create-walle 0.9.21 → 0.9.23

package/template/wall-e/eval/harvester.js

@@ -1,685 +0,0 @@
-'use strict';
-
-const fs = require('fs');
-const path = require('path');
-const { createHash } = require('crypto');
-const { execFileSync } = require('child_process');
-
-let claudeDesktopSessions = null;
-function getClaudeDesktopSessions() {
-  if (claudeDesktopSessions) return claudeDesktopSessions;
-  try {
-    claudeDesktopSessions = require('../../claude-task-manager/lib/claude-desktop-sessions');
-  } catch {
-    return null;
-  }
-  return claudeDesktopSessions;
-}
-
-// --- Task type classification ---
-
-function classifyTaskType(content) {
-  if (!content) return 'chat';
-  const lower = content.toLowerCase();
-  // Coding signals
-  if (/\b(function|class|const |let |var |import |require\(|def |async |await |return )\b/.test(content)) return 'coding';
-  if (/\b(fix|bug|error|implement|refactor|test|deploy|build)\b/i.test(lower)) return 'coding';
-  // Planning signals
-  if (/\b(plan|design|architect\w*|roadmap|strategy|prioritize|brainstorm|spec\b|tradeoffs?|trade-offs?|discuss|proposal)\b/i.test(lower)) return 'planning';
-  // QA signals
-  if (/\b(what is|how do|explain|why does|describe|compare)\b/i.test(lower)) return 'qa';
-  // Slack signals
-  if (/\b(slack|reply|respond|message|dm|channel)\b/i.test(lower)) return 'slack-reply';
-  return 'chat';
-}
-
-// --- Content hash for dedup ---
-
-function contentHash(source, prompt) {
-  return createHash('sha256').update(`${source}:${prompt}`).digest('hex').slice(0, 16);
-}
-
-// --- Claude Code Session Harvesting ---
-
-async function harvestClaudeCodeSessions(since) {
-  const claudeDir = path.join(process.env.HOME, '.claude', 'projects');
-  if (!fs.existsSync(claudeDir)) return [];
-
-  const samples = [];
-  const jsonlFiles = findJsonlFiles(claudeDir);
-
-  for (const file of jsonlFiles) {
-    const stat = fs.statSync(file);
-    if (since && stat.mtime < new Date(since)) continue;
-
-    try {
-      const lines = fs.readFileSync(file, 'utf8').split('\n').filter(Boolean);
-      const messages = lines.map(line => {
-        try { return JSON.parse(line); } catch { return null; }
-      }).filter(Boolean);
-
-      // Pair user->assistant turns
-      for (let i = 0; i < messages.length - 1; i++) {
-        const msg = messages[i];
-        const next = messages[i + 1];
-
-        if (msg.type !== 'human' && msg.role !== 'user') continue;
-        if (next.type !== 'assistant' && next.role !== 'assistant') continue;
-
-        const userContent = extractContent(msg);
-        const assistantContent = extractContent(next);
-        if (!userContent || userContent.length < 20) continue;
-        if (!assistantContent || assistantContent.length < 20) continue;
-
-        // Extract tool calls from assistant message
-        const toolCalls = extractToolCalls(next);
-
-        const id = contentHash('claude-code', userContent);
-        samples.push({
-          id,
-          source: 'claude-code',
-          session_id: path.basename(file, '.jsonl'),
-          timestamp: msg.timestamp || stat.mtime.toISOString(),
-          task_type: classifyTaskType(userContent),
-          prompt: userContent,
-          response: assistantContent,
-          tool_calls: toolCalls,
-          outcome: 'unknown',
-          outcome_signal: { git_committed: false, git_diff: null, task_status: null, user_corrected: false },
-          model: next.model || 'unknown',
-          quality_label: 0.5,
-        });
-      }
-    } catch (err) {
-      console.error(`[harvester] Error parsing ${file}:`, err.message);
-    }
-  }
-  return samples;
-}
-
-// --- Claude Desktop Session Harvesting ---
-
-async function harvestClaudeDesktopSessions(since) {
-  const reader = getClaudeDesktopSessions();
-  if (!reader) return [];
-
-  const sessions = reader.listSessions();
-  const samples = [];
-
-  for (const session of sessions) {
-    if (since && session.updatedAt && session.updatedAt <= since) continue;
-    const messages = Array.isArray(session.messages) ? session.messages : [];
-    for (let i = 0; i < messages.length - 1; i++) {
-      const userMsg = messages[i];
-      const assistantMsg = messages[i + 1];
-      if (userMsg.role !== 'user' || assistantMsg.role !== 'assistant') continue;
-      const userContent = userMsg.text || '';
-      const assistantContent = assistantMsg.text || '';
-      if (!userContent || userContent.length < 20) continue;
-      if (!assistantContent || assistantContent.length < 20) continue;
-
-      samples.push({
-        id: contentHash('claude-desktop', `${session.uuid}:${i}:${userContent}`),
-        source: 'claude-desktop',
-        session_id: session.uuid,
-        timestamp: userMsg.timestamp || session.updatedAt || session.createdAt || new Date().toISOString(),
-        task_type: classifyTaskType(userContent),
-        prompt: userContent,
-        response: assistantContent,
-        tool_calls: [],
-        outcome: 'unknown',
-        outcome_signal: { git_committed: false, git_diff: null, task_status: null, user_corrected: false },
-        model: session.model || 'unknown',
-        quality_label: 0.5,
-      });
-    }
-  }
-  return samples;
-}
-
-// --- Codex Session Harvesting ---
-
-async function harvestCodexSessions(since) {
-  const codexDir = path.join(process.env.HOME, '.codex', 'sessions');
-  if (!fs.existsSync(codexDir)) return [];
-
-  const samples = [];
-  const jsonlFiles = findJsonlFiles(codexDir);
-
-  for (const file of jsonlFiles) {
-    const stat = fs.statSync(file);
-    if (since && stat.mtime < new Date(since)) continue;
-
-    try {
-      const lines = fs.readFileSync(file, 'utf8').split('\n').filter(Boolean);
-      const messages = lines.map(line => {
-        try { return JSON.parse(line); } catch { return null; }
-      }).filter(Boolean);
-
-      // Codex format: response_item with role
-      const userMsgs = messages.filter(m => m.role === 'user' || (m.type === 'response_item' && m.item?.role === 'user'));
-      const assistantMsgs = messages.filter(m => m.role === 'assistant' || (m.type === 'response_item' && m.item?.role === 'assistant'));
-
-      // Pair them up sequentially
-      const pairs = Math.min(userMsgs.length, assistantMsgs.length);
-      for (let i = 0; i < pairs; i++) {
-        const userContent = extractContent(userMsgs[i]);
-        const assistantContent = extractContent(assistantMsgs[i]);
-        if (!userContent || userContent.length < 20) continue;
-        if (!assistantContent || assistantContent.length < 20) continue;
-
-        const toolCalls = extractToolCalls(assistantMsgs[i]);
-        const id = contentHash('codex', userContent);
-
-        samples.push({
-          id,
-          source: 'codex',
-          session_id: path.basename(file, '.jsonl'),
-          timestamp: userMsgs[i].timestamp || stat.mtime.toISOString(),
-          task_type: classifyTaskType(userContent),
-          prompt: userContent,
-          response: assistantContent,
-          tool_calls: toolCalls,
-          outcome: 'unknown',
-          outcome_signal: { git_committed: false, git_diff: null, task_status: null, user_corrected: false },
-          model: assistantMsgs[i].model || 'unknown',
-          quality_label: 0.5,
-        });
-      }
-    } catch (err) {
-      console.error(`[harvester] Error parsing ${file}:`, err.message);
-    }
-  }
-  return samples;
-}
-
-// --- CTM Session Harvesting ---
-
-async function harvestCtmSessions(since, dataDirOverride = null) {
-  const dataDir = dataDirOverride || process.env.WALL_E_DATA_DIR || path.join(process.env.HOME, '.walle', 'data');
-  const ctmDbPath = path.join(dataDir, 'task-manager.db');
-  if (!fs.existsSync(ctmDbPath)) return [];
-
-  let Database;
-  try { Database = require('better-sqlite3'); } catch { return []; }
-  let ctmDb;
-  try {
-    ctmDb = new Database(ctmDbPath, { readonly: true, fileMustExist: true });
-  } catch {
-    return [];
-  }
-
-  try {
-    let sql = `
-      SELECT u.session_id, u.message_text AS prompt, u.executed_at, u.project_path,
-             a.message_text AS response, a.tool_uses
-      FROM prompt_executions u
-      JOIN prompt_executions a ON a.session_id = u.session_id AND a.message_index = u.message_index + 1
-      WHERE u.role = 'user' AND a.role = 'assistant'
-        AND length(u.message_text) >= 20 AND length(a.message_text) >= 20
-    `;
-    if (since) { sql += ` AND u.executed_at > ?`; }
-    sql += ' ORDER BY u.executed_at DESC LIMIT 500';
-
-    const rows = since ? ctmDb.prepare(sql).all(since) : ctmDb.prepare(sql).all();
-    ctmDb.close();
-
-    return rows.map(row => ({
-      id: contentHash('ctm', row.prompt),
-      source: 'ctm-session',
-      session_id: row.session_id,
-      timestamp: row.executed_at,
-      task_type: classifyTaskType(row.prompt),
-      prompt: row.prompt,
-      response: row.response,
-      tool_calls: (() => { try { return JSON.parse(row.tool_uses || '[]'); } catch { return []; } })(),
-      outcome: 'unknown',
-      outcome_signal: { git_committed: false, git_diff: null, task_status: null, user_corrected: false },
-      model: 'claude',
-      quality_label: 0.7,
-    }));
-  } catch (err) {
-    try { ctmDb.close(); } catch {}
-    console.error('[harvester] Error reading CTM sessions:', err.message);
-    return [];
-  }
-}
-
-// --- Wall-E Chat Harvesting ---
-
-async function harvestWalleChat(brain, since) {
-  const messages = brain.listChatMessages ? brain.listChatMessages() : [];
-  const samples = [];
-
-  // Group by session
-  const sessions = new Map();
-  for (const msg of messages) {
-    if (since && msg.created_at && msg.created_at < since) continue;
-    const sid = msg.session_id || '__default__';
-    if (!sessions.has(sid)) sessions.set(sid, []);
-    sessions.get(sid).push(msg);
-  }
-
-  for (const [sid, msgs] of sessions) {
-    for (let i = 0; i < msgs.length - 1; i++) {
-      const userMsg = msgs[i];
-      const assistantMsg = msgs[i + 1];
-      if (userMsg.role !== 'user' || assistantMsg.role !== 'assistant') continue;
-      if (!userMsg.content || userMsg.content.length < 20) continue;
-      if (!assistantMsg.content || assistantMsg.content.length < 20) continue;
-      // Skip tool-use-only messages
-      if (isToolUseOnly(assistantMsg.content)) continue;
-
-      const id = contentHash('walle-chat', userMsg.content);
-      samples.push({
-        id,
-        source: 'walle-chat',
-        session_id: sid,
-        timestamp: userMsg.created_at || new Date().toISOString(),
-        task_type: classifyTaskType(userMsg.content),
-        prompt: userMsg.content,
-        response: assistantMsg.content,
-        tool_calls: [],
-        outcome: 'unknown',
-        outcome_signal: { git_committed: false, git_diff: null, task_status: null, user_corrected: false },
-        model: assistantMsg.model_id || 'unknown',
-        quality_label: 0.5,
-      });
-    }
-  }
-  return samples;
-}
-
-// --- Wall-E Task Harvesting ---
-
-async function harvestWalleTasks(brain, since) {
-  const tasks = brain.listTasks ? brain.listTasks({ status: 'completed' }) : [];
-  const samples = [];
-
-  for (const task of tasks) {
-    if (!task.result || !task.description) continue;
-    if (task.description.length < 20) continue;
-    if (since && task.created_at && task.created_at < since) continue;
-
-    const id = contentHash('walle-task', task.description);
-    const isSlack = task.source === 'slack';
-    const quality = isSlack ? 0.8 : 0.7;
-
-    samples.push({
-      id,
-      source: 'walle-task',
-      session_id: task.id,
-      timestamp: task.created_at || new Date().toISOString(),
-      task_type: isSlack ? 'slack-reply' : classifyTaskType(task.description),
-      prompt: task.description,
-      response: task.result,
-      tool_calls: [],
-      outcome: 'completed',
-      outcome_signal: { git_committed: false, git_diff: null, task_status: task.status, user_corrected: false },
-      model: 'unknown',
-      quality_label: quality,
-    });
-  }
-  return samples;
-}
-
-// --- Git Correlation ---
-
-function findCommitsInWindow(cwd, startTime, endTime) {
-  try {
-    const result = execFileSync(
-      'git', ['log', `--after=${startTime}`, `--before=${endTime}`, '--format=%H|%s|%aI'],
-      { cwd, encoding: 'utf8', timeout: 10000, stdio: ['pipe', 'pipe', 'pipe'] }
-    ).trim();
-    if (!result) return [];
-    return result.split('\n').map(line => {
-      const [hash, subject, date] = line.split('|');
-      return { hash, subject, date };
-    });
-  } catch {
-    return [];
-  }
-}
-
-function getCommitDiff(cwd, hash) {
-  try {
-    if (!/^[0-9a-f]{7,40}$/i.test(hash)) return null;
-    const result = execFileSync(
-      'git', ['show', '--stat', '--format=', hash],
-      { cwd, encoding: 'utf8', timeout: 10000, stdio: ['pipe', 'pipe', 'pipe'] }
-    ).trim();
-    return result;
-  } catch {
-    return null;
-  }
-}
-
-async function correlateGitOutcomes(samples, cwd) {
-  if (!cwd || !fs.existsSync(path.join(cwd, '.git'))) return samples;
-
-  for (const sample of samples) {
-    if (sample.source !== 'claude-code' && sample.source !== 'codex') continue;
-
-    const ts = new Date(sample.timestamp);
-    const startTime = ts.toISOString();
-    const endTime = new Date(ts.getTime() + 3600000).toISOString(); // +1 hour window
-
-    const commits = findCommitsInWindow(cwd, startTime, endTime);
-    if (commits.length > 0) {
-      sample.outcome = 'committed';
-      sample.outcome_signal.git_committed = true;
-      sample.outcome_signal.git_diff = getCommitDiff(cwd, commits[0].hash);
-      sample.quality_label = 1.0;
-    }
-  }
-  return samples;
-}
-
-// --- Helpers ---
-
-function findJsonlFiles(dir) {
-  const results = [];
-  try {
-    const entries = fs.readdirSync(dir, { withFileTypes: true, recursive: true });
-    for (const entry of entries) {
-      if (entry.isFile() && entry.name.endsWith('.jsonl')) {
-        // node 20+ returns parentPath, older returns path property
-        const parent = entry.parentPath || entry.path || dir;
-        results.push(path.join(parent, entry.name));
-      }
-    }
-  } catch {}
-  return results;
-}
-
-function extractContent(msg) {
-  if (typeof msg.content === 'string') return msg.content;
-  if (Array.isArray(msg.content)) {
-    const textParts = msg.content
-      .filter(b => b.type === 'text')
-      .map(b => b.text);
-    // Count user-attached non-text blocks (images, documents) so the
-    // replay path can warn the agent that input was incomplete. Without
-    // this marker, an agent replaying a multimodal session sees only
-    // the text portion of the user's prompt and goes off-target — saw
-    // this burn a real cc-replay (ba1efb8b: user attached an image
-    // showing a UI bug, walle got just "why auto approver missed this?"
-    // and edited the wrong file). tool_use/tool_result are assistant
-    // turn artifacts, not user attachments — don't count them.
-    const imageCount = msg.content.filter(b => b.type === 'image').length;
-    const docCount = msg.content.filter(b => b.type === 'document').length;
-    let text = textParts.join('\n');
-    if (imageCount > 0 || docCount > 0) {
-      const parts = [];
-      if (imageCount > 0) parts.push(`${imageCount} image${imageCount > 1 ? 's' : ''}`);
-      if (docCount > 0) parts.push(`${docCount} document${docCount > 1 ? 's' : ''}`);
-      const verb = (imageCount + docCount) > 1 ? 'are' : 'is';
-      text = `[NOTE: User originally attached ${parts.join(' + ')} that ${verb} NOT visible in this replay. The text below may reference visual content you cannot see (e.g. "this button", "the screenshot shows"). Make BEST-EFFORT PROGRESS on the textual prompt — investigate the codebase, find the most likely intent, and make a concrete change. State your assumption explicitly in the final summary. Do NOT freeze or refuse to edit; producing a wrong edit is more useful for review than no edit at all.]\n\n${text}`;
-    }
-    return text;
-  }
-  if (msg.message?.content) return extractContent(msg.message);
-  if (msg.item?.content) {
-    if (Array.isArray(msg.item.content)) {
-      return msg.item.content
-        .filter(b => b.type === 'input_text' || b.type === 'output_text' || b.type === 'text')
-        .map(b => b.text)
-        .join('\n');
-    }
-  }
-  return '';
-}
-
-function extractToolCalls(msg) {
-  if (Array.isArray(msg.content)) {
-    return msg.content
-      .filter(b => b.type === 'tool_use')
-      .map(b => ({ name: b.name, input: b.input }));
-  }
-  return [];
-}
-
-function isToolUseOnly(content) {
-  if (typeof content !== 'string') return false;
-  try {
-    const parsed = JSON.parse(content);
-    if (Array.isArray(parsed)) return parsed.every(b => b.type === 'tool_use');
-  } catch {}
-  return false;
-}
-
-// --- Coding Agent Session Harvesting ---
-
-function scoreCodingSession(session) {
-  let score = 0;
-  const toolCalls = session.tool_calls || [];
-  if (toolCalls.length > 3) score += 0.2;
-  if (session.git_committed) score += 0.3;
-  if ((session.turns || 0) > 5) score += 0.1;
-  // Check for review phase: tool calls that include grep/read after edit/write suggest review
-  const toolNames = toolCalls.map(t => typeof t === 'string' ? t : t.name || '');
-  const lastEditIdx = Math.max(
-    toolNames.lastIndexOf('edit_file'),
-    toolNames.lastIndexOf('write_file'),
-  );
-  const hasReviewAfterEdit = lastEditIdx >= 0 && toolNames.slice(lastEditIdx + 1).some(
-    n => /read_file|grep|test|npm/.test(n)
-  );
-  if (hasReviewAfterEdit) score += 0.1;
-  if ((session.files_modified || []).length > 1) score += 0.1;
-  if (!session.user_corrected) score += 0.2;
-  return Math.min(score, 1);
-}
-
-async function harvestCodingAgentSessions(since) {
-  const dataDir = process.env.WALL_E_DATA_DIR || path.join(process.env.HOME, '.walle', 'data');
-  const logDir = path.join(dataDir, 'coding-logs');
-  if (!fs.existsSync(logDir)) return [];
-
-  const sessions = [];
-  let files;
-  try {
-    files = fs.readdirSync(logDir).filter(f => f.endsWith('.jsonl'));
-  } catch { return []; }
-
-  for (const file of files) {
-    const filePath = path.join(logDir, file);
-    const stat = fs.statSync(filePath);
-    if (since && stat.mtime < new Date(since)) continue;
-
-    try {
-      const lines = fs.readFileSync(filePath, 'utf8').split('\n').filter(Boolean);
-      const turns = lines.map(line => {
-        try { return JSON.parse(line); } catch { return null; }
-      }).filter(Boolean);
-
-      if (turns.length === 0) continue;
-
-      // First turn content is the initial prompt
-      const firstTurn = turns[0];
-      const prompt = firstTurn.content || '';
-      if (!prompt || prompt.length < 20) continue;
-
-      // Collect all tool calls across turns
-      const allToolCalls = [];
-      const filesModified = new Set();
-      let hasUserCorrection = false;
-
-      for (const t of turns) {
-        const tcs = t.toolCalls || [];
-        for (const tc of tcs) {
-          allToolCalls.push(tc);
-          // Track files modified by write/edit tools
-          const name = tc.name || '';
-          const input = tc.input || {};
-          if (/edit_file|write_file|write_to_file/.test(name)) {
-            const fp = input.path || input.file_path || input.filePath || '';
-            if (fp) filesModified.add(fp);
-          }
-        }
-      }
-
-      // Check for git commits in the tool calls
-      const gitCommitted = allToolCalls.some(tc => {
-        const name = tc.name || '';
-        const input = tc.input || {};
-        const cmd = typeof input === 'string' ? input : input.command || '';
-        return name === 'run_shell' && /git\s+commit/.test(cmd);
-      });
-
-      // Extract git diff if available
-      let gitDiff = null;
-      const diffCall = allToolCalls.find(tc => {
-        const name = tc.name || '';
-        const input = tc.input || {};
-        const cmd = typeof input === 'string' ? input : input.command || '';
-        return name === 'run_shell' && /git\s+diff/.test(cmd);
-      });
-      if (diffCall && diffCall.output) gitDiff = diffCall.output;
-
-      const sessionId = path.basename(file, '.jsonl');
-      const id = contentHash('coding-agent', `${sessionId}:${prompt}`);
-
-      const session = {
-        id,
-        session_id: sessionId,
-        prompt,
-        tool_calls: allToolCalls,
-        turns: turns.length,
-        files_modified: [...filesModified],
-        git_committed: gitCommitted,
-        git_diff: gitDiff,
-        user_corrected: hasUserCorrection,
-        created_at: stat.mtime.toISOString(),
-      };
-
-      session.significance_score = scoreCodingSession(session);
-      sessions.push(session);
-    } catch (err) {
-      console.error(`[harvester] Error parsing coding log ${file}:`, err.message);
-    }
-  }
-
-  return sessions;
-}
-
-// --- Main Orchestrator ---
-
-async function runHarvest({ incremental = true, brain, dataDir } = {}) {
-  if (!brain) throw new Error('brain is required');
-  if (!dataDir) {
-    dataDir = process.env.WALL_E_DATA_DIR || path.join(process.env.HOME, '.walle', 'data');
-  }
-
-  const outputDir = path.join(dataDir, 'training', 'harvested');
-  if (!fs.existsSync(outputDir)) fs.mkdirSync(outputDir, { recursive: true });
-
-  const allSamples = [];
-  const stats = { total: 0, bySource: {}, byTaskType: {} };
-
-  // Determine "since" for incremental harvesting
-  const getSince = (source) => {
-    if (!incremental) return null;
-    const state = brain.getHarvestState(source);
-    return state?.last_processed_at || null;
-  };
-
-  // Harvest from each source
-  const claudeSamples = await harvestClaudeCodeSessions(getSince('claude-code'));
-  const claudeDesktopSamples = await harvestClaudeDesktopSessions(getSince('claude-desktop'));
-  const codexSamples = await harvestCodexSessions(getSince('codex'));
-  const chatSamples = await harvestWalleChat(brain, getSince('walle-chat'));
-  const taskSamples = await harvestWalleTasks(brain, getSince('walle-task'));
-  const ctmSamples = await harvestCtmSessions(getSince('ctm-sessions'), dataDir);
-
-  // Harvest coding agent sessions and store in brain
-  const codingAgentSessions = await harvestCodingAgentSessions(getSince('coding-agent'));
-  for (const session of codingAgentSessions) {
-    try {
-      const { classifyCodingType } = require('./benchmark-generator');
-      session.classified_type = classifyCodingType(session);
-    } catch { /* benchmark-generator not loaded yet — use default */ }
-    if (typeof brain.insertCodingSession === 'function') brain.insertCodingSession(session);
-  }
-
-  allSamples.push(...claudeSamples, ...claudeDesktopSamples, ...codexSamples, ...chatSamples, ...taskSamples, ...ctmSamples);
-
-  // Deduplicate by content hash
-  const seen = new Set();
-  const dedupedSamples = allSamples.filter(s => {
-    if (seen.has(s.id)) return false;
-    seen.add(s.id);
-    return true;
-  });
-
-  // Build stats
-  for (const s of dedupedSamples) {
-    stats.bySource[s.source] = (stats.bySource[s.source] || 0) + 1;
-    stats.byTaskType[s.task_type] = (stats.byTaskType[s.task_type] || 0) + 1;
-  }
-  stats.total = dedupedSamples.length;
-
-  // Write to output file (EAGAIN-safe — data dir may be on Dropbox)
-  if (dedupedSamples.length > 0) {
-    const today = new Date().toISOString().slice(0, 10);
-    const outputPath = path.join(outputDir, `${today}.jsonl`);
-    const jsonl = dedupedSamples.map(s => JSON.stringify(s)).join('\n') + '\n';
-    try {
-      fs.appendFileSync(outputPath, jsonl);
-    } catch (writeErr) {
-      console.error(`[harvester] Write failed (${writeErr.code || writeErr.message}), skipping persist`);
-    }
-  }
-
-  // Update harvest state for each source (EAGAIN-safe)
-  const now = new Date().toISOString();
-  try {
-    if (claudeSamples.length > 0) {
-      brain.updateHarvestState('claude-code', { lastProcessedAt: now, totalHarvested: claudeSamples.length });
-    }
-    if (claudeDesktopSamples.length > 0) {
-      brain.updateHarvestState('claude-desktop', { lastProcessedAt: now, totalHarvested: claudeDesktopSamples.length });
-    }
-    if (codexSamples.length > 0) {
-      brain.updateHarvestState('codex', { lastProcessedAt: now, totalHarvested: codexSamples.length });
-    }
-    if (chatSamples.length > 0) {
-      brain.updateHarvestState('walle-chat', { lastProcessedAt: now, totalHarvested: chatSamples.length });
-    }
-    if (taskSamples.length > 0) {
-      brain.updateHarvestState('walle-task', { lastProcessedAt: now, totalHarvested: taskSamples.length });
-    }
-    if (ctmSamples.length > 0) {
-      brain.updateHarvestState('ctm-sessions', { lastProcessedAt: now, totalHarvested: ctmSamples.length });
-    }
-    if (codingAgentSessions.length > 0) {
-      brain.updateHarvestState('coding-agent', { lastProcessedAt: now, totalHarvested: codingAgentSessions.length });
-    }
-  } catch (stateErr) {
-    console.error(`[harvester] State update failed (${stateErr.code || stateErr.message}), will retry next cycle`);
-  }
-
-  return { samples: dedupedSamples, stats, codingAgentSessions: codingAgentSessions.length };
-}
-
-module.exports = {
-  classifyTaskType,
-  contentHash,
-  harvestClaudeCodeSessions,
-  harvestClaudeDesktopSessions,
-  harvestCodexSessions,
-  harvestCtmSessions,
-  harvestWalleChat,
-  harvestWalleTasks,
-  harvestCodingAgentSessions,
-  scoreCodingSession,
-  correlateGitOutcomes,
-  findCommitsInWindow,
-  getCommitDiff,
-  runHarvest,
-  // Helpers (exported for testing)
-  extractContent,
-  extractToolCalls,
-  isToolUseOnly,
-  findJsonlFiles,
-};