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

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

English | 中文

Subagent model pinning UI, browser half: decorates the host /subagent-model command with a popupSelect. Options load from the session's model directory (session.models), so the box always reflects the providers configured in settings; each row shows the display name on the label line and the provider route plus model id on the detail line, satisfying the three-field rule (provider / model id / display name). The row matching the session's current pin — read from the model-roles session projection — carries the active check. Picking a row executes the host command /subagent-model /, which logs the pin event and renders the command node in the conversation; the argued manual invocation keeps working unchanged because the popup only decorates the bare command.

Addressed subagent sessions (the @-referenced children) expose no popup: their pins are decided by the delegating model, not the user.

The /client exports are the plugin body (apply/inject) plus the pure optionsOf row builder.

Settings page

The browser half also registers a Subagent Models settings section (settings.section, id model-roles, order 20) once the settings shell, locale, and remote services are composed. It lists the whole model directory — the same llm.models source the Models settings page renders — with one row per model. Each row edits two fields:

• Capability description: free text describing what the model is good at. It is required — an empty description cannot be saved. This is the text the routing system-prompt section shows the delegating model.
• Allow automatic subagent assignment: the subagent switch. A model with the switch off is excluded from the auto-assignment candidates the routing guidance lists.

Rows stay in sync with the Models page because they enumerate the same model directory: a provider or model added there appears here on the next load. Role values live in the model-roles settings namespace, and every write goes through settings.mutate with the expected revision, so a concurrent edit is refused rather than silently overwritten. The host plugin watches its namespace and hot-applies it — a saved row is effective immediately, no restart. Pushed invalidations (settings/document-updated for the model-roles namespace, llm/adapters-updated, connection/reset) refetch the page after it has loaded at least once, so an open panel converges without polling.

The section activates in an optional child fiber over ['slots', 'locale', 'connection', 'remote']: a composition without the settings shell still gets the popup decoration and nothing else. A read-only settings provider disables every write control and says why.

Model Experience

No direct prompt content. Picking a row runs the /subagent-model command, whose handler appends the model-roles/subagent-pin log event; the routing system-prompt section then instructs the delegating model to pass the pinned agentOptions on each subagent tool call.

KV Cache effect

The command lifecycle records (command/run/command/done) are ordinary log events; the pin never enters a model request body by itself. Selecting an already-pinned row is a no-op that appends nothing.

Known Limitations and Deferred Work

• Soft execution only: the pin is advisory — the model receives guidance, never an enforced route. The settings page edits each model's subagent switch (which the auto-assignment guidance honors), but the popup still lists every configured model regardless of that switch; filtering subagent: false models out of the picker is deferred until a dedicated pin RPC exists.
• Catalog absence renders an empty box: a session whose provider catalog failed to load offers no rows (the shared popup shell shows its empty state); the typed command remains the fallback.