Feature: Minerva Travel Laptop

Purpose

Defines the offline inference plan for the laptop minerva.lan (minerva.lan, WiFi, SSH available per the homelab inventory).

When travelling, the laptop has no LAN access to beast or athena, so it must run a small model locally and keep aider and llama.vscode working offline.

Requirements

• MINERVA-R001 — Before any deployment, discover the GPU (lspci, vulkaninfo --summary, nvidia-smi), the RAM (free -h), and the disk (df -h ~); record the findings in records/ and update documents/07-endpoints-map.md.
• MINERVA-R002 — Keep the travel model at or below 4B parameters, for battery life and thermals.
• MINERVA-R003 — Install llama.cpp on minerva with the backend the discovered GPU needs (Vulkan for AMD and Intel iGPUs, CUDA for NVIDIA), following the pattern in sources/models/install.sh adapted to a single preset or a small router preset on 127.0.0.1:2001.
• MINERVA-R004 — Point aider (--openai-api-base http://127.0.0.1:2001/v1) and llama.vscode at the local server.
• MINERVA-R005 — Predownload the chosen GGUF into the HuggingFace cache before travelling.
• MINERVA-R006 — The travel plan must state the model, the exact GGUF, its size on disk, and the expected context, each verified on minerva before the plan is marked done.
• MINERVA-R007 — Use the homelab Ansible inventory (minerva.lan, SSH) for provisioning once the specifications are known.

Behavior

• With no LAN access, aider and llama.vscode reach a local llama.cpp server on 127.0.0.1:2001 and complete an edit and a chat turn.
• The plan is marked done only after the stated model, GGUF, size on disk, and context have been observed on minerva.

Dependencies

• prompts/features/03-dsh-provider-integration.md

Hardware (verified 2026-08-26)

• CPU: AMD Ryzen 7 7840U (16 threads).
• GPU: AMD Radeon 780M (RDNA3 iGPU, shared memory) — Vulkan via Mesa RADV expected; confirm with vulkaninfo --summary.
• RAM: 14 GB total (~11 GB available); iGPU VRAM comes out of this pool.
• Disk: 935 GB, ~691 GB free.
• OS: Ubuntu 24.04.4 LTS. llama.cpp not yet installed.
• SSH: ssh -i ~/.ssh/keys/key-athena owner@minerva.lan works (key-athena is the access key).

Model Candidates (from the existing catalog)

• Qwen/Qwen2.5-Coder-3B-Instruct-GGUF:Q8_0 — recommended default: coding + editing, FIM-capable; fits the shared-memory iGPU comfortably.
• unsloth/gemma-4-E2B-it-qat-GGUF:Q4_K_XL — 2B QAT gemma, small and capable.
• unsloth/Qwen3.5-4B-GGUF:Q4_K_M — general chat alternative.
• Qwen/Qwen2.5-Coder-1.5B-Instruct-GGUF:Q8_0 — minimal fallback; fastest.
• unsloth/gemma-4-E4B-it-qat-GGUF:Q4_K_XL — 4B, upper bound for the laptop.

Deployment Shape

• The same llm-pi-ai provider shape from prompts/features/03-dsh-provider-integration.md applies if a DSH session ever runs on the laptop.
• The local server mirrors the router preset shape already used on beast and athena, so the client configuration differs only in baseURL.