dsh-more-models-thinking-level

DSH's model picker only surfaces a reasoning-level control when a model carries an explicit reasoningEfforts declaration in the bundled catalog. Models served through custom gateways or relays such as pptoken, one-api, new-api, or LiteLLM typically use vendor-specific, date-stamped, or aliased IDs that are absent from that catalog, so DSH defaults them to no reasoning capability and hides the picker entirely. This plugin closes that gap without touching Harness core files: a one-time helper script writes the missing reasoningEfforts map into ~/.dsh/settings.yaml, and the host (dsh-llm-pi-ai) is then free to surface the level selector and handle per-request wire-value conversion.

The helper creates a timestamped backup, then assigns declarations by model family: image models are marked non-reasoning, Gemini models receive an off-through-high five-level map, and every other non-image ID (codex, claude, qwen, deepseek, moonshot, or arbitrary gateway aliases) gets the full off-through-xhigh six-level map. Existing declarations are left untouched, and the plugin runtime itself only logs an enable message — all real capability comes from the persistent settings edit rather than from any runtime code path.

This plugin is most useful when you route requests through a custom gateway or relay and those model IDs fall outside DSH's bundled catalog, yet you still want to pick a reasoning intensity in the UI just as you would with an official provider. If you exclusively use standard catalog IDs through DSH's built-in provider route, the levels already exist and this plugin is unnecessary. Keep in mind that the helper currently requires Windows PowerShell and defaults to the desktop profile, and whether the gateway actually honors the reasoning parameter is ultimately the gateway's own responsibility.