@luanpdd/kit-mcp 1.33.0 → 1.35.0

package/src/mcp-server/index.js

@@ -1,794 +1,794 @@
-// kit-mcp server — exposes 9 tools, each with action-based dispatch (or none).
-//
-//   kit              action: list-agents | list-commands | list-skills | get | search
-//   sync             action: targets | status | install | remove
-//   gates            action: list | get | for-stage
-//   forensics        action: collect | summarize | write-learnings | list-replays | record-replay | load-replay
-//   install          action: targets | install | dry-run                    (registers this MCP into an IDE)
-//   metrics-snapshot (parameterless)                                          (OBS-18 four-golden-signals readout)
-//
-// Transport: stdio (MCP standard).
-
-import { Server } from '@modelcontextprotocol/sdk/server/index.js';
-import { StdioServerTransport } from '@modelcontextprotocol/sdk/server/stdio.js';
-import { CallToolRequestSchema, ListToolsRequestSchema } from '@modelcontextprotocol/sdk/types.js';
-
-import { readFileSync } from 'node:fs';
-import { fileURLToPath } from 'node:url';
-import path from 'node:path';
-
-import { listKit, searchKit, findItem, BUNDLED_KIT_ROOT } from '../core/kit.js';
-import { listTargets } from '../core/registry.js';
-import { syncTo, statusOf, removeFrom, summarize } from '../core/sync.js';
-import { detectReverse, applyReverse } from '../core/reverse-sync.js';
-import { validateProjectRoot } from '../core/path-safety.js';
-import { sanitizeMcpError } from '../core/error-redaction.js';
-import { listGates, getGate, gatesForStage } from '../core/gates.js';
-import { runGate } from '../core/gate-runner.js';
-import { collectFailures, summarizeByAgent, writeLearnings } from '../core/failures.js';
-import { reflect } from '../core/reflect.js';
-import { recordReplay, listReplays, loadReplay, annotateReplay } from '../core/replays.js';
-import { installMcp, listInstallTargets } from './install.js';
-import { ensureSidecar } from '../ui/auto-spawn.js';
-import { wrapProgressForUi } from '../ui/wrapper.js';
-import { incrementInvocation, recordLatency, snapshot as metricsSnapshot, persistSnapshot } from '../core/metrics.js';
-import { logEvent } from '../core/logger.js';
-import { notify, isNotifyEnabled } from '../core/notify.js';
-// Note: roots.js is imported dynamically inside handlers that need it
-// (handleAutoInstall, handleAckRestart) — keeps boot path minimal.
-
-const TOOLS = [
-  {
-    name: 'kit',
-    // Phase 170 (v1.29): description enriched with trigger keywords so MCP
-    // hosts route here on relevant intents even in MCP-pure mode (before
-    // auto-install made .claude/ native). Keep under 1024 chars (host limit).
-    // v1.30.3 (#5): counts are {{PLACEHOLDERS}} — injected live at ListTools
-    // time from listKit() so they never drift from the actual catalog.
-    description: 'Browse the personal kit: {{AGENTS}} agents, {{COMMANDS}} commands, {{SKILLS}} skills. Call this when the user mentions Supabase (RLS, branching, migrations, Edge Functions, Custom Claims, Postgres Roles, Storage, Realtime, pgvector), multi-tenant SaaS, agentic harness, characterization tests, legacy refactor, observability (SLO, golden signals, error budgets), DDIA topics (consistency, replication lag, schema evolution), SRE (postmortems, toil, PRR), CI/CD (hermetic builds, pipelines), or any workflow that benefits from the canonical patterns. Use action=search to discover, action=get to read the full prompt/skill.',
-    inputSchema: {
-      type: 'object',
-      properties: {
-        action: { type: 'string', enum: ['list-agents', 'list-commands', 'list-skills', 'get', 'search'] },
-        kind:   { type: 'string', enum: ['agent', 'command', 'skill'], description: 'For action=get' },
-        name:   { type: 'string', description: 'For action=get' },
-        query:  { type: 'string', description: 'For action=search' },
-        terse:  { type: 'boolean', description: 'For action=list-*: omit description, return only {kind, name}. Default false (PERF-15-01).' },
-        tier:   { type: 'string', enum: ['core', 'specialized'], description: 'For action=list-agents: filter by tier. core = workflow backbone (~13); specialized = domain-specific (~54). Omit for all.' },
-      },
-      required: ['action'],
-    },
-  },
-  {
-    name: 'sync',
-    description: 'Project the kit into an IDE-specific layout (markdown references by default).',
-    inputSchema: {
-      type: 'object',
-      properties: {
-        action:      { type: 'string', enum: ['targets', 'status', 'install', 'remove'] },
-        target:      { type: 'string', description: 'IDE id (e.g. claude-code, cursor, codex). Use action=targets to list.' },
-        projectRoot: { type: 'string', description: 'Defaults to cwd' },
-        mode:        { type: 'string', enum: ['reference', 'copy'], description: 'Default: reference' },
-        dryRun:      { type: 'boolean' },
-        autoSpawn:   { type: 'boolean', description: 'On action=install: auto-start the sidecar UI (kit ui) if not running and stream progress to it.' },
-      },
-      required: ['action'],
-    },
-  },
-  {
-    name: 'reverse-sync',
-    description: 'Detect and apply edits made directly in an IDE back to the canonical kit/.',
-    inputSchema: {
-      type: 'object',
-      properties: {
-        action:      { type: 'string', enum: ['detect', 'apply'] },
-        target:      { type: 'string', description: 'IDE id (e.g. claude-code, cursor)' },
-        projectRoot: { type: 'string' },
-        strategy:    { type: 'string', enum: ['skip', 'overwrite', 'merge', 'rename'], description: 'For action=apply' },
-        only:        { type: 'array', items: { type: 'string' }, description: 'For action=apply: limit to these kind/name pairs' },
-        dryRun:      { type: 'boolean' },
-        autoSpawn:   { type: 'boolean', description: 'On action=apply: auto-start the sidecar UI (kit ui) if not running and stream progress to it.' },
-      },
-      required: ['action', 'target'],
-    },
-  },
-  {
-    name: 'gates',
-    description: 'List, fetch, or execute reusable workflow gates (regression, confidence, etc).',
-    inputSchema: {
-      type: 'object',
-      properties: {
-        action:      { type: 'string', enum: ['list', 'get', 'for-stage', 'run'] },
-        id:          { type: 'string', description: 'For action=get or action=run' },
-        stage:       { type: 'string', enum: ['pre-plan', 'pre-execute', 'pre-verify', 'post-verify', 'any'], description: 'For action=for-stage' },
-        projectRoot: { type: 'string', description: 'For action=run' },
-        autoSpawn:   { type: 'boolean', description: 'On action=run: auto-start the sidecar UI (kit ui) if not running and stream progress to it.' },
-      },
-      required: ['action'],
-    },
-  },
-  {
-    name: 'forensics',
-    description: 'Failure dataset & replays — close the learning loop on failed agent runs.',
-    inputSchema: {
-      type: 'object',
-      properties: {
-        action:      { type: 'string', enum: ['collect', 'summarize', 'write-learnings', 'list-replays', 'record-replay', 'load-replay', 'annotate-replay', 'reflect'] },
-        projectRoot: { type: 'string' },
-        replayId:    { type: 'string' },
-        payload:     { type: 'object', description: 'For action=record-replay: the Task() payload to store.' },
-        outcome:     { type: 'object', description: 'For action=annotate-replay' },
-        agent:       { type: 'string', description: 'For action=reflect: agent name (e.g. executor)' },
-        dryRun:      { type: 'boolean', description: 'For action=reflect: only save the assembled prompt, no API call' },
-      },
-      required: ['action'],
-    },
-  },
-  {
-    name: 'install',
-    description: 'Register this kit-mcp server into an IDE\'s MCP config (Claude/Cursor/Codex/Gemini/Windsurf).',
-    inputSchema: {
-      type: 'object',
-      properties: {
-        action:      { type: 'string', enum: ['targets', 'install', 'dry-run'] },
-        target:      { type: 'string', description: 'IDE id. Use action=targets to list.' },
-        scope:       { type: 'string', enum: ['user', 'project'], description: 'Default: user' },
-        name:        { type: 'string', description: 'Server name in the IDE config. Default: kit' },
-        via:         { type: 'string', enum: ['local', 'npx', 'global'], description: 'How the IDE will invoke the server. Default: local (this clone)' },
-        pkg:         { type: 'string', description: 'npm package name (only with via=npx). Default: @luanpdd/kit-mcp' },
-        force:       { type: 'boolean', description: 'Overwrite existing entry with same name' },
-        projectRoot: { type: 'string' },
-      },
-      required: ['action'],
-    },
-  },
-  {
-    // OBS-18 (Phase 94.01): expose four-golden-signals data for the MCP server itself.
-    // Read-only (no auth needed beyond the underlying transport): returns counters
-    // keyed `${tool}:${status}` and per-tool latency p50/p95/p99/count.
-    name: 'metrics-snapshot',
-    description: 'Read in-memory golden-signals metrics for this MCP server (counters + latency p50/p95/p99 per tool).',
-    inputSchema: {
-      type: 'object',
-      properties: {},
-    },
-  },
-  {
-    // Phase 167 (v1.29): auto-sync the kit content into the host project's
-    // .claude/ (or equivalent) directory so agents become real subagent_types,
-    // skills get native auto-trigger, and commands appear as slash-commands.
-    // Idempotent — re-runs are no-ops if .claude/.kit-mcp-version matches the
-    // running server's package version. Permission-gated by the host.
-    name: 'auto-install',
-    description: 'IMPORTANT for first contact: project kit/ into the host\'s native layout (.claude/agents/, skills/, commands/) so {{AGENTS}} agents become real subagent_types in the Agent tool, {{SKILLS}} skills get native auto-trigger via descriptions, and {{COMMANDS}} commands appear as /slash-commands in the IDE. Idempotent — re-running is a no-op if already in sync. Run once per project on first kit-mcp contact; restart the IDE session after to load the new agents/skills/commands. After restart, call ack-restart to clear the marker.',
-    inputSchema: {
-      type: 'object',
-      properties: {
-        action:      { type: 'string', enum: ['install', 'check'], description: 'install: write files. check: read-only drift report. Default: install.' },
-        target:      { type: 'string', description: 'IDE id (claude-code, cursor, …). Defaults to claude-code.' },
-        projectRoot: { type: 'string', description: 'Override the auto-detected project root. Usually omitted — server reads it from MCP roots capability.' },
-        force:       { type: 'boolean', description: 'Re-write even if .kit-mcp-version already matches. Default: false.' },
-      },
-    },
-  },
-  {
-    // Phase 168 (v1.29): acknowledge the restart-required marker after the
-    // user reloads the IDE session. Removes .claude/.kit-mcp-restart-required
-    // so doctor stops flagging it.
-    name: 'ack-restart',
-    description: 'Acknowledge that the IDE session was restarted after kit:auto-install. Removes the .kit-mcp-restart-required marker so kit:doctor stops warning. Called automatically by the harness when it detects the marker after reload, or manually by the user.',
-    inputSchema: {
-      type: 'object',
-      properties: {
-        projectRoot: { type: 'string' },
-      },
-    },
-  },
-];
-
-// DRIFT-13-03: read version from package.json at module load (NOT inside
-// createServer — re-reading on every call adds zero value). Same pattern as
-// bin/cli.js:43-51. Both files are 2 levels deep from repo root, so the
-// '..', '..' resolution works identically. Falls back to 'unknown' if the
-// package.json lookup fails (unusual install layout).
-function readPkgVersion() {
-  try {
-    const here = path.dirname(fileURLToPath(import.meta.url));
-    const pkgPath = path.resolve(here, '..', '..', 'package.json');
-    return JSON.parse(readFileSync(pkgPath, 'utf8')).version;
-  } catch {
-    return 'unknown';
-  }
-}
-
-export const PKG_VERSION = readPkgVersion();
-
-// --- handlers ---
-
-async function handleKit(args) {
-  const kit = await listKit();
-  // PERF-15-01: terse mode skips description payload entirely. Backward-compat:
-  // args.terse undefined/false preserves slim()+summarize() cap-80 behavior.
-  const variant = args.terse === true ? slimTerse : slim;
-  switch (args.action) {
-    case 'list-agents': {
-      // #6 (v1.30.4): optional tier filter — core vs specialized. Agents
-      // without a `tier` frontmatter field are treated as specialized so a
-      // tier=core query never accidentally surfaces an untagged agent.
-      let agents = kit.agents;
-      if (args.tier === 'core' || args.tier === 'specialized') {
-        agents = agents.filter((a) => (a.frontmatter?.tier ?? 'specialized') === args.tier);
-      }
-      return agents.map(variant);
-    }
-    case 'list-commands': return kit.commands.map(variant);
-    case 'list-skills':   return [...kit.skills, ...kit.skillsExtras].map(variant);
-    case 'get': {
-      const item = findItem(kit, args.kind, args.name);
-      if (!item) return { error: `Not found: ${args.kind}/${args.name}` };
-      return { kind: item.kind, name: item.name, absPath: item.absPath, content: item.content ?? item.skillContent };
-    }
-    case 'search': return searchKit(kit, args.query ?? '');
-    default: return { error: `Unknown action: ${args.action}` };
-  }
-}
-
-// withAutoSpawn — if args.autoSpawn is set, ensure the sidecar is up and wrap
-// the user-supplied onProgress so events flow there. Otherwise pass-through.
-async function withAutoSpawn(args, tool, run) {
-  const projectRoot = args.projectRoot || process.cwd();
-  let wrapped = null;
-  let sidecarInfo = null;
-
-  if (args.autoSpawn) {
-    sidecarInfo = await ensureSidecar({ projectRoot, openBrowserOnSpawn: true });
-    if (sidecarInfo?.ready) {
-      wrapped = wrapProgressForUi(null, { projectRoot, tool });
-    }
-  }
-
-  // run(onProgress) — pass our wrapped callback (or undefined to no-op)
-  try {
-    const result = await run(wrapped);
-    if (wrapped?.done) wrapped.done({ ok: true });
-    return sidecarInfo ? { ...result, _sidecar: sidecarInfo } : result;
-  } catch (err) {
-    if (wrapped?.error) wrapped.error(err);
-    throw err;
-  }
-}
-
-async function handleSync(args) {
-  switch (args.action) {
-    case 'targets': return listTargets();
-    case 'status':
-    case 'install':
-    case 'remove': {
-      // SEC-14-03: MCP message must specify a path inside a git workspace.
-      // CLI bypasses this — bin/cli.js trusts whoever invoked it (same trust
-      // model as Phase 79.01's gates.run guard). status is read-only but
-      // included for defense-in-depth and a single uniform error surface.
-      const guard = await validateProjectRoot(args.projectRoot);
-      if (!guard.ok) return { error: guard.reason };
-      const projectRoot = guard.resolvedPath;
-      if (args.action === 'status') return statusOf(args.target, { projectRoot });
-      if (args.action === 'install')
-        return withAutoSpawn({ ...args, projectRoot }, 'sync.install', (onProgress) =>
-          syncTo(args.target, { projectRoot, mode: args.mode, dryRun: args.dryRun, onProgress }));
-      // action === 'remove'
-      return removeFrom(args.target, { projectRoot });
-    }
-    default: return { error: `Unknown action: ${args.action}` };
-  }
-}
-
-async function handleReverseSync(args) {
-  switch (args.action) {
-    case 'detect':
-    case 'apply': {
-      // SEC-14-03: same guard as handleSync — reverse-sync apply also writes
-      // to disk (kit/<file>) so it must be on the same allowlist as sync.
-      const guard = await validateProjectRoot(args.projectRoot);
-      if (!guard.ok) return { error: guard.reason };
-      const projectRoot = guard.resolvedPath;
-      if (args.action === 'detect') return detectReverse(args.target, { projectRoot });
-      // action === 'apply'
-      return withAutoSpawn({ ...args, projectRoot }, 'reverse-sync.apply', (onProgress) =>
-        applyReverse(args.target, {
-          projectRoot,
-          strategy: args.strategy, only: args.only, dryRun: args.dryRun,
-          onProgress,
-        }));
-    }
-    default: return { error: `Unknown action: ${args.action}` };
-  }
-}
-
-async function handleGates(args) {
-  switch (args.action) {
-    case 'list':      return listGates();
-    case 'get':       return getGate(args.id);
-    case 'for-stage': return gatesForStage(args.stage);
-    case 'run':
-      // SEC-13-01: MCP transport must never execute shell — runGate spawns bash with
-      // arbitrary content from gates/*.md (which reverse-sync can rewrite). Even with
-      // {yes: true}, this skips the interactive "y/N before exec" promise. The CLI
-      // entry point (`kit gates run <id>` via bin/cli.js) preserves the prompt and
-      // remains the only path to executing gates.
-      return {
-        error: 'MCP gates.run requires interactive TTY confirmation; use `kit gates run` from CLI instead.',
-      };
-    default: return { error: `Unknown action: ${args.action}` };
-  }
-}
-
-async function handleForensics(args) {
-  const projectRoot = args.projectRoot;
-  switch (args.action) {
-    case 'collect':         return collectFailures({ projectRoot });
-    case 'summarize': {
-      const failures = await collectFailures({ projectRoot });
-      return summarizeByAgent(failures);
-    }
-    case 'write-learnings': {
-      const failures = await collectFailures({ projectRoot });
-      return writeLearnings(failures, { projectRoot });
-    }
-    case 'list-replays':    return listReplays({ projectRoot });
-    case 'record-replay':   return recordReplay(args.payload, { projectRoot });
-    case 'load-replay':     return loadReplay(args.replayId, { projectRoot });
-    case 'annotate-replay': return annotateReplay(args.replayId, args.outcome, { projectRoot });
-    case 'reflect': return reflect({
-      agent: args.agent, projectRoot, dryRun: args.dryRun,
-      apply: false, interactive: false,  // MCP never auto-applies
-    });
-    default: return { error: `Unknown action: ${args.action}` };
-  }
-}
-
-async function handleInstall(args) {
-  switch (args.action) {
-    case 'targets':  return listInstallTargets();
-    case 'install':  return installMcp(args.target, { scope: args.scope, name: args.name, via: args.via, pkg: args.pkg, force: args.force, projectRoot: args.projectRoot });
-    case 'dry-run':  return installMcp(args.target, { scope: args.scope, name: args.name, via: args.via, pkg: args.pkg, force: args.force, projectRoot: args.projectRoot, dryRun: true });
-    default: return { error: `Unknown action: ${args.action}` };
-  }
-}
-
-// OBS-18 (Phase 94.01): metrics-snapshot is parameterless and read-only.
-// Returns the live snapshot synchronously — no auth, no projectRoot guard
-// (no disk reads, no shell). Wraps in an async fn for handler-API uniformity.
-//
-// OBS-20-01 (Phase 102): auto-persist throttle — clients polling rapidly
-// shouldn't create N files per second. 1s is generous vs typical 30s+ polls.
-// State is in-memory; resets on server restart. Closes the operational gap
-// where snapshots dir was empty until someone manually triggered persist.
-let _lastAutoPersistTs = 0;
-const AUTO_PERSIST_THROTTLE_MS = 1000;
-
-// Phase 167 (v1.29): auto-install handler.
-// Bridges MCP → host-native integration by writing kit/ files into .claude/
-// (or whatever the target IDE expects). Idempotent via .kit-mcp-version marker.
-// Phase 168 hooks add the restart_recommended signal to the result.
-async function handleAutoInstall(args) {
-  const action = args.action || 'install';
-  const target = args.target || 'claude-code';
-  const force = !!args.force;
-
-  // Resolve projectRoot: explicit arg > cwd fallback. (Future v1.30 will add
-  // MCP `roots` capability consumer for a tighter projectRoot signal — for now
-  // we use cwd to keep the boot path race-free with the SDK init handshake.)
-  let projectRoot = args.projectRoot;
-  const _rootsSource = projectRoot ? 'explicit' : 'cwd';
-  if (!projectRoot) projectRoot = process.cwd();
-
-  // SEC-14-03: validate project root (allowlist-of-1 — must be a real dir).
-  const guard = await validateProjectRoot(projectRoot);
-  if (!guard.ok) {
-    return { ok: false, reason: guard.reason, projectRoot, rootsSource: _rootsSource };
-  }
-  projectRoot = guard.resolvedPath;
-
-  const fs = await import('node:fs/promises');
-  const path = await import('node:path');
-  const markerPath = path.join(projectRoot, '.claude', '.kit-mcp-version');
-
-  // Read current marker if present.
-  let currentVersion = null;
-  try {
-    currentVersion = (await fs.readFile(markerPath, 'utf8')).trim();
-  } catch { /* not installed yet */ }
-
-  const targetVersion = PKG_VERSION;
-  const inSync = currentVersion === targetVersion;
-
-  // Phase 167 — action=check: read-only drift report.
-  if (action === 'check') {
-    return {
-      ok: true,
-      action: 'check',
-      target,
-      projectRoot,
-      rootsSource: _rootsSource,
-      installedVersion: currentVersion,
-      currentVersion: targetVersion,
-      inSync,
-      restartRecommended: false,
-    };
-  }
-
-  // action=install: skip if in sync and not forced.
-  if (inSync && !force) {
-    return {
-      ok: true,
-      action: 'install',
-      target,
-      projectRoot,
-      rootsSource: _rootsSource,
-      version: targetVersion,
-      skipped: true,
-      reason: 'already in sync',
-      restartRecommended: false,
-    };
-  }
-
-  // Run the sync.
-  let syncResult;
-  try {
-    syncResult = await syncTo(target, { projectRoot, mode: 'reference', dryRun: false });
-  } catch (e) {
-    return {
-      ok: false,
-      action: 'install',
-      target,
-      projectRoot,
-      rootsSource: _rootsSource,
-      reason: `sync_failed: ${e.message}`,
-    };
-  }
-
-  // Write/update marker file (.claude/.kit-mcp-version).
-  try {
-    await fs.mkdir(path.dirname(markerPath), { recursive: true });
-    await fs.writeFile(markerPath, targetVersion + '\n', 'utf8');
-  } catch (e) {
-    // Marker is best-effort — sync already succeeded. Just warn.
-    process.stderr.write(`[kit-mcp] auto-install marker write failed: ${e.message}\n`);
-  }
-
-  // v1.30.2/v1.30.4: register kit-mcp UserPromptSubmit hooks in
-  // .claude/settings.local.json. Idempotent — only adds hooks not already
-  // present. Without this step the hooks ship but never fire (user would have
-  // to edit settings.json manually per-project).
-  //   - kit-router: detects domain keywords → injects delegation directive.
-  //   - kit-attribution-reminder: injects the 1-line attribution directive.
-  try {
-    const settingsPath = path.join(projectRoot, '.claude', 'settings.local.json');
-    const HOOKS = ['kit-router', 'kit-attribution-reminder'];
-    let settings = {};
-    try {
-      const raw = await fs.readFile(settingsPath, 'utf8');
-      settings = JSON.parse(raw);
-    } catch { /* file may not exist yet */ }
-    settings.hooks = settings.hooks || {};
-    settings.hooks.UserPromptSubmit = settings.hooks.UserPromptSubmit || [];
-    let changed = false;
-    for (const hookName of HOOKS) {
-      const hookCmd = `node ${path.join(projectRoot, '.claude', 'hooks', hookName + '.cjs').replace(/\\/g, '/')}`;
-      const already = settings.hooks.UserPromptSubmit.some((entry) =>
-        Array.isArray(entry?.hooks) &&
-        entry.hooks.some((h) => typeof h?.command === 'string' && h.command.includes(hookName)));
-      if (!already) {
-        settings.hooks.UserPromptSubmit.push({
-          matcher: '*',
-          hooks: [{ type: 'command', command: hookCmd }],
-        });
-        changed = true;
-      }
-    }
-    if (changed) {
-      await fs.mkdir(path.dirname(settingsPath), { recursive: true });
-      await fs.writeFile(settingsPath, JSON.stringify(settings, null, 2) + '\n', 'utf8');
-    }
-  } catch (e) {
-    process.stderr.write(`[kit-mcp] hook registration failed (non-fatal): ${e.message}\n`);
-  }
-
-  // Phase 168 (v1.29): write .kit-mcp-restart-required so doctor/host can detect
-  // pending restart even if the user closes/reopens kit-mcp without restarting IDE.
-  try {
-    const restartMarker = path.join(projectRoot, '.claude', '.kit-mcp-restart-required');
-    const payload = JSON.stringify({
-      version: targetVersion,
-      previousVersion: currentVersion,
-      writtenAt: new Date().toISOString(),
-      reason: currentVersion
-        ? `Kit updated ${currentVersion} → ${targetVersion}`
-        : 'Initial kit install',
-    }, null, 2);
-    await fs.writeFile(restartMarker, payload + '\n', 'utf8');
-  } catch (e) {
-    process.stderr.write(`[kit-mcp] restart marker write failed: ${e.message}\n`);
-  }
-
-  return {
-    ok: true,
-    action: 'install',
-    target,
-    projectRoot,
-    rootsSource: _rootsSource,
-    version: targetVersion,
-    previousVersion: currentVersion,
-    written: (syncResult.written || []).length,
-    restartRecommended: true,
-    _kit_action: 'session_restart_recommended',
-    _kit_reason: currentVersion
-      ? `Kit updated from ${currentVersion} to ${targetVersion} — restart the IDE session so agents/skills/commands reload.`
-      : `Kit installed (v${targetVersion}) into ${path.join('.claude', '')} — restart the IDE session for native subagent_type + slash-command + skill auto-trigger integration.`,
-  };
-}
-
-// Phase 168 (v1.29): acknowledge that the IDE was restarted after a kit:auto-install.
-// Removes the .kit-mcp-restart-required marker. Read by kit:doctor (Phase 171)
-// to stop flagging "pending restart".
-async function handleAckRestart(args) {
-  let projectRoot = args.projectRoot || process.cwd();
-
-  const guard = await validateProjectRoot(projectRoot);
-  if (!guard.ok) return { ok: false, reason: guard.reason, projectRoot };
-  projectRoot = guard.resolvedPath;
-
-  const fs = await import('node:fs/promises');
-  const path = await import('node:path');
-  const restartMarker = path.join(projectRoot, '.claude', '.kit-mcp-restart-required');
-
-  let acked = false;
-  try {
-    await fs.unlink(restartMarker);
-    acked = true;
-  } catch (e) {
-    if (e.code !== 'ENOENT') {
-      return { ok: false, reason: `unlink_failed: ${e.message}`, projectRoot };
-    }
-    // ENOENT — nothing to ack, already clean. Not an error.
-  }
-
-  return {
-    ok: true,
-    projectRoot,
-    acked,
-    reason: acked ? 'restart marker removed' : 'no restart marker present (nothing to ack)',
-  };
-}
-
-async function handleMetricsSnapshot() {
-  const payload = metricsSnapshot();
-  const now = Date.now();
-  if (now - _lastAutoPersistTs >= AUTO_PERSIST_THROTTLE_MS) {
-    try {
-      await persistSnapshot();
-      _lastAutoPersistTs = now;
-    } catch (err) {
-      // OBS-20-01: graceful — log to stderr, do NOT fail the handler.
-      // In-memory snapshot still returned normally so the client tool call
-      // contract is preserved even when fs is read-only or quota-exhausted.
-      process.stderr.write(`[kit-mcp] auto-snapshot persist failed: ${err.message}\n`);
-    }
-  }
-  return payload;
-}
-
-const HANDLERS = {
-  kit:                handleKit,
-  sync:               handleSync,
-  'reverse-sync':     handleReverseSync,
-  gates:              handleGates,
-  forensics:          handleForensics,
-  install:            handleInstall,
-  'metrics-snapshot': handleMetricsSnapshot,
-  'auto-install':     handleAutoInstall,
-  'ack-restart':      handleAckRestart,
-};
-
-// Phase 167+168 test affordances — exported for unit coverage.
-// Production callers should go through the MCP dispatch (HANDLERS map).
-export const __TEST_HANDLERS = {
-  handleAutoInstall,
-  handleAckRestart,
-};
-
-function slim(x) {
-  // absPath omitted by design — list-* tools are AI-consumed in tight context budgets.
-  // Use action=get to fetch the absPath (and content) for a specific item.
-  // PERF-13-01 (TOK-02): truncate description via SUMMARY_MAX_CHARS (80) cap shared
-  // with src/core/sync.js — full description lives in each item's file under kit/.
-  // #6 (v1.30.4): include `tier` when present (agents only) so clients can group.
-  const out = { kind: x.kind, name: x.name };
-  if (x.frontmatter?.tier) out.tier = x.frontmatter.tier;
-  out.description = summarize(x.description);
-  return out;
-}
-
-// PERF-15-01: terse variant — omits description entirely. Used when MCP client
-// only needs name discovery (e.g. populating UI lists, validating slug references).
-// Default action=list-* still returns description capped via slim()/summarize().
-function slimTerse(x) {
-  const out = { kind: x.kind, name: x.name };
-  if (x.frontmatter?.tier) out.tier = x.frontmatter.tier;
-  return out;
-}
-
-// --- server bootstrap ---
-
-export async function createServer() {
-  const server = new Server(
-    { name: 'kit-mcp', version: PKG_VERSION },
-    { capabilities: { tools: {} } },
-  );
-
-  // v1.30.3 (#5): inject live catalog counts into tool descriptions so the
-  // {{AGENTS}}/{{COMMANDS}}/{{SKILLS}} placeholders never drift from reality.
-  // Falls back to the placeholder-as-literal if listKit() fails (defensive —
-  // a broken kit dir must not break tool discovery).
-  server.setRequestHandler(ListToolsRequestSchema, async () => {
-    let counts = null;
-    try {
-      const kit = await listKit(BUNDLED_KIT_ROOT);
-      counts = {
-        AGENTS: String(kit.agents.length),
-        COMMANDS: String(kit.commands.length),
-        SKILLS: String(kit.skills.length + kit.skillsExtras.length),
-      };
-    } catch { /* leave placeholders untouched */ }
-    if (!counts) return { tools: TOOLS };
-    const tools = TOOLS.map((t) => ({
-      ...t,
-      description: t.description.replace(
-        /\{\{(AGENTS|COMMANDS|SKILLS)\}\}/g,
-        (_, key) => counts[key],
-      ),
-    }));
-    return { tools };
-  });
-
-  // v1.30.1: open browser tab on FIRST kit-mcp tool invocation (not on boot —
-  // would spam tabs on IDE start). Provides visual feedback that kit-mcp is
-  // actively being used. Escape hatch: KIT_MCP_NO_UI=1 (same as auto-spawn) or
-  // KIT_MCP_NO_BROWSER=1 (sidecar runs but no browser open).
-  let kitFirstToolBrowserOpened = false;
-  server.setRequestHandler(CallToolRequestSchema, async (req) => {
-    const { name, arguments: args } = req.params;
-    const handler = HANDLERS[name];
-
-    // v1.30.1: first-tool browser open — fire-and-forget; never blocks handler.
-    // Suppressed in same conditions as boot-time sidecar (test/CI/no-ui).
-    if (!kitFirstToolBrowserOpened) {
-      kitFirstToolBrowserOpened = true;
-      const noUi = process.env.KIT_MCP_NO_UI === '1' || process.env.KIT_MCP_NO_UI === 'true';
-      const noBrowser = process.env.KIT_MCP_NO_BROWSER === '1' || process.env.KIT_MCP_NO_BROWSER === 'true';
-      const isTestRun = (process.execArgv || []).some(
-        (a) => a === '--test' || a === '--experimental-test-coverage',
-      ) || process.env.NODE_TEST_CONTEXT !== undefined;
-      const isCi = process.env.CI === 'true' || process.env.CI === '1';
-      if (!noUi && !noBrowser && !isTestRun && !isCi) {
-        const projectRoot = args?.projectRoot || process.cwd();
-        ensureSidecar({ projectRoot, openBrowserOnSpawn: true }).catch(() => {});
-      }
-    }
-
-    if (!handler) {
-      // OBS-18 (Phase 94.01): unknown-tool path counts as an error against
-      // the unknown name itself — useful signal if a client is mis-spelling
-      // a tool name in production. No latency observation (handler never ran).
-      incrementInvocation(name || 'unknown', 'error');
-      return { content: [{ type: 'text', text: JSON.stringify({ error: `Unknown tool: ${name}` }) }], isError: true };
-    }
-    // OBS-18 (Phase 94.01): timestamp the dispatch boundary. The four-golden-signals
-    // skill cares about the *user-facing* latency, which for the MCP server is the
-    // time from request receipt (we are inside the SDK callback) to the JSON envelope
-    // being ready. Date.now() is sub-millisecond-cheap and aligns with the bucket
-    // granularity we report (50/100/250/500ms thresholds in CONTEXT.md).
-    const start = Date.now();
-    const argsSize = args ? JSON.stringify(args).length : 0;
-    try {
-      const result = await handler(args ?? {});
-      const duration = Date.now() - start;
-      recordLatency(name, duration);
-      incrementInvocation(name, 'ok');
-      // Phase 158 (v1.28): JSONL log per tool call → ~/.kit-mcp/logs/*.log.
-      // Fire-and-forget; never blocks the handler.
-      try {
-        const ev = {
-          tool: name,
-          action: args?.action,
-          args_size: argsSize,
-          result_size: result ? JSON.stringify(result).length : 0,
-          duration_ms: duration,
-          status: 'ok',
-        };
-        // Phase 163 (v1.28): when KIT_MCP_INSPECT=1, also capture raw args/result
-        // so `kit inspect` can render full request/response live. Off by default
-        // because payloads can be large and may contain user paths.
-        if (process.env.KIT_MCP_INSPECT === '1' || process.env.KIT_MCP_INSPECT === 'true') {
-          ev.args = args ?? null;
-          ev.result = result ?? null;
-        }
-        logEvent(ev);
-      } catch { /* swallow */ }
-      // Phase 164 (v1.28): opt-in OS notification on success path.
-      if (isNotifyEnabled()) {
-        try { notify({ title: `kit-mcp ${name}`, body: `${args?.action ?? ''} ok (${duration}ms)` }); } catch { /* swallow */ }
-      }
-      return { content: [{ type: 'text', text: JSON.stringify(result, null, 2) }] };
-    } catch (e) {
-      // OBS-18: still record latency on the error path — half the value of a
-      // latency histogram is catching tail-latency-then-fail patterns. Status
-      // 'error' covers any thrown exception, including Phase 79.01 gates guard
-      // and the validateProjectRoot rejection (Phase 83.01).
-      const duration = Date.now() - start;
-      recordLatency(name, duration);
-      incrementInvocation(name, 'error');
-      try {
-        logEvent({
-          tool: name,
-          action: args?.action,
-          args_size: argsSize,
-          duration_ms: duration,
-          status: 'error',
-          error_type: e?.code || e?.name || 'Error',
-        });
-      } catch { /* swallow */ }
-      if (isNotifyEnabled()) {
-        try { notify({ title: `kit-mcp ${name} (error)`, body: e?.code || e?.name || 'Error' }); } catch { /* swallow */ }
-      }
-      // SEC-14-06: full stack stays in stderr for operator debug; client envelope is sanitized.
-      // sanitizeMcpError redacts secrets/paths from e.message, preserves e.code (Phase 83
-      // EMANIFESTMISMATCH invariant), and emits NO stack field.
-      console.error('[mcp-server] error in handler:', e?.stack ?? e);
-      return {
-        content: [{ type: 'text', text: JSON.stringify(sanitizeMcpError(e), null, 2) }],
-        isError: true,
-      };
-    }
-  });
-
-  return server;
-}
-
-export async function startStdio() {
-  const server = await createServer();
-  const transport = new StdioServerTransport();
-  await server.connect(transport);
-
-
-  // SRE-20-02 (Phase 105): pre-warm the kit cache to push MCP dispatch p95
-  // below 100ms. Without this, the very first tools/call against `kit` pays
-  // the full disk read (~144ms baseline on the v1.17 reference machine; ~96ms
-  // on faster hardware). Fire-and-forget: failure here is non-fatal — the
-  // next dispatch will lazily populate via the same listKit code path.
-  // This shifts the cold-path work from the first user-visible request to
-  // the boot path, where it's invisible behind IDE startup. See skill
-  // production-readiness-review (Performance axe) for the rationale.
-  listKit(BUNDLED_KIT_ROOT).catch(() => {});
-
-  // Phase 157 (v1.28): sidecar UI auto-spawn ON by default. Resolves the
-  // "kit-mcp has no terminal feedback" pain — operators can now see live
-  // tool calls in a browser without needing to set autoSpawn: true per tool.
-  // Escape hatch: KIT_MCP_NO_UI=1 (CI, headless, opt-out). Fire-and-forget:
-  // sidecar failure must never block the MCP transport (spec requires clean
-  // stdout). Errors are swallowed silently — kit doctor will surface them.
-  //
-  // Skip when running under node --test (or coverage) to avoid leaking detached
-  // child processes that hold the event loop and cause non-zero exits. Same
-  // for CI=true (GitHub Actions, etc.) — sidecar is interactive and meaningless
-  // there. End users running locally still get the default-on behavior.
-  const isTestRun = (process.execArgv || []).some(
-    (a) => a === '--test' || a === '--experimental-test-coverage',
-  ) || process.env.NODE_TEST_CONTEXT !== undefined;
-  const noUi = process.env.KIT_MCP_NO_UI === '1' || process.env.KIT_MCP_NO_UI === 'true';
-  const isCi = process.env.CI === 'true' || process.env.CI === '1';
-  if (!noUi && !isTestRun && !isCi) {
-    const projectRoot = process.cwd();
-    ensureSidecar({ projectRoot, openBrowserOnSpawn: false }).catch(() => {});
-  }
-}
+// kit-mcp server — exposes 9 tools, each with action-based dispatch (or none).
+//
+//   kit              action: list-agents | list-commands | list-skills | get | search
+//   sync             action: targets | status | install | remove
+//   gates            action: list | get | for-stage
+//   forensics        action: collect | summarize | write-learnings | list-replays | record-replay | load-replay
+//   install          action: targets | install | dry-run                    (registers this MCP into an IDE)
+//   metrics-snapshot (parameterless)                                          (OBS-18 four-golden-signals readout)
+//
+// Transport: stdio (MCP standard).
+
+import { Server } from '@modelcontextprotocol/sdk/server/index.js';
+import { StdioServerTransport } from '@modelcontextprotocol/sdk/server/stdio.js';
+import { CallToolRequestSchema, ListToolsRequestSchema } from '@modelcontextprotocol/sdk/types.js';
+
+import { readFileSync } from 'node:fs';
+import { fileURLToPath } from 'node:url';
+import path from 'node:path';
+
+import { listKit, searchKit, findItem, BUNDLED_KIT_ROOT } from '../core/kit.js';
+import { listTargets } from '../core/registry.js';
+import { syncTo, statusOf, removeFrom, summarize } from '../core/sync.js';
+import { detectReverse, applyReverse } from '../core/reverse-sync.js';
+import { validateProjectRoot } from '../core/path-safety.js';
+import { sanitizeMcpError } from '../core/error-redaction.js';
+import { listGates, getGate, gatesForStage } from '../core/gates.js';
+import { runGate } from '../core/gate-runner.js';
+import { collectFailures, summarizeByAgent, writeLearnings } from '../core/failures.js';
+import { reflect } from '../core/reflect.js';
+import { recordReplay, listReplays, loadReplay, annotateReplay } from '../core/replays.js';
+import { installMcp, listInstallTargets } from './install.js';
+import { ensureSidecar } from '../ui/auto-spawn.js';
+import { wrapProgressForUi } from '../ui/wrapper.js';
+import { incrementInvocation, recordLatency, snapshot as metricsSnapshot, persistSnapshot } from '../core/metrics.js';
+import { logEvent } from '../core/logger.js';
+import { notify, isNotifyEnabled } from '../core/notify.js';
+// Note: roots.js is imported dynamically inside handlers that need it
+// (handleAutoInstall, handleAckRestart) — keeps boot path minimal.
+
+const TOOLS = [
+  {
+    name: 'kit',
+    // Phase 170 (v1.29): description enriched with trigger keywords so MCP
+    // hosts route here on relevant intents even in MCP-pure mode (before
+    // auto-install made .claude/ native). Keep under 1024 chars (host limit).
+    // v1.30.3 (#5): counts are {{PLACEHOLDERS}} — injected live at ListTools
+    // time from listKit() so they never drift from the actual catalog.
+    description: 'Browse the personal kit: {{AGENTS}} agents, {{COMMANDS}} commands, {{SKILLS}} skills. Call this when the user mentions Supabase (RLS, branching, migrations, Edge Functions, Custom Claims, Postgres Roles, Storage, Realtime, pgvector), multi-tenant SaaS, agentic harness, characterization tests, legacy refactor, observability (SLO, golden signals, error budgets), DDIA topics (consistency, replication lag, schema evolution), SRE (postmortems, toil, PRR), CI/CD (hermetic builds, pipelines), or any workflow that benefits from the canonical patterns. Use action=search to discover, action=get to read the full prompt/skill.',
+    inputSchema: {
+      type: 'object',
+      properties: {
+        action: { type: 'string', enum: ['list-agents', 'list-commands', 'list-skills', 'get', 'search'] },
+        kind:   { type: 'string', enum: ['agent', 'command', 'skill'], description: 'For action=get' },
+        name:   { type: 'string', description: 'For action=get' },
+        query:  { type: 'string', description: 'For action=search' },
+        terse:  { type: 'boolean', description: 'For action=list-*: omit description, return only {kind, name}. Default false (PERF-15-01).' },
+        tier:   { type: 'string', enum: ['core', 'specialized'], description: 'For action=list-agents: filter by tier. core = workflow backbone (~13); specialized = domain-specific (~54). Omit for all.' },
+      },
+      required: ['action'],
+    },
+  },
+  {
+    name: 'sync',
+    description: 'Project the kit into an IDE-specific layout (markdown references by default).',
+    inputSchema: {
+      type: 'object',
+      properties: {
+        action:      { type: 'string', enum: ['targets', 'status', 'install', 'remove'] },
+        target:      { type: 'string', description: 'IDE id (e.g. claude-code, cursor, codex). Use action=targets to list.' },
+        projectRoot: { type: 'string', description: 'Defaults to cwd' },
+        mode:        { type: 'string', enum: ['reference', 'copy'], description: 'Default: reference' },
+        dryRun:      { type: 'boolean' },
+        autoSpawn:   { type: 'boolean', description: 'On action=install: auto-start the sidecar UI (kit ui) if not running and stream progress to it.' },
+      },
+      required: ['action'],
+    },
+  },
+  {
+    name: 'reverse-sync',
+    description: 'Detect and apply edits made directly in an IDE back to the canonical kit/.',
+    inputSchema: {
+      type: 'object',
+      properties: {
+        action:      { type: 'string', enum: ['detect', 'apply'] },
+        target:      { type: 'string', description: 'IDE id (e.g. claude-code, cursor)' },
+        projectRoot: { type: 'string' },
+        strategy:    { type: 'string', enum: ['skip', 'overwrite', 'merge', 'rename'], description: 'For action=apply' },
+        only:        { type: 'array', items: { type: 'string' }, description: 'For action=apply: limit to these kind/name pairs' },
+        dryRun:      { type: 'boolean' },
+        autoSpawn:   { type: 'boolean', description: 'On action=apply: auto-start the sidecar UI (kit ui) if not running and stream progress to it.' },
+      },
+      required: ['action', 'target'],
+    },
+  },
+  {
+    name: 'gates',
+    description: 'List, fetch, or execute reusable workflow gates (regression, confidence, etc).',
+    inputSchema: {
+      type: 'object',
+      properties: {
+        action:      { type: 'string', enum: ['list', 'get', 'for-stage', 'run'] },
+        id:          { type: 'string', description: 'For action=get or action=run' },
+        stage:       { type: 'string', enum: ['pre-plan', 'pre-execute', 'pre-verify', 'post-verify', 'any'], description: 'For action=for-stage' },
+        projectRoot: { type: 'string', description: 'For action=run' },
+        autoSpawn:   { type: 'boolean', description: 'On action=run: auto-start the sidecar UI (kit ui) if not running and stream progress to it.' },
+      },
+      required: ['action'],
+    },
+  },
+  {
+    name: 'forensics',
+    description: 'Failure dataset & replays — close the learning loop on failed agent runs.',
+    inputSchema: {
+      type: 'object',
+      properties: {
+        action:      { type: 'string', enum: ['collect', 'summarize', 'write-learnings', 'list-replays', 'record-replay', 'load-replay', 'annotate-replay', 'reflect'] },
+        projectRoot: { type: 'string' },
+        replayId:    { type: 'string' },
+        payload:     { type: 'object', description: 'For action=record-replay: the Task() payload to store.' },
+        outcome:     { type: 'object', description: 'For action=annotate-replay' },
+        agent:       { type: 'string', description: 'For action=reflect: agent name (e.g. executor)' },
+        dryRun:      { type: 'boolean', description: 'For action=reflect: only save the assembled prompt, no API call' },
+      },
+      required: ['action'],
+    },
+  },
+  {
+    name: 'install',
+    description: 'Register this kit-mcp server into an IDE\'s MCP config (Claude/Cursor/Codex/Gemini/Windsurf).',
+    inputSchema: {
+      type: 'object',
+      properties: {
+        action:      { type: 'string', enum: ['targets', 'install', 'dry-run'] },
+        target:      { type: 'string', description: 'IDE id. Use action=targets to list.' },
+        scope:       { type: 'string', enum: ['user', 'project'], description: 'Default: user' },
+        name:        { type: 'string', description: 'Server name in the IDE config. Default: kit' },
+        via:         { type: 'string', enum: ['local', 'npx', 'global'], description: 'How the IDE will invoke the server. Default: local (this clone)' },
+        pkg:         { type: 'string', description: 'npm package name (only with via=npx). Default: @luanpdd/kit-mcp' },
+        force:       { type: 'boolean', description: 'Overwrite existing entry with same name' },
+        projectRoot: { type: 'string' },
+      },
+      required: ['action'],
+    },
+  },
+  {
+    // OBS-18 (Phase 94.01): expose four-golden-signals data for the MCP server itself.
+    // Read-only (no auth needed beyond the underlying transport): returns counters
+    // keyed `${tool}:${status}` and per-tool latency p50/p95/p99/count.
+    name: 'metrics-snapshot',
+    description: 'Read in-memory golden-signals metrics for this MCP server (counters + latency p50/p95/p99 per tool).',
+    inputSchema: {
+      type: 'object',
+      properties: {},
+    },
+  },
+  {
+    // Phase 167 (v1.29): auto-sync the kit content into the host project's
+    // .claude/ (or equivalent) directory so agents become real subagent_types,
+    // skills get native auto-trigger, and commands appear as slash-commands.
+    // Idempotent — re-runs are no-ops if .claude/.kit-mcp-version matches the
+    // running server's package version. Permission-gated by the host.
+    name: 'auto-install',
+    description: 'IMPORTANT for first contact: project kit/ into the host\'s native layout (.claude/agents/, skills/, commands/) so {{AGENTS}} agents become real subagent_types in the Agent tool, {{SKILLS}} skills get native auto-trigger via descriptions, and {{COMMANDS}} commands appear as /slash-commands in the IDE. Idempotent — re-running is a no-op if already in sync. Run once per project on first kit-mcp contact; restart the IDE session after to load the new agents/skills/commands. After restart, call ack-restart to clear the marker.',
+    inputSchema: {
+      type: 'object',
+      properties: {
+        action:      { type: 'string', enum: ['install', 'check'], description: 'install: write files. check: read-only drift report. Default: install.' },
+        target:      { type: 'string', description: 'IDE id (claude-code, cursor, …). Defaults to claude-code.' },
+        projectRoot: { type: 'string', description: 'Override the auto-detected project root. Usually omitted — server reads it from MCP roots capability.' },
+        force:       { type: 'boolean', description: 'Re-write even if .kit-mcp-version already matches. Default: false.' },
+      },
+    },
+  },
+  {
+    // Phase 168 (v1.29): acknowledge the restart-required marker after the
+    // user reloads the IDE session. Removes .claude/.kit-mcp-restart-required
+    // so doctor stops flagging it.
+    name: 'ack-restart',
+    description: 'Acknowledge that the IDE session was restarted after kit:auto-install. Removes the .kit-mcp-restart-required marker so kit:doctor stops warning. Called automatically by the harness when it detects the marker after reload, or manually by the user.',
+    inputSchema: {
+      type: 'object',
+      properties: {
+        projectRoot: { type: 'string' },
+      },
+    },
+  },
+];
+
+// DRIFT-13-03: read version from package.json at module load (NOT inside
+// createServer — re-reading on every call adds zero value). Same pattern as
+// bin/cli.js:43-51. Both files are 2 levels deep from repo root, so the
+// '..', '..' resolution works identically. Falls back to 'unknown' if the
+// package.json lookup fails (unusual install layout).
+function readPkgVersion() {
+  try {
+    const here = path.dirname(fileURLToPath(import.meta.url));
+    const pkgPath = path.resolve(here, '..', '..', 'package.json');
+    return JSON.parse(readFileSync(pkgPath, 'utf8')).version;
+  } catch {
+    return 'unknown';
+  }
+}
+
+export const PKG_VERSION = readPkgVersion();
+
+// --- handlers ---
+
+async function handleKit(args) {
+  const kit = await listKit();
+  // PERF-15-01: terse mode skips description payload entirely. Backward-compat:
+  // args.terse undefined/false preserves slim()+summarize() cap-80 behavior.
+  const variant = args.terse === true ? slimTerse : slim;
+  switch (args.action) {
+    case 'list-agents': {
+      // #6 (v1.30.4): optional tier filter — core vs specialized. Agents
+      // without a `tier` frontmatter field are treated as specialized so a
+      // tier=core query never accidentally surfaces an untagged agent.
+      let agents = kit.agents;
+      if (args.tier === 'core' || args.tier === 'specialized') {
+        agents = agents.filter((a) => (a.frontmatter?.tier ?? 'specialized') === args.tier);
+      }
+      return agents.map(variant);
+    }
+    case 'list-commands': return kit.commands.map(variant);
+    case 'list-skills':   return [...kit.skills, ...kit.skillsExtras].map(variant);
+    case 'get': {
+      const item = findItem(kit, args.kind, args.name);
+      if (!item) return { error: `Not found: ${args.kind}/${args.name}` };
+      return { kind: item.kind, name: item.name, absPath: item.absPath, content: item.content ?? item.skillContent };
+    }
+    case 'search': return searchKit(kit, args.query ?? '');
+    default: return { error: `Unknown action: ${args.action}` };
+  }
+}
+
+// withAutoSpawn — if args.autoSpawn is set, ensure the sidecar is up and wrap
+// the user-supplied onProgress so events flow there. Otherwise pass-through.
+async function withAutoSpawn(args, tool, run) {
+  const projectRoot = args.projectRoot || process.cwd();
+  let wrapped = null;
+  let sidecarInfo = null;
+
+  if (args.autoSpawn) {
+    sidecarInfo = await ensureSidecar({ projectRoot, openBrowserOnSpawn: true });
+    if (sidecarInfo?.ready) {
+      wrapped = wrapProgressForUi(null, { projectRoot, tool });
+    }
+  }
+
+  // run(onProgress) — pass our wrapped callback (or undefined to no-op)
+  try {
+    const result = await run(wrapped);
+    if (wrapped?.done) wrapped.done({ ok: true });
+    return sidecarInfo ? { ...result, _sidecar: sidecarInfo } : result;
+  } catch (err) {
+    if (wrapped?.error) wrapped.error(err);
+    throw err;
+  }
+}
+
+async function handleSync(args) {
+  switch (args.action) {
+    case 'targets': return listTargets();
+    case 'status':
+    case 'install':
+    case 'remove': {
+      // SEC-14-03: MCP message must specify a path inside a git workspace.
+      // CLI bypasses this — bin/cli.js trusts whoever invoked it (same trust
+      // model as Phase 79.01's gates.run guard). status is read-only but
+      // included for defense-in-depth and a single uniform error surface.
+      const guard = await validateProjectRoot(args.projectRoot);
+      if (!guard.ok) return { error: guard.reason };
+      const projectRoot = guard.resolvedPath;
+      if (args.action === 'status') return statusOf(args.target, { projectRoot });
+      if (args.action === 'install')
+        return withAutoSpawn({ ...args, projectRoot }, 'sync.install', (onProgress) =>
+          syncTo(args.target, { projectRoot, mode: args.mode, dryRun: args.dryRun, onProgress }));
+      // action === 'remove'
+      return removeFrom(args.target, { projectRoot });
+    }
+    default: return { error: `Unknown action: ${args.action}` };
+  }
+}
+
+async function handleReverseSync(args) {
+  switch (args.action) {
+    case 'detect':
+    case 'apply': {
+      // SEC-14-03: same guard as handleSync — reverse-sync apply also writes
+      // to disk (kit/<file>) so it must be on the same allowlist as sync.
+      const guard = await validateProjectRoot(args.projectRoot);
+      if (!guard.ok) return { error: guard.reason };
+      const projectRoot = guard.resolvedPath;
+      if (args.action === 'detect') return detectReverse(args.target, { projectRoot });
+      // action === 'apply'
+      return withAutoSpawn({ ...args, projectRoot }, 'reverse-sync.apply', (onProgress) =>
+        applyReverse(args.target, {
+          projectRoot,
+          strategy: args.strategy, only: args.only, dryRun: args.dryRun,
+          onProgress,
+        }));
+    }
+    default: return { error: `Unknown action: ${args.action}` };
+  }
+}
+
+async function handleGates(args) {
+  switch (args.action) {
+    case 'list':      return listGates();
+    case 'get':       return getGate(args.id);
+    case 'for-stage': return gatesForStage(args.stage);
+    case 'run':
+      // SEC-13-01: MCP transport must never execute shell — runGate spawns bash with
+      // arbitrary content from gates/*.md (which reverse-sync can rewrite). Even with
+      // {yes: true}, this skips the interactive "y/N before exec" promise. The CLI
+      // entry point (`kit gates run <id>` via bin/cli.js) preserves the prompt and
+      // remains the only path to executing gates.
+      return {
+        error: 'MCP gates.run requires interactive TTY confirmation; use `kit gates run` from CLI instead.',
+      };
+    default: return { error: `Unknown action: ${args.action}` };
+  }
+}
+
+async function handleForensics(args) {
+  const projectRoot = args.projectRoot;
+  switch (args.action) {
+    case 'collect':         return collectFailures({ projectRoot });
+    case 'summarize': {
+      const failures = await collectFailures({ projectRoot });
+      return summarizeByAgent(failures);
+    }
+    case 'write-learnings': {
+      const failures = await collectFailures({ projectRoot });
+      return writeLearnings(failures, { projectRoot });
+    }
+    case 'list-replays':    return listReplays({ projectRoot });
+    case 'record-replay':   return recordReplay(args.payload, { projectRoot });
+    case 'load-replay':     return loadReplay(args.replayId, { projectRoot });
+    case 'annotate-replay': return annotateReplay(args.replayId, args.outcome, { projectRoot });
+    case 'reflect': return reflect({
+      agent: args.agent, projectRoot, dryRun: args.dryRun,
+      apply: false, interactive: false,  // MCP never auto-applies
+    });
+    default: return { error: `Unknown action: ${args.action}` };
+  }
+}
+
+async function handleInstall(args) {
+  switch (args.action) {
+    case 'targets':  return listInstallTargets();
+    case 'install':  return installMcp(args.target, { scope: args.scope, name: args.name, via: args.via, pkg: args.pkg, force: args.force, projectRoot: args.projectRoot });
+    case 'dry-run':  return installMcp(args.target, { scope: args.scope, name: args.name, via: args.via, pkg: args.pkg, force: args.force, projectRoot: args.projectRoot, dryRun: true });
+    default: return { error: `Unknown action: ${args.action}` };
+  }
+}
+
+// OBS-18 (Phase 94.01): metrics-snapshot is parameterless and read-only.
+// Returns the live snapshot synchronously — no auth, no projectRoot guard
+// (no disk reads, no shell). Wraps in an async fn for handler-API uniformity.
+//
+// OBS-20-01 (Phase 102): auto-persist throttle — clients polling rapidly
+// shouldn't create N files per second. 1s is generous vs typical 30s+ polls.
+// State is in-memory; resets on server restart. Closes the operational gap
+// where snapshots dir was empty until someone manually triggered persist.
+let _lastAutoPersistTs = 0;
+const AUTO_PERSIST_THROTTLE_MS = 1000;
+
+// Phase 167 (v1.29): auto-install handler.
+// Bridges MCP → host-native integration by writing kit/ files into .claude/
+// (or whatever the target IDE expects). Idempotent via .kit-mcp-version marker.
+// Phase 168 hooks add the restart_recommended signal to the result.
+async function handleAutoInstall(args) {
+  const action = args.action || 'install';
+  const target = args.target || 'claude-code';
+  const force = !!args.force;
+
+  // Resolve projectRoot: explicit arg > cwd fallback. (Future v1.30 will add
+  // MCP `roots` capability consumer for a tighter projectRoot signal — for now
+  // we use cwd to keep the boot path race-free with the SDK init handshake.)
+  let projectRoot = args.projectRoot;
+  const _rootsSource = projectRoot ? 'explicit' : 'cwd';
+  if (!projectRoot) projectRoot = process.cwd();
+
+  // SEC-14-03: validate project root (allowlist-of-1 — must be a real dir).
+  const guard = await validateProjectRoot(projectRoot);
+  if (!guard.ok) {
+    return { ok: false, reason: guard.reason, projectRoot, rootsSource: _rootsSource };
+  }
+  projectRoot = guard.resolvedPath;
+
+  const fs = await import('node:fs/promises');
+  const path = await import('node:path');
+  const markerPath = path.join(projectRoot, '.claude', '.kit-mcp-version');
+
+  // Read current marker if present.
+  let currentVersion = null;
+  try {
+    currentVersion = (await fs.readFile(markerPath, 'utf8')).trim();
+  } catch { /* not installed yet */ }
+
+  const targetVersion = PKG_VERSION;
+  const inSync = currentVersion === targetVersion;
+
+  // Phase 167 — action=check: read-only drift report.
+  if (action === 'check') {
+    return {
+      ok: true,
+      action: 'check',
+      target,
+      projectRoot,
+      rootsSource: _rootsSource,
+      installedVersion: currentVersion,
+      currentVersion: targetVersion,
+      inSync,
+      restartRecommended: false,
+    };
+  }
+
+  // action=install: skip if in sync and not forced.
+  if (inSync && !force) {
+    return {
+      ok: true,
+      action: 'install',
+      target,
+      projectRoot,
+      rootsSource: _rootsSource,
+      version: targetVersion,
+      skipped: true,
+      reason: 'already in sync',
+      restartRecommended: false,
+    };
+  }
+
+  // Run the sync.
+  let syncResult;
+  try {
+    syncResult = await syncTo(target, { projectRoot, mode: 'reference', dryRun: false });
+  } catch (e) {
+    return {
+      ok: false,
+      action: 'install',
+      target,
+      projectRoot,
+      rootsSource: _rootsSource,
+      reason: `sync_failed: ${e.message}`,
+    };
+  }
+
+  // Write/update marker file (.claude/.kit-mcp-version).
+  try {
+    await fs.mkdir(path.dirname(markerPath), { recursive: true });
+    await fs.writeFile(markerPath, targetVersion + '\n', 'utf8');
+  } catch (e) {
+    // Marker is best-effort — sync already succeeded. Just warn.
+    process.stderr.write(`[kit-mcp] auto-install marker write failed: ${e.message}\n`);
+  }
+
+  // v1.30.2/v1.30.4: register kit-mcp UserPromptSubmit hooks in
+  // .claude/settings.local.json. Idempotent — only adds hooks not already
+  // present. Without this step the hooks ship but never fire (user would have
+  // to edit settings.json manually per-project).
+  //   - kit-router: detects domain keywords → injects delegation directive.
+  //   - kit-attribution-reminder: injects the 1-line attribution directive.
+  try {
+    const settingsPath = path.join(projectRoot, '.claude', 'settings.local.json');
+    const HOOKS = ['kit-router', 'kit-attribution-reminder'];
+    let settings = {};
+    try {
+      const raw = await fs.readFile(settingsPath, 'utf8');
+      settings = JSON.parse(raw);
+    } catch { /* file may not exist yet */ }
+    settings.hooks = settings.hooks || {};
+    settings.hooks.UserPromptSubmit = settings.hooks.UserPromptSubmit || [];
+    let changed = false;
+    for (const hookName of HOOKS) {
+      const hookCmd = `node ${path.join(projectRoot, '.claude', 'hooks', hookName + '.cjs').replace(/\\/g, '/')}`;
+      const already = settings.hooks.UserPromptSubmit.some((entry) =>
+        Array.isArray(entry?.hooks) &&
+        entry.hooks.some((h) => typeof h?.command === 'string' && h.command.includes(hookName)));
+      if (!already) {
+        settings.hooks.UserPromptSubmit.push({
+          matcher: '*',
+          hooks: [{ type: 'command', command: hookCmd }],
+        });
+        changed = true;
+      }
+    }
+    if (changed) {
+      await fs.mkdir(path.dirname(settingsPath), { recursive: true });
+      await fs.writeFile(settingsPath, JSON.stringify(settings, null, 2) + '\n', 'utf8');
+    }
+  } catch (e) {
+    process.stderr.write(`[kit-mcp] hook registration failed (non-fatal): ${e.message}\n`);
+  }
+
+  // Phase 168 (v1.29): write .kit-mcp-restart-required so doctor/host can detect
+  // pending restart even if the user closes/reopens kit-mcp without restarting IDE.
+  try {
+    const restartMarker = path.join(projectRoot, '.claude', '.kit-mcp-restart-required');
+    const payload = JSON.stringify({
+      version: targetVersion,
+      previousVersion: currentVersion,
+      writtenAt: new Date().toISOString(),
+      reason: currentVersion
+        ? `Kit updated ${currentVersion} → ${targetVersion}`
+        : 'Initial kit install',
+    }, null, 2);
+    await fs.writeFile(restartMarker, payload + '\n', 'utf8');
+  } catch (e) {
+    process.stderr.write(`[kit-mcp] restart marker write failed: ${e.message}\n`);
+  }
+
+  return {
+    ok: true,
+    action: 'install',
+    target,
+    projectRoot,
+    rootsSource: _rootsSource,
+    version: targetVersion,
+    previousVersion: currentVersion,
+    written: (syncResult.written || []).length,
+    restartRecommended: true,
+    _kit_action: 'session_restart_recommended',
+    _kit_reason: currentVersion
+      ? `Kit updated from ${currentVersion} to ${targetVersion} — restart the IDE session so agents/skills/commands reload.`
+      : `Kit installed (v${targetVersion}) into ${path.join('.claude', '')} — restart the IDE session for native subagent_type + slash-command + skill auto-trigger integration.`,
+  };
+}
+
+// Phase 168 (v1.29): acknowledge that the IDE was restarted after a kit:auto-install.
+// Removes the .kit-mcp-restart-required marker. Read by kit:doctor (Phase 171)
+// to stop flagging "pending restart".
+async function handleAckRestart(args) {
+  let projectRoot = args.projectRoot || process.cwd();
+
+  const guard = await validateProjectRoot(projectRoot);
+  if (!guard.ok) return { ok: false, reason: guard.reason, projectRoot };
+  projectRoot = guard.resolvedPath;
+
+  const fs = await import('node:fs/promises');
+  const path = await import('node:path');
+  const restartMarker = path.join(projectRoot, '.claude', '.kit-mcp-restart-required');
+
+  let acked = false;
+  try {
+    await fs.unlink(restartMarker);
+    acked = true;
+  } catch (e) {
+    if (e.code !== 'ENOENT') {
+      return { ok: false, reason: `unlink_failed: ${e.message}`, projectRoot };
+    }
+    // ENOENT — nothing to ack, already clean. Not an error.
+  }
+
+  return {
+    ok: true,
+    projectRoot,
+    acked,
+    reason: acked ? 'restart marker removed' : 'no restart marker present (nothing to ack)',
+  };
+}
+
+async function handleMetricsSnapshot() {
+  const payload = metricsSnapshot();
+  const now = Date.now();
+  if (now - _lastAutoPersistTs >= AUTO_PERSIST_THROTTLE_MS) {
+    try {
+      await persistSnapshot();
+      _lastAutoPersistTs = now;
+    } catch (err) {
+      // OBS-20-01: graceful — log to stderr, do NOT fail the handler.
+      // In-memory snapshot still returned normally so the client tool call
+      // contract is preserved even when fs is read-only or quota-exhausted.
+      process.stderr.write(`[kit-mcp] auto-snapshot persist failed: ${err.message}\n`);
+    }
+  }
+  return payload;
+}
+
+const HANDLERS = {
+  kit:                handleKit,
+  sync:               handleSync,
+  'reverse-sync':     handleReverseSync,
+  gates:              handleGates,
+  forensics:          handleForensics,
+  install:            handleInstall,
+  'metrics-snapshot': handleMetricsSnapshot,
+  'auto-install':     handleAutoInstall,
+  'ack-restart':      handleAckRestart,
+};
+
+// Phase 167+168 test affordances — exported for unit coverage.
+// Production callers should go through the MCP dispatch (HANDLERS map).
+export const __TEST_HANDLERS = {
+  handleAutoInstall,
+  handleAckRestart,
+};
+
+function slim(x) {
+  // absPath omitted by design — list-* tools are AI-consumed in tight context budgets.
+  // Use action=get to fetch the absPath (and content) for a specific item.
+  // PERF-13-01 (TOK-02): truncate description via SUMMARY_MAX_CHARS (80) cap shared
+  // with src/core/sync.js — full description lives in each item's file under kit/.
+  // #6 (v1.30.4): include `tier` when present (agents only) so clients can group.
+  const out = { kind: x.kind, name: x.name };
+  if (x.frontmatter?.tier) out.tier = x.frontmatter.tier;
+  out.description = summarize(x.description);
+  return out;
+}
+
+// PERF-15-01: terse variant — omits description entirely. Used when MCP client
+// only needs name discovery (e.g. populating UI lists, validating slug references).
+// Default action=list-* still returns description capped via slim()/summarize().
+function slimTerse(x) {
+  const out = { kind: x.kind, name: x.name };
+  if (x.frontmatter?.tier) out.tier = x.frontmatter.tier;
+  return out;
+}
+
+// --- server bootstrap ---
+
+export async function createServer() {
+  const server = new Server(
+    { name: 'kit-mcp', version: PKG_VERSION },
+    { capabilities: { tools: {} } },
+  );
+
+  // v1.30.3 (#5): inject live catalog counts into tool descriptions so the
+  // {{AGENTS}}/{{COMMANDS}}/{{SKILLS}} placeholders never drift from reality.
+  // Falls back to the placeholder-as-literal if listKit() fails (defensive —
+  // a broken kit dir must not break tool discovery).
+  server.setRequestHandler(ListToolsRequestSchema, async () => {
+    let counts = null;
+    try {
+      const kit = await listKit(BUNDLED_KIT_ROOT);
+      counts = {
+        AGENTS: String(kit.agents.length),
+        COMMANDS: String(kit.commands.length),
+        SKILLS: String(kit.skills.length + kit.skillsExtras.length),
+      };
+    } catch { /* leave placeholders untouched */ }
+    if (!counts) return { tools: TOOLS };
+    const tools = TOOLS.map((t) => ({
+      ...t,
+      description: t.description.replace(
+        /\{\{(AGENTS|COMMANDS|SKILLS)\}\}/g,
+        (_, key) => counts[key],
+      ),
+    }));
+    return { tools };
+  });
+
+  // v1.30.1: open browser tab on FIRST kit-mcp tool invocation (not on boot —
+  // would spam tabs on IDE start). Provides visual feedback that kit-mcp is
+  // actively being used. Escape hatch: KIT_MCP_NO_UI=1 (same as auto-spawn) or
+  // KIT_MCP_NO_BROWSER=1 (sidecar runs but no browser open).
+  let kitFirstToolBrowserOpened = false;
+  server.setRequestHandler(CallToolRequestSchema, async (req) => {
+    const { name, arguments: args } = req.params;
+    const handler = HANDLERS[name];
+
+    // v1.30.1: first-tool browser open — fire-and-forget; never blocks handler.
+    // Suppressed in same conditions as boot-time sidecar (test/CI/no-ui).
+    if (!kitFirstToolBrowserOpened) {
+      kitFirstToolBrowserOpened = true;
+      const noUi = process.env.KIT_MCP_NO_UI === '1' || process.env.KIT_MCP_NO_UI === 'true';
+      const noBrowser = process.env.KIT_MCP_NO_BROWSER === '1' || process.env.KIT_MCP_NO_BROWSER === 'true';
+      const isTestRun = (process.execArgv || []).some(
+        (a) => a === '--test' || a === '--experimental-test-coverage',
+      ) || process.env.NODE_TEST_CONTEXT !== undefined;
+      const isCi = process.env.CI === 'true' || process.env.CI === '1';
+      if (!noUi && !noBrowser && !isTestRun && !isCi) {
+        const projectRoot = args?.projectRoot || process.cwd();
+        ensureSidecar({ projectRoot, openBrowserOnSpawn: true }).catch(() => {});
+      }
+    }
+
+    if (!handler) {
+      // OBS-18 (Phase 94.01): unknown-tool path counts as an error against
+      // the unknown name itself — useful signal if a client is mis-spelling
+      // a tool name in production. No latency observation (handler never ran).
+      incrementInvocation(name || 'unknown', 'error');
+      return { content: [{ type: 'text', text: JSON.stringify({ error: `Unknown tool: ${name}` }) }], isError: true };
+    }
+    // OBS-18 (Phase 94.01): timestamp the dispatch boundary. The four-golden-signals
+    // skill cares about the *user-facing* latency, which for the MCP server is the
+    // time from request receipt (we are inside the SDK callback) to the JSON envelope
+    // being ready. Date.now() is sub-millisecond-cheap and aligns with the bucket
+    // granularity we report (50/100/250/500ms thresholds in CONTEXT.md).
+    const start = Date.now();
+    const argsSize = args ? JSON.stringify(args).length : 0;
+    try {
+      const result = await handler(args ?? {});
+      const duration = Date.now() - start;
+      recordLatency(name, duration);
+      incrementInvocation(name, 'ok');
+      // Phase 158 (v1.28): JSONL log per tool call → ~/.kit-mcp/logs/*.log.
+      // Fire-and-forget; never blocks the handler.
+      try {
+        const ev = {
+          tool: name,
+          action: args?.action,
+          args_size: argsSize,
+          result_size: result ? JSON.stringify(result).length : 0,
+          duration_ms: duration,
+          status: 'ok',
+        };
+        // Phase 163 (v1.28): when KIT_MCP_INSPECT=1, also capture raw args/result
+        // so `kit inspect` can render full request/response live. Off by default
+        // because payloads can be large and may contain user paths.
+        if (process.env.KIT_MCP_INSPECT === '1' || process.env.KIT_MCP_INSPECT === 'true') {
+          ev.args = args ?? null;
+          ev.result = result ?? null;
+        }
+        logEvent(ev);
+      } catch { /* swallow */ }
+      // Phase 164 (v1.28): opt-in OS notification on success path.
+      if (isNotifyEnabled()) {
+        try { notify({ title: `kit-mcp ${name}`, body: `${args?.action ?? ''} ok (${duration}ms)` }); } catch { /* swallow */ }
+      }
+      return { content: [{ type: 'text', text: JSON.stringify(result, null, 2) }] };
+    } catch (e) {
+      // OBS-18: still record latency on the error path — half the value of a
+      // latency histogram is catching tail-latency-then-fail patterns. Status
+      // 'error' covers any thrown exception, including Phase 79.01 gates guard
+      // and the validateProjectRoot rejection (Phase 83.01).
+      const duration = Date.now() - start;
+      recordLatency(name, duration);
+      incrementInvocation(name, 'error');
+      try {
+        logEvent({
+          tool: name,
+          action: args?.action,
+          args_size: argsSize,
+          duration_ms: duration,
+          status: 'error',
+          error_type: e?.code || e?.name || 'Error',
+        });
+      } catch { /* swallow */ }
+      if (isNotifyEnabled()) {
+        try { notify({ title: `kit-mcp ${name} (error)`, body: e?.code || e?.name || 'Error' }); } catch { /* swallow */ }
+      }
+      // SEC-14-06: full stack stays in stderr for operator debug; client envelope is sanitized.
+      // sanitizeMcpError redacts secrets/paths from e.message, preserves e.code (Phase 83
+      // EMANIFESTMISMATCH invariant), and emits NO stack field.
+      console.error('[mcp-server] error in handler:', e?.stack ?? e);
+      return {
+        content: [{ type: 'text', text: JSON.stringify(sanitizeMcpError(e), null, 2) }],
+        isError: true,
+      };
+    }
+  });
+
+  return server;
+}
+
+export async function startStdio() {
+  const server = await createServer();
+  const transport = new StdioServerTransport();
+  await server.connect(transport);
+
+
+  // SRE-20-02 (Phase 105): pre-warm the kit cache to push MCP dispatch p95
+  // below 100ms. Without this, the very first tools/call against `kit` pays
+  // the full disk read (~144ms baseline on the v1.17 reference machine; ~96ms
+  // on faster hardware). Fire-and-forget: failure here is non-fatal — the
+  // next dispatch will lazily populate via the same listKit code path.
+  // This shifts the cold-path work from the first user-visible request to
+  // the boot path, where it's invisible behind IDE startup. See skill
+  // production-readiness-review (Performance axe) for the rationale.
+  listKit(BUNDLED_KIT_ROOT).catch(() => {});
+
+  // Phase 157 (v1.28): sidecar UI auto-spawn ON by default. Resolves the
+  // "kit-mcp has no terminal feedback" pain — operators can now see live
+  // tool calls in a browser without needing to set autoSpawn: true per tool.
+  // Escape hatch: KIT_MCP_NO_UI=1 (CI, headless, opt-out). Fire-and-forget:
+  // sidecar failure must never block the MCP transport (spec requires clean
+  // stdout). Errors are swallowed silently — kit doctor will surface them.
+  //
+  // Skip when running under node --test (or coverage) to avoid leaking detached
+  // child processes that hold the event loop and cause non-zero exits. Same
+  // for CI=true (GitHub Actions, etc.) — sidecar is interactive and meaningless
+  // there. End users running locally still get the default-on behavior.
+  const isTestRun = (process.execArgv || []).some(
+    (a) => a === '--test' || a === '--experimental-test-coverage',
+  ) || process.env.NODE_TEST_CONTEXT !== undefined;
+  const noUi = process.env.KIT_MCP_NO_UI === '1' || process.env.KIT_MCP_NO_UI === 'true';
+  const isCi = process.env.CI === 'true' || process.env.CI === '1';
+  if (!noUi && !isTestRun && !isCi) {
+    const projectRoot = process.cwd();
+    ensureSidecar({ projectRoot, openBrowserOnSpawn: false }).catch(() => {});
+  }
+}