Reviewing a day of AI-assisted work
An AI session may keep running after you move to another task. Counting its whole span as your working time makes a report look busy and hides what you did. Hina keeps your prompts, your computer activity and your commits apart so each can answer its own question.
What Hina reads from your AI tools
Prompt history is off until you turn it on. In the Privacy tab, switch on Read agent sessions and Hina reads the logs that Claude Code, Codex, Cursor and Antigravity keep on your Mac. From them it takes your prompts, the time of each session, its working directory and branch, and the file paths that editing tools touched. It does not read the assistant's replies or the output of tools.
The prompts then appear in two places. The Day view lists What you asked AI, and the daily report's Evidence puts each prompt under the project it ran in. Sessions that ran outside any repository are gathered in a separate section called AI sessions outside repositories.
Three clocks that answer different questions
Recorded computer activity says when you were at the keyboard. An AI session's span says when the tool was working, and that span can continue while you are in a meeting or reading something else. A commit timestamp says when a change was written to the local repository. Hina keeps these separate. A prompt contributes no activity duration of its own, and in search results a prompt is labelled AI prompt and carries no time spent.
The practical consequence is that a long session does not make a long working day. If a commit lands during a stretch when you were away, you can say the two overlap, and that is all you can say. It does not show who wrote the change, and it does not turn the absence into work. Hina checks only the local repositories you chose. It does not look up pull requests, issues or releases on the remote.
A short routine for reviewing the day
Open the day's report, expand Evidence, choose a project, and read three kinds of evidence in order. Start with its Local commit record, since it shows changes that exist. Then read What you asked AI, which shows what you wanted. Last, read Files touched for the size of the change. When a project still has uncommitted work or untagged commits and no commit followed your latest prompt within half an hour, the report adds that prompt under In progress and unconfirmed with a line such as “Latest visible request, outcome unconfirmed”, and it does not infer completion from the request. In the sample report on the Hina site, a prompt about permission text being cut off in a narrow welcome window sits next to a commit with a matching message, and that pairing is what you are looking for.
Version tags are handled the same way. A local tag confirms that the tag was created. It does not prove a release, an installer or a live website. For each prompt, ask whether a commit, a file or a delivery channel confirms the result, and mark the rest as still to check.
If you want a count of how much AI was used, there is a separate switch, AI usage metadata, in the same Privacy tab. It shows Claude Code and Codex models and token counts by project in an AI in your work section of the daily report. The report states that this is recorded usage and not time worked, completed tasks or a bill, and input includes cached tokens, so do not add the cache figure on top.
Finding an earlier prompt
When you remember a phrase from a request but not the day, use Search records. It searches titles, URLs and the prompts you sent, over the last 7, 30 or 90 days, and several keywords must all match. A search such as yesterday followed by a project name works, and so does a date or a few words from the prompt. Double quotes search for the exact phrase.
Each hit shows the records within ten minutes before and after it, labelled as nearby records that are not necessarily related. Treat them as context for jogging your memory and nothing more. Being close in time does not mean one thing caused the other.
Use the result to improve the next request
Pick one task that needed rework and compare your first prompt with what you finally accepted. The gap is often something you knew and did not say, such as the input format, the range of supported versions, or how you would check the work. Those are worth more attention than the number of sessions you opened.
Carry the requirements you confirmed into the next task description. A line in your prompt about what done looks like and how it will be checked can save a round of correction. Hina can bring back the evidence, but deciding how to work with AI still depends on what you know about the task.
Limits to keep in mind
A prompt shows your intent and nothing else. The report cannot know whether the code ran, whether the tests passed or whether the answer was right, because it never reads the replies. Missing logs also count as unknown. If a source could not be read, the report says so and does not guess.
The report is plain Markdown that you can copy or export. If you paste it into another AI tool to ask about the day, the text is then governed by that tool's privacy practice and no longer by Hina's. Share the day you need, not a whole history.