---
title: "A review comment on code your pull request did not change | Proofline"
description: "When an AI reviewer flags a problem that predates your pull request, ask whether your change widens it."
url: https://proofline.sh/learn/ai-review-comment-on-unchanged-code/
---

[Learning center](/learn/)

# A review comment on code your pull request did not change

When an AI reviewer flags a problem that predates your pull request, ask whether your change widens it. If the pull request copies the defect to new code or gives it new callers, fix it now. If the pull request only moves or borders the code, open a follow-up issue that someone owns and link it from the thread. Dismiss the comment only when the claim is wrong or the problem is already tracked.

Code review The Proofline team November 22, 2026 4 min read

## Check what your diff did to the code

A comment can sit on a changed line and still be about unchanged behavior. Diff tools show moved code as deleted in one place and added in another, so a reviewer sees new lines where the logic is old. Before you decide, find out which of these your change did:

-   moved or reformatted the code without changing what it does;
-   added a new caller, entity or code path that runs the old code;
-   copied the old pattern into new code.

The first leaves the risk where it was. The other two make it larger, and the pull request that makes it larger is the cheapest place to fix it.

## A pre-existing problem the pull request extended

In September 2026, a Home Assistant pull request moved Hikvision's device-availability logic and its pyhik callback registration from the binary sensor and event platforms into a shared `HikvisionEntity`, so that camera entities would also reflect the device connection. Copilot flagged the registration: "Unregister this callback when the entity is removed." [Review thread](https://github.com/home-assistant/core/pull/182306#discussion_r4014934509)

The author agreed the problem was real and said it was older than the change: "it is pre-existing (binary sensor and event have registered without unregistering since the config-entry rewrite; this PR moves that and adds the camera)." The library had no way to unregister a callback, so the author opened a pyHik pull request to add `remove_update_callback()` and planned a follow-up. [Author's reply](https://github.com/home-assistant/core/pull/182306#issuecomment-5682615464)

The diff supports both halves of that reply. The removed lines in `binary_sensor.py` call `add_update_callback` in `async_added_to_hass` and never remove the callback. The new `entity.py` moves that registration into the base class, so every Hikvision entity runs it, cameras included. [Pull request diff](https://github.com/home-assistant/core/pull/182306/files) The pull request did not create the leak, but it gave the leak a new kind of entity.

On 24 September the pyHik maintainer reported that 0.4.7, with the new method, was on PyPI. The author then changed course: "Bumped to 0.4.7 and wired it up here in 5a49c52 rather than a follow-up". The commit calls `remove_update_callback()` from `async_on_remove` and adds a test that removing the camera entity unregisters exactly the callback it added. [Author's update](https://github.com/home-assistant/core/pull/182306#issuecomment-5807028685) A maintainer merged the pull request the same day, 24 September 2026.

The comment deserved a code change, and a test, because the pull request extended the defect. Deferring would have been reasonable only while the library offered no fix.

## A pre-existing problem the pull request only moved

In May 2026, a vLLM pull request moved the AXK1 model's weight loading to `AutoWeightsLoader` and relocated several methods. Gemini Code Assist commented on one of them, saying that `get_expert_mapping` was inconsistent with `AXK1Model.load_weights`: it ignored shared experts when ROCm AITER fusion was enabled and hardcoded `num_redundant_experts` to 0. [Review comment](https://github.com/vllm-project/vllm/pull/41901#discussion_r3199155434)

The author replied that the inconsistency was pre-existing, not introduced by this pull request. They traced it to an earlier pull request and noted that `DeepseekV2ForCausalLM` had the same issue, at line 1472. [Author's reply](https://github.com/vllm-project/vllm/pull/41901#discussion_r3200037201)

Both statements check out against the diff. The `get_expert_mapping` body that the pull request removed and the one it added are identical, line for line, including `num_redundant_experts=0`. In the merged file, `load_weights` still passes `n_routed_experts` plus `n_shared_experts` when AITER fusion is on, and `self.num_redundant_experts`. [Merged AXK1.py](https://github.com/vllm-project/vllm/blob/af6b982cd2f814ee2f182ac24de6fd4e4a44fbad/vllm/model_executor/models/AXK1.py) The bot was right about the code and the author was right about its origin. Because the same defect existed in a second model, a fix belonged in a change that covered both.

The author opened [issue #41920](https://github.com/vllm-project/vllm/issues/41920) for both models the same morning, and the pull request merged on 7 May 2026. No maintainer replied on the issue. A bot marked it stale on 6 August and closed it as not planned on 7 September. When we fetched vLLM's main branch on 22 November 2026, `AXK1.py` still passed `num_redundant_experts=0` from `get_expert_mapping`.

Deferring was the right call for this pull request. The follow-up still failed, because an issue that nobody owns can lapse even when it is correct.

## Decide for one comment

| What you find | Decision | What to leave in the thread |
| --- | --- | --- |
| Your change copies the pattern or gives the old code a new caller | Fix it in this pull request, with a test | The commit that fixes it |
| The fix needs something you cannot ship yet, such as a library release | Defer with a linked issue, and revisit before merge | The issue, or the upstream change you are waiting for |
| Your change only moves or reformats the code | Open a follow-up issue with an owner or a milestone | The issue link and the commit that introduced the problem |
| The problem is already tracked | Dismiss as out of scope | The existing issue |
| The claim is wrong at the reviewed commit | Dismiss | The code or documentation that shows it, as in [a dismissal reply](/learn/reply-to-dismiss-ai-review-comment/) |

GitHub's documentation suggests the same route for out-of-scope feedback: "you can open a new issue that tracks the feedback and links back to the original comment." [Resolving conversations](https://docs.github.com/en/pull-requests/how-tos/review-pull-requests/commenting-on-a-pull-request#resolving-conversations) Whatever you choose, record which case applied. A reviewer trial can then count the comment as a valid, deferred finding, as the [guide to scoring dismissed findings](/learn/reconstruct-dismissed-review-finding/) describes, rather than as noise.

## One month of free code review.

New organizations get one calendar month of hosted code review for private and public GitHub repositories. No card or cloud connection required. [Eligibility and fair-use terms](/free-code-review/).

[Start free with GitHub](https://proofline.sh/orgs/plan?offer=free_month)

## Related reading

-   ### [What should a code reviewer do after the author pushes a fix?](/learn/review-finding-follow-through/)
    
    Code review The Proofline team November 20, 2026 3 min read
    
-   ### [Read a dismissed code review finding before you score it](/learn/reconstruct-dismissed-review-finding/)
    
    Code review The Proofline team November 20, 2026 4 min read
