Agentic Refactor

Agentic Refactor

The mirror image of agentic-unit-test: there, tests change and code is frozen; here, code changes and tests are frozen. The agent tests are the definition of correct behavior.

3 RULES

1. NEVER edit any test — not agent tests, not user tests, not to "fix" a failure. A failing test means YOUR change is wrong.
2. Behavior must not change. Refactor = same observable behavior, better internals. If a desired change WOULD alter behavior, STOP and tell the user to use the agentic-test-update skill after deciding intentionally.
3. Work in small steps; every step ends with a fully green suite and a checkpoint commit.

Step 1 — Ask the user

1. "What should be refactored/optimized (files + goal: readability, performance, structure, ...)?"
2. "Is current agent-test coverage of those files enough to refactor safely?" — run the agent-tests-only coverage command (recipes: ../agentic-unit-test/docs/coverage-recipes.md) and show the per-file numbers for the target files. If any target file is poorly covered, recommend using the agentic-unit-test skill on that file first and let the user decide.

Step 2 — Preconditions (STOP if any fails)

1. All agent tests AND all user tests pass before touching anything.
2. Git repo, clean tree (git status --porcelain empty apart from plan files). Refactoring without git checkpoints is not allowed — STOP and say why.

Step 3 — Plan

Write agentic-refactor-plan.md at the repo root (resume it if present — continue from the first non-done entry after re-running the suite):

# Agentic refactor plan
Goal: 

- [ ] step 1:  — pending
- [ ] step 2: ...

Statuses: pending | in-progress | done | REVERTED (3 attempts): . Update after every step.

Step 4 — Refactor loop (per plan step)

1. Mark in-progress. Make ONE small change.
2. Build + run the FULL test suite (agent + user tests).
3. All green?

• Yes → lint changed files, fix lint in main code (allowed here), commit: git commit -am "refactor: [agentic-refactor]". Mark done.
• No → fix the CODE (never a test). 3 attempts. After the 3rd: git checkout -- . back to the last checkpoint, mark REVERTED (3 attempts): , move on.

1. Performance goal? Measure before/after (the suite's runtime is not a benchmark — use the project's benchmark or a quick timed script in the scratchpad) and record numbers in the plan entry.

Step 5 — Report

• Goal: restated, achieved or not.
• Steps: done vs reverted (with reasons).
• Verification: final full-suite run result (must be all green).
• Measurements: before/after numbers if performance was the goal.
• Commits made, in order.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: jpbaking
• Source: jpbaking/agentic-tests
• License: MIT

Install and usage instructions live in the source repository linked above.