Skip to content

01_reasoning_content.md

Latest commit

 

History

History
64 lines (50 loc) · 2.66 KB

01_reasoning_content.md

File metadata and controls

64 lines (50 loc) · 2.66 KB
Raw

§1 — reasoning_content lifecycle

1.1 Symptom (verbatim from probe_2 trial 0, 2026-05-09 official endpoint)

HTTP 400 Bad Request
{
  "error": {
    "message": "The `reasoning_content` in the thinking mode must be passed back to the API.",
    "type": "invalid_request_error",
    "param": null,
    "code": "invalid_request_error"
  }
}

Triggered when a multi-turn loop's PRIOR assistant message (the one carrying tool_calls) is replayed to DeepSeek without its original reasoning_content field.

Reproduction rate (probe_2, n=3, model=deepseek-v4-pro): 3/3 = 100%.

Important corollary: this bug fires ONLY when thinking is enabled (which is the default for deepseek-v4-pro). If your client sets extra_body={"thinking":{"type":"disabled"}} it will never see this 400, but it also gives up the model's deep-reasoning behaviour.

1.2 Citations

Source What it says
microsoft/agent-framework#5538 reasoning_content must be echoed back in tool-call loops
NousResearch/hermes-agent#15353 tool-call message missing reasoning_content → 400
cline/cline PR #7888 adds addReasoningContent + isNextGenModelProvider list including DeepSeek — gold-standard fix
cline/cline#8365, #8130 V3.2/V4 ALSO pack XML tool calls inside reasoning_content (cross-cuts §2)

Probes: probe_2_reasoning_lifecycle phase B.

1.3 Normative rules

  1. The adapter MUST preserve the reasoning_content field on every assistant message it receives from DeepSeek thinking-mode.
  2. When the adapter re-sends an assistant message that has tool_calls != null AND the IMMEDIATELY-NEXT message in the request body has role == "tool", the assistant message MUST carry the original reasoning_content.
  3. Across user-turn boundaries (i.e. once a NEW message with role == "user" arrives), the adapter SHOULD strip reasoning_content from prior assistant messages, because:
    • the server does not require it across turns, AND
    • keeping it bloats the prefix and breaks byte-for-byte cache equality (§4).
  4. Streaming chunks MAY contain delta.reasoning_content even when delta.content is null. The adapter MUST NOT skip such chunks.
  5. The terminal SSE chunk MAY have choices == [] and only carry usage. The adapter MUST tolerate this. (Cross-ref §5, cline #1594.)

1.4 Reference implementation

src/deepseek_v4pro_kit/reasoning.py::ReasoningLifecycle and normalize.prepare_for_new_user_turn.

1.5 Negative test

probe_2's phase B is allowed to fail (we WANT the 400). If it stops failing, we have evidence DeepSeek changed the contract — open an issue and re-spec.