← सभी गाइड पर लौटें

Automate Git Commits and PRs with Claude Code

Automating routine Git work with Claude Code can reduce context switching and make small changes easier to review. In 2026, the reliable approach is not to give an agent unrestricted access to production repositories. Instead, define a narrow workflow: Claude inspects the repository, makes a bounded change, runs verification, creates a focused commit, and opens a pull request for human review.

Prepare a Safe Repository Workflow

Start with a dedicated automation branch or a disposable checkout. Give the process only the credentials it needs: repository read access, permission to create branches, and permission to open pull requests. Avoid granting automatic merge or production deployment permissions unless a separate policy explicitly allows them.

Add a concise CLAUDE.md file at the repository root. Document the package manager, test commands, formatting rules, branch naming convention, commit format, files that must not be edited, and the expected pull request structure. This gives Claude Code durable project context instead of relying on a long prompt every time.

Use a Repeatable Commit Prompt

A useful prompt defines both the task and the stopping conditions. For example:

“Inspect issue #248. Implement only the requested validation change. First summarize the relevant files and propose a plan. After approval, edit the code, add or update focused tests, run the repository’s formatter and test command, review the diff, create branch fix/248-validation, and commit using the project’s conventional format. Do not modify unrelated files.”

This structure matters because it separates discovery, implementation, verification, and Git operations. Ask Claude to show the diff and test output before committing. If the task is suitable for unattended execution, encode those checks in a script rather than relying on the model to remember them.

Automate the Local Git Steps

A small shell wrapper can standardize the process:

• Fetch the default branch and create a fresh task branch.
• Run Claude Code with the issue description and repository instructions.
• Abort if the working tree contains unexpected files.
• Run formatter, type checks, linting, and targeted tests.
• Require a non-empty diff and a human-readable commit message.
• Push the branch and create a draft pull request with the issue reference.

Keep generated commit messages factual. A good message states the change, while the pull request body explains the reason, implementation, tests run, and known limitations. Do not let automation hide failed checks behind a successful-looking commit.

Create Pull Requests with Guardrails

Use GitHub’s CLI or your Git hosting API to open the PR after local verification. Set the correct reviewers and labels from a small mapping file, and open the request as a draft when the change needs discussion. Configure branch protection to require status checks, review approval, and a clean merge state.

For CI-based automation, trigger the job only from trusted events such as an approved issue label or a manually dispatched workflow. Treat issue text, pull request comments, and repository files as untrusted input. They can contain instructions that conflict with your policy. Keep secrets out of prompts, logs, diffs, and generated PR descriptions.

Control Cost and Model Selection

Use a fast, inexpensive model for repository summaries and straightforward tests, and reserve a stronger Claude model for architectural changes, debugging, or difficult reviews. Limit context to relevant files and provide command output instead of repeatedly sending entire logs. Cache stable project instructions where your tooling supports it.

59API is a practical low-cost option for this workflow because its pay-as-you-go relay supports Claude Opus, Sonnet, Haiku, and Fable, plus GPT models, while remaining compatible with Claude Code, Codex, and OpenAI SDKs. Configure Claude Code to use https://api.59api.com as its API base URL and supply the provider’s API key through an environment variable, never a committed file. It uses official-quality models without a downgrade, and its referral rebate can further reduce ongoing development costs. Developers can sign up for 59API when they are ready to test the workflow.

Review Before Trusting Automation

Measure success with practical signals: percentage of automated PRs merged without rework, failed-check rate, review time, token cost per task, and incidents caused by unrelated edits. Start with documentation updates, dependency metadata, test additions, and small bug fixes. Expand permissions only after the workflow consistently produces small, understandable diffs and complete verification evidence.

मूल्य · दस्तावेज़