The core philosophy of DeepSeek Harness (DSH) is “everything is a plugin.” When developing or debugging agents, models often output complete code directly, which can cause developers to lose their coding intuition. The dsh-human-coding plugin introduces a “Traditional Coding” mode that returns control of model output to the user, allowing the model to only pose questions and perform review, thereby improving programming skills through interactive practice.
It is a DSH workflow plugin maintained by gameswu. It uses AI to generate skeletons and tests, guides users to complete the implementation, and has AI perform acceptance scoring, addressing the problem of eroded development intuition caused by over-reliance on AI-generated code.
Core Features¶
- AI question generation: AI writes the skeleton and tests into the project, then presents a question. Triggering requires a request that involves real code (pure Q&A, configuration adjustments, and single-line edits do not trigger it).
- Answering mode: Once answering mode begins, AI only answers questions and gives hints. It never writes code on your behalf.
- Hint consumption: Each question deducts one hint from the budget (the budget for each question is set by AI when generating the question), and hints become progressively more specific—direction → approach → near the answer. After the budget is exhausted, AI will guide you to submit or give up.
- Acceptance scoring: After submission, AI runs the tests, reviews your implementation, and gives a completion score from 0 to 1.
- Persistent statistics: The result of each challenge is saved as a long-term record and persists across restarts.
- Record management: Supports exporting records and importing them (overwrite or append).
Installation and Enablement¶
Before installing, ensure the DSH version is ≥ 0.1.5-rc.1 (this package v0.6.0 is compatible). Run the following command in the repository root:
bash scripts/install.sh # Linux / macOS / Git Bash
.\scripts\install.ps1 # Windows PowerShell
The script builds the plugin package, installs it into the Profile, and prepares the “Traditional Coding” preset. After updating the plugin, rerun the installation script.
After installation, restart the Profile, refresh the page, and choose “Traditional Coding” when creating a new session to get started. Note: after a session is created, refresh the page to see the status bar.
Typical Usage¶
- Accept the challenge: Enter answering mode; from then on, AI only answers questions and gives hints, and never writes code.
- Ask questions: If you are stuck, ask for help; hints become more and more specific.
- Submit or give up: You can click Submit or Give Up at any time (typing it directly also works).
- Acceptance choices: After submission, AI runs tests and reviews the implementation. You can then choose:
- Keep your implementation: Your implementation is retained, and AI provides a summary and commentary.
- Minor AI edits: AI makes small modifications based on your implementation and explains each change.
- Full AI rewrite: AI rewrites the solution in its own style and compares/explains the differences.
Applicable Scenarios and Notes¶
Suitable for developers who want to improve their coding intuition through interactive practice. The plugin runs with the permissions of the current DSH process; you should check the source code and license before installing.
- Note: Question state is persisted and does not leave extra files in the project; Traditional Coding mode is a “one-on-one practice session with the human user” and is not distributed to teammates.
With this mechanism, developers can obtain immediate feedback and scoring while maintaining a normal development workflow. For more details, see the plugin directory page and the GitHub repository.