Accessing the Full Agent Transcript in Cursor

Support Help
,
1

Hi,

I was going through the agent transcripts in Cursor (the JSONL files) and noticed that they only contain the tool inputs, not the outputs (also what does [REDACTED] mean?). I wanted to check if there is a way to get the full trace of the agent, including the tool outputs.

My use case is a bit unusual. I want to analyse a few markdown files and process them directly in Cursor’s agent chat, but I also want to allow Cursor to look back at the agent transcripts for every document it processed, and use those to come up with a script for a repeatable workflow or pipeline that it can then code up, without any further chatting. Essentially, the idea is to have it learn a pipeline from its own chat history.

The reason I was looking for the tool outputs is so that in a new session, I can reference those transcript files and ask Cursor to analyse them and design a repeatable script based on all the nuances and edge cases it encountered while processing the documents manually.

I understand I could just ask Cursor to write the script directly from the raw files, but I don’t think that approach is very efficient. I already tried it and the output wasn’t great. The core idea is that I want Cursor to actually have the experience of processing the files itself, acting as a proxy for a human doing the task manually, and then distil everything it did into a repeatable, generalised script or AI pipeline. Or even better, a full skill package with scripts, references, and supporting material. Writing a script from raw files skips that experiential step, which I think is where a lot of the nuance gets lost for my use-case.

The alternative I am considering is building my own agent harness using the Claude Agents SDK or something similar, and logging everything from there. But I wanted to first see if I could make it work with Cursor’s agent directly, because there is little point spending time on a custom agent setup when Cursor is already right there.

Any ideas for this use case are welcome, whether that is something native to Cursor, using the ACP, the CLI, or any other workaround for capturing the full trace. Very open to alternatives here.

Thanks!

4

Hi @nerdimite , I filed a bug on this a while back – Transcripts no longer exported in full . Unfortunately it doesn’t seem to have been solved yet.

5

Hey @nerdimite and @rsb!

The JSONL transcript files include your messages, the assistant’s text responses, and tool call inputs (tool name + arguments). They intentionally do not include tool call outputs, which can be very large.

For your use case of capturing full tool outputs, you could use a postToolUse hook to log them yourself. Hooks receive context about each tool call and can write to files, so you could build your own full-trace log alongside the existing transcripts!

[REDACTED] comes from the model provider (Anthropic/OpenAI), not Cursor. Some models return “extended thinking” / chain-of-thought reasoning, but the provider redacts portions of it before sending it back to us. When that happens, Cursor records [REDACTED] as a placeholder.