@nathapp/nax 0.59.1 → 0.59.3

package/dist/nax.js

@@ -18060,7 +18060,7 @@ function isLegacyFlatModels(val) {
   }
   return false;
 }
-var TokenPricingSchema, ModelDefSchema, ModelEntrySchema, PerAgentModelMapSchema, ModelMapSchema, ModelTierSchema, TierConfigSchema, AutoModeConfigSchema, RectificationConfigSchema, RegressionGateConfigSchema, SmartTestRunnerConfigSchema, SMART_TEST_RUNNER_DEFAULT, smartTestRunnerFieldSchema, ExecutionConfigSchema, QualityConfigSchema, TddConfigSchema, ConstitutionConfigSchema, AnalyzeConfigSchema, SemanticReviewConfigSchema, ReviewDialogueConfigSchema, ReviewConfigSchema, PlanConfigSchema, AcceptanceFixConfigSchema, AcceptanceConfigSchema, TestCoverageConfigSchema, ContextAutoDetectConfigSchema, ContextConfigSchema, LlmRoutingConfigSchema, RoutingConfigSchema, OptimizerConfigSchema, PluginConfigEntrySchema, HooksConfigSchema, InteractionConfigSchema, StorySizeGateConfigSchema, PromptAuditConfigSchema, AgentConfigSchema, PrecheckConfigSchema, PromptsConfigSchema, ProjectProfileSchema, VALID_AGENT_TYPES, GenerateConfigSchema, DebaterSchema, toObject = (val) => val === undefined || val === null ? {} : val, RESOLVER_TYPES, makeResolverSchema = (defaultType) => exports_external.preprocess(toObject, exports_external.object({
+var TokenPricingSchema, ModelDefSchema, ModelEntrySchema, PerAgentModelMapSchema, ModelMapSchema, ModelTierSchema, TierConfigSchema, AutoModeConfigSchema, RectificationConfigSchema, RegressionGateConfigSchema, SmartTestRunnerConfigSchema, SMART_TEST_RUNNER_DEFAULT, smartTestRunnerFieldSchema, ExecutionConfigSchema, QualityConfigSchema, TddConfigSchema, ConstitutionConfigSchema, AnalyzeConfigSchema, SemanticReviewConfigSchema, ReviewDialogueConfigSchema, ReviewConfigSchema, PlanConfigSchema, AcceptanceFixConfigSchema, AcceptanceConfigSchema, TestCoverageConfigSchema, ContextAutoDetectConfigSchema, ContextConfigSchema, LlmRoutingConfigSchema, RoutingConfigSchema, OptimizerConfigSchema, PluginConfigEntrySchema, HooksConfigSchema, InteractionConfigSchema, StorySizeGateConfigSchema, PromptAuditConfigSchema, AgentConfigSchema, PrecheckConfigSchema, PromptsConfigSchema, ProjectProfileSchema, VALID_AGENT_TYPES, GenerateConfigSchema, DebaterPersonaEnum, DebaterSchema, toObject = (val) => val === undefined || val === null ? {} : val, RESOLVER_TYPES, makeResolverSchema = (defaultType) => exports_external.preprocess(toObject, exports_external.object({
   type: exports_external.enum(RESOLVER_TYPES).default(defaultType),
   agent: exports_external.string().min(1).optional(),
   tieBreaker: exports_external.string().min(1).optional(),
@@ -18072,7 +18072,8 @@ var TokenPricingSchema, ModelDefSchema, ModelEntrySchema, PerAgentModelMapSchema
   rounds: exports_external.number().int().min(1).default(defaults.rounds),
   mode: exports_external.enum(["panel", "hybrid"]).default("panel"),
   debaters: exports_external.array(DebaterSchema).min(2, "debaters must have at least 2 entries").optional(),
-  timeoutSeconds: exports_external.number().int().positive().default(600)
+  timeoutSeconds: exports_external.number().int().positive().default(600),
+  autoPersona: exports_external.boolean().default(false)
 })), DebateConfigSchema, NaxConfigSchema;
 var init_schemas3 = __esm(() => {
   init_zod();
@@ -18431,9 +18432,11 @@ var init_schemas3 = __esm(() => {
   GenerateConfigSchema = exports_external.object({
     agents: exports_external.array(exports_external.enum(VALID_AGENT_TYPES)).optional()
   });
+  DebaterPersonaEnum = exports_external.enum(["challenger", "pragmatist", "completionist", "security", "testability"]);
   DebaterSchema = exports_external.object({
     agent: exports_external.string().min(1, "debater.agent must be non-empty"),
-    model: exports_external.string().min(1, "debater.model must be non-empty").optional()
+    model: exports_external.string().min(1, "debater.model must be non-empty").optional(),
+    persona: DebaterPersonaEnum.optional()
   });
   RESOLVER_TYPES = ["synthesis", "majority-fail-closed", "majority-fail-open", "custom"];
   DebateConfigSchema = exports_external.preprocess(toObject, exports_external.object({
@@ -18722,7 +18725,8 @@ var init_schemas3 = __esm(() => {
           sessionMode: "stateful",
           rounds: 3,
           mode: "panel",
-          timeoutSeconds: 600
+          timeoutSeconds: 600,
+          autoPersona: false
         },
         review: {
           enabled: true,
@@ -18730,7 +18734,8 @@ var init_schemas3 = __esm(() => {
           sessionMode: "one-shot",
           rounds: 2,
           mode: "panel",
-          timeoutSeconds: 600
+          timeoutSeconds: 600,
+          autoPersona: false
         },
         acceptance: {
           enabled: false,
@@ -18738,7 +18743,8 @@ var init_schemas3 = __esm(() => {
           sessionMode: "one-shot",
           rounds: 1,
           mode: "panel",
-          timeoutSeconds: 600
+          timeoutSeconds: 600,
+          autoPersona: false
         },
         rectification: {
           enabled: false,
@@ -18746,7 +18752,8 @@ var init_schemas3 = __esm(() => {
           sessionMode: "one-shot",
           rounds: 1,
           mode: "panel",
-          timeoutSeconds: 600
+          timeoutSeconds: 600,
+          autoPersona: false
         },
         escalation: {
           enabled: false,
@@ -18754,7 +18761,8 @@ var init_schemas3 = __esm(() => {
           sessionMode: "one-shot",
           rounds: 1,
           mode: "panel",
-          timeoutSeconds: 600
+          timeoutSeconds: 600,
+          autoPersona: false
         }
       }
     })),
@@ -21365,28 +21373,85 @@ var init_config = __esm(() => {
   init_profile();
 });
 
-// src/debate/prompts.ts
-function buildCritiquePrompt(taskPrompt, allProposals, debaterIndex) {
-  const othersProposals = allProposals.filter((_, i) => i !== debaterIndex);
-  const proposalsSection = othersProposals.map((p, i) => `### Proposal ${i + 1}
-${p}`).join(`
-
-`);
-  return `You are reviewing proposals from other agents for the following task.
-
-## Task
-${taskPrompt}
+// src/utils/llm-json.ts
+function extractJsonFromMarkdown(text) {
+  const match = text.match(/```(?:json)?\s*\n([\s\S]*?)\n?\s*```/);
+  if (match) {
+    return match[1] ?? text;
+  }
+  return text;
+}
+function stripTrailingCommas(text) {
+  return text.replace(/,\s*([}\]])/g, "$1");
+}
+function extractJsonObject(text) {
+  const objStart = text.indexOf("{");
+  const arrStart = text.indexOf("[");
+  let start;
+  let closeChar;
+  if (objStart === -1 && arrStart === -1)
+    return null;
+  if (objStart === -1) {
+    start = arrStart;
+    closeChar = "]";
+  } else if (arrStart === -1) {
+    start = objStart;
+    closeChar = "}";
+  } else if (objStart < arrStart) {
+    start = objStart;
+    closeChar = "}";
+  } else {
+    start = arrStart;
+    closeChar = "]";
+  }
+  const end = text.lastIndexOf(closeChar);
+  if (end <= start)
+    return null;
+  return text.slice(start, end + 1);
+}
+function wrapJsonPrompt(prompt) {
+  return `IMPORTANT: Your entire response must be a single JSON object or array. Do not explain your reasoning. Do not use markdown formatting. Output ONLY the JSON.
 
-## Other Agents' Proposals
-${proposalsSection}
+${prompt.trim()}
 
-Please critique these proposals and provide your refined analysis, identifying strengths, weaknesses, and your own updated position.`;
+YOUR RESPONSE MUST START WITH { OR [ AND END WITH } OR ]. No other text.`;
 }
-function buildSynthesisPrompt(proposals, critiques) {
-  const proposalsSection = proposals.map((p, i) => `### Proposal ${i + 1}
+function parseLLMJson(text) {
+  const trimmed = text.trim();
+  try {
+    return JSON.parse(trimmed);
+  } catch {}
+  const fromFence = extractJsonFromMarkdown(trimmed);
+  if (fromFence !== trimmed) {
+    try {
+      return JSON.parse(stripTrailingCommas(fromFence));
+    } catch {}
+  }
+  const bareJson = extractJsonObject(trimmed);
+  if (bareJson) {
+    try {
+      return JSON.parse(stripTrailingCommas(bareJson));
+    } catch {}
+  }
+  throw new SyntaxError("[llm-json] Failed to parse LLM response as JSON");
+}
+function tryParseLLMJson(text) {
+  try {
+    return parseLLMJson(text);
+  } catch {
+    return null;
+  }
+}
+
+// src/debate/resolvers.ts
+function buildProposalsSection(proposals) {
+  return proposals.map((p, i) => `### Proposal ${i + 1}
 ${p}`).join(`
 
 `);
+}
+function buildSynthesisPrompt(proposals, critiques) {
+  const proposalsSection = buildProposalsSection(proposals);
   const critiquesSection = critiques.length > 0 ? `
 
 ## Critiques
@@ -21402,10 +21467,7 @@ ${proposalsSection}${critiquesSection}
 Please synthesize these into the best possible unified response, incorporating the strongest elements from each proposal.`;
 }
 function buildJudgePrompt(proposals, critiques) {
-  const proposalsSection = proposals.map((p, i) => `### Proposal ${i + 1}
-${p}`).join(`
-
-`);
+  const proposalsSection = buildProposalsSection(proposals);
   const critiquesSection = critiques.length > 0 ? `
 
 ## Critiques
@@ -21420,28 +21482,6 @@ ${proposalsSection}${critiquesSection}
 
 As the judge, provide your final verdict with clear reasoning, selecting or synthesizing the best approach.`;
 }
-function buildRebuttalContext(prompt, proposals, rebuttalOutputs, currentDebaterIndex) {
-  const proposalsSection = proposals.map((p, i) => `### Proposal ${i + 1} (${p.debater.agent})
-${p.output}`).join(`
-
-`);
-  const rebuttalsSection = rebuttalOutputs.length > 0 ? `
-
-## Previous Rebuttals
-${rebuttalOutputs.map((r, i) => `${i + 1}. ${r}`).join(`
-
-`)}` : "";
-  const debaterNumber = currentDebaterIndex + 1;
-  return `${prompt}
-
-## Proposals
-${proposalsSection}${rebuttalsSection}
-
-## Your Task
-You are debater ${debaterNumber}. Provide your rebuttal to the proposals and previous rebuttals above.`;
-}
-
-// src/debate/resolvers.ts
 function stripMarkdownFence(text) {
   const match = text.match(/^```(?:json)?\s*\n?([\s\S]*?)\n?```\s*$/);
   return match ? match[1] ?? text : text;
@@ -21475,7 +21515,10 @@ function majorityResolver(proposals, failOpen) {
   return passCount > failCount ? "passed" : "failed";
 }
 async function synthesisResolver(proposals, critiques, opts) {
-  const prompt = buildSynthesisPrompt(proposals, critiques);
+  const base = buildSynthesisPrompt(proposals, critiques);
+  const prompt = opts.promptSuffix ? `${base}
+
+${opts.promptSuffix}` : base;
   return opts.adapter.complete(prompt, opts.completeOptions);
 }
 async function judgeResolver(proposals, critiques, resolverConfig, opts) {
@@ -21488,7 +21531,6 @@ async function judgeResolver(proposals, critiques, resolverConfig, opts) {
   return adapter.complete(prompt, opts.completeOptions);
 }
 var DEFAULT_FALLBACK_AGENT = "claude";
-var init_resolvers = () => {};
 
 // src/debate/session-helpers.ts
 function resolveDebaterModel(debater, config2) {
@@ -21564,7 +21606,7 @@ function resolveModelDefForDebater(debater, tier, config2) {
     return resolveModelForAgent(configModels, debater.agent, "fast", configDefaultAgent);
   }
 }
-async function resolveOutcome(proposalOutputs, critiqueOutputs, stageConfig, config2, storyId, timeoutMs, workdir, featureName, reviewerSession, resolverContext) {
+async function resolveOutcome(proposalOutputs, critiqueOutputs, stageConfig, config2, storyId, timeoutMs, workdir, featureName, reviewerSession, resolverContext, promptSuffix) {
   const resolverConfig = stageConfig.resolver;
   const logger = _debateSessionDeps.getSafeLogger();
   if (reviewerSession && resolverContext) {
@@ -21578,21 +21620,13 @@ async function resolveOutcome(proposalOutputs, critiqueOutputs, stageConfig, con
         let passCount = 0;
         let failCount = 0;
         for (const proposal of proposalOutputs) {
-          try {
-            const stripped = proposal.trim().replace(/^```(?:json)?\s*\n?/, "").replace(/\n?```\s*$/, "");
-            const parsed = JSON.parse(stripped);
-            if (typeof parsed.passed === "boolean" && parsed.passed)
-              passCount++;
-            else if (failOpen)
-              passCount++;
-            else
-              failCount++;
-          } catch {
-            if (failOpen)
-              passCount++;
-            else
-              failCount++;
-          }
+          const parsed = tryParseLLMJson(proposal);
+          if (parsed !== null && typeof parsed.passed === "boolean" && parsed.passed)
+            passCount++;
+          else if (failOpen)
+            passCount++;
+          else
+            failCount++;
         }
         debateCtx.majorityVote = { passed: rawOutcome === "passed", passCount, failCount };
       }
@@ -21633,20 +21667,23 @@ async function resolveOutcome(proposalOutputs, critiqueOutputs, stageConfig, con
       resolverCostUsd: 0
     };
   }
-  const implementerSessionName = workdir !== undefined ? buildSessionName(workdir, featureName, storyId, "implementer") : undefined;
   if (resolverConfig.type === "synthesis") {
     const agentName = resolverConfig.agent ?? RESOLVER_FALLBACK_AGENT;
     const adapter = _debateSessionDeps.getAgent(agentName, config2);
     if (adapter) {
+      const synthesisSessionName = workdir !== undefined ? buildSessionName(workdir, featureName, storyId, "synthesis") : undefined;
       const resolverResult = await synthesisResolver(proposalOutputs, critiqueOutputs, {
         adapter,
+        promptSuffix,
         completeOptions: {
           model: resolveDebaterModel({ agent: agentName }, config2),
           config: config2,
           storyId,
+          featureName,
+          workdir,
           sessionRole: "synthesis",
           timeoutMs,
-          ...implementerSessionName !== undefined && { sessionName: implementerSessionName }
+          ...synthesisSessionName !== undefined && { sessionName: synthesisSessionName }
         }
       });
       return {
@@ -21659,6 +21696,7 @@ async function resolveOutcome(proposalOutputs, critiqueOutputs, stageConfig, con
   }
   if (resolverConfig.type === "custom") {
     const agentName = resolverConfig.agent ?? RESOLVER_FALLBACK_AGENT;
+    const judgeSessionName = workdir !== undefined ? buildSessionName(workdir, featureName, storyId, "judge") : undefined;
     const resolverResult = await judgeResolver(proposalOutputs, critiqueOutputs, resolverConfig, {
       getAgent: (name) => _debateSessionDeps.getAgent(name, config2),
       defaultAgentName: RESOLVER_FALLBACK_AGENT,
@@ -21666,9 +21704,11 @@ async function resolveOutcome(proposalOutputs, critiqueOutputs, stageConfig, con
         model: resolveDebaterModel({ agent: agentName }, config2),
         config: config2,
         storyId,
+        featureName,
+        workdir,
         sessionRole: "judge",
         timeoutMs,
-        ...implementerSessionName !== undefined && { sessionName: implementerSessionName }
+        ...judgeSessionName !== undefined && { sessionName: judgeSessionName }
       }
     });
     return {
@@ -21685,7 +21725,6 @@ var init_session_helpers = __esm(() => {
   init_registry();
   init_config();
   init_logger2();
-  init_resolvers();
   _debateSessionDeps = {
     getAgent: (name, config2) => config2 ? createAgentRegistry(config2).getAgent(name) : getAgent(name),
     getSafeLogger,
@@ -21719,6 +21758,171 @@ async function allSettledBounded(tasks, limit) {
   return results;
 }
 
+// src/debate/personas.ts
+function resolvePersonas(debaters, stage, autoPersona) {
+  if (!autoPersona)
+    return debaters;
+  const rotation = stage === "plan" ? PLAN_ROTATION : REVIEW_ROTATION;
+  let rotationIndex = 0;
+  return debaters.map((d) => {
+    if (d.persona)
+      return d;
+    const assigned = rotation[rotationIndex % rotation.length];
+    rotationIndex++;
+    return { ...d, persona: assigned };
+  });
+}
+var PERSONA_FRAGMENTS, PLAN_ROTATION, REVIEW_ROTATION;
+var init_personas = __esm(() => {
+  PERSONA_FRAGMENTS = {
+    challenger: {
+      identity: "You are the challenger \u2014 your job is to stress-test proposals and find weaknesses.",
+      lens: "Question every assumption. Look for missing edge cases, unhandled error states, " + "and scenarios where the proposed approach could break under real-world conditions. " + "If a proposal lacks justification for a design choice, call it out."
+    },
+    pragmatist: {
+      identity: "You are the pragmatist \u2014 your job is to find the simplest path that satisfies the spec.",
+      lens: "Favour minimal scope, fewest files changed, and lowest complexity. " + "Challenge any proposal that adds abstraction, configuration, or code beyond what the spec requires. " + "If something can be done in 5 lines instead of 50, advocate for the 5-line version."
+    },
+    completionist: {
+      identity: "You are the completionist \u2014 your job is to ensure nothing is missed.",
+      lens: "Verify every acceptance criterion is addressed. Check that edge cases have tests, " + "that error messages are user-friendly, and that the implementation handles all status/state variants. " + "If the spec is ambiguous, flag it and propose the safer interpretation."
+    },
+    security: {
+      identity: "You are the security reviewer \u2014 your job is to surface risks before they ship.",
+      lens: "Evaluate input validation, secret handling, injection vectors, and trust boundaries. " + "Check that user-supplied data is never used unsanitised in commands, queries, or file paths. " + "If the proposal touches auth, permissions, or external APIs, apply extra scrutiny."
+    },
+    testability: {
+      identity: "You are the testability advocate \u2014 your job is to ensure the design is verifiable.",
+      lens: "Assess whether the proposed implementation can be tested without mocks, " + "whether test boundaries are clean, and whether the acceptance criteria are machine-verifiable. " + "Challenge any design that makes testing harder (global state, tight coupling, hidden side effects)."
+    }
+  };
+  PLAN_ROTATION = ["challenger", "pragmatist", "completionist", "security", "testability"];
+  REVIEW_ROTATION = ["security", "completionist", "testability", "challenger", "pragmatist"];
+});
+
+// src/debate/prompt-builder.ts
+class DebatePromptBuilder {
+  stageContext;
+  options;
+  constructor(stageContext, options) {
+    this.stageContext = stageContext;
+    this.options = options;
+  }
+  buildProposalPrompt(debaterIndex) {
+    const personaBlock = this.buildPersonaBlock(debaterIndex);
+    return `${this.stageContext.taskContext}${personaBlock}
+
+${this.stageContext.outputFormat}`;
+  }
+  buildCritiquePrompt(debaterIndex, proposals) {
+    const otherProposals = proposals.filter((_, i) => i !== debaterIndex);
+    const proposalsSection = this.buildProposalsSection(otherProposals);
+    const personaBlock = this.buildPersonaBlock(debaterIndex);
+    return `You are reviewing proposals for a ${this.stageContext.stage} task.
+
+## Task
+${this.stageContext.taskContext}${personaBlock}
+
+## Other Agents' Proposals
+${proposalsSection}
+
+Please critique these proposals and provide your refined analysis, identifying strengths, weaknesses, and your own updated position.`;
+  }
+  buildRebuttalPrompt(debaterIndex, proposals, priorRebuttals) {
+    const contextBlock = this.options.sessionMode === "one-shot" ? `${this.stageContext.taskContext}
+
+` : "";
+    const proposalsSection = this.buildProposalsSection(proposals);
+    const rebuttalsSection = this.buildRebuttalsSection(priorRebuttals);
+    const personaBlock = this.buildPersonaBlock(debaterIndex);
+    const debaterNumber = debaterIndex + 1;
+    return `${contextBlock}## Proposals
+${proposalsSection}${rebuttalsSection}${personaBlock}
+
+## Your Task
+You are debater ${debaterNumber}. Provide your critique in prose.
+Identify strengths, weaknesses, and specific improvements for each proposal.
+Do NOT output JSON \u2014 focus on analysis only.`;
+  }
+  buildSynthesisPrompt(proposals, critiques, promptSuffix) {
+    const proposalsSection = this.buildProposalsSection(proposals);
+    const critiquesSection = this.buildCritiquesSection(critiques);
+    return `You are a synthesis agent. Your task is to combine the strongest elements from multiple proposals into a single, optimal response.
+
+${this.stageContext.taskContext}
+
+## Proposals
+${proposalsSection}
+
+## Critiques
+${critiquesSection}
+
+Please synthesize these into the best possible unified response, incorporating the strongest elements from each proposal.
+${this.stageContext.outputFormat}${promptSuffix ? `
+${promptSuffix}` : ""}`;
+  }
+  buildJudgePrompt(proposals, critiques) {
+    const proposalsSection = this.buildProposalsSection(proposals);
+    const critiquesSection = this.buildCritiquesSection(critiques);
+    return `You are a judge evaluating multiple proposals. Select the best proposal or synthesize the optimal response.
+
+${this.stageContext.taskContext}
+
+## Proposals
+${proposalsSection}
+
+## Critiques
+${critiquesSection}
+
+Evaluate each proposal against the critiques and provide the best possible response.
+${this.stageContext.outputFormat}`;
+  }
+  buildClosePrompt() {
+    return "Close this debate session.";
+  }
+  buildPersonaBlock(debaterIndex) {
+    const debater = this.options.debaters[debaterIndex];
+    if (!debater?.persona)
+      return "";
+    const { identity, lens } = PERSONA_FRAGMENTS[debater.persona];
+    return `
+
+## Your Role
+${identity}
+${lens}`;
+  }
+  buildProposalsSection(proposals) {
+    return proposals.map((p, i) => `### Proposal ${i + 1} (${this.buildDebaterLabel(p.debater)})
+${p.output}`).join(`
+
+`);
+  }
+  buildRebuttalsSection(rebuttals) {
+    if (rebuttals.length === 0)
+      return "";
+    return `
+
+## Previous Rebuttals
+${rebuttals.map((r, i) => `${i + 1}. ${r.output}`).join(`
+
+`)}`;
+  }
+  buildCritiquesSection(critiques) {
+    if (critiques.length === 0)
+      return "";
+    return critiques.map((c, i) => `### Critique ${i + 1} (${this.buildDebaterLabel(c.debater)})
+${c.output}`).join(`
+
+`);
+  }
+  buildDebaterLabel(debater) {
+    return debater.persona ? `${debater.agent} (${debater.persona})` : debater.agent;
+  }
+}
+var init_prompt_builder = __esm(() => {
+  init_personas();
+});
+
 // src/debate/session-stateful.ts
 async function runStatefulTurn(ctx, adapter, debater, prompt, roleKey, keepSessionOpen) {
   const modelTier = modelTierFromDebater(debater);
@@ -21774,7 +21978,9 @@ async function closeStatefulSession(ctx, adapter, debater, roleKey) {
 async function runStateful(ctx, prompt) {
   const logger = _debateSessionDeps.getSafeLogger();
   const config2 = ctx.stageConfig;
-  const debaters = config2.debaters ?? [];
+  const personaStage = ctx.stage === "plan" ? "plan" : "review";
+  const rawDebaters = config2.debaters ?? [];
+  const debaters = resolvePersonas(rawDebaters, personaStage, config2.autoPersona ?? false);
   let totalCostUsd = 0;
   const resolved = [];
   for (const debater of debaters) {
@@ -21871,8 +22077,9 @@ async function runStateful(ctx, prompt) {
   }
   let critiqueOutputs = [];
   if (config2.rounds > 1) {
-    const proposalOutputs2 = successfulProposals.map((s) => s.output);
-    const critiqueSettled = await allSettledBounded(successfulProposals.map((proposal, successfulIdx) => () => runStatefulTurn(ctx, proposal.adapter, proposal.debater, buildCritiquePrompt(prompt, proposalOutputs2, successfulIdx), proposal.roleKey ?? `debate-${ctx.stage}-${successfulIdx}`, false)), concurrencyLimit);
+    const proposals2 = successfulProposals.map((s) => ({ debater: s.debater, output: s.output }));
+    const critiqueBuilder = new DebatePromptBuilder({ taskContext: prompt, outputFormat: "", stage: ctx.stage }, { debaters: proposals2.map((p) => p.debater), sessionMode: ctx.stageConfig.sessionMode ?? "one-shot" });
+    const critiqueSettled = await allSettledBounded(successfulProposals.map((proposal, successfulIdx) => () => runStatefulTurn(ctx, proposal.adapter, proposal.debater, critiqueBuilder.buildCritiquePrompt(successfulIdx, proposals2), proposal.roleKey ?? `debate-${ctx.stage}-${successfulIdx}`, false)), concurrencyLimit);
     for (const r of critiqueSettled) {
       if (r.status === "fulfilled") {
         totalCostUsd += r.value.cost;
@@ -21908,11 +22115,13 @@ async function runStateful(ctx, prompt) {
   };
 }
 var init_session_stateful = __esm(() => {
+  init_personas();
+  init_prompt_builder();
   init_session_helpers();
 });
 
 // src/debate/session-hybrid.ts
-async function runRebuttalLoop(ctx, proposals, originalPrompt, sessionRolePrefix) {
+async function runRebuttalLoop(ctx, proposals, builder, sessionRolePrefix) {
   const logger = _debateSessionDeps.getSafeLogger();
   const config2 = ctx.stageConfig;
   const rebuttals = [];
@@ -21920,7 +22129,7 @@ async function runRebuttalLoop(ctx, proposals, originalPrompt, sessionRolePrefix
   const proposalList = proposals.map((s) => ({ debater: s.debater, output: s.output }));
   try {
     for (let round = 1;round <= config2.rounds; round++) {
-      const priorRebuttals = rebuttals.filter((r) => r.round < round).map((r) => r.output);
+      const priorRebuttals = rebuttals.filter((r) => r.round < round);
       for (let debaterIdx = 0;debaterIdx < proposals.length; debaterIdx++) {
         const proposal = proposals[debaterIdx];
         const sessionRole = `${sessionRolePrefix}-${debaterIdx}`;
@@ -21929,7 +22138,7 @@ async function runRebuttalLoop(ctx, proposals, originalPrompt, sessionRolePrefix
           round,
           debaterIndex: debaterIdx
         });
-        const rebuttalPrompt = buildRebuttalContext(originalPrompt, proposalList, priorRebuttals, debaterIdx);
+        const rebuttalPrompt = builder.buildRebuttalPrompt(debaterIdx, proposalList, priorRebuttals);
         try {
           const turnResult = await runStatefulTurn(ctx, proposal.adapter, proposal.debater, rebuttalPrompt, sessionRole, true);
           costUsd += turnResult.cost;
@@ -21959,7 +22168,9 @@ async function runRebuttalLoop(ctx, proposals, originalPrompt, sessionRolePrefix
 async function runHybrid(ctx, prompt) {
   const logger = _debateSessionDeps.getSafeLogger();
   const config2 = ctx.stageConfig;
-  const debaters = config2.debaters ?? [];
+  const personaStage = ctx.stage === "plan" ? "plan" : "review";
+  const rawDebaters = config2.debaters ?? [];
+  const debaters = resolvePersonas(rawDebaters, personaStage, config2.autoPersona ?? false);
   let totalCostUsd = 0;
   const resolved = [];
   for (const debater of debaters) {
@@ -22024,7 +22235,8 @@ async function runHybrid(ctx, prompt) {
   }
   const proposalOutputs = successfulProposals.map((s) => s.output);
   const proposalList = successfulProposals.map((s) => ({ debater: s.debater, output: s.output }));
-  const { rebuttals, costUsd: rebuttalCost } = await runRebuttalLoop(ctx, successfulProposals, prompt, "debate-hybrid");
+  const rebuttalBuilder = new DebatePromptBuilder({ taskContext: prompt, outputFormat: "", stage: ctx.stage }, { debaters: successfulProposals.map((s) => s.debater), sessionMode: "stateful" });
+  const { rebuttals, costUsd: rebuttalCost } = await runRebuttalLoop(ctx, successfulProposals, rebuttalBuilder, "debate-hybrid");
   totalCostUsd += rebuttalCost;
   const critiqueOutputs = rebuttals.map((r) => r.output);
   const fullResolverContext = ctx.resolverContextInput ? {
@@ -22046,6 +22258,8 @@ async function runHybrid(ctx, prompt) {
   };
 }
 var init_session_hybrid = __esm(() => {
+  init_personas();
+  init_prompt_builder();
   init_session_helpers();
   init_session_stateful();
 });
@@ -22054,7 +22268,9 @@ var init_session_hybrid = __esm(() => {
 async function runOneShot(ctx, prompt) {
   const logger = _debateSessionDeps.getSafeLogger();
   const config2 = ctx.stageConfig;
-  const debaters = config2.debaters ?? [];
+  const personaStage = ctx.stage === "plan" ? "plan" : "review";
+  const rawDebaters = config2.debaters ?? [];
+  const debaters = resolvePersonas(rawDebaters, personaStage, config2.autoPersona ?? false);
   let totalCostUsd = 0;
   const resolved = [];
   for (const debater of debaters) {
@@ -22157,8 +22373,9 @@ async function runOneShot(ctx, prompt) {
   }
   let critiqueOutputs = [];
   if (config2.rounds > 1) {
-    const proposalOutputs2 = successful.map((p) => p.output);
-    const critiqueSettled = await allSettledBounded(successful.map(({ debater, adapter }, i) => () => runComplete(adapter, buildCritiquePrompt(prompt, proposalOutputs2, i), {
+    const proposals2 = successful.map((p) => ({ debater: p.debater, output: p.output }));
+    const critiqueBuilder = new DebatePromptBuilder({ taskContext: prompt, outputFormat: "", stage: ctx.stage }, { debaters: proposals2.map((p) => p.debater), sessionMode: ctx.stageConfig.sessionMode ?? "one-shot" });
+    const critiqueSettled = await allSettledBounded(successful.map(({ debater, adapter }, i) => () => runComplete(adapter, critiqueBuilder.buildCritiquePrompt(i, proposals2), {
       model: resolveDebaterModel(debater, ctx.config),
       featureName: ctx.stage,
       config: ctx.config,
@@ -22201,15 +22418,18 @@ async function runOneShot(ctx, prompt) {
   };
 }
 var init_session_one_shot = __esm(() => {
+  init_personas();
+  init_prompt_builder();
   init_session_helpers();
 });
 
 // src/debate/session-plan.ts
 import { join as join10 } from "path";
-async function runPlan2(ctx, basePrompt, opts) {
+async function runPlan2(ctx, taskContext, outputFormat, opts) {
   const logger = _debateSessionDeps.getSafeLogger();
   const config2 = ctx.stageConfig;
-  const debaters = config2.debaters ?? [];
+  const rawDebaters = config2.debaters ?? [];
+  const debaters = resolvePersonas(rawDebaters, "plan", config2.autoPersona ?? false);
   let totalCostUsd = 0;
   const resolved = [];
   for (const debater of debaters) {
@@ -22227,14 +22447,15 @@ async function runPlan2(ctx, basePrompt, opts) {
   });
   const debate = ctx.config?.debate;
   const concurrencyLimit = debate?.maxConcurrentDebaters ?? 2;
-  const settled = await allSettledBounded(resolved.map(({ debater, adapter }, i) => async () => {
+  const proposalBuilder = new DebatePromptBuilder({ taskContext, outputFormat, stage: "plan" }, { debaters: resolved.map((r) => r.debater), sessionMode: ctx.stageConfig.sessionMode ?? "one-shot" });
+  const settled = await allSettledBounded(resolved.map(({ debater: rd, adapter }, i) => async () => {
     const tempOutputPath = join10(opts.outputDir, `prd-debate-${i}.json`);
-    const debaterPrompt = `${basePrompt}
+    const debaterPrompt = `${proposalBuilder.buildProposalPrompt(i)}
 
 Write the PRD JSON directly to this file path: ${tempOutputPath}
 Do NOT output the JSON to the conversation. Write the file, then reply with a brief confirmation.`;
-    const modelTier = modelTierFromDebater(debater);
-    const modelDef = resolveModelDefForDebater(debater, modelTier, ctx.config);
+    const modelTier = modelTierFromDebater(rd);
+    const modelDef = resolveModelDefForDebater(rd, modelTier, ctx.config);
     const planResult = await adapter.plan({
       prompt: debaterPrompt,
       workdir: opts.workdir,
@@ -22250,7 +22471,7 @@ Do NOT output the JSON to the conversation. Write the file, then reply with a br
       sessionRole: `plan-${i}`
     });
     const output = await _debateSessionDeps.readFile(tempOutputPath);
-    return { debater, adapter, output, cost: planResult.costUsd ?? 0 };
+    return { debater: rd, adapter, output, cost: planResult.costUsd ?? 0 };
   }), concurrencyLimit);
   const successful = [];
   for (let i = 0;i < settled.length; i++) {
@@ -22323,7 +22544,8 @@ Do NOT output the JSON to the conversation. Write the file, then reply with a br
       featureName: opts.feature,
       timeoutSeconds: opts.timeoutSeconds ?? 600
     };
-    const { rebuttals, costUsd } = await runRebuttalLoop(hybridCtx, successful, basePrompt, "plan-hybrid");
+    const rebuttalBuilder = new DebatePromptBuilder({ taskContext, outputFormat: "", stage: "plan" }, { debaters: successful.map((p) => p.debater), sessionMode });
+    const { rebuttals, costUsd } = await runRebuttalLoop(hybridCtx, successful, rebuttalBuilder, "plan-hybrid");
     critiqueOutputs = rebuttals.map((r) => r.output);
     rebuttalList = rebuttals;
     totalCostUsd += costUsd;
@@ -22331,7 +22553,8 @@ 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 outcome = await resolveOutcome(proposalOutputs, critiqueOutputs, ctx.stageConfig, ctx.config, ctx.storyId, resolverTimeoutMs, opts.workdir, opts.feature);
+  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 outcome = await resolveOutcome(proposalOutputs, critiqueOutputs, ctx.stageConfig, ctx.config, ctx.storyId, resolverTimeoutMs, opts.workdir, opts.feature, undefined, undefined, planSynthesisSuffix);
   const winningOutput = outcome.output ?? successful[0].output;
   const proposals = successful.map((p) => ({ debater: p.debater, output: p.output }));
   logger?.info("debate", "debate:result", {
@@ -22353,6 +22576,8 @@ Do NOT output the JSON to the conversation. Write the file, then reply with a br
   };
 }
 var init_session_plan = __esm(() => {
+  init_personas();
+  init_prompt_builder();
   init_session_helpers();
   init_session_hybrid();
 });
@@ -22438,13 +22663,13 @@ class DebateSession {
       resolverContextInput: this.resolverContextInput
     }, prompt);
   }
-  async runPlan(basePrompt, opts) {
+  async runPlan(taskContext, outputFormat, opts) {
     return runPlan2({
       storyId: this.storyId,
       stage: this.stage,
       stageConfig: this.stageConfig,
       config: this.config
-    }, basePrompt, opts);
+    }, taskContext, outputFormat, opts);
   }
 }
 var DEFAULT_TIMEOUT_SECONDS = 600;
@@ -22461,8 +22686,8 @@ var init_session = __esm(() => {
 var init_debate = __esm(() => {
   init_session();
   init_session_helpers();
-  init_resolvers();
-  init_session_helpers();
+  init_prompt_builder();
+  init_personas();
 });
 
 // src/interaction/bridge-builder.ts
@@ -23643,50 +23868,6 @@ var init_init = __esm(() => {
   init_webhook();
 });
 
-// src/utils/llm-json.ts
-function extractJsonFromMarkdown(text) {
-  const match = text.match(/```(?:json)?\s*\n([\s\S]*?)\n?\s*```/);
-  if (match) {
-    return match[1] ?? text;
-  }
-  return text;
-}
-function stripTrailingCommas(text) {
-  return text.replace(/,\s*([}\]])/g, "$1");
-}
-function extractJsonObject(text) {
-  const objStart = text.indexOf("{");
-  const arrStart = text.indexOf("[");
-  let start;
-  let closeChar;
-  if (objStart === -1 && arrStart === -1)
-    return null;
-  if (objStart === -1) {
-    start = arrStart;
-    closeChar = "]";
-  } else if (arrStart === -1) {
-    start = objStart;
-    closeChar = "}";
-  } else if (objStart < arrStart) {
-    start = objStart;
-    closeChar = "}";
-  } else {
-    start = arrStart;
-    closeChar = "]";
-  }
-  const end = text.lastIndexOf(closeChar);
-  if (end <= start)
-    return null;
-  return text.slice(start, end + 1);
-}
-function wrapJsonPrompt(prompt) {
-  return `IMPORTANT: Your entire response must be a single JSON object or array. Do not explain your reasoning. Do not use markdown formatting. Output ONLY the JSON.
-
-${prompt.trim()}
-
-YOUR RESPONSE MUST START WITH { OR [ AND END WITH } OR ]. No other text.`;
-}
-
 // src/prd/validate.ts
 function validateStoryId(id) {
   if (!id || id.length === 0) {
@@ -25792,7 +25973,7 @@ 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.
 - Output raw code only \u2014 no markdown fences, start directly with the language's import or package declaration
-- **Path anchor (CRITICAL)**: This test file will be saved at \`<repo-root>/.nax/features/${featureName}/${resolvedTestPath}\` and will ALWAYS run from the repo root. The repo root is exactly 4 \`../\` levels above \`__dirname\`: \`join(__dirname, '..', '..', '..', '..')\`. For monorepo projects, navigate into packages from root (e.g. \`join(root, 'apps/api/src')\`).`;
+- **Path anchor (CRITICAL)**: This test file will be saved at \`<repo-root>/.nax/features/${featureName}/${resolvedTestPath}\` and will ALWAYS run from the repo root. The repo root is exactly 3 \`../\` levels above \`__dirname\`: \`join(__dirname, '..', '..', '..')\`. For monorepo projects, navigate into packages from root (e.g. \`join(root, 'apps/api/src')\`).`;
 }
 async function generateAcceptanceTests(adapter, options) {
   const logger = getLogger();
@@ -26950,6 +27131,7 @@ function buildReviewPrompt(diff, story, _semanticConfig) {
     "## 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(`
 `);
@@ -26971,7 +27153,7 @@ function buildReReviewPrompt(updatedDiff, previousFindings) {
   ].join(`
 `);
 }
-function buildProposalsSection(proposals) {
+function buildProposalsSection2(proposals) {
   return proposals.map((p) => `### ${p.debater}
 ${p.output}`).join(`
 
@@ -27015,7 +27197,7 @@ function buildDebateResolverPrompt(proposals, critiques, diff, story, _semanticC
 `);
   const framing = buildResolverFraming(resolverContext);
   const voteTally = buildVoteTallyLine(resolverContext);
-  const proposalsSection = buildProposalsSection(proposals);
+  const proposalsSection = buildProposalsSection2(proposals);
   const critiquesSection = buildCritiquesSection(critiques);
   return [
     framing,
@@ -27041,7 +27223,7 @@ function buildDebateReReviewPrompt(proposals, critiques, updatedDiff, previousFi
   const framing = buildResolverFraming(resolverContext);
   const findingsList = previousFindings.length > 0 ? previousFindings.map((f) => `- ${f.ruleId}: ${f.message}`).join(`
 `) : "(none)";
-  const proposalsSection = buildProposalsSection(proposals);
+  const proposalsSection = buildProposalsSection2(proposals);
   const critiquesSection = buildCritiquesSection(critiques);
   return [
     `${framing} This is a re-review after implementer changes.`,
@@ -27064,12 +27246,10 @@ function buildDebateReReviewPrompt(proposals, critiques, updatedDiff, previousFi
 
 // src/review/dialogue.ts
 function extractDeltaSummary(rawOutput, previousFindings, newFindings) {
-  try {
-    const parsed = JSON.parse(rawOutput);
-    if (typeof parsed.deltaSummary === "string" && parsed.deltaSummary.length > 0) {
-      return parsed.deltaSummary;
-    }
-  } catch {}
+  const parsed = tryParseLLMJson(rawOutput);
+  if (parsed && typeof parsed.deltaSummary === "string" && parsed.deltaSummary.length > 0) {
+    return parsed.deltaSummary;
+  }
   const newIds = new Set(newFindings.map((f) => f.ruleId));
   const prevIds = new Set(previousFindings.map((f) => f.ruleId));
   const resolved = previousFindings.filter((f) => !newIds.has(f.ruleId));
@@ -27109,7 +27289,7 @@ function compactHistory(history) {
 function parseReviewResponse(output) {
   let parsed;
   try {
-    parsed = JSON.parse(output);
+    parsed = parseLLMJson(output);
   } catch {
     throw new NaxError("[dialogue] Failed to parse reviewer JSON response", "REVIEWER_PARSE_FAILED", {
       stage: "review",
@@ -27688,23 +27868,11 @@ function validateLLMShape(parsed) {
   return { passed: obj.passed, findings: obj.findings };
 }
 function parseLLMResponse(raw) {
-  const text = raw.trim();
   try {
-    return validateLLMShape(JSON.parse(text));
-  } catch {}
-  const fromFence = extractJsonFromMarkdown(text);
-  if (fromFence !== text) {
-    try {
-      return validateLLMShape(JSON.parse(stripTrailingCommas(fromFence)));
-    } catch {}
-  }
-  const bareJson = extractJsonObject(text);
-  if (bareJson) {
-    try {
-      return validateLLMShape(JSON.parse(stripTrailingCommas(bareJson)));
-    } catch {}
+    return validateLLMShape(tryParseLLMJson(raw));
+  } catch {
+    return null;
   }
-  return null;
 }
 function formatFindings(findings) {
   return findings.map((f) => `[${f.severity}] ${f.file}:${f.line} \u2014 ${f.issue}
@@ -28679,6 +28847,7 @@ async function runAgentRectification(ctx, lintFixCmd, formatFixCmd, effectiveWor
         config: ctx.config,
         projectDir: ctx.projectDir,
         maxInteractionTurns: ctx.config.agent?.maxInteractionTurns,
+        featureName: ctx.prd.feature,
         storyId: ctx.story.id,
         sessionRole: "implementer"
       });
@@ -30456,61 +30625,34 @@ var init_executor = __esm(() => {
 });
 
 // src/verification/parser.ts
-function parseBunTestOutput(output) {
-  if (isJestLikeOutput(output)) {
-    return parseJestOutput(output);
-  }
-  return parseBunOutput(output);
-}
-function isJestLikeOutput(output) {
-  return /^\s*Tests:\s+\d+/m.test(output) || /^\s*Test Files\s+\d+/m.test(output);
+function detectFramework(output) {
+  if (/^\s*Test Files\s+\d+/m.test(output))
+    return "vitest";
+  if (/^\s*Tests:\s+\d+/m.test(output))
+    return "jest";
+  if (/={3,}\s+\d+\s+(?:failed|passed).*in\s+[\d.]+s\s*={3,}/m.test(output))
+    return "pytest";
+  if (/^--- (?:FAIL|PASS):/m.test(output) || /^(?:ok|FAIL)\s+\t/m.test(output))
+    return "go";
+  if (/^\(fail\)\s/m.test(output) || /^bun test/m.test(output) || /[\u2713\u2714\u2717\u2718]/m.test(output))
+    return "bun";
+  return "unknown";
 }
-function parseJestOutput(output) {
-  const failures = [];
-  let passed = 0;
-  let failed = 0;
-  const summaryMatches = Array.from(output.matchAll(/^\s*Tests:\s+(.*)/gm));
-  if (summaryMatches.length > 0) {
-    const summaryLine = summaryMatches[summaryMatches.length - 1][1];
-    const failedMatch = summaryLine.match(/(\d+)\s+failed/);
-    const passedMatch = summaryLine.match(/(\d+)\s+passed/);
-    if (failedMatch)
-      failed = Number.parseInt(failedMatch[1], 10);
-    if (passedMatch)
-      passed = Number.parseInt(passedMatch[1], 10);
-  }
-  let currentFile = "unknown";
-  const lines = output.split(`
-`);
-  for (let i = 0;i < lines.length; i++) {
-    const line = lines[i];
-    const fileMatch = line.match(/^\s*(?:FAIL|PASS)\s+(\S+\.[jt]sx?)/);
-    if (fileMatch) {
-      currentFile = fileMatch[1];
-      continue;
-    }
-    const bulletMatch = line.match(/^\s+\u25CF\s+(.+)$/);
-    if (bulletMatch) {
-      const testName = bulletMatch[1].trim();
-      let error48 = "";
-      for (let j = i + 1;j < lines.length && j < i + 10; j++) {
-        const next = lines[j].trim();
-        if (!next)
-          continue;
-        if (next.startsWith("\u25CF") || /^(?:FAIL|PASS)\s/.test(next))
-          break;
-        error48 = next;
-        break;
-      }
-      failures.push({
-        file: currentFile,
-        testName,
-        error: error48 || "Unknown error",
-        stackTrace: []
-      });
-    }
+function parseTestOutput(output) {
+  const framework = detectFramework(output);
+  switch (framework) {
+    case "bun":
+      return parseBunOutput(output);
+    case "jest":
+    case "vitest":
+      return parseJestOutput(output);
+    case "pytest":
+      return parsePytestOutput(output);
+    case "go":
+      return parseGoTestOutput(output);
+    default:
+      return parseCommonOutput(output);
   }
-  return { passed, failed, failures };
 }
 function parseBunOutput(output) {
   const lines = output.split(`
@@ -30572,6 +30714,122 @@ function parseBunOutput(output) {
   }
   return { passed, failed, failures };
 }
+function parseJestOutput(output) {
+  const failures = [];
+  let passed = 0;
+  let failed = 0;
+  const summaryMatches = Array.from(output.matchAll(/^\s*Tests:\s+(.*)/gm));
+  if (summaryMatches.length > 0) {
+    const summaryLine = summaryMatches[summaryMatches.length - 1][1];
+    const failedMatch = summaryLine.match(/(\d+)\s+failed/);
+    const passedMatch = summaryLine.match(/(\d+)\s+passed/);
+    if (failedMatch)
+      failed = Number.parseInt(failedMatch[1], 10);
+    if (passedMatch)
+      passed = Number.parseInt(passedMatch[1], 10);
+  }
+  let currentFile = "unknown";
+  const lines = output.split(`
+`);
+  for (let i = 0;i < lines.length; i++) {
+    const line = lines[i];
+    const fileMatch = line.match(/^\s*(?:FAIL|PASS)\s+(\S+\.[jt]sx?)/);
+    if (fileMatch) {
+      currentFile = fileMatch[1];
+      continue;
+    }
+    const bulletMatch = line.match(/^\s+\u25CF\s+(.+)$/);
+    if (bulletMatch) {
+      const testName = bulletMatch[1].trim();
+      let error48 = "";
+      for (let j = i + 1;j < lines.length && j < i + 10; j++) {
+        const next = lines[j].trim();
+        if (!next)
+          continue;
+        if (next.startsWith("\u25CF") || /^(?:FAIL|PASS)\s/.test(next))
+          break;
+        error48 = next;
+        break;
+      }
+      failures.push({
+        file: currentFile,
+        testName,
+        error: error48 || "Unknown error",
+        stackTrace: []
+      });
+    }
+  }
+  return { passed, failed, failures };
+}
+function parsePytestOutput(output) {
+  const common = parseCommonOutput(output);
+  const failures = [];
+  for (const line of output.split(`
+`)) {
+    const m = line.match(/^FAILED\s+(\S+)(?:\s+-\s+(.*))?$/);
+    if (m) {
+      const [, location, reason] = m;
+      const parts = location.split("::");
+      failures.push({
+        file: parts[0] ?? location,
+        testName: parts.slice(1).join(" > ") || location,
+        error: reason?.trim() || "Unknown error",
+        stackTrace: []
+      });
+    }
+  }
+  return {
+    passed: common.passed,
+    failed: common.failed,
+    failures: failures.length > 0 ? failures : common.failures
+  };
+}
+function parseGoTestOutput(output) {
+  const common = parseCommonOutput(output);
+  const failures = [];
+  for (const line of output.split(`
+`)) {
+    const m = line.match(/^--- FAIL:\s+(\S+)\s+\([\d.]+s\)/);
+    if (m) {
+      failures.push({
+        file: "unknown",
+        testName: m[1],
+        error: "Unknown error",
+        stackTrace: []
+      });
+    }
+  }
+  return {
+    passed: common.passed,
+    failed: common.failed,
+    failures: failures.length > 0 ? failures : common.failures
+  };
+}
+function parseCommonOutput(output) {
+  let passed = 0;
+  let failed = 0;
+  const patterns = [
+    /(\d+)\s+pass(?:ed)?(?:,\s*|\s+)(\d+)\s+fail/i,
+    /Tests:\s+(\d+)\s+passed,\s+(\d+)\s+failed/i,
+    /(\d+)\s+pass/i
+  ];
+  for (const pattern of patterns) {
+    const matches = Array.from(output.matchAll(new RegExp(pattern, "gi")));
+    if (matches.length > 0) {
+      const last = matches[matches.length - 1];
+      passed = Number.parseInt(last[1], 10);
+      failed = last[2] ? Number.parseInt(last[2], 10) : 0;
+      break;
+    }
+  }
+  if (failed === 0) {
+    const failMatches = Array.from(output.matchAll(/(\d+)\s+fail/gi));
+    if (failMatches.length > 0) {
+      failed = Number.parseInt(failMatches[failMatches.length - 1][1], 10);
+    }
+  }
+  return { passed, failed, failures: [] };
+}
 function formatFailureSummary(failures, maxChars = 2000) {
   if (failures.length === 0) {
     return "No test failures";
@@ -30585,48 +30843,24 @@ function formatFailureSummary(failures, maxChars = 2000) {
     const errorLine = `   Error: ${failure.error}`;
     const stackLine = failure.stackTrace.length > 0 ? `   ${failure.stackTrace[0]}` : "";
     const blockLines = [header, errorLine];
-    if (stackLine) {
+    if (stackLine)
       blockLines.push(stackLine);
-    }
     blockLines.push("");
     const block = blockLines.join(`
 `);
-    const blockLength = block.length;
-    if (totalChars + blockLength > maxChars && lines.length > 0) {
-      const remaining = failures.length - i;
+    if (totalChars + block.length > maxChars && lines.length > 0) {
       lines.push(`
-... and ${remaining} more failure(s) (truncated)`);
+... and ${failures.length - i} more failure(s) (truncated)`);
       break;
     }
     lines.push(...blockLines);
-    totalChars += blockLength;
+    totalChars += block.length;
   }
   return lines.join(`
 `).trim();
 }
-function parseTestOutput(output, exitCode) {
-  const patterns = [
-    /(\d+)\s+pass(?:ed)?(?:,\s+|\s+)(\d+)\s+fail/i,
-    /Tests:\s+(\d+)\s+passed,\s+(\d+)\s+failed/i,
-    /(\d+)\s+pass/i
-  ];
-  let passCount = 0;
-  let failCount = 0;
-  for (const pattern of patterns) {
-    const matches = Array.from(output.matchAll(new RegExp(pattern, "gi")));
-    if (matches.length > 0) {
-      const lastMatch = matches[matches.length - 1];
-      passCount = Number.parseInt(lastMatch[1], 10);
-      failCount = lastMatch[2] ? Number.parseInt(lastMatch[2], 10) : 0;
-      break;
-    }
-  }
-  if (failCount === 0) {
-    const failMatches = Array.from(output.matchAll(/(\d+)\s+fail/gi));
-    if (failMatches.length > 0) {
-      failCount = Number.parseInt(failMatches[failMatches.length - 1][1], 10);
-    }
-  }
+function analyzeTestExitCode(output, exitCode) {
+  const { passed: passCount, failed: failCount } = parseCommonOutput(output);
   const allTestsPassed = passCount > 0 && failCount === 0;
   const isEnvironmentalFailure = allTestsPassed && exitCode !== 0;
   const result = {
@@ -30697,7 +30931,7 @@ async function runVerificationCore(options) {
   }
   const exitCode = execution.exitCode ?? 1;
   if (exitCode !== 0 && execution.output) {
-    const analysis = parseTestOutput(execution.output, exitCode);
+    const analysis = analyzeTestExitCode(execution.output, exitCode);
     if (analysis.isEnvironmentalFailure) {
       return {
         status: "ENVIRONMENTAL_FAILURE",
@@ -30990,7 +31224,7 @@ async function runFullSuiteGate(story, config2, workdir, agent, implementerTier,
   });
   const fullSuitePassed = fullSuiteResult.success && fullSuiteResult.exitCode === 0;
   if (!fullSuitePassed && fullSuiteResult.output) {
-    const testSummary = _rectificationGateDeps.parseBunTestOutput(fullSuiteResult.output);
+    const testSummary = _rectificationGateDeps.parseTestOutput(fullSuiteResult.output);
     if (testSummary.failed > 0) {
       return await runRectificationLoop(story, config2, workdir, agent, implementerTier, contextMarkdown, lite, logger, testSummary, rectificationConfig, testCmd, fullSuiteTimeout, featureName, projectDir);
     }
@@ -31124,7 +31358,7 @@ async function runRectificationLoop(story, config2, workdir, agent, implementerT
         return true;
       }
       if (retryFullSuite.output) {
-        const newTestSummary = _rectificationGateDeps.parseBunTestOutput(retryFullSuite.output);
+        const newTestSummary = _rectificationGateDeps.parseTestOutput(retryFullSuite.output);
         state.currentFailures = newTestSummary.failed;
         testSummary.failures = newTestSummary.failures;
         testSummary.failed = newTestSummary.failed;
@@ -31172,7 +31406,7 @@ var init_rectification_gate = __esm(() => {
   init_prompts();
   _rectificationGateDeps = {
     executeWithTimeout,
-    parseBunTestOutput,
+    parseTestOutput,
     shouldRetryRectification
   };
 });
@@ -33142,9 +33376,6 @@ function calculateMaxIterations(tierOrder) {
   return tierOrder.reduce((sum, t) => sum + t.attempts, 0);
 }
 
-// src/execution/test-output-parser.ts
-var init_test_output_parser = () => {};
-
 // src/verification/rectification-loop.ts
 async function _defaultRunDebate(storyId, stageConfig, prompt, config2) {
   const logger = getSafeLogger();
@@ -33197,7 +33428,7 @@ async function runRectificationLoop2(opts) {
   } = opts;
   const logger = getSafeLogger();
   const rectificationConfig = config2.execution.rectification;
-  const testSummary = parseBunTestOutput(testOutput);
+  const testSummary = parseTestOutput(testOutput);
   const label = promptPrefix ? "regression rectification" : "rectification";
   const rectificationState = {
     attempt: 0,
@@ -33332,13 +33563,13 @@ ${rectificationPrompt}`;
         return true;
       }
       if (retryVerification.output) {
-        const newTestSummary = parseBunTestOutput(retryVerification.output);
+        const newTestSummary = parseTestOutput(retryVerification.output);
         state.currentFailures = newTestSummary.failed;
         state.lastExitCode = retryVerification.status === "SUCCESS" ? 0 : 1;
         testSummary.failures = newTestSummary.failures;
         testSummary.failed = newTestSummary.failed;
         testSummary.passed = newTestSummary.passed;
-        if (newTestSummary.failed === 0) {
+        if (newTestSummary.failed === 0 && (retryVerification.status === "SUCCESS" || newTestSummary.passed > 0)) {
           state.lastExitCode = 0;
           logger?.info("rectification", `[OK] ${label} succeeded after parsing retry output`, {
             storyId: story.id,
@@ -33478,7 +33709,6 @@ var init_rectification_loop = __esm(() => {
   init_cost();
   init_registry();
   init_config();
-  init_test_output_parser();
   init_logger2();
   init_prd();
   init_rectification();
@@ -33744,7 +33974,7 @@ class RegressionStrategy {
     });
     const durationMs = Date.now() - start;
     if (result.success) {
-      const parsed2 = result.output ? parseBunTestOutput(result.output) : { passed: 0, failed: 0, failures: [] };
+      const parsed2 = result.output ? parseTestOutput(result.output) : { passed: 0, failed: 0, failures: [] };
       return makePassResult(ctx.storyId, "regression", {
         rawOutput: result.output,
         passCount: parsed2.passed,
@@ -33760,7 +33990,7 @@ class RegressionStrategy {
     if (result.status === "TIMEOUT") {
       return makeFailResult(ctx.storyId, "regression", "TIMEOUT", { rawOutput: result.output, durationMs });
     }
-    const parsed = result.output ? parseBunTestOutput(result.output) : { passed: 0, failed: 0, failures: [] };
+    const parsed = result.output ? parseTestOutput(result.output) : { passed: 0, failed: 0, failures: [] };
     return makeFailResult(ctx.storyId, "regression", "TEST_FAILURE", {
       rawOutput: result.output,
       passCount: parsed.passed,
@@ -33984,7 +34214,7 @@ class ScopedStrategy {
     });
     const durationMs = Date.now() - start;
     if (result.success) {
-      const parsed2 = result.output ? parseBunTestOutput(result.output) : { passed: 0, failed: 0, failures: [] };
+      const parsed2 = result.output ? parseTestOutput(result.output) : { passed: 0, failed: 0, failures: [] };
       return makePassResult(ctx.storyId, "scoped", {
         rawOutput: result.output,
         passCount: parsed2.passed,
@@ -34000,7 +34230,7 @@ class ScopedStrategy {
         scopeTestFallback
       });
     }
-    const parsed = result.output ? parseBunTestOutput(result.output) : { passed: 0, failed: 0, failures: [] };
+    const parsed = result.output ? parseTestOutput(result.output) : { passed: 0, failed: 0, failures: [] };
     return makeFailResult(ctx.storyId, "scoped", "TEST_FAILURE", {
       rawOutput: result.output,
       passCount: parsed.passed,
@@ -34334,12 +34564,11 @@ function stripCodeFences(text) {
   return trimmed;
 }
 function parseRoutingResponse(output, story, config2) {
-  const jsonText = extractJsonFromMarkdown(output.trim());
-  const parsed = JSON.parse(jsonText);
+  const parsed = parseLLMJson(output);
   return validateRoutingDecision(parsed, config2, story);
 }
 function parseBatchResponse(output, stories, config2) {
-  const parsed = JSON.parse(extractJsonFromMarkdown(output.trim()));
+  const parsed = parseLLMJson(output);
   if (!Array.isArray(parsed)) {
     throw new Error("Batch LLM response must be a JSON array");
   }
@@ -36133,7 +36362,7 @@ var package_default;
 var init_package = __esm(() => {
   package_default = {
     name: "@nathapp/nax",
-    version: "0.59.1",
+    version: "0.59.3",
     description: "AI Coding Agent Orchestrator \u2014 loops until done",
     type: "module",
     bin: {
@@ -36213,8 +36442,8 @@ var init_version = __esm(() => {
   NAX_VERSION = package_default.version;
   NAX_COMMIT = (() => {
     try {
-      if (/^[0-9a-f]{6,10}$/.test("b8492d03"))
-        return "b8492d03";
+      if (/^[0-9a-f]{6,10}$/.test("0c763972"))
+        return "0c763972";
     } catch {}
     try {
       const result = Bun.spawnSync(["git", "rev-parse", "--short", "HEAD"], {
@@ -36877,28 +37106,17 @@ function parseDiagnosisResult(output) {
   if (!output || output.trim() === "") {
     return null;
   }
-  try {
-    const cleaned = output.trim();
-    let jsonStr = cleaned;
-    const firstBrace = cleaned.indexOf("{");
-    const lastBrace = cleaned.lastIndexOf("}");
-    if (firstBrace !== -1 && lastBrace !== -1 && lastBrace > firstBrace) {
-      jsonStr = cleaned.slice(firstBrace, lastBrace + 1);
-    }
-    const parsed = JSON.parse(jsonStr);
-    if (typeof parsed.verdict === "string" && typeof parsed.reasoning === "string" && typeof parsed.confidence === "number") {
-      return {
-        verdict: parsed.verdict,
-        reasoning: parsed.reasoning,
-        confidence: parsed.confidence,
-        testIssues: parsed.testIssues,
-        sourceIssues: parsed.sourceIssues
-      };
-    }
-    return null;
-  } catch {
-    return null;
+  const parsed = tryParseLLMJson(output);
+  if (parsed && typeof parsed.verdict === "string" && typeof parsed.reasoning === "string" && typeof parsed.confidence === "number") {
+    return {
+      verdict: parsed.verdict,
+      reasoning: parsed.reasoning,
+      confidence: parsed.confidence,
+      testIssues: Array.isArray(parsed.testIssues) ? parsed.testIssues : undefined,
+      sourceIssues: Array.isArray(parsed.sourceIssues) ? parsed.sourceIssues : undefined
+    };
   }
+  return null;
 }
 var MAX_SOURCE_FILES = 5, MAX_FILE_LINES = 500, MAX_TEST_OUTPUT_CHARS = 2000;
 var init_fix_diagnosis = __esm(() => {
@@ -37096,6 +37314,12 @@ async function regenerateAcceptanceTest(testPath, acceptanceContext) {
   logger?.info("acceptance", `Backed up acceptance test -> ${bakPath}`);
   const { unlink: unlink3 } = await import("fs/promises");
   await unlink3(testPath);
+  if (acceptanceContext.featureDir) {
+    const metaPath = path15.join(acceptanceContext.featureDir, "acceptance-meta.json");
+    try {
+      await unlink3(metaPath);
+    } catch {}
+  }
   let implementationContext;
   const storyGitRef = acceptanceContext.storyGitRef;
   const workdir = acceptanceContext.workdir;
@@ -37736,7 +37960,7 @@ async function runDeferredRegression(options) {
       affectedStories: []
     };
   }
-  const testSummary = _regressionDeps.parseBunTestOutput(fullSuiteResult.output);
+  const testSummary = _regressionDeps.parseTestOutput(fullSuiteResult.output);
   if (testSummary.failed === 0 && testSummary.passed === 0) {
     logger?.warn("regression", "No test results parsed from output \u2014 test runner likely crashed or errored (not a regression, accepting as pass)", { output: fullSuiteResult.output.slice(0, 500) });
     return {
@@ -37882,7 +38106,7 @@ var init_run_regression = __esm(() => {
   _regressionDeps = {
     runVerification: fullSuite,
     runRectificationLoop: runRectificationLoop2,
-    parseBunTestOutput,
+    parseTestOutput,
     reverseMapTestToSource
   };
 });
@@ -38863,6 +39087,11 @@ async function handlePipelineSuccess(ctx, pipelineResult) {
         const diffSummary = await captureDiffSummary(ctx.workdir, ctx.storyGitRef, completedStory.workdir);
         if (diffSummary) {
           completedStory.diffSummary = diffSummary;
+        } else {
+          logger?.debug("context-chain", "No diff summary captured (agent may not have committed yet)", {
+            storyId: completedStory.id,
+            storyGitRef: ctx.storyGitRef
+          });
         }
       } catch {}
     }
@@ -72799,13 +73028,16 @@ function validateStory(raw, index, allIds) {
     throw new Error(`[schema] story[${index}].routing.complexity "${rawComplexity}" is invalid. Valid values: ${VALID_COMPLEXITY.join(", ")}`);
   }
   const rawTestStrategy = routing.testStrategy ?? s.testStrategy;
-  const testStrategy = resolveTestStrategy(typeof rawTestStrategy === "string" ? rawTestStrategy : undefined);
+  let testStrategy = resolveTestStrategy(typeof rawTestStrategy === "string" ? rawTestStrategy : undefined);
   const rawJustification = routing.noTestJustification ?? s.noTestJustification;
   if (testStrategy === "no-test") {
     if (!rawJustification || typeof rawJustification !== "string" || rawJustification.trim() === "") {
       throw new Error(`[schema] story[${index}].routing.noTestJustification is required when testStrategy is "no-test"`);
     }
   }
+  if (testStrategy !== "no-test" && typeof rawJustification === "string" && rawJustification.trim() !== "") {
+    testStrategy = "no-test";
+  }
   const noTestJustification = typeof rawJustification === "string" && rawJustification.trim() !== "" ? rawJustification.trim() : undefined;
   const rawDeps = s.dependencies;
   const dependencies = Array.isArray(rawDeps) ? rawDeps : [];
@@ -72965,7 +73197,7 @@ async function planCommand(workdir, config2, options) {
   let rawResponse;
   const debateEnabled = config2?.debate?.enabled && config2?.debate?.stages?.plan?.enabled;
   if (debateEnabled) {
-    const basePrompt = buildPlanningPrompt(specContent, codebaseContext, undefined, relativePackages, packageDetails, config2?.project);
+    const { taskContext: planTaskContext, outputFormat: planOutputFormat } = buildPlanningPrompt(specContent, codebaseContext, undefined, relativePackages, packageDetails, config2?.project);
     const resolvedPerm = resolvePermissions(config2, "plan");
     const planStageConfig = config2?.debate?.stages.plan;
     const debateSession = _planDeps.createDebateSession({
@@ -72982,7 +73214,7 @@ async function planCommand(workdir, config2, options) {
       rounds: planStageConfig.rounds,
       feature: options.feature
     });
-    const debateResult = await debateSession.runPlan(basePrompt, {
+    const debateResult = await debateSession.runPlan(planTaskContext, planOutputFormat, {
       workdir,
       feature: options.feature,
       outputDir,
@@ -73001,7 +73233,10 @@ async function planCommand(workdir, config2, options) {
     }
   } else if (options.auto) {
     const isAcp = config2?.agent?.protocol === "acp";
-    const prompt = buildPlanningPrompt(specContent, codebaseContext, isAcp ? outputPath : undefined, relativePackages, packageDetails, config2?.project);
+    const { taskContext: autoTaskCtx, outputFormat: autoOutputFmt } = buildPlanningPrompt(specContent, codebaseContext, isAcp ? outputPath : undefined, relativePackages, packageDetails, config2?.project);
+    const prompt = `${autoTaskCtx}
+
+${autoOutputFmt}`;
     const adapter = _planDeps.getAgent(agentName, config2);
     if (!adapter)
       throw new Error(`[plan] No agent adapter found for '${agentName}'`);
@@ -73087,7 +73322,10 @@ async function planCommand(workdir, config2, options) {
     rawResponse = await runInteractivePlan();
   }
   async function runInteractivePlan() {
-    const prompt = buildPlanningPrompt(specContent, codebaseContext, outputPath, relativePackages, packageDetails, config2?.project);
+    const { taskContext: interactiveTaskCtx, outputFormat: interactiveOutputFmt } = buildPlanningPrompt(specContent, codebaseContext, outputPath, relativePackages, packageDetails, config2?.project);
+    const prompt = `${interactiveTaskCtx}
+
+${interactiveOutputFmt}`;
     const adapter = _planDeps.getAgent(agentName, config2);
     if (!adapter)
       throw new Error(`[plan] No agent adapter found for '${agentName}'`);
@@ -73294,7 +73532,7 @@ ${packageDetailsSection}
 For each user story, set the "workdir" field to the relevant package path (e.g. "packages/api"). Stories that span the root should omit "workdir".` : "";
   const workdirField = isMonorepo ? `
       "workdir": "string \u2014 optional, relative path to package (e.g. \\"packages/api\\"). Omit for root-level stories.",` : "";
-  return `You are a senior software architect generating a product requirements document (PRD) as JSON.
+  const taskContext = `You are a senior software architect generating a product requirements document (PRD) as JSON.
 
 ## Step 1: Understand the Spec
 
@@ -73322,6 +73560,8 @@ If this is a greenfield project (empty or minimal codebase):
 
 Record ALL findings in the "analysis" field of the output JSON. This analysis is provided to every implementation agent as context \u2014 be thorough.
 
+**Important:** The codebase context below contains file names and structure only \u2014 no file content. Do NOT assert specific line numbers. The implementer will read the actual files via contextFiles.
+
 ## Codebase Context
 
 ${codebaseContext}${monorepoHint}
@@ -73338,9 +73578,8 @@ For each story, set "contextFiles" to the key source files the agent should read
 
 ${COMPLEXITY_GUIDE}
 
-${TEST_STRATEGY_GUIDE}
-
-## Output Schema
+${TEST_STRATEGY_GUIDE}`;
+  const outputFormat = `## Output Schema
 
 Generate a JSON object with this exact structure (no markdown, no explanation \u2014 JSON only):
 
@@ -73376,6 +73615,7 @@ Generate a JSON object with this exact structure (no markdown, no explanation \u
 
 ${outputFilePath ? `Write the PRD JSON directly to this file path: ${outputFilePath}
 Do NOT output the JSON to the conversation. Write the file, then reply with a brief confirmation.` : "Output ONLY the JSON object. Do not wrap in markdown code blocks."}`;
+  return { taskContext, outputFormat };
 }
 async function planDecomposeCommand(workdir, config2, options) {
   const prdPath = join11(workdir, ".nax", "features", options.feature, "prd.json");