Skip to content

chainlit.md

Latest commit

 

History

History
197 lines (146 loc) · 11.1 KB

chainlit.md

File metadata and controls

197 lines (146 loc) · 11.1 KB

Agent Tide Logo

Welcome to Agent Tide! 🚀🤖

█████╗  ██████╗ ███████╗███╗   ██╗████████╗    ████████╗██╗██████╗ ███████╗
██╔══██╗██╔════╝ ██╔════╝████╗  ██║╚══██╔══╝    ╚══██╔══╝██║██╔══██╗██╔════╝
███████║██║  ███╗█████╗  ██╔██╗ ██║   ██║          ██║   ██║██║  ██║█████╗  
██╔══██║██║   ██║██╔══╝  ██║╚██╗██║   ██║          ██║   ██║██║  ██║██╔══╝  
██║  ██║╚██████╔╝███████╗██║ ╚████║   ██║          ██║   ██║██████╔╝███████╗
╚═╝  ╚═╝ ╚═════╝ ╚══════╝╚═╝  ╚═══╝   ╚═╝          ╚═╝   ╚═╝╚═════╝ ╚══════╝

What is Agent Tide?

Agent Tide is a next-generation, precision-driven software engineering agent built on top of CodeTide. It enables you to interact directly with your local codebase—everything is local and stays local. No code, context, or metadata ever leaves your machine.

Unlike most agent frameworks that rely on hidden inner logic, opaque planning, or LLMs doing everything in a single pass, Agent Tide offers a transparent, stepwise, and human-in-the-loop workflow. Think of it as "ChatGPT, but connected to your codebase"—with full visibility and control over every change.

How Does Agent Tide Work?

  1. Direct Codebase Connection & Context Loading

    • Agent Tide uses CodeTide's fast, symbolic codebase parser to build a tree of all code identifiers (functions, classes, methods, etc.).
    • When you make a request, Agent Tide analyzes it and loads only the relevant identifiers and their dependencies—retrieving just the code snippets needed for your task.
    • This context loading is fast, deterministic, and fully local.

  2. Three Modes of Operation

    • Clarification: If your request is ambiguous, Agent Tide will ask for more information before proceeding.
    • Direct Implementation: For simple, clear requests, Agent Tide will immediately generate a patch to fulfill your request.
    • Planning Mode: For complex tasks, Agent Tide enters a planning mode—decomposing your request into a step-by-step plan. You can review, edit, or reorder steps before execution.

  3. Human-in-the-Loop, Stepwise Execution

    • You always see the plan before any code is changed. You can modify steps, request clarifications, or approve the plan.
    • When ready, click "Proceed to next step" to have Agent Tide implement each step sequentially.
    • After each step, you can review the patch, provide feedback, or continue to the next step—keeping you in control at all times.

  4. Patch-Based, Token-Efficient Editing

    • All code changes are generated as atomic, high-precision diffs (patches), not full file rewrites.
    • This means you see exactly what will change, and can stop or correct the agent at any point—no waiting for a huge batch of changes to finish before intervening.
    • This approach is also highly efficient for LLM token usage and latency.

  5. Privacy & Local-First Philosophy

    • No code, context, or metadata is sent to any external service other than the selected LLM provider.
    • All code analysis, context building, and patching is performed locally.

  6. Model Compatibility

    • Agent Tide works best with GPT-4.1, Claude Sonnet 4, and Opus 4.1.
    • Some models (e.g., DeepSeek Chat) may struggle with patch/diff generation.

  7. Terminal & MCP Integration

    • Terminal access is not available yet.
    • MCP (Multi-Component Patch) integration via aicore is deactivated by default in this demo, but can be enabled for more advanced use cases.

Why Agent Tide?

  • Transparency: Every step, plan, and patch is visible and editable before execution.
  • Control: You decide when and how changes are applied—no "black box" agent behavior.
  • Precision: Code changes are atomic, minimal, and always shown as diffs.
  • Speed: Context loading and patching are fast, thanks to CodeTide's symbolic engine.
  • Privacy: Your code never leaves your machine.

Example Things to Ask Agent Tide

You can ask Agent Tide to perform a wide variety of code-related tasks. Here are some example prompts to get you started:

  • Add new functionality

    • "Add a function to calculate the factorial of a number in utils.py."
    • "Implement a REST API endpoint for user registration."

  • Refactor or improve code

    • "Refactor the process_data function to improve readability and performance."
    • "Rename the variable x to user_id throughout models/user.py."

  • Fix bugs

    • "Fix the bug where the login form crashes on empty input."
    • "Resolve the off-by-one error in the get_page method."

  • Write or update tests

    • "Add unit tests for the EmailSender class."
    • "Increase test coverage for api/views.py."

  • Documentation

    • "Generate a docstring for the parse_config function."
    • "Update the README with installation instructions."

  • Code analysis and suggestions

    • "List all functions in main.py that are missing type annotations."
    • "Suggest performance improvements for the data_loader module."

  • Other codebase tasks

    • "Delete the deprecated old_utils.py file."
    • "Move the helpers directory into core/."

Feel free to be specific or general in your requests. Agent Tide will analyze your codebase and generate precise, production-ready patches to fulfill your needs!

Usage Tips & Advanced Workflows

  • Direct Context Control:
    If you already know the exact code context you want Agent Tide to use, you can specify identifiers directly in your request.
    Use the format:

    module.submodule.file_withoutextension.object

    For example:

    Please update codetide.agents.tide.ui.app.agent_loop to support async.

    This helps Agent Tide load only the relevant code blocks, making the process faster and more precise.

  • Planning and Human-in-the-Loop:
    Agent Tide allows you to interact directly with your local codebase—everything is local and will stay local.
    Most agent frameworks are filled with hidden logic, letting the LLM do all the planning and execution at once.
    Agent Tide offers a more direct perspective:

    • The first thing Agent Tide does is use the CodeTide repo tree to load identifiers related to your request, then builds context blocks with the relevant code snippets (and their dependencies).
    • Once context is loaded, Agent Tide can:
      1. Ask for more information or clarification if your request is ambiguous.
      2. Start implementing changes right away if the request is simple.
      3. Enter planning mode and present you with a step-by-step plan (you can also request this explicitly: "create me a plan to do X").
    • You can edit the plan, reorder steps, or request changes until you are satisfied.
    • Click "Start implementing steps" and Agent Tide will go step by step, letting you review, correct, or continue after each one.
    • This keeps you in the loop and gives you direct control and visualization of every change.
    • Code is generated as diffs for optimized token usage and latency—so you can intervene at any point, even if the agent is working on many files.

  • Model and Integration Notes:

    • Agent Tide works best with GPT-4.1, Claude Sonnet 4, and Opus 4.1.
    • Models like DeepSeek Chat may struggle with patch generation.
    • Terminal access is available for advanced workflows.
    • MCP integration (via aicore) is deactivated by default in this demo.

Available Slash Commands

Agent Tide supports special commands that can be invoked directly from the chat input using a /command syntax. These commands provide quick access to advanced features and workflows.

How to Use

Type /command (replace command with the actual command name) in the text input box and press Enter. Agent Tide will recognize and execute the command.

Available Commands

  1. /test

    • Description: Request test implementations for a specific file, function, or object. Use this to add or improve test coverage for any part of your codebase.
    • When to use: When you want Agent Tide to generate or enhance tests for a specific area.
    • Usage:
      /test add coverage for apply_patch tool

  2. /review

    • Description: Request a code review for a specific file, function, or recent patch. Agent Tide will analyze the code and provide feedback or suggestions.
    • When to use: When you want a review of code quality, style, or correctness.
    • Usage:
      /review codetide/agents/tide/ui/app.py

  3. /commit

    • Description: Commit changed files. This command will stage and commit all recent changes, generating a conventional commit message. You can also stage files manually and use this command to write a message and commiting them.
    • When to use: After reviewing and accepting code changes, to save your work with a proper commit message.
    • Usage:
      /commit the changes we just made

  4. /plan

    • Description: Create a step-by-step task plan for your request. Agent Tide will decompose your request into actionable steps, which you can review and edit before execution.
    • When to use: For complex or multi-step tasks where you want to see and control the implementation plan.
    • Usage:
      /plan add a new authentication system to the project

  5. /brainstorm

    • Description: Brainstorm and discuss solutions for a problem or feature, without generating code. Use this to explore ideas, approaches, or architectural decisions.
    • When to use: When you want to discuss possible solutions, strategies, or designs before writing code.
    • Usage:
      /brainstorm ways to improve the performance of the data pipeline

  6. /direct_mode

    • Description: Skip repository analysis and jump straight into code generation with the specified context (identifiers or paths). Use this if you know exactly what you want to change and want to bypass the agent's context loading.
    • When to use: For advanced users who want to target specific code blocks directly and speed up the process.
    • Usage:
      /direct_mode codetide.agents.tide.ui.app.agent_loop

You can use these commands at any time to guide Agent Tide's workflow, request reviews, generate commit messages, brainstorm, or create implementation plans. More commands may be added in the future—refer to this section for updates.

Original repository: https://github.com/BrunoV21/CodeTide

This README was written by AgentTide!