@gobing-ai/superskill 0.2.4 → 0.2.5

package/dist/index.js

@@ -2226,13 +2226,20 @@ import { rmSync } from "fs";
 async function backupFile(filePath) {
   let backupPath = `${filePath}.bak`;
   if (await Bun.file(backupPath).exists()) {
-    const ts = new Date().toISOString().replace(/[:.]/g, "").slice(0, 15);
+    const ts = new Date().toISOString().replace(/[:.]/g, "").slice(0, 18);
     backupPath = `${filePath}.bak.${ts}`;
+    let counter = 1;
+    while (await Bun.file(backupPath).exists()) {
+      backupPath = `${filePath}.bak.${ts}.${counter++}`;
+    }
   }
   await Bun.write(backupPath, Bun.file(filePath));
   return backupPath;
 }
 async function restoreFromBackup(backupPath, originalPath) {
+  if (!await Bun.file(backupPath).exists()) {
+    throw new Error(`Backup file not found: ${backupPath} \u2014 cannot restore ${originalPath}`);
+  }
   await Bun.write(originalPath, Bun.file(backupPath));
   rmSync(backupPath, { force: true });
 }
@@ -9230,11 +9237,10 @@ var init_dist2 = __esm(() => {
 
 // ../../packages/core/src/content/frontmatter.ts
 function parseFrontmatter(content) {
-  if (!content.startsWith(`---
-`)) {
+  if (!/^---\r?\n/.test(content)) {
     throw new FrontmatterError("Missing frontmatter: content must start with ---");
   }
-  const closerMatch = content.slice(4).match(/\n---(?=\n|$)/);
+  const closerMatch = content.slice(4).match(/\r?\n---(?=\r?\n|$)/);
   if (closerMatch?.index === undefined) {
     throw new FrontmatterError("Missing frontmatter closing delimiter (---)");
   }
@@ -9445,7 +9451,7 @@ function walkFrontmatter(content, opts) {
     }
     out.push(line);
   }
-  if (!injectedName) {
+  if (!injectedName || inFrontmatter) {
     return `${opts.fallbackBlock}
 
 ${content}`;
@@ -9469,6 +9475,12 @@ var init_rewrite_references = __esm(() => {
   SLASH_COMMAND_LINE_RE = /^\s*\/[a-zA-Z0-9._-]+:[a-zA-Z0-9._-]+(\s|$)/;
 });
 
+// ../../packages/core/src/pipeline/yaml-utils.ts
+function quoteYaml(value) {
+  const escaped = value.replace(/\\/g, "\\\\").replace(/"/g, "\\\"");
+  return `"${escaped}"`;
+}
+
 // ../../packages/core/src/pipeline/adapt-command.ts
 function adaptCommandToSkill(source, expectedName, pluginPrefix) {
   let result;
@@ -9480,7 +9492,7 @@ function adaptCommandToSkill(source, expectedName, pluginPrefix) {
     const description = firstLine?.trim() || `${expectedName} command`;
     result = `---
 name: ${expectedName}
-description: "${description}"
+description: ${quoteYaml(description)}
 disable-model-invocation: true
 ---
 
@@ -9502,7 +9514,8 @@ function normalizeCommandFrontmatter(content, expectedName) {
           const value = line.slice(line.indexOf(":") + 1).trim();
           if (value.startsWith("["))
             return line;
-          const parts = value.split(/,\s*/).map((p) => p.trim()).filter(Boolean);
+          const unquoted = value.replace(/^["']|["']$/g, "");
+          const parts = unquoted.split(/,\s*/).map((p) => p.trim()).filter(Boolean);
           return `allowed-tools: [${parts.join(", ")}]`;
         }
       }
@@ -9517,10 +9530,6 @@ disable-model-invocation: true
 ---`
   });
 }
-function quoteYaml(value) {
-  const escaped = value.replace(/\\/g, "\\\\").replace(/"/g, "\\\"");
-  return `"${escaped}"`;
-}
 var init_adapt_command = __esm(() => {
   init_rewrite_references();
 });
@@ -9603,7 +9612,7 @@ function adaptSubagentToSkill(source, expectedName, pluginPrefix) {
     const description = firstLine?.trim() || `${expectedName} subagent`;
     result = `---
 name: ${expectedName}
-description: "${description}"
+description: ${quoteYaml(description)}
 ---
 
 ${source}`;
@@ -9637,17 +9646,17 @@ ${source}`;
   const piTools = normalizePiToolList(rawToolsStr);
   const description = asString(data.description) || `${expectedName} subagent`;
   let model = asString(data.model);
-  if (model === "inherit")
+  if (model.toLowerCase() === "inherit")
     model = "";
   const skillsList = resolvePiSkills(data, body, pluginPrefix, skillExists);
   const skillsCsv = skillsList.join(", ");
   const fields = [`name: ${expectedName}`];
   if (description)
-    fields.push(`description: ${description}`);
+    fields.push(`description: ${quoteYaml(description)}`);
   if (piTools)
     fields.push(`tools: ${piTools}`);
   if (model)
-    fields.push(`model: ${model}`);
+    fields.push(`model: ${quoteYaml(model)}`);
   if (skillsCsv)
     fields.push(`skill: ${skillsCsv}`);
   const runtimeNotes = buildPiRuntimeNotes(parseToolsList(rawToolsStr), skillsCsv);
@@ -9731,8 +9740,9 @@ var init_adapt_subagent = __esm(() => {
 });
 
 // ../../packages/core/src/mapper.ts
-import { existsSync as existsSync3, mkdirSync, readdirSync, readFileSync as readFileSync2, rmSync as rmSync2, statSync as statSync2, writeFileSync } from "fs";
-import { join as join3 } from "path";
+import { existsSync as existsSync3, lstatSync, mkdirSync, readdirSync, readFileSync as readFileSync2, rmSync as rmSync2, writeFileSync } from "fs";
+import { homedir as homedir2 } from "os";
+import { join as join3, resolve } from "path";
 function convertClaudeHooksToCanonical(claudeJson) {
   const claudeHooks = claudeJson.hooks;
   if (!claudeHooks || typeof claudeHooks !== "object")
@@ -9771,10 +9781,15 @@ function convertClaudeHooksToCanonical(claudeJson) {
   return { hooks: canonical };
 }
 function setSkillName(content, newName) {
-  return content.replace(/^name:\s*.+$/m, `name: ${newName}`);
+  if (/^name:\s*.+$/m.test(content)) {
+    return content.replace(/^name:\s*.+$/m, `name: ${newName}`);
+  }
+  return content.replace(/^---\s*$/m, `---
+name: ${newName}`);
 }
 function mapPluginToRulesync(pluginPath, pluginName, outputDir) {
   assertSafePathSegment(pluginName, "plugin name");
+  assertSafeOutputDir(outputDir);
   if (existsSync3(outputDir)) {
     rmSync2(outputDir, { recursive: true, force: true });
   }
@@ -9875,7 +9890,12 @@ function deepMergeJsonFile(sourcePath, targetPath) {
 `);
 }
 function readJsonObject(path) {
-  const value = JSON.parse(readFileSync2(path, "utf-8"));
+  let value;
+  try {
+    value = JSON.parse(readFileSync2(path, "utf-8"));
+  } catch (e) {
+    throw new Error(`Failed to parse JSON in ${path}: ${e instanceof Error ? e.message : String(e)}`);
+  }
   if (!isJsonObject(value)) {
     throw new Error(`Expected JSON object in ${path}`);
   }
@@ -9892,20 +9912,44 @@ function deepMerge2(target, source) {
 function isJsonObject(value) {
   return typeof value === "object" && value !== null && !Array.isArray(value);
 }
+function assertSafeOutputDir(outputDir) {
+  const resolved = resolve(outputDir);
+  const home = homedir2();
+  const dangerous = [home, process.cwd(), "/", "/home", "/Users"];
+  for (const danger of dangerous) {
+    if (resolved === resolve(danger)) {
+      throw new Error(`Refusing to delete output directory '${resolved}': resolves to a protected path ` + `(home, cwd, or filesystem root). Use a subdirectory.`);
+    }
+  }
+}
 function copyAndRewriteDirectory(source, destination, pluginName) {
   mkdirSync(destination, { recursive: true });
   for (const entry of readdirSync(source)) {
     const srcPath = join3(source, entry);
     const destPath = join3(destination, entry);
-    if (statSync2(srcPath).isDirectory()) {
+    const stat = lstatSync(srcPath);
+    if (stat.isSymbolicLink())
+      continue;
+    if (stat.isDirectory()) {
       copyAndRewriteDirectory(srcPath, destPath, pluginName);
     } else {
-      const content = readFileSync2(srcPath, "utf-8");
-      writeFileSync(destPath, rewriteSkillReferences(content, pluginName));
+      const bytes = readFileSync2(srcPath);
+      if (isTextFile(entry, bytes)) {
+        const content = bytes.toString("utf-8");
+        writeFileSync(destPath, rewriteSkillReferences(content, pluginName));
+      } else {
+        writeFileSync(destPath, bytes);
+      }
     }
   }
 }
-var CLAUDE_TO_CANONICAL_EVENT;
+function isTextFile(filename, bytes) {
+  const ext = filename.slice(filename.lastIndexOf("."));
+  if (TEXT_EXTENSIONS.has(ext))
+    return true;
+  return !bytes.subarray(0, 8192).includes(0);
+}
+var CLAUDE_TO_CANONICAL_EVENT, TEXT_EXTENSIONS;
 var init_mapper = __esm(() => {
   init_identity();
   init_adapt_command();
@@ -9927,6 +9971,34 @@ var init_mapper = __esm(() => {
     WorktreeRemove: "worktreeRemove",
     MessageDisplay: "messageDisplay"
   };
+  TEXT_EXTENSIONS = new Set([
+    ".md",
+    ".ts",
+    ".js",
+    ".json",
+    ".yaml",
+    ".yml",
+    ".txt",
+    ".sh",
+    ".bash",
+    ".py",
+    ".rb",
+    ".go",
+    ".rs",
+    ".java",
+    ".c",
+    ".cpp",
+    ".h",
+    ".css",
+    ".html",
+    ".xml",
+    ".toml",
+    ".ini",
+    ".cfg",
+    ".env",
+    ".gitignore",
+    ".dockerfile"
+  ]);
 });
 
 // ../../node_modules/.bun/zod@3.25.76/node_modules/zod/v3/helpers/util.js
@@ -13897,20 +13969,20 @@ var init_zod = __esm(() => {
 
 // ../../packages/core/src/marketplace.ts
 import { existsSync as existsSync4, readdirSync as readdirSync2, readFileSync as readFileSync3 } from "fs";
-import { join as join4, resolve } from "path";
+import { join as join4, resolve as resolve2 } from "path";
 function resolvePlugin(marketplacePath, pluginName) {
   let manifestPath = null;
   if (marketplacePath) {
     if (marketplacePath.endsWith("marketplace.json")) {
-      manifestPath = resolve(marketplacePath);
+      manifestPath = resolve2(marketplacePath);
     } else {
-      manifestPath = resolve(join4(marketplacePath, "marketplace.json"));
+      manifestPath = resolve2(join4(marketplacePath, "marketplace.json"));
     }
     if (!existsSync4(manifestPath)) {
       throw new Error(`Marketplace manifest not found at: ${manifestPath}`);
     }
   } else {
-    const cwdManifest = resolve(".claude-plugin", "marketplace.json");
+    const cwdManifest = resolve2(".claude-plugin", "marketplace.json");
     if (existsSync4(cwdManifest)) {
       manifestPath = cwdManifest;
     }
@@ -13937,12 +14009,15 @@ function resolvePlugin(marketplacePath, pluginName) {
   if (!source.startsWith("./")) {
     throw new Error(`Remote sources not yet supported for plugin '${pluginName}'. Phase 1 only supports relative-path sources (starting with './').`);
   }
-  if (/(?:^|\/)\.\.(?:\/|$)/.test(source)) {
+  if (/(?:^|[\\/])\.\.(?:[\\/]|$)/.test(source)) {
     throw new Error(`Plugin source for '${pluginName}' escapes the marketplace root: '${source}'.`);
   }
-  const marketplaceRoot = resolve(manifestPath, "..", "..");
   const pluginRootBase = manifest.metadata?.pluginRoot ?? "";
-  const pluginRoot = resolve(marketplaceRoot, pluginRootBase, source);
+  if (pluginRootBase && /(?:^|[\\/])\.\.(?:[\\/]|$)/.test(pluginRootBase)) {
+    throw new Error(`Plugin root for '${pluginName}' escapes the marketplace root: '${pluginRootBase}'.`);
+  }
+  const marketplaceRoot = resolve2(manifestPath, "..", "..");
+  const pluginRoot = resolve2(marketplaceRoot, pluginRootBase, source);
   let dirents;
   try {
     dirents = readdirSync2(pluginRoot);
@@ -13962,11 +14037,11 @@ function resolvePlugin(marketplacePath, pluginName) {
 function listResolvablePlugins(marketplacePath) {
   let manifestPath = null;
   if (marketplacePath) {
-    manifestPath = marketplacePath.endsWith("marketplace.json") ? resolve(marketplacePath) : resolve(join4(marketplacePath, "marketplace.json"));
+    manifestPath = marketplacePath.endsWith("marketplace.json") ? resolve2(marketplacePath) : resolve2(join4(marketplacePath, "marketplace.json"));
     if (!existsSync4(manifestPath))
       return [];
   } else {
-    const cwdManifest = resolve(".claude-plugin", "marketplace.json");
+    const cwdManifest = resolve2(".claude-plugin", "marketplace.json");
     if (existsSync4(cwdManifest))
       manifestPath = cwdManifest;
   }
@@ -14106,7 +14181,7 @@ var init_migrate = __esm(() => {
 });
 
 // ../../packages/core/src/operations/package.ts
-import { cpSync, existsSync as existsSync6, mkdirSync as mkdirSync3, statSync as statSync3 } from "fs";
+import { cpSync, existsSync as existsSync6, mkdirSync as mkdirSync3, rmSync as rmSync3, statSync as statSync2 } from "fs";
 import { basename as basename2, dirname as dirname3, join as join5 } from "path";
 import { cwd as cwd3 } from "process";
 function resolveSkillDir(name) {
@@ -14114,7 +14189,7 @@ function resolveSkillDir(name) {
   if (!skillPath) {
     throw Object.assign(new Error(`Skill not found: ${name}`), { code: "ENOENT" });
   }
-  const dir = statSync3(skillPath).isDirectory() ? skillPath : dirname3(skillPath);
+  const dir = statSync2(skillPath).isDirectory() ? skillPath : dirname3(skillPath);
   return { dir, name: basename2(dir) };
 }
 function copyDirIfExists(src, dest) {
@@ -14130,6 +14205,9 @@ function copyFileIfExists(src, dest) {
 async function packageSkill(name, opts = {}) {
   const { dir: skillDir, name: skillName } = resolveSkillDir(name);
   const outputDir = join5(opts.output ?? cwd3(), skillName);
+  if (existsSync6(outputDir)) {
+    rmSync3(outputDir, { recursive: true, force: true });
+  }
   mkdirSync3(outputDir, { recursive: true });
   copyFileIfExists(join5(skillDir, "SKILL.md"), join5(outputDir, "SKILL.md"));
   copyDirIfExists(join5(skillDir, "references"), join5(outputDir, "references"));
@@ -14152,7 +14230,7 @@ var init_package = __esm(() => {
 
 // ../../packages/core/src/operations/scaffold.ts
 import { existsSync as existsSync7, mkdirSync as mkdirSync4, readFileSync as readFileSync5, writeFileSync as writeFileSync3 } from "fs";
-import { homedir as homedir2 } from "os";
+import { homedir as homedir3 } from "os";
 import { join as join6 } from "path";
 import { cwd as cwd4 } from "process";
 function parseList(value) {
@@ -14192,8 +14270,10 @@ function mergeFrontmatterList(content, key, items) {
   return content.slice(0, fmStart) + newFmBody + content.slice(closePos);
 }
 function resolveTemplate(type, tier) {
-  const homeDir = process.env.HOME ?? homedir2();
+  const homeDir = process.env.HOME ?? homedir3();
   const tierName = tier?.trim();
+  if (tierName)
+    assertSafePathSegment(tierName, "template tier");
   if (tierName && tierName !== "default") {
     const userTierPath = join6(homeDir, ".superskill", "templates", type, `${tierName}.md`);
     if (existsSync7(userTierPath)) {
@@ -14296,10 +14376,7 @@ function scoreClarityFromDensities(body) {
 function parseFrontmatterSafe(content) {
   try {
     return parseFrontmatter(content).data;
-  } catch (e) {
-    if (e instanceof FrontmatterError) {
-      return null;
-    }
+  } catch {
     return null;
   }
 }
@@ -14413,7 +14490,7 @@ function parseHookEntries(raw) {
 function scoreCorrectness(entries) {
   let valid = 0;
   for (const e of entries) {
-    if (e.type === "command" && e.command.length > 0 && e.matcher.length > 0)
+    if ((e.type === "command" || e.type === "prompt") && e.command.length > 0 && e.matcher.length > 0)
       valid++;
   }
   const score = entries.length > 0 ? clamp(valid / entries.length) : 0;
@@ -14423,7 +14500,7 @@ function scoreCorrectness(entries) {
 function scoreEventCoverage(entries) {
   const events = new Set(entries.map((e) => e.event));
   const covered = KNOWN_HOOK_EVENTS.filter((e) => events.has(e)).length;
-  const score = clamp(Math.min(events.size / 3, 1));
+  const score = clamp(Math.min(covered / 3, 1));
   const note = `${covered} of ${KNOWN_HOOK_EVENTS.length} known events covered (${events.size} total)`;
   return { score, note };
 }
@@ -14578,7 +14655,7 @@ var init_hook = __esm(() => {
 });
 
 // ../../packages/core/src/operations/validate.ts
-import { existsSync as existsSync8, statSync as statSync4 } from "fs";
+import { existsSync as existsSync8, statSync as statSync3 } from "fs";
 import { dirname as dirname4, join as join7 } from "path";
 function checkBodyLinks(body, baseDir) {
   const findings = [];
@@ -14616,7 +14693,7 @@ async function validate(type, nameOrPath, opts) {
   }
   let stat;
   try {
-    stat = statSync4(resolvedPath);
+    stat = statSync3(resolvedPath);
   } catch {
     return sentinelResult(`File not found: ${resolvedPath}`);
   }
@@ -14638,13 +14715,14 @@ async function validate(type, nameOrPath, opts) {
   const baseDir = dirname4(resolvedPath);
   const result = _validateContent(type, content, {
     ...opts,
-    referenceChecker: (refType, refName) => resolveContentPath(refType, refName, { baseDir }) !== null
+    referenceChecker: opts?.referenceChecker ?? ((refType, refName) => resolveContentPath(refType, refName, { baseDir }) !== null)
   });
-  const bodyStart = content.indexOf(`
----
-`);
-  const body = bodyStart >= 0 ? content.slice(bodyStart + 5) : content;
-  const linkFindings = checkBodyLinks(body, baseDir);
+  let bodyForLinks = content;
+  try {
+    const { body } = parseFrontmatter(content);
+    bodyForLinks = body;
+  } catch {}
+  const linkFindings = checkBodyLinks(bodyForLinks, baseDir);
   result.findings.push(...linkFindings);
   return result;
 }
@@ -14652,7 +14730,9 @@ function _validateContent(type, content, opts) {
   const findings = [];
   let data;
   let body;
-  const frontmatterPresent = /^---\s*$/m.test(content);
+  const frontmatterPresent = content.startsWith(`---
+`) || content.startsWith(`---\r
+`);
   try {
     if (frontmatterPresent) {
       const parsed = parseFrontmatter(content);
@@ -14832,7 +14912,12 @@ function strictChecks(type, data, body) {
       message: `Body content is too short (${trimmedBody.length} chars). Recommended minimum is 20 characters.`
     });
   }
-  for (const key of Object.keys(data)) {
+  const recognized = new Set([
+    ...Object.keys(FIELD_TYPES[type] ?? {}),
+    ...REQUIRED_FIELDS[type],
+    ...KNOWN_OPTIONAL[type] ?? []
+  ]);
+  for (const [key, value] of Object.entries(data)) {
     const replacement = DEPRECATED_FIELDS[key];
     if (replacement) {
       findings.push({
@@ -14841,31 +14926,20 @@ function strictChecks(type, data, body) {
         message: `'${key}' is deprecated. ${replacement}`
       });
     }
-  }
-  for (const [field, value] of Object.entries(data)) {
     if (typeof value === "string" && value !== value.trimEnd()) {
       findings.push({
         severity: "warning",
-        field,
-        message: `'${field}' has trailing whitespace`
+        field: key,
+        message: `'${key}' has trailing whitespace`
+      });
+    }
+    if (!recognized.has(key) && DEPRECATED_FIELDS[key] === undefined) {
+      findings.push({
+        severity: "warning",
+        field: key,
+        message: `Unknown frontmatter key '${key}'`
       });
     }
-  }
-  const recognized = new Set([
-    ...Object.keys(FIELD_TYPES[type] ?? {}),
-    ...REQUIRED_FIELDS[type],
-    ...KNOWN_OPTIONAL[type] ?? []
-  ]);
-  for (const key of Object.keys(data)) {
-    if (recognized.has(key))
-      continue;
-    if (DEPRECATED_FIELDS[key] !== undefined)
-      continue;
-    findings.push({
-      severity: "warning",
-      field: key,
-      message: `Unknown frontmatter key '${key}'`
-    });
   }
   return findings;
 }
@@ -17290,8 +17364,12 @@ function getTelemetryConfig(configPartial = {}) {
 function getResolvedConfig() {
   return resolvedConfig;
 }
-var CONFIG_DEFAULTS, resolvedConfig;
+function getTracer() {
+  return import_api.trace.getTracer(TRACER_NAME, TRACER_VERSION);
+}
+var import_api, CONFIG_DEFAULTS, TRACER_NAME = "@gobing-ai/ts-infra", TRACER_VERSION = "0.1.0", resolvedConfig;
 var init_sdk = __esm(() => {
+  import_api = __toESM(require_src(), 1);
   CONFIG_DEFAULTS = {
     enabled: true,
     serviceName: "ts-libs",
@@ -17300,13 +17378,41 @@ var init_sdk = __esm(() => {
   resolvedConfig = getTelemetryConfig();
 });
 
+// ../../node_modules/.bun/@gobing-ai+ts-infra@0.3.21+3246782889eb34e0/node_modules/@gobing-ai/ts-infra/dist/telemetry/tracing.js
+function isSuppressed(tracer) {
+  return tracer === undefined && !getResolvedConfig().enabled;
+}
+function nonRecordingSpan() {
+  return import_api2.trace.wrapSpanContext(import_api2.INVALID_SPAN_CONTEXT);
+}
+async function traceAsync(name, fn, options2, tracer) {
+  if (isSuppressed(tracer))
+    return fn(nonRecordingSpan());
+  const resolvedTracer = tracer ?? getTracer();
+  return resolvedTracer.startActiveSpan(name, options2 ?? {}, async (span) => {
+    try {
+      return await fn(span);
+    } catch (err) {
+      span.setStatus({ code: 2, message: err instanceof Error ? err.message : String(err) });
+      throw err;
+    } finally {
+      span.end();
+    }
+  });
+}
+var import_api2;
+var init_tracing = __esm(() => {
+  init_sdk();
+  import_api2 = __toESM(require_src(), 1);
+});
+
 // ../../node_modules/.bun/@gobing-ai+ts-infra@0.3.21+3246782889eb34e0/node_modules/@gobing-ai/ts-infra/dist/telemetry/metrics.js
 function getMeter() {
-  return import_api.metrics.getMeter(METER_NAME, METER_VERSION);
+  return import_api3.metrics.getMeter(METER_NAME, METER_VERSION);
 }
 function noopCounter() {
   if (!_noopCounter)
-    _noopCounter = import_api.createNoopMeter().createCounter("noop");
+    _noopCounter = import_api3.createNoopMeter().createCounter("noop");
   return _noopCounter;
 }
 function getOrCreateCounter(key, name, description, unit = "{operation}") {
@@ -17323,10 +17429,10 @@ function getEventbusEmitsTotal() {
 function getEventbusErrorsTotal() {
   return getOrCreateCounter("ebErr", "eventbus.errors.total", "Event bus errors", "{error}");
 }
-var import_api, METER_NAME = "@gobing-ai/ts-infra", METER_VERSION = "0.1.0", instruments, _noopCounter;
+var import_api3, METER_NAME = "@gobing-ai/ts-infra", METER_VERSION = "0.1.0", instruments, _noopCounter;
 var init_metrics = __esm(() => {
   init_sdk();
-  import_api = __toESM(require_src(), 1);
+  import_api3 = __toESM(require_src(), 1);
   instruments = {};
 });
 
@@ -17580,10 +17686,16 @@ var init_event_bus2 = __esm(() => {
   init_event_bus();
 });
 
+// ../../node_modules/.bun/@gobing-ai+ts-infra@0.3.21+3246782889eb34e0/node_modules/@gobing-ai/ts-infra/dist/telemetry/index.js
+var init_telemetry = __esm(() => {
+  init_tracing();
+});
+
 // ../../node_modules/.bun/@gobing-ai+ts-infra@0.3.21+3246782889eb34e0/node_modules/@gobing-ai/ts-infra/dist/index.js
 var init_dist3 = __esm(() => {
   init_event_bus2();
   init_logger2();
+  init_telemetry();
 });
 
 // ../../node_modules/.bun/@gobing-ai+ts-ai-runner@0.3.21+5a4f8cc6c6b198a4/node_modules/@gobing-ai/ts-ai-runner/dist/agents/shims.js
@@ -32516,7 +32628,7 @@ var init_config = __esm(() => {
 });
 
 // ../../node_modules/.bun/@gobing-ai+ts-runtime@0.3.21/node_modules/@gobing-ai/ts-runtime/dist/file-system-node.js
-import { appendFileSync, cpSync as cpSync2, createWriteStream, existsSync as existsSync9, mkdirSync as mkdirSync5, readdirSync as readdirSync3, readFileSync as readFileSync6, renameSync, rmSync as rmSync3, statSync as statSync5, writeFileSync as writeFileSync4 } from "fs";
+import { appendFileSync, cpSync as cpSync2, createWriteStream, existsSync as existsSync9, mkdirSync as mkdirSync5, readdirSync as readdirSync3, readFileSync as readFileSync6, renameSync, rmSync as rmSync4, statSync as statSync4, writeFileSync as writeFileSync4 } from "fs";
 import { dirname as dirname5, resolve as resolvePath } from "path";
 function createNodeFileSystem(root) {
   const projectRoot = root ?? findProjectRoot(process.cwd());
@@ -32538,7 +32650,7 @@ function createNodeFileSystem(root) {
     },
     readDir: (path) => readdirSync3(path),
     deleteFile: (path) => {
-      rmSync3(path, { recursive: true, force: true });
+      rmSync4(path, { recursive: true, force: true });
     },
     rename: (src, dest) => {
       renameSync(src, dest);
@@ -32552,7 +32664,7 @@ function createNodeFileSystem(root) {
     },
     stat: (path) => {
       try {
-        const s = statSync5(path);
+        const s = statSync4(path);
         return {
           isFile: () => s.isFile(),
           isDirectory: () => s.isDirectory(),
@@ -33518,12 +33630,12 @@ var require_isexe = __commonJS((exports, module) => {
       if (typeof Promise !== "function") {
         throw new TypeError("callback not provided");
       }
-      return new Promise(function(resolve2, reject) {
+      return new Promise(function(resolve3, reject) {
         isexe(path, options2 || {}, function(er, is) {
           if (er) {
             reject(er);
           } else {
-            resolve2(is);
+            resolve3(is);
           }
         });
       });
@@ -33585,27 +33697,27 @@ var require_which = __commonJS((exports, module) => {
       opt = {};
     const { pathEnv, pathExt, pathExtExe } = getPathInfo(cmd, opt);
     const found = [];
-    const step = (i) => new Promise((resolve2, reject) => {
+    const step = (i) => new Promise((resolve3, reject) => {
       if (i === pathEnv.length)
-        return opt.all && found.length ? resolve2(found) : reject(getNotFoundError(cmd));
+        return opt.all && found.length ? resolve3(found) : reject(getNotFoundError(cmd));
       const ppRaw = pathEnv[i];
       const pathPart = /^".*"$/.test(ppRaw) ? ppRaw.slice(1, -1) : ppRaw;
       const pCmd = path.join(pathPart, cmd);
       const p = !pathPart && /^\.[\\\/]/.test(cmd) ? cmd.slice(0, 2) + pCmd : pCmd;
-      resolve2(subStep(p, i, 0));
+      resolve3(subStep(p, i, 0));
     });
-    const subStep = (p, i, ii) => new Promise((resolve2, reject) => {
+    const subStep = (p, i, ii) => new Promise((resolve3, reject) => {
       if (ii === pathExt.length)
-        return resolve2(step(i + 1));
+        return resolve3(step(i + 1));
       const ext = pathExt[ii];
       isexe(p + ext, { pathExt: pathExtExe }, (er, is) => {
         if (!er && is) {
           if (opt.all)
             found.push(p + ext);
           else
-            return resolve2(p + ext);
+            return resolve3(p + ext);
         }
-        return resolve2(subStep(p, i, ii + 1));
+        return resolve3(subStep(p, i, ii + 1));
       });
     });
     return cb ? step(0).then((res) => cb(null, res), cb) : step(0);
@@ -34411,7 +34523,7 @@ var normalizeForceKillAfterDelay = (forceKillAfterDelay) => {
     throw new TypeError(`Expected the \`forceKillAfterDelay\` option to be a non-negative integer, got \`${forceKillAfterDelay}\` (${typeof forceKillAfterDelay})`);
   }
   return forceKillAfterDelay;
-}, DEFAULT_FORCE_KILL_TIMEOUT, subprocessKill = ({ kill, options: { forceKillAfterDelay, killSignal }, onInternalError, context, controller }, signalOrError, errorArgument) => {
+}, DEFAULT_FORCE_KILL_TIMEOUT, subprocessKill = ({ kill, options: { forceKillAfterDelay, killSignal }, onInternalError, context: context2, controller }, signalOrError, errorArgument) => {
   const { signal, error: error51 } = parseKillArguments(signalOrError, errorArgument, killSignal);
   emitKillError(error51, onInternalError);
   const killResult = kill(signal);
@@ -34421,7 +34533,7 @@ var normalizeForceKillAfterDelay = (forceKillAfterDelay) => {
     forceKillAfterDelay,
     killSignal,
     killResult,
-    context,
+    context: context2,
     controller
   });
   return killResult;
@@ -34438,23 +34550,23 @@ var normalizeForceKillAfterDelay = (forceKillAfterDelay) => {
   if (error51 !== undefined) {
     onInternalError.reject(error51);
   }
-}, setKillTimeout = async ({ kill, signal, forceKillAfterDelay, killSignal, killResult, context, controller }) => {
+}, setKillTimeout = async ({ kill, signal, forceKillAfterDelay, killSignal, killResult, context: context2, controller }) => {
   if (signal === killSignal && killResult) {
     killOnTimeout({
       kill,
       forceKillAfterDelay,
-      context,
+      context: context2,
       controllerSignal: controller.signal
     });
   }
-}, killOnTimeout = async ({ kill, forceKillAfterDelay, context, controllerSignal }) => {
+}, killOnTimeout = async ({ kill, forceKillAfterDelay, context: context2, controllerSignal }) => {
   if (forceKillAfterDelay === false) {
     return;
   }
   try {
     await setTimeout2(forceKillAfterDelay, undefined, { signal: controllerSignal });
     if (kill("SIGKILL")) {
-      context.isForcefullyTerminated ??= true;
+      context2.isForcefullyTerminated ??= true;
     }
   } catch {}
 };
@@ -34478,9 +34590,9 @@ var validateCancelSignal = ({ cancelSignal }) => {
   if (cancelSignal !== undefined && Object.prototype.toString.call(cancelSignal) !== "[object AbortSignal]") {
     throw new Error(`The \`cancelSignal\` option must be an AbortSignal: ${String(cancelSignal)}`);
   }
-}, throwOnCancel = ({ subprocess, cancelSignal, gracefulCancel, context, controller }) => cancelSignal === undefined || gracefulCancel ? [] : [terminateOnCancel(subprocess, cancelSignal, context, controller)], terminateOnCancel = async (subprocess, cancelSignal, context, { signal }) => {
+}, throwOnCancel = ({ subprocess, cancelSignal, gracefulCancel, context: context2, controller }) => cancelSignal === undefined || gracefulCancel ? [] : [terminateOnCancel(subprocess, cancelSignal, context2, controller)], terminateOnCancel = async (subprocess, cancelSignal, context2, { signal }) => {
   await onAbortedSignal(cancelSignal, signal);
-  context.terminationReason ??= "cancel";
+  context2.terminationReason ??= "cancel";
   subprocess.kill();
   throw cancelSignal.reason;
 };
@@ -34544,8 +34656,8 @@ var init_validation = __esm(() => {
 // ../../node_modules/.bun/execa@9.6.1/node_modules/execa/lib/utils/deferred.js
 var createDeferred = () => {
   const methods = {};
-  const promise2 = new Promise((resolve2, reject) => {
-    Object.assign(methods, { resolve: resolve2, reject });
+  const promise2 = new Promise((resolve3, reject) => {
+    Object.assign(methods, { resolve: resolve3, reject });
   });
   return Object.assign(promise2, methods);
 };
@@ -35005,25 +35117,25 @@ var validateGracefulCancel = ({ gracefulCancel, cancelSignal, ipc, serialization
   cancelSignal,
   gracefulCancel,
   forceKillAfterDelay,
-  context,
+  context: context2,
   controller
 }) => gracefulCancel ? [sendOnAbort({
   subprocess,
   cancelSignal,
   forceKillAfterDelay,
-  context,
+  context: context2,
   controller
-})] : [], sendOnAbort = async ({ subprocess, cancelSignal, forceKillAfterDelay, context, controller: { signal } }) => {
+})] : [], sendOnAbort = async ({ subprocess, cancelSignal, forceKillAfterDelay, context: context2, controller: { signal } }) => {
   await onAbortedSignal(cancelSignal, signal);
   const reason = getReason(cancelSignal);
   await sendAbort(subprocess, reason);
   killOnTimeout({
     kill: subprocess.kill,
     forceKillAfterDelay,
-    context,
+    context: context2,
     controllerSignal: signal
   });
-  context.terminationReason ??= "gracefulCancel";
+  context2.terminationReason ??= "gracefulCancel";
   throw cancelSignal.reason;
 }, getReason = ({ reason }) => {
   if (!(reason instanceof DOMException)) {
@@ -35050,9 +35162,9 @@ var validateTimeout = ({ timeout }) => {
   if (timeout !== undefined && (!Number.isFinite(timeout) || timeout < 0)) {
     throw new TypeError(`Expected the \`timeout\` option to be a non-negative integer, got \`${timeout}\` (${typeof timeout})`);
   }
-}, throwOnTimeout = (subprocess, timeout, context, controller) => timeout === 0 || timeout === undefined ? [] : [killAfterTimeout(subprocess, timeout, context, controller)], killAfterTimeout = async (subprocess, timeout, context, { signal }) => {
+}, throwOnTimeout = (subprocess, timeout, context2, controller) => timeout === 0 || timeout === undefined ? [] : [killAfterTimeout(subprocess, timeout, context2, controller)], killAfterTimeout = async (subprocess, timeout, context2, { signal }) => {
   await setTimeout3(timeout, undefined, { signal });
-  context.terminationReason ??= "timeout";
+  context2.terminationReason ??= "timeout";
   subprocess.kill();
   throw new DiscardedError;
 };
@@ -35180,7 +35292,7 @@ var init_encoding_option = __esm(() => {
 });
 
 // ../../node_modules/.bun/execa@9.6.1/node_modules/execa/lib/arguments/cwd.js
-import { statSync as statSync6 } from "fs";
+import { statSync as statSync5 } from "fs";
 import path4 from "path";
 import process6 from "process";
 var normalizeCwd = (cwd5 = getDefaultCwd()) => {
@@ -35200,7 +35312,7 @@ ${error51.message}`;
   }
   let cwdStat;
   try {
-    cwdStat = statSync6(cwd5);
+    cwdStat = statSync5(cwd5);
   } catch (error51) {
     return `The "cwd" option is invalid: ${cwd5}.
 ${error51.message}
@@ -37409,9 +37521,9 @@ var init_all_sync = __esm(() => {
 
 // ../../node_modules/.bun/execa@9.6.1/node_modules/execa/lib/resolve/exit-async.js
 import { once as once4 } from "events";
-var waitForExit = async (subprocess, context) => {
+var waitForExit = async (subprocess, context2) => {
   const [exitCode, signal] = await waitForExitOrError(subprocess);
-  context.isForcefullyTerminated ??= false;
+  context2.isForcefullyTerminated ??= false;
   return [exitCode, signal];
 }, waitForExitOrError = async (subprocess) => {
   const [spawnPayload, exitPayload] = await Promise.allSettled([
@@ -39020,14 +39132,14 @@ var waitForSubprocessResult = async ({
     ipc,
     ipcInput
   },
-  context,
+  context: context2,
   verboseInfo,
   fileDescriptors,
   originalStreams,
   onInternalError,
   controller
 }) => {
-  const exitPromise = waitForExit(subprocess, context);
+  const exitPromise = waitForExit(subprocess, context2);
   const streamInfo = {
     originalStreams,
     fileDescriptors,
@@ -39080,12 +39192,12 @@ var waitForSubprocessResult = async ({
       ]),
       onInternalError,
       throwOnSubprocessError(subprocess, controller),
-      ...throwOnTimeout(subprocess, timeout, context, controller),
+      ...throwOnTimeout(subprocess, timeout, context2, controller),
       ...throwOnCancel({
         subprocess,
         cancelSignal,
         gracefulCancel,
-        context,
+        context: context2,
         controller
       }),
       ...throwOnGracefulCancel({
@@ -39093,12 +39205,12 @@ var waitForSubprocessResult = async ({
         cancelSignal,
         gracefulCancel,
         forceKillAfterDelay,
-        context,
+        context: context2,
         controller
       })
     ]);
   } catch (error51) {
-    context.terminationReason ??= "other";
+    context2.terminationReason ??= "other";
     return Promise.all([
       { error: error51 },
       exitPromise,
@@ -39144,10 +39256,10 @@ var initializeConcurrentStreams = () => ({
   const promises = weakMap.get(stream);
   const promise2 = createDeferred();
   promises.push(promise2);
-  const resolve2 = promise2.resolve.bind(promise2);
-  return { resolve: resolve2, promises };
-}, waitForConcurrentStreams = async ({ resolve: resolve2, promises }, subprocess) => {
-  resolve2();
+  const resolve3 = promise2.resolve.bind(promise2);
+  return { resolve: resolve3, promises };
+}, waitForConcurrentStreams = async ({ resolve: resolve3, promises }, subprocess) => {
+  resolve3();
   const [isSubprocessExit] = await Promise.race([
     Promise.allSettled([true, subprocess]),
     Promise.all([false, ...promises])
@@ -39529,13 +39641,13 @@ var execaCoreAsync = (rawFile, rawArguments, rawOptions, createNested) => {
   const originalStreams = [...subprocess.stdio];
   pipeOutputAsync(subprocess, fileDescriptors, controller);
   cleanupOnExit(subprocess, options2, controller);
-  const context = {};
+  const context2 = {};
   const onInternalError = createDeferred();
   subprocess.kill = subprocessKill.bind(undefined, {
     kill: subprocess.kill.bind(subprocess),
     options: options2,
     onInternalError,
-    context,
+    context: context2,
     controller
   });
   subprocess.all = makeAllStream(subprocess, options2);
@@ -39550,12 +39662,12 @@ var execaCoreAsync = (rawFile, rawArguments, rawOptions, createNested) => {
     originalStreams,
     command,
     escapedCommand,
-    context,
+    context: context2,
     onInternalError,
     controller
   });
   return { subprocess, promise: promise2 };
-}, handlePromise = async ({ subprocess, options: options2, startTime, verboseInfo, fileDescriptors, originalStreams, command, escapedCommand, context, onInternalError, controller }) => {
+}, handlePromise = async ({ subprocess, options: options2, startTime, verboseInfo, fileDescriptors, originalStreams, command, escapedCommand, context: context2, onInternalError, controller }) => {
   const [
     errorInfo,
     [exitCode, signal],
@@ -39565,7 +39677,7 @@ var execaCoreAsync = (rawFile, rawArguments, rawOptions, createNested) => {
   ] = await waitForSubprocessResult({
     subprocess,
     options: options2,
-    context,
+    context: context2,
     verboseInfo,
     fileDescriptors,
     originalStreams,
@@ -39583,22 +39695,22 @@ var execaCoreAsync = (rawFile, rawArguments, rawOptions, createNested) => {
     stdio,
     all,
     ipcOutput,
-    context,
+    context: context2,
     options: options2,
     command,
     escapedCommand,
     startTime
   });
   return handleResult2(result, verboseInfo, options2);
-}, getAsyncResult = ({ errorInfo, exitCode, signal, stdio, all, ipcOutput, context, options: options2, command, escapedCommand, startTime }) => ("error" in errorInfo) ? makeError({
+}, getAsyncResult = ({ errorInfo, exitCode, signal, stdio, all, ipcOutput, context: context2, options: options2, command, escapedCommand, startTime }) => ("error" in errorInfo) ? makeError({
   error: errorInfo.error,
   command,
   escapedCommand,
-  timedOut: context.terminationReason === "timeout",
-  isCanceled: context.terminationReason === "cancel" || context.terminationReason === "gracefulCancel",
-  isGracefullyCanceled: context.terminationReason === "gracefulCancel",
+  timedOut: context2.terminationReason === "timeout",
+  isCanceled: context2.terminationReason === "cancel" || context2.terminationReason === "gracefulCancel",
+  isGracefullyCanceled: context2.terminationReason === "gracefulCancel",
   isMaxBuffer: errorInfo.error instanceof MaxBufferError,
-  isForcefullyTerminated: context.isForcefullyTerminated,
+  isForcefullyTerminated: context2.isForcefullyTerminated,
   exitCode,
   signal,
   stdio,
@@ -39903,18 +40015,18 @@ class ObservedPipeProcess {
   inner;
   killedWith;
   exited;
-  constructor(inner, events, context) {
+  constructor(inner, events, context2) {
     this.inner = inner;
     this.exited = inner.exited.then((exitCode) => {
       events?.emit("process.exited", {
-        command: context.command,
-        args: context.args,
+        command: context2.command,
+        args: context2.args,
         exitCode,
         ...this.killedWith !== undefined ? { signal: String(this.killedWith) } : {},
-        durationMs: Date.now() - context.startedAt,
+        durationMs: Date.now() - context2.startedAt,
         reason: this.killedWith !== undefined ? "signal" : "exit",
         timestamp: new Date().toISOString(),
-        ...context.label !== undefined ? { label: context.label } : {}
+        ...context2.label !== undefined ? { label: context2.label } : {}
       });
       return exitCode;
     });
@@ -40001,8 +40113,11 @@ function isTimedOut(error51) {
 function errorMessage(error51) {
   return error51 instanceof Error ? error51.message : String(error51);
 }
+var NodeProcessExecutor;
 var init_process_executor = __esm(() => {
   init_execa();
+  NodeProcessExecutor = class NodeProcessExecutor extends ProcessExecutor {
+  };
 });
 
 // ../../node_modules/.bun/@gobing-ai+ts-runtime@0.3.21/node_modules/@gobing-ai/ts-runtime/dist/context.js
@@ -40159,6 +40274,7 @@ var init_types3 = () => {};
 var init_dist4 = __esm(() => {
   init_file_system_node();
   init_process_executor();
+  init_process_executor();
   init_config();
   init_context2();
   init_path();
@@ -40234,19 +40350,101 @@ var init_slash_command = __esm(() => {
 });
 
 // ../../node_modules/.bun/@gobing-ai+ts-ai-runner@0.3.21+5a4f8cc6c6b198a4/node_modules/@gobing-ai/ts-ai-runner/dist/ai-runner.js
+class AiRunner {
+  processExecutor;
+  defaultCwd;
+  defaultTimeout;
+  logger;
+  events;
+  constructor(options2 = {}) {
+    this.processExecutor = options2.processExecutor ?? new NodeProcessExecutor({
+      ...options2.processEvents !== undefined ? {
+        events: {
+          emit: (event, detail) => {
+            options2.processEvents?.emit(event, detail);
+          }
+        }
+      } : {},
+      tracer: options2.tracer ?? {
+        traceAsync: async (name, fn) => await traceAsync(name, (span) => fn(span))
+      }
+    });
+    this.defaultCwd = options2.defaultCwd;
+    this.defaultTimeout = options2.defaultTimeout;
+    this.logger = options2.logger ?? getLogger2("ai-runner");
+    this.events = options2.events;
+  }
+  runHelpCommand(agent, options2 = {}) {
+    return this.invoke(agent, "help", getAgentShim(agent).getHelpCommand(), options2, true);
+  }
+  runVersionCommand(agent, options2 = {}) {
+    return this.invoke(agent, "version", getAgentShim(agent).getVersionCommand(), options2, true);
+  }
+  runPromptCommand(agent, promptOptions, options2 = {}) {
+    return this.invoke(agent, "prompt", this.buildPromptCommand(agent, promptOptions, options2), options2, false);
+  }
+  runSlashCommand(agent, input, promptOptions, options2 = {}) {
+    return this.runPromptCommand(agent, { ...promptOptions, input: translateSlashCommand(agent, input) }, options2);
+  }
+  buildPromptCommand(agent, promptOptions, options2 = {}) {
+    return buildAgentCommand(agent, promptOptions, {
+      workspace: options2.cwd ?? this.defaultCwd ?? getProcessCwd()
+    });
+  }
+  runAuthCommand(agent, options2 = {}) {
+    const command = getAgentShim(agent).getAuthCommand();
+    return command === null ? null : this.invoke(agent, "auth", command, options2, true);
+  }
+  async invoke(agent, operation, command, options2, forceBuffered) {
+    const label = `ai-runner.${agent}.${operation}`;
+    this.logger.debug("invoke", { label, command: command.command, args: command.args.join(" ") });
+    this.events?.emit("agent.invoke.start", { agent, operation, label });
+    const result = await this.processExecutor.run({
+      command: command.command,
+      args: command.args,
+      label,
+      rejectOnError: false,
+      forceBuffered,
+      cwd: options2.cwd ?? this.defaultCwd,
+      timeout: options2.timeout ?? this.defaultTimeout
+    });
+    if (result.exitCode !== 0) {
+      this.logger.error("invoke exited non-zero", {
+        label,
+        exitCode: result.exitCode,
+        signal: result.signal
+      });
+    }
+    this.events?.emit("agent.invoke.exit", {
+      agent,
+      operation,
+      label,
+      exitCode: result.exitCode,
+      ...result.signal !== undefined ? { signal: result.signal } : {},
+      durationMs: result.durationMs
+    });
+    return {
+      exitCode: result.exitCode,
+      stdout: result.stdout,
+      stderr: result.stderr,
+      ...result.signal !== undefined ? { signal: result.signal } : {},
+      durationMs: result.durationMs
+    };
+  }
+}
 function hasIdentityOptions(options2) {
   return options2.purpose !== undefined || options2.systemPrompt !== undefined || options2.taskId !== undefined || options2.peers !== undefined && options2.peers.length > 0;
 }
-function buildAgentCommand(agent, promptOptions, context2) {
-  return getAgentShim(agent).getPromptCommand(applyIdentityPreamble(agent, promptOptions, context2));
+function buildAgentCommand(agent, promptOptions, context3) {
+  return getAgentShim(agent).getPromptCommand(applyIdentityPreamble(agent, promptOptions, context3));
 }
-function applyIdentityPreamble(agent, promptOptions, context2) {
+function applyIdentityPreamble(agent, promptOptions, context3) {
   if (!hasIdentityOptions(promptOptions))
     return promptOptions;
   const preamble = buildIdentityPreamble({
     agentId: agent,
     agentType: agent,
-    workspace: context2.workspace,
+    workspace: context3.workspace,
     purpose: promptOptions.purpose,
     systemPrompt: promptOptions.systemPrompt,
     taskId: promptOptions.taskId,
@@ -40257,8 +40455,11 @@ ${promptOptions.input}`;
   return { ...promptOptions, input };
 }
 var init_ai_runner = __esm(() => {
+  init_dist3();
+  init_dist4();
   init_shims();
   init_identity2();
+  init_slash_command();
 });
 
 // ../../node_modules/.bun/@gobing-ai+ts-ai-runner@0.3.21+5a4f8cc6c6b198a4/node_modules/@gobing-ai/ts-ai-runner/dist/agent-detector.js
@@ -40409,8 +40610,8 @@ class TeamAgentProcess {
       this.warn("stdin close failed", "stop.endStdin", error51);
     }
     process11.kill("SIGTERM");
-    const timeout = new Promise((resolve2) => {
-      setTimeout(() => resolve2("timeout"), 5000);
+    const timeout = new Promise((resolve3) => {
+      setTimeout(() => resolve3("timeout"), 5000);
     });
     const result = await Promise.race([process11.exited, timeout]);
     if (result === "timeout") {
@@ -40766,7 +40967,7 @@ function scoreSkillLinkage(body) {
 function scoreModelFit(data) {
   const model = data.model;
   const recognizedAliases = { inherit: true, sonnet: true, opus: true, haiku: true };
-  const wellFormedPattern = /^claude-(sonnet|opus|haiku)-/i;
+  const wellFormedPattern = /^claude-(\d+-\d+-)?(sonnet|opus|haiku)-/i;
   let score;
   let note;
   if (typeof model === "string") {
@@ -40957,6 +41158,127 @@ var init_command3 = __esm(() => {
   init_types2();
 });
 
+// ../../packages/core/src/quality/eval-cases.ts
+import { existsSync as existsSync10, readFileSync as readFileSync9 } from "fs";
+import { join as join8 } from "path";
+function isRuleJudge(value) {
+  return typeof value === "object" && value !== null && "checks" in value && Array.isArray(value.checks);
+}
+function isRubricRef(value) {
+  return typeof value === "object" && value !== null && "criterion" in value && typeof value.criterion === "string" && !("checks" in value);
+}
+function resolveEvalCasesPath(name, opts) {
+  if (opts?.path) {
+    if (!existsSync10(opts.path))
+      return null;
+    return opts.path;
+  }
+  assertSafePathSegment(name, "eval cases skill name");
+  const coLocated = join8(process.cwd(), "skills", name, "eval", "cases.yaml");
+  if (existsSync10(coLocated))
+    return coLocated;
+  return null;
+}
+function loadEvalCases(name, opts) {
+  const resolvedPath = resolveEvalCasesPath(name, opts);
+  if (!resolvedPath)
+    return null;
+  const raw = readFileSync9(resolvedPath, "utf-8");
+  let parsed;
+  try {
+    parsed = $parse(raw);
+  } catch (e) {
+    throw new EvalCaseError(resolvedPath, null, `Failed to parse YAML: ${e instanceof Error ? e.message : String(e)}`);
+  }
+  const result = EvalCaseSetSchema.safeParse(parsed);
+  if (!result.success) {
+    const issue2 = result.error.issues[0];
+    if (!issue2) {
+      throw new EvalCaseError(resolvedPath, null, "Schema validation failed with no issue reported.");
+    }
+    const caseId = extractCaseIdFromPath(issue2.path, parsed);
+    const field = issue2.path.length > 0 ? issue2.path.join(".") : "root";
+    throw new EvalCaseError(resolvedPath, caseId, `Schema validation failed at "${field}": ${issue2.message}`);
+  }
+  const caseSet = result.data;
+  const seen = new Set;
+  for (const c3 of caseSet.cases) {
+    if (seen.has(c3.id)) {
+      throw new EvalCaseError(resolvedPath, c3.id, `Duplicate case id "${c3.id}"`);
+    }
+    seen.add(c3.id);
+  }
+  return caseSet;
+}
+function extractCaseIdFromPath(path7, parsed) {
+  const casesIdx = path7.indexOf("cases");
+  if (casesIdx >= 0 && path7.length > casesIdx + 1 && typeof path7[casesIdx + 1] === "number") {
+    const caseIndex = path7[casesIdx + 1];
+    if (typeof parsed === "object" && parsed !== null && "cases" in parsed && Array.isArray(parsed.cases)) {
+      const candidate = parsed.cases[caseIndex];
+      if (typeof candidate === "object" && candidate !== null && "id" in candidate && typeof candidate.id === "string") {
+        return candidate.id;
+      }
+    }
+    return `cases[${caseIndex}]`;
+  }
+  return null;
+}
+var RuleCheckSchema, RubricRefSchema, RuleJudgeSchema, EvalCaseSchema, EvalCaseSetSchema, EvalCaseError;
+var init_eval_cases = __esm(() => {
+  init_dist2();
+  init_zod();
+  init_identity();
+  RuleCheckSchema = exports_external.object({
+    op: exports_external.enum(["contains", "regex", "equals", "not_contains", "tool_called"]),
+    arg: exports_external.string().min(1)
+  });
+  RubricRefSchema = exports_external.object({
+    criterion: exports_external.string().min(1),
+    excellent: exports_external.string().optional(),
+    poor: exports_external.string().optional()
+  });
+  RuleJudgeSchema = exports_external.object({
+    checks: exports_external.array(RuleCheckSchema).min(1)
+  });
+  EvalCaseSchema = exports_external.object({
+    id: exports_external.string().min(1),
+    split: exports_external.enum(["train", "holdout"]),
+    prompt: exports_external.string().min(1),
+    reference_kind: exports_external.enum(["exact", "rule", "rubric"]),
+    reference: exports_external.union([exports_external.string().min(1), RuleJudgeSchema, RubricRefSchema]),
+    tags: exports_external.array(exports_external.string().min(1)).optional()
+  }).refine((c3) => {
+    if (c3.reference_kind === "exact" && typeof c3.reference !== "string") {
+      return false;
+    }
+    if (c3.reference_kind === "rule" && !isRuleJudge(c3.reference)) {
+      return false;
+    }
+    if (c3.reference_kind === "rubric" && !isRubricRef(c3.reference)) {
+      return false;
+    }
+    return true;
+  }, {
+    message: 'reference must match reference_kind: string for "exact", RuleJudge for "rule", RubricRef for "rubric"'
+  });
+  EvalCaseSetSchema = exports_external.object({
+    version: exports_external.literal(1),
+    cases: exports_external.array(EvalCaseSchema).min(1)
+  });
+  EvalCaseError = class EvalCaseError extends Error {
+    file;
+    caseId;
+    constructor(file2, caseId, message) {
+      const prefix = caseId ? `${file2}: case "${caseId}" \u2014 ` : `${file2}: `;
+      super(`${prefix}${message}`);
+      this.file = file2;
+      this.caseId = caseId;
+      this.name = "EvalCaseError";
+    }
+  };
+});
+
 // ../../packages/core/src/quality/magent.ts
 function scoreCompleteness3(body) {
   let found = 0;
@@ -40993,7 +41315,7 @@ function scorePlatformCoverage(data, body) {
       [/opencode/i, "opencode"],
       [/openclaw/i, "openclaw"],
       [/antigravity/i, "antigravity"],
-      [/pi\b/i, "pi"]
+      [/\bpi\b/i, "pi"]
     ];
     for (const [re, name] of platformPatterns) {
       if (re.test(body))
@@ -41038,7 +41360,9 @@ function scoreSafety2(body) {
 }
 function evaluateMagent(content, target) {
   const body = extractBody(content);
-  const hasFrontmatter = /^---\s*$/m.test(content);
+  const hasFrontmatter = content.startsWith(`---
+`) || content.startsWith(`---\r
+`);
   const fmResult = hasFrontmatter ? parseFrontmatterSafe(content) : undefined;
   const data = fmResult ?? {};
   const fmNote = hasFrontmatter && fmResult === null ? parseErrorNote(content, "Frontmatter parse error") : null;
@@ -41180,7 +41504,7 @@ function countTriggerPhrases(body) {
     if (headingMatch) {
       const depth = headingMatch[1]?.length ?? 0;
       const title = (headingMatch[2] ?? "").toLowerCase();
-      if (/\btrigger|when to use/i.test(title)) {
+      if (/\b(?:trigger|when to use)/i.test(title)) {
         inTriggerSection = true;
         sectionDepth = depth;
       } else if (inTriggerSection && depth <= sectionDepth) {
@@ -41196,7 +41520,7 @@ function countTriggerPhrases(body) {
   }
   if (count2 === 0) {
     for (const line of lines) {
-      if (/^\s*[-*+]\s/.test(line)) {
+      if (/^\s*[-*+]\s/.test(line) || /^\s*\d+[.)]\s/.test(line)) {
         count2++;
       }
     }
@@ -41228,29 +41552,77 @@ var init_evaluate = __esm(() => {
   };
 });
 
+// ../../packages/core/src/quality/replay.ts
+function scoreExact(output2, reference) {
+  const normOut = normalize(output2);
+  const normRef = normalize(reference);
+  if (normOut === normRef)
+    return 1;
+  if (normOut.includes(normRef))
+    return 1;
+  return 0;
+}
+function scoreRule(output2, judge, toolsCalled) {
+  for (const check2 of judge.checks) {
+    if (!evaluateCheck(output2, check2.op, check2.arg, toolsCalled))
+      return 0;
+  }
+  return 1;
+}
+function aggregateHard(scores) {
+  if (scores.length === 0)
+    return 0;
+  const sum = scores.reduce((acc, s) => acc + s, 0);
+  return sum / scores.length;
+}
+function normalize(text) {
+  return text.trim().toLowerCase().replace(/\s+/g, " ");
+}
+function evaluateCheck(output2, op, arg, toolsCalled) {
+  switch (op) {
+    case "contains":
+      return output2.toLowerCase().includes(arg.toLowerCase());
+    case "regex": {
+      try {
+        return new RegExp(arg).test(output2);
+      } catch {
+        return false;
+      }
+    }
+    case "equals":
+      return normalize(output2) === normalize(arg);
+    case "not_contains":
+      return !output2.toLowerCase().includes(arg.toLowerCase());
+    case "tool_called":
+      return (toolsCalled ?? []).includes(arg);
+    default:
+      return false;
+  }
+}
+
 // ../../packages/core/src/quality/rubric.ts
-import { existsSync as existsSync10, readFileSync as readFileSync9 } from "fs";
-import { homedir as homedir3 } from "os";
-import { join as join8 } from "path";
+import { existsSync as existsSync11, readFileSync as readFileSync10 } from "fs";
+import { homedir as homedir4 } from "os";
+import { join as join9 } from "path";
 function resolveRubricContent(type, opts) {
   if (opts?.path) {
-    if (!existsSync10(opts.path)) {
+    if (!existsSync11(opts.path)) {
       throw new RubricError("path", `Rubric file not found: ${opts.path}`, opts.path);
     }
-    return readFileSync9(opts.path, "utf-8");
+    return readFileSync10(opts.path, "utf-8");
   }
-  const homeDir = process.env.HOME ?? homedir3();
-  const userPath = join8(homeDir, ".superskill", "rubrics", `${type}.yaml`);
-  if (existsSync10(userPath)) {
-    return readFileSync9(userPath, "utf-8");
+  const homeDir = process.env.HOME || homedir4();
+  const userPath = join9(homeDir, ".superskill", "rubrics", `${type}.yaml`);
+  if (existsSync11(userPath)) {
+    return readFileSync10(userPath, "utf-8");
   }
-  const devPath = join8(import.meta.dir, "..", "rubrics", `${type}.yaml`);
-  if (existsSync10(devPath)) {
-    return readFileSync9(devPath, "utf-8");
+  const devPath = join9(import.meta.dir, "..", "rubrics", `${type}.yaml`);
+  if (existsSync11(devPath)) {
+    return readFileSync10(devPath, "utf-8");
   }
-  const prodPath = join8(import.meta.dir, "..", "..", "rubrics", `${type}.yaml`);
-  if (existsSync10(prodPath)) {
-    return readFileSync9(prodPath, "utf-8");
+  const prodPath = join9(import.meta.dir, "..", "..", "rubrics", `${type}.yaml`);
+  if (existsSync11(prodPath)) {
+    return readFileSync10(prodPath, "utf-8");
   }
   throw new RubricError("path", `No rubric found for type "${type}". Searched: ${opts?.path ?? "(no explicit path)"}, ${userPath}, ${devPath}, ${prodPath}`, type);
 }
@@ -47079,11 +47451,11 @@ var require_out = __commonJS((exports) => {
     async.read(path7, getSettings(optionsOrSettingsOrCallback), callback);
   }
   exports.stat = stat;
-  function statSync7(path7, optionsOrSettings) {
+  function statSync6(path7, optionsOrSettings) {
     const settings = getSettings(optionsOrSettings);
     return sync.read(path7, settings);
   }
-  exports.statSync = statSync7;
+  exports.statSync = statSync6;
   function getSettings(settingsOrOptions = {}) {
     if (settingsOrOptions instanceof settings_1.default) {
       return settingsOrOptions;
@@ -47487,11 +47859,11 @@ var require_reusify = __commonJS((exports, module) => {
 // ../../node_modules/.bun/fastq@1.20.1/node_modules/fastq/queue.js
 var require_queue = __commonJS((exports, module) => {
   var reusify = require_reusify();
-  function fastqueue(context2, worker, _concurrency) {
-    if (typeof context2 === "function") {
+  function fastqueue(context3, worker, _concurrency) {
+    if (typeof context3 === "function") {
       _concurrency = worker;
-      worker = context2;
-      context2 = null;
+      worker = context3;
+      context3 = null;
     }
     if (!(_concurrency >= 1)) {
       throw new Error("fastqueue concurrency must be equal to or greater than 1");
@@ -47578,7 +47950,7 @@ var require_queue = __commonJS((exports, module) => {
     }
     function push(value, done) {
       var current = cache.get();
-      current.context = context2;
+      current.context = context3;
       current.release = release;
       current.value = value;
       current.callback = done || noop4;
@@ -47594,12 +47966,12 @@ var require_queue = __commonJS((exports, module) => {
         }
       } else {
         _running++;
-        worker.call(context2, current.value, current.worked);
+        worker.call(context3, current.value, current.worked);
       }
     }
     function unshift(value, done) {
       var current = cache.get();
-      current.context = context2;
+      current.context = context3;
       current.release = release;
       current.value = value;
       current.callback = done || noop4;
@@ -47615,7 +47987,7 @@ var require_queue = __commonJS((exports, module) => {
         }
       } else {
         _running++;
-        worker.call(context2, current.value, current.worked);
+        worker.call(context3, current.value, current.worked);
       }
     }
     function release(holder) {
@@ -47630,7 +48002,7 @@ var require_queue = __commonJS((exports, module) => {
           }
           queueHead = next.next;
           next.next = null;
-          worker.call(context2, next.value, next.worked);
+          worker.call(context3, next.value, next.worked);
           if (queueTail === null) {
             self2.empty();
           }
@@ -47661,14 +48033,14 @@ var require_queue = __commonJS((exports, module) => {
         var callback = current.callback;
         var errorHandler2 = current.errorHandler;
         var val = current.value;
-        var context3 = current.context;
+        var context4 = current.context;
         current.value = null;
         current.callback = noop4;
         current.errorHandler = null;
         if (errorHandler2) {
           errorHandler2(new Error("abort"), val);
         }
-        callback.call(context3, new Error("abort"));
+        callback.call(context4, new Error("abort"));
         current.release(current);
         current = next;
       }
@@ -47700,18 +48072,18 @@ var require_queue = __commonJS((exports, module) => {
       self2.release(self2);
     };
   }
-  function queueAsPromised(context2, worker, _concurrency) {
-    if (typeof context2 === "function") {
+  function queueAsPromised(context3, worker, _concurrency) {
+    if (typeof context3 === "function") {
       _concurrency = worker;
-      worker = context2;
-      context2 = null;
+      worker = context3;
+      context3 = null;
     }
     function asyncWrapper(arg, cb) {
       worker.call(this, arg).then(function(res) {
         cb(null, res);
       }, cb);
     }
-    var queue = fastqueue(context2, asyncWrapper, _concurrency);
+    var queue = fastqueue(context3, asyncWrapper, _concurrency);
     var pushCb = queue.push;
     var unshiftCb = queue.unshift;
     queue.push = push;
@@ -47719,42 +48091,42 @@ var require_queue = __commonJS((exports, module) => {
     queue.drained = drained;
     return queue;
     function push(value) {
-      var p = new Promise(function(resolve2, reject) {
+      var p = new Promise(function(resolve3, reject) {
         pushCb(value, function(err, result) {
           if (err) {
             reject(err);
             return;
           }
-          resolve2(result);
+          resolve3(result);
         });
       });
       p.catch(noop4);
       return p;
     }
     function unshift(value) {
-      var p = new Promise(function(resolve2, reject) {
+      var p = new Promise(function(resolve3, reject) {
         unshiftCb(value, function(err, result) {
           if (err) {
             reject(err);
             return;
           }
-          resolve2(result);
+          resolve3(result);
         });
       });
       p.catch(noop4);
       return p;
     }
     function drained() {
-      var p = new Promise(function(resolve2) {
+      var p = new Promise(function(resolve3) {
         process.nextTick(function() {
           if (queue.idle()) {
-            resolve2();
+            resolve3();
           } else {
             var previousDrain = queue.drain;
             queue.drain = function() {
               if (typeof previousDrain === "function")
                 previousDrain();
-              resolve2();
+              resolve3();
               queue.drain = previousDrain;
             };
           }
@@ -48215,9 +48587,9 @@ var require_stream3 = __commonJS((exports) => {
       });
     }
     _getStat(filepath) {
-      return new Promise((resolve2, reject) => {
+      return new Promise((resolve3, reject) => {
         this._stat(filepath, this._fsStatSettings, (error51, stats) => {
-          return error51 === null ? resolve2(stats) : reject(error51);
+          return error51 === null ? resolve3(stats) : reject(error51);
         });
       });
     }
@@ -48239,10 +48611,10 @@ var require_async5 = __commonJS((exports) => {
       this._readerStream = new stream_1.default(this._settings);
     }
     dynamic(root, options2) {
-      return new Promise((resolve2, reject) => {
+      return new Promise((resolve3, reject) => {
         this._walkAsync(root, options2, (error51, entries) => {
           if (error51 === null) {
-            resolve2(entries);
+            resolve3(entries);
           } else {
             reject(error51);
           }
@@ -48252,10 +48624,10 @@ var require_async5 = __commonJS((exports) => {
     async static(patterns, options2) {
       const entries = [];
       const stream = this._readerStream.static(patterns, options2);
-      return new Promise((resolve2, reject) => {
+      return new Promise((resolve3, reject) => {
         stream.once("error", reject);
         stream.on("data", (entry) => entries.push(entry));
-        stream.once("end", () => resolve2(entries));
+        stream.once("end", () => resolve3(entries));
       });
     }
   }
@@ -49783,7 +50155,7 @@ var import_fast_glob2, import_ignore, defaultIgnoredDirectories, ignoreFilesGlob
     }
     throw createGitConfigReadError(normalizedPath, error51);
   }
-}, getExcludesFileFromGitConfigSync = (filePath, readFileSync10, gitDirectory, options2 = {}) => {
+}, getExcludesFileFromGitConfigSync = (filePath, readFileSync11, gitDirectory, options2 = {}) => {
   const { suppressErrors, includeStack = new Set, depth = 0 } = options2;
   const normalizedPath = path9.resolve(filePath);
   if (includeStack.has(normalizedPath)) {
@@ -49793,14 +50165,14 @@ var import_fast_glob2, import_ignore, defaultIgnoredDirectories, ignoreFilesGlob
     return;
   }
   includeStack.add(normalizedPath);
-  const content = readGitConfigFile(normalizedPath, readFileSync10, suppressErrors);
+  const content = readGitConfigFile(normalizedPath, readFileSync11, suppressErrors);
   if (content === undefined) {
     includeStack.delete(normalizedPath);
     return;
   }
   let { excludesFile, includePaths } = parseGitConfigForExcludesFile(content, normalizedPath, gitDirectory);
   for (const includePath of includePaths) {
-    const includedExcludesFile = getExcludesFileFromGitConfigSync(includePath, readFileSync10, gitDirectory, { suppressErrors, includeStack, depth: depth + 1 });
+    const includedExcludesFile = getExcludesFileFromGitConfigSync(includePath, readFileSync11, gitDirectory, { suppressErrors, includeStack, depth: depth + 1 });
     if (includedExcludesFile !== undefined) {
       excludesFile = includedExcludesFile;
     }
@@ -49842,13 +50214,13 @@ var import_fast_glob2, import_ignore, defaultIgnoredDirectories, ignoreFilesGlob
     return gitFilePath;
   }
   return path9.resolve(path9.dirname(gitFilePath), match[1]);
-}, getGitDirectorySync = (gitRoot, readFileSync10) => {
+}, getGitDirectorySync = (gitRoot, readFileSync11) => {
   if (!gitRoot) {
     return;
   }
   const gitFilePath = path9.join(gitRoot, ".git");
   try {
-    return resolveGitDirectoryFromFile(gitFilePath, readFileSync10(gitFilePath, "utf8"));
+    return resolveGitDirectoryFromFile(gitFilePath, readFileSync11(gitFilePath, "utf8"));
   } catch {
     return gitFilePath;
   }
@@ -49891,18 +50263,18 @@ var import_fast_glob2, import_ignore, defaultIgnoredDirectories, ignoreFilesGlob
   }
 }, getGlobalGitignoreFile = (options2 = {}) => {
   const cwd5 = toPath2(options2.cwd) ?? process11.cwd();
-  const readFileSync10 = getReadFileSyncMethod(options2.fs);
+  const readFileSync11 = getReadFileSyncMethod(options2.fs);
   const gitRoot = findGitRootSync(cwd5, options2.fs);
-  const gitDirectory = getGitDirectorySync(gitRoot, readFileSync10);
+  const gitDirectory = getGitDirectorySync(gitRoot, readFileSync11);
   let excludesFileConfig;
   for (const gitConfigPath of getGitConfigPaths()) {
-    const value = getExcludesFileFromGitConfigSync(gitConfigPath, readFileSync10, gitDirectory, { suppressErrors: options2.suppressErrors });
+    const value = getExcludesFileFromGitConfigSync(gitConfigPath, readFileSync11, gitDirectory, { suppressErrors: options2.suppressErrors });
     if (value !== undefined) {
       excludesFileConfig = value;
     }
   }
   const filePath = resolveExcludesFilePath(excludesFileConfig);
-  return filePath === undefined ? undefined : readGlobalGitignoreContent(filePath, readFileSync10, options2.suppressErrors);
+  return filePath === undefined ? undefined : readGlobalGitignoreContent(filePath, readFileSync11, options2.suppressErrors);
 }, getGlobalGitignoreFileAsync = async (options2 = {}) => {
   const cwd5 = toPath2(options2.cwd) ?? process11.cwd();
   const readFile = getReadFileMethod(options2.fs);
@@ -58242,7 +58614,7 @@ import {
   writeFile
 } from "fs/promises";
 import os2 from "os";
-import { dirname as dirname6, isAbsolute, join as join22, relative, resolve as resolve2 } from "path";
+import { dirname as dirname6, isAbsolute, join as join22, relative, resolve as resolve4 } from "path";
 import { isAbsolute as isAbsolute2, resolve as resolve22 } from "path";
 import { basename as basename3, join as join44, relative as relative3 } from "path";
 import { extname as extname2 } from "path";
@@ -58250,8 +58622,8 @@ import { isDeepStrictEqual } from "util";
 import { join as join62 } from "path";
 import { join as join42 } from "path";
 import { join as join52 } from "path";
-import path10, { relative as relative2, resolve as resolve4 } from "path";
-import { basename as basename4, join as join9 } from "path";
+import path10, { relative as relative2, resolve as resolve42 } from "path";
+import { basename as basename4, join as join92 } from "path";
 import { join as join72 } from "path";
 import { join as join82 } from "path";
 import { join as join10 } from "path";
@@ -58335,7 +58707,7 @@ import { join as join88 } from "path";
 import { join as join89 } from "path";
 import { join as join90 } from "path";
 import { join as join91 } from "path";
-import { join as join92 } from "path";
+import { join as join922 } from "path";
 import { join as join93 } from "path";
 import { join as join94 } from "path";
 import { join as join95 } from "path";
@@ -58510,9 +58882,9 @@ function checkPathTraversal({
   if (segments.includes("..")) {
     throw new Error(`Path traversal detected: ${relativePath}`);
   }
-  const resolved = resolve2(intendedRootDir, relativePath);
+  const resolved = resolve4(intendedRootDir, relativePath);
   const rel = relative(intendedRootDir, resolved);
-  if (rel.startsWith("..") || resolve2(resolved) !== resolved) {
+  if (rel.startsWith("..") || resolve4(resolved) !== resolved) {
     throw new Error(`Path traversal detected: ${relativePath}`);
   }
 }
@@ -58520,7 +58892,7 @@ function resolvePath3(relativePath, outputRoot) {
   if (!outputRoot)
     return relativePath;
   checkPathTraversal({ relativePath, intendedRootDir: outputRoot });
-  return resolve2(outputRoot, relativePath);
+  return resolve4(outputRoot, relativePath);
 }
 async function directoryExists(dirPath) {
   try {
@@ -58633,7 +59005,7 @@ function validateOutputRoot(outputRoot) {
     if (segments.includes("..")) {
       throw new Error(`Path traversal detected: ${outputRoot}`);
     }
-    const normalized = resolve2(outputRoot);
+    const normalized = resolve4(outputRoot);
     if (normalized !== outputRoot) {
       throw new Error(`outputRoot must be a normalized absolute path: ${outputRoot} (normalized: ${normalized})`);
     }
@@ -63025,8 +63397,8 @@ var import_gray_matter, BaseLogger = class {
   }
   getFilePath() {
     const fullPath = path10.join(this.outputRoot, this.relativeDirPath, this.relativeFilePath);
-    const resolvedFull = resolve4(fullPath);
-    const resolvedBase = resolve4(this.outputRoot);
+    const resolvedFull = resolve42(fullPath);
+    const resolvedBase = resolve42(this.outputRoot);
     const rel = relative2(resolvedBase, resolvedFull);
     if (rel.startsWith("..") || path10.isAbsolute(rel)) {
       throw new Error(`Path traversal detected: Final path escapes outputRoot. outputRoot="${this.outputRoot}", relativeDirPath="${this.relativeDirPath}", relativeFilePath="${this.relativeFilePath}"`);
@@ -63958,7 +64330,7 @@ var init_chunk_QKQNZ6C3 = __esm(() => {
       if (rest2.validate) {
         const result = AntigravityCommandFrontmatterSchema.safeParse(frontmatter);
         if (!result.success) {
-          throw new Error(`Invalid frontmatter in ${join9(rest2.relativeDirPath, rest2.relativeFilePath)}: ${formatError2(result.error)}`);
+          throw new Error(`Invalid frontmatter in ${join92(rest2.relativeDirPath, rest2.relativeFilePath)}: ${formatError2(result.error)}`);
         }
       }
       super({
@@ -64054,7 +64426,7 @@ ${body}${turboDirective}`;
       } else {
         return {
           success: false,
-          error: new Error(`Invalid frontmatter in ${join9(this.relativeDirPath, this.relativeFilePath)}: ${formatError2(result.error)}`)
+          error: new Error(`Invalid frontmatter in ${join92(this.relativeDirPath, this.relativeFilePath)}: ${formatError2(result.error)}`)
         };
       }
     }
@@ -64069,7 +64441,7 @@ ${body}${turboDirective}`;
       relativeFilePath,
       validate: validate2 = true
     }) {
-      const filePath = join9(outputRoot, _AntigravityCommand.getSettablePaths().relativeDirPath, relativeFilePath);
+      const filePath = join92(outputRoot, _AntigravityCommand.getSettablePaths().relativeDirPath, relativeFilePath);
       const fileContent = await readFileContent(filePath);
       const { frontmatter, body: content } = parseFrontmatter2(fileContent, filePath);
       const result = AntigravityCommandFrontmatterSchema.safeParse(frontmatter);
@@ -71258,7 +71630,7 @@ prompt = ""`;
         try {
           this.json = JSON.parse(this.fileContent);
         } catch (error51) {
-          throw new Error(`Failed to parse Cursor MCP config at ${join92(this.relativeDirPath, this.relativeFilePath)}: ${formatError2(error51)}`, { cause: error51 });
+          throw new Error(`Failed to parse Cursor MCP config at ${join922(this.relativeDirPath, this.relativeFilePath)}: ${formatError2(error51)}`, { cause: error51 });
         }
       } else {
         this.json = {};
@@ -71282,13 +71654,13 @@ prompt = ""`;
       global: global3 = false
     }) {
       const paths = this.getSettablePaths({ global: global3 });
-      const filePath = join92(outputRoot, paths.relativeDirPath, paths.relativeFilePath);
+      const filePath = join922(outputRoot, paths.relativeDirPath, paths.relativeFilePath);
       const fileContent = await readFileContentOrNull(filePath) ?? '{"mcpServers":{}}';
       let json3;
       try {
         json3 = JSON.parse(fileContent);
       } catch (error51) {
-        throw new Error(`Failed to parse Cursor MCP config at ${join92(paths.relativeDirPath, paths.relativeFilePath)}: ${formatError2(error51)}`, { cause: error51 });
+        throw new Error(`Failed to parse Cursor MCP config at ${join922(paths.relativeDirPath, paths.relativeFilePath)}: ${formatError2(error51)}`, { cause: error51 });
       }
       const newJson = { ...json3, mcpServers: json3.mcpServers ?? {} };
       return new _CursorMcp({
@@ -71307,12 +71679,12 @@ prompt = ""`;
       global: global3 = false
     }) {
       const paths = this.getSettablePaths({ global: global3 });
-      const fileContent = await readOrInitializeFileContent(join92(outputRoot, paths.relativeDirPath, paths.relativeFilePath), JSON.stringify({ mcpServers: {} }, null, 2));
+      const fileContent = await readOrInitializeFileContent(join922(outputRoot, paths.relativeDirPath, paths.relativeFilePath), JSON.stringify({ mcpServers: {} }, null, 2));
       let json3;
       try {
         json3 = JSON.parse(fileContent);
       } catch (error51) {
-        throw new Error(`Failed to parse Cursor MCP config at ${join92(paths.relativeDirPath, paths.relativeFilePath)}: ${formatError2(error51)}`, { cause: error51 });
+        throw new Error(`Failed to parse Cursor MCP config at ${join922(paths.relativeDirPath, paths.relativeFilePath)}: ${formatError2(error51)}`, { cause: error51 });
       }
       const mcpServers = rulesyncMcp.getMcpServers();
       const transformedServers = convertEnvVarRefsToToolFormat({
@@ -89292,7 +89664,7 @@ var init_dist9 = __esm(() => {
 });
 
 // ../../packages/core/src/rulesync.ts
-import { homedir as homedir4 } from "os";
+import { homedir as homedir5 } from "os";
 async function runRulesync(targets, features, inputRoot, options2) {
   const mappedTargets = [];
   for (const target of targets) {
@@ -89322,7 +89694,7 @@ async function runRulesync(targets, features, inputRoot, options2) {
       hasDiff: false
     };
   }
-  const root = options2.outputRoot ?? (options2.global ? homedir4() : process.cwd());
+  const root = options2.outputRoot ?? (options2.global ? homedir5() : process.cwd());
   const rulesyncGlobal = options2.outputRoot ? false : options2.global;
   return generate2({
     targets: mappedTargets,
@@ -89361,6 +89733,7 @@ var init_src = __esm(() => {
   init_slash_command2();
   init_agent();
   init_command3();
+  init_eval_cases();
   init_evaluate();
   init_heuristics();
   init_hook();
@@ -89740,7 +90113,7 @@ var init_version = () => {};
 
 // ../../node_modules/.bun/drizzle-orm@0.44.7+a0473b45aab9bf33/node_modules/drizzle-orm/tracing.js
 var otel, rawTracer, tracer;
-var init_tracing = __esm(() => {
+var init_tracing2 = __esm(() => {
   init_tracing_utils();
   init_version();
   tracer = {
@@ -89830,7 +90203,7 @@ var init_sql = __esm(() => {
   init_entity();
   init_enum();
   init_subquery();
-  init_tracing();
+  init_tracing2();
   init_view_common();
   init_column();
   init_table();
@@ -94967,13 +95340,13 @@ __export(exports_db, {
   openStore: () => openStore,
   getDBPath: () => getDBPath
 });
-import { existsSync as existsSync11, mkdirSync as mkdirSync6 } from "fs";
+import { existsSync as existsSync12, mkdirSync as mkdirSync6 } from "fs";
 import { dirname as dirname7 } from "path";
 async function openStore(opts) {
   const url3 = getDBPath(opts);
   if (url3 !== ":memory:") {
     const dir = dirname7(url3);
-    if (!existsSync11(dir)) {
+    if (!existsSync12(dir)) {
       mkdirSync6(dir, { recursive: true });
     }
   }
@@ -95000,7 +95373,7 @@ var init_db2 = __esm(() => {
 });
 
 // src/cli.ts
-import { readFileSync as readFileSync15 } from "fs";
+import { readFileSync as readFileSync16 } from "fs";
 import { join as join226 } from "path";
 
 // ../../node_modules/.bun/commander@14.0.3/node_modules/commander/esm.mjs
@@ -95026,7 +95399,7 @@ init_dist();
 init_src();
 init_dist();
 init_db2();
-import { readFileSync as readFileSync10 } from "fs";
+import { readFileSync as readFileSync11 } from "fs";
 
 // src/store/evaluations.ts
 init_dist10();
@@ -95222,7 +95595,7 @@ async function ingestScores(type, resolvedPath, resolvedTarget, opts) {
   }
   let raw;
   try {
-    raw = readFileSync10(ingestPath, "utf-8");
+    raw = readFileSync11(ingestPath, "utf-8");
   } catch {
     throw Object.assign(new Error(`Cannot read scores file: ${ingestPath}`), { code: 2 });
   }
@@ -95342,7 +95715,7 @@ function formatEvaluationReport(report, json3) {
 // src/operations/evolve.ts
 init_src();
 init_dist();
-import { existsSync as existsSync12, mkdirSync as mkdirSync7, readFileSync as readFileSync11, rmSync as rmSync4, writeFileSync as writeFileSync6 } from "fs";
+import { existsSync as existsSync13, mkdirSync as mkdirSync7, readFileSync as readFileSync12, rmSync as rmSync5, writeFileSync as writeFileSync6 } from "fs";
 import { join as join220 } from "path";
 import { createInterface } from "readline";
 
@@ -95410,6 +95783,149 @@ function deserializeProposal(row) {
   };
 }
 
+// src/operations/pairwise-judge.ts
+init_src();
+init_dist5();
+class TsAiRunnerJudgeBackend {
+  runner;
+  agent;
+  constructor(agent2 = "claude", runner) {
+    this.runner = runner ?? new AiRunner;
+    this.agent = agent2;
+  }
+  async judge(rubric2, prompt, candOutput, baseOutput, options2) {
+    const candidateFirst = options2?.order !== "baseline-first";
+    const outputA = candidateFirst ? candOutput : baseOutput;
+    const outputB = candidateFirst ? baseOutput : candOutput;
+    const systemPrompt = [
+      "You are a strict pairwise behavior judge.",
+      "Compare two outputs for the same task against the rubric criterion.",
+      'Return only JSON: {"winner":"A"|"B"|"tie","margin":number}.',
+      "The margin must be a number from 0 to 1."
+    ].join(`
+`);
+    const input = [
+      `Criterion: ${rubric2.criterion}`,
+      rubric2.excellent ? `Excellent anchor: ${rubric2.excellent}` : "",
+      rubric2.poor ? `Poor anchor: ${rubric2.poor}` : "",
+      `Task prompt:
+${prompt}`,
+      `Output A:
+${outputA}`,
+      `Output B:
+${outputB}`
+    ].filter(Boolean).join(`
+
+`);
+    const result = await this.runner.runPromptCommand(this.agent, {
+      input,
+      systemPrompt,
+      ...options2?.seed !== undefined ? { seed: options2.seed } : {},
+      ...options2?.temperature !== undefined ? { temperature: options2.temperature } : {}
+    });
+    return parseJudgeResponse(result.stdout, candidateFirst);
+  }
+}
+function createJudgeBackend(target, injected) {
+  return injected ?? new TsAiRunnerJudgeBackend(TARGET_TO_AGENT_NAME[target]);
+}
+async function pairwiseJudge(backend, rubric2, casePrompt, candOutput, baseOutput, options2 = {}) {
+  const order = options2.order ?? (options2.seed !== undefined && options2.seed % 2 === 1 ? "baseline-first" : "candidate-first");
+  return backend.judge(rubric2, casePrompt, candOutput, baseOutput, { ...options2, order });
+}
+function signedMargin(verdict) {
+  if (verdict.winner === "candidate")
+    return verdict.margin;
+  if (verdict.winner === "baseline")
+    return -verdict.margin;
+  return 0;
+}
+function parseJudgeResponse(stdout, candidateFirst) {
+  let parsed;
+  try {
+    parsed = JSON.parse(stdout.trim());
+  } catch {
+    throw new Error(`Pairwise judge returned invalid JSON: ${stdout.slice(0, 200)}`);
+  }
+  if (typeof parsed !== "object" || parsed === null) {
+    throw new Error("Pairwise judge returned a non-object JSON payload.");
+  }
+  const value = parsed;
+  if (value.winner !== "A" && value.winner !== "B" && value.winner !== "tie") {
+    throw new Error('Pairwise judge JSON must include winner "A", "B", or "tie".');
+  }
+  if (typeof value.margin !== "number" || value.margin < 0 || value.margin > 1) {
+    throw new Error("Pairwise judge JSON must include margin in [0,1].");
+  }
+  if (value.winner === "tie")
+    return { winner: "tie", margin: 0 };
+  if (value.winner === "A") {
+    return { winner: candidateFirst ? "candidate" : "baseline", margin: value.margin };
+  }
+  return { winner: candidateFirst ? "baseline" : "candidate", margin: value.margin };
+}
+
+// src/operations/noise-floor.ts
+async function estimateNoiseFloor(judge, rubric2, casePrompt, candOutput, baseOutput, n2 = 5) {
+  const margins = [];
+  for (let i2 = 0;i2 < n2; i2++) {
+    const verdict = await pairwiseJudge(judge, rubric2, casePrompt, candOutput, baseOutput, {
+      seed: i2 + 1,
+      temperature: 0
+    });
+    margins.push(signedMargin(verdict));
+  }
+  if (margins.length < 2)
+    return 0;
+  const mean2 = margins.reduce((a2, b) => a2 + b, 0) / margins.length;
+  const variance = margins.reduce((sum2, m) => sum2 + (m - mean2) ** 2, 0) / margins.length;
+  return Math.sqrt(variance);
+}
+function rejectsWithinNoise(measuredDelta, noiseFloor) {
+  return Math.abs(measuredDelta) < noiseFloor;
+}
+
+// src/operations/replay-runner.ts
+init_src();
+init_dist5();
+class TsAiRunnerBackend {
+  runner;
+  agent;
+  constructor(agent2 = "claude", runner) {
+    this.runner = runner ?? new AiRunner;
+    this.agent = agent2;
+  }
+  async run(systemPrompt, userPrompt) {
+    const result = await this.runner.runPromptCommand(this.agent, {
+      input: userPrompt,
+      systemPrompt
+    });
+    return result.stdout;
+  }
+}
+function createReplayBackend(target, injected) {
+  return injected ?? new TsAiRunnerBackend(TARGET_TO_AGENT_NAME[target]);
+}
+async function replayCase(backend, skill2, ev, toolsCalled) {
+  const prompt = `[case:${ev.id}] ${ev.prompt}`;
+  const output2 = await backend.run(skill2, prompt);
+  let hard;
+  if (ev.reference_kind === "exact") {
+    hard = scoreExact(output2, ev.reference);
+  } else {
+    hard = scoreRule(output2, ev.reference, toolsCalled);
+  }
+  return { hard, output: output2 };
+}
+async function replaySplit(backend, skill2, cases, split, toolsCalled) {
+  const splitCases = cases.filter((c3) => c3.split === split);
+  if (splitCases.length === 0)
+    return { hard: 0, n: 0 };
+  const results = await Promise.all(splitCases.map((c3) => replayCase(backend, skill2, c3, toolsCalled)));
+  const scores = results.map((r) => r.hard);
+  return { hard: aggregateHard(scores), n: splitCases.length };
+}
+
 // src/operations/validate.ts
 init_src();
 
@@ -95591,6 +96107,106 @@ async function runGate(input) {
       reason: `\u0394-margin gate failed: \u0394${delta >= 0 ? "+" : ""}${delta.toFixed(2)} < required margin ${input.margin.toFixed(2)}.`
     };
   }
+  let empiricalScores;
+  if (input.evalGate) {
+    const cases = loadEvalCases(input.evalGate.name);
+    if (cases) {
+      const holdout = cases.cases.filter((c3) => c3.split === "holdout");
+      const rubricCases = holdout.filter((c3) => c3.reference_kind === "rubric");
+      const detCases = holdout.filter((c3) => c3.reference_kind !== "rubric");
+      const maxModelCalls = input.evalGate.judgeBudget?.maxModelCalls ?? Number.POSITIVE_INFINITY;
+      let modelCalls = 0;
+      const consumeModelCalls = (count3, reason) => {
+        modelCalls += count3;
+        if (modelCalls > maxModelCalls) {
+          return {
+            ok: false,
+            failedGate: "empirical",
+            reason: `Empirical gate budget cap hit: ${modelCalls} model call(s) for ${reason} exceeds max ${maxModelCalls}.`
+          };
+        }
+        return null;
+      };
+      const replayBackend = createReplayBackend(input.evalGate.target, input.evalGate.replayBackend);
+      let detBaseScore = 0;
+      let detCandScore = 0;
+      let detN = 0;
+      if (detCases.length > 0) {
+        const budgetResult = consumeModelCalls(detCases.length * 2, "deterministic replay");
+        if (budgetResult)
+          return budgetResult;
+        const baseDet = await replaySplit(replayBackend, input.evalGate.baselineSkillText, detCases, "holdout");
+        const candDet = await replaySplit(replayBackend, input.evalGate.candidateSkillText, detCases, "holdout");
+        detBaseScore = baseDet.hard;
+        detCandScore = candDet.hard;
+        detN = detCases.length;
+      }
+      let rubricBaseScore = 0;
+      let rubricCandScore = 0;
+      let maxNoiseFloor = 0;
+      let signedDeltaTotal = 0;
+      let rubricN = 0;
+      if (rubricCases.length > 0) {
+        const judge = createJudgeBackend(input.evalGate.target, input.evalGate.judgeBackend);
+        const judgeReplays = input.evalGate.judgeReplays ?? 3;
+        for (const rc of rubricCases) {
+          const replayBudget = consumeModelCalls(2, `rubric replay for ${rc.id}`);
+          if (replayBudget)
+            return replayBudget;
+          const prompt = `[case:${rc.id}] ${rc.prompt}`;
+          const [candOutput, baseOutput] = await Promise.all([
+            replayBackend.run(input.evalGate.candidateSkillText, prompt),
+            replayBackend.run(input.evalGate.baselineSkillText, prompt)
+          ]);
+          const rubricRef = rc.reference;
+          const judgeBudget = consumeModelCalls(1, `pairwise rubric judge for ${rc.id}`);
+          if (judgeBudget)
+            return judgeBudget;
+          const verdict = await pairwiseJudge(judge, rubricRef, prompt, candOutput, baseOutput, {
+            seed: 0,
+            temperature: 0
+          });
+          const noiseBudget = consumeModelCalls(judgeReplays, `noise-floor replay for ${rc.id}`);
+          if (noiseBudget)
+            return noiseBudget;
+          const noiseFloor = await estimateNoiseFloor(judge, rubricRef, prompt, candOutput, baseOutput, judgeReplays);
+          const delta2 = signedMargin(verdict);
+          maxNoiseFloor = Math.max(maxNoiseFloor, noiseFloor);
+          signedDeltaTotal += delta2;
+          if (rejectsWithinNoise(delta2, noiseFloor)) {
+            rubricCandScore += 0.5;
+            rubricBaseScore += 0.5;
+          } else if (verdict.winner === "candidate") {
+            rubricCandScore += 1;
+          } else if (verdict.winner === "baseline") {
+            rubricBaseScore += 1;
+          } else {
+            rubricCandScore += 0.5;
+            rubricBaseScore += 0.5;
+          }
+          rubricN++;
+        }
+      }
+      const totalN = detN + rubricN;
+      const baseHard = totalN > 0 ? (detBaseScore * detN + rubricBaseScore / Math.max(rubricN, 1) * rubricN) / totalN : 0;
+      const candHard = totalN > 0 ? (detCandScore * detN + rubricCandScore / Math.max(rubricN, 1) * rubricN) / totalN : 0;
+      const rubricDelta = rubricN > 0 ? signedDeltaTotal / rubricN : undefined;
+      if (!(candHard > baseHard && candHard - baseHard >= input.evalGate.margin)) {
+        return {
+          ok: false,
+          failedGate: "empirical",
+          reason: `Empirical gate failed: candidate behavior ${candHard.toFixed(2)} \u2264 baseline ${baseHard.toFixed(2)} (margin ${input.evalGate.margin.toFixed(2)}${rubricDelta !== undefined ? `, rubric_delta=${rubricDelta.toFixed(2)}, noise_floor=${maxNoiseFloor.toFixed(2)}` : ""}).`
+        };
+      }
+      const trainCases = cases.cases.filter((c3) => c3.split === "train");
+      empiricalScores = {
+        hard: candHard,
+        holdout_n: totalN,
+        train_n: trainCases.length,
+        ...rubricN > 0 ? { noise_floor: maxNoiseFloor, rubric_delta: rubricDelta ?? 0 } : {}
+      };
+    }
+  }
   if (input.ingestedAnchorHash !== undefined && input.baselineAnchorHash !== undefined) {
     if (input.ingestedAnchorHash !== input.baselineAnchorHash) {
       return {
@@ -95608,7 +96224,7 @@ async function runGate(input) {
       reason: `Skeptic gate failed: the Skeptic reported an anchor/invariant violation.${violations}`
     };
   }
-  return { ok: true };
+  return { ok: true, ...empiricalScores ? { empirical: empiricalScores } : {} };
 }
 function emitGenerationEnvelope(type, resolvedPath, content, baseline, trends) {
   const rubric2 = loadRubric(type);
@@ -95666,7 +96282,7 @@ function emitGenerationEnvelope(type, resolvedPath, content, baseline, trends) {
 async function ingestProposal(db2, type, name, resolvedPath, ingestPath, opts, baselineScore) {
   let raw;
   try {
-    raw = readFileSync11(ingestPath, "utf-8");
+    raw = readFileSync12(ingestPath, "utf-8");
   } catch {
     throw Object.assign(new Error(`Cannot read proposal file: ${ingestPath}`), { code: 2 });
   }
@@ -95708,10 +96324,22 @@ async function ingestProposal(db2, type, name, resolvedPath, ingestPath, opts, b
   if (opts?.acceptId) {
     const backupPath = await backupFile(resolvedPath);
     const appliedCount = await stepApply(parsed.changes, resolvedPath, proposalRecord.id, db2);
+    const evalGate = opts?.evalGate ? {
+      name,
+      candidateSkillText: await Bun.file(resolvedPath).text(),
+      baselineSkillText: readFileSync12(backupPath, "utf-8"),
+      margin: opts?.margin ?? 0.05,
+      target: opts?.target ?? "claude",
+      replayBackend: opts?.replayBackend,
+      judgeBackend: opts?.judgeBackend,
+      judgeReplays: opts?.judgeReplays,
+      judgeBudget: opts?.judgeBudget
+    } : undefined;
     const verdict = await stepVerify(type, name, resolvedPath, baselineScore, proposalRecord.id, opts, db2, {
       backupPath,
       ingestedAnchorHash: parsed.anchor_hash,
-      skeptic: parsed.skeptic
+      skeptic: parsed.skeptic,
+      evalGate
     });
     if (!verdict.rejected && verdict.backupPath) {
       await persistVersionSnapshot(verdict.backupPath, resolvedPath, proposalId);
@@ -95928,7 +96556,7 @@ async function stepVerify(type, name, filePath, baselineScore, proposalDbId, opt
   const delta = postScore - baselineScore;
   if (gate) {
     const margin = opts?.margin ?? 0.05;
-    const baselineAnchorHash = gate.ingestedAnchorHash !== undefined ? computeBaselineAnchorHash(readFileSync11(gate.backupPath, "utf-8")) : undefined;
+    const baselineAnchorHash = gate.ingestedAnchorHash !== undefined ? computeBaselineAnchorHash(readFileSync12(gate.backupPath, "utf-8")) : undefined;
     const gateResult = await runGate({
       type,
       resolvedPath: filePath,
@@ -95937,7 +96565,8 @@ async function stepVerify(type, name, filePath, baselineScore, proposalDbId, opt
       margin,
       ingestedAnchorHash: gate.ingestedAnchorHash,
       baselineAnchorHash,
-      skeptic: gate.skeptic
+      skeptic: gate.skeptic,
+      evalGate: gate.evalGate
     });
     if (!gateResult.ok) {
       await restoreFromBackup(gate.backupPath, filePath);
@@ -95945,6 +96574,31 @@ async function stepVerify(type, name, filePath, baselineScore, proposalDbId, opt
       echoError(`Gate rejected proposal \u2014 ${gateResult.reason} File restored; proposal stays draft.`);
       return { postScore: baselineScore, delta: 0, rejected: true, reason: gateResult.reason };
     }
+    if (gateResult.empirical) {
+      try {
+        await new EvaluationDao(db2).insertEvaluation({
+          content_type: type,
+          content_name: name,
+          target_agent: opts?.target ?? "claude",
+          operation: "evolve",
+          aggregate: gateResult.empirical.hard,
+          dimensions: {
+            empirical: {
+              score: gateResult.empirical.hard,
+              hard: gateResult.empirical.hard,
+              holdout_n: gateResult.empirical.holdout_n,
+              train_n: gateResult.empirical.train_n,
+              ...gateResult.empirical.noise_floor !== undefined ? { noise_floor: gateResult.empirical.noise_floor } : {},
+              ...gateResult.empirical.rubric_delta !== undefined ? { rubric_delta: gateResult.empirical.rubric_delta } : {},
+              note: `Behavior gate: holdout_n=${gateResult.empirical.holdout_n}, train_n=${gateResult.empirical.train_n}${gateResult.empirical.noise_floor !== undefined ? `, noise_floor=${gateResult.empirical.noise_floor.toFixed(2)}` : ""}${gateResult.empirical.rubric_delta !== undefined ? `, rubric_delta=${gateResult.empirical.rubric_delta.toFixed(2)}` : ""}`
+            }
+          }
+        });
+        echo(`Empirical behavior: holdout ${gateResult.empirical.hard.toFixed(2)} (holdout_n=${gateResult.empirical.holdout_n}, train_n=${gateResult.empirical.train_n}${gateResult.empirical.noise_floor !== undefined ? `, noise_floor=${gateResult.empirical.noise_floor.toFixed(2)}` : ""})`);
+      } catch {
+        echoError("Cannot persist empirical behavior score.");
+      }
+    }
   }
   try {
     const verifyEval = await new EvaluationDao(db2).getLatestEvaluation(type, name);
@@ -95974,7 +96628,7 @@ function formatAnalyze(name, analysis) {
   const { trends, baselineScore, baselineDate, evaluations: evaluations2 } = analysis;
   const grade = scoreToGrade(baselineScore);
   const verdict = baselineScore >= 0.7 ? "PASS" : "FAIL";
-  const gitAvailable = existsSync12(join220(process.cwd(), ".git"));
+  const gitAvailable = existsSync13(join220(process.cwd(), ".git"));
   const lines = [];
   lines.push("=== Evolution Analysis ===");
   lines.push(`Target: ${name}   Score: ${(baselineScore * 100).toFixed(0)}% (${grade})   Status: ${verdict}`);
@@ -96004,12 +96658,12 @@ function formatAnalyze(name, analysis) {
 async function persistVersionSnapshot(backupPath, resolvedPath, proposalId) {
   const versionPath = `${resolvedPath}.version-${proposalId}`;
   await Bun.write(versionPath, Bun.file(backupPath));
-  rmSync4(backupPath, { force: true });
+  rmSync5(backupPath, { force: true });
   return versionPath;
 }
 async function evolve(type, name, opts) {
   const resolvedPath = resolveContentPath(type, name);
-  if (!resolvedPath || !existsSync12(resolvedPath)) {
+  if (!resolvedPath || !existsSync13(resolvedPath)) {
     echoError(`File not found: ${name}`);
     return { baselineScore: 0, postScore: 0, delta: 0, changesApplied: 0, proposalPath: "" };
   }
@@ -96044,7 +96698,7 @@ async function evolve(type, name, opts) {
       const versionId = proposalId2 ?? `#${p.id}`;
       const appliedAt = p.applied_at ?? "unknown";
       const versionPath = `${resolvedPath}.version-${versionId}`;
-      const hasSnapshot = existsSync12(versionPath) ? "\u2713" : "\u2717";
+      const hasSnapshot = existsSync13(versionPath) ? "\u2713" : "\u2717";
       lines.push(`| ${versionId} | ${appliedAt} | ${hasSnapshot} |`);
     }
     echo(lines.join(`
@@ -96057,7 +96711,7 @@ async function evolve(type, name, opts) {
       return { baselineScore: 0, postScore: 0, delta: 0, changesApplied: 0, proposalPath: "" };
     }
     const versionPath = `${resolvedPath}.version-${opts.rollback}`;
-    if (!existsSync12(versionPath)) {
+    if (!existsSync13(versionPath)) {
       echoError(`Version snapshot not found: ${opts.rollback}`);
       return { baselineScore: 0, postScore: 0, delta: 0, changesApplied: 0, proposalPath: "" };
     }
@@ -96137,10 +96791,22 @@ async function evolve(type, name, opts) {
     }
     const backupPath2 = await backupFile(resolvedPath);
     const appliedCount = await stepApply(acceptedFromStore, resolvedPath, target.id, db2);
+    const evalGateCtx2 = opts?.evalGate ? {
+      name: contentName,
+      candidateSkillText: await Bun.file(resolvedPath).text(),
+      baselineSkillText: readFileSync12(backupPath2, "utf-8"),
+      margin: opts?.margin ?? 0.05,
+      target: opts?.target ?? "claude",
+      replayBackend: opts?.replayBackend,
+      judgeBackend: opts?.judgeBackend,
+      judgeReplays: opts?.judgeReplays,
+      judgeBudget: opts?.judgeBudget
+    } : undefined;
     const verdict = await stepVerify(type, contentName, resolvedPath, baselineScore, target.id, opts, db2, {
       backupPath: backupPath2,
       ingestedAnchorHash: storedAnchorHash,
-      skeptic: storedSkeptic
+      skeptic: storedSkeptic,
+      evalGate: evalGateCtx2
     });
     if (!verdict.rejected && verdict.backupPath) {
       await persistVersionSnapshot(verdict.backupPath, resolvedPath, opts.acceptId);
@@ -96163,7 +96829,18 @@ async function evolve(type, name, opts) {
   const { accepted: acceptedChanges } = await interactiveReview(changes, trends);
   const backupPath = await backupFile(resolvedPath);
   const changesApplied = await stepApply(acceptedChanges, resolvedPath, proposalDbId, db2);
-  const { postScore, delta } = await stepVerify(type, contentName, resolvedPath, baselineScore, proposalDbId, opts, db2);
+  const evalGateCtx = opts?.evalGate ? {
+    name: contentName,
+    candidateSkillText: await Bun.file(resolvedPath).text(),
+    baselineSkillText: readFileSync12(backupPath, "utf-8"),
+    margin: opts?.margin ?? 0.05,
+    target: opts?.target ?? "claude",
+    replayBackend: opts?.replayBackend,
+    judgeBackend: opts?.judgeBackend,
+    judgeReplays: opts?.judgeReplays,
+    judgeBudget: opts?.judgeBudget
+  } : undefined;
+  const { postScore, delta } = await stepVerify(type, contentName, resolvedPath, baselineScore, proposalDbId, opts, db2, evalGateCtx ? { backupPath, evalGate: evalGateCtx } : undefined);
   await persistVersionSnapshot(backupPath, resolvedPath, proposalId);
   return { baselineScore, postScore, delta, changesApplied, proposalPath };
 }
@@ -96580,7 +97257,7 @@ function addScaffoldOptions(cmd) {
   return cmd.option("-d, --description <text>", "Content description").option("-t, --target <agent>", "Target agent platform", "claude").option("-o, --output <dir>", "Output directory (default: cwd)").option("--template <tier>", "Template tier (e.g. minimal / standard / specialist)").option("--tools <list>", "Comma-separated tool names to pre-populate frontmatter").option("--force", "Overwrite existing file if present");
 }
 function addEvolveOptions(cmd) {
-  return cmd.option("-t, --target <agent>", "Target agent platform", "claude").option("--from <date>", "Analyze evaluations since date (ISO 8601)").option("--propose-only", "Generate proposal without applying").option("--accept <id>", "Accept a specific proposal by ID").option("--reject <id>", "Reject a specific proposal").option("--json", "Output machine-readable JSON (envelope-out with --propose-only)").option("--ingest <file>", "Agent-authored proposal JSON (ingest-in mode)").option("--margin <n>", "\u0394-margin gate threshold for accept (default 0.05)", Number.parseFloat, 0.05).option("--analyze", "Print analysis summary (trends, score, data sources) without writing a proposal").option("--history", "List applied proposal versions from the store").option("--rollback <id>", "Rollback to a prior version by proposal_id (requires --confirm)").option("--confirm", "Confirm a destructive operation (required for --rollback)");
+  return cmd.option("-t, --target <agent>", "Target agent platform", "claude").option("--from <date>", "Analyze evaluations since date (ISO 8601)").option("--propose-only", "Generate proposal without applying").option("--accept <id>", "Accept a specific proposal by ID").option("--reject <id>", "Reject a specific proposal").option("--json", "Output machine-readable JSON (envelope-out with --propose-only)").option("--ingest <file>", "Agent-authored proposal JSON (ingest-in mode)").option("--margin <n>", "\u0394-margin gate threshold for accept (default 0.05)", Number.parseFloat, 0.05).option("--eval-gate", "Enable empirical behavior gate (requires skills/<name>/eval/cases.yaml)").option("--analyze", "Print analysis summary (trends, score, data sources) without writing a proposal").option("--history", "List applied proposal versions from the store").option("--rollback <id>", "Rollback to a prior version by proposal_id (requires --confirm)").option("--confirm", "Confirm a destructive operation (required for --rollback)");
 }
 function addHookEvolveOptions(cmd) {
   return cmd.option("-t, --target <agent>", "Target agent platform", "claude").option("--from <date>", "Analyze evaluations since date (ISO 8601)").option("--analyze", "Print analysis summary (trends, score, data sources) \u2014 analyze-only, no apply").option("--json", "Output machine-readable JSON");
@@ -96689,7 +97366,8 @@ async function agentEvolve(opts) {
     analyze: opts.analyze,
     history: opts.history,
     rollback: opts.rollback,
-    confirm: opts.confirm
+    confirm: opts.confirm,
+    evalGate: opts.evalGate
   });
   return;
 }
@@ -96785,7 +97463,8 @@ async function commandEvolve(opts) {
     analyze: opts.analyze,
     history: opts.history,
     rollback: opts.rollback,
-    confirm: opts.confirm
+    confirm: opts.confirm,
+    evalGate: opts.evalGate
   });
   return;
 }
@@ -96826,7 +97505,7 @@ function registerCommand(program2) {
 // src/commands/hook.ts
 init_src();
 init_dist();
-import { homedir as homedir6 } from "os";
+import { homedir as homedir7 } from "os";
 import { join as join225 } from "path";
 
 // src/commands/install.ts
@@ -96834,19 +97513,19 @@ init_src();
 init_dist();
 import {
   copyFileSync as copyFileSync2,
-  existsSync as existsSync14,
+  existsSync as existsSync15,
   mkdirSync as mkdirSync9,
   readdirSync as readdirSync4,
-  readFileSync as readFileSync13,
-  rmSync as rmSync5,
-  statSync as statSync7,
+  readFileSync as readFileSync14,
+  rmSync as rmSync6,
+  statSync as statSync6,
   writeFileSync as writeFileSync8
 } from "fs";
-import { homedir as homedir5 } from "os";
+import { homedir as homedir6 } from "os";
 import { join as join223 } from "path";
 
 // src/hooks.ts
-import { copyFileSync, existsSync as existsSync13, mkdirSync as mkdirSync8, readFileSync as readFileSync12, writeFileSync as writeFileSync7 } from "fs";
+import { copyFileSync, existsSync as existsSync14, mkdirSync as mkdirSync8, readFileSync as readFileSync13, writeFileSync as writeFileSync7 } from "fs";
 import { join as join221 } from "path";
 var CANONICAL_TO_PI_EVENT = {
   sessionStart: "session_start",
@@ -96890,10 +97569,10 @@ function convertCanonicalToPiHooks(config3) {
 }
 function readCanonicalHooks(rulesyncDir) {
   const hooksPath = join221(rulesyncDir, "hooks.json");
-  if (!existsSync13(hooksPath))
+  if (!existsSync14(hooksPath))
     return null;
   try {
-    return JSON.parse(readFileSync12(hooksPath, "utf-8"));
+    return JSON.parse(readFileSync13(hooksPath, "utf-8"));
   } catch {
     return null;
   }
@@ -97051,7 +97730,7 @@ async function executeInstall(plugin, targets2, options2, dependencies = {}) {
       echo(`  Skills written: ${resultCounts.skillsCount}, Commands: ${resultCounts.commandsCount}, Subagents: ${resultCounts.subagentsCount}, Hooks: ${resultCounts.hooksCount}`);
     }
   }
-  const outputRoot = options2.outputRoot ?? (options2.global ? homedir5() : process.cwd());
+  const outputRoot = options2.outputRoot ?? (options2.global ? homedir6() : process.cwd());
   const hookEmitResults = [];
   for (const target of targets2) {
     if (target === "claude") {
@@ -97060,9 +97739,9 @@ async function executeInstall(plugin, targets2, options2, dependencies = {}) {
       if (!options2.dryRun) {
         const marketplaceName = resolution.marketplaceName ?? "superskill";
         const marketplaceRoot = resolution.marketplaceRoot ?? process.cwd();
-        const cacheDir = join223(homedir5(), ".claude", "plugins", "cache", marketplaceName);
-        if (existsSync14(cacheDir))
-          rmSync5(cacheDir, { recursive: true, force: true });
+        const cacheDir = join223(homedir6(), ".claude", "plugins", "cache", marketplaceName);
+        if (existsSync15(cacheDir))
+          rmSync6(cacheDir, { recursive: true, force: true });
         await runClaudeInstallImpl(marketplaceRoot, marketplaceName, plugin);
       }
     }
@@ -97090,7 +97769,7 @@ async function executeInstall(plugin, targets2, options2, dependencies = {}) {
       if (options2.verbose)
         echo(`  ${hookResult.message}`);
       const agentsDir = join223(pluginRoot, "agents");
-      if (existsSync14(agentsDir) && !options2.dryRun) {
+      if (existsSync15(agentsDir) && !options2.dryRun) {
         const piAgentsDir = join223(outputRoot, ".pi", "agent", "agents");
         mkdirSync9(piAgentsDir, { recursive: true });
         for (const entry of readdirSync4(agentsDir)) {
@@ -97098,8 +97777,8 @@ async function executeInstall(plugin, targets2, options2, dependencies = {}) {
             continue;
           const agentName = entry.replace(/\.md$/, "");
           const expectedName = `${plugin}-${agentName}`;
-          const source = readFileSync13(join223(agentsDir, entry), "utf-8");
-          const skillExists = (bare) => existsSync14(join223(pluginRoot, "skills", bare));
+          const source = readFileSync14(join223(agentsDir, entry), "utf-8");
+          const skillExists = (bare) => existsSync15(join223(pluginRoot, "skills", bare));
           const adapted = adaptSubagentToPi(source, expectedName, plugin, skillExists);
           writeFileSync8(join223(piAgentsDir, `${expectedName}.md`), adapted);
         }
@@ -97148,9 +97827,9 @@ function resolvePluginRoot(plugin, marketplacePath) {
     const manifestRoot = resolved.marketplaceRoot;
     const manifestPath = join223(manifestRoot, ".claude-plugin", "marketplace.json");
     let marketplaceName;
-    if (existsSync14(manifestPath)) {
+    if (existsSync15(manifestPath)) {
       try {
-        const raw = readFileSync13(manifestPath, "utf-8");
+        const raw = readFileSync14(manifestPath, "utf-8");
         const parsed = JSON.parse(raw);
         marketplaceName = parsed.name;
       } catch {}
@@ -97158,7 +97837,7 @@ function resolvePluginRoot(plugin, marketplacePath) {
     return { pluginRoot: resolved.pluginRoot, marketplaceRoot: manifestRoot, marketplaceName };
   }
   const fallback = join223("plugins", plugin);
-  if (existsSync14(fallback) && readdirSync4(fallback).some((d) => ["skills", "commands", "agents", "hooks", "hooks.json"].includes(d)))
+  if (existsSync15(fallback) && readdirSync4(fallback).some((d) => ["skills", "commands", "agents", "hooks", "hooks.json"].includes(d)))
     return { pluginRoot: fallback };
   const available = listResolvablePlugins(marketplacePath);
   const msg = available.length > 0 ? `Available: ${available.join(", ")}` : "No marketplace manifest found and no plugins/<name>/ directory.";
@@ -97167,7 +97846,7 @@ function resolvePluginRoot(plugin, marketplacePath) {
 function prepareTargetRulesyncInput(sourceRoot, target, pluginName) {
   const targetRoot = join223(sourceRoot, ".targets", target);
   const targetRulesyncRoot = join223(targetRoot, ".rulesync");
-  rmSync5(targetRoot, { recursive: true, force: true });
+  rmSync6(targetRoot, { recursive: true, force: true });
   copyDirectory(sourceRoot, targetRulesyncRoot, { skipDirectoryNames: new Set([".targets"]) });
   transformRulesyncMarkdown(targetRulesyncRoot, target, pluginName);
   return targetRoot;
@@ -97193,25 +97872,25 @@ function transformRulesyncMarkdown(root, target, pluginName) {
   transformMarkdownDirectory(join223(root, "skills"), target, pluginName);
 }
 function transformMarkdownDirectory(dir, target, pluginName) {
-  if (!existsSync14(dir))
+  if (!existsSync15(dir))
     return;
   for (const entry of readdirSync4(dir)) {
     const path11 = join223(dir, entry);
-    const stats = statSync7(path11);
+    const stats = statSync6(path11);
     if (stats.isDirectory()) {
       transformMarkdownDirectory(path11, target, pluginName);
       continue;
     }
     if (!entry.endsWith(".md"))
       continue;
-    const content = readFileSync13(path11, "utf-8");
+    const content = readFileSync14(path11, "utf-8");
     const slashTranslated = translateSlashCommands(content, target);
     const transformed = rewriteSkillReferences(slashTranslated, pluginName);
     writeFileSync8(path11, transformed);
   }
 }
 function copyDirectory(source, destination, options2 = {}) {
-  if (!existsSync14(source))
+  if (!existsSync15(source))
     return;
   mkdirSync9(destination, { recursive: true });
   for (const entry of readdirSync4(source)) {
@@ -97219,7 +97898,7 @@ function copyDirectory(source, destination, options2 = {}) {
       continue;
     const sourcePath = join223(source, entry);
     const destinationPath = join223(destination, entry);
-    if (statSync7(sourcePath).isDirectory()) {
+    if (statSync6(sourcePath).isDirectory()) {
       copyDirectory(sourcePath, destinationPath, options2);
     } else {
       copyFileSync2(sourcePath, destinationPath);
@@ -97258,7 +97937,7 @@ async function emitHook(name, opts) {
   const target = resolveTarget(opts);
   const global3 = opts.global !== false;
   const dryRun = opts.dryRun === true;
-  const outputRoot = global3 ? homedir6() : process.cwd();
+  const outputRoot = global3 ? homedir7() : process.cwd();
   const pluginRoot = resolvePluginRoot(name).pluginRoot;
   const outputDir = ".rulesync";
   mapPluginToRulesync(pluginRoot, name, outputDir);
@@ -97386,7 +98065,8 @@ async function magentEvolve(opts) {
     analyze: opts.analyze,
     history: opts.history,
     rollback: opts.rollback,
-    confirm: opts.confirm
+    confirm: opts.confirm,
+    evalGate: opts.evalGate
   });
   return;
 }
@@ -97430,9 +98110,9 @@ init_dist();
 // src/operations/migrate.ts
 init_src();
 init_dist();
-import { readFileSync as readFileSync14 } from "fs";
+import { readFileSync as readFileSync15 } from "fs";
 function readProposalId(proposalPath) {
-  const raw = readFileSync14(proposalPath, "utf-8");
+  const raw = readFileSync15(proposalPath, "utf-8");
   const parsed = JSON.parse(raw);
   if (parsed !== null && typeof parsed === "object" && "proposal_id" in parsed) {
     const id = parsed.proposal_id;
@@ -97537,7 +98217,8 @@ async function skillEvolve(opts) {
     analyze: opts.analyze,
     history: opts.history,
     rollback: opts.rollback,
-    confirm: opts.confirm
+    confirm: opts.confirm,
+    evalGate: opts.evalGate
   });
   return;
 }
@@ -97601,7 +98282,7 @@ function getPackageVersion() {
   if (packageVersion)
     return packageVersion;
   const packageJsonPath = join226(import.meta.dir, "..", "package.json");
-  const packageJson = JSON.parse(readFileSync15(packageJsonPath, "utf-8"));
+  const packageJson = JSON.parse(readFileSync16(packageJsonPath, "utf-8"));
   if (typeof packageJson.version !== "string" || packageJson.version.length === 0) {
     throw new Error(`Invalid package version in ${packageJsonPath}`);
   }