@xidong-ai/dsh-rate-limiter
DeepSeek Harness 主动限速插件：按供应商令牌桶限速，超限延迟排队而非失败
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add @xidong-ai/dsh-rate-limiter@0.3.2Rate Limiter 以 @xidong-ai/dsh-rate-limiter 发布，当前版本为 0.3.2。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
A proactive rate limiter plugin for DeepSeek Harness (dsh): it controls the request rate per provider (token bucket) before model requests are issued, and queues the request with a delay instead of failing when the limit is exceeded — avoiding upstream 429s.
It complements the official dsh-llm-retry (exponential backoff after failure): rate limiting comes first (prevention), backoff comes last (safety net); the two do not interfere with each other.
agent/request, coexists naturally with dsh-llm-retryInstall from npm:
dsh plugin --profile web add @xidong-ai/dsh-rate-limiter
npm registry URLs are case-sensitive; use the lowercase package name.
Or install directly from GitHub:
dsh plugin --profile web add github:Xidong-AI/dsh-rate-limiter
For local development, add the checkout directly:
dsh plugin --profile web add .
After installing, dsh --profile web --dump-config should show the plugin entry:
- id: rate-limiter
name: @xidong-ai/dsh-rate-limiter
config:
enabled: true
providers: {}
Configure the token bucket per provider in the profile's cordis.patch.yml (or this plugin's cordis.patch.yml):
- id: rate-limiter
config:
enabled: true
providers:
nvidia:
rate: 0.5 # tokens/second (long-term average QPS)
burst: 1 # bucket capacity (allowed burst requests)
sensenova:
rate: 0.02778
burst: 1
rate: refill rate (tokens/second), i.e. the long-term average request rate.burst: bucket capacity, the number of burst requests allowed.enabled: false disables the plugin entirely.The plugin hooks onto the agent/request waterfall: it await next() first to obtain the call config (which carries the provider), then performs a per-provider token bucket check; when tokens are insufficient, it queues the request with a delay (interrupted immediately by the abort signal when the user stops), then returns the config unchanged — it never modifies request content, never changes routing, never swallows errors. It only controls when a request is issued.
The rate-limiting algorithm is a hand-written reservation-based token bucket (concurrency-safe), with zero third-party rate-limiting dependencies.
| Plugin | Timing | Behavior |
|---|---|---|
dsh-rate-limiter | Before the request is issued | Queue with a delay when over the limit (prevents 429s) |
dsh-llm-retry | After the request fails | Exponential backoff retry (safety net) |
They mount at different points (agent/request vs agent/request-error) and coexist naturally.
dsh plugin --profile web remove @xidong-ai/dsh-rate-limiter
npm install
npm run typecheck # tsc --noEmit
npm run test # vitest run
npm run build # esbuild transpiles lib/*.ts → lib/*.js
Thanks to the Linux.do community for support.