AI Agent Hub
Back to plugins
🤖

dsh-llm-rate-limit

Model Inference Updated 2026.08.20

Run the following command in DeepSeek Harness:

dsh plugin install Asong6824/dsh-llm-rate-limit

Paste the following prompt into your AI chat to install this plugin:

Run dsh plugin install Asong6824/dsh-llm-rate-limit in your DeepSeek Harness terminal to install the rate-limiting plugin from https://github.com/Asong6824/dsh-llm-rate-limit.

About this plugin

Running multiple parallel agents, subagents, or background retry loops can flood a provider with LLM requests almost instantly, triggering HTTP 429 errors, throttling, or brief connection drops. dsh-llm-rate-limit inserts a pacing layer before any request reaches the provider, turning wall-hitting into orderly queuing.

Per-provider, it offers configurable RPM token buckets with burst capacity, an optional estimated-tokens-per-minute budget that reconciles against actual usage, concurrency slots with bounded FIFO queues (timeout and cancellation supported), and an adaptive cooldown driven by error codes, HTTP statuses, and Retry-After headers. Auxiliary background traffic is explicitly shed so it never blocks primary work; every dsh-llm-retry attempt is admitted independently, and the plugin itself never performs retries.

If you are driving DeepSeek, Volcengine Ark, or other DSH providers through multiple concurrent agents and want rate-limit control solved locally rather than relying on remote retries, this plugin is the gatekeeper that keeps things in line.

Use Cases

  • Multiple parallel agents or subagents fire LLM requests simultaneously, repeatedly hitting HTTP 429 limits
  • Serving multiple providers such as DeepSeek and Volcengine Ark with independent, configurable rate policies
  • Background retries or auxiliary traffic must not block primary foreground requests

Best For

  • Developers orchestrating multiple concurrent LLM agents through DSH
  • Platform or ops teams that need per-provider local rate policies across several LLM vendors
  • Engineering teams preferring local request pacing over reliance on remote retry logic