Introduction

Running a coding agent is relatively easy, but operating it continuously and reliably is more challenging: someone still has to decide which issues are ready, track token usage, limit concurrency, recover interrupted work, and notify the team. DeepSeek Harness Autopilot (DSH Autopilot) adds this operational management layer to DSH. It is only responsible for automating work between “approved issues” and “review-ready pull requests”; it does not interfere with final review, CI decisions, or the merge process. This plugin is currently in the Alpha stage.

Core Capabilities

This plugin translates DSH’s capabilities into an operable pipeline through a series of mechanisms:

  1. Inbound Control: It only accepts issues with a configured ready marker and currently human-authorized issues. This ensures only confirmed tasks enter the processing queue.
  2. Queueing and Scheduling: Work is queued under scheduling and concurrency limits, with configured allowance reserved to prevent exceeding provider token limits.
  3. Worktree Execution: Runs the repository’s own instructions within managed Git worktrees via DSH, keeping the environment isolated.
  4. Lifecycle Management: Supports pause, resume, and cancel operations without discarding persistent run history. Even when interrupted, it can recover retained worktrees, interrupted releases, and notifications.
  5. State Validation: Before creating a pull request, it revalidates the exact Git state to ensure the committed changes match expectations.
  6. Web Observability: Displays queues, run details, worktrees, delivery status, and settings in the DSH Web interface for easy monitoring.
  7. Human Decision Points: The workflow stops at pull request creation, with final code reviews, CI decisions, and merge authority still held by humans.

Installation and Enablement

Installing this plugin requires an existing DeepSeek Harness environment.

  1. Install the plugin (using the Alpha version):
dsh plugin --profile web add @canhta/dsh-autopilot@alpha
  1. Start the DSH Web service:
dsh --profile web
  1. Open the Web interface from the URL printed in the terminal, and select Autopilot in the sidebar.

Configuration

Open Settings → Autopilot in DSH Web to configure tracker and repository bindings, scheduling policies, concurrency limits, budgets, notifications, and retention policies.

1. Preparation

Before enabling execution, select the default Agent preset and model in DSH Web first, and configure each provider in the DSH Settings namespace. Credentials (OAuth, model keys, Tokens) should remain in DSH Credentials; do not create new authentication flows through Autopilot.

2. Configure MCP Servers

Autopilot communicates with external services through DSH’s official MCP server (@deepseek-ai/dsh-mcp-client), not third-party clones.

Edit the configuration file $DSH_HOME/profiles/<profile>/cordis.patch.yml and add the following configuration:

GitHub MCP Configuration:

- insert:
    - id: mcp-github
      name: '@deepseek-ai/dsh-mcp-client'
      config:
        serverName: github
        transport: streamable-http
        url: https://api.githubcopilot.com/mcp/
        headers:
          Authorization: !!js '`Bearer ${process.env.GITHUB_TOKEN}`'

Jira MCP Configuration:

- insert:
    - id: mcp-jira
      name: '@deepseek-ai/dsh-mcp-client'
      config:
        serverName: atlassian
        transport: streamable-http
        url: https://mcp.atlassian.com/v2/mcp?tools=all
        headers:
          Authorization: !!js >-
            `Basic ${Buffer.from(`${process.env.JIRA_EMAIL}:${process.env.JIRA_API_TOKEN}`).toString('base64')}`

After configuration is complete, restart the Web Profile (dsh --profile web).

Notes

  • Environment Dependency: DeepSeek Harness must be installed and running first.
  • Not a Replacement Integration: This plugin does not replace DSH itself or provider integrations (such as GitHub/Jira MCP servers); it only provides scheduling and lifecycle management.
  • Version Status: The current version is Alpha, and rollout to production environments is limited by provider compliance.
  • Budgeting Mechanism: The budget field reserves capacity for reconciliation but is not an exact cumulative hard limit.
  • Jira Authentication: When using Jira, API token authentication must be enabled.
  • Notification Disruptions: Disruptions in the notification service do not revoke completed Pull Requests.

Conclusion

DeepSeek Harness Autopilot provides developers with an operational layer for stable agent execution in a DSH environment. It addresses the pain points of continuously operating agents through queue management, persistent worktrees, and state validation. For more details or to contribute code, you can access its GitHub repository.