Keep response channels separate from acceptance
Use this article as a repository-policy template, informed by the supplied GitHub guidance rather than as a provider-prescribed workflow. A response beneath a Copilot comment can be read by other people, while Copilot does not receive that response. Rating controls are for sending feedback that may improve later suggestions.
Do not treat any of the four channels as merge acceptance. In its default form, a Copilot review is a Comment review and does not contribute to required approvals; an installation can be configured differently. The supplied overview also warns that Copilot may overlook issues or be incorrect, and calls for careful validation alongside human review.
Four-case channel decision table
| Case | Intended recipient or purpose | Branch-change evidence | Documented Copilot behavior | Independent acceptance evidence | Inference to reject | Policy owner | Bounded repository-owned action |
|---|---|---|---|---|---|---|---|
| Human-visible reply | Coordinate context, questions, or a decision with people reviewing the pull request. | No branch alteration is established by the reply; record any needed correction separately. | Other people can view the reply, but Copilot cannot read it. | Require the repository's chosen human approval and required-check evidence. | “My reply will alter Copilot’s next assessment.” | Reviewer and pull-request maintainer. | Reply when human context helps; independently verify the concern and amend the branch when warranted. |
| Feedback rating | Signal whether a particular Copilot comment was useful, with optional explanatory feedback for a negative rating. | A rating is not repository evidence that code changed. | GitHub presents positive and negative ratings as feedback intended to improve subsequent suggestions. | Use the same repository-defined acceptance evidence as for any other pull request. | “A rating settles the finding on this pull request.” | Comment recipient or designated reviewer. | Rate the comment when learning feedback is useful; keep the technical disposition in normal review records. |
| Manual re-review | Ask for an explicit new assessment after a relevant branch update or a resolved ambiguity. | Identify the commit or diff that addresses the issue before making the request. | After pull-request changes, a reviewer can request another Copilot review manually; repeat review is not automatic unless new-push review has been configured. | Collect the repository's separate approval, check, and escalation evidence. | “Requesting reassessment itself accepts the change.” | Pull-request maintainer, under repository review rules. | Request one reassessment after the relevant update, then route repeated or disputed findings to the repository’s escalation path. |
| Automatic re-review after a push | Apply a preselected repository or ruleset policy to newly pushed work. | The triggering evidence is a new push to the pull request. | New-push review must be enabled for automatic reassessment; otherwise, automatic review normally happens once. | Keep human acceptance and required checks independent of the automatic event. | “An automatic reassessment is a merge decision.” | Repository administrator or ruleset owner. | Enable this path only when its scope, draft handling, escalation owner, and required acceptance evidence are documented locally. |
Hypothetical failure walkthrough
Consider an explicitly hypothetical pull request that targets Python 3.14. An author replies that Boolean use of NotImplemented is acceptable and assumes Copilot will consume that reply. That assumption fails as a communication path: people can see the reply, but Copilot cannot.
A human reviewer instead checks the Python 3.14 rule: using NotImplemented as a Boolean raises TypeError. Python documents that value as a special-method signal for an unsupported operation, not as NotImplementedError. Under the proposed policy, the reviewer withholds acceptance, asks for a branch correction, and sends the changed branch through the repository’s selected manual or automatic reassessment route. This is a conditional workflow example, not a claim that the scenario was run or that Copilot would find the problem.
Boundaries, ownership, and maintenance
Assign channel choice, merge criteria, and escalation separately. A repository owner can define which human approval, automated checks, security review, or combination is required; a reviewer can decide whether a reply adds useful context; and a maintainer can supply the branch correction. These are proposed local responsibilities, not GitHub guarantees.
The decision table excludes preview agentic capabilities and stays within the supplied excerpts on GitHub.com review interactions and Python 3.14 semantics. Revisit this policy when the relevant sources change or when the repository retires, changes, or no longer targets the Python version assumed by the walkthrough.
The separation is necessary, but the policy should also treat a reply as a disclosure channel rather than merely a non-authoritative one. GitHub states that replies to Copilot comments are visible to other people but not to Copilot; therefore a reply cannot be relied upon to change the assessment, and it may expose rationale, exploit details, credentials, or internal system context to every party able to view the pull request. GitHub’s review guidance
Could the policy add a fail-closed control: replies may contain only information approved for the pull request’s audience, while sensitive remediation detail is placed in the repository’s separately authorized security or escalation channel? The acceptance record should then bind the human decision and required checks to the corrected commit, not to the existence, resolution, or wording of a public reply.