Run the AI reviewer only on /review

.gitattributes

@@ -1 +1,3 @@
-.github/workflows/*.lock.yml linguist-generated=true
+.github/workflows/*.lock.yml linguist-generated=true
+# gh aw compile doesn't recognize "## agent:" markers that end in CRLF, so keep its sources LF on Windows.
+.github/workflows/**/*.md text eol=lf

.github/workflows/pr-review-command.md

@@ -1,5 +1,5 @@
 ---
-description: Re-reviews a pull request with two reviewers on different models when someone comments /review on it.
+description: Reviews a pull request with two reviewers on different models when someone comments /review on it.
 
 on:
   slash_command:
@@ -21,6 +21,6 @@ imports:
 timeout-minutes: 40
 ---
 
-# PR re-review
+# PR review
 
-Review pull request #${{ github.event.issue.number }} in ${{ github.repository }} again, following the steps below. It has probably changed since the last review: focus on what's new, and don't repeat earlier comments that were already addressed.
+Review pull request #${{ github.event.issue.number }} in ${{ github.repository }}, following the steps below. If it was reviewed before, it has probably changed since: focus on what's new, and don't repeat earlier comments that were already addressed.

.github/workflows/shared/pr-review.md

@@ -1,5 +1,5 @@
 ---
-# Shared by pr-review.md (automatic) and pr-review-command.md (/review comment).
+# Imported by pr-review-command.md, which runs when someone comments /review on a PR.
 # A lead agent runs two independent reviewers on different models, cross-checks
 # what only one of them found, and posts the merged result.
 

README.md

@@ -25,11 +25,11 @@ To hide the bulk reformat commit from `git blame`, run `git config blame.ignoreR
 
 ## AI PR review
 
-`pr-review.md` and `pr-review-command.md` in `.github/workflows/` are an AI reviewer written as [GitHub Agentic Workflows](https://github.github.com/gh-aw/). It reviews each PR when it's opened or marked ready for review, and commenting `/review` on a PR runs it again, including on drafts. It only comments; it never approves or blocks.
+`pr-review-command.md` in `.github/workflows/` is an AI reviewer written as [GitHub Agentic Workflows](https://github.github.com/gh-aw/). It only runs when you comment `/review` on a PR (drafts included), so it never uses premium requests on its own. It only comments; it never approves or blocks.
 
 - How it works (`shared/pr-review.md`): two reviewers on different models (`claude-reviewer`, `gpt-reviewer`) review the PR independently. A lead agent merges their findings, asks each reviewer to check what only the other found, verifies disputes against the code, and posts the result with a *Review context and counts* table. `Consensus reached: False` or a low confidence score means the round deserves a second look; comment `/review` to retry.
 - The review criteria live in `.github/agents/pr-reviewer.agent.md`, which is also available in the repo's Agents tab. The sub-agents' task and output format are in `shared/pr-reviewer-subagent.md`.
-- Models are pinned: the lead's in `pr-review.md` and `pr-review-command.md`, the reviewers' in the `## agent:` blocks of `shared/pr-review.md`.
+- Models are pinned: the lead's in `pr-review-command.md`, the reviewers' in the `## agent:` blocks of `shared/pr-review.md`.
 - The agent always uses the instructions on `main`, not the PR's copy, so a PR can't rewrite its own reviewer. A PR that changes the reviewer is still reviewed by the old version; test the change with `/review` after merging.
 - Edit the `.md` files, then run `gh aw compile` (needs `gh extension install github/gh-aw`) to regenerate the `.lock.yml` files GitHub actually runs.
 - It runs Copilot with the `COPILOT_GITHUB_TOKEN` repository secret: a fine-grained personal access token with the **Copilot Requests** permission. Each run uses Copilot premium requests for three models (the lead and two reviewers). Only users with write access trigger it.