Skip to content

Disable LLM thinking via OPENAI_DISABLE_THINKING #2

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
kingkingkingson wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Disable LLM thinking via OPENAI_DISABLE_THINKING #2

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
30 changes: 28 additions & 2 deletions trigfuzz/llm_client.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,6 +9,18 @@
chat-completions endpoints are also supported by setting
`OPENAI_API_BASE_URL` or `OPENAI_BASE_URL` to a chat-completions style
endpoint.

Thinking/reasoning control
--------------------------
Set ``OPENAI_DISABLE_THINKING=1`` to suppress the model's internal
chain-of-thought before it emits the final answer.

* Responses API path – passes ``reasoning={"effort": "none"}`` per
https://api-docs.deepseek.com/zh-cn/api/create-response
* Chat completions path – passes ``extra_body={"enable_thinking": False}``
(DeepSeek chat extension).

Without this flag both paths use the model's default thinking behaviour.
"""

from __future__ import annotations
Expand All @@ -21,6 +33,7 @@
_DEFAULT_MODEL = os.environ.get("OPENAI_MODEL", "gpt-5.5")
_DEFAULT_MAX_OUTPUT_TOKENS = int(os.environ.get("OPENAI_MAX_OUTPUT_TOKENS", "4096"))
_CHAT_ENDPOINT_TAIL = ("chat", "completions")
_DISABLE_THINKING = os.environ.get("OPENAI_DISABLE_THINKING", "").strip() in ("1", "true", "yes")


class LLMClient:
Expand Down Expand Up @@ -69,22 +82,35 @@ def query(self, messages: list[dict[str, Any]]) -> str:
if self._use_chat_completions:
return self._query_chat_completions(messages)

resp = self._client.responses.create(
kwargs: dict[str, Any] = dict(
model=self._model,
instructions=SYSTEM_PROMPT,
input=self._to_response_input(messages),
max_output_tokens=_DEFAULT_MAX_OUTPUT_TOKENS,
)
if _DISABLE_THINKING:
# Responses API: effort="none" disables the reasoning chain.
# ref: https://api-docs.deepseek.com/zh-cn/api/create-response
kwargs["reasoning"] = {"effort": "none"}
resp = self._client.responses.create(**kwargs)
return self._extract_text(resp)

def _query_chat_completions(self, messages: list[dict[str, Any]]) -> str:
chat_messages = [{"role": "system", "content": SYSTEM_PROMPT}]
chat_messages.extend(self._to_chat_messages(messages))
resp = self._client.chat.completions.create(
kwargs: dict[str, Any] = dict(
model=self._model,
messages=chat_messages,
max_tokens=_DEFAULT_MAX_OUTPUT_TOKENS,
)
if _DISABLE_THINKING:
# DeepSeek chat-completions extension for disabling thinking.
# Newer models (e.g. deepseek-v4-flash) require the nested
# {"thinking": {"type": "disabled"}} format; the older
# {"enable_thinking": False} form is ignored and causes all
# completion tokens to be consumed by the reasoning chain.
kwargs["extra_body"] = {"thinking": {"type": "disabled"}}
resp = self._client.chat.completions.create(**kwargs)
return self._extract_chat_text(resp)

@staticmethod
Expand Down