fix: score CI from the latest run of each check name

CHANGELOG.md

@@ -6,6 +6,11 @@ All notable changes to this project are documented here. The format follows
 
 ## [Unreleased]
 
+### Fixed
+
+- CI badge uses the latest run of each check name, matching the PR merge box. A cancelled
+  workflow's failed `Verify / gate` no longer keeps a later green run red.
+
 ### Added
 
 - Pace icons on weekly and monthly usage bars: `>` to `>>>` when you use a quota faster than

README.md

@@ -83,7 +83,7 @@ devdash --help                     # all flags
 
 |Mark|Meaning|
 |---|---|
-|`✓ci` `●ci2` `✗ci1`|Checks passed, 2 still running, 1 failed (the first failed check is named at the end of the row)|
+|`✓ci` `●ci2` `✗ci1`|Latest run of each check passed, 2 still running, 1 failed (the first failed check is named at the end of the row)|
 |`approved` `changes` `needs review`|The PR's review decision|
 |`✓dana` `✗lee` `✎sam`|A reviewer's latest review: approved, requested changes, commented|
 |`bot✎2`|Reviews from 2 bots|

devdash.py

@@ -407,8 +407,8 @@ def render_usage(st, width, now):
   viewerLatestReview { state }
   commits(last: 1) { nodes { commit { statusCheckRollup { state
     contexts(first: 100) { nodes { __typename
-      ... on CheckRun { name status conclusion }
-      ... on StatusContext { context state } } } } } } }
+      ... on CheckRun { name status conclusion databaseId startedAt }
+      ... on StatusContext { context state createdAt } } } } } } }
 }
 """
 # gh-stack's `stack` field is not in every GitHub schema (GitHub Enterprise Server, for one).
@@ -449,14 +449,35 @@ def keep(nodes):
     return keep(d["mine"]["nodes"]), keep(d["review"]["nodes"]) if review_requested else []
 
 
+def latest_contexts(nodes):
+    """One entry per check name / status context: the newest run.
+
+    GitHub's statusCheckRollup keeps every check-run on the SHA, so a
+    cancelled in-progress run's `Verify / gate` FAILURE sits next to the
+    later SUCCESS. The PR merge box and `gh pr checks` use the latest of
+    each name; the rollup `state` does not.
+    """
+    latest = {}
+    for i, ctx in enumerate(nodes):
+        if ctx.get("__typename") == "CheckRun":
+            key = ("check", ctx["name"])
+            sort = (ctx.get("databaseId") or 0, ctx.get("startedAt") or "", i)
+        else:
+            key = ("status", ctx.get("context"))
+            sort = (0, ctx.get("createdAt") or "", i)
+        if key not in latest or sort > latest[key][0]:
+            latest[key] = (sort, ctx)
+    return [item[1] for item in latest.values()]
+
+
 def ci_badge(pr):
     """(Text badge, name of the first failed check or '')."""
     nodes = pr["commits"]["nodes"]
     roll = nodes[0]["commit"]["statusCheckRollup"] if nodes else None
     if not roll:
         return Text("○ci", style=META), ""
     failed, running = [], 0
-    for ctx in roll["contexts"]["nodes"]:
+    for ctx in latest_contexts(roll["contexts"]["nodes"]):
         if ctx["__typename"] == "CheckRun":
             if ctx["status"] != "COMPLETED":
                 running += 1
@@ -468,11 +489,9 @@ def ci_badge(pr):
             running += 1
     if failed:
         return Text(f"✗ci{len(failed)}", style=f"bold {BAD}"), failed[0]
-    if running or roll["state"] in ("PENDING", "EXPECTED"):
+    if running:
         return Text(f"●ci{running or ''}", style=WARN), ""
-    if roll["state"] == "SUCCESS":
-        return Text("✓ci", style=OK), ""
-    return Text("✗ci", style=BAD), ""
+    return Text("✓ci", style=OK), ""
 
 
 DECISION = {"APPROVED": ("approved", f"bold {OK}"),

tests/test_devdash.py

@@ -108,6 +108,53 @@ def fake_run(cmd, timeout=45):
     assert len(queries) == 3  # the fallback is remembered: no second failed attempt
 
 
+def rollup(*nodes, state="FAILURE"):
+    return {"commits": {"nodes": [{"commit": {"statusCheckRollup": {
+        "state": state, "contexts": {"nodes": list(nodes)}}}}]}}
+
+
+def check(name, conclusion, status="COMPLETED", database_id=1, started="2026-09-25T16:00:00Z"):
+    return {"__typename": "CheckRun", "name": name, "status": status,
+            "conclusion": conclusion, "databaseId": database_id, "startedAt": started}
+
+
+def test_ci_badge_uses_the_latest_run_of_a_check_name():
+    """A cancelled workflow posts Verify / gate FAILURE; the surviving run posts SUCCESS.
+    statusCheckRollup.state stays FAILURE because every run remains on the SHA."""
+    pr = rollup(
+        check("Verify / gate", "FAILURE", database_id=1, started="2026-09-25T15:57:07Z"),
+        check("Verify / gate", "SUCCESS", database_id=2, started="2026-09-25T16:04:12Z"),
+        check("Infra / gate", "FAILURE", database_id=3, started="2026-09-25T15:57:07Z"),
+        check("Infra / gate", "SUCCESS", database_id=4, started="2026-09-25T15:57:32Z"),
+        check("Infra / plan (${{ matrix.env }})", "CANCELLED", database_id=5),
+        {"__typename": "StatusContext", "context": "CodeRabbit", "state": "SUCCESS",
+         "createdAt": "2026-09-25T16:00:00Z"},
+    )
+    badge, failed = devdash.ci_badge(pr)
+    assert badge.plain == "✓ci"
+    assert failed == ""
+
+
+def test_ci_badge_a_later_failure_replaces_an_earlier_pass():
+    pr = rollup(
+        check("Verify / gate", "SUCCESS", database_id=1),
+        check("Verify / gate", "FAILURE", database_id=2),
+    )
+    badge, failed = devdash.ci_badge(pr)
+    assert badge.plain == "✗ci1"
+    assert failed == "Verify / gate"
+
+
+def test_ci_badge_a_rerun_in_progress_is_pending_not_the_old_result():
+    pr = rollup(
+        check("Verify / gate", "SUCCESS", database_id=1),
+        check("Verify / gate", None, status="IN_PROGRESS", database_id=2),
+    )
+    badge, failed = devdash.ci_badge(pr)
+    assert badge.plain == "●ci1"
+    assert failed == ""
+
+
 def test_fetch_does_not_hide_other_errors(monkeypatch):
     monkeypatch.setattr(devdash, "HAS_STACK", True)