Git AI Trace — AI skill for Claude Code

git-ai-trace

A Claude Code skill that embeds an honest, observation-only recap of human/AI collaboration inside each commit message — so the story of how the code was made travels with the code.

Why

Codebases written with AI assistants raise a question that binary disclosure ("was AI used?") can't answer: **who actually drove what?** A developer who reframed the AI's first proposal three times exercised strong creative control, even if most keystrokes came from the model. A generated file that passed through without comment is a delegation, regardless of intent.

Percentages lie about this ("70% AI-written" is meaningless). Labels flatten it ("ACE" vs "HCE" forces every session into five bins). File-based logs lose it (a separate `AI_CONTRIB.md` breaks on rebase, generates merge conflicts, and drifts from the commits it tries to describe).

`git-ai-trace` takes a different path: a short block of **observed moments** — proposals, choices, rejections, generations — embedded directly in the commit message. Each moment is one line, pointable to a specific message in the chat. Absences are never recorded, only presences. The recap is incomplete by design, but it is honest.

What a recap looks like

fix: prevent duplicate job processing under load

Workers were occasionally picking up the same job. Row-level
locking closes the race without adding infrastructure.

--- session-recap ---
What changed: Worker loop now claims jobs via SELECT FOR UPDATE SKIP
  LOCKED; batch size reduced from 100 to 20.
Moments:
- human: brought symptom (duplicate processing in prod logs) and initial hypothesis (lock contention)
- claude: diagnosed non-atomic SELECT+UPDATE across workers as root cause
- claude: proposed three remediation options (advisory lock, SKIP LOCKED, Redis queue) with trade-offs
- human: rejected Redis-queue option as over-engineering
- human: chose SKIP LOCKED option
- human: set batch size to 20
- claude: wrote the SQL patch
--- end-recap ---

Assisted-by