Git Commit

Git Commit

Create technical commit messages, one per logical unit. Use the bundled grouping script for the deterministic part; spend tokens on Fix/Feature/Refactor refinement and message wording.

Process

1. Branch check. Run git branch --show-current. If on main / master and the user did not explicitly request committing there, create a feature branch first (feat/ or fix/) and switch to it.
2. Run scripts/group-changes.sh (default: working tree). Output is JSON:

• files[] — every changed path with {status, type, top} where type ∈ {types, config, code, test, doc}
• groups[] — pre-bucketed by (type, top-level dir)
• order[] — suggested commit sequence (types → config → code → test → doc)

1. Refine code groups. The script can't tell Fix from Feature from Refactor. For each type=code group, read the diff (git diff -- ) and decide:

• Fix vs Feature vs Refactor vs Types vs Config (re-classify if the script guessed wrong)
• Split if one group mixes intents (e.g., a fix and an unrelated refactor in the same dir)
• Merge if two adjacent groups belong together (e.g., type changes + the code consuming them, when tightly coupled and 6 commit units, surface the plan and ask before executing.

Edge cases

• All changes one logical unit → one commit; ignore the group split.
• Mixed intents in a single file → git add -p to stage hunks separately.
• Script reports code for a config-like file (rare, e.g., custom build script) → manually re-classify; the heuristic is path-based.
• Renames → script picks the destination path; rename + edits in same file land in one commit.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Shavakan
• Source: Shavakan/claude-marketplace
• License: MIT

Install and usage instructions live in the source repository linked above.