Agent Provenance: Which Commits an Agent Wrote, with Coverage

Why does agent provenance matter?

Agents now write a real part of most codebases, and the history rarely says which part. A percentage with nothing behind it is a guess: most commits carry no marker at all, and counting every unmarked commit as human turns missing evidence into a claim.

• Which commits an agent wrote is a fact about the code, worth recording where it can be read.
• How much of the history carries a known source decides whether any share means anything.
• A record of the codebase must never become a score for the people working in it.

Provenance is one part of a wider practice. The AI code governance guide covers the rest, from review and ownership to a copy-paste AI coding policy, and its chapter on recording which agent wrote each change goes deeper on the method.

How does it work?

repowise treats provenance as evidence with a coverage, not a number to display. The mechanism is deterministic end to end, computed from the git log you already have.

1. Read the markers. With no configuration, indexing reads public markers from local git history. No network calls, no model, no prompts.

2. Label every commit.

3. Report with coverage. Coverage is the fraction of commits with a known source. The AI share is the agent fraction of covered commits, always shown beside its coverage. Below 30 percent coverage the share is withheld and only the counts are shown.

This is a record about files and commits, not a per-developer ledger. It reads commits; it does not watch people.

Where you read it

• In the dashboard: the commits page labels each commit, filters by authorship and shows a 12-month trend with its coverage.
• In your AI agent: get_context file cards carry the file's commits by source and their coverage, beside health and ownership.
• In the open-source CLI: repowise attribution status prints commits by label with coverage, and what your settings capture.

Raising coverage

Public markers only reach commits that carry them. With the open-source CLI, for the agents you use, edit-time capture records which lines an agent wrote as it writes them, and transcript backfill binds past commits from the transcripts those agents keep locally. Both are off until the repository and the user turn them on, store line hashes rather than prompts or code, and leave an unmatched commit unknown.

Walkthrough: from commit to provenance

Step 1: Index your repos. Index a repository to build the graph and git history. No IDE plugins.

Step 2: Open the commits. Each commit carries its label, and the trend shows agent, human-attested and unknown per month with the coverage beside it.

Step 3: Read a file. The file card shows its commits by source and coverage next to its health and owner. A share appears only above the coverage floor.

Step 4: Raise coverage where it matters. With the open-source CLI, turn on capture or backfill for the agents your team uses, so new commits carry a known source.