Record: DSH settings and cache-reuse applied — 2026-08-26

Outcome

Completed the two pending applies from the previous session, both verified.

1. DSH provider sections applied (task 03 completion)

~/.dsh/settings.yaml llm-pi-ai section replaced with the corrected routes:

• beast — 11 models by router alias (coder32, coder14, devstral, qwen9, gpt-oss-20b, gemma-12b, r1-14b, gemma-26b, gpt-oss-120b, r1-70b, llama70), contextWindow/maxTokens matching the live presets, compat {supportsDeveloperRole: false, maxTokensField: max_tokens}.
• athena — 6 models (coder15, coder3, llama3, gemma-e4, coder32-fallback, r1-32b-fallback) with the same compat block.
• agent-default-model left unchanged (deepseek-official / deepseek-v4-flash / reasoningEffort high).

Credentials in ~/.dsh/.credentials.yaml: BEAST_API_KEY (already present, non-empty) and ATHENA_API_KEY (generated random value). Both refs verified non-empty. The DSH adapter resolves credentials per stream call, so no restart is needed; if the model picker does not show the new providers, reload the GUI page.

2. cache-reuse applied

Added cache-reuse = 256 under [*] in /usr/local/etc/config.ini on both hosts (baselines backed up as config.ini.bak-20260826-cachereuse), restarted both llama.services.

Verification (all 2026-08-26)

• Beast: 13 models, child args show --cache-reuse 256, --parallel 4; athena: 15 models, same args.
• Aliases intact on both routers; captures refreshed (dataflow.in/endpoints/).
• ~/.dsh/settings.yaml parses and mirrors sources/config/dsh-settings-beast.example.yaml (validated by make test).
• make test passes.

Commits

• (pending) records + doc updates for this apply.