AI-assisted changes
How Prodgator marks a pull request or deployment as made with an AI coding tool, and the settings that control it.
Prodgator marks each pull request and each finished deployment as AI-assisted or not, so you can compare AI-assisted work with the rest. This is part of AI impact, on the Business and Enterprise plans. It never sends anything to an AI model and never uses AI credits. The AI impact page compares the two groups.
How a pull request is classified
A pull request is AI-assisted when at least one of these is true. Prodgator records which ones matched, so you can see why.
| Signal | What counts |
|---|---|
| Commit trailers | A commit in the pull request names an AI tool in its trailers, the lines at the end of the message. For example , , , , a Copilot cloud agent line or a "Generated with Claude Code" line. The same lines at the end of the pull request description count too. |
| Commit author | A commit in the pull request was authored by an AI agent's identity, such as the Copilot cloud agent (shown as ). |
| Agent author | The pull request was opened by an AI agent's account, such as , , , or , or by an account you added. |
| Labels | The pull request has one of your AI labels. The defaults are and (the label Anthropic adds to merged pull requests it matches to Claude Code sessions). |
| Claude Code telemetry | Off by default. Claude Code reported a commit in the pull request through its OpenTelemetry logs. Recorded as the signal . See Claude Code telemetry. |
| Agent branch | Off by default. The pull request's branch starts with a default agent prefix: , , , or . A weak signal, because people can use these names themselves. |
What does not count:
- A tool named in the body of a commit message, above the trailers.
- A co-author with a person's address, even if their name is Claude or Devin, and even with a GitHub no-reply address. Prodgator matches a tool only by the tool's own address.
Recognized tools: Claude, GitHub Copilot, Cursor, Devin, OpenAI Codex, Gemini, Jules, Kiro, Amazon Q Developer and Aider. An explicit trailer for another tool counts as "other".
Prodgator reads up to 250 commits of each pull request when it is opened, pushed to, relabeled, its description is edited, and when it is merged or closed.
Providers
| Provider | Commit trailers | Agent author | Labels |
|---|---|---|---|
| GitHub | Yes | Yes | Yes |
| GitLab | Yes | Yes | Yes |
| Bitbucket Cloud | Yes | Yes | No: Bitbucket pull requests have no labels |
| Azure Repos | Yes | Yes | Yes |
How a deployment is classified
When a deployment finishes, Prodgator looks at the commits since the last successful deployment to the same environment and at the pull requests that contain them. The deployment is AI-assisted when one of those pull requests is, or when one of the commits has an AI trailer (a direct push). For the first deployment to an environment, only the deployed commit and its pull request are looked at.
Settings
Org admins change these in Admin Console > AI > AI-assisted changes:
- Read commit trailers: turn off to ignore trailers and description lines and classify by agent accounts and labels only.
- Read agent branch names: turn on to count the default agent branch prefixes.
- Use Claude Code telemetry: turn on to count commits that Claude Code reports through its OpenTelemetry logs. Records are counted while this is off, so you can check the setup first. See Claude Code telemetry.
- Labels: the labels that mark a pull request as AI-assisted. Leave the list empty to ignore labels.
- Agent accounts: accounts that open pull requests for an AI agent, beyond the built-in list. Put , , or in front to limit one to a provider, for example .
Changes apply to pull requests and deployments classified after you save. Classification records are kept for 400 days.