AI Code Provenance: A Buyer's Guide to the Category We Are In
The last of five guides to the categories called 'AI monitoring'. This one asks what AI actually built and whether you can reconstruct why — and it is the one we sell, so read it with that in mind.
Four of these guides describe categories we do not compete in. This one describes ours, so treat the recommendations accordingly — and hold it to the same standard as the others, which is whether it tells you when not to buy.
The question: what did AI actually build for us, and can we reconstruct why?
Who asks it, and when
Engineering leadership, and almost always in retrospect. There are four moments, and none of them is a planning meeting.
- Review: a reviewer faces a diff with no statement of intent. They can see what changed, not what the author was trying to do or what the model was told to avoid.
- Incident: the question is which change caused this and why it looked correct at the time. Git answers the first half. The conversation answers the second, and it is the half that stops the recurrence.
- Onboarding: a new engineer inherits a codebase whose reasoning lives in transcripts that were deleted, on machines they do not have.
- Audit: a customer's security questionnaire asks how AI-generated code is reviewed and retained, and the honest answer is that nobody knows.
Why this one is different from the other four
The other four categories can be adopted late. This one cannot: the artefacts are on a deletion timer set by someone else, and what is gone cannot be reconstructed.
Claude Code removes transcripts older than 30 days by default, and the setting that controls it is absent from a normal install. Gemini CLI defaults to the same 30 days. Cursor keeps everything in one SQLite database with no export, which support threads routinely advise deleting when it breaks. Aider writes its history inside the repository, where it is normally gitignored and disappears with the working copy.
That is the argument for treating this as urgent rather than important. Everything else in this space can be bought next quarter with the same result. Here, next quarter buys you a record that starts next quarter.
How to evaluate one — including us
- Which tools does it actually read, and how? Ask for the on-disk format per tool. Editor extensions cover editors; terminal agents need something that reads their files. Anyone claiming to read Windsurf's Cascade history should be asked how, because it is protobuf with no published schema.
- Does it link the prompt to the diff, or only store the chat? A conversation without the resulting change answers half the question.
- Where does the data live by default, and what does the free tier send? If the answer is not 'nothing', that is a decision you are making for every developer.
- What happens to secrets? Prompts contain tokens and customer names because chat feels ephemeral. Masking should be on by default and should apply to anything shared publicly regardless of local settings.
- Can you leave? Plain formats, exportable. A provenance product that locks the record inside itself has failed at its own premise.
When not to buy anything
If you are one developer, or two, the correct answer is a shell script. Set the retention keys, copy the directories nightly, point the destination at storage that is already backed up. We publish the script and it works. Spend the money elsewhere.
The script stops working for reasons that are not about features: tool formats move underneath it, it fails silently, it covers one machine, the output is unreadable to anyone but its author, and 'each developer has a cron job' is not an answer to a security questionnaire. If none of those describe you, you do not have this problem yet.
The one test that decides it
Ask what happens if the developer who set it up leaves next month. If the record goes with them, you have a personal habit rather than a system of record — which is fine, right up until the moment somebody outside the team asks a question about how your software was built.
Run promptwake doctor before deciding anything. It reports what every AI tool on your machine is holding and how much of it is inside a deletion window, needs no account, writes nothing and sends nothing anywhere. Whatever you conclude about the category, that number is the one worth having.