fix: federation only on first iteration, local-only for tool result processing

- Critical fix: peers don't have tool results from previous iterations
- Running federation on tool result iterations causes inconsistent context
- Now federation is ONLY used on iteration 1 (initial planning)
- Iterations 2+ are local-only (tool result processing)
- This prevents the infinite ls loop and wrong file hallucinations
- All 41 tests passing

src/api/chat_handlers.py

@@ -562,8 +562,12 @@ async def handle_chat_completion(
         iteration += 1
         logger.info(f"--- Tool Execution Iteration {iteration} ---")
 
-        # Generate response (unified interface - handles federation automatically)
+        # Generate response
-        logger.debug(f"Generating response...")
+        # IMPORTANT: Only use federation on FIRST iteration (initial planning)
+        # Subsequent iterations process tool results which only head node has
+        if iteration == 1 and use_federation:
+            # First iteration: use federation for consensus on initial plan
+            logger.info(f"🌐 Using federation for initial generation...")
             response_text, tokens_generated, tps = await _generate_with_consensus(
                 prompt=prompt,
                 max_tokens=request.max_tokens or 1024,
@@ -571,6 +575,19 @@ async def handle_chat_completion(
                 swarm_manager=swarm_manager,
                 federated_swarm=federated_swarm
             )
+        else:
+            # Subsequent iterations: LOCAL ONLY
+            # Peers don't have tool results from previous iterations
+            # Using federation here would cause inconsistent context
+            if iteration > 1:
+                logger.debug(f"Using local generation (iteration {iteration}, tool context local only)")
+            response_text, tokens_generated, tps = await _generate_with_consensus(
+                prompt=prompt,
+                max_tokens=request.max_tokens or 1024,
+                temperature=request.temperature or 0.7,
+                swarm_manager=swarm_manager,
+                federated_swarm=None  # Force local-only
+            )
 
         logger.info(f"Generated response ({len(response_text)} chars, {tokens_generated} tokens)")
         logger.debug(f"Response: {response_text[:200]}...")
@@ -701,7 +718,7 @@ async def handle_chat_completion(
             max_tokens=request.max_tokens or 1024,
             temperature=request.temperature or 0.7,
             swarm_manager=swarm_manager,
-            federated_swarm=federated_swarm
+            federated_swarm=None  # Tool result processing is local-only
         )
 
         logger.info(f"✅ Generated with tool results ({len(response_text)} chars, {tokens_generated} tokens)")