# boxlite serve — REST API Server
`boxlite serve` starts a long-running HTTP server that exposes the full BoxLite runtime over a REST
API. SDK clients connect to it via `Boxlite.rest()` — the server holds a single `BoxliteRuntime` and
translates HTTP requests into runtime calls, WebSocket connections into bidirectional attach
sessions.
## Quick Start
```bash
boxlite serve # listen on 0.0.0.0:8100 (permissive)
boxlite serve --port 9090 # custom port
boxlite serve --host 127.0.0.1 # bind localhost only
boxlite serve --api-key dev-key # require Bearer dev-key (else 401)
```
Ctrl-C triggers graceful shutdown (`runtime.shutdown` with a 10 s timeout).
Connect from an SDK:
```rust
let rt = BoxliteRuntime::rest(BoxliteRestOptions::new("http://localhost:8100"))?;
```
## Startup Call Graph
```
execute(ServeArgs, GlobalFlags)
│
├─ GlobalFlags::resolve_runtime_options() — resolve local runtime settings
├─ GlobalFlags::create_runtime_with_options()
│ — build local BoxliteRuntime
│
├─ Arc::new(AppState { runtime, boxes, executions,
│ api_key, lifecycle, last_activity })
│ — shared state for all handlers
├─ tokio::spawn(reaper_loop(state)) — orphan reaper + lifecycle sweep
│ (30 s tick)
│
├─ build_router(state) — register 26 routes on axum::Router
│
├─ TcpListener::bind(host:port) — bind TCP socket
│
└─ axum::serve(listener, app) — start serving requests
└─ .with_graceful_shutdown(ctrl_c)
└─ runtime.shutdown(Some(10)) — stop all boxes on exit
```
## Architecture
```
┌───────────────────────────────────────────────────────┐
│ SDK Clients │
│ (Boxlite.rest() → ApiClient → reqwest / WebSocket) │
└────────────────────────┬──────────────────────────────┘
│ HTTP / WebSocket
▼
┌───────────────────────────────────────────────────────┐
│ boxlite serve (axum) │
│ │
│ ┌──────────────────────────────────────────────────┐ │
│ │ AppState (Arc) │ │
│ │ │ │
│ │ ┌─────────────────┐ │ │
│ │ │ BoxliteRuntime │ core runtime │ │
│ │ └─────────────────┘ │ │
│ │ ┌──────────────────────────────────────────┐ │ │
│ │ │ boxes: RwLock<HashMap<String, Arc<LiteBox>>> │ │
│ │ └──────────────────────────────────────────┘ │ │
│ │ ┌──────────────────────────────────────────────┐ │ │
│ │ │ executions: RwLock<HashMap<String, │ │ │
│ │ │ Arc<ActiveExecution>>> │ │ │
│ │ └──────────────────────────────────────────────┘ │ │
│ └──────────────────────────────────────────────────┘ │
│ │
│ Handlers: config · boxes · executions │
│ files · metrics · snapshots · advanced │
│ │
│ Background: reaper_loop (orphans + lifecycle sweep) │
└───────────────────────────────────────────────────────┘
│
▼
┌───────────────────────────────────────────────────────┐
│ BoxliteRuntime (core) │
│ LiteBox → ShimController → Guest VM │
└───────────────────────────────────────────────────────┘
```
`AppState.boxes` is a lazy cache — `get_or_fetch_box()` populates it from `runtime.get()` on first
access. `AppState.executions` maps execution IDs to `Arc<ActiveExecution>` for the active-session
registry. `AppState.lifecycle` holds the AutoStop/AutoDelete deadlines the sweep acts on — the
engine accepts neither on a local runtime — and `AppState.last_activity` is the idle clock, stamped
by the request middleware and by client data frames on an attach.
## Handler Reference
| `config` | `handlers/config.rs` | Capability discovery (snapshots, clone, export, import) |
| `me` | `handlers/me.rs` | Identity of the calling credential (`GET /v1/me`) |
| `boxes` | `handlers/boxes.rs` | Box CRUD and lifecycle |
| `executions` | `handlers/executions.rs` | Lifecycle: start, status, signal, kill, resize, attach |
| `files` | `handlers/files.rs` | Tar-based file upload / download into / from boxes |
| `metrics` | `handlers/metrics.rs` | Runtime-level and per-box metrics |
| `snapshots` | `handlers/snapshots.rs` | Snapshot CRUD + restore |
| `advanced` | `handlers/advanced.rs` | Clone, export, import |
## Route Reference
All paths are relative to the server root (e.g. `http://localhost:8100`).
### Auth & Config
| GET | `/v1/config` | `config::get_config` | Capability discovery (always public) |
| GET | `/v1/me` | `me::get_me` | Identity of the calling credential |
**Auth.** With `--api-key <KEY>` (or `$BOXLITE_SERVE_API_KEY`) set, every
route except `GET /v1/config` requires `Authorization: Bearer <KEY>`
(constant-time match) and returns `401` otherwise. Without it the server is
permissive (accepts any/no bearer) — the zero-config local-dev default.
### Box CRUD & Lifecycle
| POST | `/v1/boxes` | `boxes::create_box` | Create a new box |
| GET | `/v1/boxes` | `boxes::list_boxes` | List all boxes |
| GET | `/v1/boxes/{box_id}` | `boxes::get_box` | Get box info |
| HEAD | `/v1/boxes/{box_id}` | `boxes::head_box` | Check box exists (204 / 404) |
| DELETE | `/v1/boxes/{box_id}` | `boxes::remove_box` | Remove box (`?force=true`) |
| POST | `/v1/boxes/{box_id}/start` | `boxes::start_box` | Start a stopped box |
| POST | `/v1/boxes/{box_id}/stop` | `boxes::stop_box` | Stop a running box |
### Command Execution
| GET | `/v1/boxes/{box_id}/attach` | `executions::attach_box` | WebSocket attach to init |
| POST | `/v1/boxes/{box_id}/exec` | `executions::start_execution` | Start a new command |
| GET | `/v1/boxes/{box_id}/executions/{id}` | `executions::get_execution` | Get status + exit code |
| DELETE | `/v1/boxes/{box_id}/executions/{id}` | `executions::kill_execution` | SIGKILL + evict |
| GET | `/v1/boxes/{box_id}/executions/{id}/attach` | `executions::attach_execution`| WebSocket attach (bidi) |
| POST | `/v1/boxes/{box_id}/executions/{id}/signal` | `executions::send_signal` | Send cooperative signal |
| POST | `/v1/boxes/{box_id}/executions/{id}/resize` | `executions::resize_tty` | Resize PTY |
`/attach` is the box's **main command session** — the container's init, which
`run IMAGE COMMAND` replaces with COMMAND (docker semantics). It mirrors
docker's `POST /containers/{id}/attach`, distinct from exec-attach: the client
has no id for init, because init's execution id is the container id. The first
attach opens that session and registers it alongside tenant execs; the upgrade
response returns its id in `X-Boxlite-Execution-Id`, so signal/resize/kill and
reconnect then go through the `/executions/{id}/…` routes above. A second
client on an attached main session gets 409, same as for an exec.
### Files
| PUT | `/v1/boxes/{box_id}/files` | `files::upload_files` | Upload tar, extract into box |
| GET | `/v1/boxes/{box_id}/files` | `files::download_files`| Download path as tar |
### Metrics
| GET | `/v1/metrics` | `metrics::runtime_metrics`| Runtime-wide counters |
| GET | `/v1/boxes/{box_id}/metrics` | `metrics::box_metrics` | Per-box metrics + boot timing |
Serving per-box metrics boots a box that is not running, and a scrape is not
counted as use — so on a box with an AutoStop deadline it would only be stopped
again on the next tick. `box_metrics` answers 409 rather than start that cycle.
A box with no deadline still boots on scrape, as it always has.
### Snapshots
| POST | `/v1/boxes/{box_id}/snapshots` | `snapshots::create_snapshot` | Create snapshot |
| GET | `/v1/boxes/{box_id}/snapshots` | `snapshots::list_snapshots` | List snapshots |
| GET | `/v1/boxes/{box_id}/snapshots/{name}` | `snapshots::get_snapshot` | Get snapshot info |
| DELETE | `/v1/boxes/{box_id}/snapshots/{name}` | `snapshots::delete_snapshot` | Delete snapshot |
| POST | `/v1/boxes/{box_id}/snapshots/{name}/restore` | `snapshots::restore_snapshot` | Restore snapshot |
### Advanced (Clone, Export, Import)
| POST | `/v1/boxes/{box_id}/clone` | `advanced::clone_box` | Clone a box |
| POST | `/v1/boxes/{box_id}/export` | `advanced::export_box`| Export box as archive |
| POST | `/v1/boxes/import` | `advanced::import_box`| Import box from archive body |
## Execution Lifecycle
### Start
```
POST /v1/boxes/{box_id}/exec
│
├─ get_or_resume_box(state, box_id) — resolve LiteBox; 409 if the box
│ is down and AutoResume is off
├─ build_box_command(req) — JSON body → BoxCommand
├─ litebox.run(cmd) — start command (returns Execution)
│
├─ execution.stdin() — take stdin handle
│ └─ if req.stdin is Some:
│ stdin.write_all(data) + close — write inline stdin, discard handle
│
└─ ActiveExecution::new(box_id, execution, stdin) — wrap in server-side state
│
├─ execution.stdout() / .stderr() — take single-consumer stream halves
│
├─ BacklogBroadcast::new(256, 256 KiB) — stdout_bus + stderr_bus
│
├─ tokio::spawn(stdout pump) — Stream → bus.send()
├─ tokio::spawn(stderr pump) — Stream → bus.send()
│
└─ tokio::spawn(wait task)
├─ execution.wait() — block until process exits
├─ stdout_handle.await — barrier: wait for pump completion
├─ stderr_handle.await — ensures all output is broadcast
├─ done.store(true)
└─ done_tx.send(true) — notify all observers
```
### BacklogBroadcast
`BacklogBroadcast` wraps `tokio::sync::broadcast` with a byte-capped `VecDeque` backlog (256 KiB).
When a new subscriber calls `subscribe()`, it receives the full backlog snapshot first, then
switches to live broadcast — no gap, no interleaving. Both the backlog append and the broadcast
`send()` happen under the same `Mutex`, preventing duplicates. This mirrors the Go runner's
`streamBus` pattern for late-attach replay.
### Attach (WebSocket)
```
GET /v1/boxes/{box_id}/executions/{id}/attach
│
├─ mark_connected() — claim single-attach slot (409 if taken)
├─ WebSocketUpgrade → on_upgrade — HTTP → WS handshake
│ └─ on_failed_upgrade: mark_disconnected()
│
└─ run_attach_session(socket, active, state, stdin_policy)
│
├─ stdout_bus.subscribe() — BacklogReceiver (replay + live)
├─ stderr_bus.subscribe()
├─ done_rx = active.done_rx()
├─ socket.split() → (sink, stream)
│
├─ tokio::spawn(reader) — client → server
│ ├─ Binary frames → Refuse: error frame + Close 1008; else stdin.write_all()
│ ├─ Text {"type":"resize"} → Refuse: error frame; else execution.resize_tty()
│ ├─ Text {"type":"signal"} → Refuse: error frame; else execution.signal()
│ └─ Text {"type":"stdin_eof"} → Refuse: error frame; else stdin.close() + drop
│ (Refuse answers each kind once; the reader keeps draining so the
│ writer can flush its close — whichever task ends first aborts
│ the other.)
│
├─ tokio::spawn(writer) — server → client
│ ├─ if already done → drain backlog only (fast path)
│ ├─ select! loop:
│ │ stdout_rx.recv() → Binary [0x01 | data]
│ │ stderr_rx.recv() → Binary [0x02 | data]
│ │ ctrl_rx.recv() → Text (error frames), or Text + Close 1008
│ │ ping_interval → Ping (every 15 s)
│ │ done_rx.changed()→ drain remaining, break
│ └─ if done: Text {"type":"exit","exit_code":N} + Close
│
└─ select!(reader, writer) — first to finish aborts the other
└─ mark_disconnected()
```
### Attach Wire Format
```
Client WebSocket Server
│ │
│── Binary(raw bytes) ───────────────────────────────────────▶│ stdin
│── Text({"type":"resize","rows":N,"cols":N}) ──────────────▶│ resize_tty
│── Text({"type":"signal","sig":N}) ─────────────────────────▶│ signal
│── Text({"type":"stdin_eof"}) ──────────────────────────────▶│ close stdin
│ │
│◀────── Binary([0x01 | stdout data]) ─────────────────────────│ stdout
│◀────── Binary([0x02 | stderr data]) ─────────────────────────│ stderr
│◀────── Text({"type":"exit","exit_code":N}) ──────────────────│ completion
│◀────── Text({"type":"error","message":"..."}) ───────────────│ error
│◀────── Ping ─────────────────────────────────────────────────│ keepalive
│ │
```
Only one WebSocket client may be attached to an execution at a time. `mark_connected()` returns
`false` (HTTP 409) if another client is already attached or the reaper has claimed a kill. On
reconnect, escalation flags reset — a fresh disconnect starts a fresh reap clock.
## Orphan Reaper
A background task (`tokio::spawn`) that ticks every 30 s, cleaning up executions whose attach
client disconnected and never came back.
### Escalation Timeline
```
Client disconnects
──────┬────────────────────────────────────────────────────────────
│
├── 5 min (BOXLITE_RECONNECT_GRACE) ───────▶ SIGHUP
│ │
├── + 30 s (BOXLITE_SHUTDOWN_GRACE) ─────────▶ SIGTERM
│ │
├── + 30 s (BOXLITE_SHUTDOWN_GRACE) ─────────▶ SIGKILL + evict
│
──────┴────────────────────────────────────────────────────────────
Hard cap: 24 h (BOXLITE_MAX_SESSION_LIFETIME) → immediate SIGKILL
Completed: 5 min after exit → evict from map
```
### Reaper Call Graph
```
reaper_loop(state)
│
├─ resolve_duration("BOXLITE_RECONNECT_GRACE", 300 s)
├─ resolve_duration("BOXLITE_SHUTDOWN_GRACE", 30 s)
├─ resolve_duration("BOXLITE_MAX_SESSION_LIFETIME", 86400 s)
│
└─ loop { ticker.tick(); run_reap_once(...); run_lifecycle_once(...) }
run_reap_once(state, now, ...)
│
├─ snapshot candidates from executions map (read lock, then drop)
│
└─ for each (id, active):
│
├─ is_done() + !should_retain(now) → evict from map
│
├─ lifetime > max_lifetime → mark_reaping_kill + try_kill_and_evict
│
├─ already reaping_kill → try_kill_and_evict (retry from prior tick)
│
├─ try_escalate_hup() — SIGHUP if reconnect_grace expired
│ ├─ signal OK → finish_escalation()
│ └─ signal fail → escalation_failed_mark_doomed + try_kill_and_evict
│
├─ try_escalate_term() — SIGTERM if shutdown_grace expired after HUP
│ ├─ signal OK → finish_escalation()
│ └─ signal fail → escalation_failed_mark_doomed + try_kill_and_evict
│
└─ try_escalate_kill() — SIGKILL if shutdown_grace expired after TERM
└─ try_kill_and_evict
try_kill_and_evict(state, id, active)
├─ timeout(10 s, execution.kill())
├─ success → executions.write().remove(id)
└─ failure → leave in map (reaping_kill=true), retry next tick
```
### AttachState Machine
```
mark_connected() mark_disconnected()
┌──────────┐ ──────────────────▶ ┌────────────┐ ──────────────────────┐
│ Idle │ │ Connected │ │
│ (disc.) │ ◀──────────────────────────────────────────────────────────┘
└──────────┘
│
│ reconnect_grace expired
▼
┌──────────┐ shutdown_grace ┌────────────┐ shutdown_grace
│ HUP'd │ ────────────────▶ │ TERM'd │ ────────────────┐
└──────────┘ └────────────┘ │
▼
┌────────────┐
│ KILL'd │
│ (doomed) │
└────────────┘
```
TOCTOU safety: the `escalating` flag blocks `mark_connected()` during signal delivery so a late
attach cannot race the kill. `escalation_failed_mark_doomed()` atomically clears `escalating` and
sets `reaping_kill` under one lock acquisition.
## Configuration Reference
### Environment Variables
| `BOXLITE_RECONNECT_GRACE` | `300s` | `Ns/Nm/Nh`| Grace period before SIGHUP after disconnect |
| `BOXLITE_SHUTDOWN_GRACE` | `30s` | `Ns/Nm/Nh`| Grace between SIGHUP→SIGTERM and SIGTERM→SIGKILL |
| `BOXLITE_MAX_SESSION_LIFETIME` | `86400s` | `Ns/Nm/Nh`| Hard cap: SIGKILL after this lifetime |
### CLI Flags
| `--port` | `8100` | TCP port to listen on |
| `--host` | `0.0.0.0` | Address to bind |
### Internal Constants
| `REAPER_TICK` | 30 s | Reaper check interval |
| `REAPER_SIGNAL_TIMEOUT` | 10 s | Timeout for signal / kill delivery |
| `COMPLETED_RETENTION_GRACE` | 5 min | How long completed runs stay in map |
| `ATTACH_BROADCAST_CAPACITY` | 256 | Broadcast channel buffer (chunks) |
| `BACKLOG_BYTE_CAP` | 256 KiB | Max backlog retained for late replay |
| `ATTACH_KEEPALIVE_INTERVAL` | 15 s | WebSocket ping interval |
| `ATTACH_WRITE_TIMEOUT` | 20 s | WebSocket write timeout |
| `ALLOWED_SIGNALS` | 1, 2, 3, 6, 10, 12, 15, 28 | Cooperative signal whitelist |
## Error Handling
All handlers use `error_response(StatusCode, message, error_type)` to return a consistent JSON
shape:
```json
{
"error": {
"message": "box not found: abc123",
"type": "NotFoundError",
"code": 404
}
}
```
`classify_boxlite_error()` maps `BoxliteError` to HTTP status:
| `"not found"` | 404 | `NotFoundError` |
| `"already"` / `"conflict"` | 409 | `ConflictError` |
| `"unsupported"` | 400 | `UnsupportedError` |
| anything else | 500 | `InternalError` |
## See Also
- [CLI Development Guide](../../../../../docs/development/cli.md) — building and testing the CLI
- [Architecture](../../../../../docs/architecture/README.md) — core runtime architecture
- [Rust Style Guide](../../../../../docs/development/rust-style.md) — coding standards