DeepSeek Harness (DSH) uses a plugin-based architecture to address problems in agent development such as interrupted sessions, model failures, and human approval bottlenecks. dsh-auto-driving is a workflow plugin that enables unattended, non-stalling, self-healing AI coding sessions through automatic fallback, retries, and approval takeover.

Core Capabilities

Automatic Model Failure Fallback

If a request fails before any output is produced—for example, because the model does not exist, credentials are missing, quota is exhausted, rate limiting occurs, or a server-side 5xx error happens—it automatically switches to the next model in the fallback group and continues executing the same request until an available model is found.

Automatic Task Retries

For non-model issues such as network timeouts or disconnections, it retries with the same model using exponential backoff until the operation succeeds or the maximum retry count is reached.

Fully Automatic Mode

When enabled, all permission requests, confirmation prompts, and plan reviews during task execution are automatically answered using the default allow response, so the task is no longer blocked by manual validation. Each automatic decision is appended to the AUTO-MODE.md file in the workspace root directory.

Liveness Watchdog

Detects silent hangs (default: 5 minutes) and forces a retry or model switch.

Real-Time Bubble Notifications

When switching models, resuming execution, or performing automatic approvals, real-time bubble notifications are displayed above the conversation input box.

Responsive Capsules

Capsules automatically collapse to icons only, maintaining stability under input bar compression.

Installation and Activation

After installation, the plugin adds an “Autopilot” tab to the DSH settings interface. Fully automatic mode is off by default and must be explicitly enabled.

Install URL: https://www.skillhub.cn/plugins/huhaodong/dsh-auto-driving
GitHub repository: https://github.com/huhaodong/dsh-auto-driving

Typical Usage

  1. Open the “Autopilot” tab in settings and configure fallback groups and retry parameters.
  2. After enabling fully automatic mode, permission, question, and plan approvals are completed automatically without human intervention.
  3. Inspect the AUTO-MODE.md file in the workspace root directory, where all automatic decisions are recorded.
  4. Use the input bar capsule to instantly toggle fully automatic mode and model fallback for the current conversation.
  5. Configure parameters such as initial wake-up delay and backoff multiplier to control retry behavior.

Applicable Scenarios and Notes

This plugin is suitable for scenarios where agent tasks must run unattended for extended periods. Check the source code and license before installing.

  • Node.js version requirement: >= 20
  • SSE fallback: Automatically switches to interval polling when a long-lived connection is unavailable
  • State synchronization: When switching conversations, old bubbles disappear immediately; when switching back, if the state is within 30 seconds, it is restored as before