@jaimevalasek/aioson 1.7.2 → 1.9.0

package/src/commands/squad-autorun.js

@@ -1,1220 +1,1220 @@
-'use strict';
-
-/**
- * aioson squad:autorun — Autonomous squad execution
- *
- * Given a high-level goal, this command:
- *   1. Decomposes the goal into a task plan (heuristic or structured)
- *   2. Activates the intra-squad bus for real-time communication
- *   3. Runs each task through the worker-runner with optional reflection
- *   4. Coordinator monitors the bus for blocks or feedback
- *   5. Reports the final session summary
- *
- * Phase 1 additions:
- *   - Gap Closure Loop: failed tasks are retried up to 3x with failure context
- *   - Budget Gating: halts before a task if session token budget is exceeded
- *   - Heartbeat Protocol: posts bus heartbeat every 30s during long tasks
- *   - Anti-Analysis-Loop Guard: warns executor if N+ status without result
- *   - Squad STATE.md: cross-session memory updated at start/end
- *
- * Usage:
- *   aioson squad:autorun . --squad=content-team --goal="Create 3 podcast episodes"
- *   aioson squad:autorun . --squad=content-team --goal="..." --reflect --bus --mode=structured
- *   aioson squad:autorun . --squad=content-team --plan=SESSION_ID  (resume from saved plan)
- *   aioson squad:autorun . --squad=content-team --plan=SESSION_ID --dry-run
- *
- * Flags:
- *   --goal            High-level objective (required unless --plan is given)
- *   --reflect         Run reflection after each task (default: false)
- *   --bus             Enable intra-bus for inter-executor communication (default: true)
- *   --mode            Decomposition mode: heuristic (default) | structured
- *   --plan            Resume from existing session plan (session ID)
- *   --dry-run         Show plan without executing
- *   --sequential      Force sequential execution even for independent tasks (default: false)
- *   --timeout         Per-task timeout in seconds (default: 120)
- *   --max-retries     Gap closure max retry attempts (default: 3)
- *   --no-gap-closure  Disable automatic retry on task failure
- */
-
-const fs = require('node:fs/promises');
-const path = require('node:path');
-const { randomUUID } = require('node:crypto');
-const {
-  decompose,
-  getReadyTasks,
-  isPlanComplete,
-  updateTaskStatus,
-  loadPlan,
-  formatPlan
-} = require('../squad/task-decomposer');
-const bus = require('../squad/intra-bus');
-const { reflect, formatReport } = require('../squad/reflection');
-const { runWorker, listWorkers } = require('../worker-runner');
-const stateManager = require('../squad/state-manager');
-const { getUnresolvedBlocks } = require('../squad/intra-bus');
-const interSquadEvents = require('../squad/inter-squad-events');
-const { runHook, HOOK_DENY } = require('../lib/hook-protocol');
-const { extractLearnings, persistAgentMemory } = require('../squad/learning-extractor');
-const { validateBrief, autoFixBrief } = require('../squad/brief-validator');
-const { resolveEngine, translateToTeamConfig, writeTeamConfig } = require('../squad/agent-teams-adapter');
-
-const STATUS_ICON = {
-  pending: '○',
-  in_progress: '●',
-  completed: '✓',
-  failed: '✗',
-  escalated: '⚠',
-  skipped: '–'
-};
-
-function icon(status) {
-  return STATUS_ICON[status] || '?';
-}
-
-function nowIso() { return new Date().toISOString(); }
-
-function sleep(ms) { return new Promise((r) => setTimeout(r, ms)); }
-
-// ─── Budget helpers ───────────────────────────────────────────────────────────
-
-/**
- * Load token budget from squad manifest (optional field).
- * Returns Infinity if not configured.
- */
-async function loadBudget(projectDir, squadSlug) {
-  const manifestPath = path.join(
-    projectDir, '.aioson', 'squads', squadSlug, 'squad.manifest.json'
-  );
-  try {
-    const manifest = JSON.parse(await fs.readFile(manifestPath, 'utf8'));
-    const budget = manifest.budget || {};
-    return {
-      maxTokensPerSession: budget.max_tokens_per_session || Infinity,
-      maxTokensPerTask: budget.max_tokens_per_task || Infinity,
-      actionOnExceed: budget.action_on_exceed || 'pause'
-    };
-  } catch {
-    return { maxTokensPerSession: Infinity, maxTokensPerTask: Infinity, actionOnExceed: 'pause' };
-  }
-}
-
-/**
- * Estimate tokens a task will consume (heuristic: chars/4 + overhead).
- */
-function estimateTaskTokens(task) {
-  const descLen = (task.description || '').length;
-  const criteriaLen = (task.acceptance_criteria || []).join(' ').length;
-  return Math.ceil((descLen + criteriaLen) / 4) + 500; // 500 base overhead
-}
-
-// ─── Heartbeat wrapper ────────────────────────────────────────────────────────
-
-/**
- * Run a task with heartbeat pulses on the bus every 30s.
- * Clears the interval on task completion/failure regardless.
- */
-async function runTaskWithHeartbeat(projectDir, squadSlug, task, sessionId, options, runFn) {
-  const { enableBus } = options;
-  const startMs = Date.now();
-
-  let hbInterval = null;
-  if (enableBus) {
-    hbInterval = setInterval(async () => {
-      const elapsed = Math.round((Date.now() - startMs) / 1000);
-      await bus.post(projectDir, squadSlug, sessionId, {
-        from: 'coordinator',
-        to: '*',
-        type: 'heartbeat',
-        content: `${task.title} — ${elapsed}s elapsed`,
-        metadata: { task_id: task.id, elapsed_s: elapsed }
-      }).catch(() => {});
-    }, 30_000);
-  }
-
-  try {
-    return await runFn();
-  } finally {
-    if (hbInterval) clearInterval(hbInterval);
-  }
-}
-
-// ─── Anti-analysis-loop guard ─────────────────────────────────────────────────
-
-/**
- * Check if an executor appears to be in an analysis loop.
- * Posts a coordinator feedback message if the threshold is exceeded.
- */
-async function checkAntiLoop(projectDir, squadSlug, sessionId, task, enableBus, threshold = 8) {
-  if (!enableBus || !task.executor) return;
-
-  const messages = await bus.read(projectDir, squadSlug, sessionId).catch(() => []);
-  if (bus.isAnalysisLoop(messages, task.executor, threshold)) {
-    await bus.post(projectDir, squadSlug, sessionId, {
-      from: 'coordinator',
-      to: task.executor,
-      type: 'feedback',
-      content: `Analysis loop detected for "${task.title}". You have posted ${threshold}+ status updates without a result. Stop analyzing — produce concrete output now, or post a "block" message explaining why you cannot proceed.`,
-      metadata: { task_id: task.id, anti_loop: true }
-    }).catch(() => {});
-  }
-}
-
-// ─── Core task runner ─────────────────────────────────────────────────────────
-
-/**
- * Run a single task (no retry logic here — handled by gap closure wrapper).
- */
-async function runTask(projectDir, squadSlug, task, sessionId, options, logger) {
-  const { enableBus, enableReflect, timeoutMs } = options;
-  const taskCtx = {
-    projectDir,
-    squadSlug,
-    executorSlug: task.executor || 'unknown',
-    taskTitle: task.title,
-    iteration: task._attempt || 1
-  };
-
-  // Post status to bus
-  if (enableBus) {
-    await bus.post(projectDir, squadSlug, sessionId, {
-      from: task.executor || 'coordinator',
-      to: '*',
-      type: 'status',
-      content: `Starting: ${task.title}${task._attempt > 1 ? ` (retry ${task._attempt})` : ''}`,
-      metadata: { task_id: task.id, attempt: task._attempt || 1 }
-    }).catch(() => {});
-  }
-
-  await updateTaskStatus(projectDir, squadSlug, sessionId, task.id, 'in_progress');
-
-  // Build worker input with fresh context pointers (2.2: paths, not inline content)
-  const workerInput = {
-    task_id: task.id,
-    title: task.title,
-    description: task.description,
-    acceptance_criteria: task.acceptance_criteria,
-    read_first_hints: task.read_first_hints || [],  // executor reads these, not coordinator
-    must_haves: task.must_haves || null,
-    session_id: sessionId,
-    bus_enabled: enableBus,
-    ...(task._failure_context ? { failure_context: task._failure_context, attempt: task._attempt } : {})
-  };
-
-  let workerResult = null;
-  let taskOutput = null;
-  let workerRan = false;
-
-  const workers = await listWorkers(projectDir, squadSlug);
-  const workerConfig = workers.find(
-    (w) => w.slug === task.executor || w.slug === task.id
-  );
-
-  if (workerConfig) {
-    workerRan = true;
-    workerResult = await runWorker(projectDir, squadSlug, workerConfig.slug, workerInput, {
-      timeoutMs,
-      triggerType: 'autorun'
-    });
-    taskOutput = workerResult.ok
-      ? JSON.stringify(workerResult.output || '')
-      : `Worker failed: ${workerResult.error}`;
-  } else {
-    taskOutput = `[no-worker-script] Task "${task.title}" assigned to executor "${task.executor}". Run manually or scaffold a worker with: aioson squad:worker --squad=${squadSlug} create --slug=${task.executor}`;
-    workerResult = { ok: true, output: { message: taskOutput }, noScript: true };
-  }
-
-  // Anti-analysis-loop check after task runs
-  await checkAntiLoop(projectDir, squadSlug, sessionId, task, enableBus);
-
-  // Reflection pass — pass full task object so verify-gate can check must_haves
-  let reflectionResult = null;
-  if (enableReflect && taskOutput && workerResult.ok) {
-    reflectionResult = await reflect(taskOutput, {
-      ...taskCtx,
-      iteration: task._attempt || 1,
-      task                            // enables must_haves verification
-    });
-
-    if (enableBus) {
-      await bus.post(projectDir, squadSlug, sessionId, {
-        from: task.executor || 'coordinator',
-        to: '*',
-        type: 'feedback',
-        content: formatReport(reflectionResult, task.executor),
-        metadata: { task_id: task.id, verdict: reflectionResult.verdict }
-      }).catch(() => {});
-    }
-  }
-
-  // Determine final status
-  let finalStatus;
-  if (!workerResult.ok) {
-    finalStatus = 'failed';
-  } else if (reflectionResult && reflectionResult.verdict === 'ESCALATE') {
-    finalStatus = 'escalated';
-  } else {
-    finalStatus = 'completed';
-  }
-
-  await updateTaskStatus(projectDir, squadSlug, sessionId, task.id, finalStatus, {
-    worker_ran: workerRan,
-    output_summary: String(taskOutput || '').slice(0, 500),
-    reflection: reflectionResult
-      ? { verdict: reflectionResult.verdict, score: reflectionResult.score }
-      : null,
-    completed_at: nowIso()
-  });
-
-  if (enableBus) {
-    await bus.post(projectDir, squadSlug, sessionId, {
-      from: task.executor || 'coordinator',
-      to: '*',
-      type: 'result',
-      content: `${icon(finalStatus)} ${task.title} → ${finalStatus.toUpperCase()}${reflectionResult ? ` (reflection: ${reflectionResult.verdict})` : ''}`,
-      metadata: { task_id: task.id, status: finalStatus }
-    }).catch(() => {});
-  }
-
-  return { task, finalStatus, workerResult, reflectionResult };
-}
-
-// ─── Inter-squad dependency validation ───────────────────────────────────────
-
-/**
- * Validate inter-squad dependencies declared in manifest.depends_on.
- * Returns { satisfied: boolean, unmet: Array<{squad, event}> }
- */
-async function validateInterSquadDependencies(projectDir, manifest) {
-  const deps = manifest.depends_on || [];
-  if (deps.length === 0) return { satisfied: true, unmet: [] };
-
-  const unmet = [];
-  for (const dep of deps) {
-    if (!dep.event) continue;
-    try {
-      const events = await interSquadEvents.consume(projectDir, {
-        toSquad: manifest.slug,
-        subscriptions: [dep.event]
-      });
-      if (events.length === 0) {
-        unmet.push({ squad: dep.squad || '(unknown)', event: dep.event });
-      }
-    } catch {
-      unmet.push({ squad: dep.squad || '(unknown)', event: dep.event });
-    }
-  }
-
-  return { satisfied: unmet.length === 0, unmet };
-}
-
-// ─── Bus coordinator intelligence ────────────────────────────────────────────
-
-// ── Block classification helpers ─────────────────────────────────────────────
-
-function extractKeyName(content) {
-  const m = content.match(/([A-Z_]{3,}_(?:KEY|TOKEN|SECRET|ID))/i);
-  return m ? m[1].toUpperCase() : 'API key';
-}
-
-function extractMissingPath(content) {
-  const m = content.match(/(?:file|path|found)[:\s]+([./\w-]+\.\w+)/i);
-  return m ? m[1] : 'unknown file';
-}
-
-function extractDependencyId(content) {
-  const m = content.match(/task[- _](\d{2})/i);
-  return m ? `task-${m[1]}` : null;
-}
-
-function classifyBlock(content) {
-  const c = String(content || '').toLowerCase();
-  if (/api\s*key|credential|secret|token\b|\.env/.test(c))     return 'api_key';
-  if (/file\s*not\s*found|no such file|enoent|missing file/.test(c)) return 'file_missing';
-  if (/permission\s*denied|access\s*denied|forbidden|403/.test(c))   return 'permission';
-  if (/timed?\s*out|etimedout|deadline|exceeded/.test(c))             return 'timeout';
-  if (/depends\s*on|waiting\s*for|requires\s*task|blocked\s*by/.test(c)) return 'dependency';
-  return 'generic';
-}
-
-/**
- * Strategy map: each strategy returns an { action, message, metadata } object.
- */
-const BLOCK_STRATEGIES = {
-  api_key: (block, _ctx) => ({
-    action: 'human-gate',
-    message: `API key/credential required: ${extractKeyName(block.content)}. Add to .env and restart the session.`,
-    metadata: { resolution: 'human_escalation', key_hint: extractKeyName(block.content) }
-  }),
-
-  file_missing: (block, ctx) => {
-    const missingFile = extractMissingPath(block.content);
-    const prevResults = (ctx.recentResults || []).filter((r) => r.includes('.') || r.includes('/'));
-    const hint = prevResults.length > 0
-      ? `Check if a previous task created it. Recent results: ${prevResults.slice(0, 2).join(' | ')}`
-      : `Verify the file path and that the producing task ran successfully.`;
-    return {
-      action: 'retry-with-context',
-      message: `File not found: ${missingFile}. ${hint}`,
-      metadata: { resolution: 'context_hint', missing_file: missingFile }
-    };
-  },
-
-  dependency: (block, _ctx) => {
-    const depId = extractDependencyId(block.content);
-    return {
-      action: 'wait-and-retry',
-      message: depId
-        ? `Waiting for ${depId} to complete before proceeding.`
-        : `Dependency not satisfied. Wait for upstream tasks to complete.`,
-      metadata: { resolution: 'dependency_wait', dep_id: depId }
-    };
-  },
-
-  permission: (_block, _ctx) => ({
-    action: 'human-gate',
-    message: 'Permission denied — this operation requires manual intervention or elevated access.',
-    metadata: { resolution: 'human_escalation' }
-  }),
-
-  timeout: (_block, _ctx) => ({
-    action: 'abort-and-escalate',
-    message: 'Task exceeded its timeout. It may need to be decomposed into smaller steps.',
-    metadata: { resolution: 'escalate_decompose' }
-  }),
-
-  generic: (block, ctx) => {
-    const recentCtx = (ctx.recentResults || []).join('\n');
-    return {
-      action: 'coordinator-hint',
-      message: `Coordinator attempting to resolve block. Try proceeding with available information.${recentCtx ? `\nContext from other executors:\n${recentCtx}` : ''}`,
-      metadata: { resolution: 'coordinator_hint' }
-    };
-  }
-};
-
-/**
- * After each wave, inspect the bus for unresolved block messages and
- * attempt automatic resolution using the strategy pattern.
- *
- * Block types: api_key, file_missing, dependency, permission, timeout, generic
- * Actions: human-gate, retry-with-context, wait-and-retry, abort-and-escalate, coordinator-hint
- */
-async function handleBusBlocks(projectDir, squadSlug, sessionId, logger) {
-  const allMessages = await bus.read(projectDir, squadSlug, sessionId).catch(() => []);
-  const unresolved = getUnresolvedBlocks(allMessages);
-
-  if (unresolved.length === 0) return;
-
-  const recentResultMessages = allMessages
-    .filter((m) => m.type === 'result')
-    .slice(-3)
-    .map((m) => `${m.from}: ${String(m.content || '').slice(0, 100)}`);
-
-  for (const block of unresolved) {
-    const blockType = classifyBlock(block.content);
-    const strategy = BLOCK_STRATEGIES[blockType] || BLOCK_STRATEGIES.generic;
-    const resolution = strategy(block, { recentResults: recentResultMessages });
-
-    if (resolution.action === 'human-gate' || resolution.action === 'abort-and-escalate') {
-      logger.log(`  ⚠ Block [${blockType}] requires human attention: "${String(block.content || '').slice(0, 80)}"`);
-    }
-
-    await bus.post(projectDir, squadSlug, sessionId, {
-      from: 'coordinator',
-      to: block.from,
-      type: 'resolution',
-      content: resolution.message,
-      metadata: { block_id: block.id, block_type: blockType, ...resolution.metadata }
-    }).catch(() => {});
-  }
-}
-
-// ─── Gap Closure Loop ─────────────────────────────────────────────────────────
-
-/**
- * Run a task with automatic retry on failure (gap closure).
- *
- * On failure:
- *   1. Capture the error as failure context
- *   2. Re-run the task with the failure context injected into workerInput
- *   3. Repeat up to maxRetries times
- *   4. If still failing after maxRetries: escalate
- *
- * ESCALATE results are NOT retried — they require coordinator/human attention.
- */
-async function runTaskWithGapClosure(
-  projectDir, squadSlug, task, sessionId, options, logger,
-  maxRetries = 3
-) {
-  const { enableBus } = options;
-  let currentTask = { ...task, _attempt: 1 };
-  let lastError = null;
-
-  for (let attempt = 1; attempt <= maxRetries; attempt++) {
-    currentTask = { ...currentTask, _attempt: attempt };
-
-    // Announce gap closure retry on bus (only from attempt 2 onwards)
-    if (attempt > 1 && enableBus) {
-      await bus.post(projectDir, squadSlug, sessionId, {
-        from: 'coordinator',
-        to: task.executor || '*',
-        type: 'gap_closure_attempt',
-        content: `Retrying "${task.title}" (attempt ${attempt}/${maxRetries}). Previous failure: ${lastError}`,
-        metadata: { task_id: task.id, attempt, prev_error: lastError }
-      }).catch(() => {});
-    }
-
-    // Run with heartbeat
-    const result = await runTaskWithHeartbeat(
-      projectDir, squadSlug, task, sessionId, options,
-      () => runTask(projectDir, squadSlug, currentTask, sessionId, options, logger)
-    );
-
-    if (result.finalStatus === 'completed') {
-      if (attempt > 1) {
-        logger.log(`    ↩ Gap closed after ${attempt} attempt(s)`);
-      }
-      return result;
-    }
-
-    // Don't retry escalated tasks — they need human/coordinator attention
-    if (result.finalStatus === 'escalated') {
-      return result;
-    }
-
-    // Capture failure context for next attempt
-    lastError = result.workerResult?.error || `task failed on attempt ${attempt}`;
-    currentTask = {
-      ...currentTask,
-      _failure_context: lastError
-    };
-
-    if (attempt < maxRetries) {
-      logger.log(`    ↩ Gap closure attempt ${attempt}/${maxRetries} failed — retrying with context`);
-    }
-  }
-
-  // Exhausted retries — escalate
-  logger.log(`    ✗ Gap closure exhausted (${maxRetries} attempts) — escalating`);
-
-  await updateTaskStatus(projectDir, squadSlug, sessionId, task.id, 'escalated', {
-    gap_closure_exhausted: true,
-    last_error: lastError,
-    completed_at: nowIso()
-  });
-
-  if (enableBus) {
-    await bus.post(projectDir, squadSlug, sessionId, {
-      from: 'coordinator',
-      to: '*',
-      type: 'block',
-      content: `Task "${task.title}" escalated after ${maxRetries} gap closure attempts. Last error: ${lastError}`,
-      metadata: { task_id: task.id, gap_closure_exhausted: true, attempts: maxRetries }
-    }).catch(() => {});
-  }
-
-  return {
-    task,
-    finalStatus: 'escalated',
-    workerResult: { ok: false, error: lastError, gap_closure_exhausted: true },
-    reflectionResult: null
-  };
-}
-
-// ─── Sampling-and-Voting (Plan 81 §Sprint 4) ────────────────────────────────
-
-/**
- * Synthesize votes from multiple worker instances.
- * Returns { consensus, winningOutput, allVotes }.
- *
- * Consensus is the fraction of instances that agree on the same status.
- * For completed tasks, compares output similarity via a simple hash.
- */
-function synthesizeVotes(votes) {
-  const statusCounts = {};
-  for (const v of votes) {
-    const s = v.finalStatus || 'unknown';
-    statusCounts[s] = (statusCounts[s] || 0) + 1;
-  }
-
-  let bestStatus = 'unknown';
-  let bestCount = 0;
-  for (const [s, count] of Object.entries(statusCounts)) {
-    if (count > bestCount) { bestStatus = s; bestCount = count; }
-  }
-
-  const consensus = bestCount / votes.length;
-  const winningVotes = votes.filter((v) => v.finalStatus === bestStatus);
-
-  return {
-    consensus,
-    bestStatus,
-    winningOutput: winningVotes[0],
-    allVotes: votes.map((v) => ({
-      finalStatus: v.finalStatus,
-      outputSummary: String(v.workerResult?.output || '').slice(0, 200)
-    }))
-  };
-}
-
-/**
- * Run a task with sampling-and-voting for critical decisions.
- *
- * Spawns N instances of the same task in parallel, then synthesizes votes.
- * If consensus < threshold, escalates to human-gate.
- *
- * @param {object} task — task with voting config: { instances, threshold }
- * @param {number} instances — number of parallel instances (default: 3)
- * @param {number} threshold — minimum consensus fraction (default: 0.66)
- */
-async function runTaskWithVoting(
-  projectDir, squadSlug, task, sessionId, options, logger,
-  instances = 3, threshold = 0.66
-) {
-  const { enableBus } = options;
-
-  if (enableBus) {
-    await bus.post(projectDir, squadSlug, sessionId, {
-      from: 'coordinator',
-      to: '*',
-      type: 'status',
-      content: `Sampling-and-Voting: running ${instances} instances of "${task.title}"`,
-      metadata: { task_id: task.id, voting: true, instances }
-    }).catch(() => {});
-  }
-
-  // Spawn N instances in parallel
-  const instancePromises = [];
-  for (let i = 1; i <= instances; i++) {
-    const instanceTask = { ...task, _attempt: 1, _voting_instance: i };
-    instancePromises.push(
-      runTaskWithHeartbeat(
-        projectDir, squadSlug, instanceTask, sessionId, options,
-        () => runTask(projectDir, squadSlug, instanceTask, sessionId, options, logger)
-      ).catch((err) => ({
-        task: instanceTask,
-        finalStatus: 'failed',
-        workerResult: { ok: false, error: err.message },
-        reflectionResult: null
-      }))
-    );
-  }
-
-  const votes = await Promise.all(instancePromises);
-  const result = synthesizeVotes(votes);
-
-  if (enableBus) {
-    await bus.post(projectDir, squadSlug, sessionId, {
-      from: 'coordinator',
-      to: '*',
-      type: 'feedback',
-      content: `Voting result for "${task.title}": consensus=${(result.consensus * 100).toFixed(0)}% status=${result.bestStatus} (${instances} instances)`,
-      metadata: { task_id: task.id, consensus: result.consensus, bestStatus: result.bestStatus }
-    }).catch(() => {});
-  }
-
-  // If consensus below threshold, escalate
-  if (result.consensus < threshold) {
-    logger.log(`    ⚠ Voting: consensus ${(result.consensus * 100).toFixed(0)}% < ${(threshold * 100).toFixed(0)}% threshold — escalating`);
-
-    await updateTaskStatus(projectDir, squadSlug, sessionId, task.id, 'escalated', {
-      voting: { consensus: result.consensus, threshold, allVotes: result.allVotes },
-      completed_at: nowIso()
-    });
-
-    if (enableBus) {
-      await bus.post(projectDir, squadSlug, sessionId, {
-        from: 'coordinator',
-        to: '*',
-        type: 'block',
-        content: `Task "${task.title}" — voting consensus too low (${(result.consensus * 100).toFixed(0)}%). Requires human review.`,
-        metadata: { task_id: task.id, voting_escalated: true }
-      }).catch(() => {});
-    }
-
-    return {
-      task,
-      finalStatus: 'escalated',
-      workerResult: { ok: false, error: 'voting_consensus_below_threshold', voting: result },
-      reflectionResult: null
-    };
-  }
-
-  // Use the winning vote as the result
-  return result.winningOutput;
-}
-
-// ─── Evaluator-Optimizer Loop (Plan 82 §ITEM 4) ──────────────────────────────
-
-/**
- * Run a task through an evaluator-optimizer loop (dev→qa→dev...):
- *   1. Generator phase: run the task as the implementer
- *   2. Evaluator phase: run a review task against criteria (fresh context)
- *   3. If PASS → done. If FAIL → inject structured feedback → repeat
- *   4. Max iterations before escalating to human
- *
- * The key: the evaluator receives only the artifact, not the generator's reasoning.
- * This prevents confirmation bias (same principle as verify-gate).
- */
-async function runWithEvalOptimize(
-  projectDir, squadSlug, task, sessionId, options, logger,
-  maxIterations = 3
-) {
-  const { enableBus } = options;
-  let lastFeedback = null;
-  let lastResult = null;
-
-  if (enableBus) {
-    await bus.post(projectDir, squadSlug, sessionId, {
-      from: 'coordinator',
-      to: '*',
-      type: 'status',
-      content: `Evaluator-Optimizer: starting loop for "${task.title}" (max ${maxIterations} iterations)`,
-      metadata: { task_id: task.id, eval_optimize: true, max_iterations: maxIterations }
-    }).catch(() => {});
-  }
-
-  for (let iteration = 1; iteration <= maxIterations; iteration++) {
-    // ── Generator phase ──────────────────────────────────────────────────────
-    const generatorTask = {
-      ...task,
-      _attempt: iteration,
-      ...(lastFeedback ? { _eval_feedback: lastFeedback, _eval_iteration: iteration } : {})
-    };
-
-    lastResult = await runTaskWithHeartbeat(
-      projectDir, squadSlug, generatorTask, sessionId, options,
-      () => runTask(projectDir, squadSlug, generatorTask, sessionId, options, logger)
-    );
-
-    if (lastResult.finalStatus !== 'completed') {
-      logger.log(`    ↳ Eval-Optimize iteration ${iteration}: generator failed — escalating`);
-      return lastResult;
-    }
-
-    // ── Evaluator phase ──────────────────────────────────────────────────────
-    const reviewTask = {
-      id: `${task.id}-review-${iteration}`,
-      title: `Review: ${task.title} (iteration ${iteration})`,
-      description: [
-        `Evaluate the output of task "${task.title}".`,
-        ``,
-        `Criteria to check (ALL must pass):`,
-        ...(task.review_criteria || []).map((c) => `- ${c}`),
-        ``,
-        `Output PASS if all criteria are met.`,
-        `Output FAIL with structured feedback: specific file:line references, exact criterion violated, minimum change to pass.`,
-        ``,
-        `Do NOT consider how the implementer reasoned — evaluate only the artifact.`
-      ].join('\n'),
-      executor: task.reviewer || 'qa',
-      acceptance_criteria: ['Output either PASS or FAIL with structured feedback'],
-      _eval_artifact: lastResult.workerResult?.output || null
-    };
-
-    const evalResult = await runTaskWithHeartbeat(
-      projectDir, squadSlug, reviewTask, sessionId, options,
-      () => runTask(projectDir, squadSlug, reviewTask, sessionId, options, logger)
-    );
-
-    const evalOutput = String(evalResult.workerResult?.output || '');
-    const passed = /\bPASS\b/i.test(evalOutput) && !/\bFAIL\b/i.test(evalOutput);
-
-    if (enableBus) {
-      await bus.post(projectDir, squadSlug, sessionId, {
-        from: 'coordinator',
-        to: '*',
-        type: 'feedback',
-        content: `Eval-Optimize iteration ${iteration}/${maxIterations}: ${passed ? 'PASS' : 'FAIL'} — ${evalOutput.slice(0, 120)}`,
-        metadata: { task_id: task.id, iteration, verdict: passed ? 'PASS' : 'FAIL' }
-      }).catch(() => {});
-    }
-
-    if (passed) {
-      logger.log(`    ↳ Eval-Optimize PASS at iteration ${iteration}`);
-      await updateTaskStatus(projectDir, squadSlug, sessionId, task.id, 'completed', {
-        eval_optimize: { iterations: iteration, verdict: 'PASS' },
-        completed_at: nowIso()
-      });
-      return lastResult;
-    }
-
-    lastFeedback = evalOutput.slice(0, 1000);
-    logger.log(`    ↳ Eval-Optimize FAIL at iteration ${iteration} — applying feedback`);
-  }
-
-  // Max iterations reached — escalate
-  logger.log(`    ✗ Eval-Optimize: ${maxIterations} iterations exhausted — escalating`);
-
-  await updateTaskStatus(projectDir, squadSlug, sessionId, task.id, 'escalated', {
-    eval_optimize: { iterations: maxIterations, verdict: 'FAIL', last_feedback: lastFeedback },
-    completed_at: nowIso()
-  });
-
-  if (enableBus) {
-    await bus.post(projectDir, squadSlug, sessionId, {
-      from: 'coordinator',
-      to: '*',
-      type: 'block',
-      content: `Task "${task.title}" — eval-optimize loop exhausted (${maxIterations} iterations). Requires human review.`,
-      metadata: { task_id: task.id, eval_optimize_exhausted: true }
-    }).catch(() => {});
-  }
-
-  return {
-    task,
-    finalStatus: 'escalated',
-    workerResult: { ok: false, error: 'eval_optimize_exhausted', last_feedback: lastFeedback },
-    reflectionResult: null
-  };
-}
-
-// ─── Main command ─────────────────────────────────────────────────────────────
-
-async function runSquadAutorun({ args, options = {}, logger }) {
-  const targetDir = path.resolve(process.cwd(), args[0] || '.');
-  const squadSlug = String(options.squad || options.s || '').trim();
-
-  if (!squadSlug) {
-    logger.error('Error: --squad is required');
-    return { ok: false, error: 'missing_squad' };
-  }
-
-  const dryRun = Boolean(options['dry-run'] || options.dryRun);
-  const enableBus = options.bus !== false && options.bus !== 'false';
-  const enableReflect = Boolean(options.reflect);
-  const sequential = Boolean(options.sequential);
-  const mode = String(options.mode || 'heuristic').trim();
-  const timeoutMs = (options.timeout ? Number(options.timeout) : 120) * 1000;
-  const existingPlanId = String(options.plan || '').trim();
-  const enableGapClosure = options['no-gap-closure'] !== true && options['no-gap-closure'] !== 'true';
-  const maxRetries = Math.min(Math.max(Number(options['max-retries'] || 3), 1), 5);
-  const requestedEngine = String(options.engine || 'legacy').trim();
-  const ignoreDeps = Boolean(options['ignore-deps'] || options.ignoreDeps);
-  const waitDeps = Boolean(options['wait-deps'] || options.waitDeps);
-  const waitDepsTimeoutMs = (options['wait-deps-timeout'] ? Number(options['wait-deps-timeout']) : 60) * 1000;
-
-  // ── Resolve execution engine (Plan 81 §1.1) ──────────────────────────────
-  const engineResult = resolveEngine(requestedEngine);
-  if (engineResult.fallback) {
-    logger.log(`⚠ ${engineResult.reason} — falling back to legacy engine`);
-  }
-
-  // ── Load token budget ──────────────────────────────────────────────────────
-  const budget = await loadBudget(targetDir, squadSlug);
-  let sessionTokensUsed = 0;
-  let budgetExceeded = false;
-
-  // ── Load or create plan ────────────────────────────────────────────────────
-  let plan;
-  let sessionId;
-
-  if (existingPlanId) {
-    plan = await loadPlan(targetDir, squadSlug, existingPlanId);
-    if (!plan) {
-      logger.error(`Plan not found: session "${existingPlanId}" for squad "${squadSlug}"`);
-      return { ok: false, error: 'plan_not_found' };
-    }
-    sessionId = existingPlanId;
-    logger.log(`Resuming plan [${sessionId}] — ${plan.tasks.length} tasks`);
-  } else {
-    const goal = String(options.goal || '').trim();
-    if (!goal) {
-      logger.error('Error: --goal is required (or --plan to resume an existing plan)');
-      return { ok: false, error: 'missing_goal' };
-    }
-
-    sessionId = randomUUID();
-    logger.log(`Decomposing goal for squad "${squadSlug}" [${mode}]...`);
-    plan = await decompose(targetDir, squadSlug, goal, { sessionId, mode, save: !dryRun });
-    logger.log(`Plan ready: ${plan.tasks.length} tasks across ${Object.keys(plan.parallel_groups).length} parallel group(s)`);
-  }
-
-  // ── Show plan ──────────────────────────────────────────────────────────────
-  if (options.json) {
-    if (dryRun) return { ok: true, dryRun: true, plan };
-  } else {
-    logger.log('');
-    logger.log(formatPlan(plan));
-    logger.log('');
-  }
-
-  if (dryRun) {
-    logger.log('[dry-run] Plan shown above. No tasks executed.');
-    return { ok: true, dryRun: true, plan };
-  }
-
-  // ── Handle structured mode ─────────────────────────────────────────────────
-  if (mode === 'structured' && plan.structured_prompt) {
-    const promptPath = path.join(
-      targetDir, '.aioson', 'squads', squadSlug, 'sessions', sessionId, 'decompose-prompt.md'
-    );
-    const promptContent = [
-      '---',
-      `session_id: ${sessionId}`,
-      `squad: ${squadSlug}`,
-      `created_at: ${plan.created_at}`,
-      '---',
-      '',
-      plan.structured_prompt,
-      '',
-      '> After the agent fills in the JSON above, run:',
-      `> aioson squad:autorun . --squad=${squadSlug} --plan=${sessionId}`
-    ].join('\n');
-
-    const { ensureDir } = require('../utils');
-    await ensureDir(path.dirname(promptPath));
-    await fs.writeFile(promptPath, promptContent, 'utf8');
-
-    logger.log(`[structured] Decomposition prompt saved to: ${path.relative(targetDir, promptPath)}`);
-    logger.log('Activate your agent to fill in the plan, then resume with:');
-    logger.log(`  aioson squad:autorun . --squad=${squadSlug} --plan=${sessionId}`);
-    return { ok: true, mode: 'structured', sessionId, promptPath: path.relative(targetDir, promptPath), plan };
-  }
-
-  // ── Load squad manifest (for inter-squad config and hooks) ─────────────────
-  let squadManifest = {};
-  try {
-    const manifestPath = path.join(targetDir, '.aioson', 'squads', squadSlug, 'squad.manifest.json');
-    squadManifest = JSON.parse(await fs.readFile(manifestPath, 'utf8'));
-  } catch { /* manifest is optional */ }
-
-  // ── Inter-squad dependency validation ─────────────────────────────────────
-  if (!ignoreDeps && (squadManifest.depends_on || []).length > 0) {
-    const depResult = await validateInterSquadDependencies(targetDir, squadManifest);
-    if (!depResult.satisfied) {
-      const depList = depResult.unmet.map((d) => `${d.squad}/${d.event}`).join(', ');
-      if (waitDeps) {
-        logger.log(`⏳ Waiting for inter-squad events: ${depList}`);
-        const deadline = Date.now() + waitDepsTimeoutMs;
-        let resolved = false;
-        while (Date.now() < deadline) {
-          await sleep(3000);
-          const retry = await validateInterSquadDependencies(targetDir, squadManifest);
-          if (retry.satisfied) { resolved = true; break; }
-        }
-        if (!resolved) {
-          logger.error(`✗ Dependencies still unmet after ${waitDepsTimeoutMs / 1000}s: ${depList}`);
-          return { ok: false, error: 'unmet_dependencies', unmet: depResult.unmet };
-        }
-        logger.log('  ✓ Dependencies satisfied');
-      } else {
-        logger.warn(`⚠ Unmet inter-squad dependencies: ${depList}`);
-        logger.warn('  Use --ignore-deps to run anyway, or --wait-deps to poll until satisfied');
-        return { ok: false, error: 'unmet_dependencies', unmet: depResult.unmet };
-      }
-    }
-  }
-
-  // ── 3.3 Hook Exit Code Protocol — pre_run hook ─────────────────────────────
-  const preRunHook = squadManifest.hooks?.pre_run;
-  if (preRunHook) {
-    logger.log(`Running pre_run hook: ${preRunHook.slice(0, 60)}${preRunHook.length > 60 ? '...' : ''}`);
-    const hookResult = runHook(preRunHook, {
-      squad: squadSlug,
-      session_id: sessionId,
-      goal: plan.goal,
-      project_dir: targetDir
-    });
-    if (hookResult.denied) {
-      logger.error(`✗ Pre-run hook denied execution${hookResult.stderr ? ': ' + hookResult.stderr : ''}`);
-      return { ok: false, error: 'hook_denied', hook: preRunHook, reason: hookResult.stderr };
-    }
-    if (hookResult.warn) {
-      logger.log(`  ⚠ Pre-run hook exited ${hookResult.exitCode} (non-fatal)${hookResult.stderr ? ': ' + hookResult.stderr : ''}`);
-    } else {
-      logger.log('  ✓ Pre-run hook passed');
-    }
-  }
-
-  // ── 3.1 Inter-Squad Event Streaming — consume pending events ───────────────
-  const subscriptions = [
-    ...(squadManifest.subscriptions || []),
-    ...(squadManifest.depends_on || []).map((d) => d.event).filter(Boolean)
-  ];
-  let incomingEvents = [];
-  if (subscriptions.length > 0) {
-    incomingEvents = await interSquadEvents
-      .consume(targetDir, { toSquad: squadSlug, subscriptions })
-      .catch(() => []);
-    if (incomingEvents.length > 0) {
-      logger.log(`Inter-squad events received: ${incomingEvents.length}`);
-      for (const ev of incomingEvents) {
-        logger.log(`  ← [${ev.fromSquad}] ${ev.event}${ev.payload ? ' · ' + JSON.stringify(ev.payload).slice(0, 80) : ''}`);
-      }
-      logger.log('');
-    }
-  }
-
-  // Inject incoming events into the plan goal context so executors are aware
-  if (incomingEvents.length > 0 && plan.tasks.length > 0) {
-    const firstTask = plan.tasks[0];
-    const eventSummary = incomingEvents
-      .map((e) => `[${e.fromSquad}] ${e.event}: ${JSON.stringify(e.payload || {})}`)
-      .join('\n');
-    firstTask._inter_squad_events = eventSummary;
-  }
-
-  // ── Record session start in STATE.md ───────────────────────────────────────
-  await stateManager.recordSessionStart(targetDir, squadSlug, sessionId, plan.goal).catch(() => {});
-
-  // ── Agent Teams engine path (Plan 81 §1.1) ────────────────────────────────
-  if (engineResult.engine === 'agent-teams') {
-    logger.log(`Engine: agent-teams (Claude Code ${engineResult.version})`);
-    const teamConfig = translateToTeamConfig(targetDir, squadManifest, plan, {
-      budget, enableBus
-    });
-    const configPath = await writeTeamConfig(targetDir, squadSlug, teamConfig);
-    logger.log(`Team config: ${path.relative(targetDir, configPath)}`);
-    logger.log(`Teammates: ${teamConfig.teammates.map((t) => t.name).join(', ')}`);
-    logger.log(`Tasks: ${teamConfig.tasks.length}`);
-    logger.log('');
-    logger.log('Agent Teams execution is configured. Use:');
-    logger.log(`  claude --team ${path.relative(targetDir, configPath)}`);
-    logger.log('');
-
-    await stateManager.recordSessionEnd(targetDir, squadSlug, sessionId, []).catch(() => {});
-
-    return {
-      ok: true,
-      engine: 'agent-teams',
-      session_id: sessionId,
-      squad: squadSlug,
-      configPath: path.relative(targetDir, configPath),
-      teamConfig
-    };
-  }
-
-  // ── Execute tasks (legacy engine) ─────────────────────────────────────────
-  logger.log(`Starting execution — session [${sessionId}]`);
-  if (enableBus) logger.log('Intra-squad bus: enabled');
-  if (enableReflect) logger.log('Reflection: enabled');
-  if (enableGapClosure) logger.log(`Gap closure: enabled (max ${maxRetries} retries)`);
-  if (budget.maxTokensPerSession !== Infinity) {
-    logger.log(`Budget: ${budget.maxTokensPerSession.toLocaleString()} tokens/session`);
-  }
-  logger.log('');
-
-  const results = [];
-  let completedCount = 0;
-  let failedCount = 0;
-  let escalatedCount = 0;
-  const startedAt = Date.now();
-
-  const runOptions = { enableBus, enableReflect, timeoutMs };
-
-  // Run in parallel group waves (or sequentially if --sequential)
-  const groups = Object.keys(plan.parallel_groups).map(Number).sort((a, b) => a - b);
-
-  for (const group of groups) {
-    if (budgetExceeded) {
-      logger.log(`⚠ Session budget exceeded (${sessionTokensUsed.toLocaleString()} tokens used). Stopping.`);
-      break;
-    }
-
-    const groupTaskIds = plan.parallel_groups[group];
-    const groupTasks = groupTaskIds
-      .map((id) => plan.tasks.find((t) => t.id === id))
-      .filter((t) => t && t.status === 'pending');
-
-    if (groupTasks.length === 0) continue;
-
-    logger.log(`── Group ${group} (${groupTasks.length} task${groupTasks.length > 1 ? 's' : ''})${groupTasks.length > 1 && !sequential ? ' — running in parallel' : ''}`);
-
-    // ── Budget gate: check before running this group ────────────────────────
-    const groupEstimatedTokens = groupTasks.reduce((sum, t) => sum + estimateTaskTokens(t), 0);
-    if (sessionTokensUsed + groupEstimatedTokens > budget.maxTokensPerSession) {
-      logger.log(`  ⚠ Budget gate: estimated ${(sessionTokensUsed + groupEstimatedTokens).toLocaleString()} tokens would exceed session limit of ${budget.maxTokensPerSession.toLocaleString()}.`);
-      if (budget.actionOnExceed === 'abort') {
-        logger.log('  Budget action: abort — stopping execution.');
-        budgetExceeded = true;
-        break;
-      }
-      // Default: 'pause' — warn and continue (user can see in summary)
-      logger.log('  Budget action: pause — marking remaining tasks as skipped.');
-      for (const task of groupTasks) {
-        await updateTaskStatus(targetDir, squadSlug, sessionId, task.id, 'skipped', {
-          skip_reason: 'budget_exceeded',
-          estimated_tokens: estimateTaskTokens(task)
-        });
-      }
-      budgetExceeded = true;
-      break;
-    }
-
-    // ── Run tasks in this group ────────────────────────────────────────────
-    const runTask_ = async (task) => {
-      // Brief validation guard (Plan 80 §1): validate brief before spawn
-      if (task.brief_path) {
-        const briefResult = await validateBrief(task.brief_path, targetDir);
-        if (!briefResult.ready) {
-          // Auto-fix simple fields (out_of_scope only)
-          if (briefResult.issues.length === 1 && briefResult.issues[0].field === 'out_of_scope') {
-            await autoFixBrief(task.brief_path, targetDir);
-            logger.log(`  ↩ Auto-fixed brief for ${task.id} (out_of_scope)`);
-          } else {
-            if (enableBus) {
-              await bus.post(targetDir, squadSlug, sessionId, {
-                from: 'coordinator', to: '*', type: 'block',
-                content: `Brief NOT READY for ${task.id}: ${briefResult.issues.map((i) => i.message).join(', ')}`,
-                metadata: { task_id: task.id, brief_validation: briefResult }
-              }).catch(() => {});
-            }
-            logger.log(`  ⚠ ${task.id}: brief NOT READY (${briefResult.issues.length} issues) — skipping`);
-            await updateTaskStatus(targetDir, squadSlug, sessionId, task.id, 'skipped', {
-              skip_reason: 'brief_not_ready', issues: briefResult.issues
-            });
-            return { task, finalStatus: 'skipped', workerResult: null, reflectionResult: null };
-          }
-        }
-      }
-
-      logger.log(`  ${icon('in_progress')} ${task.id}: ${task.title}`);
-
-      // Sampling-and-Voting path (Plan 81 §Sprint 4)
-      if (task.voting) {
-        const votingInstances = task.voting.instances || 3;
-        const votingThreshold = task.voting.threshold || 0.66;
-        logger.log(`    ↳ Voting: ${votingInstances} instances, threshold ${(votingThreshold * 100).toFixed(0)}%`);
-        const runner = runTaskWithVoting(
-          targetDir, squadSlug, task, sessionId, runOptions, logger,
-          votingInstances, votingThreshold
-        );
-        return runner.then((r) => {
-          logger.log(`  ${icon(r.finalStatus)} ${task.id}: ${r.finalStatus.toUpperCase()}`);
-          return r;
-        });
-      }
-
-      // Evaluator-Optimizer path (Plan 82 §ITEM 4)
-      if (task.review_loop) {
-        const maxReviewIter = task.max_review_iterations || 3;
-        logger.log(`    ↳ Eval-Optimize: max ${maxReviewIter} iterations, reviewer=${task.reviewer || 'qa'}`);
-        return runWithEvalOptimize(
-          targetDir, squadSlug, task, sessionId, runOptions, logger, maxReviewIter
-        ).then((r) => {
-          logger.log(`  ${icon(r.finalStatus)} ${task.id}: ${r.finalStatus.toUpperCase()}`);
-          return r;
-        });
-      }
-
-      const runner = enableGapClosure
-        ? runTaskWithGapClosure(targetDir, squadSlug, task, sessionId, runOptions, logger, maxRetries)
-        : runTaskWithHeartbeat(targetDir, squadSlug, task, sessionId, runOptions,
-            () => runTask(targetDir, squadSlug, task, sessionId, runOptions, logger));
-      return runner.then((r) => {
-        logger.log(`  ${icon(r.finalStatus)} ${task.id}: ${r.finalStatus.toUpperCase()}`);
-        return r;
-      });
-    };
-
-    let groupResults;
-    if (sequential || groupTasks.length === 1) {
-      groupResults = [];
-      for (const task of groupTasks) {
-        groupResults.push(await runTask_(task));
-      }
-    } else {
-      groupResults = await Promise.all(groupTasks.map(runTask_));
-    }
-
-    results.push(...groupResults);
-
-    // Accumulate token usage estimate for budget tracking
-    for (const r of groupResults) {
-      sessionTokensUsed += estimateTaskTokens(r.task);
-      if (r.finalStatus === 'completed') completedCount++;
-      else if (r.finalStatus === 'failed') failedCount++;
-      else if (r.finalStatus === 'escalated') escalatedCount++;
-    }
-
-    // Bus coordinator: resolve any unresolved blocks after this wave
-    if (enableBus) {
-      await handleBusBlocks(targetDir, squadSlug, sessionId, logger).catch(() => {});
-    }
-
-    // Short pause between waves to allow bus writes to flush
-    if (groups.length > 1) await sleep(100);
-  }
-
-  // ── Record session end in STATE.md ─────────────────────────────────────────
-  await stateManager.recordSessionEnd(targetDir, squadSlug, sessionId, results).catch(() => {});
-
-  // ── 5.1 Automatic Learning Extraction ─────────────────────────────────────
-  if (completedCount > 0 || escalatedCount > 0) {
-    const allBusMessages = enableBus
-      ? await bus.read(targetDir, squadSlug, sessionId).catch(() => [])
-      : [];
-    const allReflections = results
-      .filter((r) => r.reflectionResult)
-      .map((r) => r.reflectionResult);
-
-    const learnings = await extractLearnings(targetDir, squadSlug, sessionId, {
-      busMessages: allBusMessages,
-      taskResults: results,
-      reflectionReports: allReflections
-    }).catch(() => []);
-
-    if (learnings.length > 0) {
-      // Persist learnings to per-agent memory files (Plan 81 §Sprint 4)
-      await persistAgentMemory(targetDir, squadSlug, learnings, results).catch(() => {});
-      // Will be shown in the summary section below
-      results._extractedLearnings = learnings.length;
-    }
-  }
-
-  // ── Session summary ────────────────────────────────────────────────────────
-  const elapsed = Math.round((Date.now() - startedAt) / 1000);
-  const busSummary = enableBus
-    ? await bus.summary(targetDir, squadSlug, sessionId).catch(() => null)
-    : null;
-
-  const summary = {
-    ok: failedCount === 0 && escalatedCount === 0,
-    session_id: sessionId,
-    squad: squadSlug,
-    goal: plan.goal,
-    elapsed_s: elapsed,
-    budget_used: sessionTokensUsed,
-    budget_limit: budget.maxTokensPerSession,
-    tasks: {
-      total: plan.tasks.length,
-      completed: completedCount,
-      failed: failedCount,
-      escalated: escalatedCount
-    },
-    bus: busSummary
-      ? { total_messages: busSummary.total, blocks: busSummary.blocks.length }
-      : null
-  };
-
-  if (options.json) return summary;
-
-  logger.log('');
-  logger.log('── Autorun complete ─────────────────────────────────────────');
-  logger.log(`Session:   ${sessionId}`);
-  logger.log(`Elapsed:   ${elapsed}s`);
-  logger.log(`Tasks:     ${completedCount}/${plan.tasks.length} completed  ${failedCount > 0 ? `| ${failedCount} failed` : ''}  ${escalatedCount > 0 ? `| ${escalatedCount} escalated` : ''}`);
-  if (budget.maxTokensPerSession !== Infinity) {
-    logger.log(`Tokens:    ~${sessionTokensUsed.toLocaleString()} used / ${budget.maxTokensPerSession.toLocaleString()} budget`);
-  }
-  if (busSummary && busSummary.total > 0) {
-    logger.log(`Bus:       ${busSummary.total} messages${busSummary.blocks.length > 0 ? ` | ⚠ ${busSummary.blocks.length} block(s)` : ''}`);
-  }
-  if (escalatedCount > 0) {
-    logger.log('');
-    logger.log('⚠ Escalated tasks require coordinator attention:');
-    for (const r of results.filter((r) => r.finalStatus === 'escalated')) {
-      const exhausted = r.workerResult?.gap_closure_exhausted ? ' (gap closure exhausted)' : '';
-      logger.log(`  ${r.task.id}: ${r.task.title}${exhausted}`);
-    }
-  }
-  logger.log('');
-  logger.log(`Plan:      .aioson/squads/${squadSlug}/sessions/${sessionId}/plan.json`);
-  logger.log(`State:     .aioson/squads/${squadSlug}/STATE.md`);
-  if (enableBus) logger.log(`Bus:       .aioson/squads/${squadSlug}/sessions/${sessionId}/bus.jsonl`);
-
-  return summary;
-}
-
-module.exports = { runSquadAutorun };
+'use strict';
+
+/**
+ * aioson squad:autorun — Autonomous squad execution
+ *
+ * Given a high-level goal, this command:
+ *   1. Decomposes the goal into a task plan (heuristic or structured)
+ *   2. Activates the intra-squad bus for real-time communication
+ *   3. Runs each task through the worker-runner with optional reflection
+ *   4. Coordinator monitors the bus for blocks or feedback
+ *   5. Reports the final session summary
+ *
+ * Phase 1 additions:
+ *   - Gap Closure Loop: failed tasks are retried up to 3x with failure context
+ *   - Budget Gating: halts before a task if session token budget is exceeded
+ *   - Heartbeat Protocol: posts bus heartbeat every 30s during long tasks
+ *   - Anti-Analysis-Loop Guard: warns executor if N+ status without result
+ *   - Squad STATE.md: cross-session memory updated at start/end
+ *
+ * Usage:
+ *   aioson squad:autorun . --squad=content-team --goal="Create 3 podcast episodes"
+ *   aioson squad:autorun . --squad=content-team --goal="..." --reflect --bus --mode=structured
+ *   aioson squad:autorun . --squad=content-team --plan=SESSION_ID  (resume from saved plan)
+ *   aioson squad:autorun . --squad=content-team --plan=SESSION_ID --dry-run
+ *
+ * Flags:
+ *   --goal            High-level objective (required unless --plan is given)
+ *   --reflect         Run reflection after each task (default: false)
+ *   --bus             Enable intra-bus for inter-executor communication (default: true)
+ *   --mode            Decomposition mode: heuristic (default) | structured
+ *   --plan            Resume from existing session plan (session ID)
+ *   --dry-run         Show plan without executing
+ *   --sequential      Force sequential execution even for independent tasks (default: false)
+ *   --timeout         Per-task timeout in seconds (default: 120)
+ *   --max-retries     Gap closure max retry attempts (default: 3)
+ *   --no-gap-closure  Disable automatic retry on task failure
+ */
+
+const fs = require('node:fs/promises');
+const path = require('node:path');
+const { randomUUID } = require('node:crypto');
+const {
+  decompose,
+  getReadyTasks,
+  isPlanComplete,
+  updateTaskStatus,
+  loadPlan,
+  formatPlan
+} = require('../squad/task-decomposer');
+const bus = require('../squad/intra-bus');
+const { reflect, formatReport } = require('../squad/reflection');
+const { runWorker, listWorkers } = require('../worker-runner');
+const stateManager = require('../squad/state-manager');
+const { getUnresolvedBlocks } = require('../squad/intra-bus');
+const interSquadEvents = require('../squad/inter-squad-events');
+const { runHook, HOOK_DENY } = require('../lib/hook-protocol');
+const { extractLearnings, persistAgentMemory } = require('../squad/learning-extractor');
+const { validateBrief, autoFixBrief } = require('../squad/brief-validator');
+const { resolveEngine, translateToTeamConfig, writeTeamConfig } = require('../squad/agent-teams-adapter');
+
+const STATUS_ICON = {
+  pending: '○',
+  in_progress: '●',
+  completed: '✓',
+  failed: '✗',
+  escalated: '⚠',
+  skipped: '–'
+};
+
+function icon(status) {
+  return STATUS_ICON[status] || '?';
+}
+
+function nowIso() { return new Date().toISOString(); }
+
+function sleep(ms) { return new Promise((r) => setTimeout(r, ms)); }
+
+// ─── Budget helpers ───────────────────────────────────────────────────────────
+
+/**
+ * Load token budget from squad manifest (optional field).
+ * Returns Infinity if not configured.
+ */
+async function loadBudget(projectDir, squadSlug) {
+  const manifestPath = path.join(
+    projectDir, '.aioson', 'squads', squadSlug, 'squad.manifest.json'
+  );
+  try {
+    const manifest = JSON.parse(await fs.readFile(manifestPath, 'utf8'));
+    const budget = manifest.budget || {};
+    return {
+      maxTokensPerSession: budget.max_tokens_per_session || Infinity,
+      maxTokensPerTask: budget.max_tokens_per_task || Infinity,
+      actionOnExceed: budget.action_on_exceed || 'pause'
+    };
+  } catch {
+    return { maxTokensPerSession: Infinity, maxTokensPerTask: Infinity, actionOnExceed: 'pause' };
+  }
+}
+
+/**
+ * Estimate tokens a task will consume (heuristic: chars/4 + overhead).
+ */
+function estimateTaskTokens(task) {
+  const descLen = (task.description || '').length;
+  const criteriaLen = (task.acceptance_criteria || []).join(' ').length;
+  return Math.ceil((descLen + criteriaLen) / 4) + 500; // 500 base overhead
+}
+
+// ─── Heartbeat wrapper ────────────────────────────────────────────────────────
+
+/**
+ * Run a task with heartbeat pulses on the bus every 30s.
+ * Clears the interval on task completion/failure regardless.
+ */
+async function runTaskWithHeartbeat(projectDir, squadSlug, task, sessionId, options, runFn) {
+  const { enableBus } = options;
+  const startMs = Date.now();
+
+  let hbInterval = null;
+  if (enableBus) {
+    hbInterval = setInterval(async () => {
+      const elapsed = Math.round((Date.now() - startMs) / 1000);
+      await bus.post(projectDir, squadSlug, sessionId, {
+        from: 'coordinator',
+        to: '*',
+        type: 'heartbeat',
+        content: `${task.title} — ${elapsed}s elapsed`,
+        metadata: { task_id: task.id, elapsed_s: elapsed }
+      }).catch(() => {});
+    }, 30_000);
+  }
+
+  try {
+    return await runFn();
+  } finally {
+    if (hbInterval) clearInterval(hbInterval);
+  }
+}
+
+// ─── Anti-analysis-loop guard ─────────────────────────────────────────────────
+
+/**
+ * Check if an executor appears to be in an analysis loop.
+ * Posts a coordinator feedback message if the threshold is exceeded.
+ */
+async function checkAntiLoop(projectDir, squadSlug, sessionId, task, enableBus, threshold = 8) {
+  if (!enableBus || !task.executor) return;
+
+  const messages = await bus.read(projectDir, squadSlug, sessionId).catch(() => []);
+  if (bus.isAnalysisLoop(messages, task.executor, threshold)) {
+    await bus.post(projectDir, squadSlug, sessionId, {
+      from: 'coordinator',
+      to: task.executor,
+      type: 'feedback',
+      content: `Analysis loop detected for "${task.title}". You have posted ${threshold}+ status updates without a result. Stop analyzing — produce concrete output now, or post a "block" message explaining why you cannot proceed.`,
+      metadata: { task_id: task.id, anti_loop: true }
+    }).catch(() => {});
+  }
+}
+
+// ─── Core task runner ─────────────────────────────────────────────────────────
+
+/**
+ * Run a single task (no retry logic here — handled by gap closure wrapper).
+ */
+async function runTask(projectDir, squadSlug, task, sessionId, options, logger) {
+  const { enableBus, enableReflect, timeoutMs } = options;
+  const taskCtx = {
+    projectDir,
+    squadSlug,
+    executorSlug: task.executor || 'unknown',
+    taskTitle: task.title,
+    iteration: task._attempt || 1
+  };
+
+  // Post status to bus
+  if (enableBus) {
+    await bus.post(projectDir, squadSlug, sessionId, {
+      from: task.executor || 'coordinator',
+      to: '*',
+      type: 'status',
+      content: `Starting: ${task.title}${task._attempt > 1 ? ` (retry ${task._attempt})` : ''}`,
+      metadata: { task_id: task.id, attempt: task._attempt || 1 }
+    }).catch(() => {});
+  }
+
+  await updateTaskStatus(projectDir, squadSlug, sessionId, task.id, 'in_progress');
+
+  // Build worker input with fresh context pointers (2.2: paths, not inline content)
+  const workerInput = {
+    task_id: task.id,
+    title: task.title,
+    description: task.description,
+    acceptance_criteria: task.acceptance_criteria,
+    read_first_hints: task.read_first_hints || [],  // executor reads these, not coordinator
+    must_haves: task.must_haves || null,
+    session_id: sessionId,
+    bus_enabled: enableBus,
+    ...(task._failure_context ? { failure_context: task._failure_context, attempt: task._attempt } : {})
+  };
+
+  let workerResult = null;
+  let taskOutput = null;
+  let workerRan = false;
+
+  const workers = await listWorkers(projectDir, squadSlug);
+  const workerConfig = workers.find(
+    (w) => w.slug === task.executor || w.slug === task.id
+  );
+
+  if (workerConfig) {
+    workerRan = true;
+    workerResult = await runWorker(projectDir, squadSlug, workerConfig.slug, workerInput, {
+      timeoutMs,
+      triggerType: 'autorun'
+    });
+    taskOutput = workerResult.ok
+      ? JSON.stringify(workerResult.output || '')
+      : `Worker failed: ${workerResult.error}`;
+  } else {
+    taskOutput = `[no-worker-script] Task "${task.title}" assigned to executor "${task.executor}". Run manually or scaffold a worker with: aioson squad:worker --squad=${squadSlug} create --slug=${task.executor}`;
+    workerResult = { ok: true, output: { message: taskOutput }, noScript: true };
+  }
+
+  // Anti-analysis-loop check after task runs
+  await checkAntiLoop(projectDir, squadSlug, sessionId, task, enableBus);
+
+  // Reflection pass — pass full task object so verify-gate can check must_haves
+  let reflectionResult = null;
+  if (enableReflect && taskOutput && workerResult.ok) {
+    reflectionResult = await reflect(taskOutput, {
+      ...taskCtx,
+      iteration: task._attempt || 1,
+      task                            // enables must_haves verification
+    });
+
+    if (enableBus) {
+      await bus.post(projectDir, squadSlug, sessionId, {
+        from: task.executor || 'coordinator',
+        to: '*',
+        type: 'feedback',
+        content: formatReport(reflectionResult, task.executor),
+        metadata: { task_id: task.id, verdict: reflectionResult.verdict }
+      }).catch(() => {});
+    }
+  }
+
+  // Determine final status
+  let finalStatus;
+  if (!workerResult.ok) {
+    finalStatus = 'failed';
+  } else if (reflectionResult && reflectionResult.verdict === 'ESCALATE') {
+    finalStatus = 'escalated';
+  } else {
+    finalStatus = 'completed';
+  }
+
+  await updateTaskStatus(projectDir, squadSlug, sessionId, task.id, finalStatus, {
+    worker_ran: workerRan,
+    output_summary: String(taskOutput || '').slice(0, 500),
+    reflection: reflectionResult
+      ? { verdict: reflectionResult.verdict, score: reflectionResult.score }
+      : null,
+    completed_at: nowIso()
+  });
+
+  if (enableBus) {
+    await bus.post(projectDir, squadSlug, sessionId, {
+      from: task.executor || 'coordinator',
+      to: '*',
+      type: 'result',
+      content: `${icon(finalStatus)} ${task.title} → ${finalStatus.toUpperCase()}${reflectionResult ? ` (reflection: ${reflectionResult.verdict})` : ''}`,
+      metadata: { task_id: task.id, status: finalStatus }
+    }).catch(() => {});
+  }
+
+  return { task, finalStatus, workerResult, reflectionResult };
+}
+
+// ─── Inter-squad dependency validation ───────────────────────────────────────
+
+/**
+ * Validate inter-squad dependencies declared in manifest.depends_on.
+ * Returns { satisfied: boolean, unmet: Array<{squad, event}> }
+ */
+async function validateInterSquadDependencies(projectDir, manifest) {
+  const deps = manifest.depends_on || [];
+  if (deps.length === 0) return { satisfied: true, unmet: [] };
+
+  const unmet = [];
+  for (const dep of deps) {
+    if (!dep.event) continue;
+    try {
+      const events = await interSquadEvents.consume(projectDir, {
+        toSquad: manifest.slug,
+        subscriptions: [dep.event]
+      });
+      if (events.length === 0) {
+        unmet.push({ squad: dep.squad || '(unknown)', event: dep.event });
+      }
+    } catch {
+      unmet.push({ squad: dep.squad || '(unknown)', event: dep.event });
+    }
+  }
+
+  return { satisfied: unmet.length === 0, unmet };
+}
+
+// ─── Bus coordinator intelligence ────────────────────────────────────────────
+
+// ── Block classification helpers ─────────────────────────────────────────────
+
+function extractKeyName(content) {
+  const m = content.match(/([A-Z_]{3,}_(?:KEY|TOKEN|SECRET|ID))/i);
+  return m ? m[1].toUpperCase() : 'API key';
+}
+
+function extractMissingPath(content) {
+  const m = content.match(/(?:file|path|found)[:\s]+([./\w-]+\.\w+)/i);
+  return m ? m[1] : 'unknown file';
+}
+
+function extractDependencyId(content) {
+  const m = content.match(/task[- _](\d{2})/i);
+  return m ? `task-${m[1]}` : null;
+}
+
+function classifyBlock(content) {
+  const c = String(content || '').toLowerCase();
+  if (/api\s*key|credential|secret|token\b|\.env/.test(c))     return 'api_key';
+  if (/file\s*not\s*found|no such file|enoent|missing file/.test(c)) return 'file_missing';
+  if (/permission\s*denied|access\s*denied|forbidden|403/.test(c))   return 'permission';
+  if (/timed?\s*out|etimedout|deadline|exceeded/.test(c))             return 'timeout';
+  if (/depends\s*on|waiting\s*for|requires\s*task|blocked\s*by/.test(c)) return 'dependency';
+  return 'generic';
+}
+
+/**
+ * Strategy map: each strategy returns an { action, message, metadata } object.
+ */
+const BLOCK_STRATEGIES = {
+  api_key: (block, _ctx) => ({
+    action: 'human-gate',
+    message: `API key/credential required: ${extractKeyName(block.content)}. Add to .env and restart the session.`,
+    metadata: { resolution: 'human_escalation', key_hint: extractKeyName(block.content) }
+  }),
+
+  file_missing: (block, ctx) => {
+    const missingFile = extractMissingPath(block.content);
+    const prevResults = (ctx.recentResults || []).filter((r) => r.includes('.') || r.includes('/'));
+    const hint = prevResults.length > 0
+      ? `Check if a previous task created it. Recent results: ${prevResults.slice(0, 2).join(' | ')}`
+      : `Verify the file path and that the producing task ran successfully.`;
+    return {
+      action: 'retry-with-context',
+      message: `File not found: ${missingFile}. ${hint}`,
+      metadata: { resolution: 'context_hint', missing_file: missingFile }
+    };
+  },
+
+  dependency: (block, _ctx) => {
+    const depId = extractDependencyId(block.content);
+    return {
+      action: 'wait-and-retry',
+      message: depId
+        ? `Waiting for ${depId} to complete before proceeding.`
+        : `Dependency not satisfied. Wait for upstream tasks to complete.`,
+      metadata: { resolution: 'dependency_wait', dep_id: depId }
+    };
+  },
+
+  permission: (_block, _ctx) => ({
+    action: 'human-gate',
+    message: 'Permission denied — this operation requires manual intervention or elevated access.',
+    metadata: { resolution: 'human_escalation' }
+  }),
+
+  timeout: (_block, _ctx) => ({
+    action: 'abort-and-escalate',
+    message: 'Task exceeded its timeout. It may need to be decomposed into smaller steps.',
+    metadata: { resolution: 'escalate_decompose' }
+  }),
+
+  generic: (block, ctx) => {
+    const recentCtx = (ctx.recentResults || []).join('\n');
+    return {
+      action: 'coordinator-hint',
+      message: `Coordinator attempting to resolve block. Try proceeding with available information.${recentCtx ? `\nContext from other executors:\n${recentCtx}` : ''}`,
+      metadata: { resolution: 'coordinator_hint' }
+    };
+  }
+};
+
+/**
+ * After each wave, inspect the bus for unresolved block messages and
+ * attempt automatic resolution using the strategy pattern.
+ *
+ * Block types: api_key, file_missing, dependency, permission, timeout, generic
+ * Actions: human-gate, retry-with-context, wait-and-retry, abort-and-escalate, coordinator-hint
+ */
+async function handleBusBlocks(projectDir, squadSlug, sessionId, logger) {
+  const allMessages = await bus.read(projectDir, squadSlug, sessionId).catch(() => []);
+  const unresolved = getUnresolvedBlocks(allMessages);
+
+  if (unresolved.length === 0) return;
+
+  const recentResultMessages = allMessages
+    .filter((m) => m.type === 'result')
+    .slice(-3)
+    .map((m) => `${m.from}: ${String(m.content || '').slice(0, 100)}`);
+
+  for (const block of unresolved) {
+    const blockType = classifyBlock(block.content);
+    const strategy = BLOCK_STRATEGIES[blockType] || BLOCK_STRATEGIES.generic;
+    const resolution = strategy(block, { recentResults: recentResultMessages });
+
+    if (resolution.action === 'human-gate' || resolution.action === 'abort-and-escalate') {
+      logger.log(`  ⚠ Block [${blockType}] requires human attention: "${String(block.content || '').slice(0, 80)}"`);
+    }
+
+    await bus.post(projectDir, squadSlug, sessionId, {
+      from: 'coordinator',
+      to: block.from,
+      type: 'resolution',
+      content: resolution.message,
+      metadata: { block_id: block.id, block_type: blockType, ...resolution.metadata }
+    }).catch(() => {});
+  }
+}
+
+// ─── Gap Closure Loop ─────────────────────────────────────────────────────────
+
+/**
+ * Run a task with automatic retry on failure (gap closure).
+ *
+ * On failure:
+ *   1. Capture the error as failure context
+ *   2. Re-run the task with the failure context injected into workerInput
+ *   3. Repeat up to maxRetries times
+ *   4. If still failing after maxRetries: escalate
+ *
+ * ESCALATE results are NOT retried — they require coordinator/human attention.
+ */
+async function runTaskWithGapClosure(
+  projectDir, squadSlug, task, sessionId, options, logger,
+  maxRetries = 3
+) {
+  const { enableBus } = options;
+  let currentTask = { ...task, _attempt: 1 };
+  let lastError = null;
+
+  for (let attempt = 1; attempt <= maxRetries; attempt++) {
+    currentTask = { ...currentTask, _attempt: attempt };
+
+    // Announce gap closure retry on bus (only from attempt 2 onwards)
+    if (attempt > 1 && enableBus) {
+      await bus.post(projectDir, squadSlug, sessionId, {
+        from: 'coordinator',
+        to: task.executor || '*',
+        type: 'gap_closure_attempt',
+        content: `Retrying "${task.title}" (attempt ${attempt}/${maxRetries}). Previous failure: ${lastError}`,
+        metadata: { task_id: task.id, attempt, prev_error: lastError }
+      }).catch(() => {});
+    }
+
+    // Run with heartbeat
+    const result = await runTaskWithHeartbeat(
+      projectDir, squadSlug, task, sessionId, options,
+      () => runTask(projectDir, squadSlug, currentTask, sessionId, options, logger)
+    );
+
+    if (result.finalStatus === 'completed') {
+      if (attempt > 1) {
+        logger.log(`    ↩ Gap closed after ${attempt} attempt(s)`);
+      }
+      return result;
+    }
+
+    // Don't retry escalated tasks — they need human/coordinator attention
+    if (result.finalStatus === 'escalated') {
+      return result;
+    }
+
+    // Capture failure context for next attempt
+    lastError = result.workerResult?.error || `task failed on attempt ${attempt}`;
+    currentTask = {
+      ...currentTask,
+      _failure_context: lastError
+    };
+
+    if (attempt < maxRetries) {
+      logger.log(`    ↩ Gap closure attempt ${attempt}/${maxRetries} failed — retrying with context`);
+    }
+  }
+
+  // Exhausted retries — escalate
+  logger.log(`    ✗ Gap closure exhausted (${maxRetries} attempts) — escalating`);
+
+  await updateTaskStatus(projectDir, squadSlug, sessionId, task.id, 'escalated', {
+    gap_closure_exhausted: true,
+    last_error: lastError,
+    completed_at: nowIso()
+  });
+
+  if (enableBus) {
+    await bus.post(projectDir, squadSlug, sessionId, {
+      from: 'coordinator',
+      to: '*',
+      type: 'block',
+      content: `Task "${task.title}" escalated after ${maxRetries} gap closure attempts. Last error: ${lastError}`,
+      metadata: { task_id: task.id, gap_closure_exhausted: true, attempts: maxRetries }
+    }).catch(() => {});
+  }
+
+  return {
+    task,
+    finalStatus: 'escalated',
+    workerResult: { ok: false, error: lastError, gap_closure_exhausted: true },
+    reflectionResult: null
+  };
+}
+
+// ─── Sampling-and-Voting (Plan 81 §Sprint 4) ────────────────────────────────
+
+/**
+ * Synthesize votes from multiple worker instances.
+ * Returns { consensus, winningOutput, allVotes }.
+ *
+ * Consensus is the fraction of instances that agree on the same status.
+ * For completed tasks, compares output similarity via a simple hash.
+ */
+function synthesizeVotes(votes) {
+  const statusCounts = {};
+  for (const v of votes) {
+    const s = v.finalStatus || 'unknown';
+    statusCounts[s] = (statusCounts[s] || 0) + 1;
+  }
+
+  let bestStatus = 'unknown';
+  let bestCount = 0;
+  for (const [s, count] of Object.entries(statusCounts)) {
+    if (count > bestCount) { bestStatus = s; bestCount = count; }
+  }
+
+  const consensus = bestCount / votes.length;
+  const winningVotes = votes.filter((v) => v.finalStatus === bestStatus);
+
+  return {
+    consensus,
+    bestStatus,
+    winningOutput: winningVotes[0],
+    allVotes: votes.map((v) => ({
+      finalStatus: v.finalStatus,
+      outputSummary: String(v.workerResult?.output || '').slice(0, 200)
+    }))
+  };
+}
+
+/**
+ * Run a task with sampling-and-voting for critical decisions.
+ *
+ * Spawns N instances of the same task in parallel, then synthesizes votes.
+ * If consensus < threshold, escalates to human-gate.
+ *
+ * @param {object} task — task with voting config: { instances, threshold }
+ * @param {number} instances — number of parallel instances (default: 3)
+ * @param {number} threshold — minimum consensus fraction (default: 0.66)
+ */
+async function runTaskWithVoting(
+  projectDir, squadSlug, task, sessionId, options, logger,
+  instances = 3, threshold = 0.66
+) {
+  const { enableBus } = options;
+
+  if (enableBus) {
+    await bus.post(projectDir, squadSlug, sessionId, {
+      from: 'coordinator',
+      to: '*',
+      type: 'status',
+      content: `Sampling-and-Voting: running ${instances} instances of "${task.title}"`,
+      metadata: { task_id: task.id, voting: true, instances }
+    }).catch(() => {});
+  }
+
+  // Spawn N instances in parallel
+  const instancePromises = [];
+  for (let i = 1; i <= instances; i++) {
+    const instanceTask = { ...task, _attempt: 1, _voting_instance: i };
+    instancePromises.push(
+      runTaskWithHeartbeat(
+        projectDir, squadSlug, instanceTask, sessionId, options,
+        () => runTask(projectDir, squadSlug, instanceTask, sessionId, options, logger)
+      ).catch((err) => ({
+        task: instanceTask,
+        finalStatus: 'failed',
+        workerResult: { ok: false, error: err.message },
+        reflectionResult: null
+      }))
+    );
+  }
+
+  const votes = await Promise.all(instancePromises);
+  const result = synthesizeVotes(votes);
+
+  if (enableBus) {
+    await bus.post(projectDir, squadSlug, sessionId, {
+      from: 'coordinator',
+      to: '*',
+      type: 'feedback',
+      content: `Voting result for "${task.title}": consensus=${(result.consensus * 100).toFixed(0)}% status=${result.bestStatus} (${instances} instances)`,
+      metadata: { task_id: task.id, consensus: result.consensus, bestStatus: result.bestStatus }
+    }).catch(() => {});
+  }
+
+  // If consensus below threshold, escalate
+  if (result.consensus < threshold) {
+    logger.log(`    ⚠ Voting: consensus ${(result.consensus * 100).toFixed(0)}% < ${(threshold * 100).toFixed(0)}% threshold — escalating`);
+
+    await updateTaskStatus(projectDir, squadSlug, sessionId, task.id, 'escalated', {
+      voting: { consensus: result.consensus, threshold, allVotes: result.allVotes },
+      completed_at: nowIso()
+    });
+
+    if (enableBus) {
+      await bus.post(projectDir, squadSlug, sessionId, {
+        from: 'coordinator',
+        to: '*',
+        type: 'block',
+        content: `Task "${task.title}" — voting consensus too low (${(result.consensus * 100).toFixed(0)}%). Requires human review.`,
+        metadata: { task_id: task.id, voting_escalated: true }
+      }).catch(() => {});
+    }
+
+    return {
+      task,
+      finalStatus: 'escalated',
+      workerResult: { ok: false, error: 'voting_consensus_below_threshold', voting: result },
+      reflectionResult: null
+    };
+  }
+
+  // Use the winning vote as the result
+  return result.winningOutput;
+}
+
+// ─── Evaluator-Optimizer Loop (Plan 82 §ITEM 4) ──────────────────────────────
+
+/**
+ * Run a task through an evaluator-optimizer loop (dev→qa→dev...):
+ *   1. Generator phase: run the task as the implementer
+ *   2. Evaluator phase: run a review task against criteria (fresh context)
+ *   3. If PASS → done. If FAIL → inject structured feedback → repeat
+ *   4. Max iterations before escalating to human
+ *
+ * The key: the evaluator receives only the artifact, not the generator's reasoning.
+ * This prevents confirmation bias (same principle as verify-gate).
+ */
+async function runWithEvalOptimize(
+  projectDir, squadSlug, task, sessionId, options, logger,
+  maxIterations = 3
+) {
+  const { enableBus } = options;
+  let lastFeedback = null;
+  let lastResult = null;
+
+  if (enableBus) {
+    await bus.post(projectDir, squadSlug, sessionId, {
+      from: 'coordinator',
+      to: '*',
+      type: 'status',
+      content: `Evaluator-Optimizer: starting loop for "${task.title}" (max ${maxIterations} iterations)`,
+      metadata: { task_id: task.id, eval_optimize: true, max_iterations: maxIterations }
+    }).catch(() => {});
+  }
+
+  for (let iteration = 1; iteration <= maxIterations; iteration++) {
+    // ── Generator phase ──────────────────────────────────────────────────────
+    const generatorTask = {
+      ...task,
+      _attempt: iteration,
+      ...(lastFeedback ? { _eval_feedback: lastFeedback, _eval_iteration: iteration } : {})
+    };
+
+    lastResult = await runTaskWithHeartbeat(
+      projectDir, squadSlug, generatorTask, sessionId, options,
+      () => runTask(projectDir, squadSlug, generatorTask, sessionId, options, logger)
+    );
+
+    if (lastResult.finalStatus !== 'completed') {
+      logger.log(`    ↳ Eval-Optimize iteration ${iteration}: generator failed — escalating`);
+      return lastResult;
+    }
+
+    // ── Evaluator phase ──────────────────────────────────────────────────────
+    const reviewTask = {
+      id: `${task.id}-review-${iteration}`,
+      title: `Review: ${task.title} (iteration ${iteration})`,
+      description: [
+        `Evaluate the output of task "${task.title}".`,
+        ``,
+        `Criteria to check (ALL must pass):`,
+        ...(task.review_criteria || []).map((c) => `- ${c}`),
+        ``,
+        `Output PASS if all criteria are met.`,
+        `Output FAIL with structured feedback: specific file:line references, exact criterion violated, minimum change to pass.`,
+        ``,
+        `Do NOT consider how the implementer reasoned — evaluate only the artifact.`
+      ].join('\n'),
+      executor: task.reviewer || 'qa',
+      acceptance_criteria: ['Output either PASS or FAIL with structured feedback'],
+      _eval_artifact: lastResult.workerResult?.output || null
+    };
+
+    const evalResult = await runTaskWithHeartbeat(
+      projectDir, squadSlug, reviewTask, sessionId, options,
+      () => runTask(projectDir, squadSlug, reviewTask, sessionId, options, logger)
+    );
+
+    const evalOutput = String(evalResult.workerResult?.output || '');
+    const passed = /\bPASS\b/i.test(evalOutput) && !/\bFAIL\b/i.test(evalOutput);
+
+    if (enableBus) {
+      await bus.post(projectDir, squadSlug, sessionId, {
+        from: 'coordinator',
+        to: '*',
+        type: 'feedback',
+        content: `Eval-Optimize iteration ${iteration}/${maxIterations}: ${passed ? 'PASS' : 'FAIL'} — ${evalOutput.slice(0, 120)}`,
+        metadata: { task_id: task.id, iteration, verdict: passed ? 'PASS' : 'FAIL' }
+      }).catch(() => {});
+    }
+
+    if (passed) {
+      logger.log(`    ↳ Eval-Optimize PASS at iteration ${iteration}`);
+      await updateTaskStatus(projectDir, squadSlug, sessionId, task.id, 'completed', {
+        eval_optimize: { iterations: iteration, verdict: 'PASS' },
+        completed_at: nowIso()
+      });
+      return lastResult;
+    }
+
+    lastFeedback = evalOutput.slice(0, 1000);
+    logger.log(`    ↳ Eval-Optimize FAIL at iteration ${iteration} — applying feedback`);
+  }
+
+  // Max iterations reached — escalate
+  logger.log(`    ✗ Eval-Optimize: ${maxIterations} iterations exhausted — escalating`);
+
+  await updateTaskStatus(projectDir, squadSlug, sessionId, task.id, 'escalated', {
+    eval_optimize: { iterations: maxIterations, verdict: 'FAIL', last_feedback: lastFeedback },
+    completed_at: nowIso()
+  });
+
+  if (enableBus) {
+    await bus.post(projectDir, squadSlug, sessionId, {
+      from: 'coordinator',
+      to: '*',
+      type: 'block',
+      content: `Task "${task.title}" — eval-optimize loop exhausted (${maxIterations} iterations). Requires human review.`,
+      metadata: { task_id: task.id, eval_optimize_exhausted: true }
+    }).catch(() => {});
+  }
+
+  return {
+    task,
+    finalStatus: 'escalated',
+    workerResult: { ok: false, error: 'eval_optimize_exhausted', last_feedback: lastFeedback },
+    reflectionResult: null
+  };
+}
+
+// ─── Main command ─────────────────────────────────────────────────────────────
+
+async function runSquadAutorun({ args, options = {}, logger }) {
+  const targetDir = path.resolve(process.cwd(), args[0] || '.');
+  const squadSlug = String(options.squad || options.s || '').trim();
+
+  if (!squadSlug) {
+    logger.error('Error: --squad is required');
+    return { ok: false, error: 'missing_squad' };
+  }
+
+  const dryRun = Boolean(options['dry-run'] || options.dryRun);
+  const enableBus = options.bus !== false && options.bus !== 'false';
+  const enableReflect = Boolean(options.reflect);
+  const sequential = Boolean(options.sequential);
+  const mode = String(options.mode || 'heuristic').trim();
+  const timeoutMs = (options.timeout ? Number(options.timeout) : 120) * 1000;
+  const existingPlanId = String(options.plan || '').trim();
+  const enableGapClosure = options['no-gap-closure'] !== true && options['no-gap-closure'] !== 'true';
+  const maxRetries = Math.min(Math.max(Number(options['max-retries'] || 3), 1), 5);
+  const requestedEngine = String(options.engine || 'legacy').trim();
+  const ignoreDeps = Boolean(options['ignore-deps'] || options.ignoreDeps);
+  const waitDeps = Boolean(options['wait-deps'] || options.waitDeps);
+  const waitDepsTimeoutMs = (options['wait-deps-timeout'] ? Number(options['wait-deps-timeout']) : 60) * 1000;
+
+  // ── Resolve execution engine (Plan 81 §1.1) ──────────────────────────────
+  const engineResult = resolveEngine(requestedEngine);
+  if (engineResult.fallback) {
+    logger.log(`⚠ ${engineResult.reason} — falling back to legacy engine`);
+  }
+
+  // ── Load token budget ──────────────────────────────────────────────────────
+  const budget = await loadBudget(targetDir, squadSlug);
+  let sessionTokensUsed = 0;
+  let budgetExceeded = false;
+
+  // ── Load or create plan ────────────────────────────────────────────────────
+  let plan;
+  let sessionId;
+
+  if (existingPlanId) {
+    plan = await loadPlan(targetDir, squadSlug, existingPlanId);
+    if (!plan) {
+      logger.error(`Plan not found: session "${existingPlanId}" for squad "${squadSlug}"`);
+      return { ok: false, error: 'plan_not_found' };
+    }
+    sessionId = existingPlanId;
+    logger.log(`Resuming plan [${sessionId}] — ${plan.tasks.length} tasks`);
+  } else {
+    const goal = String(options.goal || '').trim();
+    if (!goal) {
+      logger.error('Error: --goal is required (or --plan to resume an existing plan)');
+      return { ok: false, error: 'missing_goal' };
+    }
+
+    sessionId = randomUUID();
+    logger.log(`Decomposing goal for squad "${squadSlug}" [${mode}]...`);
+    plan = await decompose(targetDir, squadSlug, goal, { sessionId, mode, save: !dryRun });
+    logger.log(`Plan ready: ${plan.tasks.length} tasks across ${Object.keys(plan.parallel_groups).length} parallel group(s)`);
+  }
+
+  // ── Show plan ──────────────────────────────────────────────────────────────
+  if (options.json) {
+    if (dryRun) return { ok: true, dryRun: true, plan };
+  } else {
+    logger.log('');
+    logger.log(formatPlan(plan));
+    logger.log('');
+  }
+
+  if (dryRun) {
+    logger.log('[dry-run] Plan shown above. No tasks executed.');
+    return { ok: true, dryRun: true, plan };
+  }
+
+  // ── Handle structured mode ─────────────────────────────────────────────────
+  if (mode === 'structured' && plan.structured_prompt) {
+    const promptPath = path.join(
+      targetDir, '.aioson', 'squads', squadSlug, 'sessions', sessionId, 'decompose-prompt.md'
+    );
+    const promptContent = [
+      '---',
+      `session_id: ${sessionId}`,
+      `squad: ${squadSlug}`,
+      `created_at: ${plan.created_at}`,
+      '---',
+      '',
+      plan.structured_prompt,
+      '',
+      '> After the agent fills in the JSON above, run:',
+      `> aioson squad:autorun . --squad=${squadSlug} --plan=${sessionId}`
+    ].join('\n');
+
+    const { ensureDir } = require('../utils');
+    await ensureDir(path.dirname(promptPath));
+    await fs.writeFile(promptPath, promptContent, 'utf8');
+
+    logger.log(`[structured] Decomposition prompt saved to: ${path.relative(targetDir, promptPath)}`);
+    logger.log('Activate your agent to fill in the plan, then resume with:');
+    logger.log(`  aioson squad:autorun . --squad=${squadSlug} --plan=${sessionId}`);
+    return { ok: true, mode: 'structured', sessionId, promptPath: path.relative(targetDir, promptPath), plan };
+  }
+
+  // ── Load squad manifest (for inter-squad config and hooks) ─────────────────
+  let squadManifest = {};
+  try {
+    const manifestPath = path.join(targetDir, '.aioson', 'squads', squadSlug, 'squad.manifest.json');
+    squadManifest = JSON.parse(await fs.readFile(manifestPath, 'utf8'));
+  } catch { /* manifest is optional */ }
+
+  // ── Inter-squad dependency validation ─────────────────────────────────────
+  if (!ignoreDeps && (squadManifest.depends_on || []).length > 0) {
+    const depResult = await validateInterSquadDependencies(targetDir, squadManifest);
+    if (!depResult.satisfied) {
+      const depList = depResult.unmet.map((d) => `${d.squad}/${d.event}`).join(', ');
+      if (waitDeps) {
+        logger.log(`⏳ Waiting for inter-squad events: ${depList}`);
+        const deadline = Date.now() + waitDepsTimeoutMs;
+        let resolved = false;
+        while (Date.now() < deadline) {
+          await sleep(3000);
+          const retry = await validateInterSquadDependencies(targetDir, squadManifest);
+          if (retry.satisfied) { resolved = true; break; }
+        }
+        if (!resolved) {
+          logger.error(`✗ Dependencies still unmet after ${waitDepsTimeoutMs / 1000}s: ${depList}`);
+          return { ok: false, error: 'unmet_dependencies', unmet: depResult.unmet };
+        }
+        logger.log('  ✓ Dependencies satisfied');
+      } else {
+        logger.warn(`⚠ Unmet inter-squad dependencies: ${depList}`);
+        logger.warn('  Use --ignore-deps to run anyway, or --wait-deps to poll until satisfied');
+        return { ok: false, error: 'unmet_dependencies', unmet: depResult.unmet };
+      }
+    }
+  }
+
+  // ── 3.3 Hook Exit Code Protocol — pre_run hook ─────────────────────────────
+  const preRunHook = squadManifest.hooks?.pre_run;
+  if (preRunHook) {
+    logger.log(`Running pre_run hook: ${preRunHook.slice(0, 60)}${preRunHook.length > 60 ? '...' : ''}`);
+    const hookResult = runHook(preRunHook, {
+      squad: squadSlug,
+      session_id: sessionId,
+      goal: plan.goal,
+      project_dir: targetDir
+    });
+    if (hookResult.denied) {
+      logger.error(`✗ Pre-run hook denied execution${hookResult.stderr ? ': ' + hookResult.stderr : ''}`);
+      return { ok: false, error: 'hook_denied', hook: preRunHook, reason: hookResult.stderr };
+    }
+    if (hookResult.warn) {
+      logger.log(`  ⚠ Pre-run hook exited ${hookResult.exitCode} (non-fatal)${hookResult.stderr ? ': ' + hookResult.stderr : ''}`);
+    } else {
+      logger.log('  ✓ Pre-run hook passed');
+    }
+  }
+
+  // ── 3.1 Inter-Squad Event Streaming — consume pending events ───────────────
+  const subscriptions = [
+    ...(squadManifest.subscriptions || []),
+    ...(squadManifest.depends_on || []).map((d) => d.event).filter(Boolean)
+  ];
+  let incomingEvents = [];
+  if (subscriptions.length > 0) {
+    incomingEvents = await interSquadEvents
+      .consume(targetDir, { toSquad: squadSlug, subscriptions })
+      .catch(() => []);
+    if (incomingEvents.length > 0) {
+      logger.log(`Inter-squad events received: ${incomingEvents.length}`);
+      for (const ev of incomingEvents) {
+        logger.log(`  ← [${ev.fromSquad}] ${ev.event}${ev.payload ? ' · ' + JSON.stringify(ev.payload).slice(0, 80) : ''}`);
+      }
+      logger.log('');
+    }
+  }
+
+  // Inject incoming events into the plan goal context so executors are aware
+  if (incomingEvents.length > 0 && plan.tasks.length > 0) {
+    const firstTask = plan.tasks[0];
+    const eventSummary = incomingEvents
+      .map((e) => `[${e.fromSquad}] ${e.event}: ${JSON.stringify(e.payload || {})}`)
+      .join('\n');
+    firstTask._inter_squad_events = eventSummary;
+  }
+
+  // ── Record session start in STATE.md ───────────────────────────────────────
+  await stateManager.recordSessionStart(targetDir, squadSlug, sessionId, plan.goal).catch(() => {});
+
+  // ── Agent Teams engine path (Plan 81 §1.1) ────────────────────────────────
+  if (engineResult.engine === 'agent-teams') {
+    logger.log(`Engine: agent-teams (Claude Code ${engineResult.version})`);
+    const teamConfig = translateToTeamConfig(targetDir, squadManifest, plan, {
+      budget, enableBus
+    });
+    const configPath = await writeTeamConfig(targetDir, squadSlug, teamConfig);
+    logger.log(`Team config: ${path.relative(targetDir, configPath)}`);
+    logger.log(`Teammates: ${teamConfig.teammates.map((t) => t.name).join(', ')}`);
+    logger.log(`Tasks: ${teamConfig.tasks.length}`);
+    logger.log('');
+    logger.log('Agent Teams execution is configured. Use:');
+    logger.log(`  claude --team ${path.relative(targetDir, configPath)}`);
+    logger.log('');
+
+    await stateManager.recordSessionEnd(targetDir, squadSlug, sessionId, []).catch(() => {});
+
+    return {
+      ok: true,
+      engine: 'agent-teams',
+      session_id: sessionId,
+      squad: squadSlug,
+      configPath: path.relative(targetDir, configPath),
+      teamConfig
+    };
+  }
+
+  // ── Execute tasks (legacy engine) ─────────────────────────────────────────
+  logger.log(`Starting execution — session [${sessionId}]`);
+  if (enableBus) logger.log('Intra-squad bus: enabled');
+  if (enableReflect) logger.log('Reflection: enabled');
+  if (enableGapClosure) logger.log(`Gap closure: enabled (max ${maxRetries} retries)`);
+  if (budget.maxTokensPerSession !== Infinity) {
+    logger.log(`Budget: ${budget.maxTokensPerSession.toLocaleString()} tokens/session`);
+  }
+  logger.log('');
+
+  const results = [];
+  let completedCount = 0;
+  let failedCount = 0;
+  let escalatedCount = 0;
+  const startedAt = Date.now();
+
+  const runOptions = { enableBus, enableReflect, timeoutMs };
+
+  // Run in parallel group waves (or sequentially if --sequential)
+  const groups = Object.keys(plan.parallel_groups).map(Number).sort((a, b) => a - b);
+
+  for (const group of groups) {
+    if (budgetExceeded) {
+      logger.log(`⚠ Session budget exceeded (${sessionTokensUsed.toLocaleString()} tokens used). Stopping.`);
+      break;
+    }
+
+    const groupTaskIds = plan.parallel_groups[group];
+    const groupTasks = groupTaskIds
+      .map((id) => plan.tasks.find((t) => t.id === id))
+      .filter((t) => t && t.status === 'pending');
+
+    if (groupTasks.length === 0) continue;
+
+    logger.log(`── Group ${group} (${groupTasks.length} task${groupTasks.length > 1 ? 's' : ''})${groupTasks.length > 1 && !sequential ? ' — running in parallel' : ''}`);
+
+    // ── Budget gate: check before running this group ────────────────────────
+    const groupEstimatedTokens = groupTasks.reduce((sum, t) => sum + estimateTaskTokens(t), 0);
+    if (sessionTokensUsed + groupEstimatedTokens > budget.maxTokensPerSession) {
+      logger.log(`  ⚠ Budget gate: estimated ${(sessionTokensUsed + groupEstimatedTokens).toLocaleString()} tokens would exceed session limit of ${budget.maxTokensPerSession.toLocaleString()}.`);
+      if (budget.actionOnExceed === 'abort') {
+        logger.log('  Budget action: abort — stopping execution.');
+        budgetExceeded = true;
+        break;
+      }
+      // Default: 'pause' — warn and continue (user can see in summary)
+      logger.log('  Budget action: pause — marking remaining tasks as skipped.');
+      for (const task of groupTasks) {
+        await updateTaskStatus(targetDir, squadSlug, sessionId, task.id, 'skipped', {
+          skip_reason: 'budget_exceeded',
+          estimated_tokens: estimateTaskTokens(task)
+        });
+      }
+      budgetExceeded = true;
+      break;
+    }
+
+    // ── Run tasks in this group ────────────────────────────────────────────
+    const runTask_ = async (task) => {
+      // Brief validation guard (Plan 80 §1): validate brief before spawn
+      if (task.brief_path) {
+        const briefResult = await validateBrief(task.brief_path, targetDir);
+        if (!briefResult.ready) {
+          // Auto-fix simple fields (out_of_scope only)
+          if (briefResult.issues.length === 1 && briefResult.issues[0].field === 'out_of_scope') {
+            await autoFixBrief(task.brief_path, targetDir);
+            logger.log(`  ↩ Auto-fixed brief for ${task.id} (out_of_scope)`);
+          } else {
+            if (enableBus) {
+              await bus.post(targetDir, squadSlug, sessionId, {
+                from: 'coordinator', to: '*', type: 'block',
+                content: `Brief NOT READY for ${task.id}: ${briefResult.issues.map((i) => i.message).join(', ')}`,
+                metadata: { task_id: task.id, brief_validation: briefResult }
+              }).catch(() => {});
+            }
+            logger.log(`  ⚠ ${task.id}: brief NOT READY (${briefResult.issues.length} issues) — skipping`);
+            await updateTaskStatus(targetDir, squadSlug, sessionId, task.id, 'skipped', {
+              skip_reason: 'brief_not_ready', issues: briefResult.issues
+            });
+            return { task, finalStatus: 'skipped', workerResult: null, reflectionResult: null };
+          }
+        }
+      }
+
+      logger.log(`  ${icon('in_progress')} ${task.id}: ${task.title}`);
+
+      // Sampling-and-Voting path (Plan 81 §Sprint 4)
+      if (task.voting) {
+        const votingInstances = task.voting.instances || 3;
+        const votingThreshold = task.voting.threshold || 0.66;
+        logger.log(`    ↳ Voting: ${votingInstances} instances, threshold ${(votingThreshold * 100).toFixed(0)}%`);
+        const runner = runTaskWithVoting(
+          targetDir, squadSlug, task, sessionId, runOptions, logger,
+          votingInstances, votingThreshold
+        );
+        return runner.then((r) => {
+          logger.log(`  ${icon(r.finalStatus)} ${task.id}: ${r.finalStatus.toUpperCase()}`);
+          return r;
+        });
+      }
+
+      // Evaluator-Optimizer path (Plan 82 §ITEM 4)
+      if (task.review_loop) {
+        const maxReviewIter = task.max_review_iterations || 3;
+        logger.log(`    ↳ Eval-Optimize: max ${maxReviewIter} iterations, reviewer=${task.reviewer || 'qa'}`);
+        return runWithEvalOptimize(
+          targetDir, squadSlug, task, sessionId, runOptions, logger, maxReviewIter
+        ).then((r) => {
+          logger.log(`  ${icon(r.finalStatus)} ${task.id}: ${r.finalStatus.toUpperCase()}`);
+          return r;
+        });
+      }
+
+      const runner = enableGapClosure
+        ? runTaskWithGapClosure(targetDir, squadSlug, task, sessionId, runOptions, logger, maxRetries)
+        : runTaskWithHeartbeat(targetDir, squadSlug, task, sessionId, runOptions,
+            () => runTask(targetDir, squadSlug, task, sessionId, runOptions, logger));
+      return runner.then((r) => {
+        logger.log(`  ${icon(r.finalStatus)} ${task.id}: ${r.finalStatus.toUpperCase()}`);
+        return r;
+      });
+    };
+
+    let groupResults;
+    if (sequential || groupTasks.length === 1) {
+      groupResults = [];
+      for (const task of groupTasks) {
+        groupResults.push(await runTask_(task));
+      }
+    } else {
+      groupResults = await Promise.all(groupTasks.map(runTask_));
+    }
+
+    results.push(...groupResults);
+
+    // Accumulate token usage estimate for budget tracking
+    for (const r of groupResults) {
+      sessionTokensUsed += estimateTaskTokens(r.task);
+      if (r.finalStatus === 'completed') completedCount++;
+      else if (r.finalStatus === 'failed') failedCount++;
+      else if (r.finalStatus === 'escalated') escalatedCount++;
+    }
+
+    // Bus coordinator: resolve any unresolved blocks after this wave
+    if (enableBus) {
+      await handleBusBlocks(targetDir, squadSlug, sessionId, logger).catch(() => {});
+    }
+
+    // Short pause between waves to allow bus writes to flush
+    if (groups.length > 1) await sleep(100);
+  }
+
+  // ── Record session end in STATE.md ─────────────────────────────────────────
+  await stateManager.recordSessionEnd(targetDir, squadSlug, sessionId, results).catch(() => {});
+
+  // ── 5.1 Automatic Learning Extraction ─────────────────────────────────────
+  if (completedCount > 0 || escalatedCount > 0) {
+    const allBusMessages = enableBus
+      ? await bus.read(targetDir, squadSlug, sessionId).catch(() => [])
+      : [];
+    const allReflections = results
+      .filter((r) => r.reflectionResult)
+      .map((r) => r.reflectionResult);
+
+    const learnings = await extractLearnings(targetDir, squadSlug, sessionId, {
+      busMessages: allBusMessages,
+      taskResults: results,
+      reflectionReports: allReflections
+    }).catch(() => []);
+
+    if (learnings.length > 0) {
+      // Persist learnings to per-agent memory files (Plan 81 §Sprint 4)
+      await persistAgentMemory(targetDir, squadSlug, learnings, results).catch(() => {});
+      // Will be shown in the summary section below
+      results._extractedLearnings = learnings.length;
+    }
+  }
+
+  // ── Session summary ────────────────────────────────────────────────────────
+  const elapsed = Math.round((Date.now() - startedAt) / 1000);
+  const busSummary = enableBus
+    ? await bus.summary(targetDir, squadSlug, sessionId).catch(() => null)
+    : null;
+
+  const summary = {
+    ok: failedCount === 0 && escalatedCount === 0,
+    session_id: sessionId,
+    squad: squadSlug,
+    goal: plan.goal,
+    elapsed_s: elapsed,
+    budget_used: sessionTokensUsed,
+    budget_limit: budget.maxTokensPerSession,
+    tasks: {
+      total: plan.tasks.length,
+      completed: completedCount,
+      failed: failedCount,
+      escalated: escalatedCount
+    },
+    bus: busSummary
+      ? { total_messages: busSummary.total, blocks: busSummary.blocks.length }
+      : null
+  };
+
+  if (options.json) return summary;
+
+  logger.log('');
+  logger.log('── Autorun complete ─────────────────────────────────────────');
+  logger.log(`Session:   ${sessionId}`);
+  logger.log(`Elapsed:   ${elapsed}s`);
+  logger.log(`Tasks:     ${completedCount}/${plan.tasks.length} completed  ${failedCount > 0 ? `| ${failedCount} failed` : ''}  ${escalatedCount > 0 ? `| ${escalatedCount} escalated` : ''}`);
+  if (budget.maxTokensPerSession !== Infinity) {
+    logger.log(`Tokens:    ~${sessionTokensUsed.toLocaleString()} used / ${budget.maxTokensPerSession.toLocaleString()} budget`);
+  }
+  if (busSummary && busSummary.total > 0) {
+    logger.log(`Bus:       ${busSummary.total} messages${busSummary.blocks.length > 0 ? ` | ⚠ ${busSummary.blocks.length} block(s)` : ''}`);
+  }
+  if (escalatedCount > 0) {
+    logger.log('');
+    logger.log('⚠ Escalated tasks require coordinator attention:');
+    for (const r of results.filter((r) => r.finalStatus === 'escalated')) {
+      const exhausted = r.workerResult?.gap_closure_exhausted ? ' (gap closure exhausted)' : '';
+      logger.log(`  ${r.task.id}: ${r.task.title}${exhausted}`);
+    }
+  }
+  logger.log('');
+  logger.log(`Plan:      .aioson/squads/${squadSlug}/sessions/${sessionId}/plan.json`);
+  logger.log(`State:     .aioson/squads/${squadSlug}/STATE.md`);
+  if (enableBus) logger.log(`Bus:       .aioson/squads/${squadSlug}/sessions/${sessionId}/bus.jsonl`);
+
+  return summary;
+}
+
+module.exports = { runSquadAutorun };