From afa50b0ff7d27592b7e770e25a80d7685b8b65be Mon Sep 17 00:00:00 2001
From: Jon Saad-Falcon <41205309+jonsaadfalcon@users.noreply.github.com>
Date: Sun, 15 Mar 2026 12:12:20 -0700
Subject: [PATCH] fix: parse tool_call arguments as dicts for OpenAI-compatible
 engines (#69)

messages_to_dicts() serializes tool_call arguments as JSON strings,
but OpenAI-compatible servers (vLLM, SGLang, llama.cpp, etc.) expect
them as JSON objects. This caused 400 Bad Request errors on every
multi-turn tool-calling conversation, breaking GAIA, DeepPlanning,
and LifelongAgent benchmarks for all local models.

The fix adds _fix_tool_call_arguments() to _OpenAICompatibleEngine
which json.loads() any string-typed arguments back to dicts before
sending. Applied to both generate() and stream() methods. This is
the same fix as commit 45c0e44 (Ollama), now applied to all
OpenAI-compatible engines.

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---
 src/openjarvis/engine/_openai_compat.py | 26 +++++++++++++++++++++++--
 1 file changed, 24 insertions(+), 2 deletions(-)

diff --git a/src/openjarvis/engine/_openai_compat.py b/src/openjarvis/engine/_openai_compat.py
index 447cd2de..fcc08034 100644
--- a/src/openjarvis/engine/_openai_compat.py
+++ b/src/openjarvis/engine/_openai_compat.py
@@ -32,6 +32,26 @@ class _OpenAICompatibleEngine(InferenceEngine):
 
     # -- InferenceEngine interface ------------------------------------------
 
+    @staticmethod
+    def _fix_tool_call_arguments(msg_dicts: list) -> list:
+        """Ensure tool_call arguments are dicts, not JSON strings.
+
+        OpenAI-compatible servers (vLLM, SGLang, llama.cpp, etc.) expect
+        tool_call arguments as JSON objects.  ``messages_to_dicts`` may
+        serialize them as strings, which causes 400 errors on multi-turn
+        tool-calling conversations.
+        """
+        for md in msg_dicts:
+            for tc in md.get("tool_calls", []):
+                fn = tc.get("function", {})
+                args = fn.get("arguments")
+                if isinstance(args, str):
+                    try:
+                        fn["arguments"] = json.loads(args)
+                    except (json.JSONDecodeError, TypeError):
+                        pass
+        return msg_dicts
+
     def generate(
         self,
         messages: Sequence[Message],
@@ -41,9 +61,10 @@ class _OpenAICompatibleEngine(InferenceEngine):
         max_tokens: int = 1024,
         **kwargs: Any,
     ) -> Dict[str, Any]:
+        msg_dicts = self._fix_tool_call_arguments(messages_to_dicts(messages))
         payload: Dict[str, Any] = {
             "model": model,
-            "messages": messages_to_dicts(messages),
+            "messages": msg_dicts,
             "temperature": temperature,
             "max_tokens": max_tokens,
             "stream": False,
@@ -105,9 +126,10 @@ class _OpenAICompatibleEngine(InferenceEngine):
         max_tokens: int = 1024,
         **kwargs: Any,
     ) -> AsyncIterator[str]:
+        msg_dicts = self._fix_tool_call_arguments(messages_to_dicts(messages))
         payload: Dict[str, Any] = {
             "model": model,
-            "messages": messages_to_dicts(messages),
+            "messages": msg_dicts,
             "temperature": temperature,
             "max_tokens": max_tokens,
             "stream": True,