[arnav] disable Claude Code's bundled plugin skills in agent sessions

Pass `skills=[]` to ClaudeAgentOptions so the SDK's built-in plugin
skills (/init, /review, /security-review, /simplify, /loop, /schedule,
/update-config, /keybindings-help, /fewer-permission-prompts,
/claude-api) are hidden from the model and rejected by the Skill tool.
These skills are inappropriate in OpenSwarm: half mutate ~/.claude
config files (settings.json, keybindings.json) that OpenSwarm doesn't
read, and the rest expose slash commands the backend never intercepts —
causing the model to falsely claim capabilities it can't actually use.
OpenSwarm's own skills system injects skill content directly into the
user prompt via _resolve_attached_skills, bypassing the Skill tool
entirely, so user-attached skills are unaffected.
Also adds a regression test that pins the `skills=[]` assignment so it
can't be silently dropped in a future refactor.

backend/apps/agents/agent_manager.py

@@ -1921,6 +1921,20 @@ class AgentManager:
                 "type": "preset",
                 "preset": "claude_code",
             }
+            # Suppress Claude Code's bundled plugin skills (/init, /review,
+            # /security-review, /simplify, /loop, /schedule, /update-config,
+            # /keybindings-help, /fewer-permission-prompts, /claude-api).
+            # OpenSwarm has its own skills system that injects skill content
+            # into the user prompt via _resolve_attached_skills, bypassing the
+            # SDK's Skill tool entirely — so the bundled skills only cause
+            # confusion: half mutate ~/.claude state OpenSwarm doesn't use
+            # (settings.json, keybindings.json), and the rest tell the model
+            # it can invoke slash commands the backend doesn't intercept.
+            # Empty list is the SDK's documented "skills off" signal — see
+            # the `skills` field on ClaudeAgentOptions in claude_agent_sdk
+            # types.py. User-attached skills are unaffected because they
+            # don't go through the Skill tool.
+            options_kwargs["skills"] = []
             # exclude_dynamic_sections=True tells the CLI to keep
             # per-user/per-machine grounding (cwd, git status, recent
             # commits, OS info) out of the cached system prompt prefix

backend/tests/test_agent_manager_unit.py

@@ -416,6 +416,26 @@ def test_compose_system_prompt_all_none_returns_none():
     assert AgentManager()._compose_system_prompt(None, None, None) is None
 
 
+def test_run_agent_loop_disables_bundled_claude_code_skills():
+    """Regression guard for `options_kwargs["skills"] = []` in _run_agent_loop.
+
+    The claude-agent-sdk ships a Skill tool that surfaces Claude Code's
+    bundled plugin skills (/init, /review, /security-review, /simplify,
+    /loop, /schedule, /update-config, /keybindings-help,
+    /fewer-permission-prompts, /claude-api). These are inappropriate in
+    OpenSwarm — half mutate ~/.claude config the backend doesn't read, and
+    the rest reference slash commands the backend never intercepts. The
+    SDK's documented "skills off" signal is `skills=[]` on
+    ClaudeAgentOptions; user-attached skills bypass the Skill tool and
+    are unaffected. If this assignment ever gets dropped during a rebase
+    or refactor the bundled skills come back, so we pin it here.
+    """
+    import inspect
+
+    src = inspect.getsource(AgentManager._run_agent_loop)
+    assert 'options_kwargs["skills"] = []' in src
+
+
 def test_resolve_context_paths_empty_returns_empty():
     assert AgentManager()._resolve_context_paths(None) == ""
     assert AgentManager()._resolve_context_paths([]) == ""