Paste a raw Server-Sent Events dump from a streaming LLM response. Get every event parsed, the assistant text and tool calls reassembled, and the failure named.
Works with what curl -N prints, what you copy out of the browser devtools EventStream or Response tab, and what your own request logger writes. CRLF and LF line endings, blank-line event separation, multi-line data: fields, event: names, : comment keep-alives, and a trailing data: [DONE] sentinel are all handled. A dump cut off mid-line is expected input, not an error.
Nothing you paste leaves this page. A streamed response carries the user prompt, retrieved documents, and tool arguments, which is exactly the data you should not upload to a stranger's parser, so all of the work here happens in your browser and the page makes no network requests of any kind.
This is wire-level debugging for streaming responses: the layer where "the model stopped mid-sentence" and "my tool call arguments are invalid JSON" actually get diagnosed. From inside an SDK those two failures are nearly invisible, a short string and a JSON.parse error. In the raw event stream they are obvious, and they are usually the same bug.
What you get back:
The Anthropic sample is a stream that was cut mid tool call, so the truncation findings fire. The OpenAI sample is a healthy stream that ends properly, for comparison.
| # | Event type | Bytes | Size | Payload carried |
|---|