From 32c8d7bae6d42247e405f2ebccb32bbcfab60764 Mon Sep 17 00:00:00 2001
From: David Baker Effendi <david@brokk.ai>
Date: Mon, 5 Oct 2026 10:06:02 +0200
Subject: [PATCH] Show the resolved Bifrost revision in benchmark summaries

Scheduled runs against moving master can repeat the same public snapshot while private repairs are still unpublished. Include the resolved engine revision in the human-readable summary so the tested source is visible alongside the outcome.

Prefer bifrostResolvedCommit and retain legacy Bifrost report support. Preserve machine-readable count and failing-case output. Validation: baseline Rust suites (273 tests), all 59 corpus documents, 44 Python tests, formatting, and seven focused summary tests.
---
 scripts/render-benchmark-summary.sh | 25 ++++++++
 tests/benchmark_summary.rs          | 99 +++++++++++++++++++++++++++++
 2 files changed, 124 insertions(+)

diff --git a/scripts/render-benchmark-summary.sh b/scripts/render-benchmark-summary.sh
index fb959b4..723d495 100755
--- a/scripts/render-benchmark-summary.sh
+++ b/scripts/render-benchmark-summary.sh
@@ -85,8 +85,33 @@ fi
 jq -r --argjson counts "$counts_json" '
   (if has("completed") then .completed else true end) as $complete
   | def count_or_unknown($value): if $value == null then "unknown" else ($value | tostring) end;
+    def usable_revision:
+      if type == "string" then
+        gsub("^\\s+|\\s+$"; "") as $revision
+        | if ($revision | length) > 0 and (($revision | ascii_downcase) != "unknown") then
+            $revision
+          else
+            null
+          end
+      else
+        null
+      end;
+    (.bifrostResolvedCommit | usable_revision) as $resolved_commit
+    | (if $resolved_commit != null then
+         $resolved_commit
+       elif .runner.name? == "bifrost" then
+         (.runner.resolvedVersion | usable_revision)
+       else
+         null
+       end) as $bifrost_revision
+    |
     if $complete then "## Totals" else "## INCOMPLETE - Processed Totals" end,
     "",
+    (if $bifrost_revision != null then
+       "- Bifrost resolved revision: `\($bifrost_revision)`"
+     else
+       empty
+     end),
     "- Documents processed/requested: \($counts.processed_documents_count)/\(count_or_unknown($counts.requested_documents_count))",
     "- Authored cases requested: \(count_or_unknown($counts.requested_authored_cases_count))",
     "- Planned cases processed/requested: \($counts.processed_cases_count)/\(count_or_unknown($counts.requested_planned_cases_count))",
diff --git a/tests/benchmark_summary.rs b/tests/benchmark_summary.rs
index 94ded49..8168799 100644
--- a/tests/benchmark_summary.rs
+++ b/tests/benchmark_summary.rs
@@ -39,6 +39,105 @@ fn render_counts(report: serde_json::Value) -> serde_json::Value {
     serde_json::from_slice(&output.stdout).unwrap()
 }
 
+fn render_failing_cases(report: serde_json::Value) -> String {
+    let tempdir = tempfile::tempdir().unwrap();
+    let report_path = tempdir.path().join("report.json");
+    fs::write(&report_path, serde_json::to_vec(&report).unwrap()).unwrap();
+    let output = Command::new("bash")
+        .arg(concat!(
+            env!("CARGO_MANIFEST_DIR"),
+            "/scripts/render-benchmark-summary.sh"
+        ))
+        .arg("--failing-cases")
+        .arg(report_path)
+        .output()
+        .unwrap();
+    assert!(output.status.success());
+    String::from_utf8(output.stdout).unwrap()
+}
+
+#[test]
+fn summary_reports_the_exact_new_bifrost_revision() {
+    let revision = "0123456789abcdef0123456789abcdef01234567";
+    let summary = render(json!({
+        "runner": {
+            "name": "bifrost",
+            "resolvedVersion": "legacy-runner-value"
+        },
+        "bifrostResolvedCommit": revision,
+        "totals": {},
+        "documents": []
+    }));
+
+    assert!(summary.contains(&format!("- Bifrost resolved revision: `{revision}`")));
+    assert!(!summary.contains("legacy-runner-value"));
+}
+
+#[test]
+fn summary_uses_legacy_runner_revision_and_omits_unusable_values() {
+    let legacy_summary = render(json!({
+        "runner": {
+            "name": "bifrost",
+            "resolvedVersion": " legacy-resolved-value "
+        },
+        "totals": {},
+        "documents": []
+    }));
+    assert!(legacy_summary.contains("- Bifrost resolved revision: `legacy-resolved-value`"));
+
+    let unavailable_summary = render(json!({
+        "runner": {
+            "name": "bifrost",
+            "resolvedVersion": "UNKNOWN"
+        },
+        "bifrostResolvedCommit": "  ",
+        "totals": {},
+        "documents": []
+    }));
+    assert!(!unavailable_summary.contains("Bifrost resolved revision:"));
+}
+
+#[test]
+fn provenance_does_not_change_counts_or_failing_case_modes() {
+    let report = json!({
+        "runner": {
+            "name": "bifrost",
+            "resolvedVersion": "legacy-runner-value"
+        },
+        "bifrostResolvedCommit": "0123456789abcdef0123456789abcdef01234567",
+        "completed": true,
+        "requestedCaseFiles": ["case.yaml"],
+        "totals": {
+            "documents": 1,
+            "cases": 1,
+            "passed": 0,
+            "failed": 1
+        },
+        "documents": [{
+            "caseFile": "case.yaml",
+            "cases": [{"id": "bad-case", "status": "failed"}]
+        }]
+    });
+
+    assert_eq!(
+        render_counts(report.clone()),
+        json!({
+            "report_complete": true,
+            "processed_documents_count": 1,
+            "requested_documents_count": 1,
+            "processed_cases_count": 1,
+            "requested_authored_cases_count": 1,
+            "requested_planned_cases_count": 1,
+            "requested_development_cases_count": 0,
+            "requested_evaluation_cases_count": 0
+        })
+    );
+    assert_eq!(
+        render_failing_cases(report),
+        "- failed: bad-case (case.yaml)\n"
+    );
+}
+
 #[test]
 fn incomplete_summary_distinguishes_processed_and_requested_scope() {
     let summary = render(json!({