Introduction

When processing complex engineering tasks in DeepSeek Harness (DSH), high-level objectives usually need to be broken down into sequential milestones and progress remains clearly visible. Without an autonomous tracking framework, agents may lose context between turns, fall into passive loops, or fail to notify users when complex workflows stall.

Core Features

goodandready/dsh-goal introduces Goal Mode to DSH. It is an autonomous goal execution and multi-turn task tracking engine with a Sticky Header.

State and Control

  • Sticky Top Goal Banner: a banner fixed above the input bar that displays elapsed time, status (RUNNING/PAUSED/COMPLETED), goal title, and control buttons in real time.
  • Play / Pause / Resume / Cancel: instantly pause, resume, or cancel the autonomous loop via buttons or commands.
  • Milestone Breakdown & ETA: an interactive checklist showing subtask status (pending/in_progress/completed/failed), progress bars, and dynamic completion estimates.

Agent Enhancements

  • Live Steering & Nudge: dynamically clarify, modify, or steer the agent’s current goal without terminating the loop.
  • Autonomous Agent Contract: provides the agent with goal_set_milestones, goal_update_progress, and goal_finish tools.
  • Core DSH Goal Tools Interception: seamlessly supports models that call built-in DSH goal tools. Intercepts operations such as complete, pause, and resume, bypasses permission restrictions that cause crashes, and routes all state updates to GoalEngine.

Safety and Stability

  • Tool-Failure Breaker: stops the loop to prevent token waste when consecutive tool errors reach the configured threshold (default is 3).
  • Safety Guardrails: configurable maxIterations safety limits and Smart Progress Guard for detecting and pausing idle turns with no progress.

Feedback and Export

  • Native Browser Notifications: desktop push notifications based on the HTML5 Notification API for completed long tasks or situations requiring attention.
  • Token Usage Tracking & Markdown Export: tracks prompt, completion, and total token counts during a session and supports one-click export of a Markdown summary.
  • Web Audio Chimes: synthesized audio feedback for completion or failure using the Web Audio API.

Quick Launch

  • 10 Quick Launch Presets: preset engineering prompt chips, including Fix Bug, Refactor YAGNI, Tests & Coverage, Code Review, Security Audit, New Feature, Docs, Upgrade Deps, Dead Code Cleanup, Performance.
  • Quick Launch Button: a quick launch button fixed above the message input box that opens a goal input modal on click.

Engineering and Collaboration

  • Git Checkpoint Snapshot: automatically captures the starting commit hash and provides git diff and rollback commands.
  • GitHub & Gitea PR Comment Export: generates a formatted PR comment with one click, including collapsible milestone cards and token telemetry data.

Language Support

  • Multi-Language Auto-Detection: automatically detects the language of user input (English, Chinese, Russian), with UI labels and prompt injection following automatically.

Usage

The plugin interacts with users through a command-line interface:

  • /goal <objective>: starts a new autonomous goal.
  • /goal pause: pauses the current goal and stops agent turns.
  • /goal resume: resumes execution and triggers the agent to continue.
  • /goal clear: resets the session goal to IDLE.
  • /goal: displays status, elapsed time, ETA, iteration count, token count, and active milestones.

Architecture

The plugin mainly completes tasks through the collaboration of the following components:

  1. User Interface and Interaction: handles /goal commands, the Sticky Top Goal Banner, the Goal Details & Milestone Modal, and the Quick Launch Button.
  2. Goal Engine Core: core engine (lib/goal-engine.js), responsible for managing session goals, milestone status, time calculation, token accumulators, and language detection (detectLanguage). It syncs state to persistent storage (goal-state.json).
  3. Autonomous Loop: automation driver and turn coordinator (lib/index.js), listens to turn/end events. It includes Smart Progress Guard and an iteration limiter, and injects prompts to the agent based on state.
  4. Agent Tools: toolset registered to Cordis, including goal_set_milestones, goal_update_progress, and goal_finish.

Installation and Dependencies

The plugin runs with the permissions of the current DSH process. Check the source code and license before installation.

  • Node version: Node 20+ required.
  • Dependencies: @deepseek-ai/schemastery.
  • Peer Dependencies:
    {
      "@deepseek-ai/cordis": "^4.0.1",
      "@deepseek-ai/dsh-host-webserver": "^0.1.0-rc.6",
      "@deepseek-ai/schemastery": "^3.18.1",
      "@deepseek-ai/dsh-goal": ">=0.1.5-rc.3"
    }

Use Cases and Notes

  • Use cases: suitable for long-running, multi-step automated engineering tasks, such as code refactoring, automated testing, documentation generation, or complex debugging workflows.
  • Note: because the plugin directly intervenes in DSH core tool-calling logic, ensure its implementation is reviewed before installation to meet your security and audit requirements.

Conclusion

This plugin solves the issues of context loss and lack of feedback for DSH agents in long-chain tasks by providing a Sticky Header, status visualization, and tool interception mechanisms. Developers can learn more details through GitHub or the community directory.