dsh-auto-driving

DeepSeek Harness tasks are only as reliable as the single provider they are pinned to. One quota rollover, a brief 5xx spike, or an unanswered permission dialog is enough to leave a multi-step agent job hanging at 80 percent, waiting for a human to notice. dsh-auto-driving removes that dependency by letting the task itself navigate around failures.

Under the hood the plugin adds three self-healing layers. First, model auto-fallback: the moment a request fails before it has streamed any text, reasoning, or tool output, the plugin walks a priority-ordered list of provider groups and re-dispatches the very same request to the next healthy model, so the running agent loop never sees a break. Second, task auto-retry with a liveness watchdog: transient network errors trigger exponential-backoff retries on the same model, while a configurable idle timeout (default five minutes) kills and re-fires requests that silently stall. Third, full-auto mode: when toggled on, every permission approval, confirmation question, and exit_plan_mode review is answered automatically as allow, and each decision is logged line-by-line into an AUTO-MODE.md file at the workspace root so you can audit what happened without replaying the session. All of this is controlled from a single capsule next to the send button, takes effect on the very next request, and requires no restart.

The plugin is built for DSH Desktop users who run long agent workflows across multiple providers and want those flows to survive real-world API instability unattended. If you have more than one model supplier configured and your biggest pain point is a task dying mid-stream or freezing on an approval prompt, dsh-auto-driving is the layer that keeps it moving.