Git

Git

Use this skill to perform repository operations in a safe, auditable way. The skill must only call operations listed in allowedOperations; if a requested operation is not permitted the agent should inform the user and request configuration changes.

Principles

• Safety first: Prompt for confirmation before any destructive action (reset --hard, push --force, branch -D, revert), unless requireConfirmationForDestructiveOps is false and the user explicitly allowed the operation.
• Least privilege: Only perform operations in allowedOperations.
• Explicit context: Always resolve repositoryPath before running commands, and refuse to act if the path is not a git repository.
• Auditability: Return both machine-readable results and a short human summary describing what was done.

Supported operations (examples)

• status — show git status --porcelain and a human summary of staged/unstaged/untracked files.
• add — stage files. Params: paths (array) or paths: ["."] to add all.
• commit — create a commit. Params: message, author (optional), sign (optional).
• push — push to remote. Params: remote (default defaultRemote), branch, force (boolean).
• pull — pull from remote. Params: remote, branch, rebase (boolean).
• fetch — fetch updates.
• branch — create/list/delete branches. Params for create: name, startPoint.
• checkout — switch branches or restore files. Params: branch, paths.
• merge — merge a branch into current. Params: source, noFastForward.
• rebase — rebase current branch onto another. Use carefully.
• log — show commits. Params: maxCount, format.
• diff — show diffs. Params: paths, commitRange.
• stash — push/pop/list stashes.
• tag — create/list/delete tags.
• reset — soft/mixed/hard reset. Destructive when --hard.
• revert — create a revert commit for a specified commit.

Process

1. Parse goal to determine the operation and parameters.
2. Verify the requested operation exists in allowedOperations.
3. Resolve repositoryPath (use defaultRepositoryPath if unset) and ensure it contains a git repository.
4. If operation is destructive and requireConfirmationForDestructiveOps is true, present a short explanation and ask for confirmation before proceeding.
5. Run the operation using a safe runtime helper (prefer a git tool or the shell tool exposed by the runtime). Capture stdout/stderr and exit code.
6. Return result with relevant fields (e.g. commit id, branch name, pushed refs) and summary explaining the outcome.

Examples

• Commit staged changes with message:

Goal: "Commit staged changes with message 'Fix validation bug'"
Operation: commit
Params: { "message": "Fix validation bug" }

• Create and push a branch:

Goal: "Create branch feature/auth and push to remote"
Operation: branch
Params: { "name": "feature/auth", "startPoint": "main" }
Then: push with params { "branch": "feature/auth" }

• Show last 5 commits:

Operation: log
Params: { "maxCount": 5 }

Safety notes

• Never expose secret credentials or tokens in command outputs. If the runtime returns environment or config values, redact secrets before returning them.
• Avoid running repository-altering commands on the wrong path — always confirm the repo root and current branch before destructive operations.
• If a command fails with merge conflicts, return the conflict details and suggest next steps rather than attempting automatic resolution.

Integration guidance

• The runtime should expose a git tool or shell tool to execute commands. Prefer a language-native git library when available for parseable results.
• Map skill allowedOperations to tool-level permission gates where possible.