diff --git a/scripts/src/platform-infra-observability.ts b/scripts/src/platform-infra-observability.ts
index 97e1da13..f8e8f4ca 100644
--- a/scripts/src/platform-infra-observability.ts
+++ b/scripts/src/platform-infra-observability.ts
@@ -1870,6 +1870,7 @@ TRACE_PATH_TEMPLATE = ${JSON.stringify(observability.probes.traceQueryPathTempla
 FULL = ${options.full ? "True" : "False"}
 RAW = ${options.raw ? "True" : "False"}
 LIMIT = ${options.limit}
+CANDIDATE_LIMIT = ${options.candidateLimit}
 DEADLINE = time.time() + 50
 IMPORTANT_ATTRS = [
     "traceId", "otel.trace_id", "agentrun.stage", "runId", "commandId",
@@ -2416,6 +2417,101 @@ def root_cause_candidates(http_summary, agentrun, read_model, lag_summary, error
     candidates.sort(key=lambda item: item.get("confidence", 0), reverse=True)
     return candidates
 
+def candidate_score(trace_id, meta, trace_body, trace_rc, trace_err):
+    parsed = parse_json(trace_body)
+    meta_root_service = str(meta.get("rootServiceName") or "") if isinstance(meta, dict) else ""
+    meta_root_name = str(meta.get("rootTraceName") or "") if isinstance(meta, dict) else ""
+    if not isinstance(parsed, dict):
+        return {
+            "traceId": trace_id,
+            "score": -100,
+            "reasons": ["trace parse failed"],
+            "parseOk": False,
+            "queryOk": trace_rc == 0,
+            "meta": compact_meta(meta),
+            "spanCount": 0,
+            "services": [],
+            "servicePath": {
+                "hwlab-cloud-api": "unknown",
+                "agentrun-manager": "unknown",
+                "agentrun-runner": "unknown",
+                "complete": False,
+            },
+            "rootTraceName": meta_root_name or None,
+            "rootServiceName": meta_root_service or None,
+            "stderrTail": (trace_err or "")[-1000:],
+        }, None
+    flat = flatten_trace(parsed)
+    spans = flat["spans"]
+    services = set(flat["services"])
+    names = [str(item.get("name") or "") for item in spans]
+    lowered_names = [name.lower() for name in names]
+    identity = identity_from_spans(spans)
+    agentrun = agentrun_summary(spans)
+    error_spans = flat["errorSpans"]
+    score = 0
+    reasons = []
+    def add(points, reason):
+        nonlocal score
+        score += points
+        reasons.append("%+d %s" % (points, reason))
+    if "agentrun-runner" in services:
+        add(120, "contains agentrun-runner")
+    if "agentrun-manager" in services:
+        add(90, "contains agentrun-manager")
+    if "hwlab-cloud-api" in services:
+        add(20, "contains hwlab-cloud-api")
+    span_points = min(len(spans), 80)
+    if span_points:
+        add(span_points, "span count %s" % len(spans))
+    for key, points in (("runId", 50), ("commandId", 50), ("runnerJobId", 35), ("runnerId", 30), ("attemptId", 20), ("backendProfile", 15)):
+        if identity.get(key) not in (None, ""):
+            add(points, "identity %s" % key)
+    if any("codex_stdio" in name for name in lowered_names):
+        add(45, "codex stdio span")
+    if any("runner" in name for name in lowered_names):
+        add(30, "runner span")
+    if any("provider" in name for name in lowered_names):
+        add(25, "provider span")
+    if any("tool_call" in name or "command" in name for name in lowered_names):
+        add(20, "tool/command span")
+    if any("turn_completed" in name or "runner_terminal" in name or "command_result" in name for name in lowered_names):
+        add(35, "terminal/result span")
+    if error_spans:
+        add(40 + min(len(error_spans), 20), "error span count %s" % len(error_spans))
+    if agentrun.get("terminalStatus") not in (None, ""):
+        add(25, "terminal status %s" % agentrun.get("terminalStatus"))
+    is_workbench_get_single_span = (
+        len(spans) <= 1
+        and services.issubset({"hwlab-cloud-api"})
+        and meta_root_name.startswith("GET /v1/workbench/")
+    )
+    if is_workbench_get_single_span:
+        add(-80, "single-span Workbench GET/SSE trace")
+    service_path = {
+        service: ("reached" if service in services else "missing")
+        for service in ("hwlab-cloud-api", "agentrun-manager", "agentrun-runner")
+    }
+    service_path["complete"] = all(service in services for service in ("hwlab-cloud-api", "agentrun-manager", "agentrun-runner"))
+    return {
+        "traceId": trace_id,
+        "score": score,
+        "reasons": reasons[:12],
+        "parseOk": True,
+        "queryOk": trace_rc == 0,
+        "meta": compact_meta(meta),
+        "spanCount": len(spans),
+        "services": sorted(services),
+        "servicePath": service_path,
+        "identity": {key: identity.get(key) for key in ("runId", "commandId", "runnerJobId", "runnerId", "backendProfile") if identity.get(key) not in (None, "")},
+        "terminalStatus": agentrun.get("terminalStatus"),
+        "errorSpanCount": len(error_spans),
+        "rootTraceName": meta_root_name or None,
+        "rootServiceName": meta_root_service or None,
+        "spanNamePreview": names[:8],
+        "traceCommand": "bun scripts/cli.ts platform-infra observability trace --target ${target.id} --trace-id %s --limit 80" % trace_id,
+    }, parsed
+
 def resolve_trace():
     if TRACE_ID is not None:
         return TRACE_ID, {
@@ -2433,14 +2529,39 @@ def resolve_trace():
     search_rc, search_body, search_err = run_kubectl(SEARCH_PROXY_PATH, timeout=15)
     search_parsed = parse_json(search_body)
     metas = extract_traces(search_parsed)
+    candidate_summaries = []
     selected = None
     selected_trace_id = None
+    selected_score = None
+    seen = set()
+    scanned = 0
     for meta in metas:
         trace_id = trace_id_from_meta(meta)
-        if trace_id:
+        if not trace_id or trace_id in seen:
+            continue
+        seen.add(trace_id)
+        if scanned >= CANDIDATE_LIMIT or time.time() > DEADLINE:
+            break
+        scanned += 1
+        trace_path = TRACE_PATH_TEMPLATE.replace("{{traceId}}", trace_id)
+        trace_rc, trace_body, trace_err = run_kubectl(TRACE_PROXY_PREFIX + trace_path, timeout=8)
+        summary, _parsed = candidate_score(trace_id, meta, trace_body, trace_rc, trace_err)
+        candidate_summaries.append(summary)
+        score = summary.get("score")
+        if isinstance(score, (int, float)) and (selected_score is None or score > selected_score):
             selected = meta
             selected_trace_id = trace_id
-            break
+            selected_score = score
+    if selected_trace_id is None:
+        for meta in metas:
+            trace_id = trace_id_from_meta(meta)
+            if trace_id:
+                selected = meta
+                selected_trace_id = trace_id
+                selected_score = None
+                break
+    candidate_summaries.sort(key=lambda item: item.get("score", -999999), reverse=True)
+    selected_summary = next((item for item in candidate_summaries if item.get("traceId") == selected_trace_id), None)
     mapping = {
         "mode": "business-trace-id",
         "businessTraceId": BUSINESS_TRACE_ID,
@@ -2450,7 +2571,12 @@ def resolve_trace():
         "searchOk": search_rc == 0,
         "searchParseOk": isinstance(search_parsed, dict),
         "candidateTraceCount": len(metas),
+        "scannedCandidateCount": scanned,
+        "candidateSelectionMode": "scored-service-and-span-content",
+        "selectedScore": selected_score,
+        "selectedReasons": selected_summary.get("reasons", []) if isinstance(selected_summary, dict) else [],
         "selectedMeta": compact_meta(selected),
+        "candidatePreview": candidate_summaries[:8],
         "searchStderrTail": (search_err or "")[-2000:],
     }
     if RAW: