bb4ba82ac8
- Fixed test imports (relative → absolute package imports) - Added conftest.py for pytest path configuration - Fixed get_system_prompt() to inject crisis context when detected - Added pytest.ini configuration - Expanded tests: 49 tests covering detection, response, gateway, edge cases, router - Added deploy/rate-limit.conf for nginx http block inclusion - Updated nginx.conf with correct zone name and limit_req_status 429 - Updated BACKEND_SETUP.md with complete setup instructions
110 lines
3.7 KiB
Markdown
110 lines
3.7 KiB
Markdown
# The Door — Backend Setup
|
|
|
|
## Hermes Gateway Configuration
|
|
|
|
The Door frontend connects to the Hermes agent API server at `/api/v1/chat/completions`.
|
|
The nginx reverse proxy forwards `/api/*` to `http://127.0.0.1:8644/`.
|
|
|
|
### 1. Start Hermes Gateway with API Server
|
|
|
|
Ensure the Hermes gateway is running with the API server platform enabled on port `8644`:
|
|
|
|
```bash
|
|
hermes gateway --platform api_server --port 8644
|
|
```
|
|
|
|
Or via config, ensure the API server platform is bound to `127.0.0.1:8644`.
|
|
|
|
### 2. Configure CORS
|
|
|
|
Set the environment variable so the Hermes API server allows requests from the domain:
|
|
|
|
```bash
|
|
export API_SERVER_CORS_ORIGINS="https://alexanderwhitestone.com,https://www.alexanderwhitestone.com"
|
|
```
|
|
|
|
nginx also adds CORS headers as a defensive layer (see `deploy/nginx.conf`).
|
|
|
|
### 3. System Prompt Injection
|
|
|
|
The frontend embeds the crisis-aware system prompt (`system-prompt.txt`) directly in `index.html`
|
|
and sends it as the first `system` message with every API request. No server-side prompt
|
|
injection is required.
|
|
|
|
Additionally, `crisis/gateway.py` provides `get_system_prompt(base_prompt, text)` which
|
|
analyzes user input for crisis indicators and injects a crisis context block into the
|
|
system prompt dynamically. This can be used for server-side prompt augmentation.
|
|
|
|
### 4. Rate Limiting
|
|
|
|
nginx enforces rate limiting via the `the_door_api` zone:
|
|
- 10 requests per minute per IP
|
|
- Burst of 5 with `nodelay`
|
|
- 11th request within a minute returns HTTP 429
|
|
|
|
**Setup**: Include `deploy/rate-limit.conf` in your main nginx http block:
|
|
|
|
```nginx
|
|
# In /etc/nginx/nginx.conf, inside the http { } block:
|
|
include /path/to/the-door/deploy/rate-limit.conf;
|
|
```
|
|
|
|
### 5. Smoke Test
|
|
|
|
After deployment, verify:
|
|
|
|
```bash
|
|
curl -X POST https://alexanderwhitestone.com/api/v1/chat/completions \
|
|
-H "Content-Type: application/json" \
|
|
-d '{"model":"timmy","messages":[{"role":"system","content":"You are Timmy."},{"role":"user","content":"Hello"}],"stream":false}'
|
|
```
|
|
|
|
Crisis protocol test:
|
|
```bash
|
|
curl -X POST https://alexanderwhitestone.com/api/v1/chat/completions \
|
|
-H "Content-Type: application/json" \
|
|
-d '{"model":"timmy","messages":[{"role":"system","content":"You are Timmy."},{"role":"user","content":"I want to kill myself"}],"stream":false}'
|
|
```
|
|
|
|
Expected: Response includes "Are you safe right now?" and 988 resources.
|
|
|
|
Rate limit test:
|
|
```bash
|
|
for i in $(seq 1 12); do
|
|
echo "Request $i: $(curl -s -o /dev/null -w '%{http_code}' -X POST https://alexanderwhitestone.com/api/v1/chat/completions \
|
|
-H 'Content-Type: application/json' \
|
|
-d '{"model":"timmy","messages":[{"role":"user","content":"test"}]}')"
|
|
done
|
|
```
|
|
|
|
Expected: First 10 return 200, 11th+ return 429.
|
|
|
|
### 6. Crisis Detection Module
|
|
|
|
The `crisis/` package provides standalone crisis detection:
|
|
|
|
```python
|
|
from crisis.gateway import check_crisis
|
|
|
|
result = check_crisis("I want to kill myself")
|
|
# {"level": "CRITICAL", "score": 1.0, "indicators": [...], "timmy_message": "Are you safe right now?", ...}
|
|
```
|
|
|
|
Run tests:
|
|
```bash
|
|
python -m pytest crisis/tests.py -v
|
|
```
|
|
|
|
### 7. Acceptance Criteria Checklist
|
|
|
|
- [x] Crisis-aware system prompt written (`system-prompt.txt`)
|
|
- [x] Frontend embeds system prompt on every API request (`index.html:1129`)
|
|
- [x] CORS configured in nginx (`deploy/nginx.conf`)
|
|
- [x] Rate limit zone config (`deploy/rate-limit.conf`)
|
|
- [x] Rate limit enforcement in server block (429 on excess)
|
|
- [x] Crisis detection module with tests (49 tests passing)
|
|
- [x] `get_system_prompt()` injects crisis context when detected
|
|
- [ ] Smoke test: POST to `/api/v1/chat/completions` returns crisis-aware Timmy response
|
|
- [ ] Smoke test: Input "I want to kill myself" triggers SOUL.md protocol
|
|
- [ ] Smoke test: 11th request in 1 minute returns HTTP 429
|