dsh-plugins

Long dsh sessions get unwieldy. You want to ask a follow-up about a specific passage without cluttering the main thread, clear model context without spawning a new session, or fine-tune provider timeouts, transport protocols, and per-model input types, yet none of these have a first-class UI. dsh-btw turns the casual aside into a complete workflow: select any text in the chat message area and a floating action bar appears, letting you either quote that selection (with source title and timestamp) into the current composer as a pre-filled block, or fork a one-off temporary session anchored exactly at the selected message. Walk away and it auto-archives itself, keeping the sidebar clean.

The New Context button (or /newcontext slash command) on the input toolbar performs an in-place logical split: every node on the session surface beyond the system prompt is replaced by a single notice, so each subsequent model request carries only the system prompt plus new messages, while the full conversation log remains visible and copyable in the UI. The split point is persisted in the session log, surviving restarts, restores, and forks. The Models settings page gains a Provider Advanced Config collapsible panel that visualizes every previously YAML-only field, including API protocol, connect and idle timeouts, SSE/WebSocket transport, cache retention, thinking budgets, per-model input type (text or image multi-select), three-state reasoningEfforts, and six compat toggles, all written through the official optimistic-lock RPC with server-side schema validation, while credentials stay in the API key form.

It is built for heavy dsh daily users, developers who frequently need local follow-up questions without opening new sessions, and teams running self-hosted LLM gateways that require precise per-model input modality declarations and reasoning-tier configuration. Pure client-side plugin, zero core modifications, bilingual UI, MIT licensed.