dsh-more-models-thinking-level

LIMITATIONS

已知限制

Please read these before installing — some materially affect whether you get the result you expect. - **Not listed in the plugin marketplace.** The marketplace registry is generated from `awesome-dsh-plugin`; because the corresponding PR has not been merged/reviewed yet, this plugin is *not discoverable via `dsh plugin search`*. Install it by name (npm) or by repository URL, and confirm it is registered in your profile before use. - **The helper is not invoked automatically.** Installing/enabling the plugin does not run `enable-capabilities.ps1`. You must run it manually once per profile, then restart DSH. Until you do, the plugin only adds the settings section and logs — your models will not gain reasoning levels. - **The host entry is mostly a no-op.** `lib/index.js`'s `apply` only logs `reasoning-adapter: enabled`. All real capability comes from the declaration the helper writes into `settings.yaml`, not from plugin runtime code. Treat the "plugin" as an installer + documenter around a settings edit, not as a runtime adapter. - **It persists edits to `settings.yaml`.** The helper rewrites your settings file in place (after a backup), so it is *not* a stateless runtime override. If you later change providers, switch protocols, or remove the plugin, the injected `reasoningEfforts` and the default `reasoning: medium` remain until you clean them up. Ownership of these fields is a manual concern. - **Model capability is guessed from the ID.** The helper identifies models with a regex (exclude image models and Gemini, treat everything else as GPT-style). It cannot know whether a given model *actually* supports a level. If a relay exposes non-reasoning models (e.g. embeddings, vision-only, or plain chat models) under arbitrary IDs, they will receive a reasoning declaration too, and a strict gateway may reject the extra reasoning parameter. - **Support ultimately depends on the gateway.** A declaration only makes the level *selectable*. Whether the relay actually accepts the `reasoning` / `reasoning_effort` parameter — and the specific wire value — is decided by the gateway. Some strict gateways validate requests against a fixed schema and reject fields they do not recognize. The plugin cannot guarantee a chosen level actually takes effect on a given relay. - **Windows PowerShell + `desktop` profile only.** The helpers use Windows PowerShell syntax and path conventions, and default to the `desktop` profile. Non-Windows hosts or other profile names require manual adaptation. - **The settings section is informational, not an editor.** The client section explains the mapping and points at the model settings page; it has no provider/model visual editor. Fine-grained per-provider or per-model control is done by editing `settings.yaml` yourself. - **Unfinished marketplace-readiness work.** Configurable profile names, migrating capability config to the DSH Settings API, automated tests for GPT/Gemini/generic models, explicit error text when a gateway refuses a parameter, and removing the dependency on an early core-runtime patch are all still open.