Bug: DeepSeek‑V4‑Flash with thinking mode enabled throws reasoning_content must be passed back to the API in multi‑turn tool‑call workflow #96

Environment

• OS: Windows 11
• Python: 3.11
• LiteLLM: latest
• LLM Backend: third‑party OpenAI‑compatible proxy（tokenrhythm.studio/v1）
• Model: deepseek‑v4‑flash‑0731 (thinking mode enabled by default)
• PentestAgent config (.env):

OPENAI_API_KEY=sk_tr_xxx
OPENAI_API_BASE=https://tokenrhythm.studio/v1/
PENTESTAGENT_MODEL=openai/deepseek‑v4‑flash‑0731
PENTESTAGENT_EMBEDDINGS=openai

Describe the bug

When running penetration‑testing tasks with multi‑turn tool calls, after the first round of model response, subsequent requests fail immediately:

litellm.BadRequestError: OpenAIException - The `reasoning_content` in the thinking mode must be passed back to the API.

Root cause analysis

1. DeepSeek‑V4‑Flash thinking mode returns an extra reasoning_content field inside assistant message, which must be preserved and sent back within assistant message on next request.
2. Current PentestAgent implementation in llm.py / memory.py discards the reasoning_content field when building / compressing conversation history.
3. On the next API call, the reasoning_content is missing from assistant message payload, proxy returns 400 bad request.
4. Adding litellm parameter preserve_reasoning_content into request dict does not fix this problem; it will cause another error unknown request field: preserve_reasoning_content, because this litellm‑specific parameter is mistakenly passed to upstream proxy.

Reproduction steps

1. Configure .env to use deepseek‑v4‑flash‑0731 via third‑party compatible proxy.
2. Start PentestAgent.
3. Submit a penetration‑testing task that requires multiple rounds of tool calling.
4. First LLM response succeeds; second round request triggers this error and agent stops.

Expected behavior

• Either automatically preserve reasoning_content field for DeepSeek thinking‑enabled models in message history.
• Or provide a config option to disable DeepSeek thinking mode via extra_body={"thinking":{"type":"disabled"}}.
• Agent should continue multi‑turn workflow normally without 400 error.

Workaround available

Switch to non‑thinking model e.g. openai/gpt‑4o‑mini in .env, avoid DeepSeek‑V4‑Flash thinking mode.

Additional context

• This issue only occurs on DeepSeek‑V4 series with thinking mode enabled, third‑party compatible proxy environment.
• Clearing __pycache__ cannot resolve this bug, it is pure application‑level logic defect.