claude-in-dsh

dsh web is normally driven by its own model engine, but if you already have Claude Code CLI logged in locally, some tasks feel more natural on Claude. The friction is obvious: you do not want a second terminal window or a duplicated chat UI. claude-in-dsh hands any dsh web session over to local Claude Code while every output line is rendered through dsh's native transcript, tool cards, approval panels, and command palette—one interface, two backends.

Feature coverage spans the daily workflow end to end: an engine selector (DSH | Claude Code) in the input bar toggles per session with strict mutual exclusion that survives restarts; Claude permission requests bridge into dsh's approval UI, AskUserQuestion flows through dsh's question card, and ExitPlanMode uses dsh's plan-review card; slash commands merge into dsh's palette and execute out-of-band so they never pollute the transcript; pasted images become native dsh attachments before being sent to Claude as image blocks; the Claude process is held by an independent broker via setsid, so plugin hot-reloads or dsh restarts never interrupt an in-flight turn, and any output produced during a dsh restart is replayed into the transcript on next open without invoking the model; switching models mid-run uses a control request instead of a process restart.

Built for developers who already run both dsh web and local Claude Code CLI and want to swap backends per task inside a single interface, while keeping dsh's persistence, theming, plugin ecosystem, and the session history under ~/.claude fully intact.