fix(agent): add debug diagnostics for empty model responses

agent/file_safety.py

@@ -27,6 +27,7 @@ def build_write_denied_paths(home: str) -> set[str]:
             os.path.join(home, ".ssh", "id_ed25519"),
             os.path.join(home, ".ssh", "config"),
             str(hermes_home / ".env"),
+            os.path.join(home, ".hermes", ".env"),
             os.path.join(home, ".bashrc"),
             os.path.join(home, ".zshrc"),
             os.path.join(home, ".profile"),

hermes_cli/web_server.py

@@ -338,6 +338,7 @@ _CATEGORY_MERGE: Dict[str, str] = {
     "human_delay": "display",
     "dashboard": "display",
     "code_execution": "agent",
+    "prompt_caching": "compression",
 }
 
 # Display order for tabs — unlisted categories sort alphabetically after these.

run_agent.py

@@ -12221,6 +12221,29 @@ class AIAgent:
                             _has_structured
                             and self._thinking_prefill_retries >= 2
                         )
+                        if _truly_empty:
+                            _raw_chars = len(final_response or "")
+                            _has_think_blocks = bool(
+                                final_response
+                                and re.search(
+                                    r'<(?:think|thinking|reasoning|thought|REASONING_SCRATCHPAD)\b',
+                                    final_response,
+                                    re.IGNORECASE,
+                                )
+                            )
+                            logger.debug(
+                                "Empty response diagnostics: finish_reason=%r "
+                                "raw_content_chars=%d has_think_blocks=%s "
+                                "structured_reasoning=%s tool_calls=%s model=%s "
+                                "raw_preview=%r",
+                                finish_reason,
+                                _raw_chars,
+                                _has_think_blocks,
+                                _has_structured,
+                                bool(getattr(assistant_message, "tool_calls", None)),
+                                self.model,
+                                (final_response or "")[:300],
+                            )
                         if _truly_empty and (not _has_structured or _prefill_exhausted) and self._empty_content_retries < 3:
                             self._empty_content_retries += 1
                             logger.warning(

tests/gateway/test_agent_cache.py

@@ -969,6 +969,7 @@ class TestAgentCacheIdleResume:
             session_id="hard-session",
         )
 
+        import run_agent as _ra
         vm_calls: list = []
         # AIAgent.close() calls the ``cleanup_vm`` name bound into
         # ``run_agent`` at import time, not ``tools.terminal_tool.cleanup_vm``

tests/hermes_cli/test_skills_config.py

@@ -266,33 +266,33 @@ class TestFindAllSkillsFiltering:
         skills = _find_all_skills()
         assert not any(s["name"] == "my-skill" for s in skills)
 
+    @patch("agent.skill_utils.iter_skill_index_files")
     @patch("tools.skills_tool._get_disabled_skill_names", return_value=set())
     @patch("tools.skills_tool.skill_matches_platform", return_value=True)
-    def test_enabled_skill_included(self, mock_platform, mock_disabled, tmp_path, monkeypatch):
+    @patch("tools.skills_tool.SKILLS_DIR")
+    def test_enabled_skill_included(self, mock_dir, mock_platform, mock_disabled, mock_iter, tmp_path):
         skill_dir = tmp_path / "my-skill"
         skill_dir.mkdir()
         skill_md = skill_dir / "SKILL.md"
         skill_md.write_text("---\nname: my-skill\ndescription: A test skill\n---\nContent")
-        import tools.skills_tool as _st
-        import agent.skill_utils as _su
-        monkeypatch.setattr(_st, "SKILLS_DIR", tmp_path)
-        monkeypatch.setattr(_su, "get_external_skills_dirs", lambda: [])
+        mock_dir.exists.return_value = True
+        mock_iter.return_value = [skill_md]
         from tools.skills_tool import _find_all_skills
         skills = _find_all_skills()
         assert any(s["name"] == "my-skill" for s in skills)
 
+    @patch("agent.skill_utils.iter_skill_index_files")
     @patch("tools.skills_tool._get_disabled_skill_names", return_value={"my-skill"})
     @patch("tools.skills_tool.skill_matches_platform", return_value=True)
-    def test_skip_disabled_returns_all(self, mock_platform, mock_disabled, tmp_path, monkeypatch):
+    @patch("tools.skills_tool.SKILLS_DIR")
+    def test_skip_disabled_returns_all(self, mock_dir, mock_platform, mock_disabled, mock_iter, tmp_path):
         """skip_disabled=True ignores the disabled set (for config UI)."""
         skill_dir = tmp_path / "my-skill"
         skill_dir.mkdir()
         skill_md = skill_dir / "SKILL.md"
         skill_md.write_text("---\nname: my-skill\ndescription: A test skill\n---\nContent")
-        import tools.skills_tool as _st
-        import agent.skill_utils as _su
-        monkeypatch.setattr(_st, "SKILLS_DIR", tmp_path)
-        monkeypatch.setattr(_su, "get_external_skills_dirs", lambda: [])
+        mock_dir.exists.return_value = True
+        mock_iter.return_value = [skill_md]
         from tools.skills_tool import _find_all_skills
         skills = _find_all_skills(skip_disabled=True)
         assert any(s["name"] == "my-skill" for s in skills)

tests/tools/test_file_state_registry.py

@@ -217,8 +217,14 @@ class FileToolsIntegrationTests(unittest.TestCase):
     def setUp(self) -> None:
         file_state.get_registry().clear()
         self._tmpdir = tempfile.mkdtemp(prefix="hermes_file_state_int_")
+        self._orig_terminal_env = os.environ.get("TERMINAL_ENV")
+        os.environ["TERMINAL_ENV"] = "local"
 
     def tearDown(self) -> None:
+        if self._orig_terminal_env is None:
+            os.environ.pop("TERMINAL_ENV", None)
+        else:
+            os.environ["TERMINAL_ENV"] = self._orig_terminal_env
         import shutil
         shutil.rmtree(self._tmpdir, ignore_errors=True)
         file_state.get_registry().clear()

tests/tools/test_modal_sandbox_fixes.py

@@ -13,6 +13,7 @@ import os
 import sys
 from pathlib import Path
 import pytest
+from unittest.mock import patch
 
 # Ensure repo root is importable
 _repo_root = Path(__file__).resolve().parent.parent.parent
@@ -33,6 +34,7 @@ except ImportError:
 class TestToolResolution:
     """Verify get_tool_definitions returns all expected tools for eval."""
 
+    @patch.dict("os.environ", {"TERMINAL_ENV": "local"})
     def test_terminal_and_file_toolsets_resolve_all_tools(self):
         """enabled_toolsets=['terminal', 'file'] should produce 6 tools."""
         from model_tools import get_tool_definitions
@@ -44,6 +46,7 @@ class TestToolResolution:
         expected = {"terminal", "process", "read_file", "write_file", "search_files", "patch"}
         assert expected == names, f"Expected {expected}, got {names}"
 
+    @patch.dict("os.environ", {"TERMINAL_ENV": "local"})
     def test_terminal_tool_present(self):
         """The terminal tool must be present (not silently dropped)."""
         from model_tools import get_tool_definitions

tools/file_tools.py

@@ -157,6 +157,10 @@ def _check_sensitive_path(filepath: str, task_id: str = "default") -> str | None
     except (OSError, ValueError):
         resolved = filepath
     normalized = os.path.normpath(os.path.expanduser(filepath))
+    # /private/var/folders is the macOS per-user temporary directory tree; it is
+    # safe to write and must not be treated as a sensitive system path.
+    if resolved.startswith("/private/var/folders/"):
+        return None
     _err = (
         f"Refusing to write to sensitive system path: {filepath}\n"
         "Use the terminal tool with sudo if you need to modify system files."
@@ -428,7 +432,7 @@ def read_file_tool(path: str, offset: int = 1, limit: int = 500, task_id: str =
         with _read_tracker_lock:
             task_data = _read_tracker.setdefault(task_id, {
                 "last_key": None, "consecutive": 0,
-                "read_history": set(), "dedup": {},
+                "read_history": set(), "dedup": {}, "read_timestamps": {},
             })
             cached_mtime = task_data.get("dedup", {}).get(dedup_key)
 

tools/mcp_tool.py

@@ -2988,8 +2988,9 @@ def _kill_orphaned_mcp_children() -> None:
         except (ProcessLookupError, PermissionError, OSError):
             pass
 
-    # Phase 2: Wait for graceful exit
-    _time.sleep(2)
+    # Phase 2: Wait for graceful exit (skip if nothing was signalled)
+    if pids:
+        _time.sleep(2)
 
     # Phase 3: SIGKILL any survivors
     _sigkill = getattr(_signal, "SIGKILL", _signal.SIGTERM)