Adding GitHub **Agent** Node

[satococoa]

Prototype here: #403

Idea

We at Giselle aim to provide an excellent GitHub Integration. Similar to Claude Desktop with GitHub MCP, our GitHub Agent Node will receive natural language input and interact with the GitHub API.
The GitHub Agent Node can:

• Search Code/Issues
• Create Issues/Pull requests
• Add review comments to pull requests
• ...and more

Implementation details

Reference: https://github.com/giselles-ai/giselle/blob/6897bca98f814d425cacc75c3cbe1ebae259c936/packages/github-agent/README.md

Component Architecture

classDiagram
    class Agent {
        -Octokit octokit
        -Planner planner
        -ToolRegistry toolRegistry
        -boolean isDebug
        -number maxRetries
        +execute(prompt: string)
        +static builder()
        +static fromAllTools(token)
    }
    class Planner {
        +plan(prompt: string)
        +planWithEvaluation()
    }
    class Evaluator {
        +evaluate(plan, results)
    }
    class ToolRegistry {
        +register(tool)
        +dispatchTool(input)
        +generateToolDescriptions()
        +toolInputSchema()
    }
    class Formatter {
        +format(result, options)
    }
    Agent --> Planner
    Agent --> Evaluator
    Agent --> ToolRegistry
    Agent --> Formatter
    ToolRegistry --> "n" Tool

Loading

Architecture

The project consists of the following components:

Core Components

• Agent: Main agent class that handles user requests and manages execution
• Planner: Analyzes user requests and breaks them down into executable steps
• Evaluator: Evaluates execution results and determines success/failure
• ToolRegistry: Manages available GitHub API tools
• Formatter: Formats API responses into readable tree structures

Tool Groups

The tools are organized into logical groups:

• Issues: Tools for managing issues (list, retrieve)
• PullRequests: Tools for working with pull requests (retrieve, reviews,
  comments, diffs)
• Files: Tools for file operations (retrieve contents)
• Search: Tools for searching code, repositories, users, and issues
• Repos: Tools for repository management (list commits)
• Utils: Advanced utility tools like GraphQL and REST API access

Execution Flow

flowchart TD
    A[User Request] --> B[Planner]
    B --> C[Generate Plan with Single Tool Call]
    C --> D[Execute Tool Call]
    D --> E[Evaluator]
    E --> F{Success?}
    F -->|Yes| G[Format Results]
    G --> H[Return Results]
    F -->|No| I{Retry Count < Max}
    I -->|Yes| B
    I -->|No| J[Return Error]

Loading

1. Receive user request
2. Planner analyzes request and selects appropriate tool
3. Execute single tool call
4. Evaluator assesses results
5. Format results into readable structure if successful
6. Replan and retry if necessary
7. Return final results to user

Consideration

• Can we leverage the existing MCP server?
• Do we need to create this plan -> evaluate mechanism ourselves, or could an existing agent framework provide more sophisticated functionality?