Cloud coding agent opens empty PR when invoked from a fork-backed PR review comment

[mojaza]

Describe the bug

Product

GitHub Copilot cloud coding agent (app/copilot-swe-agent), triggered from review feedback on a fork-backed pull request. This may need routing to the hosted coding-agent workflow owners.

Summary

When Copilot was started from a review comment on a fork-backed pull request, it created a draft PR from upstream main instead of the reviewed PR’s fork head. The requested file existed only in the reviewed PR, so the agent could not change it. It nevertheless opened an empty draft PR with an empty Initial plan commit.

Reproduction

1. Open a PR from a personal fork to an upstream repository.
2. Add review feedback about a file introduced by that PR and absent from upstream main.
3. Start Copilot coding agent from that review comment.
4. Copilot creates its branch from upstream main rather than the reviewed PR head.

Actual behavior

• Reviewed PR: [rush-daemon][WS2.3][2/9] Add warm engine component factory microsoft/rushstack#5949
• Empty Copilot PR: [rush-daemon] Add concurrency regression test for serialized invalidation reconciliation microsoft/rushstack#5951
• Copilot branch: microsoft:copilot/fix-review-comment-5949
• Base SHA: d563d6a09ebef0be1ff274faf8c1008767166875
• Head SHA: 7c5a2f29b89b8a8e59da5521ececbb918bd3f5e5
• Result: 0 changed files, 0 additions, 0 deletions.
• The head commit is titled Initial plan and is provably empty: both the head and parent tree SHA are 65487a74d0cb345023b9aebe08aa233d79e1097d.
• The timeline records copilot_work_started and copilot_work_finished despite producing no change.

The reviewed file, libraries/rush-daemon/src/test/WorkspaceEngineComponentFactory.test.ts, existed only on PR #5949’s fork head and not on upstream main.

Expected behavior

For review-comment tasks on fork-backed PRs, Copilot should either:

1. Base its work on the reviewed PR’s exact head; or
2. Stop before creating a PR and report that fork-head access is unsupported.

Copilot should never publish a zero-diff PR when its head and base trees are identical.

Suggested safeguards

• Preserve the reviewed PR head SHA when launching from inline review feedback.
• Verify that the referenced file and line exist in the selected base.
• Compare head and base trees before publishing; suppress an identical-tree PR.
• Report an explicit unsupported-fork error instead of silently falling back to upstream main.

Affected version

No response

Steps to reproduce the behavior

No response

Expected behavior

No response

Additional context

No response