Commit attribution stats

bernstein report commits reports git commit counts and line churn grouped by agent role, so an operator can see which roles (backend, qa, security, ...) produced how much of a repository's history without hand-parsing git log.

CLI

bernstein report commits                     # all-time stats for the current repo
bernstein report commits --since 2025-01-01  # date-bounded stats
bernstein report commits --until 2025-06-30  # date-bounded stats
bernstein report commits --repo-dir ../other-repo
bernstein report commits --json              # machine-readable output

bernstein commit-stats remains as a deprecated alias with identical behaviour until v4.0.0; it prints a migration notice on stderr.

Flag	Default	Meaning
--since TEXT	none	Passed straight to git log --since.
--until TEXT	none	Passed straight to git log --until.
--repo-dir PATH	.	Repository to run git log against.
--json	off	Print CommitStatsResult.to_dict() instead of a Rich table.

Output is a table with one row per role, plus a totals row: commit count, lines added, lines deleted.

How role is determined

There is no task/role metadata behind this command — it is entirely a git author string heuristic. For every commit, bernstein lowercases the author's name <email> string and checks it for one of a fixed set of keywords:

backend, frontend, qa, security, devops, docs, manager, architect

The first keyword found in the author string becomes that commit's role. If none match, the role label is the full lowercased author string (so a human contributor's commits show up under their own name/email rather than being dropped).

Practical effect: this only produces meaningful role buckets when agent commits carry a role-tagged author identity (for example backend-agent <backend@bernstein.local>). A human co-author whose name or email happens to contain one of the keywords (e.g. a teammate named "Devora" would not match, but an email containing devops would) is attributed to that role bucket rather than to themselves — the heuristic does not distinguish humans from agents.

Data source

Two git log invocations against --repo-dir (default: the current directory):

1. git log --numstat --format=%an <%ae> --date=short — line-level added/ deleted counts per commit, attributed by author.
2. git log --format=%an <%ae> — a plain commit count per author.

Both respect --since/--until when given. If git is not installed or the directory is not a repository, the command exits non-zero and prints the underlying error (or the error field, in --json mode).

Also surfaced in bernstein doctor

bernstein doctor runs the same collect_commit_stats() call (unfiltered by date) as one of its checks, labelled Commit attribution. It reports {total commits}: {role}: {commits} commits, +{added}/-{deleted} per role for info, or the underlying git error if the check fails.

Limitations

• Role attribution is a keyword match on the author string, not a link to Bernstein's actual per-task role assignment — it will misattribute any author whose name or email happens to contain a role keyword.
• No de-duplication or merge-commit filtering: line counts come straight from git log --numstat, so squash/merge history affects the totals the same way it affects git log itself.

Source

• src/bernstein/cli/commit_stats.py — collect_commit_stats, _author_to_role, render_commit_stats
• src/bernstein/cli/commands/status_cmd.py — commit_stats_cmd (bernstein commit-stats), _doctor_check_commit_attribution (bernstein doctor integration)