Classify repository tests and prune cases that fail the value bar

# tests-diet `tests-diet` splits a repository's test files into groups, runs a pool of pi agents that judge every test case against the `writing-tests` value bar, and writes a report. It never edits code during a run; applying the report is a separate step. ## Classify Settings live in `tests-diet.config.json` at the root (`"$schema": "https://arthur.run/tests-diet/schema.json"`); flags override them for one run. ```bash npx tests-diet plan # groups and token estimates, no model calls npx tests-diet run --limit 2 # calibrate on two groups, then review .tests-diet/tests-diet.html npx tests-diet run # classify the rest; resumable, safe to interrupt npx tests-diet status # progress, failures, usage and cost so far ``` - Repo-specific rules (protected areas, house patterns) go in the file named by `instructions`. Changing them, the model or the rubric makes earlier results stale, so a rerun classifies again. - Unusual layouts: `extraLanguages` in the config, or `--match '<glob>[:<content regex>]'` and `--case-start '<regex>'`. ## Apply a report 1. `npx tests-diet delete --dry-run`, then `npx tests-diet delete`. It deletes `delete_file` files and removes every `delete` range, only in files unchanged since they were reported. Add `--only <path>` to go one folder at a time. 2. `.tests-diet/tests-diet.md` lists what is left, per file, with ranges as `startLine-endLine` (1-based, inclusive). It is large on big repositories: every `## <path>` section stands alone, so read it one folder at a time. 3. Merges: cases sharing a merge group become one test (a table, or one flow) that keeps every assertion that matters. Find them by test name, since line numbers moved after step 1. A group whose "covered by" names cases in other files is combined across those files. 4. `rewrite`: leave for a follow-up unless asked. 5. Remove what became unused: imports, helpers, fixtures, mock files, production exports or flags that only those tests used. 6. Run the project's formatter, linter, type checker and the affected test suites. A failure in a kept test means an edit went wrong: restore it. Group the edits by package or area, one change set each, listing removed cases and lines.