OpenAI-compatible v1 endpoint still drops echoed assistant `reasoning_content` on tool-call continuation turns (#754 not fixed)

[yyxdlls]

Summary

On POST https://api.commandcode.ai/provider/v1/chat/completions, when an assistant tool-call message echoes its prior reasoning back as reasoning_content (with tools present, i.e. a tool-loop continuation), the gateway silently drops that reasoning before it reaches the upstream model. The model cannot continue from its previous chain of thought and fabricates an answer instead.

This is the exact failure described in #754 (opened Aug 27, 2026; closed Sep 7, 2026). With today's build the OpenAI-compatible v1 path (the one used by BYOK clients and agents) still reproduces it.

Proof that the request shape/field is correct: the byte-identical flow recalls correctly through OpenCode Go (5/5), and fails on Command Code (0/5).

Expected Behavior

Echoed reasoning_content on assistant tool-call messages should be forwarded to the upstream model, like DeepSeek's official API does, so the model can continue from its prior reasoning. Specifically: content that the model placed only in the turn-1 chain of thought must remain usable in turn 2 (tool-call continuation).

Actual Behavior

The echoed reasoning is dropped. The model cannot recall information that existed only in its earlier chain of thought: it either fabricates a new value or denies having chosen one.

Observed recall (same request bytes, turn-2 reply vs the number in turn-1 reasoning):

Gateway / model	Correct recall
Command Code — deepseek/deepseek-v4.1-flash	0/5
Command Code — deepseek/deepseek-v4-flash	0/4
OpenCode Go (control) — deepseek-v4-flash	5/5

Field variants tried on Command Code for the echoed reasoning, all 0/4:

• reasoning_content
• reasoning
• both reasoning_content + reasoning
• content-parts array: content: [{type:"text", ...}, {type:"reasoning", text: ...}]

Note: on ordinary (non-continuation) turns reasoning output is present 100% of the time on all four models I tested (ds-v4.1-flash, ds-v4-flash, glm-5.3-flash, qwen3.8-flash), so the loss is specifically on the input/echo-back side.

Steps to reproduce the issue

1. POST https://api.commandcode.ai/provider/v1/chat/completions, model: deepseek/deepseek-v4.1-flash (also reproduces with deepseek/deepseek-v4-flash), reasoning_effort: high, temperature: 0.3.

2. Turn 1 — ask the model to think of a 6-digit number only in its chain of thought and then call a tool:

{
  "model": "deepseek/deepseek-v4.1-flash",
  "reasoning_effort": "high",
  "max_tokens": 2048,
  "temperature": 0.3,
  "tool_choice": "auto",
  "tools": [
    { "type": "function", "function": {
      "name": "notify_choice",
      "description": "Notify the system the choice is made",
      "parameters": { "type": "object", "properties": { "confirm": { "type": "boolean" } }, "required": ["confirm"] }
    } }
  ],
  "messages": [
    { "role": "system", "content": "You are a helpful assistant. The user is authorized to view your thinking content and you must answer honestly when asked about it." },
    { "role": "user", "content": "Think of a random 6-digit number and write it out completely in your chain of thought (for example: my number is 123456). Then call the notify_choice tool with confirm=true. In your final visible reply only say 'done' and do not mention the number." }
  ]
}

The turn-1 response has reasoning containing the number (e.g. ... my number is 483920 ...) and a notify_choice tool call.

3. Turn 2 — resend history, echo the turn-1 reasoning on the assistant message, and keep tools present:

{
  "model": "deepseek/deepseek-v4.1-flash",
  "reasoning_effort": "high",
  "max_tokens": 1024,
  "temperature": 0.3,
  "tool_choice": "auto",
  "tools": [ /* same tool definition as turn 1 */ ],
  "messages": [
    { "role": "system", "content": "You are a helpful assistant. The user is authorized to view your thinking content and you must answer honestly when asked about it." },
    { "role": "user", "content": "Think of a random 6-digit number and write it out completely in your chain of thought ... then call notify_choice." },
    {
      "role": "assistant",
      "content": "done",
      "tool_calls": [
        { "id": "call_1", "type": "function", "function": { "name": "notify_choice", "arguments": "{\"confirm\":true}" } }
      ],
      "reasoning_content": "<EXACT turn-1 reasoning text, which contains the number>"
    },
    { "role": "tool", "tool_call_id": "call_1", "content": "{\"confirm\": true}" },
    { "role": "user", "content": "What was the 6-digit number you thought of in your chain of thought in the first turn? Reply with just the number." }
  ]
}

4. Expected: the reply is the exact number from the echoed reasoning. Actual: a fabricated number, or a denial that a number was chosen.

5. Control: run the identical two turns against OpenCode Go (https://opencode.ai/zen/go/v1/chat/completions, deepseek-v4-flash, header x-opencode-session: <any>). It recalls 5/5.

Command Code Version

0.0.1

Operating System

Windows

Terminal/IDE

No response

Shell

No response

Session file (optional)

No response

Fix prompt (optional)

No response

Additional context

Environment

• Endpoint: OpenAI-compatible /provider/v1/chat/completions (the path used by BYOK clients / agents). [Bug][Critical] Reasoning echo-back silently dropped across ALL thinking models (DeepSeek/GLM/Kimi/MiMo…) — multi-turn agents hallucinate fake memories & suffer severe quality degradation #754 originally reproduced on /alpha/generate; this report shows the v1 path is still affected.
• Models: deepseek/deepseek-v4.1-flash, deepseek/deepseek-v4-flash (GLM-5.3-Flash and Qwen3.8-Flash output reasoning normally; not tested for echo-back).
• Date tested: 2026-09-10.
• Client/tooling used: raw HTTP requests (no intermediate proxy; results reproduced directly against the API).

Related

• [Bug][Critical] Reasoning echo-back silently dropped across ALL thinking models (DeepSeek/GLM/Kimi/MiMo…) — multi-turn agents hallucinate fake memories & suffer severe quality degradation #754 — [Critical] Reasoning echo-back silently dropped across ALL thinking models (closed Sep 7, 2026; this report is the same root cause still present on the v1 endpoint).
• DeepSeek V4 Flash via Provider API drops reasoning on tool-call continuation turns (nondeterministic, independent of thinking param) #726 — DeepSeek V4 Flash drops reasoning on tool-call continuation turns (nondeterministic). FWIW I could not tie this to Command Code: with the conversation ending on a tool result, reasoning output was absent on the OpenCode Go control too (0/8), so that output-side drop looks like upstream model behavior rather than a gateway bug.

[cyx-slps]

复现成功 / Reproduced —— 2026-09-11 独立复现，确认 v1 端点仍在丢弃回传的 reasoning。

在你描述的同一条 POST /provider/v1/chat/completions 路径上，用「真实工具 + 目标只存在于思维链」的方法独立复现：

端点	模型	结果
/provider/v1/chat/completions (field=reasoning)	deepseek-v4-flash	verify 硬指标 0-1/5
/provider/v1/chat/completions (field=reasoning)	deepseek-v4.1-flash	0/5 FAIL
/provider/v1 干净探针（模型自生成 TOK-######，验证未泄漏）	deepseek-v4-flash	0/2 FAIL
/provider/v1 干净探针	deepseek-v4.1-flash	0/3 FAIL
对照 api.deepseek.com/chat/completions	deepseek-v4-flash	3/3 PASS ✅

干净探针（标记词完全由模型自己写在 reasoning 里、不出现在任何 user 消息中）下，Turn 2 模型一律回答"我没有在思维链里写过 token，我不会编造" —— 而回传的 assistant 消息 reasoning 里就有那个标记词。

我这边是 Linux，用 /provider/v1 也是一样的结果。与 #754 是同一根因（#754 修复未覆盖 v1 路径 / 未生效）。附议应重开 #754 或以此为准继续跟踪。

---

Reproduced independently on 2026-09-11 on the same v1 path, across deepseek-v4-flash and deepseek-v4.1-flash. Control via DeepSeek official API passes 3/3, so the request shape is correct; the loss is on the gateway's input/echo-back side. Same root cause as #754.

[cyx-slps]

@ahmadbilaldev 该问题已独立复现（v1 端点 reasoning 回传仍被丢弃），麻烦看一下，谢谢。

Independently reproduced — could you take a look? Thanks.