Claude Code and Codex are powerful, but longer sessions can become surprisingly difficult to follow.
The terminal quickly fills with:
- tool calls
- shell commands
- retries
- errors
- test output
- file changes
- long explanations
By the end of a session, the important questions are often still unclear:
What actually changed? Why did it change? What failed? What should I review?
That is why I built Lucius.
Lucius is an open-source tool for exploring coding-agent sessions in a clearer, more structured way. Instead of scrolling through a long terminal transcript, you can inspect the session and understand the agent’s work more quickly.
The goal is not to hide the raw output or blindly summarize everything. It is to make it easier to move between the high-level result and the underlying details.
I am especially interested in improving the experience around:
- understanding the sequence of actions
- seeing which files were touched
- spotting failures and retries
- reviewing important decisions
- identifying what still needs human attention
Lucius is still early, so feedback from people who regularly use Claude Code, Codex, or similar coding agents would be genuinely useful.
How do you review longer agent sessions today? Do you read the full terminal output, inspect only the final diff, or use another tool?
GitHub: https://github.com/berkantay/lucius
Feedback, issues, contributions, and stars are appreciated.
Top comments (1)
The high-level-to-raw transition is the right design target. I would make the review unit a run manifest rather than a summary: base commit, changed-file digest, tool calls grouped by phase, test commands and exit codes, retry count, and an explicit unresolved-items list. That lets a reviewer jump from “what changed?” to evidence without trusting a generated narrative. Would Lucius treat a session as complete only after a verifier records those fields, or is completion intentionally just a view-layer concern?