Configure and return a logger that writes to file (and optionally console). Args: name: Logger name. log_file: Optional file

Disconnect the MutationObserver.

Send a message to Claude and wait for the complete response. Args: text: The message text to send. image_pat

Gracefully close the browser context and playwright instance.

Check if user is logged in by looking for chat input vs login indicators. Returns True if the chat interface is visible, False if lo

Start a new conversation by navigating to /new.

POST to /v1/responses. Returns parsed JSON or raw response for streaming.

Validate that a response object has all required fields with correct types.

Return the target URL for the active provider.

Extract text from message content (handles both string and list format).

Return metadata for the latest assistant turn in Claude.ai. Claude wraps each assistant response in a div with data-is-streaming attribute.

Navigate to a URL and wait for page load.

Sleep for a random duration (milliseconds).

Inject a MutationObserver that logs DOM changes to the console. We capture those via page console listener.

Launch a persistent Chrome context with stealth and human-like settings. Automatically cleans up stale lock files from previous cras

Return metadata for the latest assistant turn (article ordered). signature format: "<article-index>:<stable-id>" where stable-id is best

Clear captured requests.

Rough token estimate (~4 chars per token).

Check the current state of the textarea and send button.

Extract latest assistant response by clicking copy on the latest turn.

Heuristic: true when text looks like transient "thinking/searching" UI status.

Convert internal ChatResponse to API ChatResponse with image data.

Check if the current page shows a browser or ChatGPT error.

Enforce cooldown between messages and start new chat if thread is full. ChatGPT's web UI degrades after ~6-8 messages in a thread (stops gene

Extract the thread/conversation ID from the current URL.

Try each selector in the fallback list. Return the first one that matches.

Get or create the global request lock (lazy init).

Read a local image and return an OpenAI-compatible base64 data URL.

Increment the thread message counter after a successful response.

Inject stealth JS into the current page via evaluate().

Wait for the chat input to become visible and interactive.

Count assistant turns (article based, newest-UI friendly).

Create required directories if they don't exist.

Extract latest assistant response by clicking copy on the latest turn. Never intentionally copies from previous_turn_signature when provided

Simulate idle mouse movement — small random movements to look alive. Call this periodically while waiting for responses.

Heuristic: true when text looks like transient thinking/analyzing status.

Save a base64-encoded image to disk. Returns the file path.

Flatten an OpenAI-style message array into a single prompt string that we can paste into ChatGPT's input box. The browser already mainta

Build a system-level instruction that tells the model about available tools. *tool_choice* controls how insistent the instructions are:

Count assistant turns that currently expose a copy button.

Check for and dismiss any blocking dialogs/overlays on the page.

Download / decode a file (image, PDF, etc.) from URL, base64 data URL, or a file attachment dict. Returns the local file path.

Dump info about all conversation turns for debugging.

Try each selector in the fallback list. Return the first one that matches.

Return up to *n* image files from IMAGE_DIR.

Try to parse tool calls from the model's response text. Uses robust brace-matching extraction (handles nested JSON, arrays, etc.) then v

Wait until latest assistant-turn signature differs from previous one.

Wait until latest assistant-turn signature differs from previous one.

Count assistant turns (based on data-is-streaming divs).

Click an element with human-like behavior: 1. Hover over element (triggers mouseover) 2. Brief pause 3. Click Uses .first to han

Insert text into a contenteditable input field. Uses execCommand('insertText') which fires proper beforeinput/input events that ProseMir

Normalize extracted assistant text for validation and comparisons.

Normalize extracted assistant text for validation and comparisons.

Parse SSE event stream text into list of {event, data} dicts.

Recover from DNS / page errors by re-navigating to ChatGPT. Tries JS navigation first (avoids DNS lookup), then page.goto(). Returns

Send a series of messages and collect results. Returns thread_id.

Validate a message output item.

Wait until ChatGPT finishes generating the current response. Uses latest-turn alignment to avoid returning stale previous-turn output.

Wait until Claude finishes generating the current response. Primary signal: data-is-streaming attribute changes from "true" to "false".

Docker-safe stealth: evaluate JS on every page and listen for navigations. add_init_script() is broken in Docker (kills DNS), so we: 1.

Build a full ResponseObject from the model output.

Check if the page is showing an error state (DNS failure, crash, etc.). Returns error description string if an error is detected, None otherwise.

Remove stale lock / journal / WAL files that prevent browser launch. After a crash, Chromium leaves behind: - SingletonLock/Socket/Cooki