Browser Automation

Gull Runtime: Browser Automation

Gull is a containerized browser automation runtime that wraps the agent-browser CLI with automatic session isolation and browser state persistence. It exposes a thin HTTP API (/exec, /exec_batch) for executing browser commands.

Container Base Image

• Base: python:3.11-slim-bookworm with uv package manager
• Browser Engine: Chromium (via Playwright)
• CLI: agent-browser (globally installed via npm)
• Node.js: v20 LTS
• Workspace: /workspace (shared Cargo Volume with Ship container)

Architecture

┌──────────────────────────────────────┐
│          Gull Container              │
│                                      │
│  ┌──────────────────────────────┐    │
│  │      FastAPI Wrapper          │    │
│  │  POST /exec                   │    │
│  │  POST /exec_batch             │    │
│  └──────────┬───────────────────┘    │
│             │ subprocess             │
│  ┌──────────┴───────────────────┐    │
│  │      agent-browser CLI        │    │
│  │  + Playwright Chromium        │    │
│  └──────────┬───────────────────┘    │
│             │                        │
│  ┌──────────┴───────────────────┐    │
│  │  /workspace (Cargo Volume)    │    │
│  │  ├── .browser/profile/        │    │
│  │  └── (shared with Ship)       │    │
│  └──────────────────────────────┘    │
└──────────────────────────────────────┘

Execution Model (Critical)

Treat cmd as a single command line that is split into arguments and executed directly. It is NOT a shell script.

What this means

• No shell operators: >, |, &&, ;, subshells, heredocs — none of these work
• No environment expansion: $VAR, ~, globbing (*.png) are not interpreted
• Quote arguments with spaces: fill @e1 "hello world" ✅ vs fill @e1 hello world ❌
• Orchestrate control flow in the agent loop, not inside cmd

Critical Rules (Non-negotiable)

1. Never include agent-browser prefix — Gull auto-injects it. Sending agent-browser open ... becomes agent-browser agent-browser open ... and fails.
2. Never use execute_shell for browser commands — Ship container does not have agent-browser installed.
3. Always re-snapshot after DOM changes — Element refs (@e1, @e2) are invalidated by navigation, form submissions, and dynamic content loading.
4. Never include --session or --profile in cmd — These are automatically injected by Gull for isolation and persistence.

Auto-injected Parameters

Core Workflow: Navigate → Snapshot → Interact → Re-snapshot

Every browser automation follows this cycle:

open https://example.com/form
snapshot -i
# Output: @e1 [input type="email"], @e2 [input type="password"], @e3 [button] "Submit"

fill @e1 "user@example.com"
fill @e2 "password123"
click @e3
wait --load networkidle
snapshot -i    # MUST re-snapshot — previous refs are now invalid

Command Reference

Navigation

open <url>                    Navigate to URL (aliases: goto, navigate)
back                          Go back
forward                       Go forward
reload                        Reload page
close                         Close browser

Snapshot (Page Analysis)

snapshot -i                   Interactive elements with refs (recommended)
snapshot -i -C                Include cursor-interactive elements (divs with onclick, cursor:pointer)
snapshot -s "#selector"       Scope to CSS selector
snapshot -i --json            JSON output for structured parsing
snapshot @e9                  Show subtree under a container ref

Interaction (use @refs from snapshot)

click @e1                     Click element
dblclick @e1                  Double-click element
fill @e2 "text"               Clear field and type text
type @e2 "text"               Type without clearing
select @e1 "option"           Select dropdown option
check @e1                     Check checkbox
uncheck @e1                   Uncheck checkbox
press Enter                   Press key
press Control+a               Key combination
hover @e1                     Hover over element
scroll down 500               Scroll page (direction + pixels)
scrollintoview @e1            Scroll element into view
drag @e1 @e2                  Drag from one element to another
upload @e1 /workspace/file.pdf  Upload file to input element

Get Information

get text @e1                  Get element text
get text body                 Get all visible page text
get html @e1                  Get element HTML
get value @e1                 Get input value
get attr @e1 href             Get element attribute
get title                     Get page title
get url                       Get current URL
get count ".item"             Count matching elements
get box @e1                   Get element bounding box
get styles @e1                Get computed styles

Wait

wait @e1                      Wait for element to appear
wait --load networkidle       Wait for network idle
wait --url "**/dashboard"     Wait for URL pattern
wait --text "Success"         Wait for text to appear
wait --fn "window.ready"      Wait for JS function to return truthy
wait 2000                     Wait milliseconds

Capture (Artifacts)

Write artifacts to /workspace so they are accessible cross-container:

screenshot                           Screenshot to temp dir
screenshot /workspace/page.png       Screenshot to specific path
screenshot --full /workspace/full.png  Full-page screenshot
pdf /workspace/page.pdf              Save page as PDF

Semantic Locators (Alternative to Refs)

When refs are unavailable or unstable:

find text "Sign In" click
find label "Email" fill "user@test.com"
find role button click --name "Submit"
find placeholder "Search" type "query"
find testid "submit-btn" click
find first ".item" click
find last ".item" click
find nth 2 "a" hover

State Checks

is visible @e1
is enabled @e1
is checked @e1

JavaScript Evaluation

eval document.title
eval 2+2

Tabs / Windows / Frames / Dialogs

tab                           List tabs
tab new [url]                 Open new tab
tab 2                         Switch to tab
tab close                     Close current tab
window new                    Open new window
frame "#iframe"               Switch to iframe
frame main                    Switch back to main frame
dialog accept [text]          Accept dialog
dialog dismiss                Dismiss dialog

Storage & Network

cookies                       List cookies
cookies set name value        Set cookie
cookies clear                 Clear all cookies
storage local                 List localStorage
storage local key             Get localStorage value
storage local set k v         Set localStorage value
storage local clear           Clear localStorage

network route <url>           Intercept requests
network route <url> --abort   Block requests
network route <url> --body '{}' Mock response body
network unroute [url]         Remove route
network requests              Show captured requests
network requests --filter api Filter by pattern

Browser State Management

state save /workspace/auth-state.json     Export browser state
state load /workspace/auth-state.json     Import browser state

Security note: State files may contain sensitive tokens. Do not commit or share them.

Ref Lifecycle (Must Understand)

Refs (@e1, @e2, etc.) are ephemeral. They become invalid when:

• Clicking links or buttons that cause navigation
• Form submissions
• Dynamic content loading (dropdowns, modals, AJAX updates)
• Page refresh

Always re-snapshot after any action that may change the DOM:

snapshot -i
click @e1            # may navigate or change DOM
snapshot -i          # MUST re-snapshot
click @e1            # use the NEW page's @e1

Persisting Text Output to Files

Because cmd is not a shell, get text body > page.txt will not work.

Correct workflow:

1. Run a get ... command and capture the returned stdout
2. Write the captured text into a file via filesystem tools (e.g., write_file)

Single vs Batch Execution

Browser State Persistence

Gull automatically persists browser state to /workspace/.browser/profile/ on the shared Cargo Volume:

This state:

• Survives container restarts within the same sandbox
• Is automatically cleaned up when the sandbox is deleted
• Enables "login once, reuse later" within the same sandbox lifecycle

Common Patterns

Form Submission

open https://example.com/signup
snapshot -i
# Analyze output to identify form fields
fill @e1 "Jane Doe"
fill @e2 "jane@example.com"
select @e3 "California"
check @e4
click @e5
wait --load networkidle
snapshot -i    # Verify submission result

Authentication Flow

# Step 1: Navigate
open https://app.example.com/dashboard
snapshot -i

# Step 2: Analyze snapshot
#   → If "Welcome" visible → already logged in → proceed
#   → If "Sign In" form visible → fill credentials:

fill @e1 "user@example.com"
fill @e2 "password"
click @e3
wait --load networkidle
snapshot -i

# Step 3: Verify login result
#   → Success → continue task
#   → Failure → analyze reason and adapt

Data Extraction

open https://example.com/products
snapshot -i
get text @e5                # Get specific element text
get text body               # Get all page text
snapshot -i --json           # JSON output for structured parsing

Screenshot and Cross-Container Processing

# Gull: take screenshot (writes to shared volume)
screenshot /workspace/page.png

# Ship: process it via Python
# execute_python: from PIL import Image; img = Image.open('page.png')

Deterministic Batch

# No intermediate reasoning needed — use batch
commands: [
    "open https://example.com",
    "wait --load networkidle",
    "screenshot --full /workspace/full.png",
    "snapshot -i"
]

Error Reference

Deep-Dive Reference

For complex workflows, conditional branching, file uploads/downloads, debugging flaky automation, and advanced topics (tabs, frames, network interception, proxy, recording), see references/browser.md.