from __future__ import annotations

from pathlib import Path


BRIEF_FILENAME = "MM1_CORE_IDEAS.md"


def write_concept_brief(base_dir: str | Path = "data") -> Path:
    generated_dir = Path(base_dir) / "generated"
    generated_dir.mkdir(parents=True, exist_ok=True)
    path = generated_dir / BRIEF_FILENAME
    path.write_text(_brief_text(), encoding="utf-8")
    return path


def _brief_text() -> str:
    return """# MM1 Core Ideas

MM1 is a local-first personal agent.

The central idea is a split between two kinds of intelligence:

## 1. Small Private Agent

The small agent lives close to the user.

It can remember personal context such as:

- name
- preferences
- goals
- constraints
- recurring projects
- values

This memory stays local. It is used to understand the person and to make answers feel personal.

## 2. Large Public Agent

The large agent is used only when the task needs more power, internet access, coding, analysis, or tool use.

It does not receive the private memory.

Before a task leaves the private layer, MM1 creates a sanitized version of the request. Personal details are removed or hidden.

## The Boundary

The private agent knows the human.

The public agent knows only the task.

That boundary is the product:

```text
Human
  -> Small private agent with local memory
  -> Sanitized task
  -> Search, hosted model, or Codex
  -> Result
  -> Private personalization back on the local side
```

## Why This Matters

Most assistants become useful by sending more personal context to cloud systems.

MM1 tries a different shape:

- personal context stays local
- outside tools get less private data
- powerful models are still usable
- the user gets an agent that feels personal without making every backend personal

## Current Prototype

This local prototype shows the core behavior:

- it learns from conversation
- it answers local memory questions locally
- it routes web lookups to search
- it routes technical build tasks to a larger backend
- it logs the privacy boundary for engineering inspection

The invention is not that a big model exists in the backend. The invention is the clean split: a small private model that knows the person, and a larger public model that only sees sanitized work.
"""