DeepSeek Reasoning Replay

Some DeepSeek-based providers (e.g., via OpenCode, ZenGo) require previous assistant reasoning content to be included in subsequent requests. This is because DeepSeek splits its output into visible content and internal reasoning/chain-of-thought, and the reasoning must be fed back into the conversation context for the model to maintain coherence across turns.

The reasoning transformer handles this automatically by capturing reasoning output from responses and replaying it in the next request.

How It Works

1. On each response, the reasoning transformer intercepts reasoning_content fields (e.g., DeepSeek's <reasoning_content>...</reasoning_content> blocks)
2. It stores the reasoning text in the conversation context
3. On the next request to the same model, it injects the stored reasoning as a reasoning block in the messages array
4. This ensures the model receives its own prior reasoning as context

Configuration

The reasoning transformer is typically configured per-model in your provider config:

{
  "Providers": [
    {
      "name": "opencode",
      "baseUrl": "https://api.opencode.com/v1/chat/completions",
      "apiKey": "$OPENCODE_API_KEY",
      "models": ["deepseek-v4"],
      "transformer": {
        "use": ["opencode-headers"],
        "deepseek-v4": {
          "use": ["reasoning"]
        }
      }
    }
  ]
}

Which Providers Need It

Provider	Requires Reasoning Replay	Notes
DeepSeek (official)	✅	Must replay reasoning_content across turns
OpenCode	✅	Uses DeepSeek models with reasoning
ZenGo	✅	DeepSeek-based, requires reasoning replay
Other OpenAI-compatible	❌	Standard Chat Completions, no reasoning field

Combining with Other Transformers

The reasoning transformer should run after other request modifiers. Order matters in transformer.use:

{
  "use": ["opencode-headers", "reasoning"]
}

The reasoning replay must see the final message list to inject the reasoning block in the right position.

Troubleshooting

Model forgets context across turns: If the model repeats itself or loses track, the reasoning transformer may not be active. Verify it's in the model-level use array.

Reasoning content visible to user: The transformer only replays reasoning internally — it should not appear in Claude Code's output. If it does, check the transformResponseOut logic.