOString();
     logger?.info("regression", "Deferred regression gate completed", {
@@ -63341,13 +63412,17 @@ function buildEscalationFailure(story, currentTier, reviewFindings, cost, pipeli
     summary,
     reviewFindings: reviewFindings && reviewFindings.length > 0 ? reviewFindings : undefined,
     cost: cost ?? 0,
-    timestamp: new Date().toISOString()
+    timestamp: new Date().toISOString(),
+    ...story.routing?.agent !== undefined ? { agent: story.routing.agent } : {},
+    ...story.routing?.agentProfileId !== undefined ? { agentProfileId: story.routing.agentProfileId } : {}
   };
 }
-function buildEscalationRecord(currentTier, nextTier, reason) {
+function buildEscalationRecord(currentTier, nextTier, reason, agents) {
   return {
     fromTier: currentTier,
     toTier: nextTier,
+    ...agents?.fromAgent !== undefined ? { fromAgent: agents.fromAgent } : {},
+    ...agents?.toAgent !== undefined ? { toAgent: agents.toAgent } : {},
     reason,
     timestamp: new Date().toISOString()
   };
@@ -63444,7 +63519,7 @@ async function handleTierEscalation(ctx) {
       const currentStoryTier = s.routing?.modelTier ?? ctx.routing.modelTier;
       const isChangingTier = currentStoryTier !== escalatedTier;
       const shouldResetAttempts = isChangingTier || shouldSwitchToTestAfter;
-      const escalationRecord = isChangingTier || shouldSwitchToTestAfter ? buildEscalationRecord(currentStoryTier, shouldSwitchToTestAfter ? currentStoryTier : escalatedTier, ctx.pipelineResult.reason ?? "Escalated to next retry path") : undefined;
+      const escalationRecord = isChangingTier || shouldSwitchToTestAfter ? buildEscalationRecord(currentStoryTier, shouldSwitchToTestAfter ? currentStoryTier : escalatedTier, ctx.pipelineResult.reason ?? "Escalated to next retry path", { fromAgent: s.routing?.agent, toAgent: nextAgent }) : undefined;
       const escalationFailure = buildEscalationFailure(s, currentStoryTier, escalateReviewFindings, ctx.attemptCost, verifiedPipelineReason, escalateFailureCategory);
       return {
         ...s,
@@ -65663,7 +65738,8 @@ async function initializeRun(ctx) {
   prd = await reconcileState(prd, ctx.prdPath, ctx.workdir, ctx.config);
   const resetRef = ctx.config.review?.semantic?.resetRefOnRerun ?? false;
   const storyIsolation = ctx.config.execution.storyIsolation;
-  const resetStories = resetFailedStoriesToPending(prd, resetRef, storyIsolation);
+  const resetMode = ctx.config.autoMode.escalation.resetMode;
+  const resetStories = resetFailedStoriesToPending(prd, { resetRef, storyIsolation, resetMode });
   if (resetStories.length > 0) {
     const resetIds = resetStories.map((s) => s.id);
     logger?.info("run-initialization", "Reset failed stories to pending for re-run", { storyIds: resetIds });
@@ -95894,7 +95970,8 @@ function finalizePrdRouting(prd, agentRouting, profileName) {
       agentProfileId: assignment.agentProfileId,
       profileModelTier: assignment.profileModelTier,
       initialAgent: story.routing?.initialAgent ?? assignment.agent,
-      initialProfileId: story.routing?.initialProfileId ?? assignment.agentProfileId
+      initialProfileId: story.routing?.initialProfileId ?? assignment.agentProfileId,
+      initialModelTier: story.routing?.initialModelTier ?? assignment.profileModelTier
     };
     return { ...story, routing };
   });

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@nathapp/nax",
-  "version": "0.70.0-canary.5",
+  "version": "0.70.0-canary.7",
   "description": "AI Coding Agent Orchestrator — loops until done",
   "type": "module",
   "bin": {