Record: Findings and task reorder — 2026-08-26

Outcome

Sorted the project’s next steps per the owner’s decision: llama service config first, then DSH settings. Documented minerva hardware, the model retention analysis, and the MTP/RAG position.

Decisions

• Ordering: task 01 tunes the llama.cpp service configs on beast and athena; task 03 (DSH provider) runs after aliases are settled there.
• Copilot Chat is suppressed; llama.vscode (0.0.63 installed) replaces it for in-editor chat and agent.
• Discard candidates are noted in documents/12-model-retention.md; nothing is deleted without the owner’s decision.
• MTP is adopted for gpt-oss-120b after an explicit benchmark confirms it (its 8.42 t/s on CPU already shows the MTP signature).
• RAG stays knowledge-only: the Nomic embeddings preset is enabled in task 01; Open WebUI RAG comes later if wanted.

Findings

• Minerva (verified 2026-08-26): Ubuntu 24.04.4, Ryzen 7 7840U + Radeon 780M iGPU (shared memory), 14 GB RAM, 935 GB disk, no llama.cpp yet.
• Minerva SSH: ssh -i ~/.ssh/keys/key-athena owner@minerva.lan works; no Host minerva.lan block in ~/.ssh/config yet (open question in TODO).
• Beast runs llama.cpp build 10628 with --spec-type including draft-mtp; athena runs build 10129.
• Both routers list far more models than the presets (cache entries); retention analysis identifies the redundant/roleless ones.

Verification

• make test passes (structure, config validation, live captures).
• Live router captures refreshed 2026-08-26 into dataflow.in/endpoints/.

Commits

• (pending) task reorder + findings documents.