kairn-cli 1.13.0 → 2.0.0

package/dist/cli.js

@@ -1,6 +1,6 @@
 // src/cli.ts
-import { Command as Command11 } from "commander";
-import chalk14 from "chalk";
+import { Command as Command12 } from "commander";
+import chalk15 from "chalk";
 
 // src/commands/init.ts
 import { Command } from "commander";
@@ -221,7 +221,7 @@ var ui = {
   // Key-value pairs
   kv: (key, value) => `  ${chalk.cyan(key.padEnd(14))} ${value}`,
   // File list
-  file: (path15) => chalk.dim(`    ${path15}`),
+  file: (path20) => chalk.dim(`    ${path20}`),
   // Tool display
   tool: (name, reason) => `    ${warmStone("\u25CF")} ${chalk.bold(name)}
       ${chalk.dim(reason)}`,
@@ -562,8 +562,6 @@ import chalk5 from "chalk";
 import fs4 from "fs/promises";
 import path4 from "path";
 import crypto from "crypto";
-import Anthropic2 from "@anthropic-ai/sdk";
-import OpenAI2 from "openai";
 
 // src/compiler/prompt.ts
 var SKELETON_PROMPT = `You are the Kairn skeleton compiler. Your job is to select tools and outline the project structure from a user's natural language description.
@@ -716,6 +714,13 @@ At the start of every session, before doing ANY work:
 4. Summarize what you see in 2-3 lines, then proceed
 
 This saves 2-5 exploratory turns. Never ask "what files are here?" \u2014 look first.
+
+## Completion Standards
+
+Never mark a task "done" without running the Completion Verification checklist.
+Tests passing is necessary but not sufficient \u2014 also verify requirements coverage,
+state cleanliness, and review changes from the perspective of a test engineer,
+code reviewer, and the requesting user.
 \`\`\`
 
 Do not add generic filler. Every line must be specific to the user's workflow.
@@ -738,6 +743,7 @@ Do not add generic filler. Every line must be specific to the user's workflow.
 14. A "Git Workflow" section in CLAUDE.md (3 rules: small commits, conventional format, <200 lines PR)
 15. "Engineering Standards", "Tool Usage Policy", and "Code Philosophy" sections in CLAUDE.md
 16. A "First Turn Protocol" section in CLAUDE.md (orient before working: pwd, ls, git status, check relevant runtimes, read task files)
+17. A "Completion Standards" section in CLAUDE.md (never mark done without verifying: requirements met, tests passing, no debug artifacts, reviewed from 3 perspectives)
 
 ## Shell-Integrated Commands
 
@@ -855,7 +861,7 @@ Merge this into the settings hooks alongside the PreToolUse and PostToolUse hook
 - \`/project:status\` command (live git status, recent commits, SPRINT.md overview using ! prefix)
 - \`/project:fix\` command (takes $ARGUMENTS as issue number, plans fix, implements, tests, commits)
 - \`/project:sprint\` command (define acceptance criteria before coding, writes to docs/SPRINT.md)
-- \`/project:develop\` command (full development pipeline \u2014 orchestrates @architect \u2192 @planner \u2192 @implementer \u2192 @verifier \u2192 @fixer \u2192 @grill \u2192 @doc-updater through spec, plan, TDD implement, review, and doc update phases)
+- \`/project:develop\` command (full development pipeline \u2014 orchestrates @architect \u2192 @planner \u2192 @implementer \u2192 @verifier \u2192 @fixer \u2192 @grill \u2192 @doc-updater through spec, plan, TDD implement, review, and doc update phases). MUST include a Phase 7 "Completion Gate" that runs a Completion Verification checklist before marking the feature done: re-read original requirements, confirm each is met with evidence, run test suite + lint/typecheck, review git diff for unexpected changes or debug artifacts, answer 3 perspective questions (test engineer, code reviewer, requesting user). If ANY check fails, loop back to fix before completing.
 - A TDD skill using the 3-phase isolation pattern (RED \u2192 GREEN \u2192 REFACTOR):
   - RED: Write failing test only. Verify it FAILS.
   - GREEN: Write MINIMUM code to pass. Nothing extra.
@@ -1023,6 +1029,13 @@ At the start of every session, before doing ANY work:
 4. Summarize what you see in 2-3 lines, then proceed
 
 This saves 2-5 exploratory turns. Never ask "what files are here?" \u2014 look first.
+
+## Completion Standards
+
+Never mark a task "done" without running the Completion Verification checklist.
+Tests passing is necessary but not sufficient \u2014 also verify requirements coverage,
+state cleanliness, and review changes from the perspective of a test engineer,
+code reviewer, and the requesting user.
 \`\`\`
 
 Do not add generic filler. Every line must be specific to the user's workflow.
@@ -1045,6 +1058,7 @@ Do not add generic filler. Every line must be specific to the user's workflow.
 14. A "Git Workflow" section in CLAUDE.md (3 rules: small commits, conventional format, <200 lines PR)
 15. "Engineering Standards", "Tool Usage Policy", and "Code Philosophy" sections in CLAUDE.md
 16. A "First Turn Protocol" section in CLAUDE.md (orient before working: pwd, ls, git status, check relevant runtimes, read task files)
+17. A "Completion Standards" section in CLAUDE.md (never mark done without verifying: requirements met, tests passing, no debug artifacts, reviewed from 3 perspectives)
 
 ## Tool Selection Rules
 
@@ -1188,76 +1202,9 @@ async function loadRegistry() {
   return Array.from(merged.values());
 }
 
-// src/compiler/compile.ts
-function buildSkeletonMessage(intent, registry) {
-  const registrySummary = registry.map(
-    (t) => `- ${t.id} (${t.type}, tier ${t.tier}, auth: ${t.auth}): ${t.description} [best_for: ${t.best_for.join(", ")}]`
-  ).join("\n");
-  return `## User Intent
-
-${intent}
-
-## Available Tool Registry
-
-${registrySummary}
-
-Generate the skeleton JSON now.`;
-}
-function buildHarnessMessage(intent, skeleton, concise) {
-  const skeletonJson = JSON.stringify(skeleton, null, 2);
-  const conciseNote = concise ? "\n\nIMPORTANT: Be concise. Maximum 80 lines for claude_md. Maximum 5 commands. Keep all content brief." : "";
-  return `## User Intent
-
-${intent}
-
-## Project Skeleton
-
-${skeletonJson}
-
-Generate the harness content JSON now.${conciseNote}`;
-}
-function parseSkeletonResponse(text) {
-  let cleaned = text.trim();
-  if (cleaned.startsWith("```")) {
-    cleaned = cleaned.replace(/^```(?:json)?\n?/, "").replace(/\n?```$/, "");
-  }
-  const jsonMatch = cleaned.match(/\{[\s\S]*\}/);
-  if (!jsonMatch) {
-    throw new Error("Pass 1 (skeleton) did not return valid JSON.");
-  }
-  try {
-    const parsed = JSON.parse(jsonMatch[0]);
-    if (!parsed.name || !parsed.tools || !Array.isArray(parsed.tools)) {
-      throw new Error("Skeleton missing required fields: name, tools");
-    }
-    return parsed;
-  } catch (err) {
-    throw new Error(
-      `Failed to parse skeleton JSON: ${err instanceof Error ? err.message : String(err)}`
-    );
-  }
-}
-function parseHarnessResponse(text) {
-  let cleaned = text.trim();
-  if (cleaned.startsWith("```")) {
-    cleaned = cleaned.replace(/^```(?:json)?\n?/, "").replace(/\n?```$/, "");
-  }
-  const jsonMatch = cleaned.match(/\{[\s\S]*\}/);
-  if (!jsonMatch) {
-    throw new Error("Pass 2 (harness) did not return valid JSON.");
-  }
-  try {
-    const parsed = JSON.parse(jsonMatch[0]);
-    if (!parsed.claude_md || !parsed.commands) {
-      throw new Error("Harness missing required fields: claude_md, commands");
-    }
-    return parsed;
-  } catch (err) {
-    throw new Error(
-      `Failed to parse harness JSON: ${err instanceof Error ? err.message : String(err)}`
-    );
-  }
-}
+// src/llm.ts
+import Anthropic2 from "@anthropic-ai/sdk";
+import OpenAI2 from "openai";
 function classifyError(err, provider) {
   const msg = err instanceof Error ? err.message : String(err);
   const status = err?.status;
@@ -1289,8 +1236,8 @@ function classifyError(err, provider) {
   return `${provider} API error: ${msg}`;
 }
 async function callLLM(config, userMessage, options) {
-  const maxTokens = options?.maxTokens ?? 8192;
-  const systemPrompt = options?.systemPrompt ?? SYSTEM_PROMPT;
+  const maxTokens = options.maxTokens ?? 8192;
+  const systemPrompt = options.systemPrompt;
   const providerName = getProviderName(config.provider);
   if (config.provider === "anthropic") {
     const client2 = new Anthropic2({ apiKey: config.api_key });
@@ -1332,6 +1279,77 @@ async function callLLM(config, userMessage, options) {
     throw new Error(classifyError(err, providerName));
   }
 }
+
+// src/compiler/compile.ts
+function buildSkeletonMessage(intent, registry) {
+  const registrySummary = registry.map(
+    (t) => `- ${t.id} (${t.type}, tier ${t.tier}, auth: ${t.auth}): ${t.description} [best_for: ${t.best_for.join(", ")}]`
+  ).join("\n");
+  return `## User Intent
+
+${intent}
+
+## Available Tool Registry
+
+${registrySummary}
+
+Generate the skeleton JSON now.`;
+}
+function buildHarnessMessage(intent, skeleton, concise) {
+  const skeletonJson = JSON.stringify(skeleton, null, 2);
+  const conciseNote = concise ? "\n\nIMPORTANT: Be concise. Maximum 80 lines for claude_md. Maximum 5 commands. Keep all content brief." : "";
+  return `## User Intent
+
+${intent}
+
+## Project Skeleton
+
+${skeletonJson}
+
+Generate the harness content JSON now.${conciseNote}`;
+}
+function parseSkeletonResponse(text) {
+  let cleaned = text.trim();
+  if (cleaned.startsWith("```")) {
+    cleaned = cleaned.replace(/^```(?:json)?\n?/, "").replace(/\n?```$/, "");
+  }
+  const jsonMatch = cleaned.match(/\{[\s\S]*\}/);
+  if (!jsonMatch) {
+    throw new Error("Pass 1 (skeleton) did not return valid JSON.");
+  }
+  try {
+    const parsed = JSON.parse(jsonMatch[0]);
+    if (!parsed.name || !parsed.tools || !Array.isArray(parsed.tools)) {
+      throw new Error("Skeleton missing required fields: name, tools");
+    }
+    return parsed;
+  } catch (err) {
+    throw new Error(
+      `Failed to parse skeleton JSON: ${err instanceof Error ? err.message : String(err)}`
+    );
+  }
+}
+function parseHarnessResponse(text) {
+  let cleaned = text.trim();
+  if (cleaned.startsWith("```")) {
+    cleaned = cleaned.replace(/^```(?:json)?\n?/, "").replace(/\n?```$/, "");
+  }
+  const jsonMatch = cleaned.match(/\{[\s\S]*\}/);
+  if (!jsonMatch) {
+    throw new Error("Pass 2 (harness) did not return valid JSON.");
+  }
+  try {
+    const parsed = JSON.parse(jsonMatch[0]);
+    if (!parsed.claude_md || !parsed.commands) {
+      throw new Error("Harness missing required fields: claude_md, commands");
+    }
+    return parsed;
+  } catch (err) {
+    throw new Error(
+      `Failed to parse harness JSON: ${err instanceof Error ? err.message : String(err)}`
+    );
+  }
+}
 function buildSettings(skeleton, registry) {
   const selectedTools = skeleton.tools.map((t) => registry.find((r) => r.id === t.tool_id)).filter(Boolean);
   const allow = ["Read", "Write", "Edit", "Bash(npm run *)", "Bash(npx *)"];
@@ -1501,7 +1519,9 @@ async function generateClarifications(intent, onProgress) {
   onProgress?.("Analyzing your request...");
   const clarificationConfig = { ...config };
   clarificationConfig.model = getCheapModel(config.provider, config.model);
-  const response = await callLLM(clarificationConfig, CLARIFICATION_PROMPT + "\n\nUser description: " + intent);
+  const response = await callLLM(clarificationConfig, CLARIFICATION_PROMPT + "\n\nUser description: " + intent, {
+    systemPrompt: SYSTEM_PROMPT
+  });
   try {
     let cleaned = response.trim();
     if (cleaned.startsWith("```")) {
@@ -1682,7 +1702,29 @@ If confidence is LOW or MEDIUM, fix issues and re-verify.
 Run /project:grill for adversarial review.
 Fix any BLOCKERs.
 
-## Phase 6: SHIP
+## Phase 6: COMPLETION GATE
+
+Before shipping, run the Completion Verification checklist:
+
+### Requirements Check
+- [ ] Re-read the ORIGINAL task description (not your interpretation)
+- [ ] Each explicit requirement is met with evidence (test output, diff)
+- [ ] Each implicit requirement (error handling, types, tests) is addressed
+
+### State Check
+- [ ] Test suite passes
+- [ ] Lint/typecheck passes
+- [ ] \`git diff --stat\` \u2014 every changed file is intentional
+- [ ] No debug artifacts (console.log, TODO, commented-out code, temp files)
+
+### Perspective Check (1 sentence each)
+- **Test engineer:** Most likely production failure mode?
+- **Code reviewer:** What would I flag in review?
+- **Requesting user:** Does this solve my actual problem?
+
+ALL pass \u2192 proceed to ship. ANY fail \u2192 fix first, then re-verify.
+
+## Phase 7: SHIP
 Run /project:commit.
 Report what was built and what's next from docs/SPRINT.md.
 
@@ -1762,11 +1804,26 @@ Run verification:
 - Run functional tests
 - If NEEDS FIXES: fix and re-verify
 
-## Phase 5: PR
+## Phase 5: COMPLETION GATE
+
+Before creating a PR, run the Completion Verification checklist:
+- [ ] Re-read the ORIGINAL spec from docs/SPRINT.md
+- [ ] Each requirement is met with evidence (test output, diff)
+- [ ] Test suite + lint/typecheck pass
+- [ ] \`git diff --stat\` \u2014 every changed file is intentional, no debug artifacts
+- **Test engineer:** Most likely production failure mode?
+- **Code reviewer:** What would I flag in review?
+- **Requesting user:** Does this solve my actual problem?
+
+ALL pass \u2192 proceed to PR. ANY fail \u2192 fix first, then re-verify.
+
+Include the checklist results in the PR description.
+
+## Phase 6: PR
 Create a pull request:
-  gh pr create --title "feat: {name}" --body "{spec + QA report}"
+  gh pr create --title "feat: {name}" --body "{spec + QA report + verification checklist}"
 
-## Phase 6: NEXT
+## Phase 7: NEXT
 Report:
   "PR #{N} ready for review: {link}
    Next priority from SPRINT.md: {next task}
@@ -1790,13 +1847,19 @@ Repeat until max features reached or stopped:
 2. Create worktree + branch
 3. Implement the feature
 4. Run verification (build, test, lint)
-5. Open PR via gh
-6. Report status
-7. Move to next feature
+5. Run Completion Verification checklist:
+   - Requirements met with evidence
+   - Tests + lint/typecheck pass
+   - No debug artifacts or unexpected file changes
+   - 3-perspective check (test engineer, reviewer, user)
+6. Open PR via gh (include verification results in PR body)
+7. Report status
+8. Move to next feature
 
 ## Stop Conditions
 - Max 5 features per autopilot session
 - Any BLOCKER from verification
+- Completion Verification checklist fails after 2 fix attempts
 - Build failure that can't be resolved in 3 attempts
 - User presses Escape`;
 var AUTOPILOT_WARNING = `
@@ -3627,8 +3690,882 @@ var keysCommand = new Command10("keys").description("Add or update API keys for
   console.log("");
 });
 
+// src/commands/evolve.ts
+import { Command as Command11 } from "commander";
+import chalk14 from "chalk";
+import ora2 from "ora";
+import fs19 from "fs/promises";
+import path19 from "path";
+import { parse as yamlParse } from "yaml";
+import { confirm as confirm3, select as select4 } from "@inquirer/prompts";
+
+// src/evolve/init.ts
+import fs15 from "fs/promises";
+import path15 from "path";
+import { stringify as yamlStringify } from "yaml";
+
+// src/evolve/templates.ts
+var EVAL_TEMPLATES = {
+  "add-feature": {
+    id: "add-feature",
+    name: "Add Feature",
+    description: "Can the agent add a new capability?",
+    bestFor: ["feature-development", "api-building", "full-stack"]
+  },
+  "fix-bug": {
+    id: "fix-bug",
+    name: "Fix Bug",
+    description: "Can the agent diagnose and fix a problem?",
+    bestFor: ["maintenance", "debugging", "qa"]
+  },
+  "refactor": {
+    id: "refactor",
+    name: "Refactor",
+    description: "Can the agent restructure code?",
+    bestFor: ["maintenance", "architecture", "backend"]
+  },
+  "test-writing": {
+    id: "test-writing",
+    name: "Test Writing",
+    description: "Can the agent write tests?",
+    bestFor: ["tdd", "qa", "backend"]
+  },
+  "config-change": {
+    id: "config-change",
+    name: "Config Change",
+    description: "Can the agent update configuration?",
+    bestFor: ["devops", "infrastructure", "backend"]
+  },
+  "documentation": {
+    id: "documentation",
+    name: "Documentation",
+    description: "Can the agent write and update docs?",
+    bestFor: ["content", "api-building", "full-stack"]
+  }
+};
+function selectTemplatesForWorkflow(workflowType) {
+  const mapping = {
+    "feature-development": ["add-feature", "test-writing", "documentation"],
+    "api-building": ["add-feature", "fix-bug", "test-writing"],
+    "full-stack": ["add-feature", "fix-bug", "test-writing"],
+    "maintenance": ["fix-bug", "refactor", "test-writing"],
+    "debugging": ["fix-bug", "test-writing"],
+    "qa": ["fix-bug", "test-writing", "add-feature"],
+    "architecture": ["refactor", "test-writing", "config-change"],
+    "backend": ["fix-bug", "refactor", "config-change", "test-writing"],
+    "devops": ["config-change", "fix-bug"],
+    "infrastructure": ["config-change", "refactor"],
+    "tdd": ["test-writing", "add-feature", "fix-bug"],
+    "content": ["documentation", "add-feature"],
+    "research": ["documentation", "add-feature"]
+  };
+  return mapping[workflowType] || ["add-feature", "fix-bug", "test-writing"];
+}
+var TASK_GENERATION_PROMPT = `You are an eval task generator for Claude Code agent environments. Given a project's CLAUDE.md, project structure, and selected eval templates, generate concrete, project-specific tasks.
+
+Each task must be realistic and testable against the actual project. Avoid generic placeholders.
+
+Return a JSON object with a "tasks" array. Each task has:
+- id: kebab-case identifier (e.g., "add-health-endpoint")
+- template: which eval template this instantiates
+- description: concrete task description the agent will receive
+- setup: shell commands to prepare the workspace (e.g., "npm install")
+- expected_outcome: multi-line string describing what success looks like
+- scoring: "pass-fail", "llm-judge", or "rubric"
+- timeout: seconds (300 for features/bugs, 600 for refactors, 180 for config/docs/tests)
+
+Return ONLY valid JSON, no markdown fences.`;
+function parseJsonResponse(raw) {
+  let cleaned = raw.trim();
+  if (cleaned.startsWith("```")) {
+    cleaned = cleaned.replace(/^```(?:json)?\n?/, "").replace(/\n?```$/, "");
+  }
+  const jsonMatch = cleaned.match(/\{[\s\S]*\}/) ?? cleaned.match(/\[[\s\S]*\]/);
+  if (!jsonMatch) {
+    throw new Error(
+      "LLM response did not contain valid JSON. Try again or use a different model."
+    );
+  }
+  try {
+    return JSON.parse(jsonMatch[0]);
+  } catch (err) {
+    throw new Error(
+      `Failed to parse LLM response as JSON: ${err instanceof Error ? err.message : String(err)}`
+    );
+  }
+}
+var REQUIRED_TASK_FIELDS = [
+  "id",
+  "template",
+  "description",
+  "setup",
+  "expected_outcome",
+  "scoring",
+  "timeout"
+];
+function validateTask(obj, index) {
+  if (typeof obj !== "object" || obj === null) {
+    throw new Error(`Task at index ${index} is not an object`);
+  }
+  const record = obj;
+  for (const field of REQUIRED_TASK_FIELDS) {
+    if (!(field in record) || record[field] === void 0 || record[field] === null) {
+      throw new Error(`Task at index ${index} is missing required field: ${field}`);
+    }
+  }
+  return record;
+}
+function buildTaskGenerationMessage(claudeMd, projectProfile, templates) {
+  const profileLines = [
+    `Language: ${projectProfile.language ?? "unknown"}`,
+    `Framework: ${projectProfile.framework ?? "none"}`,
+    `Scripts: ${Object.entries(projectProfile.scripts).map(([k, v]) => `${k}=${v}`).join(", ") || "none"}`,
+    `Key files: ${projectProfile.keyFiles.join(", ") || "none"}`
+  ];
+  const templateDescriptions = templates.map((t) => {
+    const meta = EVAL_TEMPLATES[t];
+    return `- ${t}: ${meta.description}`;
+  }).join("\n");
+  return [
+    "## CLAUDE.md",
+    "",
+    claudeMd,
+    "",
+    "## Project Profile",
+    "",
+    ...profileLines,
+    "",
+    "## Selected Eval Templates",
+    "",
+    templateDescriptions,
+    "",
+    "Generate concrete, project-specific tasks for each template above."
+  ].join("\n");
+}
+async function generateTasksFromTemplates(claudeMd, projectProfile, templates, config) {
+  const userMessage = buildTaskGenerationMessage(claudeMd, projectProfile, templates);
+  const rawResponse = await callLLM(config, userMessage, {
+    systemPrompt: TASK_GENERATION_PROMPT,
+    maxTokens: 4096
+  });
+  const parsed = parseJsonResponse(rawResponse);
+  if (typeof parsed !== "object" || parsed === null) {
+    throw new Error("LLM response is not a JSON object");
+  }
+  const tasksObj = parsed;
+  if (!Array.isArray(tasksObj.tasks)) {
+    throw new Error("LLM response does not contain a 'tasks' array");
+  }
+  const tasks = [];
+  for (let i = 0; i < tasksObj.tasks.length; i++) {
+    tasks.push(validateTask(tasksObj.tasks[i], i));
+  }
+  return tasks;
+}
+
+// src/evolve/init.ts
+async function createEvolveWorkspace(projectRoot, config) {
+  const workspace = path15.join(projectRoot, ".kairn-evolve");
+  await fs15.mkdir(path15.join(workspace, "baseline"), { recursive: true });
+  await fs15.mkdir(path15.join(workspace, "traces"), { recursive: true });
+  await fs15.mkdir(path15.join(workspace, "iterations"), { recursive: true });
+  const configObj = {
+    model: config.model,
+    proposer_model: config.proposerModel,
+    scorer: config.scorer,
+    max_iterations: config.maxIterations,
+    parallel_tasks: config.parallelTasks
+  };
+  await fs15.writeFile(
+    path15.join(workspace, "config.yaml"),
+    yamlStringify(configObj),
+    "utf-8"
+  );
+  return workspace;
+}
+async function writeTasksFile(workspacePath, tasks) {
+  const doc = {
+    tasks: tasks.map((t) => ({
+      id: t.id,
+      template: t.template,
+      description: t.description,
+      setup: t.setup,
+      expected_outcome: t.expected_outcome,
+      scoring: t.scoring,
+      ...t.rubric ? { rubric: t.rubric } : {},
+      timeout: t.timeout
+    }))
+  };
+  const header = "# .kairn-evolve/tasks.yaml\n# Auto-generated by kairn evolve init \u2014 edit freely\n";
+  await fs15.writeFile(
+    path15.join(workspacePath, "tasks.yaml"),
+    header + yamlStringify(doc),
+    "utf-8"
+  );
+}
+async function buildProjectProfile(projectRoot) {
+  const profile = {
+    language: null,
+    framework: null,
+    scripts: {},
+    keyFiles: []
+  };
+  try {
+    const pkgStr = await fs15.readFile(
+      path15.join(projectRoot, "package.json"),
+      "utf-8"
+    );
+    const pkg = JSON.parse(pkgStr);
+    profile.language = "typescript";
+    if (pkg.scripts && typeof pkg.scripts === "object") {
+      profile.scripts = pkg.scripts;
+    }
+    const deps = {
+      ...pkg.dependencies ?? {},
+      ...pkg.devDependencies ?? {}
+    };
+    if (deps.next) {
+      profile.framework = "Next.js";
+    } else if (deps.express) {
+      profile.framework = "Express";
+    } else if (deps.react) {
+      profile.framework = "React";
+    } else if (deps.vue) {
+      profile.framework = "Vue";
+    } else if (deps.commander) {
+      profile.framework = "CLI (Commander.js)";
+    }
+  } catch {
+  }
+  if (!profile.language) {
+    try {
+      await fs15.access(path15.join(projectRoot, "pyproject.toml"));
+      profile.language = "python";
+    } catch {
+      try {
+        await fs15.access(path15.join(projectRoot, "requirements.txt"));
+        profile.language = "python";
+      } catch {
+      }
+    }
+  }
+  try {
+    const entries = await fs15.readdir(projectRoot);
+    const keyPatterns = [
+      "README.md",
+      "package.json",
+      "tsconfig.json",
+      "pyproject.toml",
+      "Cargo.toml",
+      "go.mod",
+      "Makefile",
+      "Dockerfile"
+    ];
+    profile.keyFiles = entries.filter((e) => keyPatterns.includes(e));
+  } catch {
+  }
+  return profile;
+}
+async function autoGenerateTasks(projectRoot, workflowType) {
+  const config = await loadConfig();
+  if (!config) {
+    throw new Error("No config found. Run `kairn init` first.");
+  }
+  let claudeMd = "";
+  try {
+    claudeMd = await fs15.readFile(
+      path15.join(projectRoot, ".claude", "CLAUDE.md"),
+      "utf-8"
+    );
+  } catch {
+  }
+  const profile = await buildProjectProfile(projectRoot);
+  const templates = selectTemplatesForWorkflow(workflowType);
+  return generateTasksFromTemplates(claudeMd, profile, templates, config);
+}
+
+// src/evolve/baseline.ts
+import fs16 from "fs/promises";
+import path16 from "path";
+async function snapshotBaseline(projectRoot, workspacePath) {
+  const claudeDir = path16.join(projectRoot, ".claude");
+  const baselineDir = path16.join(workspacePath, "baseline");
+  const iter0Dir = path16.join(workspacePath, "iterations", "0", "harness");
+  try {
+    await fs16.access(claudeDir);
+  } catch {
+    throw new Error(`.claude/ directory not found in ${projectRoot}`);
+  }
+  await copyDir(claudeDir, baselineDir);
+  await copyDir(claudeDir, iter0Dir);
+}
+async function copyDir(src, dest) {
+  await fs16.mkdir(dest, { recursive: true });
+  const entries = await fs16.readdir(src, { withFileTypes: true });
+  for (const entry of entries) {
+    const srcPath = path16.join(src, entry.name);
+    const destPath = path16.join(dest, entry.name);
+    if (entry.isDirectory()) {
+      await copyDir(srcPath, destPath);
+    } else {
+      await fs16.copyFile(srcPath, destPath);
+    }
+  }
+}
+
+// src/evolve/runner.ts
+import { exec, spawn } from "child_process";
+import { promisify } from "util";
+import fs18 from "fs/promises";
+import os3 from "os";
+import path18 from "path";
+
+// src/evolve/trace.ts
+import fs17 from "fs/promises";
+import path17 from "path";
+async function writeTrace(traceDir, trace) {
+  await fs17.mkdir(traceDir, { recursive: true });
+  await fs17.writeFile(path17.join(traceDir, "stdout.log"), trace.stdout, "utf-8");
+  await fs17.writeFile(path17.join(traceDir, "stderr.log"), trace.stderr, "utf-8");
+  const toolCallsLines = trace.toolCalls.map((tc) => JSON.stringify(tc)).join("\n");
+  await fs17.writeFile(path17.join(traceDir, "tool_calls.jsonl"), toolCallsLines, "utf-8");
+  await fs17.writeFile(
+    path17.join(traceDir, "files_changed.json"),
+    JSON.stringify(trace.filesChanged, null, 2),
+    "utf-8"
+  );
+  await fs17.writeFile(
+    path17.join(traceDir, "timing.json"),
+    JSON.stringify(trace.timing, null, 2),
+    "utf-8"
+  );
+  await fs17.writeFile(
+    path17.join(traceDir, "score.json"),
+    JSON.stringify(trace.score, null, 2),
+    "utf-8"
+  );
+}
+async function writeScore(traceDir, score) {
+  await fs17.writeFile(
+    path17.join(traceDir, "score.json"),
+    JSON.stringify(score, null, 2),
+    "utf-8"
+  );
+}
+
+// src/evolve/runner.ts
+var execAsync = promisify(exec);
+async function runTask(task, harnessPath, traceDir, iteration) {
+  await fs18.mkdir(traceDir, { recursive: true });
+  const startedAt = (/* @__PURE__ */ new Date()).toISOString();
+  const startMs = Date.now();
+  const tmpDir = await fs18.mkdtemp(path18.join(os3.tmpdir(), "kairn-evolve-"));
+  try {
+    await copyDir(harnessPath, path18.join(tmpDir, ".claude"));
+    let setupStderr = "";
+    if (task.setup.trim()) {
+      try {
+        await execAsync(task.setup, { cwd: tmpDir, timeout: 6e4 });
+      } catch (err) {
+        setupStderr = err instanceof Error ? err.message : String(err);
+      }
+    }
+    const filesBefore = await snapshotFileList(tmpDir);
+    const spawnResult = await spawnClaude(task.description, tmpDir, task.timeout);
+    const filesAfter = await snapshotFileList(tmpDir);
+    const filesChanged = diffFileLists(filesBefore, filesAfter);
+    const toolCalls = parseToolCalls(spawnResult.stdout);
+    const completedAt = (/* @__PURE__ */ new Date()).toISOString();
+    const durationMs = Date.now() - startMs;
+    const combinedStderr = setupStderr ? `[setup] ${setupStderr}
+${spawnResult.stderr}` : spawnResult.stderr;
+    const trace = {
+      taskId: task.id,
+      iteration,
+      stdout: spawnResult.stdout,
+      stderr: combinedStderr,
+      toolCalls,
+      filesChanged,
+      score: { pass: false, details: "Pending scoring" },
+      timing: { startedAt, completedAt, durationMs }
+    };
+    await writeTrace(traceDir, trace);
+    return {
+      taskId: task.id,
+      score: trace.score,
+      traceDir
+    };
+  } finally {
+    await fs18.rm(tmpDir, { recursive: true, force: true }).catch(() => {
+    });
+  }
+}
+async function spawnClaude(instruction, cwd, timeoutSec) {
+  return new Promise((resolve) => {
+    const args = ["--print", "--output-format", "text", "--max-turns", "50"];
+    const child = spawn("claude", args, {
+      cwd,
+      stdio: ["pipe", "pipe", "pipe"],
+      timeout: timeoutSec * 1e3,
+      env: { ...process.env }
+    });
+    let stdout = "";
+    let stderr = "";
+    child.stdout.on("data", (data) => {
+      stdout += data.toString();
+    });
+    child.stderr.on("data", (data) => {
+      stderr += data.toString();
+    });
+    child.stdin.write(instruction);
+    child.stdin.end();
+    child.on("close", (code) => {
+      resolve({ stdout, stderr, exitCode: code ?? 1 });
+    });
+    child.on("error", (err) => {
+      resolve({
+        stdout,
+        stderr: stderr + `
+Spawn error: ${err.message}`,
+        exitCode: 1
+      });
+    });
+  });
+}
+async function snapshotFileList(dir) {
+  const result = {};
+  async function walk(current) {
+    let entries;
+    try {
+      entries = await fs18.readdir(current, { withFileTypes: true });
+    } catch {
+      return;
+    }
+    for (const entry of entries) {
+      const fullPath = path18.join(current, entry.name);
+      const relativePath = path18.relative(dir, fullPath);
+      if (relativePath.startsWith(".claude")) continue;
+      if (relativePath.startsWith("node_modules")) continue;
+      if (relativePath.startsWith(".git")) continue;
+      if (entry.isDirectory()) {
+        await walk(fullPath);
+      } else {
+        try {
+          const stat = await fs18.stat(fullPath);
+          result[relativePath] = stat.mtimeMs;
+        } catch {
+        }
+      }
+    }
+  }
+  await walk(dir);
+  return result;
+}
+function diffFileLists(before, after) {
+  const changes = {};
+  for (const [file, mtime] of Object.entries(after)) {
+    if (!(file in before)) {
+      changes[file] = "created";
+    } else if (before[file] !== mtime) {
+      changes[file] = "modified";
+    }
+  }
+  for (const file of Object.keys(before)) {
+    if (!(file in after)) {
+      changes[file] = "deleted";
+    }
+  }
+  return changes;
+}
+function parseToolCalls(stdout) {
+  try {
+    const lines = stdout.split("\n").filter((l) => l.trim());
+    const toolCalls = [];
+    for (const line of lines) {
+      try {
+        const obj = JSON.parse(line);
+        if (obj.type === "tool_use" || obj.tool_name) {
+          toolCalls.push(obj);
+        }
+      } catch {
+      }
+    }
+    return toolCalls;
+  } catch {
+    return [];
+  }
+}
+
+// src/evolve/exec.ts
+import { exec as exec2 } from "child_process";
+import { promisify as promisify2 } from "util";
+var execAsync2 = promisify2(exec2);
+async function execCommand(cmd, cwd, timeoutMs = 3e4) {
+  return execAsync2(cmd, { cwd, timeout: timeoutMs });
+}
+
+// src/evolve/scorers.ts
+var COMMAND_PATTERN = /^(npm |npx |node |python |make |cargo |go |git |test |ls |cat |grep |curl )/;
+var SHELL_METACHAR_PATTERN = /[;|&`$()<>]/;
+var JUDGE_SYSTEM_PROMPT = `You are an eval judge for Claude Code agent tasks. Given a task description, expected outcome, and actual execution results, determine if the task was completed successfully.
+
+Return ONLY valid JSON with this structure:
+{
+  "pass": true/false,
+  "score": 0-100,
+  "reasoning": "Brief explanation of your judgment"
+}`;
+var RUBRIC_SYSTEM_PROMPT = `You are an eval judge scoring a specific criterion. Given the task, the criterion to evaluate, and the execution results, score the criterion.
+
+Return ONLY valid JSON:
+{
+  "score": 0.0-1.0,
+  "reasoning": "Brief explanation"
+}`;
+async function passFailScorer(task, workspacePath, stdout, stderr) {
+  const outcomes = Array.isArray(task.expected_outcome) ? task.expected_outcome : task.expected_outcome.split("\n");
+  const commands = outcomes.map((line) => line.replace(/^-\s*/, "").trim()).filter((line) => COMMAND_PATTERN.test(line));
+  if (commands.length > 0) {
+    const failures = [];
+    for (const cmd of commands) {
+      if (SHELL_METACHAR_PATTERN.test(cmd)) {
+        failures.push(`Rejected unsafe command (shell metacharacters): ${cmd}`);
+        continue;
+      }
+      try {
+        await execCommand(cmd, workspacePath);
+      } catch (err) {
+        const msg = err instanceof Error ? err.message : String(err);
+        failures.push(`Command failed: ${cmd}
+${msg}`);
+      }
+    }
+    const passed2 = failures.length === 0;
+    return {
+      pass: passed2,
+      score: passed2 ? 100 : 0,
+      details: passed2 ? `All ${commands.length} verification commands passed` : failures.join("\n")
+    };
+  }
+  const hasErrors = stderr.toLowerCase().includes("error") || stderr.toLowerCase().includes("failed") || stderr.toLowerCase().includes("exception");
+  const passed = !hasErrors;
+  return {
+    pass: passed,
+    score: passed ? 100 : 0,
+    details: passed ? "No errors detected in output" : "Errors found in stderr"
+  };
+}
+async function llmJudgeScorer(task, workspacePath, stdout, stderr, config) {
+  const expectedOutcome = Array.isArray(task.expected_outcome) ? task.expected_outcome.join("\n") : task.expected_outcome;
+  const userMessage = [
+    "## Task",
+    task.description,
+    "",
+    "## Expected Outcome",
+    expectedOutcome,
+    "",
+    "## Actual stdout (last 2000 chars)",
+    stdout.slice(-2e3),
+    "",
+    "## Actual stderr (last 1000 chars)",
+    stderr.slice(-1e3)
+  ].join("\n");
+  try {
+    const response = await callLLM(config, userMessage, {
+      systemPrompt: JUDGE_SYSTEM_PROMPT,
+      maxTokens: 1024
+    });
+    let cleaned = response.trim();
+    if (cleaned.startsWith("```")) {
+      cleaned = cleaned.replace(/^```(?:json)?\n?/, "").replace(/\n?```$/, "");
+    }
+    const jsonMatch = cleaned.match(/\{[\s\S]*\}/);
+    if (!jsonMatch) {
+      return { pass: false, score: 0, reasoning: "Judge returned invalid JSON" };
+    }
+    const result = JSON.parse(jsonMatch[0]);
+    return {
+      pass: result.pass,
+      score: result.score,
+      reasoning: result.reasoning
+    };
+  } catch (err) {
+    return {
+      pass: false,
+      score: 0,
+      reasoning: `LLM judge error: ${err instanceof Error ? err.message : String(err)}`
+    };
+  }
+}
+async function rubricScorer(task, workspacePath, stdout, stderr, config) {
+  if (!task.rubric || task.rubric.length === 0) {
+    return passFailScorer(task, workspacePath, stdout, stderr);
+  }
+  const breakdown = [];
+  let weightedSum = 0;
+  for (const criterion of task.rubric) {
+    const userMessage = [
+      "## Task",
+      task.description,
+      "",
+      "## Criterion to Evaluate",
+      `"${criterion.criterion}" (weight: ${criterion.weight})`,
+      "",
+      "## Actual stdout (last 2000 chars)",
+      stdout.slice(-2e3),
+      "",
+      "## Actual stderr (last 500 chars)",
+      stderr.slice(-500)
+    ].join("\n");
+    try {
+      const response = await callLLM(config, userMessage, {
+        systemPrompt: RUBRIC_SYSTEM_PROMPT,
+        maxTokens: 512
+      });
+      let cleaned = response.trim();
+      if (cleaned.startsWith("```")) {
+        cleaned = cleaned.replace(/^```(?:json)?\n?/, "").replace(/\n?```$/, "");
+      }
+      const jsonMatch = cleaned.match(/\{[\s\S]*\}/);
+      if (jsonMatch) {
+        const result = JSON.parse(jsonMatch[0]);
+        const clampedScore = Math.max(0, Math.min(1, result.score));
+        breakdown.push({
+          criterion: criterion.criterion,
+          score: clampedScore,
+          weight: criterion.weight
+        });
+        weightedSum += clampedScore * criterion.weight;
+      } else {
+        breakdown.push({
+          criterion: criterion.criterion,
+          score: 0,
+          weight: criterion.weight
+        });
+      }
+    } catch {
+      breakdown.push({
+        criterion: criterion.criterion,
+        score: 0,
+        weight: criterion.weight
+      });
+    }
+  }
+  const totalWeight = task.rubric.reduce((sum, c) => sum + c.weight, 0);
+  const totalScore = totalWeight > 0 ? Math.round(weightedSum / totalWeight * 100) : 0;
+  return {
+    pass: totalScore >= 60,
+    score: totalScore,
+    reasoning: `Rubric score: ${totalScore}%`,
+    breakdown
+  };
+}
+async function scoreTask(task, workspacePath, stdout, stderr, config) {
+  if (task.scoring === "pass-fail") {
+    return passFailScorer(task, workspacePath, stdout, stderr);
+  }
+  if (task.scoring === "llm-judge" && config) {
+    return llmJudgeScorer(task, workspacePath, stdout, stderr, config);
+  }
+  if (task.scoring === "rubric" && config) {
+    return rubricScorer(task, workspacePath, stdout, stderr, config);
+  }
+  return passFailScorer(task, workspacePath, stdout, stderr);
+}
+
+// src/commands/evolve.ts
+var DEFAULT_CONFIG = {
+  model: "claude-sonnet-4-6",
+  proposerModel: "claude-opus-4-6",
+  scorer: "pass-fail",
+  maxIterations: 5,
+  parallelTasks: 1
+};
+var evolveCommand = new Command11("evolve").description("Evolve your agent environment through automated optimization");
+evolveCommand.command("init").description("Initialize an evolution workspace with auto-generated tasks").option("--workflow <type>", "Workflow type for template selection", "feature-development").action(async (options) => {
+  try {
+    const projectRoot = process.cwd();
+    console.log(ui.section("Evolve Init"));
+    const claudeDir = path19.join(projectRoot, ".claude");
+    try {
+      await fs19.access(claudeDir);
+    } catch {
+      console.log(ui.error("No .claude/ directory found. Run kairn describe first."));
+      process.exit(1);
+    }
+    const workspace = await createEvolveWorkspace(projectRoot, DEFAULT_CONFIG);
+    console.log(ui.success("Created .kairn-evolve/ workspace"));
+    const spinner = ora2("Generating project-specific eval tasks...").start();
+    let tasks;
+    try {
+      tasks = await autoGenerateTasks(projectRoot, options.workflow);
+      spinner.succeed(`Generated ${tasks.length} eval tasks`);
+    } catch {
+      spinner.fail("LLM task generation failed");
+      const templateIds = selectTemplatesForWorkflow(options.workflow);
+      tasks = templateIds.map((templateId, index) => ({
+        id: `${templateId}-${index + 1}`,
+        template: templateId,
+        description: `${EVAL_TEMPLATES[templateId].description} (project-specific task \u2014 edit in tasks.yaml)`,
+        setup: "npm install",
+        expected_outcome: "Task completed successfully",
+        scoring: "pass-fail",
+        timeout: 300
+      }));
+      console.log(ui.info(`Fell back to ${tasks.length} template placeholders`));
+    }
+    for (const task of tasks) {
+      console.log(chalk14.cyan(`  ${task.id}`) + chalk14.dim(` (${task.template}) \u2014 ${task.description.slice(0, 80)}`));
+    }
+    let addMore = true;
+    while (addMore) {
+      try {
+        addMore = await confirm3({ message: "Add another eval task?", default: false });
+      } catch {
+        addMore = false;
+      }
+      if (addMore) {
+        const templateId = await select4({
+          message: "Select eval template:",
+          choices: Object.values(EVAL_TEMPLATES).map((t) => ({
+            name: `${t.name} \u2014 ${t.description}`,
+            value: t.id
+          }))
+        });
+        const addSpinner = ora2("Generating task...").start();
+        try {
+          const config = await loadConfig();
+          if (config) {
+            let claudeMd = "";
+            try {
+              claudeMd = await fs19.readFile(path19.join(claudeDir, "CLAUDE.md"), "utf-8");
+            } catch {
+            }
+            const profile = await buildProjectProfile(projectRoot);
+            const newTasks = await generateTasksFromTemplates(claudeMd, profile, [templateId], config);
+            tasks.push(...newTasks);
+            addSpinner.succeed(`Added ${newTasks.length} task(s)`);
+          } else {
+            addSpinner.fail("No config found");
+          }
+        } catch {
+          addSpinner.fail("Failed to generate task");
+        }
+      }
+    }
+    await writeTasksFile(workspace, tasks);
+    console.log(ui.success(`Wrote ${tasks.length} tasks to tasks.yaml`));
+    console.log("");
+    console.log(chalk14.dim("  Next steps:"));
+    console.log(chalk14.dim("    1. Review .kairn-evolve/tasks.yaml"));
+    console.log(chalk14.dim("    2. Run: kairn evolve baseline"));
+    console.log(chalk14.dim("    3. Run: kairn evolve run"));
+  } catch (err) {
+    const msg = err instanceof Error ? err.message : String(err);
+    console.log(ui.error(msg));
+    process.exit(1);
+  }
+});
+evolveCommand.command("baseline").description("Snapshot current .claude/ directory as baseline").action(async () => {
+  try {
+    const projectRoot = process.cwd();
+    const workspace = path19.join(projectRoot, ".kairn-evolve");
+    console.log(ui.section("Evolve Baseline"));
+    try {
+      await fs19.access(workspace);
+    } catch {
+      console.log(ui.error("No .kairn-evolve/ directory found. Run kairn evolve init first."));
+      process.exit(1);
+    }
+    await snapshotBaseline(projectRoot, workspace);
+    const baselineDir = path19.join(workspace, "baseline");
+    const fileCount = await countFiles(baselineDir);
+    console.log(ui.success(`Baseline snapshot created (${fileCount} files)`));
+  } catch (err) {
+    const msg = err instanceof Error ? err.message : String(err);
+    console.log(ui.error(msg));
+    process.exit(1);
+  }
+});
+evolveCommand.command("run").description("Run tasks against the current harness").option("--task <id>", "Run a specific task by ID").action(async (options) => {
+  try {
+    const projectRoot = process.cwd();
+    const workspace = path19.join(projectRoot, ".kairn-evolve");
+    console.log(ui.section("Evolve Run"));
+    try {
+      await fs19.access(workspace);
+    } catch {
+      console.log(ui.error("No .kairn-evolve/ directory found. Run kairn evolve init first."));
+      process.exit(1);
+    }
+    const tasksPath = path19.join(workspace, "tasks.yaml");
+    let tasksContent;
+    try {
+      tasksContent = await fs19.readFile(tasksPath, "utf-8");
+    } catch {
+      console.log(ui.error("No tasks.yaml found. Run kairn evolve init first."));
+      process.exit(1);
+    }
+    const parsed = yamlParse(tasksContent);
+    if (!parsed?.tasks || parsed.tasks.length === 0) {
+      console.log(ui.error("No tasks found in tasks.yaml"));
+      process.exit(1);
+    }
+    const tasksToRun = options.task ? parsed.tasks.filter((t) => t.id === options.task) : parsed.tasks;
+    if (tasksToRun.length === 0) {
+      console.log(ui.error(`Task "${options.task}" not found in tasks.yaml`));
+      process.exit(1);
+    }
+    console.log(ui.info(`Running ${tasksToRun.length} task(s)...`));
+    console.log("");
+    const config = await loadConfig();
+    const harnessPath = path19.join(projectRoot, ".claude");
+    const results = [];
+    for (const task of tasksToRun) {
+      const traceDir = path19.join(workspace, "traces", "0", task.id);
+      const spinner = ora2(`Running: ${task.id}`).start();
+      const result = await runTask(task, harnessPath, traceDir, 0);
+      if (config) {
+        const stdout = await fs19.readFile(path19.join(traceDir, "stdout.log"), "utf-8").catch(() => "");
+        const stderr = await fs19.readFile(path19.join(traceDir, "stderr.log"), "utf-8").catch(() => "");
+        const score = await scoreTask(task, traceDir, stdout, stderr, config);
+        result.score = score;
+        await writeScore(traceDir, score);
+      }
+      results.push(result);
+      const status = result.score.pass ? chalk14.green("PASS") : chalk14.red("FAIL");
+      const scoreStr = result.score.score !== void 0 ? chalk14.dim(` (${result.score.score}%)`) : "";
+      spinner.stop();
+      console.log(`  ${status}  ${task.id}${scoreStr}${result.score.details ? chalk14.dim(` \u2014 ${result.score.details}`) : ""}`);
+    }
+    const passed = results.filter((r) => r.score.pass).length;
+    console.log("");
+    console.log(ui.info(`Results: ${passed}/${results.length} passed`));
+    console.log(ui.info("Traces written to .kairn-evolve/traces/0/"));
+  } catch (err) {
+    const msg = err instanceof Error ? err.message : String(err);
+    console.log(ui.error(msg));
+    process.exit(1);
+  }
+});
+async function countFiles(dir) {
+  let count = 0;
+  try {
+    const entries = await fs19.readdir(dir, { withFileTypes: true });
+    for (const entry of entries) {
+      if (entry.isDirectory()) {
+        count += await countFiles(path19.join(dir, entry.name));
+      } else {
+        count++;
+      }
+    }
+  } catch {
+  }
+  return count;
+}
+
 // src/cli.ts
-var program = new Command11();
+var program = new Command12();
 program.name("kairn").description(
   "Compile natural language intent into optimized Claude Code environments"
 ).version("1.9.0").option("--no-color", "Disable colored output");
@@ -3642,8 +4579,9 @@ program.addCommand(doctorCommand);
 program.addCommand(registryCommand);
 program.addCommand(templatesCommand);
 program.addCommand(keysCommand);
+program.addCommand(evolveCommand);
 if (process.argv.includes("--no-color") || process.env.NO_COLOR) {
-  chalk14.level = 0;
+  chalk15.level = 0;
 }
 program.parse();
 //# sourceMappingURL=cli.js.map