fix(agent_loop): apply agent_before hook return value to context

The plugin hook system in plugins/hooks.py already supports returning a
dict to mutate the agent context, but agent_runner_loop() discarded the
return value of _hook('agent_before', ...). This prevented plugins from
modifying system_prompt / user_input / initial_user_content.

Capture the return value; when a dict is returned, apply its overrides
to the messages list. Non-dict returns and missing keys are ignored, so
existing hooks that only inspect ctx (e.g. logging/metrics) continue to
work without modification.

Closes #537

Author: KaiLi

agent_loop.py

@@ -46,7 +46,15 @@ def agent_runner_loop(client, system_prompt, user_input, handler, tools_schema,
         {"role": "user", "content": initial_user_content if initial_user_content is not None else user_input}
     ]
     turn = 0;  handler.max_turns = max_turns
-    _hook('agent_before', locals())
+    _ctx = _hook('agent_before', locals())  # plugins may return dict to override system_prompt/user_input
+    if isinstance(_ctx, dict):
+        if 'system_prompt' in _ctx:
+            system_prompt = _ctx['system_prompt']
+            messages[0]['content'] = system_prompt
+        if _ctx.get('initial_user_content') is not None:
+            messages[1]['content'] = _ctx['initial_user_content']
+        elif 'user_input' in _ctx and initial_user_content is None:
+            messages[1]['content'] = _ctx['user_input']
     while turn < handler.max_turns:
         turn += 1; turnstr = f'LLM Running (Turn {turn}) ...'
         if handler.parent.task_dir: turnstr = f'Turn {turn} ...'