dsh-provider-rate-limit

DeepSeek Harness lets you route LLM traffic through multiple providers simultaneously, but in practice free-tier gateways often return HTTP 429 when quotas are exhausted or reject requests outright because they cannot identify a legitimate client. Without per-provider or per-model throttling and graceful backoff, a burst of parallel agents can instantly burn through an upstream quota and take the entire pipeline offline.

dsh-provider-rate-limit intercepts every outbound LLM stream at the llm/stream hook and enforces token-bucket rate limiting keyed by resolved provider and model, with configurable RPM, burst capacity, and a choice between queue-and-wait or fail-fast reject modes. When an upstream 429 arrives, the plugin enters an exponential-backoff cooldown window (with optional jitter) so the provider gets breathing room, and a concurrency cap prevents parallel sub-agents from overwhelming a route even when per-minute limits have not been reached. For identity-gated free gateways such as OpenCode Zen, built-in rules rewrite User-Agent and inject per-request dynamic headers, with a one-click preset for common setups.

Ideal for DeepSeek Harness users who fan out to multiple LLM providers—especially those reliant on free-tier or quota-limited gateways—and who need differentiated per-model limits, real-time queue-depth visibility in the composer dock, or the ability to query rate-limit statistics from other plugins in-process.