yao/sandbox/v2/DESIGN.md

Sandbox V2 Design

Positioning

Sandbox is a standalone infrastructure module in Yao, on the same level as process, store, and fs. It provides isolated execution environments with standard file I/O. Any module can use it — Agent, JSAPI scripts, Process handlers, API endpoints.

Yao Infrastructure
├── process    — process execution
├── store      — KV storage
├── fs         — host filesystem
├── stream     — streaming execution (planned)
├── workspace  — persistent user storage ← new in V2
└── sandbox    — isolated execution environments ← this module

Sandbox does NOT import or depend on Agent. Agent is one of many consumers.

Architecture

┌─────────────────────────────────────────────────┐
│  Consumers (know nothing about tai/Docker/K8s)  │
│  ├── JSAPI: sandbox.Create/Get/List/Delete      │
│  ├── JSAPI: workspace.Create/Get/List/Delete     │
│  ├── Process: sandbox.Create, sandbox.Exec      │
│  ├── Agent: uses sandbox via interface           │
│  └── API: /api/__yao/sandbox/*                   │
└──────────────────┬──────────────────────────────┘
                   │ sandbox.Manager (public API)
                   ▼
┌─────────────────────────────────────────────────┐
│  sandbox/v2                                      │
│                                                  │
│  Manager (global singleton)                      │
│  ├── Create / Get / GetOrCreate / List / Remove │
│  ├── Start / Cleanup / Close                    │
│  ├── Heartbeat (idle tracking)                  │
│  ├── AddPool / RemovePool / Pools               │
│  ├── SetWorkspaceManager (workspace integration)│
│  ├── EnsureImage / ImageExists / PullImage      │
│  └── guard rails (limits, TTL) + Box factory    │
│                                                  │
│  Box (per-instance)                              │
│  ├── Exec(cmd) → ExecResult                     │
│  ├── Stream(cmd) → ExecStream (real-time I/O)   │
│  ├── Attach(port) → ServiceConn (WS/SSE)        │
│  ├── Workspace() → workspace.FS                 │
│  ├── VNC() → url                                │
│  ├── Proxy(port) → url                          │
│  └── Start / Stop / Remove / Info               │
└──────────────────┬──────────────────────────────┘
                   │
                   ▼
┌─────────────────────────────────────────────────┐
│  tai.Client pool (lazy-initialized)             │
│  ├── "local"  → tai.New("local")      (Docker)  │
│  ├── "gpu"    → tai.New("tai://gpu")  (Remote)  │
│  ├── "k8s"    → tai.New("tai://k8s",K8s)(K8s)  │
│  └── ...                                        │
│                                                  │
│  Each tai.Client provides:                       │
│  ├── Sandbox()   → CRUD + Exec + ExecStream     │
│  ├── Image()     → Exists + Pull + Remove + List│
│  ├── Volume()    → file I/O (local disk / gRPC) │
│  ├── Workspace() → fs.FS                        │
│  ├── Proxy()     → URL resolve + Connect        │
│  └── VNC()       → VNC WebSocket                │
└─────────────────────────────────────────────────┘

Dependency Rules

sandbox/v2 → tai           ✓  (sole runtime dependency)
sandbox/v2 → workspace     ✓  (workspace integration, optional)
sandbox/v2 → agent         ✗  NEVER
sandbox/v2 → docker        ✗  NEVER (tai handles it)
agent      → sandbox/v2    ✓  (consumer, via Manager API)
jsapi      → sandbox/v2    ✓  (consumer, via Manager API)
process    → sandbox/v2    ✓  (consumer, via Manager API)

Manager

Global singleton. Manages a pool of named tai.Client connections — each pool entry targets a different runtime endpoint (local Docker, remote Tai, K8s cluster). Caller picks which pool to create a sandbox on.

Pool

type Pool struct {
    Name        string
    Addr        string        // tai.New() address: "local", "tai://host", "docker:///path"
    Options     []tai.Option  // tai.K8s, tai.WithKubeConfig(), tai.WithPorts(), etc.
    MaxPerUser  int           // max boxes per user on this pool, 0 = unlimited
    MaxTotal    int           // max boxes total on this pool, 0 = unlimited
    IdleTimeout time.Duration // 0 = no timeout
    MaxLifetime time.Duration // 0 = no limit
    StopTimeout time.Duration // SIGTERM grace period before SIGKILL; 0 = DefaultStopTimeout (2s)
}

Example configuration:

pool:
  - name: local
    addr: "local"
    max_total: 20
    idle_timeout: 30m

  - name: gpu
    addr: "tai://gpu-server.internal"
    max_per_user: 1
    max_total: 4
    idle_timeout: 10m
    max_lifetime: 2h

  - name: k8s
    addr: "tai://k8s-proxy.internal"
    max_total: 100
    idle_timeout: 1h
    options:
      runtime: k8s
      kubeconfig: /etc/yao/kubeconfig.yml

Initialization

var mgr *Manager

func Init(cfg Config) error   // create Manager from Config; at least one Pool required
func M() *Manager             // return global singleton; panics if Init not called

Startup sequence in cmd/start.go:

config.Load
sandbox.Init(config.Conf.Sandbox)  // create Manager with pool + guard rails
engine.Load
...
service.Start                      // HTTP
grpc.Start                         // gRPC
sandbox.M().Start(ctx)             // discover existing containers, start cleanup loop

Init creates the Manager from config (pool definitions + guard rails). Start connects to pools, discovers existing containers, and starts the cleanup loop. Two-step so that gRPC server is ready before Start.

Pool connections are created lazily on first use and reused across all Box instances.

Config

type Config struct {
    Pool []Pool
}

Container gRPC env vars (YAO_GRPC_ADDR, etc.) are derived automatically at creation time. Per-instance settings (image, memory, CPU, workdir, env, pool) are passed via CreateOptions.

Core API

type Manager struct {
    pool        map[string]*tai.Client // name → connection (lazy-initialized)
    poolDefs    []Pool
    defaultPool string                 // first pool name
    config      Config
    boxes       sync.Map               // id → *Box
    mu          sync.Mutex
    cancel      context.CancelFunc
    grpcPort    int
    wsManager   *workspace.Manager     // optional workspace integration
}

// --- Bootstrap ---
func (m *Manager) Start(ctx context.Context) error
func (m *Manager) Close() error
func (m *Manager) SetGRPCPort(port int)
func (m *Manager) SetWorkspaceManager(wm *workspace.Manager)

// --- Pool management ---
func (m *Manager) AddPool(ctx context.Context, p Pool) error
func (m *Manager) RemovePool(ctx context.Context, name string, force bool) error
func (m *Manager) Pools() []PoolInfo

// --- Heartbeat ---
func (m *Manager) Heartbeat(sandboxID string, active bool, processCount int) error

// --- CRUD ---
func (m *Manager) Create(ctx context.Context, opts CreateOptions) (*Box, error)
func (m *Manager) Get(ctx context.Context, id string) (*Box, error)
func (m *Manager) GetOrCreate(ctx context.Context, opts CreateOptions) (*Box, error)
func (m *Manager) List(ctx context.Context, opts ListOptions) ([]*Box, error)
func (m *Manager) Remove(ctx context.Context, id string) error
func (m *Manager) Cleanup(ctx context.Context) error

// --- Image management ---
func (m *Manager) ImageExists(ctx context.Context, pool, ref string) (bool, error)
func (m *Manager) PullImage(ctx context.Context, pool, ref string, opts ImagePullOptions) (<-chan PullProgress, error)
func (m *Manager) EnsureImage(ctx context.Context, pool, ref string, opts ImagePullOptions) error

CreateOptions

type CreateOptions struct {
    ID          string
    Owner       string
    Labels      map[string]string
    Pool        string              // which tai.Client to use; empty = default pool

    // Container spec
    Image       string              // required
    WorkDir     string              // default "/workspace"
    User        string
    Env         map[string]string
    Memory      int64               // bytes, 0 = no limit
    CPUs        float64             // 0 = no limit
    VNC         bool
    Ports       []PortMapping

    // Lifecycle
    Policy      LifecyclePolicy     // default: Session
    IdleTimeout time.Duration       // override pool default; 0 = use pool default
    StopTimeout time.Duration       // SIGTERM grace period; 0 = pool default or DefaultStopTimeout

    // Workspace integration
    WorkspaceID string              // workspace to mount; empty = no workspace
    MountMode   string              // "rw" (default) or "ro"
    MountPath   string              // container path; default "/workspace"
}

When WorkspaceID is set, the Manager resolves the workspace's bound node via workspace.Manager.NodeForWorkspace() and forces the container onto that node. The workspace directory is bind-mounted into the container at MountPath.

LifecyclePolicy

type LifecyclePolicy string

const (
    OneShot     LifecyclePolicy = "oneshot"     // destroyed after first Exec
    Session     LifecyclePolicy = "session"     // alive while active, cleaned on idle
    LongRunning LifecyclePolicy = "longrunning" // user workspace, extended TTL
    Persistent  LifecyclePolicy = "persistent"  // never auto-cleaned
)

const DefaultStopTimeout = 2 * time.Second

Box

A Box is a single sandbox instance. All operations go through it.

type Box struct {
    id            string
    containerID   string
    pool          string
    owner         string
    policy        LifecyclePolicy
    labels        map[string]string
    lastCall      atomic.Int64  // last external API call
    lastHeartbeat atomic.Int64  // last container heartbeat
    processCount  atomic.Int32  // user processes inside container
    idleTimeoutD  time.Duration
    stopTimeoutD  time.Duration
    createdAt     time.Time
    refreshToken  string
    vnc           bool
    image         string
    workspaceID   string
    ws            workspace.FS  // lazy-initialized, cached
    manager       *Manager
}

// --- Identity ---
func (b *Box) ID() string
func (b *Box) Owner() string
func (b *Box) ContainerID() string
func (b *Box) Pool() string
func (b *Box) WorkspaceID() string

// --- Execution ---
func (b *Box) Exec(ctx context.Context, cmd []string, opts ...ExecOption) (*ExecResult, error)
func (b *Box) Stream(ctx context.Context, cmd []string, opts ...ExecOption) (*ExecStream, error)
func (b *Box) Attach(ctx context.Context, port int, opts ...AttachOption) (*ServiceConn, error)

// --- Filesystem ---
func (b *Box) Workspace() workspace.FS

// --- Network ---
func (b *Box) VNC(ctx context.Context) (string, error)
func (b *Box) Proxy(ctx context.Context, port int, path string) (string, error)

// --- Lifecycle ---
func (b *Box) Start(ctx context.Context) error
func (b *Box) Stop(ctx context.Context) error
func (b *Box) Remove(ctx context.Context) error
func (b *Box) Info(ctx context.Context) (*BoxInfo, error)

ExecOption / ExecResult / ExecStream

type ExecOption func(*execConfig)

func WithWorkDir(dir string) ExecOption
func WithEnv(env map[string]string) ExecOption
func WithTimeout(d time.Duration) ExecOption

type ExecResult struct {
    ExitCode int
    Stdout   string
    Stderr   string
}

type ExecStream struct {
    Stdout io.ReadCloser
    Stderr io.ReadCloser
    Stdin  io.WriteCloser
    Wait   func() (int, error) // block until exit, return exit code
    Cancel func()              // kill the process
}

AttachOption / ServiceConn

type AttachOption func(*attachConfig)

func WithProtocol(proto string) AttachOption   // "ws", "sse"; default "ws"
func WithPath(path string) AttachOption
func WithHeaders(h map[string]string) AttachOption

type ServiceConn struct {
    Read   func() ([]byte, error) // read next message (WS mode)
    Write  func(data []byte) error
    Events <-chan []byte           // SSE event channel
    URL    string
    Close  func() error
}

port is the port the service listens on inside the container. Routing — Docker port mapping (local) or Tai HTTP proxy (remote) — is handled internally.

Image Management

type ImagePullOptions struct {
    Auth *RegistryAuth
}

type RegistryAuth struct {
    Username string
    Password string
    Server   string
}

EnsureImage first checks ImageExists; if not present, calls PullImage and blocks until complete. For K8s pools this is a no-op — kubelet manages image pulling natively via imagePullPolicy.

BoxInfo / PoolInfo

type BoxInfo struct {
    ID           string
    ContainerID  string
    Pool         string
    Owner        string
    Status       string // "running", "stopped", "creating"
    Policy       LifecyclePolicy
    Labels       map[string]string
    Image        string
    CreatedAt    time.Time
    LastActive   time.Time
    ProcessCount int
    VNC          bool
}

type PoolInfo struct {
    Name        string
    Addr        string
    Connected   bool
    Boxes       int
    MaxPerUser  int
    MaxTotal    int
    IdleTimeout time.Duration
    MaxLifetime time.Duration
}

Workspace Integration

Sandbox V2 integrates with the workspace module via Manager.SetWorkspaceManager() and CreateOptions.WorkspaceID:

// Link workspace manager at startup
sbm.SetWorkspaceManager(wsm)

// Create sandbox with workspace mount
box, err := sbm.Create(ctx, sandbox.CreateOptions{
    Image:       "yaoapp/workspace:latest",
    WorkspaceID: "ws-abc123",
    MountMode:   "rw",       // default
    MountPath:   "/workspace", // default
})

When WorkspaceID is set:

1. Manager calls workspace.Manager.NodeForWorkspace() to resolve the workspace's bound node
2. Forces the container onto that node's pool
3. Calls workspace.Manager.MountPath() to get the host-side directory
4. Adds a Docker bind mount: hostPath:mountPath:mode
5. Box.Workspace() uses the workspace ID as the volume session key

This guarantees that a workspace's container always runs on the same host where its storage lives.

Container Setup — Manager.Create()

When Manager creates a sandbox, it:

1. Validates CreateOptions (Image required)
2. Generates sandbox ID (or uses provided one)
3. Resolves workspace node binding (if WorkspaceID set)
4. Checks user limits (MaxPerUser) and total limits (MaxTotal)
5. Resolves pool (by name or default)
6. Creates OAuth token pair for container IPC
7. Builds tai.sandbox.CreateOptions:
   • Injects management labels: managed-by, sandbox-id, sandbox-owner, sandbox-pool, sandbox-policy, workspace-id
   • Sets container CMD to graceful-shutdown-aware sleep: sh -c "trap 'exit 0' TERM; while :; do sleep 86400 & wait $!; done"
   • Merges caller's Env with gRPC env vars (YAO_SANDBOX_ID, YAO_TOKEN, YAO_REFRESH_TOKEN, YAO_GRPC_ADDR, etc.)
   • Adds workspace bind mount if WorkspaceID is set
8. Calls tai.Client.Sandbox().Create() then Start()
9. Wraps in a Box, registers in boxes map

Lifecycle Management

Idle Tracking — Dual Source

box.lastActive = max(lastExternalCall, lastHeartbeat)

Source	What it tracks	Updated by
External call	Caller is using the sandbox	Box.Exec(), Box.Stream(), Box.Workspace(), Box.VNC(), Box.Proxy(), Box.Attach()
Container heartbeat	Processes running inside the container	gRPC Heartbeat RPC

Cleanup Loop

Runs every 60 seconds. Policy behavior:

Policy	Idle	Max Lifetime	Auto
OneShot	—	—	Removed after first Exec completes
Session	Remove	Remove	Default for agent chats
LongRunning	Stop (keep data)	Remove	User workspaces
Persistent	Never	Never	User-managed

Container Stop Behavior

DefaultStopTimeout = 2s. Docker ContainerStop sends SIGTERM, waits the timeout, then SIGKILL. The V2 container CMD (trap 'exit 0' TERM; ...) exits immediately on SIGTERM, so actual stop time is near-instant.

Manager.Remove() calls Sandbox().Remove(force=true) directly (SIGKILL + delete) — no redundant Stop call. This keeps remove latency under 200ms.

Tai SDK Interface

Sandbox V2 depends on these tai sub-package interfaces:

tai.Client

func New(addr string, opts ...Option) (*Client, error)
func (c *Client) Sandbox() sandbox.Sandbox
func (c *Client) Image() sandbox.Image
func (c *Client) Volume() volume.Volume
func (c *Client) Workspace(sessionID string) workspace.FS
func (c *Client) Proxy() proxy.Proxy
func (c *Client) VNC() vnc.VNC
func (c *Client) DataDir() string
func (c *Client) IsLocal() bool
func (c *Client) Close() error

Address schemes: "local" (Docker default), "docker://..." (explicit Docker), "tai://host" (remote Tai Server). Remote mode auto-discovers service ports via ServerInfo gRPC, with WithPorts() taking precedence.

sandbox.Sandbox

type Sandbox interface {
    Create(ctx, opts CreateOptions) (string, error)
    Start(ctx, id string) error
    Stop(ctx, id string, timeout time.Duration) error
    Remove(ctx, id string, force bool) error
    Exec(ctx, id string, cmd []string, opts ExecOptions) (*ExecResult, error)
    ExecStream(ctx, id string, cmd []string, opts ExecOptions) (*StreamHandle, error)
    Inspect(ctx, id string) (*ContainerInfo, error)
    List(ctx, opts ListOptions) ([]ContainerInfo, error)
    Close() error
}

Implementations: docker_core.go (local Docker), docker.go (remote Docker via Tai proxy), k8s.go (Kubernetes via Tai proxy).

sandbox.Image

type Image interface {
    Exists(ctx, ref string) (bool, error)
    Pull(ctx, ref string, opts PullOptions) (<-chan PullProgress, error)
    Remove(ctx, ref string, force bool) error
    List(ctx) ([]ImageInfo, error)
}

Docker implementation pulls via Docker SDK with real-time progress streaming. K8s implementation is a no-op — kubelet handles image pulling.

proxy.Proxy

type Proxy interface {
    URL(ctx, containerID string, port int, path string) (string, error)
    Connect(ctx, containerID string, opts ConnectOptions) (*Connection, error)
    Healthz(ctx) error
}

Local: resolves host ports via Inspect(). Remote: routes through Tai HTTP proxy which handles WebSocket upgrade and SSE streaming natively.

gRPC Token Injection

func CreateContainerTokens(sandboxID, owner string, scopes []string) (access, refresh string, err error)
func RevokeContainerTokens(refresh string) error
func BuildGRPCEnv(pool *Pool, sandboxID, access, refresh string, grpcPort int) map[string]string

Environment variables injected into each container:

# All modes
YAO_SANDBOX_ID=<sandbox_id>
YAO_TOKEN=<access_token>
YAO_REFRESH_TOKEN=<refresh_token>
YAO_GRPC_ADDR=127.0.0.1:9099

# Remote mode (tai://)
YAO_GRPC_ADDR=<tai-host>:9100

Errors

var (
    ErrNotAvailable  = errors.New("sandbox: not available (no pools configured)")
    ErrNotFound      = errors.New("sandbox: not found")
    ErrLimitExceeded = errors.New("sandbox: limit exceeded")
    ErrPoolNotFound  = errors.New("sandbox: pool not found")
    ErrPoolInUse     = errors.New("sandbox: pool has running boxes")
)

Package Structure

sandbox/v2/
├── sandbox.go              // Init, M(), global singleton
├── manager.go              // Manager: CRUD, pool management, image ops, cleanup
├── box.go                  // Box: Exec, Stream, Attach, Workspace, VNC, Proxy, lifecycle
├── types.go                // CreateOptions, ExecResult, ExecStream, ServiceConn, BoxInfo, etc.
├── config.go               // Config struct
├── errors.go               // sentinel errors
├── grpc.go                 // token creation/revocation, gRPC env var injection
├── jsapi/                  // (Phase 2) V8 JSAPI sandbox.* namespace
│   ├── jsapi.go            // RegisterObject("sandbox"), Create/Get/List/Delete
│   └── box.go              // Box JS object: Exec/Attach/VNC/Proxy/Workspace/Info/Start/Stop/Remove
├── export_test.go          // ResetForTest() for test isolation
├── testutils_test.go       // shared test helpers (multi-pool setup)
├── sandbox_test.go         // Init/M singleton tests
├── manager_test.go         // Manager CRUD tests
├── manager_lifecycle_test.go // Heartbeat, Cleanup, idle tracking tests
├── box_test.go             // Box Exec/Workspace/Info tests
├── box_attach_test.go      // Attach WS/SSE/VNC tests
├── box_workspace_test.go   // Workspace integration tests
├── box_image_test.go       // Image Pull API tests
├── bench_test.go           // Performance benchmarks
├── grpc_test.go            // Token/env building tests
├── DESIGN.md               // this document
└── IMPL.md                 // implementation status and plan

---

Workspace Module

Positioning

Workspace is a top-level module (workspace/), parallel to sandbox/v2. It provides persistent, user-managed storage that is decoupled from container lifecycle. Workspaces are pinned to a specific Tai node; containers referencing a workspace are automatically routed to that node.

┌─────────────────────┐     ┌─────────────────────┐
│    sandbox/v2        │     │     workspace        │
│  (container runtime) │◄────│  (persistent storage)│
│                      │     │                      │
│  CreateOptions {     │     │  CRUD + File I/O     │
│    WorkspaceID ──────┼────►│  Node binding        │
│  }                   │     │  fs.FS interface      │
└──────────┬───────────┘     └──────────┬───────────┘
           │                            │
           └──────────┬─────────────────┘
                      ▼
              tai.Client pool

Core Types

type Workspace struct {
    ID        string
    Name      string
    Owner     string
    Node      string              // Tai node this workspace is pinned to
    Labels    map[string]string
    CreatedAt time.Time
    UpdatedAt time.Time
}

type CreateOptions struct {
    ID     string                 // explicit ID; empty = auto-generate (ws-<uuid>)
    Name   string
    Owner  string
    Node   string                 // target Tai node (required)
    Labels map[string]string
}

type ListOptions struct {
    Owner string
    Node  string
}

type UpdateOptions struct {
    Name   *string               // nil = no change
    Labels map[string]string     // nil = no change; non-nil replaces all labels
}

type NodeInfo struct {
    Name   string
    Addr   string
    Online bool
}

type DirEntry struct {
    Name  string
    IsDir bool
    Size  int64
}

Manager API

type Manager struct {
    pools map[string]*tai.Client
    mu    sync.RWMutex
}

func NewManager(pools map[string]*tai.Client) *Manager

// --- CRUD ---
func (m *Manager) Create(ctx, opts CreateOptions) (*Workspace, error)
func (m *Manager) Get(ctx, id string) (*Workspace, error)
func (m *Manager) List(ctx, opts ListOptions) ([]*Workspace, error)
func (m *Manager) Update(ctx, id string, opts UpdateOptions) (*Workspace, error)
func (m *Manager) Delete(ctx, id string, force bool) error

// --- File I/O ---
func (m *Manager) ReadFile(ctx, id string, path string) ([]byte, error)
func (m *Manager) WriteFile(ctx, id string, path string, data []byte, perm os.FileMode) error
func (m *Manager) ListDir(ctx, id string, path string) ([]DirEntry, error)
func (m *Manager) Remove(ctx, id string, path string) error
func (m *Manager) FS(ctx, id string) (workspace.FS, error)

// --- Node management ---
func (m *Manager) Nodes() []NodeInfo
func (m *Manager) AddPool(name string, client *tai.Client)
func (m *Manager) RemovePool(name string)

// --- Sandbox integration ---
func (m *Manager) NodeForWorkspace(ctx, id string) (string, error)
func (m *Manager) MountPath(ctx, id string) (string, error)

Metadata Storage

Workspace metadata is stored as .workspace.json inside the workspace's root directory on the Tai node:

<data-dir>/
├── ws-abc123/
│   ├── .workspace.json   ← metadata (ID, Name, Owner, Node, Labels, timestamps)
│   ├── src/
│   ├── go.mod
│   └── ...
├── ws-def456/
│   └── ...

This approach collocates metadata with data — no external database required. List() scans top-level directories and reads each .workspace.json. Get() scans all nodes until the workspace is found.

Errors

var (
    ErrNotFound    = errors.New("workspace: not found")
    ErrNodeMissing = errors.New("workspace: node is required")
    ErrNodeOffline = errors.New("workspace: node is offline or not configured")
    ErrHasMounts   = errors.New("workspace: workspace has active container mounts")
)

Package Structure

workspace/
├── workspace.go         // types, metadata marshal/unmarshal
├── manager.go           // Manager: CRUD, file I/O, node management
├── errors.go            // sentinel errors
├── jsapi/               // (Phase 2) V8 JSAPI workspace.* namespace
│   ├── jsapi.go         // RegisterObject("workspace"), Create/Get/List/Delete
│   └── fs.go            // WorkspaceFS JS object: ReadFile/WriteFile/ReadDir/Stat/MkdirAll/Remove/RemoveAll/Rename
├── testutils_test.go    // shared test helpers
├── workspace_test.go    // CRUD tests (Create/Get/List/Update/Delete/Nodes)
├── fileio_test.go       // File I/O + fs.FS tests
├── bench_test.go        // Performance benchmarks
└── DESIGN.md            // detailed design document

---

Testing

Test Environment

Three pool modes configured via environment variables:

# Local — direct Docker daemon (always available)
SANDBOX_TEST_LOCAL_ADDR=local

# Remote — via Tai container (Docker backend)
SANDBOX_TEST_REMOTE_ADDR=tai://127.0.0.1:9100

# K8s — via Tai container (K8s backend)
TAI_TEST_K8S_HOST=<tai-host>
TAI_TEST_KUBECONFIG=<path>
TAI_TEST_K8S_PORT=6443
TAI_TEST_K8S_NAMESPACE=default

# Test image
SANDBOX_TEST_IMAGE=yaoapp/sandbox-v2-test:latest

Tests skip unavailable modes via t.Skip. Both sandbox/v2 and workspace tests iterate over all available pools.

Test Coverage

sandbox/v2

File	Coverage
sandbox_test.go	Init(), M(), singleton behavior
manager_test.go	Create, Get, GetOrCreate, List, Remove, pool management, limits
manager_lifecycle_test.go	Start (container discovery), Cleanup, idle tracking, Heartbeat
box_test.go	Exec, Info, Workspace (ReadFile/WriteFile), lifecycle
box_attach_test.go	Attach WS, Attach SSE, VNC URL, VNC Connect
box_workspace_test.go	Workspace file I/O through Box, workspace mount integration
box_image_test.go	ImageExists, PullImage (with progress), EnsureImage, K8s no-op
grpc_test.go	Token creation/revocation, env var building
bench_test.go	ContainerLifecycle, Create, Exec, ExecHeavy, Remove, Info, StopStart, WorkspaceReadWrite

workspace

File	Coverage
workspace_test.go	Create (auto/explicit ID, labels, invalid node), Get, List (filter owner/node), Update (name/labels), Delete, Nodes, NodeForWorkspace, AddPool, RemovePool, MountPath
fileio_test.go	ReadWriteFile, nested paths, ListDir, Remove, fs.FS (ReadFile, WriteFile, MkdirAll, Rename, WalkDir, Remove)
bench_test.go	WriteFile, ReadFile, ReadWriteCycle, WriteLargeFile, ListDir, FSWalkDir, CreateDelete

CI Integration

Consolidated into two CI jobs:

Job	Contents
SandboxV2Test	Image pre-pull → tai-test → sandbox/v2 (local+remote+k8s) → workspace (local+remote)
BenchmarkSandboxV2	Performance tests for sandbox/v2 + workspace (parallel with SandboxV2Test)

Benchmark Results (Reference)

Benchmark	Local	Remote	K8s
ContainerLifecycle	~300ms	~200ms	~10s
Create	~100ms	~80ms	~8s
Exec	~30ms	~50ms	~150ms
Remove	~180ms	~120ms	~220ms
Info	~5ms	~10ms	~30ms
StopStart	~2.2s	~2.2s	N/A (skip)

K8s StopStart is skipped because K8s Stop deletes the Pod; a subsequent Start cannot restart a deleted Pod.

Docker StopStart ~2.2s is expected: DefaultStopTimeout = 2s and Docker waits the full timeout before SIGKILL unless PID 1 exits on SIGTERM first.

---

Migration Plan

Phase 1: Core (DONE)

• tai SDK: Sandbox, ExecStream, Image, Proxy.Connect, Labels, User
• sandbox/v2: Manager, Box, all CRUD + Exec + Stream + Attach + Workspace + VNC + Proxy + Image
• workspace: Manager, CRUD, file I/O, node binding, sandbox integration
• gRPC: Heartbeat RPC (proto + handler)
• Tests: unit + integration + benchmarks
• CI: consolidated SandboxV2Test + BenchmarkSandboxV2

Phase 2: JSAPI + OAuth + Auth (PENDING)

Prerequisites

Task	Detail
Wire openapi/oauth	grpc.go currently uses random token placeholders; replace with real OAuth issue/revoke

JSAPI Design

All JSAPI methods are static — no constructors, no Go objects in V8, no bridge/Release. JS objects only hold string IDs, delegate everything to Go singletons (sandbox.M(), workspace.M()).

sandbox namespace (RegisterObject("sandbox"))

Static methods:

JS	Go	Returns
sandbox.Create(opts)	Manager.Create(ctx, CreateOptions)	Box
sandbox.Create(opts) (opts.id set)	Manager.GetOrCreate(ctx, CreateOptions)	Box
sandbox.Get(id)	Manager.Get(ctx, id)	Box | null
sandbox.List(filter?)	Manager.List(ctx, ListOptions) → Box.Info()	BoxInfo[]
sandbox.Delete(id)	Manager.Remove(ctx, id)	void

sandbox.Create(options) — JS options → Go CreateOptions:

{
  id:           string   →  CreateOptions.ID           // optional; triggers GetOrCreate
  owner:        string   →  CreateOptions.Owner        // required
  pool:         string   →  CreateOptions.Pool         // default: first pool
  image:        string   →  CreateOptions.Image        // required
  workdir:      string   →  CreateOptions.WorkDir
  user:         string   →  CreateOptions.User         // e.g. "1000:1000"
  env:          object   →  CreateOptions.Env          // map[string]string
  memory:       number   →  CreateOptions.Memory       // bytes (int64)
  cpus:         number   →  CreateOptions.CPUs         // float64
  vnc:          boolean  →  CreateOptions.VNC
  ports:        array    →  CreateOptions.Ports        // [{container, host, host_ip, protocol}] → []PortMapping
  policy:       string   →  CreateOptions.Policy       // "oneshot"|"session"|"longrunning"|"persistent"
  idle_timeout: number   →  CreateOptions.IdleTimeout  // ms → time.Duration
  stop_timeout: number   →  CreateOptions.StopTimeout  // ms → time.Duration
  workspace_id: string   →  CreateOptions.WorkspaceID
  mount_mode:   string   →  CreateOptions.MountMode    // "rw"|"ro"
  mount_path:   string   →  CreateOptions.MountPath
  labels:       object   →  CreateOptions.Labels       // map[string]string
}

sandbox.List(filter?) — JS filter → Go ListOptions:

{
  owner:  string  →  ListOptions.Owner   // empty = all
  pool:   string  →  ListOptions.Pool    // empty = all
  labels: object  →  ListOptions.Labels
}

Returns BoxInfo[] — each element:

{
  id:            string   ←  BoxInfo.ID
  container_id:  string   ←  BoxInfo.ContainerID
  pool:          string   ←  BoxInfo.Pool
  owner:         string   ←  BoxInfo.Owner
  status:        string   ←  BoxInfo.Status
  image:         string   ←  BoxInfo.Image
  vnc:           boolean  ←  BoxInfo.VNC
  policy:        string   ←  BoxInfo.Policy
  labels:        object   ←  BoxInfo.Labels
  created_at:    string   ←  BoxInfo.CreatedAt   (ISO 8601)
  last_active:   string   ←  BoxInfo.LastActive   (ISO 8601)
  process_count: number   ←  BoxInfo.ProcessCount
}

Box object

Read-only properties:

JS	Go
box.id	Box.ID()
box.owner	Box.Owner()
box.pool	Box.Pool()

Methods:

JS	Go	Returns
box.Exec(cmd, opts?)	Box.Exec(ctx, cmd, ...ExecOption)	ExecResult
box.Stream(cmd, opts?)	Box.Stream(ctx, cmd, ...ExecOption)	ExecStream
box.Attach(port, opts?)	Box.Attach(ctx, port, ...AttachOption)	ServiceConn
box.VNC()	Box.VNC(ctx)	string
box.Proxy(port, path?)	Box.Proxy(ctx, port, path)	string
box.Workspace()	Box.WorkspaceID() → NewFSObject	WorkspaceFS
box.Info()	Box.Info(ctx)	BoxInfo
box.Start()	Box.Start(ctx)	void
box.Stop()	Box.Stop(ctx)	void
box.Remove()	Box.Remove(ctx)	void

box.Exec(cmd, options?):

cmd:     string[]                         → cmd []string
options: {
  workdir: string,                        → WithWorkDir(dir)
  env:     object,                        → WithEnv(map[string]string)
  timeout: number                         → WithTimeout(ms → time.Duration)
}
returns: {
  exit_code: number,                      ← ExecResult.ExitCode
  stdout:    string,                      ← ExecResult.Stdout
  stderr:    string                       ← ExecResult.Stderr
}

box.Stream(cmd, options?):

options: same as Exec
returns: {
  stdout: ReadableStream,                 ← ExecStream.Stdout
  stderr: ReadableStream,                 ← ExecStream.Stderr
  stdin:  WritableStream,                 ← ExecStream.Stdin
  wait:   function() → number,            ← ExecStream.Wait() (int, error)
  cancel: function() → void               ← ExecStream.Cancel()
}

box.Attach(port, options?):

port:    number                           → port int
options: {
  protocol: "ws"|"sse",                  → WithProtocol(protocol)
  path:     string,                      → WithPath(path)
  headers:  object                       → WithHeaders(map[string]string)
}
returns: {
  url:    string,                         ← ServiceConn.URL
  read:   function() → Uint8Array,        ← ServiceConn.Read()
  write:  function(data) → void,          ← ServiceConn.Write(data)
  events: AsyncIterable<Uint8Array>,      ← ServiceConn.Events
  close:  function() → void               ← ServiceConn.Close()
}

box.Info() returns same structure as BoxInfo[] element above.

workspace namespace (RegisterObject("workspace"))

Static methods:

JS	Go	Returns
workspace.Create(opts)	Manager.Create(ctx, CreateOptions)	WorkspaceFS
workspace.Get(id)	Manager.Get(ctx, id)	WorkspaceFS | null
workspace.List(filter?)	Manager.List(ctx, ListOptions)	WorkspaceInfo[]
workspace.Delete(id)	Manager.Delete(ctx, id, false)	void

workspace.Create(options) — JS options → Go CreateOptions:

{
  id:     string  →  CreateOptions.ID      // optional; auto-generated if empty
  name:   string  →  CreateOptions.Name    // required
  owner:  string  →  CreateOptions.Owner   // required
  node:   string  →  CreateOptions.Node    // required
  labels: object  →  CreateOptions.Labels  // map[string]string
}

workspace.List(filter?) — JS filter → Go ListOptions:

{
  owner: string  →  ListOptions.Owner  // empty = all
  node:  string  →  ListOptions.Node   // empty = all
}

Returns WorkspaceInfo[] — each element:

{
  id:         string  ←  Workspace.ID
  name:       string  ←  Workspace.Name
  owner:      string  ←  Workspace.Owner
  node:       string  ←  Workspace.Node
  labels:     object  ←  Workspace.Labels
  created_at: string  ←  Workspace.CreatedAt (ISO 8601)
  updated_at: string  ←  Workspace.UpdatedAt (ISO 8601)
}

WorkspaceFS object

Read-only properties:

JS	Go
ws.id	workspace ID
ws.name	Workspace.Name
ws.node	Workspace.Node

Methods (1:1 to Go taiworkspace.FS + Manager shortcuts):

JS	Go	Returns
ws.ReadFile(path)	FS.ReadFile(name) / Manager.ReadFile(ctx, id, path)	string
ws.WriteFile(path, data, perm?)	FS.WriteFile(name, data, perm) / Manager.WriteFile(ctx, id, path, data, perm)	void
ws.ReadDir(path?)	FS.ReadDir(name) / Manager.ListDir(ctx, id, path)	DirEntry[]
ws.Stat(path)	FS.Stat(name)	FileInfo
ws.MkdirAll(path, perm?)	FS.MkdirAll(name, perm)	void
ws.Remove(path)	FS.Remove(name) / Manager.Remove(ctx, id, path)	void
ws.RemoveAll(path)	FS.RemoveAll(name)	void
ws.Rename(from, to)	FS.Rename(old, new)	void

Planned (not yet implemented):

JS	Go	Returns	Note
ws.ReadFileBase64(path)	FS.ReadFile → base64.StdEncoding.EncodeToString	string	Avoids V8↔Go binary bridge overhead for images, archives, etc.
ws.WriteFileBase64(path, b64, perm?)	base64.StdEncoding.DecodeString → FS.WriteFile	void	Same — base64 string transfer is far more efficient than Uint8Array across the bridge
ws.CopyFromHost(hostPath, destPath?)	Host os.Read → FS.WriteFile / FS.MkdirAll per entry	void	Copy file/dir from Yao host into workspace; destPath defaults to basename
ws.CopyFromHostArchive(hostPath, destPath?)	Zip on host → Tai Volume upload → Tai-side unarchive	void	For large directory trees; requires Tai server-side unarchive support

Return types:

DirEntry: { name: string, is_dir: boolean, size: number }
FileInfo: { name: string, size: number, is_dir: boolean, mod_time: string (ISO 8601) }

Auth

JSAPI does not enforce permissions internally. The Go Manager methods execute operations directly without owner/admin checks.

Developers retrieve the current caller identity via the gou global Authorized() function (registered by gou/runtime/v8/functions/authorized, reads from bridge.Share.Authorized / __yao_data.AUTHORIZED) and implement permission logic in their JS scripts.

Authorized() returns map[string]interface{} (or null if not set). The exact fields depend on what the caller sets via Context.WithAuthorized(). There is no fixed schema — typical fields include user_id, team_id, scope, etc.

const auth = Authorized()        // gou global — returns caller info or null
const box  = sandbox.Get(id)
// Developer decides permission logic — fields depend on application's auth setup
if (box.owner !== auth.user_id) {
    throw new Error("permission denied")
}

Permission control is the responsibility of the caller (JS scripts, Agent hooks, API middleware, etc.).

Implementation Tasks

Task	Detail
sandbox/v2/jsapi/	RegisterObject("sandbox") with Create/Get/List/Delete + Box object
workspace/jsapi/	RegisterObject("workspace") with Create/Get/List/Delete + FS object
Integration with cmd/start.go	Call sandbox.Init() + sandbox.M().Start()

Phase 3: Agent Integration (PENDING)

Task	Detail
Agent creates Box via sandbox.M().GetOrCreate()	Replace infraSandbox.Manager
Agent uses Box.Workspace() for file I/O	Replace Docker Copy/bind mount reads
Agent uses Box.Exec() for commands	Replace Docker exec
Agent uses Box.VNC() / Box.Proxy()	Replace vncproxy

Phase 4: Cutover (PENDING)

Task	Detail
Move sandbox/v2 → sandbox	Rename package
Delete old sandbox code	manager.go, ipc/, bridge/, vncproxy/, docker/
Update cmd/start.go	Use new init path
sandbox/process.go	Register sandbox.* process namespace (post-cutover)
workspace/process.go	Register workspace.* process namespace (post-cutover)

V1 vs V2 Comparison

Aspect	V1 (current)	V2 (this design)
Positioning	Agent's Claude executor	Yao infrastructure module
Runtime	Direct Docker SDK	tai.Client pool (Docker/K8s/Remote)
Execution	Exec + Stream	Exec + Stream + Attach (WS/SSE)
File I/O	bind mount + Docker Copy	workspace.FS (fs.FS compatible)
IPC	Unix socket + yao-bridge	gRPC (tai call)
Idle detection	External calls only	Dual: external calls + container heartbeat
Lifecycle	Chat session only	Policy-based (oneshot/session/longrunning/persistent)
Pool	Single Docker daemon	Multi-pool with per-pool policies
Agent coupling	Tightly coupled	Zero dependency
Workspace	None	Persistent, node-bound, decoupled from containers
Image management	None	EnsureImage + Pull with progress
K8s	Not supported	Supported via tai.Client
Multi-node	Local only	Local + Remote via Tai