my-git

Git Integration Practices for AI Coding Tools

English

中文

Original links:

1. Common Trends

AI coding tools are shifting from “completing code in the editor” to “completing tasks around the Git workspace.”

Their common direction is clear:

Tasks are triggered from Issues, Prompts, PR comments, or local commands.
Modifications occur in isolated branches, worktrees, sandboxes, or remote environments.
Output is presented as diffs, commits, PRs, logs, and validation results.
Merging still requires human review, CI, and repository rules.

This indicates that Git remains the collaboration boundary in AI programming.

2. Codex: Remote Environments, Sandboxes, PR Collaboration

The core feature of Codex is executing tasks in isolated environments and feeding results back to GitHub or local workflows.

Recommended team usage:

Give Codex a clear task; avoid letting it freely explore the entire repository.
Require it to output changed files, validation commands, and risks.
Use PRs or diffs as entry points for human review.
Maintain a default tight grip on permissions, networking, and command execution.
Split large tasks into multiple small tasks for parallel execution.

The value of the Codex sandbox lies in limiting the execution scope. The ability for AI to run commands is a source of efficiency but also a source of risk; teams must include permissions, network access, sensitive files, and external writes in their rules.

Recommended PR description addition:

AI tool:
Codex

Human intent:
Fix timeout validation for expired config.

Human checked:
- git diff --stat
- changed files
- test result
- rollback path

3. Claude Code: Worktree Isolation for Parallel Sessions

Official Claude Code documentation explicitly recommends using worktrees as a way to isolate parallel sessions.

It solves the problem of multiple AI sessions overwriting each other in the same repository.

Recommended practice:

claude --worktree feature-auth
claude --worktree bugfix-payment-timeout

Or create manually:

git worktree add ../repo-feature-auth -b ai/feature-auth
cd ../repo-feature-auth
claude

Note:

Add .claude/worktrees/ to .gitignore.
Do not copy .env, local configurations, or secret files to worktrees by default.
Check git status after each worktree is finished.
Reorganize branches to follow team naming conventions before merging.
Clean up useless worktrees to avoid too much leftover local context.

4. GitHub Copilot Cloud Agent: From Issue to PR

The typical path for GitHub Copilot Cloud Agent is triggering a task from an Issue or GitHub entry point; the Agent analyzes requirements, modifies code, commits to a branch, and creates a PR.

Suitable for:

Small fixes
Documentation additions
Test completion
Low-risk refactoring
Issues with clear acceptance criteria

Not suitable for:

Large features with unclear requirement boundaries
Problems requiring production data for judgment
High-risk changes involving security, permissions, billing, or payments
Architectural adjustments requiring cross-team approval

Teams should write Issues more like task orders:

Goal:

Scope:

Out of scope:

Acceptance criteria:

Tests to run:

Risk:

5. Aider: Git-First Local Pair Programming

Aider’s characteristic is its deep use of Git: it can automatically commit AI modifications, and you can also use /diff, /undo, /commit, and /git to manage changes.

Best for local developers who want to pair program quickly, but teams should pay close attention to automatic commit strategies.

Suggestions:

Ensure the workspace is clean before starting.
Do not mix uncommitted human changes with AI modifications.
Letting Aider commit automatically can improve the convenience of rolling back, but commits should still be manually organized before merging.
For repositories requiring pre-commit hooks, clarify whether to enable --git-commit-verify.

Recommended workflow:

git status
git switch -c ai/aider-small-fix
aider
git log --oneline -5
git diff main...HEAD
git rebase -i main

6. Cursor: Multi-Agent and Aggregated Diffs

The official Cursor 2.0 release emphasizes Composer, multi-agent parallelism, and a more centralized diff review experience.

This tool form’s requirements for Git workflows are:

Multi-agent tasks must have clear boundaries.
Each agent’s output must be individually reviewable.
Aggregated diffs can only serve as an entry point and cannot replace file-level review.
Final commits should be split by logic to avoid committing directly based on tool execution results.

Suitable for using Cursor in:

Exploring multiple solutions
Local transformations of UI or frontend
Documentation and test additions
Small-scale refactoring

7. Unified Team Rules

Regardless of the AI tool used, teams should unify these rules:

Rule	Recommended Practice
Task Boundary	Process only one clear goal at a time
Isolation Method	Branches, worktrees, sandboxes, or remote environments
Diff review	Humans first check `git diff --stat` and key files
Commit Splitting	Split by tests, implementation, documentation, and configuration
Validation Results	PR must clearly state commands and results
Merge Responsibility	Human reviewer assumes final judgment
Rollback Path	Each PR must explain how to undo it

8. Recommended Minimum Workflow

Issue / prompt
-> isolated branch or worktree
-> AI edits
-> human checks diff
-> split commits
-> run tests
-> AI review as assistant
-> human review
-> PR
-> CI
-> merge

The stronger the AI tools, the clearer the Git workflow must be.

This site is open source. Improve this page.