Replace snapshot_download with level-by-level listing for agent files.

snapshot_download causes timeouts on large repos due to recursive tree
listing. Use the same per-trial list_repo_tree approach as result.json.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

app.py

@@ -30,11 +30,11 @@ from harbor.models.job.result import JobResult
 # Instead, we parse only the specific fields we need from raw dicts.
 from huggingface_hub import (
     HfApi,
+    RepoFile,
     RepoFolder,
     WebhookPayload,
     WebhooksServer,
     hf_hub_download,
-    snapshot_download,
 )
 from litellm import model_cost
 from pydantic import UUID4, BaseModel, ConfigDict, Field, ValidationError
@@ -1131,6 +1131,24 @@ async def _download_submission_files(
         for trial_item in trial_items:
             if isinstance(trial_item, RepoFolder):
                 files_to_download.append(f"{trial_item.path}/result.json")
+                # List agent/* files for reward hacking checks
+                try:
+                    agent_items = await asyncio.to_thread(
+                        lambda tp=trial_item.path: list(
+                            api.list_repo_tree(
+                                repo_id=DATASET_REPO,
+                                path_in_repo=f"{tp}/agent",
+                                repo_type="dataset",
+                                revision=revision,
+                                recursive=True,
+                            )
+                        )
+                    )
+                    for agent_item in agent_items:
+                        if isinstance(agent_item, RepoFile):
+                            files_to_download.append(agent_item.path)
+                except Exception:
+                    pass  # agent/ dir may not exist
 
     logger.info(f"Downloading {len(files_to_download)} files for {submission_name}")
 
@@ -1147,19 +1165,6 @@ async def _download_submission_files(
         for f in files_to_download:
             tg.create_task(_safe_download(f))
 
-    # Download agent/* files for reward hacking checks
-    try:
-        await asyncio.to_thread(
-            snapshot_download,
-            repo_id=DATASET_REPO,
-            repo_type="dataset",
-            revision=revision,
-            local_dir=str(local_dir),
-            allow_patterns=[f"{sub_path}/*/trial-*/agent/*"],
-        )
-    except Exception as e:
-        logger.warning(f"Could not download agent files for {submission_name}: {e}")
-
 
 async def get_changed_submission_names(api: HfApi, pr_revision: str) -> list[str]:
     """Detect which submission folders changed in a PR vs main.