Shared memory and channels for multi-agent AI

# mistral_engineer.py — run this first in a terminal
import os
from mistralai import Mistral
from contextrelay import ContextRelay

relay   = ContextRelay(api_key=os.environ["CONTEXTRELAY_API_KEY"])
mistral = Mistral(api_key=os.environ["MISTRAL_API_KEY"])

print("Mistral engineer ready — subscribed to 'task-assigned'")

def on_task(url: str):
    print(f"  [task-assigned] architecture received")
    architecture = relay.pull(url)
    print(f"  [task-assigned] {len(architecture):,} chars — implementing...")

    code = mistral.chat.complete(
        model="mistral-large-latest",
        messages=[{
            "role": "user",
            "content": (
                "You are a senior engineer. Implement this architecture as "
                "complete, runnable code. Every file. No placeholders.\n\n"
                + architecture
            ),
        }],
    ).choices[0].message.content

    relay.push(code, channel="task-done", metadata={"role": "implementation"})
    print("  [task-done] implementation published — architect notified")

relay.subscribe("task-assigned", on_task)  # blocks, waits for Claude

You are a senior software architect in an autonomous pipeline.
Your engineer (Mistral) is already subscribed to channel "task-assigned".
The moment you push your architecture there, Mistral implements it — no human step.

Task: Design a production-ready FastAPI task management API.
Cover: User + Task models, all CRUD endpoints, JWT auth,
SQLAlchemy + SQLite, Pydantic schemas, file structure.

Steps:
1. Write the complete architecture document
2. Use push_context — set channel="task-assigned"
   Mistral's subscriber fires automatically on push
3. Subscribe / poll "task-done" to receive the implementation back

Start now. The pipeline is fully autonomous.