fix(hermes): skip Omnigent relay tools in the pre_tool_call hook

omnigent/inner/hermes_policy_hook.py

@@ -54,6 +54,13 @@ def main() -> None:
     tool_name = payload.get("tool_name") or "unknown"
     tool_input = payload.get("tool_input") or {}
 
+    # Omnigent relay tools are already gated when the relay dispatches them back
+    # through the server's tool path; gating them here too parks a duplicate approval
+    # card whose long-poll hangs. Hermes' own tools lack the prefix and stay gated.
+    if tool_name.startswith(("mcp_omnigent_", "mcp__omnigent__")):
+        json.dump({}, sys.stdout)
+        return
+
     # Build the evaluation request matching the server's EvaluationRequest
     # schema.
     eval_body: dict[str, object] = {

tests/inner/test_hermes_policy_hook.py

@@ -0,0 +1,76 @@
+"""Tests for the Hermes pre_tool_call policy hook's relay-tool skip.
+
+Omnigent relay tools (surfaced into Hermes as ``mcp_omnigent_*``) are gated when
+the relay dispatches them back through the server's tool path. The hook must NOT
+gate them a second time (that parks a duplicate approval card whose long-poll
+hangs, wedging the turn). Hermes' own tools are still gated here.
+"""
+
+from __future__ import annotations
+
+import io
+import json
+
+import pytest
+
+from omnigent.inner import hermes_policy_hook
+
+
+@pytest.fixture
+def wired_env(monkeypatch: pytest.MonkeyPatch) -> None:
+    monkeypatch.setenv("_OMNIGENT_SERVER_URL", "http://localhost:6767")
+    monkeypatch.setenv("_OMNIGENT_SESSION_ID", "conv_test")
+
+
+def _run(monkeypatch: pytest.MonkeyPatch, tool_name: str) -> tuple[dict, bool]:
+    """Run the hook with *tool_name*; return (stdout_json, server_was_called)."""
+    called = {"hit": False}
+
+    def _spy(*_args, **_kwargs):
+        called["hit"] = True
+
+        class _R:
+            def json(self) -> dict:
+                return {"result": "POLICY_ACTION_ALLOW"}
+
+        return _R()
+
+    monkeypatch.setattr("omnigent.native_policy_hook.post_evaluate_with_retry", _spy, raising=True)
+    monkeypatch.setattr(
+        "sys.stdin",
+        io.StringIO(json.dumps({"tool_name": tool_name, "tool_input": {}})),
+    )
+    out = io.StringIO()
+    monkeypatch.setattr("sys.stdout", out)
+    hermes_policy_hook.main()
+    return json.loads(out.getvalue() or "{}"), called["hit"]
+
+
+@pytest.mark.parametrize(
+    "tool_name",
+    [
+        "mcp_omnigent_sys_session_get_info",  # hermes single-underscore form
+        "mcp_omnigent_sys_os_write",
+        "mcp__omnigent__list_comments",  # native double-underscore form
+    ],
+)
+def test_relay_tools_are_skipped(
+    monkeypatch: pytest.MonkeyPatch, wired_env: None, tool_name: str
+) -> None:
+    result, server_called = _run(monkeypatch, tool_name)
+    # Allow (empty object) WITHOUT hitting /policies/evaluate — the dispatch
+    # gate is the single authoritative gate for these tools.
+    assert result == {}
+    assert server_called is False
+
+
+@pytest.mark.parametrize(
+    "tool_name", ["terminal", "str_replace_editor", "mcp_github_create_issue"]
+)
+def test_native_and_other_mcp_tools_are_still_gated(
+    monkeypatch: pytest.MonkeyPatch, wired_env: None, tool_name: str
+) -> None:
+    _result, server_called = _run(monkeypatch, tool_name)
+    # Hermes' own tools (and non-Omnigent MCP servers) do NOT round-trip the
+    # relay, so the hook stays their policy gate.
+    assert server_called is True