# marvin-turnstone

Local turnstone deployment as a **container-VM**: a podman image is turned
into a bootable VM with [boco](https://git.nubo.sh/hulthe/boco) (custom
init + ext4 qcow2 + QEMU direct kernel boot). Replaces the older
opencode-in-a-container setup (`~/marvin`).

```
host podman build ──> boco build qcow2 ──> boco boot
                                              │  QEMU/KVM (virtio, user-net)
                        ┌─────────────────────┴───────────────────┐
                        │ VM (Arch)                               │
                        │  /boco/init → start-turnstone.sh (root) │
                        │    ├─ sshd            (ops channel)     │
                        │    └─ setpriv marvin                    │
                        │        └─ turnstone-server :8080  PID 1 │
                        └─────────────────────────────────────────┘
   host :8080 ──hostfwd──> web UI        host :2222 ──> guest sshd :22
   guest ──HTTPS (double slirp egress)──> api.tensorx.ai (LLM)
```

## Prerequisites

Host (Arch):
- `podman`, `qemu-base`, `virtiofsd`, `croc` (only if fetching the kernel)
- `boco` CLI on PATH and `boco-init` at `~/.local/share/boco/boco-init`
  (build from the boco repo)
- Kernel at `~/.local/share/boco/registry/vmlinuz` (x86 bzImage with
  `CONFIG_VIRTIO_FS=y`; transfer via croc)

VM image:
- Arch base with python + turnstone (venv at `/opt/turnstone`), git,
  helix, podman (rootless, subuid range), openssh, build toolchain
- `marvin` user; sshd + turnstone-server started by
  `start-turnstone.sh` (server runs as `marvin`, PID 1)

## Build

```sh
scripts/build.sh
```

Creates `configs/ssh/id_ed25519` (ops keypair, once), builds the podman
image, then packs it as `boco` qcow2 into the boco registry.

## Run

```sh
scripts/run.sh
```

- Web UI: <http://localhost:8080> (login `marvin` / `marvin-turnstone`
  unless `ADMIN_PASSWORD` was overridden — change it)
- Ops SSH: `ssh -i configs/ssh/id_ed25519 -p 2222 marvin@localhost`
- QEMU runs foreground; quit with `ctrl-a x`

### Access from another machine (tailcat)

`localhost` above is the machine running QEMU. To reach the VM from
elsewhere, run `scripts/expose.sh` next to it — a tailcat tunnel
(WireGuard over Tailscale's DERP relays, outbound-only, no account):

```
host $ scripts/expose.sh
# 🐈 Server listening with saved key "default": tcpGFwWCCH...

your $ tailcat forward tcpGFwWCCH... 8080:8080 2222:2222
```

Then browse `http://localhost:8080` / `ssh -p 2222 marvin@localhost`
from *your* machine. The address is stable across VM restarts (saved
key) and is a bearer secret — share it only privately.

### LLM backend

`[models.tensorx]` in `configs/turnstone/config.toml` points at the
TensorX OpenAI-compatible gateway (`https://api.tensorx.ai/v1`, same
key this machine's opencode uses — see
`~/.local/share/opencode/auth.json`). Guest egress to HTTPS endpoints
works through both slirp layers; no relay needed.

Self-hosted provider later: add a `[models.*]` entry with any
OpenAI-compatible endpoint reachable from the guest (see
`config.toml.example`).

## Shares (virtiofs, survives reboots)

| Host path              | Guest path                     | Purpose                  |
|------------------------|--------------------------------|--------------------------|
| `configs/turnstone/`   | `/home/marvin/.config/turnstone` | turnstone config.toml  |
| `data/`                | `/var/lib/turnstone`           | SQLite database          |
| `work/`                | `/home/marvin/work`            | agent workspace          |

Guest-written files appear on the host owned by subuid-range UIDs
(virtiofsd runs under `podman unshare`) — cosmetic, and stable across
boots.

## Configuration

Edit `configs/turnstone/config.toml` (gitignored; see
`config.toml.example`):

- `[models.tensorx]` — TensorX gateway; `api_key` must be set
  (currently mirrored from opencode's `auth.json`).
- `[auth] jwt_secret` — required, generate once.
- `[database] path` — SQLite on the `data/` share.
- `[tools]` — `workspace_dir` surfaced to the model
  (`/home/marvin/work`, mounted from `./work`).

### Personas

Built-ins (Engineer, Writer, Researcher, Scribe, Manager, Executive)
are seeded by turnstone's migration. This deployment adds a custom
default persona `marvin` (VM-specific: persistence rules, Gitea remotes,
podman) via `scripts/persona.py` — the prompt text is versioned there
and re-applied after a DB reset:

```sh
python3 scripts/persona.py            # --db data/turnstone.db default
```

(Without a console deployment, persona CRUD is a direct SQLite write;
with `turnstone-console` running, use its admin API/Personas tab.)

### Skills

The 13 skills from the old opencode setup (`~/marvin/configs/opencode/skills`,
copied into this repo under `skills/`) are imported into turnstone's
DB-backed catalog via `scripts/import_skills.py`. The script runs
INSIDE the VM (uses turnstone's storage layer + SKILL.md parser +
risk scanner):

```sh
scp -O -i configs/ssh/id_ed25519 -P 2222 -r skills scripts/import_skills.py \
    marvin@localhost:/tmp/
ssh -i configs/ssh/id_ed25519 -p 2222 marvin@localhost \
    /opt/turnstone/bin/python /tmp/import_skills.py --dir /tmp/skills
```

Idempotent (`--update` re-imports + appends version rows). **Reboot the
VM after importing** — the running server holds a stale WAL snapshot
and won't see new rows until restart.

## Notes / trade-offs

- boco boots with `-snapshot`: **only the virtiofs shares persist.**
  Agent-installed packages, git clones outside `work/`, and podman image
  storage inside the VM are discarded on shutdown.
- boco hardcodes `-smp 8`; memory via `VM_MEMORY=...M scripts/run.sh`.
- The old setup's Traefik is dropped — hostfwd replaces it.