Skip to content

Personal AI, On Personal Devices

Python
1011 B
Raw Blame
[recipe]
name = "terminalbench-react"
kind = "discrete"
description = "ReAct agent tuned for terminal and shell tasks — benchmarks against TerminalBench"
version = "0.1.0"

[intelligence]
model = "qwen3:8b"
quantization = "q4_K_M"

[engine]
key = "ollama"

[agent]
type = "native_react"
max_turns = 20
temperature = 0.2
tools = ["shell_exec", "file_read", "file_write", "think"]
system_prompt = """You are an expert at terminal and shell operations. You solve tasks by reasoning step-by-step and executing shell commands.

When given a task:
1. Analyze what needs to be done
2. Plan your approach
3. Execute commands one at a time, observing the output
4. Verify the result before concluding

You have access to a full Linux shell environment. Use standard Unix tools (grep, sed, awk, find, etc.) and scripting when appropriate. Always check command exit codes and handle errors gracefully."""

[learning]
routing = "heuristic"

[eval]
benchmarks = ["terminalbench"]
backend = "jarvis-agent"
max_samples = 50