agilab-local-llm
Guidance for using local LLM backends (Ollama/GPT-OSS) inside AGILAB with correctness-first prompts.
When & Why to Use This Skill
The agilab-local-llm skill optimizes the integration of local LLM backends like Ollama and GPT-OSS within the AGILAB environment. It prioritizes reliability through 'correctness-first' prompting strategies, ensuring deterministic outputs for code edits while preventing silent API failures. By providing granular control over model parameters and explicit validation requirements, it enables secure, offline, and high-precision agentic workflows.
Use Cases
- Privacy-Centric Development: Securely process sensitive codebases locally using Ollama or GPT-OSS without exposing data to external cloud APIs.
- Deterministic Code Refactoring: Utilize correctness-first prompts to ensure local models provide exact patch intents and validation tests, reducing hallucinations during automated code edits.
- Robust Agentic Workflows: Implement strict guardrails and error handling to detect missing local endpoints up-front, preventing silent fallbacks and ensuring workflow stability.
- Custom Model Tuning: Fine-tune local model performance within AGILAB by adjusting temperature, top_p, and iteration limits for specific programming or automation tasks.
| name | agilab-local-llm |
|---|---|
| description | Guidance for using local LLM backends (Ollama/GPT-OSS) inside AGILAB with correctness-first prompts. |
| license | BSD-3-Clause (see repo LICENSE) |
| updated | 2026-01-09 |
Local LLM Skill (AGILAB)
Use this skill when working on local/offline engines or prompts.
Correctness-First Defaults
- Prefer deterministic settings for code edits (lower temperature, explicit constraints).
- Require the model to return:
- file list to edit
- exact patch intent
- tests/commands to validate
No Silent Fallbacks
- Detect missing local endpoints/models up-front and surface an actionable error.
- Do not auto-switch APIs or rewrite parameters silently.
Ollama Notes
- Let users select:
- model name
- temperature/top_p/top_k
- max tokens
- seed (if supported)
- “auto-run/auto-fix” loop guardrails (max iterations, stop on failure)