CLAUDE.md

Rama main · réplica de lectura del repositorio original.

@AGENTS.md

## Claude Code — additional guidance

### Preferred workflow

1. Read the relevant skill file in `.agents/skills/` before writing code for a new feature, API, database table, worker, or integration.
   - New feature or new operation: build and test the public API procedure (`api/public.ts`) **before** the UI — AI agents are the primary users (AGENTS.md invariant 23).
2. When touching `packages/database`, always run `pnpm --filter @chatbotx.io/database db:migrate` after schema changes.
3. After any code change, run `pnpm lint` and `pnpm --filter <app> check-types` before reporting done.
4. Use `pnpm fix` (Biome auto-fix) instead of manually formatting code.

### Skill → task mapping

| Task | Skill to read first |
|------|---------------------|
| Broad/ambiguous request, onboarding, "where does X live" | `chatbotx-basecode` |
| New feature / page (then `orpc-api` for the public API first, `cli-mcp-docs` last) | `feature-scaffold` |
| Builder UI component, form, table, dialog, or any user-facing string | `builder-ui-i18n` |
| New API endpoint | `orpc-api` |
| Result-returning operation (link, preview, test run/test send, stats, export, AI generation) — see "Result-returning operations" | `orpc-api` |
| After any public API change (every one surfaces in the CLI/MCP), or CLI/MCP docs/skill drift | `cli-mcp-docs` |
| Business logic, new service method, any DB read/write from app code | `business-data-access` |
| New DB table or migration | `drizzle-database` |
| New table / FK / workspace-scoped data, or any change to workspace deletion/purge/freeze | `workspace-purge` |
| New background job or queue | `worker-development` |
| New channel integration | `integration-channel` |
| Contact filter field/operator, filter SQL, or contact-based audience | `contact-filter` |
| Comment automation (auto-reply/like/hide on post comments — Messenger, Instagram, Threads, TikTok) | `comment-automation` |
| Minigame tool (Jackpot CRUD, prize draw, public play link/token, extending to a new minigame type) | `minigame` |
| New flow step with states (success/error/skip routing) | `flow-step-development` |
| New resource users build (shareable via Templates), new field/id reference on one (flow step, trigger action, contact filter, JSON column), or a write/delete path / list / editor of a template-copied kind | `template-integration` |
| Dev/build/lint commands | `turborepo-workflow` |
| Approved implementation plan | `implement-plan` |
| Security-sensitive change (auth, scoping, webhooks, AI tools, permissions) | `security-review` |
| Writing tests / verifying a change is done | `testing-workflow` |
| Fixing a bug, flaky test, or performance regression | `diagnosing-bugs` |
| Concurrent code (worker jobs, migrations, replace-writes) | `reliability-concurrency` |

### Specialist agents (`.claude/agents/`)

Dispatch these project subagents by name when relevant:

| Agent | When |
|-------|------|
| `invariant-guard` | after editing `apps/`/`packages/`/`integrations/` — checks the ChatbotX invariants no linter enforces |
| `rag-eval` | changes under `packages/ai` or the embedding repositories/handlers |
| `incident-responder` | triaging a production error or failed job (then `diagnosing-bugs` to reproduce and fix it locally) |

Reviewers, planners, build-fixers, etc. come from the `~/.claude/` global set — don't recreate them here.

### Model routing (cost)

Default to the cheapest tier that fits the task; reserve the top tier for judgment:

| Task class | Tier |
|------------|------|
| file lookup, grep, inventory, verification/refutation | Haiku / Sonnet |
| implementation, code review, debugging | Sonnet |
| architecture, synthesis, incident root-cause | Opus |

### Never do without checking

- Stage specific files only. A `PreToolUse` hook (`.claude/hooks/block-dangerous-git.mjs`) blocks `git add -A`/`.`, force push, `reset --hard`, `clean -f`, `branch -D`, and `checkout .`/`restore .`; if one is truly needed, ask the user to run it.
- Do not commit `.env` files or secrets.
- Do not skip `pnpm lint` — the CI will fail.
- Do not hardcode user-facing strings — use `useTranslations()`.
- Do not ship a UI-only feature or UI-only button — every operation, including result-returning ones like the bot simulator link, needs a public API procedure unless it is on the exempt list of invariant 23 in `docs/agents/invariants.md`.
- Do not import `db` directly in `apps/` or `integrations/` — the chain is `action | API handler → service (@chatbotx.io/business) → repository (@chatbotx.io/database/repositories) → DB`; app code calls a service, never a repository directly (the one exception is a pure read with zero business logic). See `.agents/rules/data-access.md`.
- Do not put Node-only code (`node:crypto`, `node:stream`, `fs`, any `process.*` but `env`) in a module the builder's oRPC router reaches — it lands in the Edge instrumentation bundle. Split it into its own file + package subpath (outside `index.ts`) imported only by actions/public API/workers; never gate `instrumentation.ts` instead. Guarded by `apps/builder/__tests__/edge-instrumentation-graph.test.ts`; see AGENTS.md invariant 25.
- Do not use dynamic `import()` in tsdown-built code (`packages/*`, `integrations/*`, `apps/worker`, `apps/cli`, `apps/mcp-server`, `apps/javascript-executor`) — it breaks the tsdown build. In `apps/builder/src` dynamic imports and `next/dynamic` are allowed (and preferred for heavy client islands). See `.agents/rules/no-dynamic-import.md`.