chore(ai-tool-config): tighten Claude Code rules on checks, public config and v-html

CHANGELOG.md

@@ -2,6 +2,11 @@
 
 ## Unreleased
 
+- The Claude Code rules now ask for `pnpm run typecheck` next to lint and tests,
+  and rule out getting checks green by suppressing them (`@ts-ignore`, `as any`,
+  `eslint-disable`, skipped tests, `--no-verify`). `security.md` adds that nothing
+  secret goes under `runtimeConfig.public` and that user content is never rendered
+  with `v-html`. Existing projects get this via `battlestack pull`.
 - The generated Dockerfile pins the latest `node:24-alpine` image.
 - The project name recorded in `.battlestack/manifest.json` now always wins over
   the directory name. Before, any project command in a clone, CI checkout or

packages/preset-nuxt4/src/features/ai-tool-config.ts

@@ -16,7 +16,7 @@ const SUPPORTED: AiTool[] = ['claude-code', 'gemini-cli', 'cursor', 'codex']
 /** AI coding tool rules + MCP server scaffolding. Selection via `state.aiTool` (default: `claude-code`). */
 export const aiToolConfigFeature: Feature = {
     id: 'shared:ai-tool-config',
-    version: '1.2.0',
+    version: '1.2.1',
     label: 'AI coding tool config',
     stage: STAGE.AI_TOOL_CONFIG,
 

packages/preset-nuxt4/templates/ai-tool-config/claude-code/.claude/rules/global.md

@@ -1,7 +1,7 @@
 ## Project Conventions
 
 - Default package manager is **pnpm**. `bun`, `npm`, `yarn` also work; match the project's `packageManager` field. Use `pnpm add`, `pnpm dlx`, `pnpm run`.
-- Always run `pnpm run lint` and `pnpm test` before considering work complete.
+- Always run `pnpm run lint`, `pnpm run typecheck` and `pnpm test` before considering work complete. Fix the cause of a failure; never get green with `@ts-ignore`, `as any`, `eslint-disable`, a skipped or deleted test, or `git commit --no-verify`.
 - The workflows in `.claude/skills/` only load when invoked, so reach for them yourself: `task-arbiter` before a task, `brainstorming` before a new feature, `test-driven-development` while coding, `systematic-debugging` for a bug, `verification-before-completion` before calling it done.
 
 ## Pending upstream merges: `.battlestack/pull/`

packages/preset-nuxt4/templates/ai-tool-config/claude-code/.claude/rules/security.md

@@ -55,6 +55,7 @@ const body = await readValidatedBody(event, schema.parse)
 - Secrets in `.env`; never hardcode API keys, passwords, or tokens
 - `.env` is in `.gitignore`; `.env.example` is committed (with placeholder values)
 - Use `useRuntimeConfig()` in server routes for Nuxt env access
+- Nothing secret goes under `runtimeConfig.public`: it is serialized into every page
 - AI gateway key, admin password, session secret: all from `.env`
 
 ## Headers and Cookies
@@ -69,4 +70,5 @@ const body = await readValidatedBody(event, schema.parse)
 - Don't log passwords or sensitive data (seed script is silent)
 - Don't disable ESLint security rules
 - Don't use unsafe code execution patterns
+- Don't render user-controlled content with `v-html`; `{{ }}` escapes it, `v-html` does not, and nothing on the server strips HTML
 - Don't store sessions in localStorage; use httpOnly cookies