Add btrain review code rules that flag weakened tests (#104) * Add btrain review code rules that catch weakened tests The rule scanner only saw added lines and skipped files with none, so a deleted test file, a removed assertion or a loosened check never showed up. Parser: read deleted/new/rename headers and quoted paths, and collect removed lines with old line numbers plus per-hunk added, removed and ordered entries. The existing file/added/lines output is unchanged. The git diff call now pins --find-renames, --no-color, --no-ext-diff and a/ b/ prefixes against local config. New rules (warn unless noted), scanned with string, regex and comment contents masked: - deleted-test-file: test file deleted or renamed to a non-test path - removed-assertion: a test hunk removes more assertion lines than it adds - skipped-test: new unconditional skip/todo/fixme (conditional skips pass) - focused-test (hard): new .only / fit / fdescribe / only: true - loosened-assertion: strict check on a subject replaced by a looser one in the same hunk, or throws/rejects/raises losing its matcher - lowered-threshold: number-only line changes that lower numRuns, FORMAL_RUNS, coverage floors or lower bounds, or raise retries/upper bounds - test-ignore-added: structural; new entry in an existing ignore list (testPathIgnorePatterns, testIgnore, exclude in JS test config, collect_ignore*, [review_code] ignore_list_keys, plus btrain's CANDIDATE_REASON_LABELS and KNOWN_DRIFTS), or a new --deselect / --test-skip-pattern in a test command The pre-handoff skill (both copies) now says hard findings block and each warning is fixed or recorded as a --gap bullet with a one-line justification. Acceptance: 9150811 gives 6 warnings, c7f5d77 gives 1. Over the last 80 merged commits the new rules report only those 7 findings. Co-Authored-By: Claude Opus 5.5 * Fix review findings in the weakened-test review rules P2-1: the assertion-line regex nested quantifiers after `assert` and backtracked exponentially on a long identifier (a 32-character assert_A_A... line took 126 s). It now uses one character class after a lookahead. A child-process test with a time bound scans pathological lines up to 105,000 characters; masking treats lines over 5,000 characters as data, and call parsing looks at 2,000 characters at most. P2-2 (focused-test false positives and a false negative): - Masking now runs over whole files. reviewCode loads both sides of every changed test file, so a template literal or doc comment opened above a hunk is respected in either direction. When contents are missing, focused-test is reported as warn instead of hard. - only / skip / todo options count only as top-level properties of a test call's options object, including a multi-line object. pattern.test(x) and assertion data no longer count. - A focus or skip call needs a title and a callback unless it uses .each, and fit / xit helpers that the file defines or imports from a non-test module are ignored. New forms: .only.each, .skip.each, test.concurrent.only, test.describe.serial.only, fit.each and xdescribe.each. A Jest test.todo("title") placeholder is not a skip. P3: - removed-assertion cancels identical lines across the file's hunks, so a moved test is not a removal. - test-ignore-added compares removed and added items as normalized sets, so reformatting and requoting do not fire. In vite config and package.json, `exclude` counts only under a test or coverage key. - btrain's CANDIDATE_REASON_LABELS, KNOWN_DRIFTS and FORMAL_RUNS are no longer defaults. [review_code] ignore_list_keys and run_count_keys supply them. - The coverage and fail_under checks apply to test config and command files only; retries count only in runner settings. vitest `retry` is added. - File contents come from one git cat-file --batch per side, and only files that name an applicable key are fetched. - Paths come from the ---/+++ headers (git's trailing tab is stripped), so a path containing " b/" parses correctly. Co-Authored-By: Claude Opus 5.5 * Fix round-2 review P3s in the weakened-test review rules 1. Masking drift. The masker started a regex literal after `<` and after a postfix `++` or `--`, so `render(

)` and `i++ / 2` swallowed the next backtick and flipped every later template. A focus inside a later fixture then read as hard, and a real `it.only` below it was missed. Neither position starts a regex now. 2. Non-literal callbacks. it / test / describe (and fdescribe, xdescribe, xtest) now count with a string title plus any second argument, so it.only("x", runCase), it.only("x", withDb(...)) and TypeScript-typed callbacks are caught. context, suite, specify, fit and xit still need a literal callback. A Playwright test.skip(condition, "reason") still does not count. 3. CRLF checkouts. A file on disk is matched against the diff with each line's trailing \r ignored, so a CRLF worktree against git's LF diff keeps whole-file masking and focused-test stays hard. 4. Packed lines. Each side is now joined once per run (the whole file or one hunk). All call arguments and options objects are found in two linear passes per run instead of a 2,000-character join and parse per match. 200 lines packed with `test(` went from about 8 s to under 0.3 s. The import scan behind the fit / xit check uses bounded patterns, so it stays linear. Also: the removed-item pools and removed-append lookups are counted maps. Co-Authored-By: Claude Opus 5.5 * Fix round-3 review findings in the weakened-test review rules P2: argHeads took its 300-character window before trimming blanks, and blanked comments count as blanks. An argument after a long comment (about four lines) got an empty head and dropped out, so a focused test with a comment block between its title and callback passed the gate. The window now starts at the first non-blank character. This stays linear, because the leading blank runs of different spans never overlap. P3: - Options objects are read in full, so an `only:` after six other properties counts again. Call arguments are capped at three (title, options, callback), since no framework takes a test body later. - `{expr} />` no longer starts a regex, in all three JSX variants. - The `<` guard, dead code before, now fires only when `<` sits against the slash (`

`). `<` is back in the regex-start class, so `a < /\`/` and `a << /\`/` start a regex again. All 19 leads in the reviewer's case list now give the right verdict. - A title must be one complete string or template literal, so test.skip("webkit" === browserName, "flaky") is not a skip. The check is linear. - New tests kill every surviving mutant: - the TypeScript return type on a look-alike's callback; - fdescribe, xdescribe and xtest as title receivers; - the last span of an unclosed call in hunk mode; - the head cap (300-line timing test: 93 ms vs 15.6 s without it); - the argument cap in both directions; - takeFresh using items up; - bounded imports (100 import lines: 54 ms vs 10.7 s); - lineAtOffset, via an unindented option property. A bounded mutation run kills all 22 mutants. Co-Authored-By: Claude Opus 5.5 * Count a removed matcher line under a surviving expect as a removed assertion Removing only `.toBe(1)` from `expect(value)\n .toBe(1)` left an expect with no matcher, but removed-assertion counted only lines that contain assert/expect, so the scan stayed clean. A chained matcher line now counts when the assertion statement it belongs to starts on a context line. Joining a chain onto its expect line, or splitting it off, still nets out because that expect line is itself removed and added. Co-Authored-By: Claude Opus 5.5 * Flag node, jest and pytest test-filter flags in test commands test-ignore-added knew only --deselect and --test-skip-pattern. It now also flags --test-name-pattern, --test-only, jest's --testPathIgnorePatterns and --testNamePattern, and pytest's --ignore, --ignore-glob and -k. The pytest names are shared with other tools (flake8 --ignore, ruff --ignore, curl -k), so they count only after `pytest` on the line, after a `pytest ... \` continuation line, in pytest.ini, or inside an addopts value found by walking up at most 20 lines of the hunk. A quoted value such as -k "not network" is kept whole so a changed expression is not cancelled by the old one. --test-only is a warn, like every other flag in this rule. With no `only` markers Node runs zero tests under it (checked on Node 25), so it is a real filter, but a command line is not masked to code the way a test file is, and a new `.only` marker that it would act on is already a hard focused-test finding in the test file itself. Co-Authored-By: Claude Opus 5.5 * Treat JS files under a test/ directory as test files Node's test runner and mocha collect every JS file under test/ by default, so test/smoke.js is a test even without .test in its name. Deleting it, removing its assertions or adding it.only to it used to give a clean scan. JS and TS files anywhere under a test/ directory now count, except inside helper, fixture, mock, support, data or snapshot folders, so a fixture that holds focus markers as scanner input does not become a hard focused-test finding. The rename check keeps flagging a file that loses its test name while staying under test/ (foo.test.mjs -> test/foo-helper.mjs), since a runner that collects by name stops running it. Co-Authored-By: Claude Opus 5.5 * Cancel focus and skip markers that move to another hunk of the file A committed it.only moved far enough in the same file lands in two hunks: a removal in one, an addition in the other. The marker check compared an addition only with removals in its own hunk, so the move read as a new focused test and, with whole files loaded, a hard violation. Removed markers now cancel added ones within the hunk first and then across the whole file, count by count, as removed-assertion already does. A second copy of a moved marker is still reported. Because cancellation can now draw on any hunk, focused-test is hard only when every removed line of the file was masked exactly, not just the removed lines of its own hunk. Co-Authored-By: Claude Opus 5.5 * Keep hunk line numbers right when git drops blank-line prefixes With diff.suppressBlankEmpty=true, git writes a blank context line as an empty line with no leading space. The parser skipped it, so every later line in the hunk got a lower number, the whole-file text check failed, and a real focused-test dropped from hard to warn on the wrong line. reviewCode now runs git diff with -c diff.suppressBlankEmpty=false, and the parser counts an empty line as context while the hunk header still promises lines on both sides, for diffs that come from elsewhere. Co-Authored-By: Claude Opus 5.5 * Flag t.skip(), pytest -m, and cancel filter flags moved across hunks Fixes the gh-codex findings on d0147e6, test first: - skipped-test: an unconditional `t.skip()` (node:test) or `this.skip()` (mocha) that starts a statement in a test body now warns. One inside an if, else, switch or catch block, or after a same-line condition, stays clean. The block is found by walking back over balanced braces in the masked run; a multi-line if whose `(` lies before the hunk counts as conditional. - test-ignore-added: pytest `-m MARKEXPR` counts as a filter flag in a pytest context. The pytest command on a line now ends at `|`, `;` or `&`, so `pytest && git commit -m` is not read as `pytest -m`. - test-ignore-added: a removed filter flag cancels an added one anywhere in the same file, count by count, so a script that only moved to another hunk is not a new flag. A second copy is still reported. - Pin the two mutation survivors: an allow marker on a command flag line (or the line above) suppresses the warning, and a `--deselect` injected from conftest.py (a test config that is not a command file) is scanned. Co-Authored-By: Claude Fable 5.1 * Report flags and markers moved onto another command, and literal-only assertions Fixes the two P2s from the stand-in review of d0e553d, test first: - test-ignore-added: a filter flag cancels a removed one only under the same command key (package.json script, ini/toml section + key, YAML key). A flag cut from a niche script and pasted into the main one is reported, naming the script it came from. Keyless shell and Makefile lines pair within their hunk only. Removed-side flags now use the same pytest context as added ones, so a removed flake8 --ignore does not cancel an added pytest --ignore. - focused-test and skipped-test: a marker's identity carries its test title, so `.only` moved from a one-case describe onto the big one is new while a marker that moves with its test (re-indented, re-quoted) cancels. - removed-assertion: an assertion that checks only literals (assert.ok(true), assert.equal(1, 1), expect(true).toBe(true), assert True) counts as no assertion on either side, so swapping one in for a real check reports a removed assertion and the detail says why. Cheap P3s from the same review: a braceless `if (cond)` / `else` line above a t.skip() counts as conditional; PYTEST_ADDOPTS env entries and $(PYTEST) make variables count as pytest context. Co-Authored-By: Claude Fable 5.1 --------- Co-authored-by: Claude Opus 5.5

Sunbelt Computer Software

PL/B Language Development and Support