AI provenance

Know which commits an agent wrote, and how much is still unknown.

Every commit is labelled agent, human-attested or unknown from the evidence your repository already carries. Any AI share is shown with the coverage behind it, beside health and ownership.

Start freeSee it on a real repo

No IDE plugins. No surveillance. It reads commits, not people.

ai provenance · last 90 days

Commits, read from git

coverage 60% · agent 64% of covered

  • feat(billing): retry failed charges

    a41f9c2 · agent, Claude Code · Co-authored-by trailer

  • fix(orders): handle empty carts

    9be03d1 · unknown · no marker

  • refactor(api): split routes module

    77c2e5a · agent, Codex · bot identity

  • docs(ledger): replay guarantees

    5c0d4e8 · human-attested · git-ai note

  • test(ledger): idempotent replays

    3d18b70 · agent, Cursor · commit trailer

Per file · share with its coverage

  • billing/retry.pyhealth 3.8

    agent 82% of covered · coverage 92%

  • api/routes.pyhealth 5.1

    agent 60% of covered · coverage 50%

  • orders/cart.pyhealth 8.2

    share withheld · coverage 21%

Withheld

orders/cart.py has too few commits with a known source for a share. Its 14 commits read 2 agent, 1 human-attested, 11 unknown.

Illustrative repository · labels read from git, every share shown with its coverage

At a glance

0
IDE plugins
Public markers are read from git history alone.
3
labels
Agent, human-attested or unknown. Unknown is never counted as human.
30%
coverage floor
Below it, an AI share is withheld and only counts are shown.
0
LLM calls
Labels and shares are deterministic.

Sound familiar?

Agents write code. The history rarely says so.

  • 01

    We cannot tell which commits an agent wrote.

  • 02

    Most commits carry no marker at all, so any percentage is a guess.

  • 03

    Every tool for this wants a plugin on every laptop.

What makes it different

What Repowise records, and how sure it is.

  1. Per commit

    Agent, human-attested or unknown, from the evidence in git.

    Works backwards through your existing history from day one.

  2. Per file

    Commits by source, with coverage.

    The agent share of covered commits, withheld when coverage is below 30 percent.

  3. Per repository

    A monthly trend of each label, with coverage.

    Unknown stays its own band, so the trend never reads missing evidence as human work.

  4. Not surveillance

    The unit is the commit and the file, never the engineer.

    A provenance record for the codebase, not a productivity ledger.

In plain words

Three labels, and the coverage behind every share.

Repowise reads the markers agents leave in git: bot identities, trailers, footers, git-ai notes and Agent Trace files. A commit with an agent marker is agent. A commit with an explicit human record is human-attested. Everything else is unknown, and unknown is never counted as human.

How to read it

How to read a provenance label

Agent
A marker or a captured edit ties the commit to a coding agent.
Human-attested
An explicit human record, such as a git-ai note naming only human authors.
Unknown
No marker either way. It stays unknown and is never counted as human.
Coverage
The share of commits with a known source. Below 30 percent the AI share is withheld.

How it works

Read, label, report with coverage.

A share without its coverage is a guess. Repowise prints both, and withholds the share when too little is known.
  1. 01 / Read

    Find the markers agents leave

    Bot authors, trailers, footers, git-ai notes and Agent Trace files. The open-source CLI adds opt-in edit-time capture.
  2. 02 / Label

    Three states per commit

    Agent, human-attested or unknown. No marker means unknown.
  3. 03 / Report

    Every share with its coverage

    Below 30 percent coverage the share is withheld and only the counts are shown.

Where it shows up

Shows up where the work already happens.

Questions

What people ask before they try it.

How do you know which agent wrote the code?

From markers your repository already carries: bot authors and committers, commit trailers and footers, git-ai notes and Agent Trace files. Every pattern is anchored to a service identity, never a bare name, so a contributor who happens to share an agent's name stays unlabelled. There is no IDE plugin, no keystroke capture and no model deciding who wrote what. With the open-source CLI you can also turn on edit-time capture for Claude Code, Codex and Cursor; it stays off until both the repository and the user opt in.

What do the three labels mean?

Agent means a marker or a captured edit ties the commit to a coding agent. Human-attested means an explicit human record, such as a git-ai note that names only human authors. Unknown means there is no marker either way. In many repositories most commits carry no marker, and Repowise says so rather than guessing.

Why is unknown never counted as human?

Because a missing marker is missing evidence, not evidence of a person. Counting unknown commits as human would turn every gap in the history into a claim. An AI share is the agent fraction of covered commits only, the ones with a known source, and it is always shown with that coverage.

What is coverage, and when is a share withheld?

Coverage is the fraction of commits with a known source: agent plus human-attested. A share reads like "at least 40% of covered commits, coverage 62%". Below 30 percent coverage the share is withheld and only the counts are shown, because a share over so little evidence would mislead.

Is this developer surveillance?

No. Provenance is a record about the codebase, not a per-developer productivity ledger. It reads commits, it does not watch people, and it never ranks or scores a person. Labels come from markers already in the commits, never from watching anyone work.

Does it work on our existing history?

Yes. Public markers are read from the history you already have, all the way back, so commits that carry them are labelled on day one. Commits without a marker stay unknown. With the open-source CLI, opt-in edit-time capture and transcript backfill can raise coverage for the agents you choose.

How is this different from GitClear?

GitClear reports engineering and AI-adoption trends across an organization. Repowise labels the commits in your repository from the evidence they carry, keeps unknown as its own state, and shows every share with its coverage, beside the health and ownership of the same files.

Last reviewed: October 2026

Get started

See which commits an agent wrote.

Index a public repository free and open its commits, labelled, with the coverage behind every share.

Start freeBook a demo