[BUG] Historical tool-call messages missing reasoning_content on multi-turn interleaved models (KV cache invalidation) #28333

Description

Description

When using reasoning models (like Qwen/Qwen3.6-27B or Qwen3.6-35B-A3B) natively with local backends like llama-swap/llama.cpp or via providers like DeepInfra, the KV cache completely invalidates and forces a slow, full prompt re-evaluation on Turn 2 if tool calls were executed in Turn 1.

This happens because OpenCode’s message compiler is stripping or failing to forward the historical reasoning_content blocks inside older assistant tool-call messages when compiling the chat history array. Because the text history passed back to the local runner diverges from what was originally evaluated, prefix-matching fails, causing a total KV cache eviction.

To make matters worse, removing "interleaved" to preserve raw text as a workaround causes a critical race condition where OpenCode accidentally executes tool calls the model was merely "thinking about" inside its reasoning block.

Configuration (opencode.json)

{
  "$schema": "https://opencode.ai/config.json",
  "lsp": true,
  "model": "local/qwen3.6-27b",
  "provider": {
    "local": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Llama-Swap",
      "options": {
        "baseURL": "http://127.0.0.1:8000/v1"
      },
      "models": {
        "qwen3.6-27b": {
          "name": "Qwen3.6-27B",
          "reasoning": true,
          "temperature": true,
          "tool_call": true,
          "interleaved": {
            "field": "reasoning_content"
          },
          "limit": {
            "context": 245760
          }
        }
      }
    }
  }
}

Expected Behavior

OpenCode should preserve and bundle reasoning_content in historical assistant blocks passed to the API. When an assistant message has tool calls, its payload should maintain exact history parity:

{
  "role": "assistant",
  "tool_calls": [...],
  "content": "...",
  "reasoning_content": "..." // This must be preserved, or fallback to "" instead of being stripped
}

Actual Behavior

Historical reasoning_content fields are dropped or set to undefined on prior turns that include tool_calls. This breaks OpenAI-compatible streaming schemas expecting interleaved structure continuity, ruining prefix-matching optimization for high-context local inference setups.

Environment Information

• OpenCode Version: 1.15.3
• Local Runner Platform: llama-swap / llama.cpp / Intel Arc Backend
• Operating System: Linux (nixos)

Plugins

none

OpenCode version

1.15.3

Steps to reproduce

Steps to Reproduce

1. Configure an OpenAI-compatible provider with interleaved tracking for reasoning models (e.g., Qwen3.6-27B with "field": "reasoning_content").
2. Start an OpenCode workspace session and issue a task that requires agentic tool usage (e.g., asking the model to find a bug in a local file or search a workspace directory using an internal tool).
3. Allow the model to step through its thinking process (the TUI will cleanly display the collapsing "Thinking..." dropdown block) and subsequently generate a tool call.
4. Allow OpenCode to execute the workspace tool, collect the return payload, and dispatch it back to the provider.
5. On the subsequent follow-up turn (Turn 2), inspect the network payload or local server logs (e.g., running llama-swap or llama.cpp with verbose tracking):
   • Observed Result: The server logs show a 0% KV cache prefix-match/reuse rate, resulting in a full prompt evaluation lag. Checking the exact incoming array schema reveals that OpenCode sent the historical assistant message with its content string and tool_calls array intact, but completely stripped or dropped the reasoning_content field from that message's history block.

Screenshot and/or share link

No response

Operating System

Linux (nixos 25.11)

Terminal

No response