fix(engine): surface action errors to LLM with [ACTION FAILED] prefix

crates/ironclaw_engine/orchestrator/default.py

@@ -741,7 +741,7 @@ def run_loop(context, goal, actions, state, config):
                     output = r.get("output")
                     output_str = str(output) if output is not None else "[no output]"
                     if r.get("is_error"):
-                        output_str = "[ACTION FAILED] " + output_str
+                        output_str = "[ACTION FAILED] " + action_name + ": " + output_str
                         batch_error_count += 1
                     else:
                         batch_success_count += 1
@@ -757,6 +757,10 @@ def run_loop(context, goal, actions, state, config):
                     action_call_id=call_id,
                 )
 
+            # TODO(#2325): track consecutive action errors here, mirroring the
+            # code error tracking above (lines 623-634). Needs a unified
+            # progress-tracking design across both execution paths.
+
             # Check results for auth/approval interrupts
             for r_idx, r in enumerate(results):
                 if r is None:

crates/ironclaw_engine/src/executor/loop_engine.rs

@@ -1278,6 +1278,42 @@ mod tests {
         }
     }
 
+    #[tokio::test]
+    async fn failed_action_result_is_explicit_in_next_llm_message() {
+        let (mut exec, _tx) = make_loop(
+            vec![
+                action_response("test_tool", "call_failed_tool"),
+                text_response("I could not complete that."),
+            ],
+            vec![Ok(ActionResult {
+                call_id: String::new(),
+                action_name: "test_tool".into(),
+                output: serde_json::json!({"error": "No lease for action 'test_tool'"}),
+                is_error: true,
+                duration: Duration::from_millis(1),
+            })],
+            ThreadConfig::default(),
+        )
+        .await;
+
+        exec.run().await.unwrap();
+
+        let action_results: Vec<_> = exec
+            .thread
+            .internal_messages
+            .iter()
+            .filter(|m| m.role == crate::types::message::MessageRole::ActionResult)
+            .collect();
+
+        assert!(
+            action_results.iter().any(|m| {
+                m.content.contains("[ACTION FAILED] test_tool:")
+                    && m.content.contains("No lease for action 'test_tool'")
+            }),
+            "failed ActionResult content should be explicit for the next LLM turn"
+        );
+    }
+
     /// Verify the trace analyzer does NOT flag any issues on a clean
     /// action execution (no empty call_ids).
     #[tokio::test]