Agent
Implements a whole small PR test-first in one agent — plans behavior slices, drives RED→GREEN→refactor for each through the public interface, follows the project rules, and lands one commit. Two modes — BUILD (implement the spec from scratch) or FIX (apply named review/critic blockers). Used by the make-pr-lite skill.
Agent
Scores the comments and docstrings a diff adds against comments.md — one score, a verdict, and one finding per comment that must be cut, shortened, moved, rewritten, or added, each carrying the replacement text. Judges only comments; the reviewer owns code quality and the critic owns goal-fit. Used by the make-pr…
Agent
Judges goal-fit — whether a change actually accomplishes its stated task, not whether the code is well-written. Scores 1-100, gives a verdict, lists what's missing. Distinct from the reviewer (code quality). Used by the make-pr architect and the make-pr-lite skills before declaring a task/PR done.
Agent
Reviews a diff for code quality, correctness/bugs, and security against the project rules. Reports structured findings with severity; does not fix code. Answers "is this code good?" — distinct from the critic, which answers "did it achieve the task?". Used by the make-pr architect and the make-pr-lite skills.
Agent
Writes exactly ONE failing test for ONE behavior through the public interface, runs it, and confirms it fails for the right reason (the RED step). Never writes production code. Used by the architect skill at the start of each TDD cycle.
Agent
Implements exactly one scoped task in one of three modes the architect names — GREEN (minimal code to pass one named failing test), REFACTOR (behavior-preserving restructure, tests stay green), or NON-BEHAVIORAL (an exact config/rename/docs edit) — then commits. Scope-locked: never exceeds the single task it was…
Agent
Frozen fixture agent for installer e2e snapshot scenarios.