Each criterion's evaluator scores a pull request with a set of read-only tools. You choose the tools per criterion in the Tools picker of the repository's Criteria card. See Configure a Repository.

Every tool only reads. Evaluators read tracked files at the evaluated commit and never run the code.

Default Tools per Criterion

Criterion	Default tools
Blast radius	Read file, List files, Grep, Git diff base, Git log, PR files
Evidence quality	Reviews & comments, CI checks, PR metadata, Git log
Change footprint	PR files, Git diff base

The Conventions, Coding Sessions, and Track record tools, and Git blame and Commit verification, are not used by the default criteria. Add them to a criterion to give its evaluator that evidence.

Codebase

Tool	What it reads
Read file	A tracked file's contents.
List files	The repository's tracked files.
Grep	File contents matching a pattern.
Git log	Commit history for the default branch.
Git diff base	The diff against the merge base.
Git blame	Line-level authorship for a file.

Pull Request

Tool	What it reads
PR metadata	Title, description, labels, and size.
PR files	Changed files with status and line counts.
CI checks	Check runs for the head commit.
Reviews & comments	Reviews, inline comments, and conversation, each tagged with the commit it was made on.
Commit verification	Each commit's signature-verification state and author.

The title, description, and comments are written by people and treated as signals rather than proof.

Conventions

Tool	What it reads
Steering files	The repository's steering files, such as `CLAUDE.md` and `AGENTS.md`, from the base branch.

Steering files are read from the base branch, so a pull request cannot change the conventions it is judged against. Evaluators treat them as context, not as instructions.

Coding Sessions

These tools read AI SDLC Insights data about the coding sessions that produced the change. They need coding harness telemetry in Dash0. See Set Up AI SDLC Insights.

Tool	What it reads
Harness	AI coding harnesses used by the pull request author on this branch.
Tools & Skills	Tool and skill usage and status counts from the author's coding sessions on this branch.
Engineering activity	Classified activities such as planning, reviewing, or debugging. These are not proof of completed tasks.

Sessions are matched to the pull request by the author's email address, repository, and branch, over the 30 days before the evaluation. The email address is the one the coding harness reports. Missing session data never raises a score.

Track Record

These tools read AutoMerge's own history for the repository.

Tool	What it reads
Path authorship	How much of a file or module the pull request author wrote, and change counts. Run the Learning phase backfill in the Advanced card to reconstruct this history from the default branch.
Author delivery record	The author's merged, cleanly merged, and later-reverted pull request counts.
Review agent stats	Comment and resolution counts of AI review agents.
Evaluation history	This repository's AutoMerge evaluation statistics.

Evaluator Tools