@nathapp/nax 0.60.0-canary.1 → 0.60.0

package/dist/nax.js

@@ -18334,7 +18334,11 @@ var init_schemas3 = __esm(() => {
       fixModel: "balanced",
       strategy: "diagnose-first",
       maxRetries: 2
-    })
+    }),
+    suggestedTestPath: exports_external.string().min(1).optional(),
+    hardening: exports_external.object({
+      enabled: exports_external.boolean().default(true)
+    }).optional().default({ enabled: true })
   });
   TestCoverageConfigSchema = exports_external.object({
     enabled: exports_external.boolean().default(true),
@@ -18663,7 +18667,8 @@ var init_schemas3 = __esm(() => {
         fixModel: "balanced",
         strategy: "diagnose-first",
         maxRetries: 2
-      }
+      },
+      hardening: { enabled: true }
     }),
     context: ContextConfigSchema.default({
       fileInjection: "disabled",
@@ -21929,6 +21934,130 @@ ${c.output}`).join(`
   buildDebaterLabel(debater) {
     return debater.persona ? `${debater.agent} (${debater.persona})` : debater.agent;
   }
+  buildReviewPrompt(diff, story) {
+    const criteria = story.acceptanceCriteria.map((c) => `- ${c}`).join(`
+`);
+    return [
+      `Review the following code diff for story ${story.id}: ${story.title}`,
+      "",
+      "## Acceptance Criteria",
+      criteria,
+      "",
+      "## Diff",
+      diff,
+      "",
+      "Also flag any changes in the diff not required by the acceptance criteria above as out-of-scope findings.",
+      "Respond with JSON: { passed: boolean, findings: [...], findingReasoning: { [id]: string } }"
+    ].join(`
+`);
+  }
+  buildReReviewPrompt(updatedDiff, previousFindings) {
+    const findingsList = previousFindings.length > 0 ? previousFindings.map((f) => `- ${f.ruleId}: ${f.message}`).join(`
+`) : "(none)";
+    return [
+      "This is a follow-up re-review. Please review the updated diff below.",
+      "",
+      "## Previous Findings",
+      findingsList,
+      "",
+      "## Updated Diff",
+      updatedDiff,
+      "",
+      "Respond with JSON: { passed: boolean, findings: [...], findingReasoning: { [id]: string }, deltaSummary: string }",
+      "deltaSummary should describe which previous findings are resolved vs still present."
+    ].join(`
+`);
+  }
+  buildResolverPrompt(proposals, critiques, diff, story, resolverContext) {
+    const criteria = story.acceptanceCriteria.map((c) => `- ${c}`).join(`
+`);
+    const framing = this.buildResolverFraming(resolverContext);
+    const voteTally = this.buildVoteTallyLine(resolverContext);
+    const proposalsSection = this.buildLabeledProposalsSection(proposals);
+    const critiquesSection = this.buildLabeledCritiquesSection(critiques);
+    return [
+      framing,
+      "",
+      `## Story ${story.id}: ${story.title}`,
+      "",
+      "## Acceptance Criteria",
+      criteria,
+      "",
+      "## Debater Proposals",
+      proposalsSection,
+      critiquesSection,
+      "",
+      "## Diff",
+      diff,
+      voteTally,
+      "",
+      "Respond with JSON: { passed: boolean, findings: [...], findingReasoning: { [id]: string } }"
+    ].filter((line) => line !== undefined).join(`
+`);
+  }
+  buildReResolverPrompt(proposals, critiques, updatedDiff, previousFindings, resolverContext) {
+    const framing = this.buildResolverFraming(resolverContext);
+    const findingsList = previousFindings.length > 0 ? previousFindings.map((f) => `- ${f.ruleId}: ${f.message}`).join(`
+`) : "(none)";
+    const proposalsSection = this.buildLabeledProposalsSection(proposals);
+    const critiquesSection = this.buildLabeledCritiquesSection(critiques);
+    return [
+      `${framing} This is a re-review after implementer changes.`,
+      "",
+      "## Previous Findings",
+      findingsList,
+      "",
+      "## Updated Debater Proposals",
+      proposalsSection,
+      critiquesSection,
+      "",
+      "## Updated Diff",
+      updatedDiff,
+      "",
+      "Respond with JSON: { passed: boolean, findings: [...], findingReasoning: { [id]: string }, deltaSummary: string }",
+      "deltaSummary should describe which previous findings are resolved vs still present."
+    ].filter((line) => line !== undefined).join(`
+`);
+  }
+  buildResolverFraming(ctx) {
+    switch (ctx.resolverType) {
+      case "majority-fail-closed":
+      case "majority-fail-open":
+        return "You are the authoritative reviewer resolving a debate. A preliminary vote was taken \u2014 see tally below. Verify disputed findings using tools (READ files, GREP for usage) and give your final verdict.";
+      case "synthesis":
+        return "You are a synthesis reviewer. Synthesize the debater proposals into a single, coherent, tool-verified verdict. Use READ and GREP to verify claims before ruling.";
+      case "custom":
+        return "You are the judge. Evaluate the debater proposals independently. Verify claims with tools (READ, GREP) and give your final authoritative verdict.";
+      default:
+        return "You are the reviewer. Evaluate the debater proposals and give your final authoritative verdict.";
+    }
+  }
+  buildVoteTallyLine(ctx) {
+    if (!ctx.majorityVote)
+      return "";
+    const { passCount, failCount } = ctx.majorityVote;
+    const failOpenNote = ctx.resolverType === "majority-fail-open" ? " (unparseable proposals count as pass)" : " (unparseable proposals count as fail)";
+    return `
+
+The preliminary majority vote is: **${passCount} passed, ${failCount} failed**${failOpenNote}. Verify the failing findings with tools before giving your authoritative verdict.`;
+  }
+  buildLabeledProposalsSection(proposals) {
+    return proposals.map((p) => `### ${p.debater}
+${p.output}`).join(`
+
+`);
+  }
+  buildLabeledCritiquesSection(critiques) {
+    if (critiques.length === 0)
+      return "";
+    return `
+
+## Critiques
+${critiques.map((c, i) => `### Critique ${i + 1}
+${c}`).join(`
+
+`)}`;
+  }
 }
 var init_prompt_builder = __esm(() => {
   init_personas();
@@ -22566,7 +22695,20 @@ Do NOT output the JSON to the conversation. Write the file, then reply with a br
     logger?.warn("debate", "hybrid mode requires sessionMode: stateful for plan \u2014 running as panel");
   }
   const resolverTimeoutMs = (ctx.stageConfig.timeoutSeconds ?? 600) * 1000;
-  const planSynthesisSuffix = "IMPORTANT: Your response must be a single valid JSON object in PRD format (with project, feature, branchName, userStories array, etc.). Do NOT wrap it in markdown fences. Output raw JSON only.";
+  const specAnchor = opts.specContent ? `
+
+## Original Spec
+
+${opts.specContent}
+
+## Synthesis Rules \u2014 Acceptance Criteria
+
+The spec above is the authoritative source for acceptance criteria.
+- Each story's \`acceptanceCriteria\` array MUST contain only criteria that are explicitly stated or directly implied by the spec.
+- If a debater proposed criteria beyond the spec (edge cases, error handling, implementation details), place those in a separate \`suggestedCriteria\` array on the same story object.
+- Never silently merge debater-invented criteria into \`acceptanceCriteria\`. The distinction matters: \`acceptanceCriteria\` drives automated testing; \`suggestedCriteria\` is logged for human review.
+- Preserve the spec's AC wording. You may refine for clarity but must not change semantics.` : "";
+  const planSynthesisSuffix = `IMPORTANT: Your response must be a single valid JSON object in PRD format (with project, feature, branchName, userStories array, etc.). Do NOT wrap it in markdown fences. Output raw JSON only.${specAnchor}`;
   const outcome = await resolveOutcome(proposalOutputs, critiqueOutputs, ctx.stageConfig, ctx.config, ctx.storyId, resolverTimeoutMs, opts.workdir, opts.feature, undefined, undefined, planSynthesisSuffix, successful.map((p) => p.debater));
   const winningOutput = outcome.output ?? successful[0].output;
   const proposals = successful.map((p) => ({ debater: p.debater, output: p.output }));
@@ -24929,6 +25071,24 @@ function resolveAcceptanceTestCandidates(options) {
     return [];
   return [resolveAcceptanceFeatureTestPath(options.featureDir, options.testPathConfig, options.language)];
 }
+function suggestedTestFilename(language) {
+  switch (language?.toLowerCase()) {
+    case "go":
+      return ".nax-suggested_test.go";
+    case "python":
+      return ".nax-suggested.test.py";
+    case "rust":
+      return ".nax-suggested.rs";
+    default:
+      return ".nax-suggested.test.ts";
+  }
+}
+function resolveSuggestedTestFile(language, testPathConfig) {
+  return testPathConfig ?? suggestedTestFilename(language);
+}
+function resolveSuggestedPackageFeatureTestPath(packageDir, featureName, testPathConfig, language) {
+  return path.join(packageDir, ".nax", "features", featureName, resolveSuggestedTestFile(language, testPathConfig));
+}
 async function findExistingAcceptanceTestPath(options) {
   const candidates = resolveAcceptanceTestCandidates(options);
   for (const testPath of candidates) {
@@ -25795,7 +25955,7 @@ Rules:
 - Every test MUST have real assertions that PASS when the feature is correctly implemented and FAIL when it is broken
 - **Prefer behavioral tests** \u2014 import functions and call them rather than reading source files. For example, to verify "getPostRunActions() returns empty array", import PluginRegistry and call getPostRunActions(), don't grep the source file for the method name.
 - **File output (REQUIRED)**: Write the acceptance test file DIRECTLY to the path shown below. Do NOT output the test code in your response. After writing the file, reply with a brief confirmation.
-- **Path anchor (CRITICAL)**: Write the test file to this exact path: \`${join16(options.workdir, ".nax", "features", options.featureName, resolveAcceptanceTestFile2(options.language, options.config?.acceptance?.testPath))}\`. Import from package sources using relative paths like \`../../../src/...\` (3 levels up from \`.nax/features/<name>/\` to the package root).
+- **Path anchor (CRITICAL)**: Write the test file to this exact path: \`${options.targetTestFile ?? join16(options.workdir, ".nax", "features", options.featureName, resolveAcceptanceTestFile2(options.language, options.config?.acceptance?.testPath))}\`. Import from package sources using relative paths like \`../../../src/...\` (3 levels up from \`.nax/features/<name>/\` to the package root).
 - **Process cwd**: When spawning child processes to invoke a CLI or binary, set the working directory to the **package root** (\`join(import.meta.dir, "../../..")\`) as your default \u2014 unless your Step 2 exploration reveals the CLI uses a different working directory convention (e.g. reads config from \`~/.config/\`, or resolves paths relative to a flag value). Always check how the CLI resolves file paths before assuming.`;
   const implementationSection = options.implementationContext && options.implementationContext.length > 0 ? `
 
@@ -25828,7 +25988,7 @@ Previous test failed because: ${options.previousFailure}` : "";
     outputPreview: rawOutput.slice(0, 300)
   });
   if (!testCode) {
-    const targetPath = join16(options.workdir, ".nax", "features", options.featureName, resolveAcceptanceTestFile2(options.language, options.config?.acceptance?.testPath));
+    const targetPath = options.targetTestFile ?? join16(options.workdir, ".nax", "features", options.featureName, resolveAcceptanceTestFile2(options.language, options.config?.acceptance?.testPath));
     const backupPath = `${targetPath}.llm-recovery.bak`;
     let recoveryFailed = false;
     logger.debug("acceptance", "BUG-076 recovery: checking for agent-written file", {
@@ -26193,156 +26353,6 @@ function logTestOutput(logger, stage, output, opts = {}) {
   });
 }
 
-// src/pipeline/stages/acceptance.ts
-var exports_acceptance = {};
-__export(exports_acceptance, {
-  acceptanceStage: () => acceptanceStage
-});
-function parseTestFailures(output) {
-  const failedACs = [];
-  const lines = output.split(`
-`);
-  for (const line of lines) {
-    if (line.includes("(fail)")) {
-      const acMatch = line.match(/(AC-\d+):/i);
-      if (acMatch) {
-        const acId = acMatch[1].toUpperCase();
-        if (!failedACs.includes(acId)) {
-          failedACs.push(acId);
-        }
-      }
-    }
-  }
-  return failedACs;
-}
-function areAllStoriesComplete(ctx) {
-  const counts = countStories(ctx.prd);
-  const totalComplete = counts.passed + counts.failed + counts.skipped;
-  return totalComplete === counts.total;
-}
-var acceptanceStage;
-var init_acceptance = __esm(() => {
-  init_generator();
-  init_test_path();
-  init_logger2();
-  init_prd();
-  acceptanceStage = {
-    name: "acceptance",
-    enabled(ctx) {
-      if (!ctx.config.acceptance.enabled) {
-        return false;
-      }
-      if (!areAllStoriesComplete(ctx)) {
-        return false;
-      }
-      return true;
-    },
-    async execute(ctx) {
-      const logger = getLogger();
-      logger.info("acceptance", "Running acceptance tests", { storyId: ctx.story.id });
-      if (!ctx.featureDir) {
-        logger.warn("acceptance", "No feature directory \u2014 skipping acceptance tests", { storyId: ctx.story.id });
-        return { action: "continue" };
-      }
-      const testGroups = ctx.acceptanceTestPaths ?? [
-        {
-          testPath: resolveAcceptanceFeatureTestPath(ctx.featureDir, ctx.config.acceptance.testPath, ctx.config.project?.language),
-          packageDir: ctx.workdir
-        }
-      ];
-      const allFailedACs = [];
-      const allOutputParts = [];
-      let anyError = false;
-      let errorExitCode = 0;
-      for (const { testPath, packageDir } of testGroups) {
-        const testFile = Bun.file(testPath);
-        const exists = await testFile.exists();
-        if (!exists) {
-          logger.warn("acceptance", "Acceptance test file not found \u2014 skipping", { storyId: ctx.story.id, testPath });
-          continue;
-        }
-        const testCmdParts = buildAcceptanceRunCommand(testPath, ctx.config.project?.testFramework, ctx.config.acceptance.command);
-        logger.info("acceptance", "Running acceptance command", {
-          storyId: ctx.story.id,
-          cmd: testCmdParts.join(" "),
-          packageDir
-        });
-        const proc = Bun.spawn(testCmdParts, {
-          cwd: packageDir,
-          stdout: "pipe",
-          stderr: "pipe"
-        });
-        const [exitCode, stdout, stderr] = await Promise.all([
-          proc.exited,
-          new Response(proc.stdout).text(),
-          new Response(proc.stderr).text()
-        ]);
-        const output = `${stdout}
-${stderr}`;
-        allOutputParts.push(output);
-        const failedACs = parseTestFailures(output);
-        const overrides = ctx.prd.acceptanceOverrides ?? {};
-        const actualFailures = failedACs.filter((acId) => !overrides[acId]);
-        const overriddenFailures = failedACs.filter((acId) => overrides[acId]);
-        if (overriddenFailures.length > 0) {
-          logger.warn("acceptance", "Skipped failures (overridden)", {
-            storyId: ctx.story.id,
-            overriddenFailures,
-            overrides: overriddenFailures.map((acId) => ({ acId, reason: overrides[acId] }))
-          });
-        }
-        if (failedACs.length === 0 && exitCode !== 0) {
-          logger.error("acceptance", "Tests errored with no AC failures parsed", {
-            storyId: ctx.story.id,
-            exitCode,
-            packageDir
-          });
-          logTestOutput(logger, "acceptance", output);
-          anyError = true;
-          errorExitCode = exitCode;
-          allFailedACs.push("AC-ERROR");
-          continue;
-        }
-        for (const acId of actualFailures) {
-          if (!allFailedACs.includes(acId)) {
-            allFailedACs.push(acId);
-          }
-        }
-        if (actualFailures.length > 0) {
-          logger.error("acceptance", "Acceptance tests failed", {
-            storyId: ctx.story.id,
-            failedACs: actualFailures,
-            packageDir
-          });
-          logTestOutput(logger, "acceptance", output);
-        } else if (exitCode === 0) {
-          logger.info("acceptance", "Package acceptance tests passed", { storyId: ctx.story.id, packageDir });
-        }
-      }
-      const combinedOutput = allOutputParts.join(`
-`);
-      if (allFailedACs.length === 0) {
-        logger.info("acceptance", "All acceptance tests passed", { storyId: ctx.story.id });
-        return { action: "continue" };
-      }
-      ctx.acceptanceFailures = {
-        failedACs: allFailedACs,
-        testOutput: combinedOutput
-      };
-      if (anyError) {
-        return {
-          action: "fail",
-          reason: `Acceptance tests errored (exit code ${errorExitCode}): syntax error, import failure, or unhandled exception`
-        };
-      }
-      return {
-        action: "fail",
-        reason: `Acceptance tests failed: ${allFailedACs.join(", ")}`
-      };
-    }
-  };
-});
-
 // src/acceptance/refinement.ts
 var exports_refinement = {};
 __export(exports_refinement, {
@@ -26509,6 +26519,315 @@ var init_refinement = __esm(() => {
   };
 });
 
+// src/acceptance/hardening.ts
+var exports_hardening = {};
+__export(exports_hardening, {
+  runHardeningPass: () => runHardeningPass,
+  _hardeningDeps: () => _hardeningDeps
+});
+async function runHardeningPass(ctx) {
+  const logger = getSafeLogger();
+  const result = { promoted: [], discarded: [], costUsd: 0 };
+  const storiesWithSuggested = ctx.prd.userStories.filter((s) => s.suggestedCriteria && s.suggestedCriteria.length > 0);
+  if (storiesWithSuggested.length === 0)
+    return result;
+  logger?.info("acceptance", "Starting hardening pass", {
+    storyId: storiesWithSuggested[0].id,
+    storiesWithSuggested: storiesWithSuggested.length,
+    totalSuggestedACs: storiesWithSuggested.reduce((n, s) => n + (s.suggestedCriteria?.length ?? 0), 0)
+  });
+  try {
+    const allRefined = [];
+    for (const story of storiesWithSuggested) {
+      const criteria = story.suggestedCriteria ?? [];
+      const refined = await _hardeningDeps.refine(criteria, {
+        storyId: story.id,
+        featureName: ctx.prd.feature,
+        workdir: ctx.workdir,
+        codebaseContext: "",
+        config: ctx.config
+      });
+      allRefined.push(...refined);
+    }
+    const language = ctx.config.project?.language;
+    const suggestedTestPath = resolveSuggestedPackageFeatureTestPath(ctx.workdir, ctx.prd.feature, ctx.config.acceptance?.suggestedTestPath, language);
+    let modelDef;
+    try {
+      modelDef = resolveModelForAgent(ctx.config.models, ctx.config.autoMode?.defaultAgent ?? "claude", ctx.config.acceptance?.model ?? "fast", ctx.config.autoMode?.defaultAgent ?? "claude");
+    } catch {
+      modelDef = { provider: "anthropic", model: "claude-haiku-4-5-20251001" };
+    }
+    const genResult = await _hardeningDeps.generate(storiesWithSuggested, allRefined, {
+      featureName: ctx.prd.feature,
+      workdir: ctx.workdir,
+      featureDir: ctx.featureDir,
+      codebaseContext: "",
+      modelTier: ctx.config.acceptance?.model ?? "fast",
+      modelDef,
+      config: ctx.config,
+      language,
+      targetTestFile: suggestedTestPath
+    });
+    if (genResult.testCode) {
+      await _hardeningDeps.writeFile(suggestedTestPath, genResult.testCode);
+    }
+    const testCmd = buildAcceptanceRunCommand(suggestedTestPath, ctx.config.project?.testFramework, ctx.config.acceptance?.command);
+    const proc = _hardeningDeps.spawn(testCmd, {
+      cwd: ctx.workdir,
+      stdout: "pipe",
+      stderr: "pipe"
+    });
+    const [exitCode, stdout, stderr] = await Promise.all([
+      proc.exited,
+      new Response(proc.stdout).text(),
+      new Response(proc.stderr).text()
+    ]);
+    const output = `${stdout}
+${stderr}`;
+    const failedACs = parseTestFailures(output);
+    const failedSet = new Set(failedACs.map((ac) => ac.toUpperCase()));
+    let acIndex = 0;
+    for (const story of storiesWithSuggested) {
+      const suggested = story.suggestedCriteria ?? [];
+      const toPromote = [];
+      const toDiscard = [];
+      for (const criterion of suggested) {
+        acIndex++;
+        const acId = `AC-${acIndex}`;
+        if (failedSet.has(acId) || exitCode !== 0 && failedACs.length === 0) {
+          toDiscard.push(criterion);
+        } else {
+          toPromote.push(criterion);
+        }
+      }
+      if (toPromote.length > 0) {
+        story.acceptanceCriteria = [...story.acceptanceCriteria, ...toPromote];
+        result.promoted.push(...toPromote);
+      }
+      result.discarded.push(...toDiscard);
+      story.suggestedCriteria = toDiscard.length > 0 ? toDiscard : undefined;
+    }
+    if (result.promoted.length > 0) {
+      await _hardeningDeps.savePRD(ctx.prd, ctx.prdPath);
+    }
+    logger?.info("acceptance", "Hardening pass complete", {
+      storyId: storiesWithSuggested[0].id,
+      promoted: result.promoted.length,
+      discarded: result.discarded.length,
+      costUsd: result.costUsd
+    });
+  } catch (err) {
+    logger?.warn("acceptance", "Hardening pass failed (non-blocking)", {
+      storyId: storiesWithSuggested[0].id,
+      error: err instanceof Error ? err.message : String(err)
+    });
+  }
+  return result;
+}
+var _hardeningDeps;
+var init_hardening = __esm(() => {
+  init_config();
+  init_logger2();
+  init_acceptance();
+  init_prd();
+  init_generator();
+  init_generator();
+  init_refinement();
+  init_test_path();
+  _hardeningDeps = {
+    refine: refineAcceptanceCriteria,
+    generate: generateFromPRD,
+    savePRD,
+    spawn: Bun.spawn,
+    writeFile: async (p, c) => {
+      await Bun.write(p, c);
+    }
+  };
+});
+
+// src/pipeline/stages/acceptance.ts
+var exports_acceptance = {};
+__export(exports_acceptance, {
+  parseTestFailures: () => parseTestFailures,
+  acceptanceStage: () => acceptanceStage,
+  _acceptanceStageDeps: () => _acceptanceStageDeps
+});
+function parseTestFailures(output) {
+  const failedACs = [];
+  const lines = output.split(`
+`);
+  for (const line of lines) {
+    if (line.includes("(fail)")) {
+      const acMatch = line.match(/(AC-\d+):/i);
+      if (acMatch) {
+        const acId = acMatch[1].toUpperCase();
+        if (!failedACs.includes(acId)) {
+          failedACs.push(acId);
+        }
+      }
+    }
+  }
+  return failedACs;
+}
+function areAllStoriesComplete(ctx) {
+  const counts = countStories(ctx.prd);
+  const totalComplete = counts.passed + counts.failed + counts.skipped;
+  return totalComplete === counts.total;
+}
+var _acceptanceStageDeps, acceptanceStage;
+var init_acceptance = __esm(() => {
+  init_generator();
+  init_test_path();
+  init_logger2();
+  init_prd();
+  _acceptanceStageDeps = {
+    runHardeningPass: async (ctx) => {
+      const { runHardeningPass: runHardeningPass2 } = await Promise.resolve().then(() => (init_hardening(), exports_hardening));
+      return runHardeningPass2(ctx);
+    }
+  };
+  acceptanceStage = {
+    name: "acceptance",
+    enabled(ctx) {
+      if (!ctx.config.acceptance.enabled) {
+        return false;
+      }
+      if (!areAllStoriesComplete(ctx)) {
+        return false;
+      }
+      return true;
+    },
+    async execute(ctx) {
+      const logger = getLogger();
+      logger.info("acceptance", "Running acceptance tests", { storyId: ctx.story.id });
+      if (!ctx.featureDir) {
+        logger.warn("acceptance", "No feature directory \u2014 skipping acceptance tests", { storyId: ctx.story.id });
+        return { action: "continue" };
+      }
+      const testGroups = ctx.acceptanceTestPaths ?? [
+        {
+          testPath: resolveAcceptanceFeatureTestPath(ctx.featureDir, ctx.config.acceptance.testPath, ctx.config.project?.language),
+          packageDir: ctx.workdir
+        }
+      ];
+      const allFailedACs = [];
+      const allOutputParts = [];
+      let anyError = false;
+      let errorExitCode = 0;
+      for (const { testPath, packageDir } of testGroups) {
+        const testFile = Bun.file(testPath);
+        const exists = await testFile.exists();
+        if (!exists) {
+          logger.warn("acceptance", "Acceptance test file not found \u2014 skipping", { storyId: ctx.story.id, testPath });
+          continue;
+        }
+        const testCmdParts = buildAcceptanceRunCommand(testPath, ctx.config.project?.testFramework, ctx.config.acceptance.command);
+        logger.info("acceptance", "Running acceptance command", {
+          storyId: ctx.story.id,
+          cmd: testCmdParts.join(" "),
+          packageDir
+        });
+        const proc = Bun.spawn(testCmdParts, {
+          cwd: packageDir,
+          stdout: "pipe",
+          stderr: "pipe"
+        });
+        const [exitCode, stdout, stderr] = await Promise.all([
+          proc.exited,
+          new Response(proc.stdout).text(),
+          new Response(proc.stderr).text()
+        ]);
+        const output = `${stdout}
+${stderr}`;
+        allOutputParts.push(output);
+        const failedACs = parseTestFailures(output);
+        const overrides = ctx.prd.acceptanceOverrides ?? {};
+        const actualFailures = failedACs.filter((acId) => !overrides[acId]);
+        const overriddenFailures = failedACs.filter((acId) => overrides[acId]);
+        if (overriddenFailures.length > 0) {
+          logger.warn("acceptance", "Skipped failures (overridden)", {
+            storyId: ctx.story.id,
+            overriddenFailures,
+            overrides: overriddenFailures.map((acId) => ({ acId, reason: overrides[acId] }))
+          });
+        }
+        if (failedACs.length === 0 && exitCode !== 0) {
+          logger.error("acceptance", "Tests errored with no AC failures parsed", {
+            storyId: ctx.story.id,
+            exitCode,
+            packageDir
+          });
+          logTestOutput(logger, "acceptance", output);
+          anyError = true;
+          errorExitCode = exitCode;
+          allFailedACs.push("AC-ERROR");
+          continue;
+        }
+        for (const acId of actualFailures) {
+          if (!allFailedACs.includes(acId)) {
+            allFailedACs.push(acId);
+          }
+        }
+        if (actualFailures.length > 0) {
+          logger.error("acceptance", "Acceptance tests failed", {
+            storyId: ctx.story.id,
+            failedACs: actualFailures,
+            packageDir
+          });
+          logTestOutput(logger, "acceptance", output);
+        } else if (exitCode === 0) {
+          logger.info("acceptance", "Package acceptance tests passed", { storyId: ctx.story.id, packageDir });
+        }
+      }
+      const combinedOutput = allOutputParts.join(`
+`);
+      if (allFailedACs.length === 0) {
+        logger.info("acceptance", "All acceptance tests passed", { storyId: ctx.story.id });
+        const hardeningEnabled = ctx.config.acceptance?.hardening?.enabled !== false;
+        const hasAnySuggested = ctx.prd.userStories.some((s) => s.suggestedCriteria && s.suggestedCriteria.length > 0);
+        if (hardeningEnabled && hasAnySuggested && ctx.featureDir) {
+          try {
+            const prdPath = ctx.prdPath ?? `${ctx.featureDir}/prd.json`;
+            const result = await _acceptanceStageDeps.runHardeningPass({
+              prd: ctx.prd,
+              prdPath,
+              featureDir: ctx.featureDir,
+              workdir: ctx.workdir,
+              config: ctx.config,
+              agentGetFn: ctx.agentGetFn
+            });
+            logger.info("acceptance", "Hardening pass complete", {
+              storyId: ctx.story.id,
+              promoted: result.promoted.length,
+              discarded: result.discarded.length
+            });
+          } catch (err) {
+            logger.warn("acceptance", "Hardening pass failed (non-blocking)", {
+              storyId: ctx.story.id,
+              error: err instanceof Error ? err.message : String(err)
+            });
+          }
+        }
+        return { action: "continue" };
+      }
+      ctx.acceptanceFailures = {
+        failedACs: allFailedACs,
+        testOutput: combinedOutput
+      };
+      if (anyError) {
+        return {
+          action: "fail",
+          reason: `Acceptance tests errored (exit code ${errorExitCode}): syntax error, import failure, or unhandled exception`
+        };
+      }
+      return {
+        action: "fail",
+        reason: `Acceptance tests failed: ${allFailedACs.join(", ")}`
+      };
+    }
+  };
+});
+
 // src/pipeline/stages/acceptance-setup.ts
 var exports_acceptance_setup = {};
 __export(exports_acceptance_setup, {
@@ -27132,132 +27451,6 @@ var init_agents = __esm(() => {
   init_errors();
 });
 
-// src/review/dialogue-prompts.ts
-function buildReviewPrompt(diff, story, _semanticConfig) {
-  const criteria = story.acceptanceCriteria.map((c) => `- ${c}`).join(`
-`);
-  return [
-    `Review the following code diff for story ${story.id}: ${story.title}`,
-    "",
-    "## Acceptance Criteria",
-    criteria,
-    "",
-    "## Diff",
-    diff,
-    "",
-    "Also flag any changes in the diff not required by the acceptance criteria above as out-of-scope findings.",
-    "Respond with JSON: { passed: boolean, findings: [...], findingReasoning: { [id]: string } }"
-  ].join(`
-`);
-}
-function buildReReviewPrompt(updatedDiff, previousFindings) {
-  const findingsList = previousFindings.length > 0 ? previousFindings.map((f) => `- ${f.ruleId}: ${f.message}`).join(`
-`) : "(none)";
-  return [
-    "This is a follow-up re-review. Please review the updated diff below.",
-    "",
-    "## Previous Findings",
-    findingsList,
-    "",
-    "## Updated Diff",
-    updatedDiff,
-    "",
-    "Respond with JSON: { passed: boolean, findings: [...], findingReasoning: { [id]: string }, deltaSummary: string }",
-    "deltaSummary should describe which previous findings are resolved vs still present."
-  ].join(`
-`);
-}
-function buildProposalsSection2(proposals) {
-  return proposals.map((p) => `### ${p.debater}
-${p.output}`).join(`
-
-`);
-}
-function buildCritiquesSection(critiques) {
-  if (critiques.length === 0)
-    return "";
-  return `
-
-## Critiques
-${critiques.map((c, i) => `### Critique ${i + 1}
-${c}`).join(`
-
-`)}`;
-}
-function buildVoteTallyLine(ctx) {
-  if (!ctx.majorityVote)
-    return "";
-  const { passCount, failCount } = ctx.majorityVote;
-  const failOpenNote = ctx.resolverType === "majority-fail-open" ? " (unparseable proposals count as pass)" : " (unparseable proposals count as fail)";
-  return `
-
-The preliminary majority vote is: **${passCount} passed, ${failCount} failed**${failOpenNote}. Verify the failing findings with tools before giving your authoritative verdict.`;
-}
-function buildResolverFraming(ctx) {
-  switch (ctx.resolverType) {
-    case "majority-fail-closed":
-    case "majority-fail-open":
-      return "You are the authoritative reviewer resolving a debate. A preliminary vote was taken \u2014 see tally below. Verify disputed findings using tools (READ files, GREP for usage) and give your final verdict.";
-    case "synthesis":
-      return "You are a synthesis reviewer. Synthesize the debater proposals into a single, coherent, tool-verified verdict. Use READ and GREP to verify claims before ruling.";
-    case "custom":
-      return "You are the judge. Evaluate the debater proposals independently. Verify claims with tools (READ, GREP) and give your final authoritative verdict.";
-    default:
-      return "You are the reviewer. Evaluate the debater proposals and give your final authoritative verdict.";
-  }
-}
-function buildDebateResolverPrompt(proposals, critiques, diff, story, _semanticConfig, resolverContext) {
-  const criteria = story.acceptanceCriteria.map((c) => `- ${c}`).join(`
-`);
-  const framing = buildResolverFraming(resolverContext);
-  const voteTally = buildVoteTallyLine(resolverContext);
-  const proposalsSection = buildProposalsSection2(proposals);
-  const critiquesSection = buildCritiquesSection(critiques);
-  return [
-    framing,
-    "",
-    `## Story ${story.id}: ${story.title}`,
-    "",
-    "## Acceptance Criteria",
-    criteria,
-    "",
-    "## Debater Proposals",
-    proposalsSection,
-    critiquesSection,
-    "",
-    "## Diff",
-    diff,
-    voteTally,
-    "",
-    "Respond with JSON: { passed: boolean, findings: [...], findingReasoning: { [id]: string } }"
-  ].filter((line) => line !== undefined).join(`
-`);
-}
-function buildDebateReReviewPrompt(proposals, critiques, updatedDiff, previousFindings, resolverContext) {
-  const framing = buildResolverFraming(resolverContext);
-  const findingsList = previousFindings.length > 0 ? previousFindings.map((f) => `- ${f.ruleId}: ${f.message}`).join(`
-`) : "(none)";
-  const proposalsSection = buildProposalsSection2(proposals);
-  const critiquesSection = buildCritiquesSection(critiques);
-  return [
-    `${framing} This is a re-review after implementer changes.`,
-    "",
-    "## Previous Findings",
-    findingsList,
-    "",
-    "## Updated Debater Proposals",
-    proposalsSection,
-    critiquesSection,
-    "",
-    "## Updated Diff",
-    updatedDiff,
-    "",
-    "Respond with JSON: { passed: boolean, findings: [...], findingReasoning: { [id]: string }, deltaSummary: string }",
-    "deltaSummary should describe which previous findings are resolved vs still present."
-  ].filter((line) => line !== undefined).join(`
-`);
-}
-
 // src/review/dialogue.ts
 function extractDeltaSummary(rawOutput, previousFindings, newFindings) {
   const parsed = tryParseLLMJson(rawOutput);
@@ -27333,6 +27526,7 @@ function createReviewerSession(agent, storyId, workdir, featureName, _config) {
     generation: 1,
     pendingCompactionContext: null
   };
+  const promptBuilder = new DebatePromptBuilder({ taskContext: "", outputFormat: "", stage: "review" }, { debaters: [], sessionMode: "stateful" });
   function resolveRunParams(semanticConfig) {
     const modelTier = semanticConfig.modelTier;
     const defaultAgent = _config.autoMode?.defaultAgent ?? "claude";
@@ -27367,7 +27561,7 @@ ${prompt}`,
       if (!active) {
         throw new NaxError(`[dialogue] ReviewerSession for story ${storyId} has been destroyed`, "REVIEWER_SESSION_DESTROYED", { stage: "review", storyId, featureName });
       }
-      const prompt = buildReviewPrompt(diff, story, semanticConfig);
+      const prompt = promptBuilder.buildReviewPrompt(diff, story);
       const { modelTier, modelDef, timeoutSeconds } = resolveRunParams(semanticConfig);
       const { effectivePrompt, acpSessionName } = buildEffectiveRunArgs(prompt);
       const result = await agent.run({
@@ -27405,7 +27599,7 @@ ${prompt}`,
         });
       }
       const previousFindings = lastCheckResult.checkResult.findings;
-      const prompt = buildReReviewPrompt(updatedDiff, previousFindings);
+      const prompt = promptBuilder.buildReReviewPrompt(updatedDiff, previousFindings);
       const { modelTier, modelDef, timeoutSeconds } = resolveRunParams(lastSemanticConfig);
       const { effectivePrompt, acpSessionName } = buildEffectiveRunArgs(prompt);
       const result = await agent.run({
@@ -27465,7 +27659,7 @@ ${prompt}`,
       if (!active) {
         throw new NaxError(`[dialogue] ReviewerSession for story ${storyId} has been destroyed`, "REVIEWER_SESSION_DESTROYED", { stage: "review", storyId, featureName });
       }
-      const prompt = buildDebateResolverPrompt(proposals, critiques, diff, story, semanticConfig, resolverContext);
+      const prompt = promptBuilder.buildResolverPrompt(proposals, critiques, diff, story, resolverContext);
       const { modelTier, modelDef, timeoutSeconds } = resolveRunParams(semanticConfig);
       const { effectivePrompt, acpSessionName } = buildEffectiveRunArgs(prompt);
       const result = await agent.run({
@@ -27500,7 +27694,7 @@ ${prompt}`,
         throw new NaxError(`[dialogue] reReviewDebate() called before any resolveDebate() on story ${storyId}`, "NO_REVIEW_RESULT", { stage: "review", storyId });
       }
       const previousFindings = lastCheckResult.checkResult.findings;
-      const prompt = buildDebateReReviewPrompt(proposals, critiques, updatedDiff, previousFindings, resolverContext);
+      const prompt = promptBuilder.buildReResolverPrompt(proposals, critiques, updatedDiff, previousFindings, resolverContext);
       const { modelTier, modelDef, timeoutSeconds } = resolveRunParams(lastSemanticConfig);
       const { effectivePrompt, acpSessionName } = buildEffectiveRunArgs(prompt);
       const result = await agent.run({
@@ -27552,6 +27746,7 @@ ${prompt}`,
   };
 }
 var init_dialogue = __esm(() => {
+  init_prompt_builder();
   init_errors();
 });
 
@@ -36376,7 +36571,7 @@ var package_default;
 var init_package = __esm(() => {
   package_default = {
     name: "@nathapp/nax",
-    version: "0.60.0-canary.1",
+    version: "0.60.0",
     description: "AI Coding Agent Orchestrator \u2014 loops until done",
     type: "module",
     bin: {
@@ -36456,8 +36651,8 @@ var init_version = __esm(() => {
   NAX_VERSION = package_default.version;
   NAX_COMMIT = (() => {
     try {
-      if (/^[0-9a-f]{6,10}$/.test("1de1f9cc"))
-        return "1de1f9cc";
+      if (/^[0-9a-f]{6,10}$/.test("73c9c082"))
+        return "73c9c082";
     } catch {}
     try {
       const result = Bun.spawnSync(["git", "rev-parse", "--short", "HEAD"], {
@@ -73029,6 +73224,20 @@ function validateStory(raw, index, allIds) {
       throw new Error(`[schema] story[${index}].acceptanceCriteria[${i}] must be a string`);
     }
   }
+  let suggestedCriteria;
+  if (s.suggestedCriteria !== undefined && s.suggestedCriteria !== null) {
+    if (!Array.isArray(s.suggestedCriteria)) {
+      throw new Error(`[schema] story[${index}].suggestedCriteria must be an array when present`);
+    }
+    if (s.suggestedCriteria.length > 0) {
+      for (let i = 0;i < s.suggestedCriteria.length; i++) {
+        if (typeof s.suggestedCriteria[i] !== "string") {
+          throw new Error(`[schema] story[${index}].suggestedCriteria[${i}] must be a string`);
+        }
+      }
+      suggestedCriteria = s.suggestedCriteria;
+    }
+  }
   const routing = typeof s.routing === "object" && s.routing !== null ? s.routing : {};
   const rawComplexity = routing.complexity ?? s.complexity;
   if (rawComplexity === undefined || rawComplexity === null) {
@@ -73096,7 +73305,8 @@ function validateStory(raw, index, allIds) {
       ...noTestJustification !== undefined ? { noTestJustification } : {}
     },
     ...workdir !== undefined ? { workdir } : {},
-    ...contextFiles.length > 0 ? { contextFiles } : {}
+    ...contextFiles.length > 0 ? { contextFiles } : {},
+    ...suggestedCriteria !== undefined ? { suggestedCriteria } : {}
   };
 }
 function sanitizeInvalidEscapes(text) {
@@ -73234,7 +73444,8 @@ async function planCommand(workdir, config2, options) {
       outputDir,
       timeoutSeconds,
       dangerouslySkipPermissions: resolvedPerm.skipPermissions,
-      maxInteractionTurns: config2?.agent?.maxInteractionTurns
+      maxInteractionTurns: config2?.agent?.maxInteractionTurns,
+      specContent
     });
     if (debateResult.outcome !== "failed" && debateResult.output) {
       rawResponse = debateResult.output;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@nathapp/nax",
-  "version": "0.60.0-canary.1",
+  "version": "0.60.0",
   "description": "AI Coding Agent Orchestrator — loops until done",
   "type": "module",
   "bin": {