AgentSkillsCN

deepcode

DeepCode 集成——基于论文与文本需求的自动化代码生成

SKILL.md
--- frontmatter
name: deepcode
description: "DeepCode integration - automated code generation from papers and text requirements"
metadata: {"nanobot":{"always":true}}

DeepCode - AI Code Generation Engine

You have access to DeepCode, a powerful multi-agent AI code generation engine that can:

  • •Paper2Code: Reproduce research paper algorithms as working code
  • •Chat2Code: Generate complete projects from text descriptions

Available Tools

ToolPurpose
deepcode_paper2codeSubmit a paper URL or file for code reproduction
deepcode_chat2codeSubmit text requirements for code generation
deepcode_statusCheck task progress and results
deepcode_list_tasksList active and recent tasks
deepcode_cancelCancel a running task
deepcode_respondRespond to User-in-Loop interactions

When to Use DeepCode

Automatically trigger deepcode_paper2code when user:

  • •Sends an arxiv URL (e.g. https://arxiv.org/abs/... or https://arxiv.org/pdf/...)
  • •Sends a paper URL from other academic sites
  • •Asks to "reproduce", "implement", or "replicate" a paper
  • •Sends a PDF file and asks for code generation
  • •Says something like "帮我复现这篇论文" or "把这篇论文的代码跑出来"

Automatically trigger deepcode_chat2code when user:

  • •Describes a coding project they want to build
  • •Asks to create a web app, backend service, algorithm implementation, etc.
  • •Provides detailed requirements for a software project
  • •Says something like "帮我写一个..." or "生成一个项目..."

Workflow Guidelines

1. Submitting a Task

When the user wants to generate code:

  1. •Identify if it's a paper (use deepcode_paper2code) or requirements (use deepcode_chat2code)
  2. •Submit the task and note the task_id
  3. •Tell the user the task has been submitted and the estimated wait time (10-60 minutes for papers, 5-30 minutes for chat)
  4. •Offer to check progress periodically

2. Monitoring Progress

  • •When user asks about progress, use deepcode_status with the task_id
  • •Report the progress percentage and current phase
  • •If the task is complete, share the result summary

3. Handling User-in-Loop Interactions

  • •Check deepcode_status - if status is "waiting_for_input", there's a pending interaction
  • •Read the interaction details (questions, plan review, etc.)
  • •Present the questions/plan to the user in a natural conversational way
  • •Collect the user's response
  • •Use deepcode_respond to submit the response back to DeepCode

4. Delivering Results

When a task completes:

  • •Report the generated file structure
  • •Mention key files (e.g. model.py, train.py, requirements.txt)
  • •The generated code is in the shared deepcode_lab/ directory
  • •Offer to read specific files if the user wants to review them

Response Style

  • •Be concise and informative about task status
  • •Use progress percentages to show advancement
  • •When a task completes, provide a brief summary of what was generated
  • •For Chinese-speaking users, respond in Chinese (follow the user's language)

Important Notes

  • •Code generation tasks run in the background and take time (10-60 minutes)
  • •Do NOT spawn subagents for DeepCode tasks - use the tools directly
  • •If DeepCode backend is unreachable, inform the user that the service may not be running
  • •Generated code is stored in /app/deepcode_lab/papers/ directory