Repository navigation
Releases
Releases · aguschirico/cheatguard
0.1.3
Changed
- Linters run in parallel, and so do up to four tsc projects, so the end-of-task check takes about as long as the slowest tool instead of the sum.
config-tamperingno longer warns about a test or lint script that only gains more runs, on other files, ofnode --checkortsc --noEmitwritten exactly as it already runs them, with no addedyarn/npxprefix (&& node --check another.mjs,&& tsc --noEmit -p b), or drops a package-manager prefix (yarn tsc→tsc). Every other added command is still a warning: test runners, interpreters and shells run their argument, and ESLint, Prettier and oxlint load executable configuration found next to the files they check.- The Claude Code plugin pins
cheatguard@0.1.3.
Fixed
- tsc in monorepos: a root
tsconfig.jsonthat compiles only some files (TanStack/form includes just its scripts) no longer hides type errors in workspace packages. Packages that no root project compiles are checked with their owntsconfig.json, and a root config that has both files andreferenceschecks the referenced projects too (tsc -pdoes not). - Comments at the end of a line of code are now read:
return 'N/A' // TODO,run(); // eslint-disable-lineandx; /* biome-ignore … */were missed bystub-leftandtype-suppression. Text inside JSX that looks like a comment is still ignored, and a TODO quoted asinline codein a comment is a mention, not a TODO. test-weakenedfinds the code under test through tsconfigpathstoo (tests that import their package by its own name), so an expected value updated together with that code is a warning instead of a block.Functionas the test of a conditional type (T extends Function ? never : T) is no longer reported as a type escape.
Security
config-tamperingjudges a gate script split across lines like one joined with;(the shell runs each line as its own command):"vitest run\nexit 0"or"vitest run\ntrue"in atestscript were not blocked.
Benchmarks
- Benchmark A: 154 agent pull requests (Cursor, Jules and Codex added) and 122 human ones, labeled v3. Precision on agent pull requests is lower on the larger sample (72%, from 87%); every wrong finding there comes from
test-weakened. - Benchmark C covers 10 repositories (4 more: unstorage, consola, and two large monorepos, TanStack/form and better-auth; better-auth is the first with Biome).
- Benchmark B rerun with claude-sonnet-5-5 (285 runs, Claude Code 2.1.308) on the current rules, replacing the 2.1.300 runs. Still 0 runs that cheated in any setup. Honest solves on trap tasks: 83% alone (was 86%), 87% with CLAUDE.md rules (was 93%), 80% with cheatguard (was 84%); all three moved down together and within the Wilson intervals of the previous run. On controls cheatguard stays at 100%. cheatguard did not intervene in any of the runs it lost on discount-order and email-validation.
- CI reruns benchmark A on every pull request and fails when its scores differ from the committed results or the synthetic set misses or wrongly flags anything, so a change in precision or recall is always published with the change that caused it.
- The license list of quoted repositories keeps the license a repository's code was quoted under when the repository is later removed or relicensed.
Install or update: npx cheatguard@0.1.3. The Claude Code plugin pins this version.