Hugging Face details local model workflow to triage OpenClaw repository issues and PRs

Hugging Face staff describe an agent-based workflow that uses local open-weight models to classify and triage issues and pull requests in the OpenClaw repository. The authors report using Gemma and Qwen models inside an agent harness to assign labels such as local_models, self_hosted_inference, acp, agent_runtime, codex, ui_tui, and inference_api.

The system is designed to replace periodic, quota-limited calls to closed models with continuous local inference on an NVIDIA GB10 workstation with 128 GB unified memory. The authors state both gemma-4-26b-a4b and qwen3.6-35b-a3b can generate hundreds of tokens per second locally with performance optimizations, enabling near-instant notifications for maintainers.

To constrain the agent’s actions, the workflow uses reposhell, a restricted bash-like shell that only permits read-only operations such as ls, find, cat, grep, and git commands. The agent receives the PR title, body, and a truncated diff excerpt, and can invoke reposhell to inspect the codebase before submitting a final classification via a structured output tool.

Orchestration is handled by localpager, which maintains a SQLite database of normalized PRs and issues via a local mirror (openclaw/gitcrawl). When a new item arrives, localpager creates a classification job; a worker claims the job, runs the local model, and produces a filtered Discord notification only for items relevant to the maintainer.

The authors emphasize the setup is “free” aside from electricity and existing hardware, contrasting it with usage-based quotas on closed-model APIs. They position the approach as increasingly viable as local model capabilities improve and as reliance on closed models introduces operational risk.