agent-gauntlet 0.10.0 → 0.11.0

package/dist/scripts/status.js

@@ -0,0 +1,280 @@
+#!/usr/bin/env node
+
+// src/scripts/status.ts
+import fs from "node:fs";
+import path from "node:path";
+function parseKeyValue(text) {
+  const result = {};
+  for (const match of text.matchAll(/(\w+)=(\S+)/g)) {
+    const key = match[1];
+    const value = match[2];
+    if (key && value)
+      result[key] = value;
+  }
+  return result;
+}
+function parseTimestamp(line) {
+  const m = line.match(/^\[([^\]]+)\]/);
+  return m?.[1] ?? "";
+}
+function parseEventType(line) {
+  const m = line.match(/^\[[^\]]+\]\s+(\S+)/);
+  return m?.[1] ?? "";
+}
+function parseEventBody(line) {
+  const m = line.match(/^\[[^\]]+\]\s+\S+\s*(.*)/);
+  return m?.[1] ?? "";
+}
+function parseDebugLog(content, sessionStartTime) {
+  const lines = content.split(`
+`).filter((l) => l.trim());
+  const sessions = [];
+  let current = null;
+  for (const line of lines) {
+    const event = parseEventType(line);
+    const body = parseEventBody(line);
+    const ts = parseTimestamp(line);
+    switch (event) {
+      case "RUN_START": {
+        if (sessionStartTime && new Date(ts) < sessionStartTime) {
+          current = null;
+          break;
+        }
+        const kv = parseKeyValue(body);
+        current = {
+          start: {
+            timestamp: ts,
+            mode: kv.mode ?? "unknown",
+            baseRef: kv.base_ref,
+            filesChanged: Number(kv.files_changed ?? kv.changes ?? 0),
+            linesAdded: Number(kv.lines_added ?? 0),
+            linesRemoved: Number(kv.lines_removed ?? 0),
+            gates: Number(kv.gates ?? 0)
+          },
+          gates: []
+        };
+        sessions.push(current);
+        break;
+      }
+      case "GATE_RESULT": {
+        if (!current)
+          break;
+        const gateIdMatch = body.match(/^(\S+)/);
+        const kv = parseKeyValue(body);
+        current.gates.push({
+          timestamp: ts,
+          gateId: gateIdMatch?.[1] ?? "unknown",
+          cli: kv.cli,
+          status: kv.status ?? "unknown",
+          duration: kv.duration ?? "?",
+          violations: kv.violations !== undefined ? Number(kv.violations) : undefined
+        });
+        break;
+      }
+      case "RUN_END": {
+        if (!current)
+          break;
+        const kv = parseKeyValue(body);
+        current.end = {
+          timestamp: ts,
+          status: kv.status ?? "unknown",
+          fixed: Number(kv.fixed ?? 0),
+          skipped: Number(kv.skipped ?? 0),
+          failed: Number(kv.failed ?? 0),
+          iterations: Number(kv.iterations ?? 0),
+          duration: kv.duration ?? "?"
+        };
+        break;
+      }
+      case "STOP_HOOK": {
+        if (!current)
+          break;
+        const kv = parseKeyValue(body);
+        current.stopHook = {
+          timestamp: ts,
+          decision: kv.decision ?? "unknown",
+          reason: kv.reason ?? "unknown"
+        };
+        break;
+      }
+    }
+  }
+  return sessions;
+}
+function getSessionStartTime(logDir) {
+  const entries = fs.readdirSync(logDir).filter((f) => !f.startsWith(".") && f !== "previous");
+  let earliest;
+  for (const entry of entries) {
+    const mtime = fs.statSync(path.join(logDir, entry)).mtimeMs;
+    if (earliest === undefined || mtime < earliest) {
+      earliest = mtime;
+    }
+  }
+  return earliest !== undefined ? new Date(earliest) : undefined;
+}
+function formatFileInventory(logDir) {
+  const lines = [];
+  const entries = fs.readdirSync(logDir).filter((f) => !f.startsWith(".") && f !== "previous");
+  if (entries.length === 0)
+    return lines;
+  const checks = [];
+  const reviews = [];
+  const other = [];
+  for (const entry of entries.sort()) {
+    const fullPath = path.join(logDir, entry);
+    const stat = fs.statSync(fullPath);
+    const sizeKB = (stat.size / 1024).toFixed(1);
+    const line = `- ${fullPath} (${sizeKB} KB)`;
+    if (entry.startsWith("review_")) {
+      reviews.push(line);
+    } else if (entry.startsWith("check_")) {
+      checks.push(line);
+    } else {
+      other.push(line);
+    }
+  }
+  lines.push("### Log Files");
+  lines.push("");
+  if (checks.length > 0) {
+    lines.push("**Check logs:**");
+    lines.push(...checks);
+  }
+  if (reviews.length > 0) {
+    lines.push("**Review logs/JSON:**");
+    lines.push(...reviews);
+  }
+  if (other.length > 0) {
+    lines.push("**Other:**");
+    lines.push(...other);
+  }
+  lines.push("");
+  return lines;
+}
+function formatStatusLine(end) {
+  return end.status === "pass" ? "PASSED" : end.status === "fail" ? "FAILED" : end.status.toUpperCase();
+}
+function formatAllRuns(sessions) {
+  const lines = [];
+  lines.push("### All Runs in Session");
+  lines.push("");
+  for (let i = 0;i < sessions.length; i++) {
+    const s = sessions[i];
+    if (!s)
+      continue;
+    const status = s.end ? s.end.status : "in-progress";
+    const duration = s.end ? s.end.duration : "?";
+    lines.push(`${i + 1}. [${s.start.timestamp}] mode=${s.start.mode} status=${status} duration=${duration}`);
+  }
+  lines.push("");
+  return lines;
+}
+function formatSession(sessions, logDir) {
+  if (sessions.length === 0) {
+    return "No gauntlet runs found in logs.";
+  }
+  const lastComplete = [...sessions].reverse().find((s) => s.end);
+  const session = lastComplete ?? sessions[sessions.length - 1];
+  if (!session)
+    return "No gauntlet runs found in logs.";
+  const lines = [];
+  lines.push("## Gauntlet Session Summary");
+  lines.push("");
+  if (session.end) {
+    lines.push(`**Status:** ${formatStatusLine(session.end)}`);
+    lines.push(`**Iterations:** ${session.end.iterations}`);
+    lines.push(`**Duration:** ${session.end.duration}`);
+    lines.push(`**Fixed:** ${session.end.fixed} | **Skipped:** ${session.end.skipped} | **Failed:** ${session.end.failed}`);
+  } else {
+    lines.push("**Status:** In Progress (no RUN_END found)");
+  }
+  lines.push("");
+  lines.push("### Diff Stats");
+  lines.push(`- Mode: ${session.start.mode}`);
+  if (session.start.baseRef) {
+    lines.push(`- Base ref: ${session.start.baseRef}`);
+  }
+  lines.push(`- Files changed: ${session.start.filesChanged}`);
+  lines.push(`- Lines: +${session.start.linesAdded} / -${session.start.linesRemoved}`);
+  lines.push(`- Gates: ${session.start.gates}`);
+  lines.push("");
+  lines.push("### Gate Results");
+  lines.push("");
+  lines.push("| Gate | CLI | Status | Duration | Violations |");
+  lines.push("|------|-----|--------|----------|------------|");
+  for (const gate of session.gates) {
+    const violations = gate.violations !== undefined ? String(gate.violations) : "-";
+    const statusIcon = gate.status === "pass" ? "pass" : "FAIL";
+    lines.push(`| ${gate.gateId} | ${gate.cli ?? "-"} | ${statusIcon} | ${gate.duration} | ${violations} |`);
+  }
+  lines.push("");
+  if (session.stopHook) {
+    lines.push("### Stop Hook");
+    lines.push(`- Decision: ${session.stopHook.decision}`);
+    lines.push(`- Reason: ${session.stopHook.reason}`);
+    lines.push("");
+  }
+  lines.push(...formatFileInventory(logDir));
+  if (sessions.length > 1) {
+    lines.push(...formatAllRuns(sessions));
+  }
+  return lines.join(`
+`);
+}
+function getLogDir(cwd) {
+  const configPath = path.join(cwd, ".gauntlet", "config.yml");
+  try {
+    const content = fs.readFileSync(configPath, "utf-8");
+    const match = content.match(/^log_dir:\s*(.+)$/m);
+    if (match?.[1])
+      return match[1].trim();
+  } catch {}
+  return "gauntlet_logs";
+}
+function resolveLogPaths(activeDir) {
+  const previousDir = path.join(activeDir, "previous");
+  const debugLogPath = path.join(activeDir, ".debug.log");
+  const activeHasLogs = fs.existsSync(activeDir) && fs.readdirSync(activeDir).some((f) => !f.startsWith(".") && f !== "previous");
+  if (activeHasLogs) {
+    return { logDir: activeDir, debugLogPath };
+  }
+  if (!fs.existsSync(previousDir)) {
+    console.log("No gauntlet_logs directory found.");
+    return null;
+  }
+  const logDir = resolvePreviousLogDir(previousDir);
+  if (!logDir)
+    return null;
+  return { logDir, debugLogPath };
+}
+function resolvePreviousLogDir(previousDir) {
+  const prevEntries = fs.readdirSync(previousDir);
+  const hasDirectFiles = prevEntries.some((f) => f.endsWith(".log") || f.endsWith(".json"));
+  if (hasDirectFiles)
+    return previousDir;
+  const prevDirs = prevEntries.map((d) => path.join(previousDir, d)).filter((d) => fs.statSync(d).isDirectory()).sort().reverse();
+  if (prevDirs.length === 0) {
+    console.log("No gauntlet logs found.");
+    return null;
+  }
+  return prevDirs[0];
+}
+function main() {
+  const cwd = process.cwd();
+  const logDirName = getLogDir(cwd);
+  const activeDir = path.join(cwd, logDirName);
+  const paths = resolveLogPaths(activeDir);
+  if (!paths) {
+    process.exit(0);
+  }
+  let sessions = [];
+  if (fs.existsSync(paths.debugLogPath)) {
+    const debugContent = fs.readFileSync(paths.debugLogPath, "utf-8");
+    const sessionStart = getSessionStartTime(paths.logDir);
+    sessions = parseDebugLog(debugContent, sessionStart);
+  }
+  const output = formatSession(sessions, paths.logDir);
+  console.log(output);
+}
+main();
+
+//# debugId=4BED01809E659CB464756E2164756E21

package/dist/scripts/status.js.map

@@ -0,0 +1,10 @@
+{
+  "version": 3,
+  "sources": ["../src/scripts/status.ts"],
+  "sourcesContent": [
+    "#!/usr/bin/env node\n/**\n * Gauntlet Status Script\n *\n * Parses the configured log_dir (default: gauntlet_logs/) to produce a structured\n * summary of the most recent gauntlet session from the .debug.log, plus a file\n * inventory of all log/JSON files for further inspection.\n *\n * This script handles structured data only (debug log events). Detailed failure\n * analysis (reading individual check logs, review JSONs) is left to the caller\n * (the /gauntlet-status skill) since log formats vary by check type.\n */\n\nimport fs from \"node:fs\";\nimport path from \"node:path\";\n\n// --- Types ---\n\ninterface RunStart {\n\ttimestamp: string;\n\tmode: string;\n\tbaseRef?: string;\n\tfilesChanged: number;\n\tlinesAdded: number;\n\tlinesRemoved: number;\n\tgates: number;\n}\n\ninterface GateResult {\n\ttimestamp: string;\n\tgateId: string;\n\tcli?: string;\n\tstatus: string;\n\tduration: string;\n\tviolations?: number;\n}\n\ninterface RunEnd {\n\ttimestamp: string;\n\tstatus: string;\n\tfixed: number;\n\tskipped: number;\n\tfailed: number;\n\titerations: number;\n\tduration: string;\n}\n\ninterface StopHookEntry {\n\ttimestamp: string;\n\tdecision: string;\n\treason: string;\n}\n\ninterface SessionRun {\n\tstart: RunStart;\n\tgates: GateResult[];\n\tend?: RunEnd;\n\tstopHook?: StopHookEntry;\n}\n\n// --- Parsing helpers ---\n\nfunction parseKeyValue(text: string): Record<string, string> {\n\tconst result: Record<string, string> = {};\n\tfor (const match of text.matchAll(/(\\w+)=(\\S+)/g)) {\n\t\tconst key = match[1];\n\t\tconst value = match[2];\n\t\tif (key && value) result[key] = value;\n\t}\n\treturn result;\n}\n\nfunction parseTimestamp(line: string): string {\n\tconst m = line.match(/^\\[([^\\]]+)\\]/);\n\treturn m?.[1] ?? \"\";\n}\n\nfunction parseEventType(line: string): string {\n\tconst m = line.match(/^\\[[^\\]]+\\]\\s+(\\S+)/);\n\treturn m?.[1] ?? \"\";\n}\n\nfunction parseEventBody(line: string): string {\n\tconst m = line.match(/^\\[[^\\]]+\\]\\s+\\S+\\s*(.*)/);\n\treturn m?.[1] ?? \"\";\n}\n\n// --- Debug log parsing ---\n\nfunction parseDebugLog(content: string, sessionStartTime?: Date): SessionRun[] {\n\tconst lines = content.split(\"\\n\").filter((l) => l.trim());\n\tconst sessions: SessionRun[] = [];\n\tlet current: SessionRun | null = null;\n\n\tfor (const line of lines) {\n\t\tconst event = parseEventType(line);\n\t\tconst body = parseEventBody(line);\n\t\tconst ts = parseTimestamp(line);\n\n\t\tswitch (event) {\n\t\t\tcase \"RUN_START\": {\n\t\t\t\t// Skip runs that predate the current session's log files\n\t\t\t\tif (sessionStartTime && new Date(ts) < sessionStartTime) {\n\t\t\t\t\tcurrent = null;\n\t\t\t\t\tbreak;\n\t\t\t\t}\n\t\t\t\tconst kv = parseKeyValue(body);\n\t\t\t\tcurrent = {\n\t\t\t\t\tstart: {\n\t\t\t\t\t\ttimestamp: ts,\n\t\t\t\t\t\tmode: kv.mode ?? \"unknown\",\n\t\t\t\t\t\tbaseRef: kv.base_ref,\n\t\t\t\t\t\tfilesChanged: Number(kv.files_changed ?? kv.changes ?? 0),\n\t\t\t\t\t\tlinesAdded: Number(kv.lines_added ?? 0),\n\t\t\t\t\t\tlinesRemoved: Number(kv.lines_removed ?? 0),\n\t\t\t\t\t\tgates: Number(kv.gates ?? 0),\n\t\t\t\t\t},\n\t\t\t\t\tgates: [],\n\t\t\t\t};\n\t\t\t\tsessions.push(current);\n\t\t\t\tbreak;\n\t\t\t}\n\t\t\tcase \"GATE_RESULT\": {\n\t\t\t\tif (!current) break;\n\t\t\t\tconst gateIdMatch = body.match(/^(\\S+)/);\n\t\t\t\tconst kv = parseKeyValue(body);\n\t\t\t\tcurrent.gates.push({\n\t\t\t\t\ttimestamp: ts,\n\t\t\t\t\tgateId: gateIdMatch?.[1] ?? \"unknown\",\n\t\t\t\t\tcli: kv.cli,\n\t\t\t\t\tstatus: kv.status ?? \"unknown\",\n\t\t\t\t\tduration: kv.duration ?? \"?\",\n\t\t\t\t\tviolations:\n\t\t\t\t\t\tkv.violations !== undefined ? Number(kv.violations) : undefined,\n\t\t\t\t});\n\t\t\t\tbreak;\n\t\t\t}\n\t\t\tcase \"RUN_END\": {\n\t\t\t\tif (!current) break;\n\t\t\t\tconst kv = parseKeyValue(body);\n\t\t\t\tcurrent.end = {\n\t\t\t\t\ttimestamp: ts,\n\t\t\t\t\tstatus: kv.status ?? \"unknown\",\n\t\t\t\t\tfixed: Number(kv.fixed ?? 0),\n\t\t\t\t\tskipped: Number(kv.skipped ?? 0),\n\t\t\t\t\tfailed: Number(kv.failed ?? 0),\n\t\t\t\t\titerations: Number(kv.iterations ?? 0),\n\t\t\t\t\tduration: kv.duration ?? \"?\",\n\t\t\t\t};\n\t\t\t\tbreak;\n\t\t\t}\n\t\t\tcase \"STOP_HOOK\": {\n\t\t\t\tif (!current) break;\n\t\t\t\tconst kv = parseKeyValue(body);\n\t\t\t\tcurrent.stopHook = {\n\t\t\t\t\ttimestamp: ts,\n\t\t\t\t\tdecision: kv.decision ?? \"unknown\",\n\t\t\t\t\treason: kv.reason ?? \"unknown\",\n\t\t\t\t};\n\t\t\t\tbreak;\n\t\t\t}\n\t\t}\n\t}\n\n\treturn sessions;\n}\n\n/**\n * Find the earliest mtime of non-hidden log files in the directory.\n * This marks the start of the current session.\n */\nfunction getSessionStartTime(logDir: string): Date | undefined {\n\tconst entries = fs\n\t\t.readdirSync(logDir)\n\t\t.filter((f) => !f.startsWith(\".\") && f !== \"previous\");\n\tlet earliest: number | undefined;\n\tfor (const entry of entries) {\n\t\tconst mtime = fs.statSync(path.join(logDir, entry)).mtimeMs;\n\t\tif (earliest === undefined || mtime < earliest) {\n\t\t\tearliest = mtime;\n\t\t}\n\t}\n\treturn earliest !== undefined ? new Date(earliest) : undefined;\n}\n\n// --- File inventory ---\n\nfunction formatFileInventory(logDir: string): string[] {\n\tconst lines: string[] = [];\n\tconst entries = fs\n\t\t.readdirSync(logDir)\n\t\t.filter((f) => !f.startsWith(\".\") && f !== \"previous\");\n\tif (entries.length === 0) return lines;\n\n\tconst checks: string[] = [];\n\tconst reviews: string[] = [];\n\tconst other: string[] = [];\n\n\tfor (const entry of entries.sort()) {\n\t\tconst fullPath = path.join(logDir, entry);\n\t\tconst stat = fs.statSync(fullPath);\n\t\tconst sizeKB = (stat.size / 1024).toFixed(1);\n\t\tconst line = `- ${fullPath} (${sizeKB} KB)`;\n\n\t\tif (entry.startsWith(\"review_\")) {\n\t\t\treviews.push(line);\n\t\t} else if (entry.startsWith(\"check_\")) {\n\t\t\tchecks.push(line);\n\t\t} else {\n\t\t\tother.push(line);\n\t\t}\n\t}\n\n\tlines.push(\"### Log Files\");\n\tlines.push(\"\");\n\tif (checks.length > 0) {\n\t\tlines.push(\"**Check logs:**\");\n\t\tlines.push(...checks);\n\t}\n\tif (reviews.length > 0) {\n\t\tlines.push(\"**Review logs/JSON:**\");\n\t\tlines.push(...reviews);\n\t}\n\tif (other.length > 0) {\n\t\tlines.push(\"**Other:**\");\n\t\tlines.push(...other);\n\t}\n\tlines.push(\"\");\n\n\treturn lines;\n}\n\n// --- Summary output ---\n\nfunction formatStatusLine(end: RunEnd): string {\n\treturn end.status === \"pass\"\n\t\t? \"PASSED\"\n\t\t: end.status === \"fail\"\n\t\t\t? \"FAILED\"\n\t\t\t: end.status.toUpperCase();\n}\n\nfunction formatAllRuns(sessions: SessionRun[]): string[] {\n\tconst lines: string[] = [];\n\tlines.push(\"### All Runs in Session\");\n\tlines.push(\"\");\n\tfor (let i = 0; i < sessions.length; i++) {\n\t\tconst s = sessions[i];\n\t\tif (!s) continue;\n\t\tconst status = s.end ? s.end.status : \"in-progress\";\n\t\tconst duration = s.end ? s.end.duration : \"?\";\n\t\tlines.push(\n\t\t\t`${i + 1}. [${s.start.timestamp}] mode=${s.start.mode} status=${status} duration=${duration}`,\n\t\t);\n\t}\n\tlines.push(\"\");\n\treturn lines;\n}\n\nfunction formatSession(sessions: SessionRun[], logDir: string): string {\n\tif (sessions.length === 0) {\n\t\treturn \"No gauntlet runs found in logs.\";\n\t}\n\n\tconst lastComplete = [...sessions].reverse().find((s) => s.end);\n\tconst session = lastComplete ?? sessions[sessions.length - 1];\n\tif (!session) return \"No gauntlet runs found in logs.\";\n\n\tconst lines: string[] = [];\n\n\t// Header\n\tlines.push(\"## Gauntlet Session Summary\");\n\tlines.push(\"\");\n\n\t// Overall status\n\tif (session.end) {\n\t\tlines.push(`**Status:** ${formatStatusLine(session.end)}`);\n\t\tlines.push(`**Iterations:** ${session.end.iterations}`);\n\t\tlines.push(`**Duration:** ${session.end.duration}`);\n\t\tlines.push(\n\t\t\t`**Fixed:** ${session.end.fixed} | **Skipped:** ${session.end.skipped} | **Failed:** ${session.end.failed}`,\n\t\t);\n\t} else {\n\t\tlines.push(\"**Status:** In Progress (no RUN_END found)\");\n\t}\n\tlines.push(\"\");\n\n\t// Diff stats\n\tlines.push(\"### Diff Stats\");\n\tlines.push(`- Mode: ${session.start.mode}`);\n\tif (session.start.baseRef) {\n\t\tlines.push(`- Base ref: ${session.start.baseRef}`);\n\t}\n\tlines.push(`- Files changed: ${session.start.filesChanged}`);\n\tlines.push(\n\t\t`- Lines: +${session.start.linesAdded} / -${session.start.linesRemoved}`,\n\t);\n\tlines.push(`- Gates: ${session.start.gates}`);\n\tlines.push(\"\");\n\n\t// Gate results\n\tlines.push(\"### Gate Results\");\n\tlines.push(\"\");\n\tlines.push(\"| Gate | CLI | Status | Duration | Violations |\");\n\tlines.push(\"|------|-----|--------|----------|------------|\");\n\tfor (const gate of session.gates) {\n\t\tconst violations =\n\t\t\tgate.violations !== undefined ? String(gate.violations) : \"-\";\n\t\tconst statusIcon = gate.status === \"pass\" ? \"pass\" : \"FAIL\";\n\t\tlines.push(\n\t\t\t`| ${gate.gateId} | ${gate.cli ?? \"-\"} | ${statusIcon} | ${gate.duration} | ${violations} |`,\n\t\t);\n\t}\n\tlines.push(\"\");\n\n\t// Stop hook\n\tif (session.stopHook) {\n\t\tlines.push(\"### Stop Hook\");\n\t\tlines.push(`- Decision: ${session.stopHook.decision}`);\n\t\tlines.push(`- Reason: ${session.stopHook.reason}`);\n\t\tlines.push(\"\");\n\t}\n\n\t// File inventory\n\tlines.push(...formatFileInventory(logDir));\n\n\t// All sessions summary (if multiple runs)\n\tif (sessions.length > 1) {\n\t\tlines.push(...formatAllRuns(sessions));\n\t}\n\n\treturn lines.join(\"\\n\");\n}\n\n// --- Main ---\n\n/**\n * Read the configured log_dir from .gauntlet/config.yml.\n * Falls back to \"gauntlet_logs\" if not found.\n */\nfunction getLogDir(cwd: string): string {\n\tconst configPath = path.join(cwd, \".gauntlet\", \"config.yml\");\n\ttry {\n\t\tconst content = fs.readFileSync(configPath, \"utf-8\");\n\t\tconst match = content.match(/^log_dir:\\s*(.+)$/m);\n\t\tif (match?.[1]) return match[1].trim();\n\t} catch {\n\t\t// Config not found — use default\n\t}\n\treturn \"gauntlet_logs\";\n}\n\n/**\n * Resolve the log directory and debug log path.\n * Returns null if no logs are found (after printing a message).\n */\nfunction resolveLogPaths(\n\tactiveDir: string,\n): { logDir: string; debugLogPath: string } | null {\n\tconst previousDir = path.join(activeDir, \"previous\");\n\tconst debugLogPath = path.join(activeDir, \".debug.log\");\n\n\t// Check active directory first for non-debug log files\n\tconst activeHasLogs =\n\t\tfs.existsSync(activeDir) &&\n\t\tfs\n\t\t\t.readdirSync(activeDir)\n\t\t\t.some((f) => !f.startsWith(\".\") && f !== \"previous\");\n\n\tif (activeHasLogs) {\n\t\treturn { logDir: activeDir, debugLogPath };\n\t}\n\n\tif (!fs.existsSync(previousDir)) {\n\t\tconsole.log(\"No gauntlet_logs directory found.\");\n\t\treturn null;\n\t}\n\n\t// Fall back to previous directory — cleanLogs archives files directly here\n\tconst logDir = resolvePreviousLogDir(previousDir);\n\tif (!logDir) return null;\n\n\t// Debug log stays in the main gauntlet_logs dir, not in previous/\n\treturn { logDir, debugLogPath };\n}\n\nfunction resolvePreviousLogDir(previousDir: string): string | null {\n\tconst prevEntries = fs.readdirSync(previousDir);\n\tconst hasDirectFiles = prevEntries.some(\n\t\t(f) => f.endsWith(\".log\") || f.endsWith(\".json\"),\n\t);\n\n\tif (hasDirectFiles) return previousDir;\n\n\t// Legacy: check for timestamped subdirectories\n\tconst prevDirs = prevEntries\n\t\t.map((d) => path.join(previousDir, d))\n\t\t.filter((d) => fs.statSync(d).isDirectory())\n\t\t.sort()\n\t\t.reverse();\n\n\tif (prevDirs.length === 0) {\n\t\tconsole.log(\"No gauntlet logs found.\");\n\t\treturn null;\n\t}\n\n\treturn prevDirs[0] as string;\n}\n\nfunction main(): void {\n\tconst cwd = process.cwd();\n\tconst logDirName = getLogDir(cwd);\n\tconst activeDir = path.join(cwd, logDirName);\n\n\tconst paths = resolveLogPaths(activeDir);\n\tif (!paths) {\n\t\tprocess.exit(0);\n\t}\n\n\t// Parse debug log, filtering to current session based on log file timestamps\n\tlet sessions: SessionRun[] = [];\n\tif (fs.existsSync(paths.debugLogPath)) {\n\t\tconst debugContent = fs.readFileSync(paths.debugLogPath, \"utf-8\");\n\t\tconst sessionStart = getSessionStartTime(paths.logDir);\n\t\tsessions = parseDebugLog(debugContent, sessionStart);\n\t}\n\n\t// Format and output\n\tconst output = formatSession(sessions, paths.logDir);\n\tconsole.log(output);\n}\n\nmain();\n"
+  ],
+  "mappings": ";;;AAaA;AACA;AAgDA,SAAS,aAAa,CAAC,MAAsC;AAAA,EAC5D,MAAM,SAAiC,CAAC;AAAA,EACxC,WAAW,SAAS,KAAK,SAAS,cAAc,GAAG;AAAA,IAClD,MAAM,MAAM,MAAM;AAAA,IAClB,MAAM,QAAQ,MAAM;AAAA,IACpB,IAAI,OAAO;AAAA,MAAO,OAAO,OAAO;AAAA,EACjC;AAAA,EACA,OAAO;AAAA;AAGR,SAAS,cAAc,CAAC,MAAsB;AAAA,EAC7C,MAAM,IAAI,KAAK,MAAM,eAAe;AAAA,EACpC,OAAO,IAAI,MAAM;AAAA;AAGlB,SAAS,cAAc,CAAC,MAAsB;AAAA,EAC7C,MAAM,IAAI,KAAK,MAAM,qBAAqB;AAAA,EAC1C,OAAO,IAAI,MAAM;AAAA;AAGlB,SAAS,cAAc,CAAC,MAAsB;AAAA,EAC7C,MAAM,IAAI,KAAK,MAAM,0BAA0B;AAAA,EAC/C,OAAO,IAAI,MAAM;AAAA;AAKlB,SAAS,aAAa,CAAC,SAAiB,kBAAuC;AAAA,EAC9E,MAAM,QAAQ,QAAQ,MAAM;AAAA,CAAI,EAAE,OAAO,CAAC,MAAM,EAAE,KAAK,CAAC;AAAA,EACxD,MAAM,WAAyB,CAAC;AAAA,EAChC,IAAI,UAA6B;AAAA,EAEjC,WAAW,QAAQ,OAAO;AAAA,IACzB,MAAM,QAAQ,eAAe,IAAI;AAAA,IACjC,MAAM,OAAO,eAAe,IAAI;AAAA,IAChC,MAAM,KAAK,eAAe,IAAI;AAAA,IAE9B,QAAQ;AAAA,WACF,aAAa;AAAA,QAEjB,IAAI,oBAAoB,IAAI,KAAK,EAAE,IAAI,kBAAkB;AAAA,UACxD,UAAU;AAAA,UACV;AAAA,QACD;AAAA,QACA,MAAM,KAAK,cAAc,IAAI;AAAA,QAC7B,UAAU;AAAA,UACT,OAAO;AAAA,YACN,WAAW;AAAA,YACX,MAAM,GAAG,QAAQ;AAAA,YACjB,SAAS,GAAG;AAAA,YACZ,cAAc,OAAO,GAAG,iBAAiB,GAAG,WAAW,CAAC;AAAA,YACxD,YAAY,OAAO,GAAG,eAAe,CAAC;AAAA,YACtC,cAAc,OAAO,GAAG,iBAAiB,CAAC;AAAA,YAC1C,OAAO,OAAO,GAAG,SAAS,CAAC;AAAA,UAC5B;AAAA,UACA,OAAO,CAAC;AAAA,QACT;AAAA,QACA,SAAS,KAAK,OAAO;AAAA,QACrB;AAAA,MACD;AAAA,WACK,eAAe;AAAA,QACnB,IAAI,CAAC;AAAA,UAAS;AAAA,QACd,MAAM,cAAc,KAAK,MAAM,QAAQ;AAAA,QACvC,MAAM,KAAK,cAAc,IAAI;AAAA,QAC7B,QAAQ,MAAM,KAAK;AAAA,UAClB,WAAW;AAAA,UACX,QAAQ,cAAc,MAAM;AAAA,UAC5B,KAAK,GAAG;AAAA,UACR,QAAQ,GAAG,UAAU;AAAA,UACrB,UAAU,GAAG,YAAY;AAAA,UACzB,YACC,GAAG,eAAe,YAAY,OAAO,GAAG,UAAU,IAAI;AAAA,QACxD,CAAC;AAAA,QACD;AAAA,MACD;AAAA,WACK,WAAW;AAAA,QACf,IAAI,CAAC;AAAA,UAAS;AAAA,QACd,MAAM,KAAK,cAAc,IAAI;AAAA,QAC7B,QAAQ,MAAM;AAAA,UACb,WAAW;AAAA,UACX,QAAQ,GAAG,UAAU;AAAA,UACrB,OAAO,OAAO,GAAG,SAAS,CAAC;AAAA,UAC3B,SAAS,OAAO,GAAG,WAAW,CAAC;AAAA,UAC/B,QAAQ,OAAO,GAAG,UAAU,CAAC;AAAA,UAC7B,YAAY,OAAO,GAAG,cAAc,CAAC;AAAA,UACrC,UAAU,GAAG,YAAY;AAAA,QAC1B;AAAA,QACA;AAAA,MACD;AAAA,WACK,aAAa;AAAA,QACjB,IAAI,CAAC;AAAA,UAAS;AAAA,QACd,MAAM,KAAK,cAAc,IAAI;AAAA,QAC7B,QAAQ,WAAW;AAAA,UAClB,WAAW;AAAA,UACX,UAAU,GAAG,YAAY;AAAA,UACzB,QAAQ,GAAG,UAAU;AAAA,QACtB;AAAA,QACA;AAAA,MACD;AAAA;AAAA,EAEF;AAAA,EAEA,OAAO;AAAA;AAOR,SAAS,mBAAmB,CAAC,QAAkC;AAAA,EAC9D,MAAM,UAAU,GACd,YAAY,MAAM,EAClB,OAAO,CAAC,MAAM,CAAC,EAAE,WAAW,GAAG,KAAK,MAAM,UAAU;AAAA,EACtD,IAAI;AAAA,EACJ,WAAW,SAAS,SAAS;AAAA,IAC5B,MAAM,QAAQ,GAAG,SAAS,KAAK,KAAK,QAAQ,KAAK,CAAC,EAAE;AAAA,IACpD,IAAI,aAAa,aAAa,QAAQ,UAAU;AAAA,MAC/C,WAAW;AAAA,IACZ;AAAA,EACD;AAAA,EACA,OAAO,aAAa,YAAY,IAAI,KAAK,QAAQ,IAAI;AAAA;AAKtD,SAAS,mBAAmB,CAAC,QAA0B;AAAA,EACtD,MAAM,QAAkB,CAAC;AAAA,EACzB,MAAM,UAAU,GACd,YAAY,MAAM,EAClB,OAAO,CAAC,MAAM,CAAC,EAAE,WAAW,GAAG,KAAK,MAAM,UAAU;AAAA,EACtD,IAAI,QAAQ,WAAW;AAAA,IAAG,OAAO;AAAA,EAEjC,MAAM,SAAmB,CAAC;AAAA,EAC1B,MAAM,UAAoB,CAAC;AAAA,EAC3B,MAAM,QAAkB,CAAC;AAAA,EAEzB,WAAW,SAAS,QAAQ,KAAK,GAAG;AAAA,IACnC,MAAM,WAAW,KAAK,KAAK,QAAQ,KAAK;AAAA,IACxC,MAAM,OAAO,GAAG,SAAS,QAAQ;AAAA,IACjC,MAAM,UAAU,KAAK,OAAO,MAAM,QAAQ,CAAC;AAAA,IAC3C,MAAM,OAAO,KAAK,aAAa;AAAA,IAE/B,IAAI,MAAM,WAAW,SAAS,GAAG;AAAA,MAChC,QAAQ,KAAK,IAAI;AAAA,IAClB,EAAO,SAAI,MAAM,WAAW,QAAQ,GAAG;AAAA,MACtC,OAAO,KAAK,IAAI;AAAA,IACjB,EAAO;AAAA,MACN,MAAM,KAAK,IAAI;AAAA;AAAA,EAEjB;AAAA,EAEA,MAAM,KAAK,eAAe;AAAA,EAC1B,MAAM,KAAK,EAAE;AAAA,EACb,IAAI,OAAO,SAAS,GAAG;AAAA,IACtB,MAAM,KAAK,iBAAiB;AAAA,IAC5B,MAAM,KAAK,GAAG,MAAM;AAAA,EACrB;AAAA,EACA,IAAI,QAAQ,SAAS,GAAG;AAAA,IACvB,MAAM,KAAK,uBAAuB;AAAA,IAClC,MAAM,KAAK,GAAG,OAAO;AAAA,EACtB;AAAA,EACA,IAAI,MAAM,SAAS,GAAG;AAAA,IACrB,MAAM,KAAK,YAAY;AAAA,IACvB,MAAM,KAAK,GAAG,KAAK;AAAA,EACpB;AAAA,EACA,MAAM,KAAK,EAAE;AAAA,EAEb,OAAO;AAAA;AAKR,SAAS,gBAAgB,CAAC,KAAqB;AAAA,EAC9C,OAAO,IAAI,WAAW,SACnB,WACA,IAAI,WAAW,SACd,WACA,IAAI,OAAO,YAAY;AAAA;AAG5B,SAAS,aAAa,CAAC,UAAkC;AAAA,EACxD,MAAM,QAAkB,CAAC;AAAA,EACzB,MAAM,KAAK,yBAAyB;AAAA,EACpC,MAAM,KAAK,EAAE;AAAA,EACb,SAAS,IAAI,EAAG,IAAI,SAAS,QAAQ,KAAK;AAAA,IACzC,MAAM,IAAI,SAAS;AAAA,IACnB,IAAI,CAAC;AAAA,MAAG;AAAA,IACR,MAAM,SAAS,EAAE,MAAM,EAAE,IAAI,SAAS;AAAA,IACtC,MAAM,WAAW,EAAE,MAAM,EAAE,IAAI,WAAW;AAAA,IAC1C,MAAM,KACL,GAAG,IAAI,OAAO,EAAE,MAAM,mBAAmB,EAAE,MAAM,eAAe,mBAAmB,UACpF;AAAA,EACD;AAAA,EACA,MAAM,KAAK,EAAE;AAAA,EACb,OAAO;AAAA;AAGR,SAAS,aAAa,CAAC,UAAwB,QAAwB;AAAA,EACtE,IAAI,SAAS,WAAW,GAAG;AAAA,IAC1B,OAAO;AAAA,EACR;AAAA,EAEA,MAAM,eAAe,CAAC,GAAG,QAAQ,EAAE,QAAQ,EAAE,KAAK,CAAC,MAAM,EAAE,GAAG;AAAA,EAC9D,MAAM,UAAU,gBAAgB,SAAS,SAAS,SAAS;AAAA,EAC3D,IAAI,CAAC;AAAA,IAAS,OAAO;AAAA,EAErB,MAAM,QAAkB,CAAC;AAAA,EAGzB,MAAM,KAAK,6BAA6B;AAAA,EACxC,MAAM,KAAK,EAAE;AAAA,EAGb,IAAI,QAAQ,KAAK;AAAA,IAChB,MAAM,KAAK,eAAe,iBAAiB,QAAQ,GAAG,GAAG;AAAA,IACzD,MAAM,KAAK,mBAAmB,QAAQ,IAAI,YAAY;AAAA,IACtD,MAAM,KAAK,iBAAiB,QAAQ,IAAI,UAAU;AAAA,IAClD,MAAM,KACL,cAAc,QAAQ,IAAI,wBAAwB,QAAQ,IAAI,yBAAyB,QAAQ,IAAI,QACpG;AAAA,EACD,EAAO;AAAA,IACN,MAAM,KAAK,4CAA4C;AAAA;AAAA,EAExD,MAAM,KAAK,EAAE;AAAA,EAGb,MAAM,KAAK,gBAAgB;AAAA,EAC3B,MAAM,KAAK,WAAW,QAAQ,MAAM,MAAM;AAAA,EAC1C,IAAI,QAAQ,MAAM,SAAS;AAAA,IAC1B,MAAM,KAAK,eAAe,QAAQ,MAAM,SAAS;AAAA,EAClD;AAAA,EACA,MAAM,KAAK,oBAAoB,QAAQ,MAAM,cAAc;AAAA,EAC3D,MAAM,KACL,aAAa,QAAQ,MAAM,iBAAiB,QAAQ,MAAM,cAC3D;AAAA,EACA,MAAM,KAAK,YAAY,QAAQ,MAAM,OAAO;AAAA,EAC5C,MAAM,KAAK,EAAE;AAAA,EAGb,MAAM,KAAK,kBAAkB;AAAA,EAC7B,MAAM,KAAK,EAAE;AAAA,EACb,MAAM,KAAK,iDAAiD;AAAA,EAC5D,MAAM,KAAK,iDAAiD;AAAA,EAC5D,WAAW,QAAQ,QAAQ,OAAO;AAAA,IACjC,MAAM,aACL,KAAK,eAAe,YAAY,OAAO,KAAK,UAAU,IAAI;AAAA,IAC3D,MAAM,aAAa,KAAK,WAAW,SAAS,SAAS;AAAA,IACrD,MAAM,KACL,KAAK,KAAK,YAAY,KAAK,OAAO,SAAS,gBAAgB,KAAK,cAAc,cAC/E;AAAA,EACD;AAAA,EACA,MAAM,KAAK,EAAE;AAAA,EAGb,IAAI,QAAQ,UAAU;AAAA,IACrB,MAAM,KAAK,eAAe;AAAA,IAC1B,MAAM,KAAK,eAAe,QAAQ,SAAS,UAAU;AAAA,IACrD,MAAM,KAAK,aAAa,QAAQ,SAAS,QAAQ;AAAA,IACjD,MAAM,KAAK,EAAE;AAAA,EACd;AAAA,EAGA,MAAM,KAAK,GAAG,oBAAoB,MAAM,CAAC;AAAA,EAGzC,IAAI,SAAS,SAAS,GAAG;AAAA,IACxB,MAAM,KAAK,GAAG,cAAc,QAAQ,CAAC;AAAA,EACtC;AAAA,EAEA,OAAO,MAAM,KAAK;AAAA,CAAI;AAAA;AASvB,SAAS,SAAS,CAAC,KAAqB;AAAA,EACvC,MAAM,aAAa,KAAK,KAAK,KAAK,aAAa,YAAY;AAAA,EAC3D,IAAI;AAAA,IACH,MAAM,UAAU,GAAG,aAAa,YAAY,OAAO;AAAA,IACnD,MAAM,QAAQ,QAAQ,MAAM,oBAAoB;AAAA,IAChD,IAAI,QAAQ;AAAA,MAAI,OAAO,MAAM,GAAG,KAAK;AAAA,IACpC,MAAM;AAAA,EAGR,OAAO;AAAA;AAOR,SAAS,eAAe,CACvB,WACkD;AAAA,EAClD,MAAM,cAAc,KAAK,KAAK,WAAW,UAAU;AAAA,EACnD,MAAM,eAAe,KAAK,KAAK,WAAW,YAAY;AAAA,EAGtD,MAAM,gBACL,GAAG,WAAW,SAAS,KACvB,GACE,YAAY,SAAS,EACrB,KAAK,CAAC,MAAM,CAAC,EAAE,WAAW,GAAG,KAAK,MAAM,UAAU;AAAA,EAErD,IAAI,eAAe;AAAA,IAClB,OAAO,EAAE,QAAQ,WAAW,aAAa;AAAA,EAC1C;AAAA,EAEA,IAAI,CAAC,GAAG,WAAW,WAAW,GAAG;AAAA,IAChC,QAAQ,IAAI,mCAAmC;AAAA,IAC/C,OAAO;AAAA,EACR;AAAA,EAGA,MAAM,SAAS,sBAAsB,WAAW;AAAA,EAChD,IAAI,CAAC;AAAA,IAAQ,OAAO;AAAA,EAGpB,OAAO,EAAE,QAAQ,aAAa;AAAA;AAG/B,SAAS,qBAAqB,CAAC,aAAoC;AAAA,EAClE,MAAM,cAAc,GAAG,YAAY,WAAW;AAAA,EAC9C,MAAM,iBAAiB,YAAY,KAClC,CAAC,MAAM,EAAE,SAAS,MAAM,KAAK,EAAE,SAAS,OAAO,CAChD;AAAA,EAEA,IAAI;AAAA,IAAgB,OAAO;AAAA,EAG3B,MAAM,WAAW,YACf,IAAI,CAAC,MAAM,KAAK,KAAK,aAAa,CAAC,CAAC,EACpC,OAAO,CAAC,MAAM,GAAG,SAAS,CAAC,EAAE,YAAY,CAAC,EAC1C,KAAK,EACL,QAAQ;AAAA,EAEV,IAAI,SAAS,WAAW,GAAG;AAAA,IAC1B,QAAQ,IAAI,yBAAyB;AAAA,IACrC,OAAO;AAAA,EACR;AAAA,EAEA,OAAO,SAAS;AAAA;AAGjB,SAAS,IAAI,GAAS;AAAA,EACrB,MAAM,MAAM,QAAQ,IAAI;AAAA,EACxB,MAAM,aAAa,UAAU,GAAG;AAAA,EAChC,MAAM,YAAY,KAAK,KAAK,KAAK,UAAU;AAAA,EAE3C,MAAM,QAAQ,gBAAgB,SAAS;AAAA,EACvC,IAAI,CAAC,OAAO;AAAA,IACX,QAAQ,KAAK,CAAC;AAAA,EACf;AAAA,EAGA,IAAI,WAAyB,CAAC;AAAA,EAC9B,IAAI,GAAG,WAAW,MAAM,YAAY,GAAG;AAAA,IACtC,MAAM,eAAe,GAAG,aAAa,MAAM,cAAc,OAAO;AAAA,IAChE,MAAM,eAAe,oBAAoB,MAAM,MAAM;AAAA,IACrD,WAAW,cAAc,cAAc,YAAY;AAAA,EACpD;AAAA,EAGA,MAAM,SAAS,cAAc,UAAU,MAAM,MAAM;AAAA,EACnD,QAAQ,IAAI,MAAM;AAAA;AAGnB,KAAK;",
+  "debugId": "4BED01809E659CB464756E2164756E21",
+  "names": []
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agent-gauntlet",
-  "version": "0.10.0",
+  "version": "0.11.0",
   "description": "A CLI tool for testing AI coding agents",
   "license": "Apache-2.0",
   "author": "Paul Caplan",
@@ -17,34 +17,47 @@
     "ai",
     "agent",
     "testing",
-    "bun"
+    "node"
   ],
   "files": [
-    "src",
+    "dist",
     "README.md",
     "LICENSE"
   ],
   "bin": {
-    "agent-gauntlet": "src/index.ts"
+    "agent-gauntlet": "dist/index.js"
   },
   "type": "module",
   "engines": {
-    "bun": ">=1.0.0"
+    "node": ">=18.0.0"
   },
   "scripts": {
-    "build": "bun build --compile --minify --sourcemap ./src/index.ts --outfile bin/agent-gauntlet",
+    "build:bin": "bun build --compile --minify --sourcemap ./src/index.ts --outfile bin/agent-gauntlet",
+    "build:npm": "bun build.ts",
+    "prepublishOnly": "bun run build:npm",
     "test": "bun test",
+    "test:e2e": "bun run build:npm && bun test/integration/stop-hook-e2e.ts",
     "lint": "biome check src",
     "typecheck": "tsc --noEmit && tsc --noEmit -p test/tsconfig.json",
     "changeset": "changeset",
     "version": "changeset version",
-    "release": "changeset publish"
+    "release": "changeset publish",
+    "run": "bun src/index.ts run",
+    "check": "bun src/index.ts check",
+    "clean": "bun src/index.ts clean",
+    "review": "bun src/index.ts review",
+    "detect": "bun src/index.ts detect",
+    "list": "bun src/index.ts list",
+    "health": "bun src/index.ts health",
+    "validate": "bun src/index.ts validate",
+    "wait-ci": "bun src/index.ts wait-ci"
   },
   "devDependencies": {
     "@biomejs/biome": "^2.3.11",
     "@changesets/changelog-github": "^0.5.2",
     "@changesets/cli": "^2.29.8",
-    "@types/bun": "latest"
+    "@types/bun": "latest",
+    "@types/picomatch": "^4.0.2"
   },
   "peerDependencies": {
     "typescript": "^5"
@@ -54,6 +67,7 @@
     "chalk": "^5.6.2",
     "commander": "^14.0.2",
     "gray-matter": "^4.0.3",
+    "picomatch": "^4.0.3",
     "yaml": "^2.8.2",
     "zod": "^4.3.5"
   }

package/src/built-in-reviews/code-quality.md

@@ -1,25 +0,0 @@
-# Code Quality Review
-
-You are a senior software engineer performing a code review. Your primary goal is to identify **real problems** that could cause bugs, security vulnerabilities, or performance issues in production. Do not report style, formatting, naming conventions, or maintainability suggestions unless you see something egregious.
-
-## Focus Areas (in priority order)
-
-1. **Bugs** — Logic errors, null/undefined issues, race conditions, unhandled edge cases, resource leaks
-2. **Security** — Injection vulnerabilities, auth/authz flaws, sensitive data exposure, input validation gaps
-3. **Performance** — Algorithmic complexity issues, N+1 queries, blocking operations, memory problems
-4. **Maintainability** — Unclear code, missing error handling, duplication
-
-## Do NOT Report
-
-- Style, formatting, or naming preferences
-- Missing documentation, comments, or type annotations
-- Suggestions for "better" abstractions or patterns that aren't broken
-- Hypothetical issues that require unlikely preconditions
-- Issues in code that wasn't changed in this diff
-
-## Guidelines
-
-- **Threshold**: only report issues you would block a PR over
-- Explain **why** each issue is a problem with a concrete failure scenario
-- Provide a **concrete fix** with corrected code
-- If the status quo works correctly, it's not a violation

package/src/built-in-reviews/index.ts

@@ -1,28 +0,0 @@
-// @ts-expect-error Bun text import
-import codeQualityContent from "./code-quality.md" with { type: "text" };
-
-const BUILT_IN_PREFIX = "built-in:";
-
-const builtInSources: Record<string, string> = {
-	"code-quality": codeQualityContent,
-};
-
-/**
- * Check if a review name uses the built-in prefix.
- */
-export function isBuiltInReview(name: string): boolean {
-	return name.startsWith(BUILT_IN_PREFIX);
-}
-
-/**
- * Load a built-in review prompt by name. Returns the raw markdown content.
- */
-export function loadBuiltInReview(name: string): string {
-	const source = builtInSources[name];
-
-	if (!source) {
-		throw new Error(`Unknown built-in review: "${name}"`);
-	}
-
-	return source;
-}

package/src/bun-plugins.d.ts

@@ -1,4 +0,0 @@
-declare module "*.yml" {
-	const content: string;
-	export default content;
-}