AGENTS.md — Rules for AI Agents Working in This Repo

Core Principles

Zod schemas are the contract. Every content file, API payload, and LLM output is validated by Zod. Never write data that hasn't passed schema validation.
No direct LLM calls in request handlers. All LLM work runs in the queue consumer. Handlers enqueue and return immediately (202 Accepted).
Canonical JSON everywhere. All files under content/, config/, and site-context.json must use stringifyCanonical() — sorted keys, 2-space indent, trailing newline.
No git in the orchestrator. Content persistence is filesystem-only via writeContentFile. Deployment of images happens outside this codebase.
Human confirmation before any write. The propose → summary → YES/NO → apply pipeline must be followed for all edit channels (SMS, HTTP, editor).
SQLite, not Redis. Idempotency, proposals, rate limits, and audit logs all live in a single SQLite file. No BullMQ, no Redis.

Do NOT call fs.writeFileSync directly on content files. Always use writeContentFile.
Do NOT skip Zod validation before writing.
Do NOT put LLM calls inside Express route handlers.
Do NOT add git operations (commit, push, etc.) anywhere in the orchestrator.
Do NOT use console.log — use the structured logger from server/src/logger.ts.
Do NOT expose raw JSON or LLM output in SMS replies — use SMS_TEMPLATES.