dsh-rpm

Per-provider requests-per-minute (RPM) rate limiting for DeepSeek Harness.

Model requests to a provider beyond its configured limit are delayed until the next one-minute window instead of failing — covering interactive chat, subagents, and retries alike, since enforcement sits in the llm/stream waterfall that wraps every streaming model call.

Features

• Durable per-provider limits in the dsh-rpm settings namespace (providers[routeId] = rpm); changes apply live, no restart.
• A Rate limit (requests per minute) input on provider cards in Settings → Models:
  • DeepSeek official card: rendered directly by this plugin.
  • Custom providers on the llm-pi-ai card cell (9router, cloudflare-…): rendered by dsh-short-tool-ids beneath its Use short tool-call IDs toggle, when that plugin is installed alongside. The provider-card cell is single-occupant per settings namespace, so exactly one plugin renders each card's extension area; the composite lives there to keep both controls visible.
• 0 or an empty field means unlimited (the default for every provider).

Installation

dsh plugin --profile web add "/absolute/path/to/dsh-rpm"
# restart the web profile afterwards
dsh web --no-open --port 3080

The client bundle is generated: edit client/index.mjs, then npm run build.

Development

npm run build   # generate lib/client.js from client/index.mjs
npm test        # limiter unit tests
npm run pack:check

How it works

• Host (index.mjs) registers the dsh-rpm settings namespace and listens to the host llm/stream event. Routes without a configured limit pass through with zero overhead; limited routes queue on a fixed-window counter (limiter.mjs, independently unit-tested).
• Client (client/index.mjs) binds the settings transport's settingsScope service and renders the per-card control into settings.models.provider-card.