Free OpenClaw Repository Triage Using Local AI Models
Hugging Face
Google/DeepMind
DeepSeek
Anthropic
The Hugging Face blog describes how local open-weight models (Gemma 4 26B and Qwen 3.6 35B) were used to triage issues and pull requests in the OpenClaw repository in real time, replacing costly closed models. The system uses an agent harness with a restricted shell (reposhell) for safe read-only operations and achieves high throughput on an NVIDIA GB10 system.
The blog shows how the authors replaced closed AI models with local open-weight models for triaging issues and PRs in the OpenClaw repository, motivated by the risk of closed models being taken away (citing Anthropic's Claude Fable 5 removal). Using a harness called 'localpager-agent' with reposhell (a restricted read-only shell), the system classifies each item into categories like local_models, self_hosted_inference, etc. They tested Gemma 4 26B and Qwen 3.6 35B on an NVIDIA GB10 system (128 GB unified memory). Gemma achieved higher recall (0.905) and throughput (1.41 seconds per row, 16 concurrent workers), while Qwen had higher precision (0.831) and exact match (0.540), but took 13.51 seconds per row with only 4 workers. A semi-agentic architecture uses the LLM only for classification; notifications are rule-based. The system processes items via gitcrawl mirroring, stores results in SQLite, and notifies via Discord.
- Сокращения
- PR = Pull Request — Pull Request
- LLM = Large Language Model — Large Language Model
- GPU = Graphics Processing Unit — Graphics Processing Unit
- CLI = Command-Line Interface — Command-Line Interface
- JSON = JavaScript Object Notation — JavaScript Object Notation
- SQLite = SQLite (self-contained SQL database engine) — SQLite
- API = Application Programming Interface — Application Programming Interface
- MoE = Mixture of Experts — Mixture of Experts
- FP8 = Floating Point 8-bit — 8-битное число с плавающей точкой
- KV = Key-Value — Key-Value
Source: Hugging Face blog —
original