Record: LLM-integration handoff — 2026-08-26

Written so a fresh session can resume without this conversation’s memory.

Current state (verified 2026-08-26)

Project 00-llm-integration created from 00-project-skeleton; canonical prompts copied verbatim; skeleton site-build example removed.
~/models/ reference material imported to sources/models/ (README.md, config-beast.ini, config-athena.ini, install.sh, exercise-mi25-models.sh, backup.sh). These ini copies are older than the live /usr/local/etc/config.ini on each host — treat live captures as authoritative.
Live router captures written to dataflow.in/endpoints/beast-models.json (28 models) and athena-models.json (32 models), verified 2026-08-26.
documents/07-endpoints-map.md, 08-model-catalog.md, 09-tool-integration.md, 10-routing-policy.md written from the captures.
sources/config/dsh-settings-beast.example.yaml written: DSH llm-pi-ai provider with openai-completions, compat {supportsDeveloperRole: false, maxTokensField: max_tokens}, and 7 curated model entries with capacities matching the live presets.
make test passes; make capture refreshes the endpoint facts.
DeepSeek Harness runs on this desktop (athena) at http://127.0.0.1:3080; default provider remains deepseek-official (deepseek-v4-flash, reasoningEffort high).

The llm-pi-ai.beast section is schema-valid and the compat block is correct.
Problem 1: bare id-only model entries inherit 262144 context / 32768 max-token defaults, which exceed every live preset and cause context overflow. Every entry needs explicit contextWindow/maxTokens.
Problem 2: short ids (coder32, r1-14b, gpt-oss-120b, …) do not exist on the router yet — no alias = lines in the beast preset. Requests with those ids fail until aliases are added or full router ids are used.
Problem 3: several listed ids are cache entries that exist, but a few (for example unsloth/gemma-4-12B-it-qat-GGUF:Q4_K_XL) are CPU cache entries, not the tuned GPU presets; capacities must come from the capture.
Problem 4: BEAST_API_KEY credential is not set anywhere (env, ~/.dsh/.credentials.yaml, or .env); requests fail with MISSING_CREDENTIAL until it is stored (any non-empty value works; llama.cpp ignores auth until --api-key is set on the router).
The human applies the fix; this project does not modify ~/.dsh/settings.yaml.

prompts/tasks/01-capture-endpoint-facts.md — refresh captures and re-derive the map/catalog documents.
prompts/tasks/02-fix-dsh-provider-catalog.md — reconcile the DSH provider with the live router, write the findings record, verify one live request.
prompts/tasks/03-document-tool-integration.md — finish documents/09-tool-integration.md (aider, llama.vscode, VS Code Chat).
prompts/tasks/04-plan-minerva-offline.md — SSH to minerva.lan, discover hardware, write documents/11-minerva-travel-plan.md.

Keep deepseek-official as the DSH default; use beast per session.
CPU-heavy background tasks go to unsloth/gpt-oss-120b-GGUF:Q4_K_XL (65536 ctx, ngl 0) so the MI25 stays free for GPU models.
Do not put R1-distill models in agent loops; reserve them for reasoning Q&A.