Página Principal Explorar Ajuda
Registe-se Iniciar Sessão
haiany
/
deepseek-harness
mirror de https://github.com/deepseek-ai/deepseek-harness.git
1
0
Fork 0
Ficheiros Problemas 0 Wiki
Árvore: 74ef3f0510
Ramos Etiquetas
deepseek-harnes...
/
packages
/
client
/
ui-model-selection
/
README.md

README.md 5.9 KB
Histórico Em bruto

description: "Model selection for the Web GUI: the /model popup and the composer model seat over one per-session provider-grouped directory; for users and maintainers of model routing."

kind: "package-reference"

@deepseek-ai/dsh-client-ui-model-selection

English | 中文

Summary

This package provides model selection in the Web GUI: the /model popup command and the composer's model seat, both over one per-session directory of provider-grouped models. Choosing a model submits the complete selection — provider, model, and reasoning effort — which the Host snapshots at the next prompt-assembly boundary, so the following request uses it while a running step keeps its assembled selection. The composer seat shows a two-level Model/Effort menu: models stay provider-grouped, and the selected exact model supplies its adapter-owned effort names and default. When the Host reports that no adapter serves the session's route, the composer input goes inert until a route becomes available.

Table of Contents

Use this package

Mount this plugin alongside ui-conversation and the commands package; the composer then shows the model seat next to the pending indicator, and /model opens the same directory as a popup. Both surfaces show the host-reported current selection when the exact provider/model pair remains in the advertised groups; a missing catalog row leaves the routable selection intact while the trigger prompts Select model.

Model and effort

Models stay grouped by provider. The menu shows model and effort names only; catalog descriptions remain available to other consumers. The /model popup applies the selected model's default effort; the composer can then choose any advertised effort. An adapter without reasoning metadata leaves the Effort row absent; there is no arbitrary effort input.

Unroutable sessions

When the Host reports that no adapter serves the session's route, this plugin raises a composer block and the input goes inert with its own copy; recovering clears it without a reload. A null before the first load or after one failed never blocks, and catalog membership never blocks either — a route serving a model it does not advertise is missing from the groups yet usable.

Understand the implementation

Implementation internals — click to expand Two entries over ONE per-session directory owned by `ModelDirectoryResolver` (`ctx.modelDirectories`): the `/model` popupSelect contribution (registered through `ctx.commandUi`) and the composer's named `conversation.input.model` seat both load the session's advisory directory through `session.models` and submit through `session.selectModel` via the same `ModelDirectory` instance, so a switch made in either entry is what the other shows next. Directory loads and selections share a generation counter so an older response never overwrites a newer one; a connection reset drops every resident projection and repulls the Host-restored selection. Directories are per-session, resolved lazily, and disposed with the session scope; addressed subagent sessions expose neither entry. Every resident directory refetches directly on forwarded `llm/adapters-updated` and `settings/document-updated` owner events.

Further Exploration

Read these pages when the model surface is not enough. They move from the browser surfaces to the command popup shell and the selection contract.

Model Experience

Indirectly, through the session.selectModel selection both entries submit: the Host snapshots the complete ModelSelection at the next prompt-assembly boundary and owns the model-visible effect, while a running step keeps its assembled selection.

KV Cache effect

Switching the route can reduce or invalidate provider-side cache reuse for subsequent requests; the prompt prefix itself is untouched.

Known Limitations and Deferred Work

These limits define the current model surface. They are current package constraints, not a general model-router comparison or a task backlog.

  • No create-time or addressed-subagent selection — both entries require an existing ordinary session's Agent; there is no draft-phase model choice to fold into session creation, and subagent continuation deliberately exposes no independent model-selection contract.
  • Directory names are presentation-only — selection and persistence use provider/model/effort ids; a provider whose catalog or exact-model metadata lookup fails lists as an unselectable failure row until reload.
  • No arbitrary effort input — the composer offers only the exact model's adapter-advertised levels; an adapter without reasoning metadata leaves the Effort row absent.

Dev Note

Working context for maintainers — click to expand None.

Runtime invariant: No companion is published. A single command contribution registration whose disposal is proven by the HMR-safety spec — it emits no cordis events and owns no cross-plugin mutable state.