From 528a3d233746ec4a80dfe88795a61a5791d5cce9 Mon Sep 17 00:00:00 2001
From: Jason Matthew <jasonm4130@gmail.com>
Date: Thu, 30 Apr 2026 20:54:09 +1000
Subject: [PATCH] fix: disable thinking on Moonshot calls so kimi-k2.6
 populates content
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

kimi-k2.6 is a reasoning model that writes its chain-of-thought to
`reasoning_content` and the answer to `content`. With graphify's default
`max_completion_tokens=8192`, large chunks (the default 20-file batches
frequently produce 50k+ token user messages) cause reasoning to consume
the entire output budget — `content` comes back empty, graphify parses
zero nodes, and the user pays for input + reasoning tokens without
getting a graph.

Reproduction on a 162-file repo (~125k words):
- Before: 9 chunks × 8192 tokens = ~74k completion tokens → 2 nodes total
- After:  same chunks → 138 nodes (chunk 1 alone: 42 nodes / 75 edges)

Moonshot's API exposes `extra_body={"thinking": {"type": "disabled"}}`.
For structured graph extraction we don't need chain-of-thought, so disable
it for Moonshot endpoints only — other OpenAI-compatible providers (real
OpenAI, Together, Groq) would 400 on the flag.
---
 graphify/llm.py   |  8 +++++++
 tests/test_llm.py | 57 +++++++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 65 insertions(+)
 create mode 100644 tests/test_llm.py

diff --git a/graphify/llm.py b/graphify/llm.py
index f07f8ec06..c21fb22a6 100644
--- a/graphify/llm.py
+++ b/graphify/llm.py
@@ -102,6 +102,14 @@ def _call_openai_compat(
     }
     if temperature is not None:
         kwargs["temperature"] = temperature
+    # kimi-k2.6 is a reasoning model: it writes its chain-of-thought to
+    # `reasoning_content` and the answer to `content`. When reasoning consumes
+    # the entire max_completion_tokens budget on large chunks, `content` stays
+    # empty and we parse zero nodes while still paying for the input + reasoning
+    # tokens. For structured extraction we don't need chain-of-thought, so
+    # disable thinking on Moonshot-hosted models.
+    if "moonshot" in base_url:
+        kwargs["extra_body"] = {"thinking": {"type": "disabled"}}
     resp = client.chat.completions.create(**kwargs)
     result = _parse_llm_json(resp.choices[0].message.content or "{}")
     result["input_tokens"] = resp.usage.prompt_tokens if resp.usage else 0
diff --git a/tests/test_llm.py b/tests/test_llm.py
new file mode 100644
index 000000000..71c300637
--- /dev/null
+++ b/tests/test_llm.py
@@ -0,0 +1,57 @@
+"""Tests for the direct LLM backend (Claude / Kimi)."""
+from unittest.mock import MagicMock, patch
+
+
+def _mock_response(content: str = '{"nodes":[],"edges":[]}', completion_tokens: int = 10):
+    """Build a fake OpenAI ChatCompletion response."""
+    resp = MagicMock()
+    resp.choices = [MagicMock(message=MagicMock(content=content))]
+    resp.usage = MagicMock(prompt_tokens=100, completion_tokens=completion_tokens)
+    return resp
+
+
+def test_kimi_call_disables_thinking():
+    """kimi-k2.6 is a reasoning model: when thinking is enabled and the chunk
+    is large, all of max_completion_tokens is consumed by reasoning_content,
+    leaving content empty and graphify parsing zero nodes. Disable thinking on
+    Moonshot endpoints so content is always populated for structured extraction.
+    """
+    from graphify.llm import _call_openai_compat
+
+    fake_client = MagicMock()
+    fake_client.chat.completions.create.return_value = _mock_response()
+
+    with patch("openai.OpenAI", return_value=fake_client):
+        _call_openai_compat(
+            base_url="https://api.moonshot.ai/v1",
+            api_key="sk-test",
+            model="kimi-k2.6",
+            user_message="=== example.py ===\ndef hello(): pass\n",
+            temperature=None,
+        )
+
+    kwargs = fake_client.chat.completions.create.call_args.kwargs
+    assert "extra_body" in kwargs, "moonshot calls must pass extra_body to disable thinking"
+    assert kwargs["extra_body"] == {"thinking": {"type": "disabled"}}
+
+
+def test_non_moonshot_call_does_not_disable_thinking():
+    """The thinking-disabled flag is Moonshot-specific extra_body. Other
+    OpenAI-compatible providers (real OpenAI, Together, Groq, etc.) don't
+    accept it and would 400. Only set it for Moonshot URLs."""
+    from graphify.llm import _call_openai_compat
+
+    fake_client = MagicMock()
+    fake_client.chat.completions.create.return_value = _mock_response()
+
+    with patch("openai.OpenAI", return_value=fake_client):
+        _call_openai_compat(
+            base_url="https://api.openai.com/v1",
+            api_key="sk-test",
+            model="gpt-4o",
+            user_message="hi",
+            temperature=0,
+        )
+
+    kwargs = fake_client.chat.completions.create.call_args.kwargs
+    assert "extra_body" not in kwargs, "non-Moonshot calls must not pass thinking flag"