Skip to main content
← All posts
A Southeast Asian software engineer at a dark wooden desk at night reading a long AI coding assistant transcript on a laptop, warm coral lamp glow over the keyboard, editorial over-the-shoulder composition in a Monocle-style color palette.
·7 min read

How to Find and Read Roo Code Chat History: A Step-by-Step Guide

Where Roo Code stores task transcripts on macOS, Windows, and Linux, how to export them to markdown, and how to search every past agent session.

Roo Code is one of the fastest growing agentic coding extensions for VS Code, forked from Cline and shipped under its own name since early 2026. If you have been using it for a few weeks, you already know the problem. The chat panel on the right side of your editor holds everything the agent did on a task, every file it touched, every tool call, every reasoning step. Then you close the task, open a new one, and that transcript becomes almost impossible to find again. This guide walks through where Roo Code stores your chat history on disk, how to read a specific task without the IDE loaded, and how to turn those transcripts into real documents you can archive, search, and share.

Where Roo Code stores task transcripts

Roo Code writes every task to a per-task folder inside the VS Code global storage directory. On macOS the path is ~/Library/Application Support/Code/User/globalStorage/rooveterinaryinc.roo-cline/tasks/. On Windows it lives under %APPDATA%\Code\User\globalStorage\rooveterinaryinc.roo-cline\tasks\ and on Linux the folder sits at ~/.config/Code/User/globalStorage/rooveterinaryinc.roo-cline/tasks/. If you installed the extension inside Cursor instead of VS Code, swap Code for Cursor in the path and the layout is identical. Inside each tasks/ directory you will see a flat list of folders named with ULID timestamps, one per task you have ever started.

Open any single task folder and you get three files that matter. The first is api_conversation_history.json, which contains the raw message array sent to the model, including system prompts, tool call results, and the full assistant reasoning. The second is ui_messages.json, which is what Roo Code renders inside the chat panel, trimmed and formatted for humans. The third is task_metadata.json, which stores the task name, token counts, cost, and timestamps. For most reading sessions ui_messages.json is the one you want, because it already has the thinking blocks and tool outputs collapsed into readable chunks.

Why the in-IDE panel fails for long transcripts

The chat panel inside VS Code was built for in-flight work, not for review. Scrollback is tied to the task you have active, so switching tasks forces a reload. Code blocks render in a cramped sidebar that is usually 400 pixels wide, which wraps long function signatures into unreadable zig-zag lines. Reasoning blocks collapse by default and the expand toggle resets every time you close the panel. If an agent ran for an hour across forty tool calls, scrolling back to the decision point where it chose the wrong file is painful.

The deeper problem is that your past Roo Code sessions are often more valuable than the output they produced. The agent explained why it rejected an approach. It flagged an edge case you had not thought of. It wrote a plan you want to lift into a design doc. Those artifacts need to leave the IDE and become durable, the same way you would archive a long reasoning trace from a frontier model. We cover the general pattern in how to read AI reasoning traces and the IDE-specific version in reading Cursor chat history outside the IDE.

Exporting a Roo Code task as portable markdown

The cleanest way to get a Roo Code task into a readable format is to convert the ui_messages.json file into markdown. Each message object has a type, a say or ask field, and a text body. A short script can walk the array and emit a markdown document with headings for each turn, fenced code blocks for file edits, and callouts for tool results. The VS Code extension marketplace also has a community exporter called "Roo Code Task Export" which handles this automatically, though it ships markdown with a few quirky escape characters you will want to clean up afterward.

Once you have markdown, you want a reader that respects the structure. Open the exported file in Prism MD and the long tool call chains render as collapsible sections, the embedded Mermaid plan diagrams render as pictures, and the code blocks use a monospace font sized for reading rather than for autocomplete. If your task included math or LaTeX from a planning phase, that renders too. The point is to stop treating the transcript as a chat log and start treating it as a document, which is the whole argument behind why AI-generated markdown deserves better typography.

Searching across every task you have run

Because Roo Code dumps each task into its own folder with predictable filenames, you can search your entire history with ripgrep in one command. From the tasks/ directory, run rg -l "the-thing-you-remember" . and you get a list of task folders that contain the phrase. This works for file paths, function names, error messages, or any plain string the agent typed. Combine it with jq to pull the task title from task_metadata.json and you have a usable local search index without any extra tooling. For a deeper walkthrough of this pattern applied across every AI tool you use, see how to search across all your saved AI conversations.

The patterns below are the ones I keep in a shell alias file after a few months of heavy agent use. They assume you are already inside the tasks/ folder in a terminal, and they lean on tools that any developer machine already has installed. None of them touch the extension itself, so you can run them while Roo Code is open without breaking the active task. Treat the list as a starting kit rather than a complete reference, since the schema does drift between releases.

  • Finding every task where the agent touched a specific file: rg -l "path/to/file.ts" ~/Library/Application\ Support/Code/User/globalStorage/rooveterinaryinc.roo-cline/tasks/
  • Listing tasks by cost, highest first: jq -r '[.tokensIn, .tokensOut, .totalCost, .taskName] | @tsv' */task_metadata.json | sort -k3 -n -r | head
  • Pulling the first user message from a task: jq -r '.[0].text' ui_messages.json
  • Counting total tasks per month: ls -1 | awk '{print substr($1,1,6)}' | sort | uniq -c

Keeping task history across machines

Roo Code has no cloud sync, which is a feature if you care about privacy and a problem if you work on two machines. The simplest fix is to turn the tasks/ folder into a git repository. Add a .gitignore for the embedded cache files, commit the JSON transcripts, and push to a private repo. You get version history, cross-device sync, and a backup, all for free. The one gotcha is that long transcripts can hit single-file limits on some git hosts, so split by month if you do heavy agent work. If you prefer a non-git approach, pointing a Syncthing folder at the tasks/ directory works equally well without any central server in the loop.

FAQ

Does Roo Code delete old tasks automatically?

No. The extension keeps every task indefinitely unless you manually clear it from the UI or delete the folder from disk. This is why the tasks/ directory can grow to several gigabytes if you use the agent heavily, mostly from large tool outputs stored as text. If you want to prune without losing the good runs, sort by task_metadata.json cost and keep only the top ten percent, since those tend to be the long multi-hour sessions worth revisiting.

Can I read a Roo Code task on my phone?

Not directly from the JSON files, but if you export to markdown and sync the file to a cloud folder, any markdown reader on iOS or Android can open it. Prism MD on mobile renders the long code blocks without horizontal scroll traps. The export step is a one-off per task, so pair it with a shortcut that fires whenever a task finishes if you want it to be hands-off. For long sessions on a plane, pre-sync the exports the night before so the whole archive stays readable without a connection.

What is different between Roo Code and Cline chat history?

The folder path differs because of the extension ID, but the JSON schema is nearly identical since Roo Code forked from Cline. Scripts that parse one will usually work on the other with a path change. The one meaningful divergence is that Roo Code added richer tool metadata in mid-2026, so older Cline exports will be missing a few fields. If you maintain a parser for both, treat the newer fields as optional and default to the Cline shape.

Is the raw API conversation history worth reading?

Only for debugging. It contains the full tool call schema, system prompts, and intermediate state, which is useful when you want to understand why the agent did something strange. For normal review, stick with ui_messages.json because it strips the noise. The exception is cost audits, where the API history is the only place that records every token the model saw rather than every token it rendered.

Read your Roo Code transcripts the way they deserve

Free to start — no credit card.

Open Prism MD free →

Related reading

Ready to read your own AI documents?

Open ChatGPT, Claude, Gemini, or any markdown file in the reader built for the way models write.

  • ✓Renders code, math & Mermaid out of the box
  • ✓Works offline once you've opened a doc
  • ✓Free forever for personal reading