Add two numbers together and return the result.

Evaluate a math expression (e.g., '7*8+3').

Start an interactive CATGPT terminal session.

Send a message in the current conversation.

Send a message in a specific thread. Navigates to it first.

Get the browser context.

OpenAI-compatible chat completions endpoint. Converts the message array into a single prompt, sends it to ChatGPT via browser automation

OpenAI-compatible image generation endpoint. Sends the prompt to ChatGPT which uses DALL-E to generate images. Downloads the generated i

OpenAI Responses API endpoint — compatible with Codex CLI. Accepts the Responses API format (flat tools, `input` field, `instructions`),

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

Get the current date and time. Returns ISO format datetime string.

Return the shared Stealth instance (for use with launch helpers).

Unauthenticated health-check for Docker / load-balancers.

Startup: launch browser. Shutdown: close it.

List available models — returns our single browser-backed model.

List recent conversation threads from the sidebar.

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

Method 2: Set clipboard and paste via Cmd+V.

Method 4: Set innerHTML + dispatch input event.

Method 5: document.execCommand('insertText') — fires proper input events.

Method 3: Playwright's fill() — uses CDP Input.insertText.

Method 6: Use React's internal value setter to update state.

Return the product of three numbers.

Navigate to an existing conversation thread.

Start a new conversation. Strategy order: 1. SPA button click (avoids DNS issues, preserves browser state) 2. JavaScript loca

Start a new conversation and send the first message.

Re-inject stealth JS when the main frame navigates.

Inject stealth into new pages and attach navigation listener.

Get the active page. Raises if browser not started.

Reverse the input string.

Return a fake Wikipedia summary for the query.

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

Health check — returns login status and current thread.

Error case: missing auth token should return 401.

Non-streaming: metadata and previous_response_id are accepted and echoed.

Non-streaming: multi-turn conversation in input array.

Test 1: Generate a single image via raw HTTP request (b64_json).

Codex CLI: simple coding question via streaming (no tools needed).

Non-streaming: simple string input.

Codex CLI: model should call the shell tool for a system task.

Non-streaming: message array input (role + content).

Test 2: Generate an image with a different prompt and verify save.

Codex CLI: full tool loop. Turn 1: User asks → model calls shell tool Turn 2: We send function_call_output → model gives final answer

Non-streaming: instructions field (system prompt).

Test 3: Generate an image with response_format='url'.

Codex CLI: model calls apply_patch to edit a file.

Non-streaming: developer role in input (should map to system).

Test 4: Use the official OpenAI Python SDK.

Codex CLI: streaming SSE events for a tool call response.

Non-streaming: input with content parts using input_text type.

Codex CLI: both shell and apply_patch tools available.

Streaming: validate complete SSE event sequence.

Codex CLI: tool_choice=none forces text response even with tools defined.

Streaming: validate individual event payloads.

Codex CLI: function_call_output with error — model should handle gracefully.

Non-streaming: tools are accepted and echoed in flat format.

Error case: empty input should return 400.

Simulate a 'thinking' pause before the user starts typing.

Return a fake weather forecast for a city and date.