@jackwener/opencli 1.6.1 → 1.6.2

package/autoresearch/commands/fix.ts

@@ -0,0 +1,145 @@
+#!/usr/bin/env npx tsx
+/**
+ * /autoresearch:fix — Iterative error elimination.
+ *
+ * Auto-detects broken state (build → test → browse tests) and iteratively
+ * fixes errors one at a time. Stops when error count reaches 0.
+ *
+ * Priority: build errors → test failures → browse task failures
+ *
+ * Usage:
+ *   npx tsx autoresearch/commands/fix.ts
+ *   npx tsx autoresearch/commands/fix.ts --iterations 10
+ */
+
+import { execSync } from 'node:child_process';
+import { join, dirname } from 'node:path';
+import { fileURLToPath } from 'node:url';
+import { parseArgs } from '../config.js';
+import { Engine, type ModifyContext } from '../engine.js';
+
+const __dirname = dirname(fileURLToPath(import.meta.url));
+const ROOT = join(__dirname, '..', '..');
+
+function exec(cmd: string): { ok: boolean; output: string } {
+  try {
+    const output = execSync(cmd, {
+      cwd: ROOT, timeout: 120_000, encoding: 'utf-8',
+      stdio: ['pipe', 'pipe', 'pipe'],
+    }).trim();
+    return { ok: true, output };
+  } catch (err: any) {
+    return { ok: false, output: (err.stdout ?? '') + '\n' + (err.stderr ?? '') };
+  }
+}
+
+/** Detect current broken state and return verify command + error count */
+function detectBrokenState(): { verify: string; errors: number; description: string } | null {
+  // 1. Build
+  const build = exec('npm run build 2>&1');
+  if (!build.ok) {
+    const errorCount = (build.output.match(/error TS/g) || []).length || 1;
+    return {
+      verify: 'npm run build 2>&1 | grep -c "error TS" || echo 0',
+      errors: errorCount,
+      description: `${errorCount} TypeScript build error(s)`,
+    };
+  }
+
+  // 2. Tests
+  const test = exec('npm test 2>&1');
+  if (!test.ok) {
+    const failMatch = test.output.match(/(\d+)\s+fail/i);
+    const errorCount = failMatch ? parseInt(failMatch[1], 10) : 1;
+    return {
+      verify: 'npm test 2>&1 | grep -oP "\\d+(?= fail)" || echo 0',
+      errors: errorCount,
+      description: `${errorCount} test failure(s)`,
+    };
+  }
+
+  // 3. Browse tests
+  const browse = exec('npx tsx autoresearch/eval-browse.ts 2>&1');
+  const scoreMatch = browse.output.match(/SCORE=(\d+)\/(\d+)/);
+  if (scoreMatch) {
+    const passed = parseInt(scoreMatch[1], 10);
+    const total = parseInt(scoreMatch[2], 10);
+    const failures = total - passed;
+    if (failures > 0) {
+      return {
+        verify: 'npx tsx autoresearch/eval-browse.ts 2>&1 | tail -1',
+        errors: failures,
+        description: `${failures} browse task failure(s) (${passed}/${total})`,
+      };
+    }
+  }
+
+  return null; // all clean
+}
+
+async function main() {
+  const args = parseArgs(process.argv.slice(2));
+  const maxIterations = args.iterations ?? 20;
+
+  console.log('\n🔧 AutoResearch Fix — Detecting broken state...\n');
+
+  const broken = detectBrokenState();
+  if (!broken) {
+    console.log('  ✓ All clean — nothing to fix!\n');
+    return;
+  }
+
+  console.log(`  Found: ${broken.description}`);
+  console.log(`  Verify: ${broken.verify}\n`);
+
+  const config = {
+    goal: `Fix all errors: ${broken.description}`,
+    scope: ['src/**/*.ts', 'extension/src/**/*.ts'],
+    metric: 'error_count',
+    direction: 'lower' as const,
+    verify: broken.verify,
+    guard: 'npm run build',
+    iterations: maxIterations,
+    minDelta: 1,
+  };
+
+  const logPath = join(ROOT, 'autoresearch-results.tsv');
+  const engine = new Engine(config, logPath, {
+    modify: async (ctx: ModifyContext) => {
+      const prompt = `Fix ONE error. Current error count: ${ctx.currentMetric}. Goal: 0 errors.
+
+Read the error output, understand the root cause, and make ONE focused fix.
+Do NOT fix multiple unrelated errors at once.
+Do NOT modify test files.
+
+${ctx.stuckHint ? `STUCK HINT: ${ctx.stuckHint}` : ''}`;
+
+      try {
+        const result = execSync(
+          `claude -p --dangerously-skip-permissions --allowedTools "Bash(npm:*),Bash(npx:*),Read,Edit,Write,Glob,Grep" --output-format text --no-session-persistence "${prompt.replace(/"/g, '\\"')}"`,
+          { cwd: ROOT, timeout: 180_000, encoding: 'utf-8', stdio: ['pipe', 'pipe', 'pipe'] }
+        ).trim();
+        const lines = result.split('\n').filter(l => l.trim());
+        return lines[lines.length - 1]?.trim()?.slice(0, 120) || 'fix attempt';
+      } catch {
+        return null;
+      }
+    },
+    onStatus: (msg) => console.log(msg),
+  });
+
+  try {
+    const results = await engine.run();
+    const finalMetric = results[results.length - 1]?.metric ?? broken.errors;
+    if (finalMetric === 0) {
+      console.log('\n✅ All errors fixed!\n');
+    } else {
+      console.log(`\n⚠ ${finalMetric} error(s) remaining after ${maxIterations} iterations.\n`);
+    }
+  } catch (err: any) {
+    console.error(`\n❌ ${err.message}`);
+    process.exit(1);
+  }
+}
+
+main();

package/autoresearch/commands/plan.ts

@@ -0,0 +1,88 @@
+#!/usr/bin/env npx tsx
+/**
+ * /autoresearch:plan — Interactive configuration wizard.
+ *
+ * Walks through goal, scope, metric, verify, guard settings
+ * and outputs a ready-to-paste run command.
+ *
+ * Usage:
+ *   npx tsx autoresearch/commands/plan.ts
+ */
+
+import { execSync } from 'node:child_process';
+import { createInterface } from 'node:readline';
+import { join, dirname } from 'node:path';
+import { fileURLToPath } from 'node:url';
+import { PRESETS } from '../presets/index.js';
+
+const __dirname = dirname(fileURLToPath(import.meta.url));
+const ROOT = join(__dirname, '..', '..');
+
+const rl = createInterface({ input: process.stdin, output: process.stdout });
+const ask = (q: string): Promise<string> => new Promise(r => rl.question(q, r));
+
+async function main() {
+  console.log('\n🔬 AutoResearch — Configuration Wizard\n');
+
+  // Offer presets first
+  const presetNames = Object.keys(PRESETS);
+  console.log('Available presets:');
+  presetNames.forEach((name, i) => {
+    console.log(`  [${i + 1}] ${name} — ${PRESETS[name].goal}`);
+  });
+  console.log(`  [0] Custom config\n`);
+
+  const choice = await ask('Choose preset or 0 for custom: ');
+  const idx = parseInt(choice, 10);
+
+  if (idx > 0 && idx <= presetNames.length) {
+    const name = presetNames[idx - 1];
+    const iterations = await ask('Iterations (empty = unbounded): ');
+    const iterFlag = iterations ? ` --iterations ${iterations}` : '';
+    console.log(`\n✅ Ready to run:\n`);
+    console.log(`  npx tsx autoresearch/commands/run.ts --preset ${name}${iterFlag}\n`);
+    rl.close();
+    return;
+  }
+
+  // Custom config
+  const goal = await ask('Goal (what to improve): ');
+  const scope = await ask('Scope (file globs, comma-separated): ');
+  const metric = await ask('Metric name (e.g. pass_count, coverage): ');
+  const direction = await ask('Direction (higher/lower): ') as 'higher' | 'lower';
+  const verify = await ask('Verify command (must output a number): ');
+
+  // Dry-run verify
+  console.log('\n  Dry-running verify command...');
+  try {
+    const output = execSync(verify, { cwd: ROOT, timeout: 120_000, encoding: 'utf-8', stdio: ['pipe', 'pipe', 'pipe'] }).trim();
+    const { extractMetric } = await import('../config.js');
+    const value = extractMetric(output);
+    if (value != null) {
+      console.log(`  ✓ Verify works — current ${metric}: ${value}`);
+    } else {
+      console.log(`  ⚠ Verify ran but no number extracted from output:\n    ${output.slice(0, 200)}`);
+    }
+  } catch (err: any) {
+    console.log(`  ✗ Verify failed: ${err.message?.slice(0, 100)}`);
+  }
+
+  const guard = await ask('Guard command (optional, press Enter to skip): ');
+  const iterations = await ask('Iterations (empty = unbounded): ');
+
+  const parts = ['npx tsx autoresearch/commands/run.ts'];
+  parts.push(`--goal "${goal}"`);
+  parts.push(`--scope "${scope}"`);
+  parts.push(`--metric "${metric}"`);
+  parts.push(`--direction ${direction}`);
+  parts.push(`--verify "${verify}"`);
+  if (guard) parts.push(`--guard "${guard}"`);
+  if (iterations) parts.push(`--iterations ${iterations}`);
+
+  console.log(`\n✅ Ready to run:\n`);
+  console.log(`  ${parts.join(' \\\n    ')}\n`);
+
+  rl.close();
+}
+
+main();

package/autoresearch/commands/run.ts

@@ -0,0 +1,138 @@
+#!/usr/bin/env npx tsx
+/**
+ * /autoresearch — Main autonomous iteration loop.
+ *
+ * Usage:
+ *   npx tsx autoresearch/commands/run.ts --preset operate-reliability
+ *   npx tsx autoresearch/commands/run.ts --preset operate-reliability --iterations 5
+ *   npx tsx autoresearch/commands/run.ts --goal "..." --scope "src/*.ts" --verify "..." --iterations 10
+ *
+ * The modify callback spawns Claude Code to make ONE atomic change per iteration.
+ * Engine handles commit, verify, guard, keep/discard, and logging.
+ */
+
+import { execSync } from 'node:child_process';
+import { join, dirname } from 'node:path';
+import { fileURLToPath } from 'node:url';
+import { parseArgs, type AutoResearchConfig } from '../config.js';
+import { Engine, type ModifyContext } from '../engine.js';
+import { PRESETS } from '../presets/index.js';
+
+const __dirname = dirname(fileURLToPath(import.meta.url));
+const ROOT = join(__dirname, '..', '..');
+
+function buildModifyPrompt(ctx: ModifyContext, config: AutoResearchConfig): string {
+  const recent = ctx.recentLog.slice(-10).map(r =>
+    `  ${r.status.padEnd(12)} ${r.description}`
+  ).join('\n');
+
+  return `You are an autonomous improvement agent. Make ONE atomic change to improve this metric.
+
+## Goal
+${config.goal}
+
+## Current State
+- Metric (${config.metric}): ${ctx.currentMetric} (best: ${ctx.bestMetric})
+- Iteration: ${ctx.iteration}
+- Consecutive discards: ${ctx.consecutiveDiscards}
+${ctx.stuckHint ? `\n## STUCK — Try a Different Approach\n${ctx.stuckHint}` : ''}
+
+## Recent History
+${recent || '  (no history yet)'}
+
+## Git Log (recent experiments)
+${ctx.gitLog.split('\n').slice(0, 10).join('\n')}
+
+## Scope (files you can modify)
+${ctx.scopeFiles.join('\n')}
+
+## Rules
+1. Make ONE atomic change (one logical intent, even if multiple files)
+2. Read the failing test output or code BEFORE modifying
+3. DO NOT modify test files or the verify command
+4. Describe what you changed in one sentence (no "and" linking unrelated actions)
+5. If previous approach was discarded, try something DIFFERENT
+6. Focus on the specific failures — read error messages carefully`;
+}
+
+async function modify(ctx: ModifyContext, config: AutoResearchConfig): Promise<string | null> {
+  const prompt = buildModifyPrompt(ctx, config);
+
+  console.log('  Claude Code making a change...');
+  try {
+    const result = execSync(
+      `claude -p --dangerously-skip-permissions --allowedTools "Bash(npm:*),Bash(npx:*),Bash(git:*),Read,Edit,Write,Glob,Grep" --output-format text --no-session-persistence "${prompt.replace(/"/g, '\\"')}"`,
+      {
+        cwd: ROOT,
+        timeout: 300_000,
+        encoding: 'utf-8',
+        stdio: ['pipe', 'pipe', 'pipe'],
+        env: process.env,
+      }
+    ).trim();
+
+    // Extract description from Claude's response (last non-empty line or summary)
+    const lines = result.split('\n').filter(l => l.trim());
+    const desc = lines[lines.length - 1]?.trim() || 'change made by Claude Code';
+    return desc.slice(0, 120);
+  } catch (err: any) {
+    console.error('  Claude Code failed:', err.message?.slice(0, 100));
+    return null;
+  }
+}
+
+async function main() {
+  const args = parseArgs(process.argv.slice(2));
+
+  // Resolve config from preset or CLI args
+  let config: AutoResearchConfig;
+  if (args.preset) {
+    config = PRESETS[args.preset];
+    if (!config) {
+      console.error(`Unknown preset: ${args.preset}`);
+      console.error(`Available: ${Object.keys(PRESETS).join(', ')}`);
+      process.exit(1);
+    }
+    // Allow CLI overrides
+    if (args.iterations != null) config = { ...config, iterations: args.iterations };
+    if (args.guard != null) config = { ...config, guard: args.guard };
+  } else if (args.goal && args.verify) {
+    config = {
+      goal: args.goal,
+      scope: args.scope ?? ['src/**/*.ts'],
+      metric: args.metric ?? 'score',
+      direction: args.direction ?? 'higher',
+      verify: args.verify,
+      guard: args.guard,
+      iterations: args.iterations,
+      minDelta: args.minDelta,
+    };
+  } else {
+    console.error('Usage: npx tsx autoresearch/commands/run.ts --preset <name> [--iterations N]');
+    console.error('   or: npx tsx autoresearch/commands/run.ts --goal "..." --verify "..." --scope "..."');
+    console.error(`\nAvailable presets: ${Object.keys(PRESETS).join(', ')}`);
+    process.exit(1);
+  }
+
+  console.log(`\n🔬 AutoResearch: ${config.goal}`);
+  console.log(`   Metric: ${config.metric} (${config.direction})`);
+  console.log(`   Verify: ${config.verify}`);
+  console.log(`   Guard:  ${config.guard ?? '(none)'}`);
+  console.log(`   Iterations: ${config.iterations ?? '∞'}`);
+  console.log('');
+
+  const logPath = join(ROOT, 'autoresearch-results.tsv');
+  const engine = new Engine(config, logPath, {
+    modify: (ctx) => modify(ctx, config),
+    onStatus: (msg) => console.log(msg),
+  });
+
+  try {
+    await engine.run();
+  } catch (err: any) {
+    console.error(`\n❌ ${err.message}`);
+    process.exit(1);
+  }
+}
+
+main();

package/autoresearch/config.ts

@@ -0,0 +1,82 @@
+/**
+ * AutoResearch Configuration — type definitions and CLI parsing.
+ *
+ * Based on Karpathy's autoresearch: constraint + mechanical metric + unbounded loop.
+ */
+
+export interface AutoResearchConfig {
+  /** Plain-language goal, e.g. "Increase operate pass rate to 59/59" */
+  goal: string;
+  /** Glob patterns for files the agent can modify */
+  scope: string[];
+  /** What the metric measures, e.g. "pass_count" */
+  metric: string;
+  /** Whether improvement means the number goes up or down */
+  direction: 'higher' | 'lower';
+  /** Shell command that outputs a number (the metric value) */
+  verify: string;
+  /** Optional guard command — must pass for a keep decision */
+  guard?: string;
+  /** Max iterations (undefined = unbounded) */
+  iterations?: number;
+  /** Minimum delta to count as real improvement (noise filter) */
+  minDelta?: number;
+}
+
+export type IterationStatus =
+  | 'baseline'
+  | 'keep'
+  | 'keep (reworked)'
+  | 'discard'
+  | 'crash'
+  | 'no-op'
+  | 'hook-blocked';
+
+export interface IterationResult {
+  iteration: number;
+  commit: string;
+  metric: number;
+  delta: number;
+  guard: 'pass' | 'fail' | '-';
+  status: IterationStatus;
+  description: string;
+}
+
+/** Parse CLI args into a partial config (missing fields filled by preset or prompts) */
+export function parseArgs(argv: string[]): Partial<AutoResearchConfig> & { preset?: string; task?: string } {
+  const config: Partial<AutoResearchConfig> & { preset?: string; task?: string } = {};
+  for (let i = 0; i < argv.length; i++) {
+    const arg = argv[i];
+    const next = argv[i + 1];
+    switch (arg) {
+      case '--preset': config.preset = next; i++; break;
+      case '--goal': config.goal = next; i++; break;
+      case '--scope': config.scope = next?.split(','); i++; break;
+      case '--metric': config.metric = next; i++; break;
+      case '--direction': config.direction = next as 'higher' | 'lower'; i++; break;
+      case '--verify': config.verify = next; i++; break;
+      case '--guard': config.guard = next; i++; break;
+      case '--iterations': config.iterations = parseInt(next, 10); i++; break;
+      case '--min-delta': config.minDelta = parseFloat(next); i++; break;
+      case '--task': config.task = next; i++; break;
+    }
+  }
+  return config;
+}
+
+/** Extract a number from command output using common patterns */
+export function extractMetric(output: string): number | null {
+  // Try: last line that looks like a number
+  const lines = output.trim().split('\n');
+  for (let i = lines.length - 1; i >= 0; i--) {
+    const line = lines[i].trim();
+    // Match standalone numbers: "56", "95.2", "SCORE=56/59" → 56
+    const scoreMatch = line.match(/SCORE[=:]\s*(\d+)/i);
+    if (scoreMatch) return parseFloat(scoreMatch[1]);
+    const numMatch = line.match(/^[\d.]+$/);
+    if (numMatch) return parseFloat(numMatch[0]);
+  }
+  // Fallback: first number in output
+  const fallback = output.match(/(\d+(?:\.\d+)?)/);
+  return fallback ? parseFloat(fallback[1]) : null;
+}