docs: DevOps documentation — complete docs/devops/ set

Adds the full devops-documentation OpenSpec change implementation.
Separate from docs/developers/ — serves a different audience (operators,
not API consumers).

docs/devops/:
- README.md          — index and system overview
- architecture.md    — components, ports, data flow, Redis key patterns
- environment-variables.md — all 7 env vars (required + optional, formats, .env example)
- database.md        — 4-table schema, indexes, constraints, migration runner
- local-development.md — docker-compose setup, health checks, startup, Dockerfile gap noted
- security.md        — RSA key generation/rotation, CORS, bcrypt, secret storage guidance
- operations.md      — startup order, graceful shutdown, log reference, troubleshooting

QA gates: 48/48 tasks complete. All env vars verified against source.
All table names verified against migrations. All ports verified against
docker-compose.yml. All internal links resolve.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
SentryAgent.ai Developer
2026-03-28 14:28:55 +00:00
parent 61ea975c79
commit d94a8cedc0
15 changed files with 1353 additions and 0 deletions

154
docs/devops/security.md Normal file
View File

@@ -0,0 +1,154 @@
# Security
Security configuration for AgentIdP — JWT key management, CORS, and secret storage.
---
## JWT Key Management
AgentIdP uses RS256 (RSA + SHA-256) to sign and verify JWT access tokens. This asymmetric scheme means:
- The **private key** signs tokens — must be kept secret, known only to the server
- The **public key** verifies tokens — can be shared with any system that needs to validate tokens
### Generate a keypair
Generate a 2048-bit RSA keypair:
```bash
# Generate private key
openssl genrsa -out private.pem 2048
# Extract public key
openssl rsa -in private.pem -pubout -out public.pem
```
Verify the files:
```bash
# Confirm private key is valid RSA
openssl rsa -in private.pem -check -noout
# Expected: RSA key ok
# Confirm public key is readable
openssl rsa -in public.pem -pubin -noout -text | head -5
```
### Load keys into environment
**Option 1 — Inline in `.env` (development only)**
Encode newlines as `\n` and wrap in double quotes:
```bash
echo "JWT_PRIVATE_KEY=\"$(awk 'NF {sub(/\r/, ""); printf "%s\\n",$0;}' private.pem)\"" >> .env
echo "JWT_PUBLIC_KEY=\"$(awk 'NF {sub(/\r/, ""); printf "%s\\n",$0;}' public.pem)\"" >> .env
```
**Option 2 — Load from file at runtime (recommended for production)**
In the startup script, read the key files and export as environment variables before running the server:
```bash
export JWT_PRIVATE_KEY="$(cat /run/secrets/jwt-private.pem)"
export JWT_PUBLIC_KEY="$(cat /run/secrets/jwt-public.pem)"
npm start
```
With Docker secrets or a secrets manager (Vault, AWS Secrets Manager), mount the key as a file and read it this way.
### Key rotation
Rotating the JWT keys invalidates all currently active tokens — every authenticated request will fail until clients re-authenticate. Plan rotation for low-traffic windows.
**Rotation procedure:**
1. Generate a new RSA keypair:
```bash
openssl genrsa -out private-new.pem 2048
openssl rsa -in private-new.pem -pubout -out public-new.pem
```
2. Update `JWT_PRIVATE_KEY` and `JWT_PUBLIC_KEY` in your environment or secrets store.
3. Restart the application:
```bash
# Graceful restart — send SIGTERM, let in-flight requests complete, then start with new keys
kill -SIGTERM <pid>
npm start # or docker restart <container>
```
4. All previously issued tokens are now invalid (wrong signature). Clients will receive `401 UNAUTHORIZED` and must call `POST /token` again with their `client_id` and `client_secret` to get a new token.
5. Remove the old key files:
```bash
rm private-old.pem public-old.pem
```
**Important:** There is no grace period or dual-key support in Phase 1. All tokens issued with the old private key are immediately rejected after rotation. If zero-downtime key rotation is required, it is a Phase 2 feature.
---
## CORS Configuration
Cross-Origin Resource Sharing is configured via the `CORS_ORIGIN` environment variable.
| Value | Behaviour |
|-------|-----------|
| `*` (default) | All origins permitted — appropriate for a public API |
| `https://app.example.ai` | Only the specified origin permitted |
Set in `.env`:
```
CORS_ORIGIN=https://app.example.ai
```
The CORS header is set by the `cors` middleware applied globally in `src/app.ts`. Credentials (cookies) are not used — all auth is Bearer token.
For production deployments where the API is only called server-to-server (agent to AgentIdP), setting `CORS_ORIGIN` to a specific origin or removing browser-facing CORS entirely is recommended.
---
## Client Secret Storage
Client secrets are **never stored in plaintext**. The flow:
1. On credential generation or rotation, AgentIdP generates a random secret string (`sk_live_...`)
2. The plaintext is returned to the caller **once only** in the API response
3. AgentIdP immediately hashes the secret with **bcrypt** (cost factor from `bcryptjs` defaults) and stores only the hash in the `credentials.secret_hash` column
4. On every `POST /token` call, the provided `client_secret` is verified against the stored hash using `bcrypt.compare()`
**Implication:** If a client loses their `client_secret`, it cannot be recovered. They must rotate the credential to get a new one.
---
## Secret Storage Guidance
| Environment | Recommendation |
|-------------|---------------|
| Local development | `.env` file, not committed to git |
| CI/CD | Environment variables injected by the CI platform (GitHub Actions secrets, GitLab CI variables, etc.) |
| Production (Docker) | Docker secrets or bind-mounted files from a secrets manager |
| Production (cloud) | AWS Secrets Manager, GCP Secret Manager, HashiCorp Vault (Phase 2) |
**Never:**
- Commit `.env` to version control
- Log environment variables
- Pass secrets as command-line arguments (visible in `ps aux`)
- Store keys in the database
Add `.env` to `.gitignore`:
```bash
echo ".env" >> .gitignore
echo "*.pem" >> .gitignore
```
---
## Token Lifetime
JWT access tokens expire after **3600 seconds (1 hour)**. This is hardcoded in `src/utils/jwt.ts`. There is no refresh token — clients must re-authenticate via `POST /token` when the token expires.
The 1-hour lifetime is a balance between security (short-lived tokens limit exposure if stolen) and operational load (clients don't need to authenticate every few minutes).