From 4d0390c4fa207d28b212ea14a9dbe595426adf73 Mon Sep 17 00:00:00 2001
From: openhands <openhands@all-hands.dev>
Date: Sun, 15 Mar 2026 17:02:19 +0000
Subject: [PATCH] fix: address review findings for cross-candidate red-team
 sweep (#822)

- red-team-sweep.sh: reset CROSS_PATTERNS_FILE at sweep start to prevent
  stale patterns from prior invocations contaminating a fresh run
- red-team-sweep.sh: wrap pattern-extraction Python in set +e/set -e and
  capture output so log() prefix is applied; move memory truncation outside
  the if-block so it runs unconditionally even if Python fails
- red-team.sh: filter entries where candidate == current_candidate before
  grouping, removing self-referential cross-candidate evidence
- red-team.sh: skip entries with empty pattern key (both pattern and
  strategy fields empty) to prevent spurious bucket merging

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 scripts/harb-evaluator/red-team-sweep.sh | 19 +++++++++++++++----
 scripts/harb-evaluator/red-team.sh       |  5 +++++
 2 files changed, 20 insertions(+), 4 deletions(-)

diff --git a/scripts/harb-evaluator/red-team-sweep.sh b/scripts/harb-evaluator/red-team-sweep.sh
index a392518..f59023b 100755
--- a/scripts/harb-evaluator/red-team-sweep.sh
+++ b/scripts/harb-evaluator/red-team-sweep.sh
@@ -20,6 +20,9 @@ die() { log "FATAL: $*" >&2; exit 1; }
 
 [[ -f "$INJECT" ]] || die "inject.sh not found at $INJECT"
 
+# Reset cross-patterns file for this sweep invocation (prevents stale data from prior runs)
+> "$CROSS_PATTERNS_FILE"
+
 # Load progress
 completed=()
 if [[ -f "$PROGRESS_FILE" ]]; then
@@ -118,7 +121,8 @@ PYEOF
 
   # 4b. Extract abstract patterns into cross-candidate file, then clear raw memory
   if [[ -f "$MEMORY_FILE" && -s "$MEMORY_FILE" ]]; then
-    python3 - "$MEMORY_FILE" "$CROSS_PATTERNS_FILE" <<'PYEOF'
+    set +e
+    _extract_out=$(python3 - "$MEMORY_FILE" "$CROSS_PATTERNS_FILE" <<'PYEOF'
 import json, sys
 
 mem_file   = sys.argv[1]
@@ -135,16 +139,23 @@ with open(mem_file) as f:
                 pass
 
 if not new_entries:
-    print("  No memory entries to extract")
+    print("No memory entries to extract")
     sys.exit(0)
 
 with open(cross_file, 'a') as f:
     for e in new_entries:
         f.write(json.dumps(e) + '\n')
 
-print(f"  Extracted {len(new_entries)} entr{'y' if len(new_entries)==1 else 'ies'} to cross-patterns file")
+print(f"Extracted {len(new_entries)} entr{'y' if len(new_entries)==1 else 'ies'} to cross-patterns file")
 PYEOF
-    # Clear raw memory so the next candidate starts with a fresh tactical state
+    )
+    _py_exit=$?
+    set -e
+    [[ -n "$_extract_out" ]] && log "$_extract_out"
+    [[ $_py_exit -ne 0 ]] && log "WARNING: cross-pattern extraction failed (exit $_py_exit) — continuing"
+  fi
+  # Always clear raw memory so the next candidate starts with a fresh tactical state
+  if [[ -f "$MEMORY_FILE" ]]; then
     > "$MEMORY_FILE"
     log "Cleared raw memory for next candidate"
   fi
diff --git a/scripts/harb-evaluator/red-team.sh b/scripts/harb-evaluator/red-team.sh
index a3a1add..a951bb9 100755
--- a/scripts/harb-evaluator/red-team.sh
+++ b/scripts/harb-evaluator/red-team.sh
@@ -486,10 +486,15 @@ with open(cross_file) as f:
 if not entries:
     sys.exit(0)
 
+# Exclude entries from the current candidate (they are cross-candidate evidence, not self-evidence)
+entries = [e for e in entries if e.get("candidate", "unknown") != current_candidate]
+
 # Group by abstract pattern; track worked/failed per candidate
 by_pattern = defaultdict(lambda: {"worked": {}, "failed": {}, "insight": ""})
 for e in entries:
     pat    = e.get("pattern", "") or e.get("strategy", "")[:80]
+    if not pat:
+        continue  # skip entries with no identifiable pattern
     cand   = e.get("candidate", "unknown")
     prof   = e.get("optimizer_profile", "unknown")
     result = e.get("result", "HELD")