gitagenthq/git-agent-cli/using-git-agent

Git Agent CLI

AI-powered Git CLI for atomic AI commits. It is online and LLM-backed, so every install configures its own endpoint: base_url and model, plus an api_key for a hosted provider.

Start here

Use the bare command by default. Every commit is two steps — gather context first, then assemble and pass the intent. Never build the intent from memory alone:

1. Gather context: use the harness session-context tool (session_context

in pi) to collect the user's requests and decisions since the last commit; if no such tool exists, fall back to reading the diff yourself.

1. Assemble the intent: condense that context into one concise sentence

describing what changed and why — never a raw file list or diff dump.

1. Commit with the assembled intent:

git-agent --intent "<one-sentence intent assembled from the gathered context>"

The bare command is the autonomous agent entry point introduced by the recent CLI workflow. It inspects the repository and runs the commit workflow. Do not default to git-agent commit; use that subcommand only for an explicit commit case.

This file is a discovery stub, not the usage guide. Before running any git-agent command, load the actual workflow content from the CLI:

git-agent skills get core

The CLI serves skill content that always matches the installed version, so instructions never go stale.

Available documents

Why git-agent

• Autonomous workflow: the bare command inspects the repository and runs the commit workflow
• Atomic commits: splits changes into up to 5 logically distinct commits, each hook-validated
• Auto-scope: generates commit scopes from git history automatically
• Scope & `.gitignore` optimization: regenerates scopes from latest history (init --scope --force) and re-derives .gitignore while preserving custom rules (init --gitignore)
• Hook validation: conventional-commit validation built in; custom hooks via shell scripts
• Config precedence: CLI flag > git config --local > ~/.config/git-agent/config.yml. There is no built-in endpoint: set base_url and model, plus an api_key for a hosted endpoint or cloudflare_ai_gateway_id to route through your own AI Gateway. Agent-session env vars (PI_MODEL, CLAUDE_CODE_MODEL, CODEX_MODEL) never set the inference model — it resolves only from flag / local git config / user config; the session model is read separately for Co-Authored-By attribution
• Structured output: -o json on every read command for scripting