raindrop

Use when the user wants to replay a prod trace locally against their real agent code (not a tool-mock) — typically triggered from Workshop's "Local Agent" replay button or directly via /raindrop:add-readonly-replay. Wires a `RAINDROP_READONLY=1` switch into the user's agent, classifies every side-effecting tool (fail / stub / real), scaffolds an HTTP `/replay` endpoint that flips the switch before invoking the agent, and registers the agent in `~/.raindrop/agents.json` so Workshop knows where to send replay requests. Run once per agent — re-run when the tool surface changes.

The plugin's primary skill. Use when the user wants to read, reason about, or annotate an AI agent trace in Raindrop Workshop. Also dispatches inbound messages from the Workshop UI's chat pane when channels are enabled. Requires the daemon to be running — if `get_active_run` errors with "backend unreachable," tell the user to run `/raindrop:setup` first.

Run this first after installing the raindrop plugin. Installs the raindrop binary if missing, runs `raindrop workshop init` (which writes RAINDROP_LOCAL_DEBUGGER into the project's ./.env, starts the daemon at http://localhost:5899, and opens the UI), inventories the project's agents and reports which are instrumented with a `@raindrop-ai/*` SDK, then tells the user what to do next. Idempotent — safe to re-run any time the daemon might be down or to re-bootstrap a project.