e2e Files

Minor Changes

• #762 83c17e2 Thanks @okwasniewski! - Anonymous telemetry covers e2e mcp and e2e explore. Each MCP session sends one e2e_mcp_session event when it closes or fails to open: the client's self-reported name and version, the platform and engine, how it ended, and its tool calls counted by the runner's own tool names with the error codes they failed with. e2e_run_completed adds command, counts and option ids of the config features the run used, and for e2e explore why it stopped and its steps and findings by kind and severity. Beyond the client's own name, no project names, paths, tool arguments or results, goals, or findings are sent. The notice shows again once; E2E_TELEMETRY_DEBUG=1 prints every event, and the opt-outs are unchanged.

Patch Changes

• #760 b3c9832 Thanks @devin-ai-integration! - e2e models openai asks the ChatGPT backend for its model list as Codex CLI 0.160.0 (was 0.155.1). The backend hides models from clients it considers too old, so models unlocked since 0.155.1 now show up and can be passed to chatgpt().

• #767 5609f5e Thanks @okwasniewski! - e2e models github-copilot marks the models copilot() cannot call: not enabled for a model your plan has not turned on, and no chat completions for a model Copilot serves only over other APIs such as Responses. These used to be listed like any other model and then failed the first agent step with "The requested model is not supported".

• #579 c6c39ad Thanks @okwasniewski! - Locator matchers honor Playwright's { checked: false }, { enabled: false }, { visible: false }, and { attached: false }, which flip the matcher, and { ignoreCase } on toHaveText, toContainText, toHaveAccessibleName, and toHaveAttribute(name, value). Before, these were ignored at runtime, so toBeChecked({ checked: false }) passed on a checked box and not.toContainText('error', { ignoreCase: true }) passed on Error. Any other option key, or a flag that is not a boolean, is now INVALID_ARGUMENT before the first read, from JavaScript too.

e2e/engine exports rejectUnknownOptions(api, options, known, code?), the check the runner's own fixtures use, so an engine refuses an option its fixture methods and matchers do not take the same way.

• #759 0cda95d Thanks @devin-ai-integration! - e2e mcp opens a session in a project without the optional ai package. Without the AI SDK, the catalog and the argument checks read each tool's Standard Schema (zod's), so a deterministic-only project (no agents, e2e init with no model gateway) can open_session, observe, locate, and drive the page.

• #727 dec903b Thanks @okwasniewski! - A registered secret value passed as a plain string is now redacted in more places: test and describe titles (and the test ids, reporter output, and artifact and failure-page file names derived from them), step labels, agent.act and agent.assert instructions and params as executors and the model receive them, agentContext, judgment prompts, and an e2e explore goal. Error messages, observations, and worker console output already masked it. A test whose title spells out a secret gets a new id, so --last-failed from an earlier run does not select it once. Not covered: a value only a secret provider returns stays in titles, and a Playwright trace of a session no secret reached is kept as recorded.

• #763 b573756 Thanks @okwasniewski! - Installing e2e pulls in 29 packages instead of 117 and takes about 31MB instead of 36MB. e2e mcp now runs on the split MCP SDK (@modelcontextprotocol/server 2.2.0) in place of @modelcontextprotocol/sdk, which brought in express, hono, and the rest of an HTTP server stack that stdio never used. The server keeps the same protocol version, so existing clients connect as before. Packages are built and published without sourcemaps, which pointed at a src/ that was never shipped. Stack traces show dist/ positions.