Test cleanup

Find the tests that cost more than they prove, and let a person decide what goes. This skill changes no test. It writes reports, asks the person, and hands the approved cleanup to a task. The task makes the edits through the normal stages, so each edit is reviewed like any other change.

This skill never runs the suite, and never runs a test. A test's worth is judged by reading it. Running the suite here would cost time and prove nothing about which tests to keep.

Every call below runs cleanup-actions.sh, named in this skill's own grant, so it runs without asking. Any other Bash command still asks.

1. Find the project

"${CLAUDE_PLUGIN_ROOT}"/skills/test-cleanup/scripts/cleanup-actions.sh where

It prints project: and one framework: line per framework. Exit 1 means no project owns this directory. Say so in one line, name /aida:project, and stop.

2. Take the test globs from the implement recipe

Run this once, with one --framework flag per framework:

"${CLAUDE_PLUGIN_ROOT}"/skills/project/scripts/project-actions.sh recipe-point "<project folder>" \
  --point implement --point test-authoring --framework <framework>

Implementation asks for the same two recipes. Then pass each implement answer= to the script:

"${CLAUDE_PLUGIN_ROOT}"/skills/test-cleanup/scripts/cleanup-actions.sh where --implement-lookup "<framework>=<recipe path or reason>"...

Pass one flag per framework. The value is the recipe path, or no-recipe, listing-unreachable or fetch-failed when the lookup found no recipe. The script reads the globs from the recipe's ## Oracle files block, the same globs the freeze of a test takes. It prints one glob: line per glob. Use each glob as printed. Do not change one, and do not add one.

A no-glob: line names a framework that gave no glob. Ask the person where the tests of that framework live and how they are named, as globs. Do not guess a glob from file names.

3. Group the test files

"${CLAUDE_PLUGIN_ROOT}"/skills/test-cleanup/scripts/cleanup-actions.sh groups [--max-lines <n>] --glob "<glob>"...

Pass one --glob per glob from step 2, and --max-lines only when the person gave one. The script decides the groups: it fills each group in folder order, up to the lines one reader can read. It prints run:, code: and one group: line per group. Exit 2 means no tracked file matched the globs. Show the globs to the person and ask which ones are wrong.

Say how many groups there are before the next step. Each group costs one dispatch.

4. Read each group

Dispatch test-reader once per group, naming the role, all in parallel. The message holds four lines: the run folder, the group id, the code folder, and the test-authoring recipe paths, or none. The role holds the rubric, and writes <run folder>/reports/<group id>.json.

5. Add up the reports

"${CLAUDE_PLUGIN_ROOT}"/skills/test-cleanup/scripts/cleanup-actions.sh summary "<run folder>"

Exit 3 names each group whose report is missing or malformed. Dispatch the reader again for those groups only, then run summary again. Exit 0 prints the path of summary.md. Show the person its category table and its causes.

6. Ask the person, one group at a time

For each group with proposals, show its proposals from the group's report: the action, the target and the reason. Ask what to do with them. The person can approve all, approve some, change an action, or refuse all. Ask about one group per turn. Skip a group with no proposals, and say so.

Write each answer in <run folder>/decisions.md as the person gives it: the group id, then each target with the action the person approved. A refused proposal is left out. Never write an answer the person did not give. Nothing in the code changes in this step.

7. Hand the cleanup to a task

When every group is answered, and at least one change was approved, offer a task. With a yes, run /aida:task create. Propose the name test-cleanup. Propose this goal, with the run folder as groups printed it: "Apply the test cleanup the person approved in <run folder>/decisions.md. A merge, a shrink or a rewrite of a test is a tests order that owns only test files. So is one small test kept for a behaviour that only a removed test checked. A check script that is rewritten is code, with tests of its own. Whole files to delete, check scripts included, are one confirm order that owns only those files. It names each file in a done-when row and depends on every tests order. A runner configuration or script entry that names a deleted file goes to another order, which depends on the deletion order. On a light run, wo1 is the first tests order, or the deletion order when there is none. A removed end-to-end test that is a registered surface goes with surfaces-actions.sh unregister <id>." Then ask whether to run it light, and on a yes run /aida:task set-run-mode <task-id> light. The task skill asks for the person's answer there.

With no approved change, say that nothing will change, and stop.

What this skill never does

It never edits, moves or deletes a test. The task does that, after the person approved it.

It never runs the suite or a test. The task's own stages run the suite once at the end.

It never counts a category itself. The script adds the counts, so the summary says what the reports hold.