SKILL.md
Git Agent CLI
AI-powered Git CLI for atomic AI commits. It is online and LLM-backed; official releases provide a shared gateway, so no personal API key is required.
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:
- 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.
- Assemble the intent: condense that context into one concise sentence
describing what changed and why — never a raw file list or diff dump.
- 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 coreThe CLI serves skill content that always matches the installed version, so instructions never go stale.
Available documents
| Document | Command |
|---|---|
| Main usage guide (triggers, workflows, flags, exit codes) | git-agent skills get core |
| Full command reference (all flags, subcommands, config scopes, hook types) | git-agent skills get cli |
| All available documents | git-agent skills list |
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 &
.gitignoreoptimization: regenerates scopes from latest history (init --scope --force) and re-derives.gitignorewhile 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> free shared-gateway default;--freeforces the gateway and ignores the rest. Agent-session env vars (PIMODEL,CLAUDECODEMODEL,CODEXMODEL) never set the inference model — it resolves only from flag / local git config / user config; the session model is read separately forCo-Authored-Byattribution - Structured output:
-o jsonon every read command for scripting