Try to click the send button using selector fallbacks.

Try to click the send button using selector fallbacks.

Count assistant turns that currently expose a copy button.

Extract file attachments from message content. Supported content part format: {"type": "file", "file": {"filename": "test.pdf", "data"

Extract any text content from the latest turn (for image responses). Image turns may contain a title/description like: "Crea

Extract image URLs from message content (OpenAI vision format).

Extract a JSON object containing *anchor* key from *text*. Uses two strategies: 1. Look inside markdown code blocks (```json ... ```)

Extract the thread/conversation ID from the current URL.

Fallback extraction: innerText from latest assistant turn only.

Fallback extraction: innerText from latest assistant turn only.

Read any file and return its base64-encoded content.

Return test files (non-image) from the downloads directory.

Return model ID based on active provider.

Pre-resolve key domains via the OS and return a --host-resolver-rules string for Chrome. Chrome's built-in DNS client (even with --disab

Convert Responses API `input` (string or item array) into a list of ChatMessage objects compatible with our existing _build_prompt(). Ha

Convert flat Responses API tool definitions to nested Chat Completions ToolDefinition format so we can reuse _build_tool_system_prompt().

Monitor network requests, WebSockets, and JS errors for debugging.

Yield SSE events for a streaming Responses API call. Since the browser backend doesn't truly stream, we emit the full response as a burs

Upload files (images, PDFs, docs, etc.) to ChatGPT's input area. ChatGPT has a hidden <input type="file"> that accepts various file

Upload files to Claude's input area.

Wait for copy button to appear on the latest assistant turn.

Wait for either copy-button readiness or generated image on the latest turn. Returns "copy", "image", or None if timed out.

Wait for Claude's data-is-streaming attribute to become "false". This is the primary and most reliable completion signal for Claude.

Wait for stop button appear -> disappear cycle.

Last resort: poll latest assistant-turn text and wait until stable. If previous_turn_signature is provided, ignores stabilization on that ol

Last resort: poll latest assistant-turn text and wait until stable.

Apply stealth patches to a browser context. In Docker: uses page.evaluate() + navigation listener (safe for DNS). Outside Docker: uses t

Apply stealth patches to the browser context. Must be called AFTER the first page navigation, not before. In Docker containe

Connect to the already-running browser via CDP.

Check the last conversation turn for generated images. ChatGPT DALL-E image responses do NOT use data-message-author-role. Instead, imag

Download an image from a URL using the browser's fetch API. Uses the browser context so cookies/auth are preserved (required for OpenAI-

Check if the user is logged in. If not, guide them through login. This replaces the need to manually run `scripts/first_login.py`. Opens

Full pipeline: detect images in the last response, download them, and return ImageInfo objects with both URLs and local paths.

Return all captured request entries.

Get the current page URL (contains thread ID if in a conversation).

Return signature for the latest assistant turn, if available.

Return signature for the latest assistant turn, if available.

Scrape the sidebar for recent conversation threads. Returns a list of dicts: [{id, title, url}, ...]

Entry point.

Method 1: keyboard.type() — sends keyDown/keyPress/keyUp per char.

Navigate to an existing conversation thread.

Called by server.py to inject the client instance.

Called by server.py to inject the client.

Start recording network requests and responses.

Stop recording (note: Playwright doesn't support removing listeners easily).

Disable all console log handlers (call before any setup_logging).

Test tool calling with add_numbers.

Test complex multi-tool scenarios.

Test sending non-image file attachments (e.g. PDF).

Test sending images via the OpenAI vision content format.

Test a single input method: type text, check state, click send, check network.

Test that /v1/models returns our model.

Test a simple chat without tools.

Test tool/function calling with get_current_time.

Validate test result. Returns (passed, reason).

Check if the response meets expectations. Returns (passed, reason).

Validate usage object.

Wait for a backend-api/conversation POST and return True if seen.

Clear Chrome's in-memory DNS host cache via chrome://net-internals.

Check if the newest assistant turn (not previous turn) contains an image.

Handle browser console messages from our observer.