docs: README updated for Overlord Agent — host-side service, security stack, deploy flow

Adds an AI Assistant section covering:
- Architecture (host-side vs Docker, dedicated overlord-agent user)
- The 12 MCP tools available to the model
- 10-layer security stack (cookie auth, rate limit, audit log, allowed/disallowed tools, settings.json deny, system prompt, SQL parser, RO PG role, systemd hardening, dedicated UID)
- Full file inventory under agent/
- Routing via nginx /api/agent/*
- Cost / quota notes (subscription auth, reactive only)

Plus features-section blurb, agent env vars table (/etc/overlord/agent.env),
and deploy-flow recipes (code-only restart, requirements update, unit
file change, first-time install).
This commit is contained in:
Erik 2026-04-25 22:50:54 +02:00
parent 0633865598
commit 6de4bfe03e

119
README.md
View file

@ -17,6 +17,7 @@ all driven by WebSocket events from the companion [MosswartMassacre](https://git
- [WebSocket Contract](#websocket-contract) - [WebSocket Contract](#websocket-contract)
- [HTTP API Reference](#http-api-reference) - [HTTP API Reference](#http-api-reference)
- [Frontend](#frontend) - [Frontend](#frontend)
- [AI Assistant (Overlord Agent)](#ai-assistant-overlord-agent)
- [Database Schema](#database-schema) - [Database Schema](#database-schema)
- [Operations & Health](#operations--health) - [Operations & Health](#operations--health)
- [Contributing](#contributing) - [Contributing](#contributing)
@ -38,7 +39,7 @@ The frontend is a React + Vite app served at `/` with a live map, draggable wind
│ WebSocket /ws/position (authenticated) │ WebSocket /ws/position (authenticated)
┌────────────────────────────────────────────────────────┐ ┌────────────────────────────────────────────────────────┐
│ dereth-tracker (FastAPI) │ dereth-tracker (FastAPI, Docker)
│ • main.py — WS routing, analytics, broadcasts │ │ • main.py — WS routing, analytics, broadcasts │
│ • idle/death detection → Discord webhook │ │ • idle/death detection → Discord webhook │
│ • combat stats delta/lifetime accumulation │ │ • combat stats delta/lifetime accumulation │
@ -49,18 +50,29 @@ The frontend is a React + Vite app served at `/` with a live map, draggable wind
▼ ▼ ▼ ▼ ▼ ▼
┌──────────┐ ┌──────────────────┐ ┌──────────────────┐ ┌──────────┐ ┌──────────────────┐ ┌──────────────────┐
│ Browsers │ │ inventory-svc │ │ Discord bot │ │ Browsers │ │ inventory-svc │ │ Discord bot │
│ (React) │ │ (FastAPI) │ │ (rare monitor) │ │ (React) │ │ (FastAPI, Docker)│ │ (rare monitor) │
└──────────┘ └────────┬─────────┘ └──────────────────┘ └────┬─────┘ └────────┬─────────┘ └──────────────────┘
│ ▼
│ ┌──────────────┐
│ │ inventory-db │
│ └──────────────┘
│ /api/agent/* (host-side, OUTSIDE Docker)
┌──────────────┐ ┌────────────────────────────────────────┐
│ inventory-db │ │ overlord-agent (FastAPI, systemd) │ ← runs as dedicated unprivileged user
└──────────────┘ │ • shells out to `claude -p ...` │ /var/lib/overlord-agent home,
│ • MCP server: live-state Q&A tools │ strict settings, no /home/erik
└────────────────────────────────────────┘
┌──────────────┐ ┌──────────────┐
│ dereth-db │ ← TimescaleDB (telemetry, spawns, rares, portals) │ dereth-db │ ← TimescaleDB (telemetry, spawns, rares, portals)
└──────────────┘ └──────────────┘
``` ```
All services run via Docker Compose. Most services run via Docker Compose. **`overlord-agent` is host-side**
(systemd) because it shells out to the `claude` CLI which depends on
host-side credentials — see [AI Assistant](#ai-assistant-overlord-agent).
## Features ## Features
@ -86,6 +98,12 @@ All services run via Docker Compose.
- Plugin broadcasts its own vitals and consumes peer vitals - Plugin broadcasts its own vitals and consumes peer vitals
- In-game `DxHud` overlay shows peer health/stamina/mana bars with direction arrows - In-game `DxHud` overlay shows peer health/stamina/mana bars with direction arrows
### AI Assistant
- 🤖 chat window in the dashboard backed by `claude -p` running headless on the server
- Read-only access to live game state via 12 MCP tools (live players, inventory cross-search, combat stats, quests, suitbuilder, read-only SQL, etc.)
- Per-browser persistent session, "New Chat" button, history rehydration on reload
- Hardened: dedicated unprivileged Linux user, systemd lockdown, strict tool whitelist, audit log, rate limit. See [AI Assistant section](#ai-assistant-overlord-agent) for the full security stack.
### Discord Integration ### Discord Integration
- **Rare Monitor Bot** — posts rares (split by common/great) to configured channels - **Rare Monitor Bot** — posts rares (split by common/great) to configured channels
- **Death Alerts** — webhook to `#alerts` when a character's vitae goes from 0 → >0 (rate-limited to one per character per 5 min) - **Death Alerts** — webhook to `#alerts` when a character's vitae goes from 0 → >0 (rate-limited to one per character per 5 min)
@ -149,6 +167,18 @@ All secrets go in `.env`:
| `ACLOG_CHANNEL_ID` | Discord channel ID for the rare bot's status/vortex messages | | `ACLOG_CHANNEL_ID` | Discord channel ID for the rare bot's status/vortex messages |
| `MONITOR_CHARACTER` | Which character's chat the bot monitors | | `MONITOR_CHARACTER` | Which character's chat the bot monitors |
The Overlord Agent has its own env file at `/etc/overlord/agent.env` (root:overlord-agent 0640) so it doesn't share the tracker's secrets:
| Variable | Purpose |
|---|---|
| `SECRET_KEY` | Same value as the tracker — validates browser session cookies |
| `AGENT_DB_DSN` | Read-only connection string `postgresql://overlord_agent_ro:<pw>@127.0.0.1:5432/dereth` |
| `TRACKER_URL` | Loopback to the tracker container (default `http://127.0.0.1:8765`) |
| `AGENT_RATE_MAX` | Per-user rate limit (default 60/hour) |
| `AGENT_RATE_WINDOW_S` | Rate-limit window in seconds (default 3600) |
| `AGENT_AUDIT_LOG` | Path to audit JSONL (default `/var/log/overlord-agent/audit.jsonl`) |
| `CLAUDE_TIMEOUT_S` | Max seconds per `claude -p` invocation (default 240) |
## Deploying Changes ## Deploying Changes
Live backend host: `overlord.snakedesert.se` (SSH user `erik`, key-based auth). Live backend host: `overlord.snakedesert.se` (SSH user `erik`, key-based auth).
@ -187,6 +217,34 @@ ssh erik@overlord.snakedesert.se "cd /home/erik/MosswartOverlord && \
`BUILD_VERSION` is displayed in the sidebar of the live frontend. Format is CalVer: `YYYY.M.D.HHMM-gitshorthash`. `BUILD_VERSION` is displayed in the sidebar of the live frontend. Format is CalVer: `YYYY.M.D.HHMM-gitshorthash`.
### Overlord Agent deploy
Code changes to `agent/` only:
```bash
ssh erik@overlord.snakedesert.se "cd /home/erik/MosswartOverlord && \
git pull --ff-only origin master && \
sudo systemctl restart overlord-agent"
journalctl -u overlord-agent -f # tail logs to verify
```
`agent/requirements.txt` changed (new pip deps):
```bash
ssh erik@overlord.snakedesert.se "cd /home/erik/MosswartOverlord && \
git pull --ff-only origin master && \
agent/.venv/bin/pip install -r agent/requirements.txt && \
sudo systemctl restart overlord-agent"
```
systemd unit changed:
```bash
ssh erik@overlord.snakedesert.se "cd /home/erik/MosswartOverlord && \
git pull --ff-only origin master && \
sudo cp agent/overlord-agent.service /etc/systemd/system/ && \
sudo systemctl daemon-reload && sudo systemctl restart overlord-agent"
```
First-time install: `bash agent/install.sh` — see `agent/README.md` for the full bootstrap procedure (creating the `overlord-agent` user, copying claude auth, granting filesystem access, populating `/etc/overlord/agent.env`).
## WebSocket Contract ## WebSocket Contract
### `/ws/position` (plugin → backend) ### `/ws/position` (plugin → backend)
@ -253,6 +311,53 @@ See `EVENT_FORMATS.json` for event schemas. Major HTTP endpoints:
### Classic v1 (preserved at `/classic`) ### Classic v1 (preserved at `/classic`)
The original vanilla JS frontend with element-pooling optimization is kept for fallback and reference. The original vanilla JS frontend with element-pooling optimization is kept for fallback and reference.
## AI Assistant (Overlord Agent)
A draggable chat window in the dashboard (🤖 Assistant button). Powered by `claude -p` running headless on the server, with read-only access to live game state via an MCP server.
### Architecture
- **Host-side service** (`agent/`, systemd unit `overlord-agent`) runs OUTSIDE Docker because the `claude` CLI binary lives on the host (`/home/erik/.local/bin/claude`) and depends on host-side authentication credentials.
- **Dedicated UNIX user** (`overlord-agent`, system account, `/var/lib/overlord-agent` home, no shell) — kernel-level isolation from the operator's `erik` account. Cannot read `/home/erik/.claude`, `~/.ssh`, `.bash_history`, `.env`, etc.
- **MCP stdio server** (`agent/mcp_overlord.py`) exposes 12 tools that wrap the tracker's HTTP endpoints + read-only DB queries. Claude only sees these tools; no `Bash`, `Read`, `Write`, etc.
- **Frontend** (`AgentWindow.tsx`) — per-browser session UUID in localStorage, "New Chat" button, on-mount rehydration from `/agent/sessions/{id}/history`.
### MCP tools available to the assistant
`get_live_players`, `get_player_state`, `get_combat_stats`, `get_equipment_cantrips`, `get_inventory`, `get_inventory_search`, `search_items` (cross-character), `get_recent_rares`, `get_quest_status`, `get_server_health`, `query_telemetry_db` (read-only SQL via sqlglot parser + GRANT-SELECT-only PG role), `suitbuilder_search`. Plus `WebFetch(domain:acpedia.org)` for AC info lookups.
### Security stack (defense-in-depth)
1. **Cookie auth** on `/agent/ask` (same session cookie the tracker issues)
2. **Per-user rate limit** (60 req/h default) and **concurrency cap** (1 in-flight)
3. **JSONL audit log** at `/var/log/overlord-agent/audit.jsonl` (every prompt + result)
4. **CLI flags**`--allowed-tools` (just our 12 MCP tools), `--disallowed-tools` (Bash, Write, Read, Edit, Agent, ToolSearch, Monitor, scheduling, Gmail/Drive/Calendar, etc.), `--permission-mode dontAsk`
5. **`/var/lib/overlord-agent/.claude/settings.json`** — strict deny rules (server-side only, NOT in repo)
6. **System-prompt scope rules** in `CLAUDE.md` — instruct the model not to probe, not to suggest workarounds
7. **SQL parser** (`sqlglot`) rejects any non-SELECT statement on `query_telemetry_db`
8. **Read-only PG role** `overlord_agent_ro` (GRANT SELECT only) — even a parser bypass can't mutate
9. **systemd hardening**`ProtectSystem=strict`, `ProtectHome=read-only`, `InaccessiblePaths=/etc/shadow,/root,~/.ssh,…`, `NoNewPrivileges=true`, `CapabilityBoundingSet=` (empty), `PrivateTmp=true`, `PrivateDevices=true`, `RestrictAddressFamilies=AF_UNIX AF_INET AF_INET6`, `SystemCallFilter=@system-service ~@privileged ~@reboot ~@mount`, `MemoryMax=512M`, `TasksMax=128`
10. **Secrets out of /home**`/etc/overlord/agent.env` (root:overlord-agent 0640) for SECRET_KEY + AGENT_DB_DSN
### Files
| Path | What |
|------|------|
| `agent/service.py` | FastAPI app: `/agent/health`, `/agent/sessions/new`, `/agent/ask`, `/agent/sessions/{id}/history` |
| `agent/auth.py` | Session cookie validation (mirrors `main.py:1013-1019`) |
| `agent/claude_wrapper.py` | `asyncio.create_subprocess_exec("claude", "-p", …)` with allowed/disallowed-tools |
| `agent/tools.py` | Pure tool implementations |
| `agent/mcp_overlord.py` | MCP stdio server registering tools |
| `agent/sql/0001_overlord_agent_ro.sql` | Read-only PG role |
| `agent/overlord-agent.service` | systemd unit (the hardening directives) |
| `agent/install.sh` | venv + systemd setup |
| `agent/README.md` | Operator's deeper reference |
| `.mcp.json` (repo root) | Project-level MCP config Claude Code auto-loads |
| `CLAUDE.md` "Overlord Assistant Mode" section | System-prompt briefing |
### Routing
nginx forwards `/api/agent/*` to `127.0.0.1:8767` (the host-side service) with a 300s read/send timeout (suitbuilder runs can be slow). Other `/api/*` continues to the dereth-tracker container at `127.0.0.1:8765`.
### Cost / quota
Subscription auth (no API key); per-call cost is informational only. Each `/agent/ask` invocation = one `claude -p` subprocess with shared session cache. Reactive only — no background polling, no scheduled tasks.
## Database Schema ## Database Schema
### Telemetry DB (`dereth`, TimescaleDB) ### Telemetry DB (`dereth`, TimescaleDB)