analyzthis_design 2.0.0 → 2.1.0

package/dist/lib/chunk-synthesis.js

@@ -1 +1,176 @@
-'use strict';var chunkModels=require('./chunk-models'),chunkRouter=require('./chunk-router'),synthesis=require('./synthesis');function buildSynthesisPrompt(a,b,c){var d=['You\x20are\x20the\x20synthesis\x20engine.\x20Merge\x20the\x20following\x20persona\x20chunk\x20outputs\x20into\x20a\x20final\x20verdict.','','Task:\x20'+(c['task']||a['task_summary']||''),'','Chunk\x20outputs:'];for(var e=0x0;e<b['length'];e++){var f=b[e];d['push']('---\x20'+f['chunk_id']+'\x20('+f['persona']+')\x20---'),d['push'](f['output']['slice'](0x0,0x4b0));}return d['push']('','Produce:','-\x20Verdict:\x20SHIP\x20|\x20REVISE\x20|\x20BLOCK','-\x20Composite\x20score\x20out\x20of\x205\x20if\x20applicable','-\x20Top\x203\x20actionable\x20changes','-\x20Brief\x20reasoning','','Use\x20the\x20same\x20format\x20as\x20the\x20design-critic\x20synthesis.'),{'system':'You\x20synthesize\x20multi-persona\x20design\x20critique\x20outputs.\x20Be\x20concise\x20and\x20cite\x20the\x20most\x20important\x20conflicts.','user':d['join']('\x0a')};}function hasContradictions(a){var b=[];for(var c=0x0;c<a['length'];c++){var d=a[c]['structured_output']&&a[c]['structured_output']['verdict'];if(d)b['push'](d);}if(b['length']<0x2)return![];var e=b['every'](function(f){return f===b[0x0];});return!e;}async function synthesize(a){a=a||{};var b=a['plan'],c=a['results']||[],d=a['contextPack'],f=a['config']||chunkModels['loadConfig'](),g=a['budget']||'auto',h=a['callLlmFn'];if(!h){var i=require('./orchestrator/run');h=i['callLlm'];}var j=c['filter'](function(t){return t['success'];});if(!j['length'])return{'verdict':'BLOCK','total':0x0,'max_total':0x0,'top3':[{'change':'All\x20chunk\x20models\x20failed','owner':'user','reason':'No\x20model\x20produced\x20output'}],'composite':{'verdict':'BLOCK','total':0x0,'max_total':0x0,'top3':[]},'markdown':'##\x20Synthesis\x0a\x0aAll\x20chunk\x20models\x20failed.\x20No\x20verdict\x20possible.','model_used':null,'contradictions':![]};var k=hasContradictions(j),l=b['synthesis']||{'persona':'raj','goal':'resolve\x20conflicts','system_prompt':'','user_prompt':''},m=buildSynthesisPrompt(b,j,d),n;try{var o=await chunkRouter['resolveChunkModel']({'effort':'hard','chunkType':'synthesis','budget':g,'config':f});n=o['selected'];}catch(t){n=chunkModels['resolvePlannerModel'](f);}k&&(n=chunkModels['resolvePlannerModel'](f));var p='';try{p=await h({'provider':n['provider'],'model':n['model'],'system':m['system'],'user':m['user'],'maxTokens':0x708,'personaId':l['persona']||'raj'});}catch(u){return fallbackSynthesis(j);}var q=Math['ceil']((m['system']['length']+m['user']['length'])/0x4),r=Math['ceil'](p['length']/0x4),s=synthesis['buildSynthesis']?synthesis['buildSynthesis'](j):null;if(!s)s=fallbackSynthesis(j);return{'verdict':s['composite']['verdict']||'REVISE','total':s['composite']['total']||0x0,'max_total':s['composite']['max_total']||0x0,'top3':s['composite']['top3']||[],'composite':s['composite'],'markdown':s['markdown']||p,'model_used':{'provider':n['provider'],'model':n['model'],'cost':n['cost']},'contradictions':k,'tokens':{'input':q,'output':r},'cost':chunkModels['estimateCost'](n,q,r)};}function fallbackSynthesis(a){var b=['##\x20Synthesis\x0a'],c='REVISE',d=0x0,e=0x0;for(var f=0x0;f<a['length'];f++){var g=a[f];b['push']('###\x20'+g['persona']),b['push'](g['output']['slice'](0x0,0x1f4));g['structured_output']&&g['structured_output']['score']&&(d+=g['structured_output']['score'],e++);if(g['structured_output']&&g['structured_output']['verdict']){if(g['structured_output']['verdict']==='BLOCK')c='BLOCK';else{if(g['structured_output']['verdict']==='REVISE'&&c==='SHIP')c='REVISE';}}}var h=e?d/e:0x0;return b['push']('','**Verdict:**\x20'+c,'**Score:**\x20'+h['toFixed'](0x1)+'/5'),{'composite':{'verdict':c,'total':Math['round'](h),'max_total':0x5,'top3':[]},'markdown':b['join']('\x0a')};}module['exports']={'synthesize':synthesize,'buildSynthesisPrompt':buildSynthesisPrompt,'hasContradictions':hasContradictions};
+'use strict';
+
+/**
+ * Chunk synthesis (v2.0).
+ *
+ * Merges chunk outputs into a final verdict. Uses the cheapest capable model
+ * from the chunk pool. Escalates to the planner-tier model if outputs
+ * contradict or if chunk success rate is low.
+ *
+ * CommonJS, 'use strict', var.
+ */
+
+var chunkModels = require('./chunk-models');
+var chunkRouter = require('./chunk-router');
+var synthesis = require('./synthesis');
+
+function buildSynthesisPrompt(plan, results, contextPack) {
+  var lines = [
+    'You are the synthesis engine. Merge the following persona chunk outputs into a final verdict.',
+    '',
+    'Task: ' + (contextPack.task || plan.task_summary || ''),
+    '',
+    'Chunk outputs:',
+  ];
+
+  for (var i = 0; i < results.length; i++) {
+    var r = results[i];
+    lines.push('--- ' + r.chunk_id + ' (' + r.persona + ') ---');
+    lines.push(r.output.slice(0, 1200));
+  }
+
+  lines.push('',
+    'Produce:',
+    '- Verdict: SHIP | REVISE | BLOCK',
+    '- Composite score out of 5 if applicable',
+    '- Top 3 actionable changes',
+    '- Brief reasoning',
+    '',
+    'Use the same format as the design-critic synthesis.'
+  );
+
+  return {
+    system: 'You synthesize multi-persona design critique outputs. Be concise and cite the most important conflicts.',
+    user: lines.join('\n'),
+  };
+}
+
+function hasContradictions(results) {
+  var verdicts = [];
+  for (var i = 0; i < results.length; i++) {
+    var v = results[i].structured_output && results[i].structured_output.verdict;
+    if (v) verdicts.push(v);
+  }
+  if (verdicts.length < 2) return false;
+  var allSame = verdicts.every(function(v) { return v === verdicts[0]; });
+  return !allSame;
+}
+
+async function synthesize(opts) {
+  opts = opts || {};
+  var plan = opts.plan;
+  var results = opts.results || [];
+  var contextPack = opts.contextPack;
+  var config = opts.config || chunkModels.loadConfig();
+  var budget = opts.budget || 'auto';
+  var callLlmFn = opts.callLlmFn;
+
+  if (!callLlmFn) {
+    var orchestrator = require('./orchestrator/run');
+    callLlmFn = orchestrator.callLlm;
+  }
+
+  var successful = results.filter(function(r) { return r.success; });
+  if (!successful.length) {
+    return {
+      verdict: 'BLOCK',
+      total: 0,
+      max_total: 0,
+      top3: [{ change: 'All chunk models failed', owner: 'user', reason: 'No model produced output' }],
+      composite: { verdict: 'BLOCK', total: 0, max_total: 0, top3: [] },
+      markdown: '## Synthesis\n\nAll chunk models failed. No verdict possible.',
+      model_used: null,
+      contradictions: false,
+    };
+  }
+
+  var contradictions = hasContradictions(successful);
+  var synChunk = plan.synthesis || { persona: 'raj', goal: 'resolve conflicts', system_prompt: '', user_prompt: '' };
+
+  // Try cheap synthesis first.
+  var prompts = buildSynthesisPrompt(plan, successful, contextPack);
+  var model;
+  try {
+    var resolved = await chunkRouter.resolveChunkModel({
+      effort: 'hard',
+      chunkType: 'synthesis',
+      budget: budget,
+      config: config,
+    });
+    model = resolved.selected;
+  } catch (e) {
+    // Fall back to planner-tier.
+    model = chunkModels.resolvePlannerModel(config);
+  }
+
+  if (contradictions) {
+    // Escalate to planner-tier for arbitration.
+    model = chunkModels.resolvePlannerModel(config);
+  }
+
+  var text = '';
+  try {
+    text = await callLlmFn({
+      provider: model.provider,
+      model: model.model,
+      system: prompts.system,
+      user: prompts.user,
+      maxTokens: 1800,
+      personaId: synChunk.persona || 'raj',
+    });
+  } catch (e) {
+    // Last-resort: try to use the existing synthesis module on raw text.
+    return fallbackSynthesis(successful);
+  }
+
+  var inTok = Math.ceil((prompts.system.length + prompts.user.length) / 4);
+  var outTok = Math.ceil(text.length / 4);
+
+  var syn = synthesis.buildSynthesis ? synthesis.buildSynthesis(successful) : null;
+  if (!syn) syn = fallbackSynthesis(successful);
+
+  return {
+    verdict: syn.composite.verdict || 'REVISE',
+    total: syn.composite.total || 0,
+    max_total: syn.composite.max_total || 0,
+    top3: syn.composite.top3 || [],
+    composite: syn.composite,
+    markdown: syn.markdown || text,
+    model_used: { provider: model.provider, model: model.model, cost: model.cost },
+    contradictions: contradictions,
+    tokens: { input: inTok, output: outTok },
+    cost: chunkModels.estimateCost(model, inTok, outTok),
+  };
+}
+
+function fallbackSynthesis(results) {
+  var lines = ['## Synthesis\n'];
+  var verdict = 'REVISE';
+  var score = 0;
+  var count = 0;
+  for (var i = 0; i < results.length; i++) {
+    var r = results[i];
+    lines.push('### ' + r.persona);
+    lines.push(r.output.slice(0, 500));
+    if (r.structured_output && r.structured_output.score) {
+      score += r.structured_output.score;
+      count++;
+    }
+    if (r.structured_output && r.structured_output.verdict) {
+      if (r.structured_output.verdict === 'BLOCK') verdict = 'BLOCK';
+      else if (r.structured_output.verdict === 'REVISE' && verdict === 'SHIP') verdict = 'REVISE';
+    }
+  }
+  var avgScore = count ? (score / count) : 0;
+  lines.push('', '**Verdict:** ' + verdict, '**Score:** ' + avgScore.toFixed(1) + '/5');
+  return {
+    composite: { verdict: verdict, total: Math.round(avgScore), max_total: 5, top3: [] },
+    markdown: lines.join('\n'),
+  };
+}
+
+module.exports = {
+  synthesize: synthesize,
+  buildSynthesisPrompt: buildSynthesisPrompt,
+  hasContradictions: hasContradictions,
+};

package/dist/lib/chunk-telemetry.js

@@ -1 +1,88 @@
-'use strict';var fs=require('fs'),path=require('path'),os=require('os'),TELEMETRY_FILE=path['join'](os['homedir'](),'.analyzthis_design','chunk-telemetry.jsonl'),MAX_LINES=0x2710;function ensureDir(){fs['mkdirSync'](path['dirname'](TELEMETRY_FILE),{'recursive':!![]});}function loadLines(){if(!fs['existsSync'](TELEMETRY_FILE))return[];try{var a=fs['readFileSync'](TELEMETRY_FILE,'utf8');return a['trim']()['split']('\x0a')['filter'](function(b){return b['trim']();});}catch(b){return[];}}function append(a){ensureDir();var b=JSON['stringify'](a);fs['appendFileSync'](TELEMETRY_FILE,b+'\x0a');var c=loadLines();if(c['length']>MAX_LINES){var d=c['slice'](Math['floor'](MAX_LINES/0x2));fs['writeFileSync'](TELEMETRY_FILE,d['join']('\x0a')+'\x0a');}}function record(a){var b=Object['assign']({'at':new Date()['toISOString']()},a);append(b);}function getStatsForChunkType(a){var b=loadLines(),c={};for(var d=0x0;d<b['length'];d++){try{var f=JSON['parse'](b[d]);if(!f['model'])continue;var g=f['model'];if(!c[g])c[g]={'model':g,'attempts':0x0,'successful':0x0,'total_cost':0x0};c[g]['attempts']++;if(f['success'])c[g]['successful']++;c[g]['total_cost']+=f['cost']||0x0;}catch(h){}}return Object['keys'](c)['map'](function(j){return c[j];});}function getRecentFailures(a){a=a||0x14;var b=loadLines(),c=[];for(var d=b['length']-0x1;d>=0x0&&c['length']<a;d--){try{var f=JSON['parse'](b[d]);if(!f['success'])c['push'](f);}catch(g){}}return c;}module['exports']={'record':record,'getStatsForChunkType':getStatsForChunkType,'getRecentFailures':getRecentFailures,'TELEMETRY_FILE':TELEMETRY_FILE};
+'use strict';
+
+/**
+ * Chunk telemetry (v2.0).
+ *
+ * Records per-chunk model performance so the router can learn which
+ * model+persona combinations work best.
+ *
+ * CommonJS, 'use strict', var, try/catch around JSON.
+ */
+
+var fs = require('fs');
+var path = require('path');
+var os = require('os');
+
+var TELEMETRY_FILE = path.join(os.homedir(), '.analyzthis_design', 'chunk-telemetry.jsonl');
+var MAX_LINES = 10000;
+
+function ensureDir() {
+  fs.mkdirSync(path.dirname(TELEMETRY_FILE), { recursive: true });
+}
+
+function loadLines() {
+  if (!fs.existsSync(TELEMETRY_FILE)) return [];
+  try {
+    var text = fs.readFileSync(TELEMETRY_FILE, 'utf8');
+    return text.trim().split('\n').filter(function(l) { return l.trim(); });
+  } catch (e) { return []; }
+}
+
+function append(record) {
+  ensureDir();
+  var line = JSON.stringify(record);
+  fs.appendFileSync(TELEMETRY_FILE, line + '\n');
+
+  // Simple rotation: if file gets too large, keep last half.
+  var lines = loadLines();
+  if (lines.length > MAX_LINES) {
+    var keep = lines.slice(Math.floor(MAX_LINES / 2));
+    fs.writeFileSync(TELEMETRY_FILE, keep.join('\n') + '\n');
+  }
+}
+
+function record(record) {
+  var r = Object.assign({
+    at: new Date().toISOString(),
+  }, record);
+  append(r);
+}
+
+function getStatsForChunkType(chunkType) {
+  var lines = loadLines();
+  var byModel = {};
+
+  for (var i = 0; i < lines.length; i++) {
+    try {
+      var r = JSON.parse(lines[i]);
+      if (!r.model) continue;
+      var key = r.model;
+      if (!byModel[key]) byModel[key] = { model: key, attempts: 0, successful: 0, total_cost: 0 };
+      byModel[key].attempts++;
+      if (r.success) byModel[key].successful++;
+      byModel[key].total_cost += r.cost || 0;
+    } catch (e) { /* skip corrupt line */ }
+  }
+
+  return Object.keys(byModel).map(function(k) { return byModel[k]; });
+}
+
+function getRecentFailures(limit) {
+  limit = limit || 20;
+  var lines = loadLines();
+  var failures = [];
+  for (var i = lines.length - 1; i >= 0 && failures.length < limit; i--) {
+    try {
+      var r = JSON.parse(lines[i]);
+      if (!r.success) failures.push(r);
+    } catch (e) { /* skip */ }
+  }
+  return failures;
+}
+
+module.exports = {
+  record: record,
+  getStatsForChunkType: getStatsForChunkType,
+  getRecentFailures: getRecentFailures,
+  TELEMETRY_FILE: TELEMETRY_FILE,
+};