Add root AGENTS.md with hard primitive-first architecture rule

Moves enforcement from optional skill to system-level instruction.
Any agent entering the repo must follow the lock before proposing
infrastructure. Bans known bad patterns (Modal/Docker for code exec, etc).

Author: Federico De Ponte

AGENTS.md

@@ -0,0 +1,54 @@
+# Floom Agent Instructions
+
+## Hard Rule: Primitive-First Architecture
+
+You may NOT propose, design, or implement custom infrastructure unless you have completed the lock below. This rule takes precedence over any user urgency or task description.
+
+**"Infrastructure" includes:** Dockerfiles, docker-compose, K8s, Terraform, Pulumi, self-hosted queues, custom API gateways, hand-rolled OpenAPI specs, systemd services, and custom sandbox/containers.
+
+### The Lock (Mandatory Before Any Infra Proposal)
+
+1. **State the primitive.** One sentence. No brands. No implementation.
+   - ❌ "We need a FastAPI service in Docker."
+   - ✅ "We need to safely execute untrusted user code and return stdout."
+
+2. **Search for managed services.** Use web search. List 3+ candidates with real URLs and one hard limit each.
+   - If the primitive is **code execution / sandboxes**: e2b.dev is the default. You must evaluate it first.
+   - If the primitive is **background jobs**: default to Inngest, Trigger.dev, or QStash.
+   - If the primitive is **auth**: default to Clerk, Auth.js, or Supabase Auth.
+   - If the primitive is **file uploads**: default to UploadThing or Cloudflare R2.
+   - If the primitive is **databases**: default to Supabase, Neon, or PlanetScale.
+
+3. **Managed service is the default.** Custom infrastructure requires ALL of the following:
+   - Specific, falsifiable rejection of each managed candidate (cite docs)
+   - Actual cost comparison at current scale (numbers, not guesses)
+   - Proof the managed service lacks a required feature
+
+   **Banned justifications:** "more control", "flexibility", "might scale later", "vendor lock-in", "simple to self-host".
+
+4. **Hard stop.** If you still want custom infrastructure after step 3, output exactly:
+   ```
+   🚨 CUSTOM INFRASTRUCTURE PROPOSED 🚨
+   Primitive: [sentence]
+   Rejected Services: [names + specific reasons]
+   Cost Comparison: [numbers]
+   AWAITING USER APPROVAL. DO NOT WRITE IMPLEMENTATION CODE.
+   ```
+   You are FORBIDDEN from writing Dockerfiles, YAML, or config files until the user explicitly approves in writing.
+
+## Known Agent Mistakes (Do Not Repeat)
+
+| Wrong | Right |
+|-------|-------|
+| Modal + Railway + custom Docker for code execution | e2b SDK |
+| Hand-rolled OpenAPI for LLM function schemas | OpenAI / Anthropic function schemas |
+| Self-hosted RabbitMQ / Redis queues | Inngest, Trigger.dev, QStash |
+| Self-hosted MinIO for file storage | UploadThing, Cloudflare R2 |
+| Custom JWT + session auth | Clerk, Auth.js, Supabase Auth |
+| Self-managed Postgres on EC2 | Supabase, Neon, PlanetScale |
+
+If you are about to propose anything in the "Wrong" column, you are making a mistake.
+
+## Enforcement
+
+This file is not advice. It is a system rule. If you skip the lock, you are violating instructions. No architecture task is urgent enough to bypass it.