How it works

Diffs are slugs

A diff is base..head, where each side is a git ref, WT (working tree), or STAGED:

staff main..WT              # current main → uncommitted working tree
staff abc1234..HEAD         # a specific commit → HEAD
staff main..STAGED          # main → what you’ve git-added
staff                       # just open the UI on the active diff

The slug is also how the skills reference a diff: /staff-review main..WT targets (and if needed creates) that exact diff and makes it active.

The web UI

A GitHub-style review of the active diff:

• Click a line number to anchor a comment; drag or shift-click to select a range.
• Hover a line and click the + to comment; reply to build a thread.
• Resolve a thread as Fixed or Skipped, or flag it with Document so /staff-resolve captures it as a reusable lesson.
• The diff and the comment sidebar scroll independently, with light/dark themes, split/unified view, syntax highlighting, the /staff-loop round cap, and the /staff-review agent count — all in the gear menu.

Comments are stored under .staffreview/diffs/<slug>.json; the running UI picks up changes over a WebSocket and re-renders live as the agent posts.

The skills

staff install writes nine skills to .agents/skills/ (symlinked into .claude/skills/ so Claude Code picks them up as slash commands) — six you run directly (/staff-review, /staff-resolve, /staff-comment, /staff-document, /staff-docs, /staff-loop) plus three building blocks the orchestrators fan out to (/staff-review-find, /staff-review-verify, /staff-docs-scout). They’re just Markdown — open .agents/skills/staff-review/SKILL.md and tune it to your team’s standards. See the Skills reference for what each one does.

The docs

.staffreview/docs/ is your team’s captured review wisdom. /staff-review reads it on every pass and re-flags any recurrence of a documented mistake. The docs is meant to be committed; the session data (.staffreview/diffs/, attachments/, active.json) is gitignored automatically. See The docs.

The loop

/staff-loop chains review and resolution in subagents until a review surfaces nothing new, capped by the loopMaxRounds setting (default 5, adjustable in the gear menu). It’s the “set it going and come back to a higher-quality result” button.