Docker Sandbox for Agent Tools

Isolated execution of

claude

codex

docker sandbox

ADR: ADR-0023

Prerequisites

• Docker Desktop running (OrbStack works)

• docker sandbox version

  returns ≥0.11.0
• Auth secrets stored in

  agent-secrets

  :

  • claude_setup_token

    — from

    claude setup-token

    (1-year token, Max subscription)

  • codex_auth_json

    — contents of

    ~/.codex/auth.json

    (ChatGPT Pro subscription)

Quick Reference

# Create a sandbox
docker sandbox create --name my-sandbox claude /path/to/project

# Run a command in it
docker sandbox exec -e "CLAUDE_CODE_OAUTH_TOKEN=..." -w /path/to/project my-sandbox \
  claude -p "implement the feature" --output-format text --dangerously-skip-permissions

# List sandboxes
docker sandbox ls

# Remove
docker sandbox rm my-sandbox

Auth Setup (One-Time)

Claude (Max subscription)

Run interactively on the host (needs browser for OAuth):

claude setup-token

This opens a browser, completes OAuth, and prints a token like

sk-ant-oat01-...

Store it:

secrets add claude_setup_token --value "sk-ant-oat01-..."

Use in sandbox:

TOKEN=$(secrets lease claude_setup_token --ttl 1h --raw)
docker sandbox exec -e "CLAUDE_CODE_OAUTH_TOKEN=$TOKEN" my-sandbox claude auth status
# → loggedIn: true, authMethod: oauth_token

Codex (ChatGPT Pro subscription)

Authenticate codex locally (needs browser):

codex  # Select "Sign in with ChatGPT", complete OAuth

The auth file at

~/.codex/auth.json

secrets add codex_auth_json --value "$(cat ~/.codex/auth.json)"

Inject into sandbox:

AUTH=$(secrets lease codex_auth_json --ttl 1h --raw)
docker sandbox exec my-sandbox bash -c "mkdir -p ~/.codex && cat > ~/.codex/auth.json << 'EOF'
${AUTH}
EOF"

Token Refresh

claude_setup_token

claude setup-token

codex_auth_json

codex

Agent Loop Integration

Pre-warm Pattern

Create sandbox(es) at loop start, reuse for all stories, destroy at loop end.

PLANNER (loop start)
  ├── docker sandbox create --name loop-{loopId}-claude claude {workDir}
  ├── docker sandbox create --name loop-{loopId}-codex codex {workDir}  # if needed
  └── inject auth into both

IMPLEMENTOR / TEST-WRITER / REVIEWER (per story)
  └── docker sandbox exec -w {workDir} -e CLAUDE_CODE_OAUTH_TOKEN=... loop-{loopId}-{tool} \
        {tool command}
      # ~90ms overhead, workspace changes visible on host immediately

COMPLETE / CANCEL (loop end)
  ├── docker sandbox rm loop-{loopId}-claude
  └── docker sandbox rm loop-{loopId}-codex

Timing

Net overhead per loop: ~14s create + ~90ms × N stories = negligible for loops running 5-10 stories at 5-15min each.

Workspace Mount

The workspace is bidirectional — same path on host and in sandbox:

• File created in sandbox → visible on host at same path
• File created on host → visible in sandbox
• Git operations work normally (host sees sandbox changes, sandbox sees host commits)

Sandbox Templates

claude

codex

Neither includes

bun

Env Vars

Pass via

docker sandbox exec -e

docker sandbox exec \
  -e "CLAUDE_CODE_OAUTH_TOKEN=$TOKEN" \
  -e "NODE_ENV=development" \
  -w /path/to/project \
  my-sandbox \
  claude -p "prompt" --output-format text --dangerously-skip-permissions

Network Control

Sandboxes have network access by default. Restrict with proxy rules:

# Allow only API endpoints
docker sandbox network proxy my-sandbox --policy deny
docker sandbox network proxy my-sandbox --allow-host api.anthropic.com
docker sandbox network proxy my-sandbox --allow-host api.openai.com

Fallback to Host Mode

If Docker is unavailable:

# Check availability
docker info >/dev/null 2>&1 || echo "Docker not available"

# Force host mode
export AGENT_LOOP_HOST=1

Saving Custom Templates

If you install additional tools in a sandbox, save it as a template:

# Install tools
docker sandbox exec my-sandbox bash -c 'npm i -g @anthropic-ai/claude-code @openai/codex'

# Save as template
docker sandbox save my-sandbox my-agent-template:v1

# Use the template for future sandboxes
docker sandbox create --name fast-sandbox -t my-agent-template:v1 claude /path/to/project

Implementation in utils.ts

New Functions (ADR-0023)

// Create sandbox for a loop
async function createLoopSandbox(
  loopId: string,
  tool: "claude" | "codex",
  workDir: string
): Promise<string>  // returns sandbox name

// Execute command in existing sandbox
async function execInSandbox(
  sandboxName: string,
  command: string[],
  opts: { env?: Record<string, string>; workDir?: string; timeout?: number }
): Promise<{ exitCode: number; output: string }>

// Destroy loop sandbox(es)
async function destroyLoopSandbox(loopId: string): Promise<void>

Replacing spawnTool()

Current

spawnTool()

AGENT_LOOP_HOST

isDockerAvailable()

1. Check if sandbox

   loop-{loopId}-{tool}

   exists (created by planner)
2. If yes →

   execInSandbox()

   with auth env vars
3. If no → fall back to

   spawnToolHost()

   (current host-mode behavior)

Troubleshooting

"Not logged in" in sandbox

Auth not injected. Check:

docker sandbox exec my-sandbox bash -c 'claude auth status'
docker sandbox exec my-sandbox bash -c 'cat ~/.codex/auth.json | head -3'

Sandbox creation slow

First pull downloads ~500MB image. Subsequent creates use cached image (~14s). Use

docker sandbox save

File not visible between host and sandbox

Only the workspace path is mounted. Files outside the workspace directory are not shared.

"docker sandbox: command not found"

Docker Desktop must be running. Check version:

docker sandbox version