deepseek-harness-leetcode

DeepSeek Harness LeetCode

Learn algorithms inside DeepSeek Harness: lessons, all of LeetCode Hot 100, a sandboxed judge, and an AI tutor.

English | 中文

Why

• All of LeetCode Hot 100, plus 8 more classics — each with an original summary, three progressive hints, and an explained solution. Statements stay on LeetCode; every problem links to the official page.
• A real judge on your machine. Code runs in Docker with no network. Hidden boundary, random, and large cases catch wrong and slow solutions — including design classes, in-place edits, and linked-list/tree node problems.
• Lessons before problems. 16 topics, from hashing to dynamic programming, each with a worked example you can trace step by step. Then the tutor writes a short check about the lesson, reviews your answers, and decides when you're ready for the practice problem.
• Add problems by just describing them. "3 medium BFS problems I haven't done" — an AI agent finds them on LeetCode, writes the tests, and adds them only if they pass validation.

A closer look

Hot 100 — LeetCode's own grouping, with your progress per group

Solve — editor, Run and Submit, hidden tests	Submissions — runtime, memory, and every past attempt

Your record — solved by difficulty, a yearly calendar, streaks, and recent passes

Learn — worked examples you can trace	Hints — direction, key idea, outline, solution

Tutor check — questions written for the lesson, answers reviewed by the tutor (real DeepSeek output)

Also included: an AI tutor chat, spaced-repetition reviews with fresh validated variants, progress by concept, and optional daily reminder email.

Quick start

1. Install Docker Desktop. Learn then guides the rest: it can start Docker and download the runner images with one click, or run docker pull python:3.13-alpine node:24-alpine yourself.

   

2. Install the plugin from github:kapelame/deepseek-harness-leetcode:

   • Desktop: Plugins → install from GitHub.

   • Web:

     dsh plugin --profile web add github:kapelame/deepseek-harness-leetcode
     dsh web
     

3. Open Learn in the sidebar. Explanations default to Chinese; switch to English in Learn → Settings.

Needs DSH 0.2.0-rc.2 or newer, below 0.3 (tested on 0.2.0-rc.2 and 0.2.1-alpha.3). A default model (for example, signed in to DeepSeek) is optional: it powers the tutor, tutor-written lesson checks, AI feedback, and Add problems. Running, judging, written hints, and the lessons' built-in check questions work without one.

"/Applications/DeepSeek Harness.app/Contents/Resources/runtime/cli/bin/dsh" plugin --profile desktop add github:kapelame/deepseek-harness-leetcode

How judging works

Every verdict comes from test execution only. Your code runs in a Docker container with no network, a read-only filesystem, and memory, CPU, and process limits; expected answers never enter the container. Each suite combines official example inputs with hidden boundary, random, and large cases, built from a reference solution and checked against an independent brute-force solution. A pass means "passed local tests", not LeetCode acceptance. The optional AI judge adds feedback but can never change the verdict.

Add problems

Practice → Add problems. Describe what you want and pick how many (1–5). Progress streams live, and you can stop at any time; problems already added stay.

• The agent follows the bundled skill skills/learn-problem-author/SKILL.md, which is visible only to the finder agent, never in your normal Sessions.
• Number, title, difficulty, tags, signature, and examples always come from LeetCode's metadata, never from the model. Paid-only problems and problems that need a judge Learn does not provide are skipped.
• Each problem must pass the same validation as the built-in ones before it is added.
• In our live tests with DeepSeek, a run took 70–300 seconds and added every problem requested.

Privacy

• Your code runs only inside offline Docker containers.
• With a default model, Submit (unless agents.judge: false), hints, tutor chat, reflections, and variant generation send the problem summary and your code to that model through DSH. Lesson checks send the lesson and your answers. Add problems sends your description and LeetCode metadata, never your code. Each run appears in your Session history.
• Add problems queries LeetCode's public GraphQL API for problem metadata only: no login, no statements, paced requests with a per-run budget.
• Nothing else leaves your machine, except reminder email if you configure SMTP.

- id: learn
  config:
    dataRoot: !!js dshHomePath('learn')
    defaultLocale: en
    reviewDailyLimit: 3
    agents:
      judge: false              # deterministic judging only, no AI feedback
      variantGeneration: true
    finder:
      enabled: true
      maxProblems: 5            # problems per Add problems run
      maxSubmissionsPerProblem: 3
      maxLeetcodeRequests: 40   # LeetCode requests per run
      requestIntervalMs: 1000
    email:
      transport: capture        # 'smtp' sends only with liveSendingAuthorized: true

git clone https://github.com/deepseek-ai/deepseek-harness.git dsh
git -C dsh checkout dsh-v0.2.0-rc.2
node scripts/build.mjs ./dsh    # copy sources in, install, build, assemble lib/

License

MIT