How are you preserving context from AI coding sessions during code review?
I’ve been thinking about a gap in AI-assisted PRs.
The review artifact is usually just the final diff, commit message, and PR description. But the prompt, response, tool usage, and intermediate reasoning often stay in the agent UI or local transcript. Once the session is gone, reviewers have to infer intent from the patch.
One approach I’ve been experimenting with is storing commit-level session context in Git notes (refs/notes/...) instead of a hosted service or a separate database.
The data model I’m trying to keep close to the commit is roughly:
- prompt / response pairs
- files touched by the agent
- a rough AI involvement estimate per commit
- bounded context around short prompts
- machine-readable reviewer context
- a way to jump from
git blameto the recorded commit context
This is narrower than broader checkpoint/session-history tools like Entire. I’m mostly interested in PR review and commit-level traceability, not rewind/resume/search across full sessions.
Curious how others are handling this. Do you store agent session context anywhere today, or is the final diff still the only artifact that survives into review?
For context, this is the open-source tool I’ve been building: https://github.com/wasabeef/AgentNote