Read PR author association with the kernel-internal app token

.github/workflows/pr-risk.yml

@@ -10,8 +10,12 @@ name: PR Risk
 # rewritten so it never shows a stale "yes".
 #
 # Credentials are split by step. The model runs with no tools and no GitHub token, reading
-# only the PR data gathered with the caller's read-only token. The kernel-internal token is
-# minted after the model has finished, scoped to this repo and to pull-request writes.
+# only the PR data gathered with a read-only kernel-internal token. The token that can comment
+# is minted after the model has finished, scoped to this repo and to pull-request writes.
+#
+# Gather uses the app rather than the workflow token because both the workflow token and the
+# pull_request event payload report private org members as CONTRIBUTOR, and every Kernel
+# member is private. An org-installed app's token sees the real MEMBER/OWNER association.
 
 on:
   workflow_call:
@@ -61,12 +65,23 @@ jobs:
           curl -fsSL "$base/scripts/pr_risk.py" -o /tmp/pr_risk.py
           curl -fsSL "$base/.github/workflows/pr-risk/classify-prompt.md" -o /tmp/pr-risk-prompt.md
 
+      - name: Generate read token
+        id: read-token
+        # Pinned to a commit because this step is handed the app's private key. d72941d is v1.12.0.
+        uses: actions/create-github-app-token@d72941d797fd3113feb6b93fd0dec494b13a2547
+        with:
+          app-id: ${{ secrets.ADMIN_APP_ID }}
+          private-key: ${{ secrets.ADMIN_APP_PRIVATE_KEY }}
+          owner: ${{ github.repository_owner }}
+          repositories: ${{ github.event.repository.name }}
+          permission-contents: read
+          permission-pull-requests: read
+
       - name: Gather
         id: gather
         env:
-          GH_TOKEN: ${{ github.token }}
+          GH_TOKEN: ${{ steps.read-token.outputs.token }}
           DENY_PATHS: ${{ inputs.deny-paths }}
-          AUTHOR_ASSOCIATION: ${{ github.event.pull_request.author_association }}
         run: python3 /tmp/pr_risk.py gather --repo "$REPO" --number "$NUMBER" --dir "$WORKDIR"
 
       - name: Install Claude Code CLI

scripts/pr_risk.py

@@ -160,18 +160,16 @@ def deny_patterns(extra: str) -> list[str]:
     return list(DEFAULT_DENY_PATHS) + [line for line in lines if line and not line.startswith("#")]
 
 
-def ineligible_reason(pr: dict[str, Any], files: list[dict[str, Any]], patterns: list[str],
-                      association: str) -> str | None:
+def ineligible_reason(pr: dict[str, Any], files: list[dict[str, Any]], patterns: list[str]) -> str | None:
     if pr.get("draft"):
         return "draft PR"
     head_repo = (pr["head"].get("repo") or {}).get("full_name")
     if head_repo != pr["base"]["repo"]["full_name"]:
         return "PR comes from a fork"
     # Only org members' own changes are candidates. Bots are never members, so their PRs are
     # skipped too: no person wrote them, and approving them unread is a separate decision.
-    # The association comes from the event payload: the API answers for the caller's token,
-    # and the workflow token cannot see private org membership, so it reports CONTRIBUTOR.
-    if association not in ("OWNER", "MEMBER"):
+    # This is only right when read with the app token; see the note in pr-risk.yml.
+    if pr.get("author_association") not in ("OWNER", "MEMBER"):
         return "author is not an org member"
     changed = pr["additions"] + pr["deletions"]
     if changed > MAX_CHANGED_LINES:
@@ -283,12 +281,12 @@ def previous_data(body: str) -> dict[str, Any] | None:
         return None
 
 
-def gather(gh: GitHub, number: int, extra_deny: str, association: str, out: Path) -> dict[str, Any]:
+def gather(gh: GitHub, number: int, extra_deny: str, out: Path) -> dict[str, Any]:
     out.mkdir(parents=True, exist_ok=True)
     pr = gh.pull(number)
     files = gh.files(number)
     existing = find_comment(gh.comments(number))
-    reason = ineligible_reason(pr, files, deny_patterns(extra_deny), association)
+    reason = ineligible_reason(pr, files, deny_patterns(extra_deny))
     gate = {
         "sha": pr["head"]["sha"],
         "eligible": reason is None,
@@ -439,7 +437,7 @@ def main() -> int:
         return 0
     gh = GitHub(args.repo)
     if args.cmd == "gather":
-        gather(gh, args.number, os.environ.get("DENY_PATHS", ""), os.environ["AUTHOR_ASSOCIATION"], args.dir)
+        gather(gh, args.number, os.environ.get("DENY_PATHS", ""), args.dir)
         return 0
     return comment(gh, args.number, args.dir, os.environ.get("RUN_URL", ""))
 

scripts/test_pr_risk.py

@@ -38,10 +38,10 @@
 PATCH = "@@ -10,3 +10,4 @@ flags:\n   - a\n   - b\n+  - c\n   - d"
 
 
-def pr(additions=1, deletions=0, draft=False, fork=False, sha="abc1234def"):
+def pr(additions=1, deletions=0, draft=False, fork=False, sha="abc1234def", association="MEMBER"):
     return {
         "number": 7, "draft": draft, "title": "Add an org to a flag", "body": "Adds one org.",
-        "user": {"login": "someone"}, "author_association": "CONTRIBUTOR",
+        "user": {"login": "someone"}, "author_association": association,
         "additions": additions, "deletions": deletions,
         "head": {"sha": sha, "repo": {"full_name": "external/infra" if fork else "kernel/infra"}},
         "base": {"sha": "base000", "ref": "main", "repo": {"full_name": "kernel/infra"}},
@@ -97,10 +97,10 @@ def visible_words(body):
 
 
 class PRRiskTests(unittest.TestCase):
-    def run_pr(self, gh, verdict=VERDICT, result=None, extra_deny="", association="MEMBER"):
+    def run_pr(self, gh, verdict=VERDICT, result=None, extra_deny=""):
         with tempfile.TemporaryDirectory() as tmp:
             work = Path(tmp)
-            gate = pr_risk.gather(gh, 7, extra_deny, association, work)
+            gate = pr_risk.gather(gh, 7, extra_deny, work)
             if gate["eligible"]:
                 (work / "result.json").write_text(json.dumps(result or {"is_error": False, "structured_output": verdict}))
             code = pr_risk.comment(gh, 7, work, "https://example.com/run")
@@ -130,22 +130,21 @@ def test_ineligible_reasons(self):
         renamed = file("b.yaml", status="renamed", previous_filename="a.yaml", additions=0, changes=0)
         del renamed["patch"]
         cases = {
-            "draft": (pr(draft=True), [file()], "MEMBER"),
-            "fork": (pr(fork=True), [file()], "MEMBER"),
-            "not an org member": (pr(), [file()], "CONTRIBUTOR"),
-            "author is not an org member": (pr(), [file()], "NONE"),
-            "51 changed lines": (pr(additions=30, deletions=21), [file()], "MEMBER"),
-            "deny list (`identity/**`)": (pr(), [file("identity/access.go")], "MEMBER"),
-            "deny list (`.github/**`)": (pr(), [file("ci.yml", previous_filename=".github/workflows/ci.yml")], "MEMBER"),
-            "`bin/helper` has no reviewable text diff": (pr(additions=0), [binary], "MEMBER"),
-            "`b.yaml` has no reviewable text diff": (pr(additions=0), [renamed], "MEMBER"),
+            "draft": (pr(draft=True), [file()]),
+            "fork": (pr(fork=True), [file()]),
+            "not an org member": (pr(association="CONTRIBUTOR"), [file()]),
+            "author is not an org member": (pr(association="NONE"), [file()]),
+            "51 changed lines": (pr(additions=30, deletions=21), [file()]),
+            "deny list (`identity/**`)": (pr(), [file("identity/access.go")]),
+            "deny list (`.github/**`)": (pr(), [file("ci.yml", previous_filename=".github/workflows/ci.yml")]),
+            "`bin/helper` has no reviewable text diff": (pr(additions=0), [binary]),
+            "`b.yaml` has no reviewable text diff": (pr(additions=0), [renamed]),
         }
-        for expected, (p, files, association) in cases.items():
+        for expected, (p, files) in cases.items():
             with self.subTest(expected=expected):
-                self.assertIn(expected, pr_risk.ineligible_reason(p, files, patterns, association))
-        # The API reports CONTRIBUTOR for private members; only the event payload's value counts.
-        self.assertIsNone(pr_risk.ineligible_reason(pr(additions=25, deletions=25), [file()], patterns, "MEMBER"))
-        self.assertIsNone(pr_risk.ineligible_reason(pr(), [file()], patterns, "OWNER"))
+                self.assertIn(expected, pr_risk.ineligible_reason(p, files, patterns))
+        self.assertIsNone(pr_risk.ineligible_reason(pr(additions=25, deletions=25), [file()], patterns))
+        self.assertIsNone(pr_risk.ineligible_reason(pr(association="OWNER"), [file()], patterns))
 
     def test_ineligible_without_comment_is_silent(self):
         gh = FakeGitHub(pr(additions=200), [file()])
@@ -281,8 +280,12 @@ def test_workflow_isolates_the_model(self):
         names = list(steps)
         self.assertLess(names.index("Classify"), names.index("Generate app token"))
         self.assertEqual(steps["Generate app token"]["with"]["permission-pull-requests"], "write")
-        self.assertEqual(steps["Gather"]["env"]["AUTHOR_ASSOCIATION"],
-                         "${{ github.event.pull_request.author_association }}")
+        # The workflow token and event payload both report private members as CONTRIBUTOR.
+        self.assertEqual(steps["Gather"]["env"]["GH_TOKEN"], "${{ steps.read-token.outputs.token }}")
+        read = steps["Generate read token"]["with"]
+        self.assertEqual({k: v for k, v in read.items() if k.startswith("permission-")},
+                         {"permission-contents": "read", "permission-pull-requests": "read"})
+        self.assertLess(names.index("Generate read token"), names.index("Gather"))
         triggers = workflow[True]["workflow_call"]
         self.assertEqual(set(triggers["secrets"]), {"ADMIN_APP_ID", "ADMIN_APP_PRIVATE_KEY", "ANTHROPIC_API_KEY"})