MicroHarness

中文版

Minimal Agent Harness built on LangGraph. Multi-provider, safety-guarded, memory-enabled.

Features

• Multi-provider support (9 providers: Anthropic, OpenAI, DeepSeek, Kimi, Qwen, GLM, MiniMax, Xiaomi, Custom)
• 3-tier tool safety guard (auto-approve / always-confirm / keyword-check)
• Cross-session long-term memory (persisted to memory.json)
• Minimal 6-layer architecture (~500 lines total)

Quick Start

# 1. Install core dependencies
pip install -r requirements.txt

# For non-Anthropic providers, also install:
pip install langchain-openai

# 2. Edit .env — set your Provider and API Key
# 3. Run
python harness.py

Usage Examples

Example 1: Write and run a Python script

Task: Write a Python script that prints Fibonacci numbers up to 100, then run it

[HARNESS] Step 1/10 — Agent thinking...
[HARNESS] Step 2/10 — Agent thinking...
=======================================================
  FINAL RESPONSE
=======================================================
I wrote fib.py and ran it. Output: 1 1 2 3 5 8 13 21 34 55 89
=======================================================
  Total steps used: 2/10
=======================================================
[HARNESS] Memory saved: User asked to write and run a Fibonacci script.

Example 2: Resume with long-term memory

On the second run, MicroHarness automatically loads the previous session summary:

[HARNESS] Found 1 long-term memory record(s).
          Last: 2024-01-15 — User asked to write and run a Fibonacci script...

Task: Improve the script from last time

[HARNESS] Step 1/10 — Agent thinking...
(Agent recalls previous session context from injected memory)

Example 3: Safety guard blocks a dangerous write

When the agent attempts to write code containing dangerous operations, the guard pauses and requests human confirmation:

[HARNESS] Step 1/10 — Agent thinking...

=======================================================
  [HARNESS GUARD] 📝 WRITE OP
  Tool   : write_file
  filename: cleanup.py
  content : import shutil; shutil.rmtree('/tmp/sandbox')
=======================================================
  Approve? (yes / no): no
  ❌ Rejected. Operation cancelled.

Provider Configuration

Edit .env to select your provider and supply the API key:

Anthropic (default)

PROVIDER=anthropic
ANTHROPIC_API_KEY=your_key
MAIN_MODEL=claude-sonnet-4-20250514
MEMORY_MODEL=claude-haiku-4-5-20251001

DeepSeek

PROVIDER=deepseek
OPENAI_COMPATIBLE_API_KEY=your_key
MAIN_MODEL=deepseek-chat
MEMORY_MODEL=deepseek-chat

Kimi (Moonshot)

PROVIDER=kimi
OPENAI_COMPATIBLE_API_KEY=your_key
MAIN_MODEL=moonshot-v1-8k
MEMORY_MODEL=moonshot-v1-8k

Qwen (Alibaba)

PROVIDER=qwen
OPENAI_COMPATIBLE_API_KEY=your_key
MAIN_MODEL=qwen-plus
MEMORY_MODEL=qwen-turbo

GLM (Zhipu)

PROVIDER=glm
OPENAI_COMPATIBLE_API_KEY=your_key
MAIN_MODEL=glm-4-air
MEMORY_MODEL=glm-4-flash

MiniMax

PROVIDER=minimax
OPENAI_COMPATIBLE_API_KEY=your_key
MAIN_MODEL=abab6.5s-chat
MEMORY_MODEL=abab6.5s-chat

OpenAI / GPT

PROVIDER=openai
OPENAI_COMPATIBLE_API_KEY=your_key
MAIN_MODEL=gpt-4o
MEMORY_MODEL=gpt-4o-mini

Xiaomi

PROVIDER=xiaomi
OPENAI_COMPATIBLE_API_KEY=your_key
MAIN_MODEL=your-model-name
MEMORY_MODEL=your-model-name

Custom Provider (any OpenAI-compatible API)

PROVIDER=custom
OPENAI_COMPATIBLE_API_KEY=your_key
OPENAI_COMPATIBLE_BASE_URL=https://your-api-endpoint/v1
MAIN_MODEL=your-model-name
MEMORY_MODEL=your-model-name

Tool Safety Levels

Tool	Level	Behavior
list_files	AUTO_APPROVE	Automatically allowed
read_file	AUTO_APPROVE	Automatically allowed
get_file_info	AUTO_APPROVE	Automatically allowed
write_file	ALWAYS_CONFIRM	Always requires human approval
delete_file	ALWAYS_CONFIRM	Always requires human approval
run_python	KEYWORD_CHECK	Blocked if dangerous keywords detected

Architecture

Layer	File	Responsibility
Config	config.py	Reads .env, builds LLM instance by provider
Prompts	prompts.py	System prompt + long-term memory injection
Tools	tools.py	6 tools, sandboxed to /tmp/sandbox/
Guard	guard.py	3-tier classification, human confirmation
Lifecycle	harness.py	MAX_STEPS, routing control
Memory	memory.py	Cross-session persistence, memory.json

License

This project is licensed under the MIT License.
When sharing or republishing this project or its content, please credit the source: jingw2/microharness.

Star History