OSS-Onboarding Skill is an open-source Agent Skill that lets a small, local open-weight model (Gemma 4 via Ollama) onboard onto an unfamiliar repository and issue, then produce a grounded repo map, setup steps, ranked candidate files, a draft patch, and a PR description — the exact workflow a GSSoC/NSoC/SSoC contributor burns hours on. The skill is a portable SKILL.md plus five context-shrinking helper scripts (repo_map.py, issue_context.py, candidate_files.py, setup_detector.py, file_excerpt.py) written in Python 3.11 with uv, tree-sitter for symbol extraction, rank_bm25 for candidate ranking, and Pydantic-validated tool schemas; the harness wraps Ollama behind a three-tier model ladder (gemma4:e4b → gemma4:26b → hosted open weights) with a dispatcher that validates every tool call, detects narration-instead-of-invocation, and escalates rather than silently failing, while OpenTelemetry GenAI spans stream to a trace.jsonl that serves as the submission evidence. The 2-hour deliverable is the with-skill versus baseline comparison on one pinned repo and one real issue: baseline sees README + directory listing and gives generic advice, while the skill produces a trace showing 3 real files referenced, correct setup commands, a verified draft diff, and — ideally — one visible escalation event proving the harness works, all without a vector database, agent framework, or hosted API.