feat(adapter/local-llm): Ollama / llama.cpp / vLLM skill runtime for sovereign and air-gapped pilots

[potiuk]

Goal: Define and land a skill runtime that works against local LLMs (Ollama, llama.cpp, vLLM, LM Studio, Jan). Grounds: RFC-AI-0004 Principle 3 — Vendor neutrality, which explicitly names "a local-Ollama wrapper" as a target runtime, and RFC-AI-0004's "no cloud-only skills" consequence.

Why local LLM:

• Sovereign deployments — projects bound by data-residency rules (EU, government, defence-orbit OSS) that can't send issue / mail content to a US cloud API
• Air-gapped triage — security teams that need to assess <security-list> traffic without an internet connection
• Cost ceiling — projects unable to commit to a per-token cloud bill, but happy to run a 70B model on shared infra

What "parity" means (calibrated for what local LLMs can realistically do):

• Skills under .claude/skills/<name>/SKILL.md are invokable via a local-LLM-backed agent loop
• The tools/* bridges are reachable — these are language-agnostic CLI calls, no LLM context needed
• Sandbox / HITL primitives map to the same bubblewrap baseline; the local-LLM runtime is the new piece
• Realistic scoping: some skills (multi-step reasoning over long contexts — security-issue-triage, pr-management-code-review) may need a high-capability local model (Llama 70B, Qwen 72B, DeepSeek). Document the per-skill model-size floor so adopters can pick a model that actually works.

Suggested approach:

• Pick an agent-loop frontend that supports local LLMs natively — candidates: aider --model ollama/..., goose with a local backend, continue.dev, a fresh thin wrapper around llama.cpp's OpenAI-compatible API
• Document the model-size floor empirically (per skill, against the existing eval suite at tools/skill-evals/)
• Land a setup-local-llm skill family alongside setup-isolated-* for the runtime-side install

Reference:

• RFC-AI-0004: docs/rfcs/RFC-AI-0004.md
• Existing skill shape: .claude/skills/
• Eval suite (for model-floor calibration): tools/skill-evals/
• Ollama: https://ollama.com
• llama.cpp: https://github.com/ggml-org/llama.cpp (OpenAI-compatible server)
• vLLM: https://github.com/vllm-project/vllm