@apkawa/dsh-model-settings

Per-model reasoning effort and context window management for llama.cpp models in dsh's settings.

Screenshot

Features

• Model settings section — a top-level settings page with a per-model accordion:
  • the seven thinking levels (off, minimal, low, medium, high, xhigh, max) as checkboxes with wire values;
  • a context-window field with a Refresh button that force-writes the size from the provider's live /v1/models listing (rounded down to the thousand);
  • one Save commits both parts together.
• Clearing: unchecking every stored level activates Save and, on save, clears the model's reasoning mapping (dsh's canonical reasoningEfforts: false); emptying the context-window field of a sized model drops the key so the model inherits again.
• Context-window autofill (host side): fills every unsized model's contextWindow from the provider's INI preset (ctx-size, else fit-ctx), rounded down to the thousand.

How it works

• Watcher on settings/updated (host side, filtered to the llm-pi-ai section): for each candidate route one GET /v1/models, parse INI presets, fill every model that has no nonzero contextWindow yet.
• Self-cycle guard: a route whose models are all already sized is skipped entirely, so our own write never schedules another round.
• The section's rows and saves go over the plugin-owned /model-settings RPC channel (host handler on the same settings seam); failures surface as human-readable lines in the UI.

Install

dsh plugin --profile web add github:apkawa/dsh-plugins#path:/packages/dsh-model-settings

Debugging

Logs are quiet by default. Run dsh with DEBUG=1 to see the plugin's trace output (fetchModels, autofill rounds); error lines are always printed regardless of the flag.

Development

See AGENTS.md for principles + pitfalls, and docs/ai/key-moments.md for architectural reference.