docs/services/secrets-manager.md

# Secrets Manager — Design & Integration

## Problem statement

Secrets are currently scattered across:

| Location | Count | Risk |
|---|---|---|
| `/opt/stacks/ai/.env` on CT 104 | ~30 keys | Readable by any root-equivalent process; not versioned |
| Coder `main.tf` (hardcoded env vars) | 6 | Committed to Gitea; visible in template history |
| Zoraxy proxy configs | 0 (plain routes) | Low |
| HA `secrets.yaml` on HAOS VM 100 | unknown | Not backed up centrally |
| `/root/garage-chat-artifacts.creds` | 1 | `0600`, but on Proxmox host root |
| Browser/Vaultwarden (manual) | Many | Human-only access |

Goal: single secret store that agents, Docker services, and Coder workspaces can all read programmatically, with OIDC-gated access via Pocket-ID.

---

## Candidate solutions

### 1. Infisical (recommended)

Open-source HashiCorp Vault alternative. Docker-compose deployable. Native integrations for:

- Docker secrets injection (Infisical agent sidecar)
- Kubernetes/Coder workspace env injection
- REST API + OIDC machine identities
- OIDC SSO (can federate with Pocket-ID)
- Web UI

**Deployment**: separate LXC (recommended — see rationale below). Postgres backend → candidate for CT 113 consolidation once second NVMe lands.

**Port**: 8080 internally; no external exposure needed (LAN-only MCP sidecar pattern).

---

### 2. HashiCorp Vault (OSS)

Industry standard. Steeper ops overhead (unsealing, audit logs, lease renewal). Overkill for a homelab unless you need HSM-grade guarantees.

---

### 3. Doppler (SaaS)

Managed; free tier; native CLI and Docker integration. Outbound dependency; secrets leave the homelab. **Not suitable** given the DLR/LUMEN data handling doctrine.

---

### 4. Stay with Vaultwarden + manual OIDC

Already deployed. Works for human access. No programmatic injection without writing custom code. Dead end for agent/pipeline automation.

---

## Recommendation: Infisical on a dedicated LXC

### Why a separate LXC?

1. **Blast radius isolation** — if the CT 104 Docker stack is compromised, the secret store is not on the same attack surface.
2. **Minimal footprint** — Infisical + Postgres is the only workload; nothing else can mess with the process space.
3. **Simpler audit** — outbound connections from CT 104 are numerous; a dedicated secrets LXC should have near-zero egress.
4. **HA restart independence** — CT 104 restarts (image gen, GPU passthrough experiments) don't affect secret availability.

Suggested: **CT 112** (next available), 2 vCPU / 2 GB RAM, 8 GB disk on local-zfs.

---

## Architecture

```mermaid
flowchart TD
  subgraph CT112["CT 112 — secrets"]
    infisical["Infisical Server\n:8080"]
    pg_sec["Postgres\n(infisical DB)"]
    infisical --- pg_sec
  end

  subgraph CT104["CT 104 — Docker host"]
    agent["Infisical Agent\n(sidecar per stack)"]
    env_file[".env (templated)\nrendered at startup"]
    agent -->|pull on start| infisical
    agent --> env_file
  end

  subgraph CT111["CT 111 — Coder"]
    coder["Coder server"]
    ws["Workspace containers\n(env injected at provision)"]
    coder -->|agent token| infisical
    coder --> ws
  end

  subgraph CT110["CT 110 — Pocket-ID"]
    oidc["OIDC IdP"]
  end

  oidc -->|machine identity| infisical
  oidc -->|user SSO| infisical
```

---

## Setup plan

### Phase 1 — Deploy Infisical on CT 112

```bash
# On Proxmox host
pvesh create /nodes/nuc/lxc \
  --ostemplate local:vztmpl/debian-12-standard_12.7-1_amd64.tar.zst \
  --vmid 112 --hostname secrets --memory 2048 --cores 2 \
  --rootfs local-zfs:8 --net0 name=eth0,bridge=vmbr0,ip=dhcp

# Inside CT 112
apt-get install -y docker.io docker-compose-plugin
# Deploy from https://github.com/Infisical/infisical (official compose)
```

Infisical requires a Redis instance alongside Postgres. The official `docker-compose.yml` bundles both.

### Phase 2 — Migrate CT 104 secrets

1. Create an Infisical project `homelab/ct104`
2. Import existing `.env` keys via Infisical CLI: `infisical import --env prod --path /homelab/ct104 < .env`
3. Add Infisical agent to each compose stack as a sidecar that renders `.env` from templates at container start

### Phase 3 — Coder workspace injection

Infisical has a native Coder integration (via machine identity token). Replace hardcoded env vars in `main.tf` with:
```hcl
data "external" "secrets" {
  program = ["infisical", "export", "--env", "prod", "--path", "/coder/workspaces", "--format", "dotenv-export"]
}
```

Or use the Infisical agent installed in the workspace image.

### Phase 4 — OIDC SSO via Pocket-ID

Infisical supports OIDC SSO under Settings → Authentication → OIDC. Register a new client in Pocket-ID with callback `https://secrets.nuclide.systems/api/v1/sso/oidc/callback` (or LAN-only URL).

---

## Prioritisation

| Item | Effort | Value |
|---|---|---|
| Create CT 112, deploy Infisical | 1h | High — removes .env sprawl |
| Migrate CT 104 .env | 30min | High |
| Coder template injection | 1h | Medium — workspaces currently work |
| HA secrets.yaml → Infisical | 30min (HA add-on) | Medium |
| OIDC SSO via Pocket-ID | 30min | Low (convenience) |

---

## Short-term mitigation (before CT 112 is built)

1. Remove secrets from Coder `main.tf` that appear in Gitea history — use Coder template variables or env injection instead.
2. Keep `/opt/stacks/ai/.env` mode `0600`; ensure it is in `.gitignore` for any repo that mounts that directory.
3. Use Vaultwarden as the authoritative human copy; rotate any key that has appeared in Claude Code context (see `audit-claude-code-meta.md`).

---

## Related

- [[postgres-consolidation]] — CT 113 will consolidate Postgres instances including Infisical's DB
- [[feedback_zoraxy_confirm]] — Zoraxy route for `secrets.nuclide.systems` will need operator approval when CT 112 is ready
- `security/audit-claude-code-meta.md` — lists secrets that transited Claude API this session