Tool call content in output

[thomas-0816]

Description

Got tool call output in content:

Plugins

none

OpenCode version

1.14.40

Steps to reproduce

start llama.cpp server (b9041)

./llama-server -hf unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF:Q4_K_XL --threads -1 --parallel 1 --ctx-size 131072 --temp 0.7 --min-p 0.0 --top-p 0.8 --top-k 20 --repeat-penalty 1.05 --predict 131072 --cache-type-k q8_0 --cache-type-v q8_0

set opencode.json

{"$schema": "https://opencode.ai/config.json",
  "enabled_providers": ["llama.cpp"],
  "lsp": true,
  "permission": {"bash": "ask", "edit": "ask"},
  "provider": {
    "llama.cpp": {
      "npm": "@ai-sdk/openai-compatible", "name": "llama-server",
      "options": {"baseURL": "http://127.0.0.1:8080/v1"},
      "models": {
        "llama.cpp": {"name": "llama.cpp", "limit": {"context": 131072, "output": 131072}}
}}}}

start opencode, run a prompt like "which package is currently installed?"

The problem is not happening on every attempt, but ~ 5/10.

output:

see screenshot

Screenshot and/or share link

Operating System

Ubuntu 24.04.4 LTS

Terminal

konsole

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• Progress halts with qwen 3.6 35b-a3b with naked tool call in the console #24316: Similar issue with llama.cpp + Qwen model where naked tool call XML/content appears in the terminal output instead of being executed
• The <tool_call> tag cannot be rendered correctly, results in the conversation being interrupted and failing to proceed automatically #9674: <tool_call> tag not rendered correctly, tool call content leaking into the visible output
• Default agent is brittle against local OpenAI-compatible tool-call quirks #20669: General brittleness of local OpenAI-compatible backends (like llama.cpp) with tool calls

If your issue is distinct from these, please note what's different.

[thomas-0816]

As a workaround, I'm now using legacy functions/function_call format as suggested in https://gist.github.com/daniel-farina/87dc1c394b94e45bb700d27e9ea03193 :

{"$schema": "https://opencode.ai/config.json",
  "enabled_providers": ["llama.cpp"],
  "lsp": true,
  "permission": {"bash": "ask", "edit": "ask"},
  "provider": {
    "llama.cpp": {
      "npm": "@ai-sdk/openai-compatible", "name": "llama-server",
      "options": {"baseURL": "http://127.0.0.1:8080/v1", "toolParser": [{"type": "raw-function-call" }, {"type": "json"}]},
      "models": {
        "llama.cpp": {"name": "llama.cpp", "tool_call": true, "limit": {"context": 131072, "output": 131072}}
}}}}

This workaround only works when -kvu is not set.

[thomas-0816]

In general very onnoying if tool calls work sometimes and sometimes not: