diff --git a/devagent/core/llm.py b/devagent/core/llm.py
index 6d6e8a0..3e30c6a 100644
--- a/devagent/core/llm.py
+++ b/devagent/core/llm.py
@@ -971,5 +971,8 @@ def get_llm_for_task(config: DevAgentConfig, task: str) -> LLMClient:
         base_url=config.llm.base_url,
         temperature=config.llm.temperature,
         api_key=config.llm.api_key,
+        effort=config.llm.effort,
+        extended_thinking=config.llm.extended_thinking,
+        thinking_budget_tokens=config.llm.thinking_budget_tokens,
     )
     return LLMClient(routed_cfg)
diff --git a/tests/test_phase5.py b/tests/test_phase5.py
index 2ed544f..e966601 100644
--- a/tests/test_phase5.py
+++ b/tests/test_phase5.py
@@ -345,6 +345,19 @@ def test_router_config_inherits_api_key():
     assert client.cfg.api_key == "sk-test-key"
 
 
+def test_router_config_inherits_effort_and_thinking():
+    """get_llm_for_task keeps effort / extended-thinking settings from the main llm config."""
+    from devagent.core.llm import get_llm_for_task
+    cfg = _make_config(provider="anthropic", model="claude-3-haiku", api_key="sk-test-key")
+    cfg.llm.effort = "max"
+    cfg.llm.extended_thinking = True
+    cfg.llm.thinking_budget_tokens = 4_000
+    client = get_llm_for_task(cfg, "coding")
+    assert client.cfg.effort == "max"
+    assert client.cfg.extended_thinking is True
+    assert client.cfg.thinking_budget_tokens == 4_000
+
+
 def test_router_config_fallback_on_unknown_task():
     """Unknown task falls back to RouterConfig.fallback."""
     from devagent.core.router import MultiModelRouter