Runs flake-detector's scorer over the ingested run history (.flake/history.db in the repository) and returns the most flaky tests first, each with a 0-1 score, run count, the signals behind the score and a plain-language explanation. Use it before retrying a failing test or deciding whether a red build is a regression.

Returns the score of one test with every component that contributed to it (flip rate, rerun recovery, failure-message entropy, confidence interval, correlations with runner and hour) and the explanation flake-detector gives. Test ids look like com.acme.CheckoutTest#appliesCoupon.

Reads the quarantine ledger (.flake/quarantine.yaml) and returns every entry with its reason, owner, expiry and whether it has expired. Expired entries fail the build until they are renewed or removed.

Runs flake-detector's gate over a directory of JUnit XML reports (for example target/surefire-reports). Failures of quarantined or known-flaky tests do not block; any other failure does. Returns the verdict, each failure with the decision taken, and the reasoning.

Runs test-quality-linter over test sources and returns findings with rule id, severity, location, message and fix hint. Catches tautological or constant assertions, tests with no assertion, assertions on stubs, hard-coded sleeps, swallowed exceptions, duplicate bodies and more. Paths are relative to the repository.

Returns the catalogue of rules the linter applies: id, name, default severity and a one-line description. Use tql.explain or the toolkit://tql/rules/{id} resource for the full documentation of one rule.

Returns the documentation of one rule: what it catches, a bad example, a fixed example and its options. Rule ids look like TQL003.

Runs llm-eval-harness over a golden dataset with one prompt version and returns every case with its checks, score, tokens and latency, plus the pass rate. Mode replay (default) serves recorded responses and needs no API key; record calls the model and saves recordings. The result file path is returned for llmeval.drift.

Compares two llm-eval-harness result files and returns the cases whose score dropped by more than the threshold or that flipped from pass to fail. Use it after llmeval.run to see whether a prompt change or model upgrade regressed anything.