fix(review): recheck paginated file inventory versions

loopx/capabilities/pr_review_queue/github_source.py

@@ -194,13 +194,15 @@ def same_snapshot() -> bool:
         pages = payload if all(isinstance(page, list) for page in payload) else [payload]
         files: list[dict[str, Any]] = []
         observed: list[dict[str, Any]] = []
+        paths: set[str] = set()
         for page in pages:
             for item in page:
                 if not isinstance(item, dict):
                     return None
                 path = item.get("filename") or item.get("path")
-                if not isinstance(path, str) or not path:
+                if not isinstance(path, str) or not path or path in paths:
                     return None
+                paths.add(path)
                 if recover and any(type(item.get(key)) is not int for key in ("additions", "deletions")):
                     return None
                 additions = int(item.get("additions") or 0)
@@ -212,7 +214,9 @@ def same_snapshot() -> bool:
                 observed.append(row | {"status": item.get("status"),
                                        "previous_filename": item.get("previous_filename")})
         if len(files) == expected_count:
-            return files if not recover or same_snapshot() else None
+            # REST pages are not pinned to the GraphQL detail's revision. Even
+            # an ordinary PR can change without changing its file count.
+            return files if snapshot is None or same_snapshot() else None
         if not recover or not files or len(files) > expected_count:
             return None
         recovered = _recover_git_pr_files(

tests/test_pr_review_github_scan.py

@@ -271,13 +271,16 @@ def test_pr_list_paginates_files_when_graphql_detail_is_truncated(monkeypatch) -
         {"filename": f"src/file_{index}.py", "additions": index, "deletions": 0}
         for index in range(101)
     ]
+    row.update(additions=sum(item["additions"] for item in rest_files), deletions=0)
     calls: list[list[str]] = []
 
     def fake(args: list[str], *, cwd: Path | None = None):
         calls.append(args)
         if args[:2] == ["pr", "list"]:
             return [row]
         if args[:2] == ["pr", "view"]:
+            if "files" not in args[args.index("--json") + 1].split(","):
+                return dict(row)
             details = _fake_run_gh_json(args, cwd=cwd)
             details["files"] = details["files"] * 100
             return details
@@ -298,6 +301,74 @@ def fake(args: list[str], *, cwd: Path | None = None):
     assert ["api", "--paginate", "--slurp"] in [call[:3] for call in calls]
 
 
+@pytest.mark.parametrize("exact_target", [False, True])
+@pytest.mark.parametrize("drift", [
+    None, "headRefOid", "baseRefOid", "changedFiles", "additions", "deletions",
+    "unavailable", "duplicate_path",
+])
+def test_paginated_files_require_a_complete_unchanged_snapshot(
+    monkeypatch, exact_target: bool, drift: str | None,
+) -> None:
+    listed = _rows()[0] | {"changedFiles": 101, "additions": 101, "deletions": 0}
+    rest_files = [
+        {"filename": f"src/file_{index}.py", "additions": 1, "deletions": 0}
+        for index in range(101)
+    ]
+    if drift == "duplicate_path":
+        rest_files[-1] = dict(rest_files[0])
+    inventory_read = False
+    calls = []
+
+    def github(args, *, cwd=None):
+        nonlocal inventory_read
+        calls.append(args)
+        assert "statusCheckRollup" not in ",".join(args)
+        if args[:2] == ["pr", "list"]:
+            return [dict(listed)]
+        if args[:3] == ["api", "--paginate", "--slurp"]:
+            inventory_read = True
+            return [rest_files[:100], rest_files[100:]]
+        fields = args[args.index("--json") + 1].split(",")
+        if "files" in fields:
+            # The initial detail belongs to the original head; the REST pages
+            # can come from a later revision with the same number of files.
+            return _fake_run_gh_json(args, cwd=cwd) | {"files": [{}] * 100}
+        current = dict(listed)
+        if inventory_read:
+            if drift == "unavailable":
+                raise RuntimeError("version readback unavailable")
+            if drift in {"headRefOid", "baseRefOid"}:
+                current[drift] = "d" * 40
+            elif drift in {"changedFiles", "additions", "deletions"}:
+                current[drift] += 1
+        return current
+
+    if exact_target:
+        def scan():
+            return github_source_module.scan_github_pull_request_targets(
+                repository="owner/repo", exact_heads=[f"1@{HEAD_1}"],
+                run_gh_json=github, wait_for_ci=False,
+            )
+
+        if drift is not None:
+            with pytest.raises(RuntimeError, match="detail read was incomplete"):
+                scan()
+            return
+        result = scan()
+    else:
+        monkeypatch.setattr(pr_review_module, "_run_gh_json", github)
+        result = pr_review_module.scan_github_pull_requests(
+            repo="owner/repo", limit=10, wait_for_ci=False,
+        )
+        if drift is not None:
+            assert result["complete"] is False
+            assert result["states"][0]["detail_read_failures"] == 1
+            assert "files" not in result["pull_requests"][0]
+            return
+    assert result["complete"] is True
+    assert len(result["pull_requests"][0]["files"]) == 101
+
+
 def test_pr_list_marks_source_incomplete_when_rest_files_are_still_truncated(
     monkeypatch,
 ) -> None: