@nathapp/nax 0.64.2-canary.1 → 0.64.2

package/dist/nax.js

@@ -5304,7 +5304,11 @@ class AgentManager {
     const sessionRole = handle.role ?? opts.sessionRole ?? "main";
     const start = Date.now();
     try {
-      const result = await this._sendPrompt(handle, prompt, opts);
+      const rawResult = await this._sendPrompt(handle, prompt, opts);
+      const result = {
+        ...rawResult,
+        protocolIds: rawResult.protocolIds ?? handle.protocolIds
+      };
       const event = {
         kind: "session-turn",
         sessionName: handle.id,
@@ -29564,7 +29568,43 @@ Fix in priority order. After fixing each priority, re-run the failing check(s) a
     return parts.join(`
 `);
   }
-  static testWriterRectification(testFileFindings, story) {
+  static testWriterRectification(findings, story, options) {
+    if (options?.mode === "write-failing-test") {
+      return RectifierPromptBuilder._testWriterWriteFailingTest(findings, story);
+    }
+    return RectifierPromptBuilder._testWriterFixTestFiles(findings, story);
+  }
+  static _testWriterWriteFailingTest(findings, story) {
+    const acList = story.acceptanceCriteria.map((ac, i) => `${i + 1}. ${ac}`).join(`
+`);
+    const findingLines = findings.flatMap((c) => (c.findings ?? []).map((f) => `- [${f.severity}] ${f.file ?? "unknown"}:${f.line ?? "?"} \u2014 ${f.message}`)).join(`
+`);
+    const scopeConstraint = story.workdir ? `
+
+IMPORTANT: Only create or modify test files within \`${story.workdir}/\`. Do NOT touch source files.` : `
+
+IMPORTANT: Only create or modify test files. Do NOT touch source implementation files.`;
+    return `You are writing a failing test that documents spec-correct behavior.
+
+Story: ${story.title} (${story.id})
+
+### Acceptance Criteria
+${acList}
+
+### Source Bugs Found by Adversarial Review
+${findingLines}
+
+**Task:** For each bug above, write a NEW failing test that asserts the spec-correct behavior described in the finding. The test should FAIL with the current (buggy) implementation and PASS once the implementer fixes the source.
+
+Rules:
+1. Write the test against the SPECIFICATION, not the current behavior.
+2. Do NOT fix the source files \u2014 only write test files.
+3. Do NOT modify existing passing tests.
+4. The test must fail with the current code.
+
+Commit your new tests when done.${scopeConstraint}`;
+  }
+  static _testWriterFixTestFiles(testFileFindings, story) {
     const scopeConstraint = story.workdir ? `
 
 IMPORTANT: Only modify test files within \`${story.workdir}/\`. Do NOT touch source files.` : `
@@ -29587,10 +29627,14 @@ IMPORTANT: Only modify test files. Do NOT touch source implementation files.`;
 `);
     const acList = story.acceptanceCriteria.map((ac, i) => `${i + 1}. ${ac}`).join(`
 `);
-    const importantNote = isLintOnly ? "**Important:** Fix the lint errors in the test files listed above. Do NOT modify source implementation files." : `**Important:** These findings are in test files. Before making any changes:
-1. Read the flagged test files to verify each finding is a real issue
-2. Only fix findings that are genuinely incorrect or missing \u2014 do NOT remove tests
-3. Do NOT modify source implementation files`;
+    const importantNote = isLintOnly ? "**Important:** Fix the lint errors in the test files listed above. Do NOT modify source implementation files." : `**Important:** You are encoding the SPECIFICATION, not the current behavior.
+
+Before making any changes:
+1. Read the flagged test files to verify each finding is a real issue.
+2. Do NOT loosen assertions to match current implementation behavior. If a test is failing because the source is wrong, the source is the suspect \u2014 not the test.
+3. Do NOT delete a failing test because the implementation makes it hard to assert on. Refactor the test structure if needed; never silently drop coverage.
+4. If the current behavior disagrees with the acceptance criteria, write the test against the spec and let the implementer fix the source.
+5. Do NOT modify source implementation files.`;
     return `${opener}
 
 Story: ${story.title} (${story.id})
@@ -30355,7 +30399,9 @@ function turnResultToAgentResult(r) {
     durationMs: 0,
     estimatedCostUsd: r.estimatedCostUsd ?? 0,
     exactCostUsd: r.exactCostUsd,
-    tokenUsage: r.tokenUsage
+    tokenUsage: r.tokenUsage,
+    protocolIds: r.protocolIds,
+    internalRoundTrips: r.internalRoundTrips
   };
 }
 function buildHopCallback(ctx, sessionId, _initialOptions) {
@@ -32442,7 +32488,9 @@ var init_autofix_test_writer = __esm(() => {
     session: { role: "test-writer", lifetime: "fresh" },
     config: autofixConfigSelector,
     build(input, _ctx) {
-      const prompt = RectifierPromptBuilder.testWriterRectification(input.failedChecks, input.story);
+      const prompt = RectifierPromptBuilder.testWriterRectification(input.failedChecks, input.story, {
+        mode: input.mode
+      });
       return {
         role: { id: "role", content: "", overridable: false },
         task: { id: "task", content: prompt, overridable: false }
@@ -36840,9 +36888,9 @@ async function runTrackedSession(state, id, runner, request) {
     ...request,
     runOptions: {
       ...request.runOptions,
-      onSessionEstablished: (protocolIds, sessionName2) => {
+      onSessionEstablished: (protocolIds2, sessionName2) => {
         try {
-          state.bindHandle(id, sessionName2, protocolIds);
+          state.bindHandle(id, sessionName2, protocolIds2);
         } catch (err) {
           getLogger().warn("session", "bindHandle via onSessionEstablished failed", {
             storyId: state.sessions.get(id)?.storyId,
@@ -36850,7 +36898,7 @@ async function runTrackedSession(state, id, runner, request) {
             error: err instanceof Error ? err.message : String(err)
           });
         }
-        callerCallback?.(protocolIds, sessionName2);
+        callerCallback?.(protocolIds2, sessionName2);
       }
     }
   };
@@ -36925,6 +36973,8 @@ async function runTrackedSession(state, id, runner, request) {
   if (current?.state === "RUNNING") {
     state.transition(id, result.success ? "COMPLETED" : "FAILED");
   }
+  const protocolIds = result.protocolIds ?? current?.protocolIds;
+  const turn = Math.max(result.internalRoundTrips ?? 1, 1);
   const sessionName = state.nameFor({
     workdir: pre.workdir,
     featureName: pre.featureName,
@@ -36943,10 +36993,10 @@ async function runTrackedSession(state, id, runner, request) {
     featureName: pre.featureName,
     workdir: pre.workdir,
     resolvedPermissions,
-    turn: result.internalRoundTrips ?? 0,
+    turn,
     protocolIds: {
-      sessionId: result.protocolIds?.sessionId ?? null,
-      recordId: result.protocolIds?.recordId ?? null
+      sessionId: protocolIds?.sessionId ?? null,
+      recordId: protocolIds?.recordId ?? null
     },
     origin: "runTrackedSession",
     tokenUsage: result.tokenUsage,
@@ -37336,11 +37386,12 @@ class SessionManager {
     }
     this._busySessions.add(handle.id);
     try {
-      return await adapter.sendTurn(handle, prompt, {
+      const result = await adapter.sendTurn(handle, prompt, {
         interactionHandler: opts?.interactionHandler ?? NO_OP_INTERACTION_HANDLER,
         signal: opts?.signal,
         maxTurns: opts?.maxTurns
       });
+      return { ...result, protocolIds: result.protocolIds ?? handle.protocolIds };
     } catch (err) {
       if (opts?.signal?.aborted || err instanceof Error && err.name === "AbortError") {
         this._cancelledSessions.add(handle.id);
@@ -37648,7 +37699,9 @@ function createSessionRunHop(sessionManager) {
           durationMs: Date.now() - startMs,
           estimatedCostUsd: turnResult.estimatedCostUsd ?? 0,
           exactCostUsd: turnResult.exactCostUsd,
-          tokenUsage: turnResult.tokenUsage
+          tokenUsage: turnResult.tokenUsage,
+          protocolIds: handle.protocolIds,
+          internalRoundTrips: turnResult.internalRoundTrips
         }
       };
     } catch (err) {
@@ -40071,9 +40124,66 @@ async function runFixCycle(cycle, ctx, cycleName, _deps = {}) {
         op: strategy.fixOp.name,
         targetFiles: extracted.targetFiles ?? [],
         summary: extracted.summary ?? "",
+        ...extracted.unresolved ? { unresolved: extracted.unresolved } : {},
         costUsd: extracted.costUsd
       });
     }
+    const unresolvedFa = fixesApplied.find((fa) => fa.unresolved);
+    if (unresolvedFa) {
+      const finishedAt2 = now();
+      cycle.iterations.push({
+        iterationNum: cycle.iterations.length + 1,
+        findingsBefore,
+        fixesApplied,
+        findingsAfter: cycle.findings,
+        outcome: "unchanged",
+        startedAt,
+        finishedAt: finishedAt2
+      });
+      logger?.info("findings.cycle", "cycle exited \u2014 agent gave up", {
+        storyId,
+        packageDir,
+        cycleName,
+        reason: "agent-gave-up",
+        strategyName: unresolvedFa.strategyName,
+        unresolvedDetail: unresolvedFa.unresolved
+      });
+      return {
+        iterations: cycle.iterations,
+        finalFindings: cycle.findings,
+        exitReason: "agent-gave-up",
+        unresolvedDetail: unresolvedFa.unresolved,
+        costUsd: totalCostUsd
+      };
+    }
+    const provisionalIterations = [...cycle.iterations, { fixesApplied }];
+    const allExhausted = group.every((s) => countStrategyAttempts(provisionalIterations, s.name) >= s.maxAttempts);
+    if (allExhausted) {
+      const finishedAt2 = now();
+      cycle.iterations.push({
+        iterationNum: cycle.iterations.length + 1,
+        findingsBefore,
+        fixesApplied,
+        findingsAfter: cycle.findings,
+        outcome: "unchanged",
+        startedAt,
+        finishedAt: finishedAt2
+      });
+      logger?.info("findings.cycle", "cycle exited \u2014 strategy attempt cap reached (skipped final validate)", {
+        storyId,
+        packageDir,
+        cycleName,
+        reason: "max-attempts-per-strategy",
+        exhaustedStrategy: group[0]?.name
+      });
+      return {
+        iterations: cycle.iterations,
+        finalFindings: cycle.findings,
+        exitReason: "max-attempts-per-strategy",
+        exhaustedStrategy: group[0]?.name,
+        costUsd: totalCostUsd
+      };
+    }
     let findingsAfter;
     let validatorAttempt = 0;
     for (;; ) {
@@ -41636,7 +41746,13 @@ function collectCurrentFindings(ctx) {
   });
 }
 function collectTestTargetedChecks(ctx) {
-  return collectFailedChecks(ctx).filter((c) => c.findings?.some((f) => f.fixTarget === "test"));
+  return collectFailedChecks(ctx).map((c) => ({ ...c, findings: (c.findings ?? []).filter((f) => f.fixTarget === "test") })).filter((c) => c.findings.length > 0);
+}
+function collectAdversarialSourceChecks(ctx) {
+  return collectFailedChecks(ctx).map((c) => ({
+    ...c,
+    findings: (c.findings ?? []).filter((f) => (f.fixTarget ?? "source") === "source" && f.severity === "error" && f.source === "adversarial-review")
+  })).filter((c) => c.check === "adversarial" && c.findings.length > 0);
 }
 function buildAutofixStrategies(ctx, maxAttempts) {
   const implementer = {
@@ -41650,31 +41766,41 @@ function buildAutofixStrategies(ctx, maxAttempts) {
       story: ctx.story
     }),
     extractApplied: (output) => ({
-      summary: output.unresolvedReason ?? ""
+      summary: output.unresolvedReason ?? "",
+      unresolved: output.unresolvedReason
     })
   };
   const testWriter = {
     name: "autofix-test-writer",
-    appliesTo: (f) => f.fixTarget === "test",
+    appliesTo: (f) => f.fixTarget === "test" || (f.fixTarget ?? "source") === "source" && f.severity === "error" && f.source === "adversarial-review",
     fixOp: testWriterRectifyOp,
     maxAttempts: 1,
     coRun: "co-run-sequential",
-    buildInput: (_findings, _prior, _cycleCtx) => ({
-      failedChecks: collectTestTargetedChecks(ctx),
-      story: ctx.story
-    })
-  };
-  return [implementer, testWriter];
-}
-function findUnresolvedReason(result) {
-  for (const iter of result.iterations) {
-    for (const fa of iter.fixesApplied) {
-      if (fa.strategyName === "autofix-implementer" && fa.summary) {
-        return fa.summary;
+    buildInput: (findings, _prior, _cycleCtx) => {
+      const hasSourceBug = findings.some((f) => (f.fixTarget ?? "source") === "source" && f.source === "adversarial-review");
+      if (hasSourceBug) {
+        return { failedChecks: collectAdversarialSourceChecks(ctx), story: ctx.story, mode: "write-failing-test" };
       }
+      return { failedChecks: collectTestTargetedChecks(ctx), story: ctx.story };
     }
+  };
+  return [testWriter, implementer];
+}
+function buildEscalationDigest(findings) {
+  const byFile = new Map;
+  for (const f of findings) {
+    const file3 = f.file ?? "unknown";
+    const list = byFile.get(file3) ?? [];
+    list.push(f);
+    byFile.set(file3, list);
   }
-  return;
+  const lines = [...byFile.entries()].map(([file3, fs]) => {
+    const categories = fs.map((f) => f.category ?? f.source).join(", ");
+    return `  - ${categories} in ${file3}`;
+  });
+  return `Autofix exhausted: ${findings.length} finding${findings.length !== 1 ? "s" : ""} remain
+${lines.join(`
+`)}`;
 }
 async function writeShadowReport(ctx, result, initialFindingsCount) {
   const logger = getLogger();
@@ -41728,7 +41854,8 @@ async function runAgentRectificationV2(ctx, _lintFixCmd, _formatFixCmd, _effecti
   const result = await runFixCycle(cycle, cycleCtx, "autofix-v2");
   ctx.autofixPriorIterations = result.iterations;
   await writeShadowReport(ctx, result, initialFindings.length);
-  const unresolvedReason = findUnresolvedReason(result);
+  const unresolvedReason = result.exitReason === "agent-gave-up" ? result.unresolvedDetail : undefined;
+  const escalationDigest = result.exitReason === "max-attempts-per-strategy" && result.finalFindings.length > 0 ? buildEscalationDigest(result.finalFindings) : undefined;
   const succeeded = result.exitReason === "resolved" || result.finalFindings.length === 0;
   logger.info("autofix-cycle", "V2 fix cycle complete", {
     storyId,
@@ -41736,9 +41863,15 @@ async function runAgentRectificationV2(ctx, _lintFixCmd, _formatFixCmd, _effecti
     iterations: result.iterations.length,
     finalFindingsCount: result.finalFindings.length,
     succeeded,
-    ...unresolvedReason ? { unresolvedReason } : {}
+    ...unresolvedReason ? { unresolvedReason } : {},
+    ...escalationDigest ? { escalationDigest } : {}
   });
-  return { succeeded, cost: 0, ...unresolvedReason ? { unresolvedReason } : {} };
+  return {
+    succeeded,
+    cost: 0,
+    ...unresolvedReason ? { unresolvedReason } : {},
+    ...escalationDigest ? { escalationDigest } : {}
+  };
 }
 var init_autofix_cycle = __esm(() => {
   init_findings();
@@ -43132,10 +43265,7 @@ async function runAdversarialReview(opts) {
     const durationMs2 = Date.now() - startTime;
     logger?.warn("review", `Adversarial review failed: ${blockingFindings.length} blocking findings`, {
       storyId: story.id,
-      durationMs: durationMs2
-    });
-    logger?.debug("review", "Adversarial review findings", {
-      storyId: story.id,
+      durationMs: durationMs2,
       findings: blockingFindings.map((f) => ({
         severity: f.severity,
         category: f.category,
@@ -45106,9 +45236,10 @@ var init_autofix = __esm(() => {
       const {
         succeeded: agentFixed,
         cost: agentCost,
-        unresolvedReason
+        unresolvedReason,
+        escalationDigest
       } = await _autofixDeps.runAgentRectification(ctx, lintFixCmd, formatFixCmd, ctx.workdir);
-      if (unresolvedReason) {
+      if (!agentFixed && unresolvedReason) {
         if (ctx.mechanicalFailedOnly) {
           logger.warn("autofix", "Mechanical-only failure unfixable \u2014 proceeding (LLM review passed)", {
             storyId: ctx.story.id,
@@ -45155,7 +45286,7 @@ var init_autofix = __esm(() => {
       logger.warn("autofix", "Autofix exhausted \u2014 escalating", { storyId: ctx.story.id });
       return {
         action: "escalate",
-        reason: "Autofix exhausted: review still failing after fix attempts",
+        reason: escalationDigest ?? "Autofix exhausted: review still failing after fix attempts",
         cost: agentCost
       };
     }
@@ -46763,7 +46894,8 @@ async function runThreeSessionTdd(options) {
     lite = false,
     _recursionDepth = 0,
     projectDir,
-    agentManager
+    agentManager,
+    runtime
   } = options;
   const logger = getLogger();
   const MAX_RECURSION_DEPTH = 2;
@@ -46906,7 +47038,7 @@ async function runThreeSessionTdd(options) {
     };
   }
   const implementerBinding = getTddSessionBinding?.("implementer");
-  const { cost: fullSuiteGateCost, fullSuiteGatePassed } = await runFullSuiteGate(story, config2, workdir, agentManager, implementerTier, lite, logger, featureName, projectDir, implementerBinding?.sessionManager, implementerBinding?.sessionId);
+  const { cost: fullSuiteGateCost, fullSuiteGatePassed } = await runFullSuiteGate(story, config2, workdir, agentManager, implementerTier, lite, logger, featureName, projectDir, implementerBinding?.sessionManager, implementerBinding?.sessionId, runtime);
   const session3Ref = await captureGitRef(workdir) ?? "HEAD";
   const verifierBundle = await getTddContextBundle?.("verifier") ?? tddContextBundles?.verifier;
   const session3 = await runTddSessionOp(verifyTddOp, options, session3Ref, verifierBundle, getTddSessionBinding?.("verifier"));
@@ -47162,7 +47294,8 @@ async function runThreeSessionTddFromCtx(ctx, opts) {
     interactionChain: ctx.interaction,
     projectDir: ctx.projectDir,
     abortSignal: ctx.abortSignal,
-    agentManager: ctx.agentManager
+    agentManager: ctx.agentManager,
+    runtime: ctx.runtime
   });
 }
 var init_orchestrator_ctx = __esm(() => {
@@ -50141,7 +50274,7 @@ var package_default;
 var init_package = __esm(() => {
   package_default = {
     name: "@nathapp/nax",
-    version: "0.64.2-canary.1",
+    version: "0.64.2",
     description: "AI Coding Agent Orchestrator \u2014 loops until done",
     type: "module",
     bin: {
@@ -50225,8 +50358,8 @@ var init_version = __esm(() => {
   NAX_VERSION = package_default.version;
   NAX_COMMIT = (() => {
     try {
-      if (/^[0-9a-f]{6,10}$/.test("b8feb3bf"))
-        return "b8feb3bf";
+      if (/^[0-9a-f]{6,10}$/.test("7a4c7325"))
+        return "7a4c7325";
     } catch {}
     try {
       const result = Bun.spawnSync(["git", "rev-parse", "--short", "HEAD"], {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@nathapp/nax",
-  "version": "0.64.2-canary.1",
+  "version": "0.64.2",
   "description": "AI Coding Agent Orchestrator — loops until done",
   "type": "module",
   "bin": {