forge-workflow 0.0.4 → 0.0.6

package/scripts/improve-command.js

@@ -0,0 +1,375 @@
+/**
+ * Semi-autonomous improvement loop: analyze eval failures, rewrite command,
+ * re-evaluate, and stop on regression or plateau.
+ *
+ * Usage:
+ *   bun scripts/improve-command.js <command-path> --eval-set <path> [--max-iterations <N>]
+ */
+
+const { execFileSync } = require('child_process');
+const fs = require('fs');
+const { DEFAULT_BASE_PATH, loadEvalHistory } = require('./lib/eval-storage');
+
+// ---------------------------------------------------------------------------
+// analyzeFailures
+// ---------------------------------------------------------------------------
+
+/**
+ * Extract failing assertions with their query context from an eval result.
+ *
+ * @param {object} evalResult - result from runEvalPipeline
+ * @returns {Array<{ query: string, assertion: string, reasoning: string }>}
+ */
+function analyzeFailures(evalResult) {
+  const failures = [];
+
+  for (const queryResult of evalResult.results) {
+    for (const assertion of queryResult.assertions) {
+      if (!assertion.pass) {
+        failures.push({
+          query: queryResult.prompt,
+          assertion: assertion.check,
+          reasoning: assertion.reasoning,
+        });
+      }
+    }
+  }
+
+  return failures;
+}
+
+// ---------------------------------------------------------------------------
+// buildRewritePrompt
+// ---------------------------------------------------------------------------
+
+/**
+ * Build a prompt asking for a command rewrite that fixes the identified failures.
+ *
+ * @param {string} commandContent - current command markdown content
+ * @param {Array<{ query: string, assertion: string, reasoning: string }>} failures
+ * @param {object[]} history - array of prior eval results (from loadEvalHistory)
+ * @returns {string}
+ */
+function buildRewritePrompt(commandContent, failures, history) {
+  let prompt = '';
+
+  prompt += 'You are a command prompt engineer. Rewrite the following command to fix the failing assertions.\n\n';
+
+  prompt += '## Current Command\n\n';
+  prompt += commandContent + '\n\n';
+
+  prompt += '## Failing Assertions\n\n';
+  for (const failure of failures) {
+    prompt += `- Query: "${failure.query}"\n`;
+    prompt += `  Assertion: "${failure.assertion}"\n`;
+    prompt += `  Reasoning: "${failure.reasoning}"\n\n`;
+  }
+
+  if (history && history.length > 0) {
+    prompt += '## Prior Eval Attempts\n\n';
+    prompt += 'Avoid repeating approaches that have already been tried. Here is a summary of prior attempts:\n\n';
+    for (let i = 0; i < history.length; i++) {
+      const attempt = history[i];
+      const failCount = attempt.results
+        ? attempt.results.filter((result) => result.assertions && result.assertions.some((assertion) => !assertion.pass)).length
+        : 0;
+      prompt += `- Attempt ${i + 1}: score ${attempt.overall_score}, ${failCount} failing queries\n`;
+    }
+    prompt += '\n';
+
+    // Detect flaky assertions: the same check both passed and failed across sessions.
+    const assertionOutcomes = new Map();
+    for (const attempt of history) {
+      if (!attempt.results) continue;
+      for (const result of attempt.results) {
+        if (!result.assertions) continue;
+        for (const assertion of result.assertions) {
+          if (!assertionOutcomes.has(assertion.check)) {
+            assertionOutcomes.set(assertion.check, { pass: 0, fail: 0 });
+          }
+          const entry = assertionOutcomes.get(assertion.check);
+          if (assertion.pass) {
+            entry.pass++;
+          } else {
+            entry.fail++;
+          }
+        }
+      }
+    }
+
+    const flakyAssertions = [];
+    for (const [check, outcomes] of assertionOutcomes) {
+      if (outcomes.pass > 0 && outcomes.fail > 0) {
+        flakyAssertions.push({ check, pass: outcomes.pass, fail: outcomes.fail });
+      }
+    }
+
+    if (flakyAssertions.length > 0) {
+      prompt += '## Flaky/Inconsistent Assertions\n\n';
+      prompt += 'These assertions are flaky - they pass in some sessions and fail in others. ';
+      prompt += 'Do not waste iterations on these; they may depend on environment rather than command quality.\n\n';
+      for (const flakyAssertion of flakyAssertions) {
+        prompt += `- "${flakyAssertion.check}" - passed ${flakyAssertion.pass}x, failed ${flakyAssertion.fail}x across sessions\n`;
+      }
+      prompt += '\n';
+    }
+  }
+
+  prompt += '## Instructions\n\n';
+  prompt += 'Return ONLY the rewritten command markdown content. No explanation, no code fences.\n';
+
+  return prompt;
+}
+
+// ---------------------------------------------------------------------------
+// generateDiff
+// ---------------------------------------------------------------------------
+
+/**
+ * Simple line-by-line diff showing added (+) and removed (-) lines.
+ *
+ * @param {string} original
+ * @param {string} modified
+ * @returns {string}
+ */
+function generateDiff(original, modified) {
+  const originalLines = original.split('\n');
+  const modifiedLines = modified.split('\n');
+
+  if (original === modified) {
+    return '';
+  }
+
+  const lines = [];
+  let i = 0;
+  let j = 0;
+
+  while (i < originalLines.length || j < modifiedLines.length) {
+    if (i < originalLines.length && j < modifiedLines.length && originalLines[i] === modifiedLines[j]) {
+      lines.push(' ' + originalLines[i]);
+      i++;
+      j++;
+    } else {
+      const originalLineInModified = j < modifiedLines.length ? modifiedLines.indexOf(originalLines[i], j) : -1;
+      const modifiedLineInOriginal = i < originalLines.length ? originalLines.indexOf(modifiedLines[j], i) : -1;
+
+      if (i >= originalLines.length) {
+        lines.push('+' + modifiedLines[j]);
+        j++;
+      } else if (j >= modifiedLines.length) {
+        lines.push('-' + originalLines[i]);
+        i++;
+      } else if (
+        originalLineInModified !== -1 &&
+        (modifiedLineInOriginal === -1 || originalLineInModified - j <= modifiedLineInOriginal - i)
+      ) {
+        lines.push('+' + modifiedLines[j]);
+        j++;
+      } else {
+        lines.push('-' + originalLines[i]);
+        i++;
+      }
+    }
+  }
+
+  return lines.join('\n');
+}
+
+/**
+ * Default rewrite invocation via `claude -p`.
+ *
+ * @param {string} prompt
+ * @param {{ timeout?: number, _execFileSync?: Function }} [options]
+ * @returns {Promise<string>}
+ */
+async function defaultRewriteCommand(prompt, options = {}) {
+  const timeout = options.timeout || 120_000;
+  const execFile = options._execFileSync || execFileSync;
+
+  return execFile(
+    'claude',
+    ['-p', prompt, '--output-format', 'text', '--no-session-persistence'],
+    {
+      encoding: 'utf-8',
+      timeout,
+      maxBuffer: 10 * 1024 * 1024,
+    }
+  );
+}
+
+// ---------------------------------------------------------------------------
+// runImprovementLoop
+// ---------------------------------------------------------------------------
+
+/**
+ * Main improvement loop orchestrator.
+ *
+ * @param {string} commandPath - path to the command markdown file
+ * @param {string} evalSetPath - path to the .eval.json file
+ * @param {object} [options]
+ * @param {number} [options.maxIterations=3]
+ * @param {Function} [options._runEval] - injectable eval function for testing
+ * @param {Function} [options._rewriteCommand] - injectable rewriter for testing
+ * @param {string} [options._basePath] - eval-logs base path for testing
+ * @returns {Promise<{ original: string, best: string, originalScore: number, bestScore: number, iterations: number, reason: string, diff: string }>}
+ */
+async function runImprovementLoop(commandPath, evalSetPath, options = {}) {
+  const maxIterations = options.maxIterations || 3;
+  const runEval = options._runEval;
+  const rewriteCommand = options._rewriteCommand || ((prompt) => defaultRewriteCommand(prompt, options));
+  const basePath = options._basePath || DEFAULT_BASE_PATH;
+
+  const originalContent = fs.readFileSync(commandPath, 'utf8');
+  let bestContent = originalContent;
+
+  try {
+    const baselineResult = await runEval(evalSetPath);
+    const originalScore = baselineResult.overall_score;
+    const history = loadEvalHistory(baselineResult.command, basePath);
+
+    let bestScore = originalScore;
+    let latestResult = baselineResult;
+    let previousScore = originalScore;
+    let iterations = 0;
+    let reason = 'max_iterations';
+
+    for (let iter = 1; iter <= maxIterations; iter++) {
+      iterations = iter;
+
+      const failures = analyzeFailures(latestResult);
+      const prompt = buildRewritePrompt(
+        fs.readFileSync(commandPath, 'utf8'),
+        failures,
+        history
+      );
+
+      const newContent = await rewriteCommand(prompt);
+      fs.writeFileSync(commandPath, newContent, 'utf8');
+
+      const newResult = await runEval(evalSetPath);
+      const newScore = newResult.overall_score;
+
+      if (newScore < bestScore) {
+        fs.writeFileSync(commandPath, bestContent, 'utf8');
+        reason = 'regression';
+        break;
+      }
+
+      if (newScore === previousScore) {
+        if (newScore >= bestScore) {
+          bestContent = newContent;
+          bestScore = newScore;
+        }
+        fs.writeFileSync(commandPath, bestContent, 'utf8');
+        reason = 'plateau';
+        break;
+      }
+
+      if (newScore > bestScore) {
+        bestContent = newContent;
+        bestScore = newScore;
+      }
+
+      previousScore = newScore;
+      latestResult = newResult;
+
+      if (iter === maxIterations) {
+        reason = 'max_iterations';
+      }
+    }
+
+    fs.writeFileSync(commandPath, bestContent, 'utf8');
+
+    return {
+      original: originalContent,
+      best: bestContent,
+      originalScore,
+      bestScore,
+      iterations,
+      reason,
+      diff: generateDiff(originalContent, bestContent),
+    };
+  } catch (err) {
+    try {
+      fs.writeFileSync(commandPath, bestContent, 'utf8');
+    } catch (_restoreErr) {
+      // Preserve the original failure if restoring also fails.
+    }
+    throw err;
+  }
+}
+
+// ---------------------------------------------------------------------------
+// CLI entry point
+// ---------------------------------------------------------------------------
+
+/**
+ * Parse CLI arguments.
+ *
+ * @param {string[]} argv - process.argv.slice(2)
+ * @returns {{ commandPath: string, evalSetPath: string, maxIterations: number }}
+ */
+function parseCliArgs(argv) {
+  let commandPath = null;
+  let evalSetPath = null;
+  let maxIterations = 3;
+
+  for (let i = 0; i < argv.length; i++) {
+    if (argv[i] === '--eval-set' && i + 1 < argv.length) {
+      evalSetPath = argv[i + 1];
+      i++;
+    } else if (argv[i] === '--max-iterations' && i + 1 < argv.length) {
+      maxIterations = Number(argv[i + 1]);
+      i++;
+    } else if (!argv[i].startsWith('--')) {
+      commandPath = argv[i];
+    }
+  }
+
+  if (!commandPath) {
+    throw new Error('Usage: improve-command <command-path> --eval-set <path> [--max-iterations <N>]');
+  }
+  if (!evalSetPath) {
+    throw new Error('Missing required --eval-set <path>');
+  }
+
+  return { commandPath, evalSetPath, maxIterations };
+}
+
+if (require.main === module) {
+  const args = parseCliArgs(process.argv.slice(2));
+  const { runEvalPipeline } = require('./run-command-eval');
+
+  runImprovementLoop(args.commandPath, args.evalSetPath, {
+    maxIterations: args.maxIterations,
+    _runEval: (evalPath) => runEvalPipeline(evalPath),
+  })
+    .then((result) => {
+      console.log('\n=== Improvement Summary ===');
+      console.log(`Original score: ${result.originalScore.toFixed(2)}`);
+      console.log(`Best score:     ${result.bestScore.toFixed(2)}`);
+      console.log(`Iterations:     ${result.iterations}`);
+      console.log(`Reason:         ${result.reason}`);
+
+      if (result.diff) {
+        console.log('\n=== Diff (original -> best) ===');
+        console.log(result.diff);
+      } else {
+        console.log('\nNo changes made.');
+      }
+
+      console.log('\nDiff shown above. Review and decide whether to apply.');
+    })
+    .catch((err) => {
+      console.error(`Error: ${err.message}`);
+      process.exit(1);
+    });
+}
+
+module.exports = {
+  analyzeFailures,
+  buildRewritePrompt,
+  defaultRewriteCommand,
+  generateDiff,
+  runImprovementLoop,
+};

package/scripts/lib/eval-runner.js

@@ -0,0 +1,268 @@
+/**
+ * Eval runner core — worktree isolation + command execution.
+ *
+ * Provides building blocks for the eval pipeline:
+ *   - createEvalWorktree()   — spin up an isolated worktree
+ *   - destroyEvalWorktree()  — tear it down (force, even if dirty)
+ *   - resetWorktree()        — reset between eval queries
+ *   - executeCommand()       — run a claude CLI command in a worktree
+ */
+
+const path = require('path');
+const { execSync } = require('node:child_process');
+
+// ── active worktree tracking (cleanup on crash) ─────────────────────
+// Tracks active eval worktrees so we can clean up on process exit/crash.
+// Prevents orphaned eval-* branches when interrupted.
+// Note: execSync is safe here — all paths are internally generated, never user input.
+const activeEvalWorktrees = new Map(); // path -> branch
+
+function cleanupActiveWorktrees() {
+  if (activeEvalWorktrees.size === 0) return;
+  let repoRoot;
+  try { repoRoot = getRepoRoot(); } catch (_err) { return; }
+  for (const [wtPath, branch] of activeEvalWorktrees) {
+    try {
+      execSync(`git worktree remove --force "${wtPath}"`, { cwd: repoRoot, stdio: 'pipe' });
+    } catch (_err) { /* already removed */ }
+    if (branch && branch.startsWith('eval-')) {
+      try {
+        execSync(`git branch -D "${branch}"`, { cwd: repoRoot, stdio: 'pipe' });
+      } catch (_err) { /* already deleted */ }
+    }
+  }
+  try { execSync('git worktree prune', { cwd: repoRoot, stdio: 'pipe' }); } catch (_err) { /* ignore */ }
+  activeEvalWorktrees.clear();
+}
+
+process.on('exit', cleanupActiveWorktrees);
+process.on('SIGINT', () => {
+  const hadWork = activeEvalWorktrees.size > 0;
+  cleanupActiveWorktrees();
+  if (hadWork) process.exit(130);
+});
+process.on('SIGTERM', () => {
+  const hadWork = activeEvalWorktrees.size > 0;
+  cleanupActiveWorktrees();
+  if (hadWork) process.exit(143);
+});
+
+// ── helpers ──────────────────────────────────────────────────────────
+
+/**
+ * Detect the repo root by walking up from cwd.
+ * Works from both the main repo and from within worktrees.
+ */
+function getRepoRoot() {
+  const root = execSync('git rev-parse --show-toplevel', {
+    encoding: 'utf-8',
+    stdio: ['pipe', 'pipe', 'pipe'],
+  }).trim();
+  return root;
+}
+
+/**
+ * Get the .worktrees directory path for eval worktrees.
+ * Eval worktrees live under <repo-root>/.worktrees/
+ */
+function getWorktreesDir() {
+  const root = getRepoRoot();
+  return path.join(root, '.worktrees');
+}
+
+// ── createEvalWorktree ───────────────────────────────────────────────
+
+/**
+ * Create a git worktree with a unique name for eval isolation.
+ *
+ * @returns {Promise<{ path: string, branch: string }>}
+ */
+async function createEvalWorktree() {
+  const timestamp = Date.now();
+  const pid = process.pid;
+  const name = `eval-${timestamp}-${pid}`;
+  const branch = `eval-${timestamp}-${pid}`;
+  const worktreesDir = getWorktreesDir();
+  const wtPath = path.join(worktreesDir, name);
+
+  // Create the worktree with a detached HEAD first, then create branch
+  execSync(`git worktree add -b "${branch}" "${wtPath}" HEAD`, {
+    cwd: getRepoRoot(),
+    encoding: 'utf-8',
+    stdio: ['pipe', 'pipe', 'pipe'],
+  });
+
+  activeEvalWorktrees.set(wtPath, branch);
+  return { path: wtPath, branch };
+}
+
+// ── destroyEvalWorktree ──────────────────────────────────────────────
+
+/**
+ * Remove a worktree and its temporary branch.
+ * Succeeds even if the worktree is dirty.
+ *
+ * @param {string} worktreePath — absolute path to the worktree
+ * @returns {Promise<void>}
+ */
+async function destroyEvalWorktree(worktreePath) {
+  const repoRoot = getRepoRoot();
+
+  // Query actual branch for this worktree (more reliable than inferring from dir name)
+  let branch;
+  try {
+    branch = execSync('git branch --show-current', {
+      cwd: worktreePath,
+      encoding: 'utf-8',
+      stdio: ['pipe', 'pipe', 'pipe'],
+    }).trim();
+  } catch (_err) {
+    // Worktree may be corrupted — fall back to directory name
+    branch = path.basename(worktreePath);
+  }
+
+  // Remove the worktree (--force handles dirty state)
+  execSync(`git worktree remove --force "${worktreePath}"`, {
+    cwd: repoRoot,
+    encoding: 'utf-8',
+    stdio: ['pipe', 'pipe', 'pipe'],
+  });
+
+  // Prune to clean up references
+  execSync('git worktree prune', {
+    cwd: repoRoot,
+    encoding: 'utf-8',
+    stdio: ['pipe', 'pipe', 'pipe'],
+  });
+
+  activeEvalWorktrees.delete(worktreePath);
+
+  // Delete the temporary branch (force in case it's not fully merged)
+  if (branch && branch.startsWith('eval-')) {
+    try {
+      execSync(`git branch -D "${branch}"`, {
+        cwd: repoRoot,
+        encoding: 'utf-8',
+        stdio: ['pipe', 'pipe', 'pipe'],
+      });
+    } catch (_err) {
+      // Branch may already be gone — ignore
+    }
+  }
+}
+
+// ── resetWorktree ────────────────────────────────────────────────────
+
+/**
+ * Reset a worktree to a clean state (tracked files restored, untracked removed).
+ *
+ * @param {string} worktreePath — absolute path to the worktree
+ * @returns {Promise<void>}
+ */
+async function resetWorktree(worktreePath) {
+  // Restore tracked files
+  execSync('git checkout -- .', {
+    cwd: worktreePath,
+    encoding: 'utf-8',
+    stdio: ['pipe', 'pipe', 'pipe'],
+  });
+
+  // Remove untracked files, directories, and ignored files (full reset between runs)
+  execSync('git clean -fdx', {
+    cwd: worktreePath,
+    encoding: 'utf-8',
+    stdio: ['pipe', 'pipe', 'pipe'],
+  });
+}
+
+// ── executeCommand ───────────────────────────────────────────────────
+
+/**
+ * Execute a command in an eval worktree.
+ *
+ * In production, runs `claude -p "<prompt>" --output-format stream-json --verbose --no-session-persistence`.
+ * Accepts an optional `cmdOverride` array for testing (avoids invoking real LLM).
+ *
+ * @param {string} _command — label only (e.g., "/status")
+ * @param {string} prompt — the prompt to send
+ * @param {string} worktreePath — absolute path to the worktree (used as cwd)
+ * @param {number} [timeout=120000] — timeout in milliseconds
+ * @param {string[]} [cmdOverride] — optional command array for testing
+ * @returns {Promise<{ stdout: string, stderr: string, exitCode: number, timedOut: boolean }>}
+ */
+async function executeCommand(_command, prompt, worktreePath, timeout = 120000, cmdOverride) {
+  // Build the command to run
+  const cmd = cmdOverride || [
+    'claude',
+    '-p',
+    prompt,
+    '--output-format',
+    'stream-json',
+    '--verbose',
+    '--no-session-persistence',
+  ];
+
+  // Build environment: inherit current env, strip CLAUDECODE, set FORGE_EVAL
+  const env = { ...process.env };
+  delete env.CLAUDECODE;
+  env.FORGE_EVAL = '1';
+
+  // Spawn with Bun.spawn (array form, no shell interpolation)
+  const proc = Bun.spawn(cmd, { // eslint-disable-line no-undef -- Bun global provided by runtime
+    cwd: worktreePath,
+    env,
+    stdout: 'pipe',
+    stderr: 'pipe',
+  });
+
+  let timedOut = false;
+  let timeoutId;
+
+  // Set up timeout
+  const timeoutPromise = new Promise((resolve) => {
+    timeoutId = setTimeout(() => {
+      timedOut = true;
+      proc.kill();
+      resolve();
+    }, timeout);
+  });
+
+  // Wait for process to exit (or timeout)
+  const exitPromise = proc.exited.then(() => {
+    clearTimeout(timeoutId);
+  });
+
+  await Promise.race([exitPromise, timeoutPromise]);
+
+  // Read stdout and stderr
+  let stdout = '';
+  let stderr = '';
+
+  try {
+    stdout = await new Response(proc.stdout).text();
+  } catch (_err) {
+    // stream may be closed on kill — ignore
+  }
+
+  try {
+    stderr = await new Response(proc.stderr).text();
+  } catch (_err) {
+    // stream may be closed on kill — ignore
+  }
+
+  const exitCode = timedOut ? (proc.exitCode ?? 1) : proc.exitCode;
+
+  return {
+    stdout,
+    stderr,
+    exitCode,
+    timedOut,
+  };
+}
+
+module.exports = {
+  createEvalWorktree,
+  destroyEvalWorktree,
+  resetWorktree,
+  executeCommand,
+};