DeepSeek V4 Flash via Provider API drops reasoning on tool-call continuation turns (nondeterministic, independent of thinking param) #726

Summary

When using deepseek/deepseek-v4-flash through the CommandCode Provider API, requests whose conversation ends with a tool result message (i.e. a tool-loop continuation) return 0 reasoning tokens — but only sometimes. Identical payloads flip between ~0 and ~20 reasoning tokens on consecutive calls, and the behavior is independent of the thinking parameter (also fails with thinking: {type: "enabled"}).

This is distinct from #697: that issue is about thinking: {type: "disabled"} being ignored (model reasons when you asked it not to). Our issue is the opposite — the model stops reasoning on tool-loop continuations even when thinking is explicitly enabled.

Repro

POST https://api.commandcode.ai/provider/v1/chat/completions, model deepseek/deepseek-v4-flash, with:

{
  "model": "deepseek/deepseek-v4-flash",
  "thinking": { "type": "enabled" },
  "reasoning_effort": "high",
  "max_tokens": 250,
  "stream": false,
  "messages": [
    { "role": "system", "content": "You are an agentic coding assistant. After a tool result comes back, continue by reasoning about it and then take the next action." },
    { "role": "user", "content": "What is the weather?" },
    { "role": "assistant", "content": "", "tool_calls": [{ "id": "call_1", "type": "function", "function": { "name": "get_weather", "arguments": "{}" } }] },
    { "role": "tool", "tool_call_id": "call_1", "content": "{\"temp\":72,\"sunny\":true}" }
  ],
  "tools": [{ "type": "function", "function": { "name": "get_weather", "description": "Get weather", "parameters": { "type": "object", "properties": {} } } }]
}

The last message is the tool result — no trailing user turn. This is the standard shape an agent harness sends to continue after a tool call.

Observed behavior

Run the same body twice, direct to the Provider API, and twice via an intermediate OpenAI-compatible proxy (byte-identical forwarding):

• Single-turn chat (no tools): reasons reliably (123 reasoning tokens).
• Tool-loop continuation (last msg = tool result): reasoning tokens flip between 0 and ~20 on identical payloads.
• Same result with thinking: {type: "disabled"}, thinking: {type: "enabled"}, or no thinking key at all.
• Observed on deepseek/deepseek-v4-flash, poolside/laguna-s-2.1-free (drops to 0), and heavily reduced on deepseek/deepseek-v4-pro and xiaomi/mimo-v2.5-pro. Control: same conversation shape via a different provider (Z.ai) reasons normally.

Expected behavior

Reasoning (or its absence) should be deterministic for a given request shape and controlled by the documented parameters — not flip randomly between calls on identical payloads.

Notes

• The nondeterminism is confirmed upstream: the same byte-identical payload behaves differently when sent direct vs. via a transparent proxy, and differently across identical repeat calls.
• We spent roughly five hours debugging this on our side (routers, proxies, client adapters) and burning API credits before isolating it to the Provider API's handling of tool-loop continuations — wanted to share the repro so others don't repeat that.

Related

• Support non-thinking mode for DeepSeek V4 Flash in Provider API #697 — non-thinking mode not honored (related, but opposite direction)