AI code review needs to identify the version it actually read
Connect AI code review findings to the compared versions, files actually read and checks after fixes.
Article language: English

In this article
"I reviewed the change. Two minor comments, otherwise no concerns." That report might close the matter if not for an awkward question: which change? During the review, the author added validation, replaced a configuration file and addressed one comment. The assistant finished reading, but the code did not stop changing.
This is a hypothetical example. We are interested in evidence of review scope, not an assessment of a particular repository. Here, code review means reading a change and its context to look for problems. It is not the same as running tests or agreeing to a deployment. "I found no errors" has meaning only alongside information about where and in which version the reviewer looked.
A company can ask Aeneia Audit to examine how AI is used and what evidence supports acceptance of its work. A specialist assessment of the correctness or security of the code itself requires suitable expertise and a separate scope. Here we are organising the relationship between the material the reviewer received and the conclusion given to the person making a decision.
One request number can contain several different changes
Suppose a studio is extending an equipment-booking portal. The change should exclude retired devices from new bookings while keeping them in the history. It changes five files: equipment selection, availability checking, booking persistence, configuration and a test. Historical bookings must still recognise devices that can no longer be offered to new customers. That distinction is the focus of the reading.
A pull request, a proposal to change code, can keep the same number all day. Inside it, however, successive saved versions appear. The request identifier is therefore not an exact label for the bytes that were read. The report needs reference points: the starting state and the proposed state assessed.
A delta is the difference between those states. If the starting point is unspecified, "I reviewed the change" might mean the last fix, the entire set since work began, or local files not yet included in a saved version. Each scope can be justified. None should be selected accidentally by a tool's default behaviour.
The git diff documentation distinguishes comparing working files with the staging area, comparing the staging area with a version, and comparing two saved states. The word "diff" alone therefore does not identify the pair of materials assessed.
For the portal example, let A stand for the state before the change and B for the state containing the five changed files. The reviewer should receive the difference from A to B and the necessary context. The letters are labels in this article, not real version identifiers. In a real report they must be replaced by precise information that makes the assessed material retrievable.
When working with local files, a saved repository identifier may not include current edits. A preserved copy or another agreed way to fix the assessed state is then needed. Adding a new identifier after the reading does not make the earlier review cover new files.
A hash in an AI report (in Polish) (in Polish), a digest calculated from data, helps identify specific material. By itself it proves neither that the material was read nor that the conclusion is true. A version label must be accompanied by the scope actually read. These are separate parts of the evidence, even when they appear in adjacent report columns.
Preserve the start of the comparison after updating a branch too. If the base changes, the same final file can belong to a differently defined difference. The report should explain whether the next reading covers the fix from B to C or the whole proposal against the new starting point. Otherwise the reader does not know why two previously visible files suddenly disappeared from the list.
The version should remain stable while it is read or be reproducible later. If the author is still working, a reviewer can read a preserved copy of B while the current proposal moves on to C. That is a reasonable way to work in parallel, provided the report is not later described as a review of C.
A list of five names does not mean five files were read
In the hypothetical report, the assistant lists five names from the change summary. Its detailed comments, however, concern only equipment selection and availability. Did it read the other three files, skip them as irrelevant or never receive their contents? A list of names does not answer that question. It may come from an index rather than from reading the content.
Ask for a simple scope description for each item. For example: the diff was read with the calling function; the entire file was read; a generated file was skipped; content was unavailable; or a tool response was truncated. These states mean different things. They do not require a claim that the reviewer read everything around a small fix.
In our portal, the configuration file determines which equipment categories are visible. If the assistant did not receive it, it should not claim to have checked that retired devices are completely excluded. It may still offer a valid comment on the availability function. The conclusion should be as broad as the dependency actually examined, not as broad as the task title.
GitHub lets a reviewer mark a file as Viewed and removes that mark if the file changes. This tracks progress in the interface. It is not independent proof of the depth or completeness of a model's analysis.
If a tool shows only part of a large file, keep that fact in the report. Do not turn a command invocation alone into "the file was read". The response may have stopped at an output limit. With long AI context (in Polish) (in Polish), there is also a difference between supplying content and using it correctly. For a review, we first need to establish at least the former.

Context outside the changed lines matters too, but it should be selected for the question. A new equipment check may require looking at the place that calls the function. Renaming a parameter requires checking its callers. That does not mean a reviewer must read the whole repository from the first licence file to the last icon.
A record can separate supporting context from the change being assessed. If the reviewer opened project instructions, record which version of the rules was considered. If they did not inspect the code of an external dependency, do not add it to the scope merely because the function uses it. A clear boundary helps the next person direct attention where evidence is still missing.
Treat generated summaries and previews with particular care. A dependency-change preview may not show every change to the raw file. An image or binary file does not become assessed content because a tool recorded its new size. A report can honestly identify such material for a different kind of check.
Aeneia Audit: Let us organise the evidence for an automated review: the material's version, what was actually read and acceptance after fixes. The team can then distinguish comments on an earlier patch from an assessment of the current proposal. Connect an AI review report to a specific change.
Fixing a comment creates the next material to assess
The assistant notices that a retired device can still be selected through a direct URL. The author fixes booking persistence. At the same time, they tidy a function parameter, change a test and add a new category to the configuration. State C is created. Removing the reported problem does not yet prove the remaining changes are correct.
There are therefore two questions in the next acceptance check. Was the particular problem in B removed? What else appeared between B and C? Limiting the check to the first can miss a defect introduced during the fix. Reviewing all of A to C without recognising the previous comment can, conversely, lose track of what especially needed confirmation.
GitHub's description of creating a review through its API allows the identifier of the reviewed commit, a saved state of code, to be specified. The documentation warns that a later change to the indicated line can make a comment outdated. The location of a comment is tied to a version of the material.
Do not treat a shifted line number as resolution of the problem. A comment can become outdated in the interface while the defect still exists elsewhere. It can also remain visible even though the author fixed a dependent function. The person accepting the work should address the substance of the comment and current behaviour, not merely the appearance of the thread.
A result of rereading might say: "Reviewed the difference from B to C in three files and reread the call to the availability function. The other two files match copy B." This is an example of report wording. It may be used only after an actual comparison, not because the author says "nothing else changed there".
Unchanged content does not, however, remove an earlier gap. If configuration in B was not read, saying "unchanged from B" still does not confirm its review. Fill that gap or keep it as an open limitation. In our example, adding a configuration category also requires understanding its earlier rules, not reading only the new line without its dependencies.
The Copilot code review instructions say that a new push does not automatically trigger another review without the appropriate configuration. A rereview can be requested manually or reviews of new pushes can be enabled. A further commit appearing does not by itself mean a new review took place.
For automation, preserve the version to which a completed run applies. A process can start for B, finish after C is published and still correctly report on B. The problem arises when a summary attaches that assessment to the current version merely because both events occurred in the same request.
If the review runs in a CI process for a pull request (in Polish) (in Polish), separately define which instructions and content it uses. CI is an automated run triggered by a repository change. Evidence of what was read does not settle whether the environment had the right access boundaries. Reading scope and environmental access require separate answers even if the tool displays one result.
After files are renamed, preserve the link to the earlier scope. Disappearance of an old path does not necessarily mean removal of the behaviour; code may have moved. The reviewer should find where the fix went and check the necessary callers. A report saying "old name absent, problem closed" could miss the same defect in a new module.
What can be accepted from such a report?
A report can honestly say it found no problems within the described scope. It does not prove that no errors exist. Nor does it automatically confirm tests were run. Evidence of tests run with AI (in Polish) (in Polish) includes the command, conditions and result; reading code is a different activity. Both can inform the final decision, but they should not impersonate each other.
Instead of collecting everything under the word "checked", leave a short result with separate boundaries:
- The version and comparison covered by the review.
- The files and necessary context actually read.
- Problems found and how they were later resolved.
- Omitted or unavailable material and the effect of that limit on the conclusion.
- Separate evidence of tests, if tests were part of the task.
That record also helps when accepting code and documentation from a provider (in Polish) (in Polish). It links the assessment to the material the company actually receives. A branch name, meeting date and number of comments may help orientation, but they do not replace that link.
Before deciding, compare the current proposal with the last accepted scope. If an equipment-category configuration was added, send it for reading and retain the earlier report as an assessment of B. If the next version concerns only an agreed fix, describe the check of that fix and its dependencies. There is no need to stage a new large review of unchanged files. There is a need to stop pretending an old report read the future.







