sleepy
e1781947f4
Reverted incorrect changes - Qwen3.5-35B-A3B IS a real model: - 35B total / 3B active parameters (MoE) - 262k native context (up to 1M extended) - Apache 2.0 license - Available on HuggingFace: Qwen/Qwen3.5-35B-A3B Updated files: - opencode/opencode/feedback/localllm/local-llm-feedback.md - opencode/opencode/feedback/SUMMARY.md - FEEDBACK_TEMPLATE.md Added correct specs: - MMLU-Pro: 85.3% - SWE-bench Verified: 69.2% - Context: 262k native, 1M extended
Coding Harness Feedback Analysis
Research on four coding agent harnesses to understand what works best for different model sizes, particularly smaller/local models.
Folder Structure
├── AGENTS.md # Project overview and data collection strategy
├── Research*.md # Prompt research and orchestration strategies
│
├── opencode/ # Go-based coding agent
│ ├── feedback/
│ │ ├── frontier/ # GPT-5.4, Claude Opus, Gemini feedback
│ │ └── localllm/ # Local model feedback (prompting, tool handling)
│ └── repo/ # Source code (submodule)
│
├── pi/ # Minimal terminal coding harness by Mario Zechner
│ ├── feedback/
│ │ ├── frontier/ # (empty - in progress)
│ │ └── localllm/ # (empty - in progress)
│ └── repo/ # Source code (submodule)
│
├── hermes/ # Nous Research's agent
│ ├── feedback/
│ │ ├── frontier/ # Claude, GPT, budget provider feedback
│ │ ├── localllm/ # Qwen, Gemma, local model feedback
│ │ └── general/ # Bug reports, benchmarks, features
│ └── repo/ # Source code (submodule)
│
└── forgecode/ # AI pair programmer with sub-agents
├── feedback/
│ ├── frontier/ # GPT-5.4, Claude, Gemini, pricing, security
│ └── localllm/ # Qwen, MiniMax, GLM, DeepSeek feedback
└── repo/ # Source code (submodule)
Quick Navigation
| Harness | Feedback Location | Key Topics |
|---|---|---|
| opencode | opencode/feedback/ |
Tool calling, local model prompting |
| pi | pi/feedback/ |
(Being researched) |
| hermes | hermes/feedback/ |
Terminal-bench results, local setup |
| forgecode | forgecode/feedback/ |
Pricing, benchmarks, security |
Feedback Format
Each feedback file includes:
- Model name/size/provider
- Task performance or benchmark results
- Issues encountered
- What worked well
- Source reference (URL, Discord, GitHub issues)
Research Focus
- Tool handling and capabilities
- Skills system effectiveness
- Prompt engineering strategies
- Context management
- Error recovery