docs/chat.md

# DeepSeek V4 Chat — Module Documentation

Status: **integrated** (v1.0.0, 2026-08-09) · Registry: `modules/chat-module.yaml`

## 1. Overview

The chat module integrates the DeepSeek-V4-Sovereign model into the
Sovereign stack as a conversational core: session management (bounded
history), sovereign-context system prompt, OpenAI-compatible API client,
and a deterministic local-echo fallback so the module runs — and is
testable — with zero external dependencies.

## 2. Architecture

```
chat/
├── core.py            ChatMessage · ChatSession · DeepSeekV4Client
└── api/
    ├── routes.py      /api/v1/chat/* (5 endpoints)
    └── app.py         standalone app (port 8003)
site/chat.html         chat web UI (self-contained, responsive)
```

Wired into `src/sovereign/main.py` with a guarded include (same pattern
as matrix/kyrexis/pec5d/coldcase).

## 3. Behavior

- **API mode** — `DEEPSEEK_API_KEY` set: OpenAI-compatible
  `/chat/completions` call; errors surface as 502.
- **Local mode** — no key: deterministic sovereign echo, no network.
- **Sessions** — in-memory, bounded to `max_history` (50), system prompt
  always retained.

## 4. Configuration

`config/chat.yaml` — model id/base/temperature/max_tokens, session bounds,
system prompt. Keys via env only.

## 5. Verification

- `tests/test_chat.py` — 7 tests (defaults, sessions, bounding, local
  echo, env config, model list, API surface).
- Registry validator covers the `chat` module entry.

## 6. Production notes

- In production set `DEEPSEEK_API_KEY` (or a gateway URL) and serve behind
  TLS; the UI (`site/chat.html`) talks to `/api/v1/chat` same-origin.
- The local echo is a smoke path, not a production model answerer.