Add CWE/model sidebar selectors + scan logs/ subdirs

app.py

@@ -14,10 +14,10 @@ from pathlib import Path
 
 import streamlit as st
 
-# Local: scripts/trajectory_viewer.py → repo_root/logs/github_red
-# Space: /app/app.py → /app/logs/github_red
-_local_logs = Path(__file__).parent.parent / "logs" / "github_red"
-_space_logs = Path(__file__).parent / "logs" / "github_red"
+# Local: scripts/trajectory_viewer.py → repo_root/logs
+# Space: /app/app.py → /app/logs
+_local_logs = Path(__file__).parent.parent / "logs"
+_space_logs = Path(__file__).parent / "logs"
 LOGS_DIR = _space_logs if _space_logs.exists() else _local_logs
 
 # ── Colours & labels ────────────────────────────────────────────────────────
@@ -707,6 +707,29 @@ def render_steps(steps: list) -> None:
                 )
 
 
+# ── Log index ─────────────────────────────────────────────────────────────────
+
+
+@st.cache_data(show_spinner=False)
+def _log_index(logs_dir: str) -> dict[str, dict[str, list[str]]]:
+    """Scan logs_dir recursively; return {cwe: {model: [path, ...]}} newest-first."""
+    index: dict[str, dict[str, list[str]]] = {}
+    for p in sorted(Path(logs_dir).rglob("*.eval"), reverse=True):
+        cwe = p.parent.name  # directory name used as CWE/task label
+        model = "unknown"
+        try:
+            with zipfile.ZipFile(str(p)) as z:
+                if "header.json" in z.namelist():
+                    h = json.loads(z.read("header.json"))
+                    usage = h.get("stats", {}).get("model_usage", {})
+                    if usage:
+                        model = next(iter(usage))
+        except Exception:
+            pass
+        index.setdefault(cwe, {}).setdefault(model, []).append(str(p))
+    return index
+
+
 # ── Main app ──────────────────────────────────────────────────────────────────
 
 
@@ -735,12 +758,17 @@ def main() -> None:
         if uploaded is not None:
             eval_source = uploaded
         else:
-            log_files = sorted(glob.glob(str(LOGS_DIR / "*.eval")), reverse=True)
-            if log_files:
+            index = _log_index(str(LOGS_DIR))
+            if index:
+                cwe_options = sorted(index.keys())
+                selected_cwe = st.selectbox("CWE / Task", cwe_options)
+                model_options = sorted(index[selected_cwe].keys())
+                selected_model = st.selectbox("Model", model_options)
+                file_list = index[selected_cwe][selected_model]
                 selected_path = st.selectbox(
-                    "Or pick a local run log",
-                    log_files,
-                    format_func=lambda p: Path(p).stem[:50],
+                    "Run",
+                    file_list,
+                    format_func=lambda p: Path(p).stem[:60],
                 )
                 eval_source = selected_path