From 907bd88c8f7e23cc57360e34e1fc5a63695d0120 Mon Sep 17 00:00:00 2001
From: Kaloyan Nikolov <konik98@gmail.com>
Date: Wed, 25 Feb 2026 23:56:29 +0100
Subject: [PATCH] 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 | 37 +++++++++++++++++++++++++++----------
 1 file changed, 27 insertions(+), 10 deletions(-)

diff --git a/src/api/chat_handlers.py b/src/api/chat_handlers.py
index ed8b881..31183eb 100644
--- a/src/api/chat_handlers.py
+++ b/src/api/chat_handlers.py
@@ -562,15 +562,32 @@ async def handle_chat_completion(
         iteration += 1
         logger.info(f"--- Tool Execution Iteration {iteration} ---")
 
-        # Generate response (unified interface - handles federation automatically)
-        logger.debug(f"Generating response...")
-        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=federated_swarm
-        )
+        # Generate 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,
+                temperature=request.temperature or 0.7,
+                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)")