Letter2025/dsh-model-failover

This is a native plugin built exclusively for DeepSeek Harness that adds two-level circuit breaking and automatic failover for LLM requests. You can install it directly to your DSH profile with a single command: dsh plugin --profile web add dsh-model-failover, and no core DSH code changes are required to use it. It hooks into two official DSH agent loop extension points to track failures and reroute requests when outages occur. A companion configuration skill is installed automatically with the plugin to help you set up fallback models without manual file editing.

The plugin tracks failed LLM requests and opens a model-level circuit when a set number of failures occur within a configurable burst window. If enough individual models from a single provider open circuits, it triggers a provider-level circuit to mark the entire platform as unhealthy. When a circuit is open, all new requests are routed to the first healthy fallback model in your preconfigured ordered list. After a cooldown period, it automatically probes the unhealthy model to check for recovery, and closes the circuit automatically if the probe succeeds.

The plugin is released under the open-source MIT license, and it works with the bundled DSH llm-retry policy, only acting on failures that escape retries. It has several known limitations that you should be aware of before installation: circuit state is stored in process-local memory and resets when the plugin reloads, and only main conversation loop requests are routed through the failover system. Auxiliary LLM calls like session title generation or context compaction are not covered, and you must select fallback models that match your context window size requirements.