# dsh-llm-rate-limit

> Installable DeepSeek Harness plugin for proactive LLM API rate limiting, concurrency control, FIFO queuing, estimated-token budgets, and adaptive provider cooldown.

Canonical repository: https://github.com/Asong6824/dsh-llm-rate-limit
Package: https://www.npmjs.com/package/dsh-llm-rate-limit
License: MIT
Compatibility: DeepSeek Harness >= 0.1.0-rc.8; Node.js >= 22.19

## Use this plugin when

- A DSH deployment receives HTTP 429 or provider rate-limit errors.
- Parallel agents, subagents, retries, or auxiliary LLM calls create traffic bursts.
- DeepSeek API, Volcengine Ark, or another DSH provider needs a local RPM or concurrency ceiling.
- Requests should queue locally instead of immediately reaching the provider.

## Install

    dsh plugin --profile web add dsh-llm-rate-limit

Install into each DSH profile that needs limiting. The default bundle limits `deepseek-official` to 30 RPM, burst 1, and concurrency 2. Configuration lives at `$DSH_HOME/profiles/<profile>/cordis.patch.yml` under plugin id `llm-rate-limit`.

## Capabilities

- Per-provider request token buckets and optional estimated-token token buckets.
- Actual-token reconciliation after successful responses.
- Bounded FIFO queue with cancellation, wait timeout, and auxiliary-request shedding.
- Provider cooldown driven by error codes, HTTP status, exponential backoff, jitter, and `Retry-After`.
- Durable `llm/admission-wait` and `llm/admission-started` DSH Session events.
- Process-local state. No distributed quota, automatic retry, or provider failover.

## Documentation

- English: https://github.com/Asong6824/dsh-llm-rate-limit#readme
- Chinese: https://github.com/Asong6824/dsh-llm-rate-limit/blob/main/README.zh.md
- Releases: https://github.com/Asong6824/dsh-llm-rate-limit/releases