Patches failing guardrail or test cases one root-cause class at a time under a regression gate.
npx skills add https://github.com/xlinh2301/EditCTC --skill blue-team-xlinh2301
Please help me install this Agent Skill. Skill: blue-team Source: https://github.com/xlinh2301/EditCTC/tree/main/.agents/skills/blue-team Command: npx skills add https://github.com/xlinh2301/EditCTC --skill blue-team-xlinh2301
tools/ingest.py normalizes red-team failures.jsonl, JUnit XML reports, or pytest node-id lists into one failure catalogue grouped by root-cause class. - Gated fix loop: each iteration patches one class, then tools/verify.py checks that the class closed and zero regressions appeared (oracle mode uses a frozen ground-truth oracle plus a benign holdout; tests mode re-runs the suite and reads the JUnit report); bad patches are reverted. - Pull-request handoff: kept fixes become one commit per class on a PR branch opened with gh pr create, degrading to git format-patch plus a PR_BODY.md when git or gh is unavailable. - Use Case: A red-team run found 5 bypass classes in your prompt guardrail. Point this loop at the failures file and the guardrail source; it normalizes inputs, broadens the policy, tightens over-broad rules, and opens a PR with each fix as a separate commit. ## Quick Start Ask the AI to run the blue-team loop on your failing tests or red-team failures file, naming the target source files it may edit, and let it iterate until every failure class is closed or the budget runs out.High-intent search queries and answers about installing and using this skill.
Run your suite with pytest --junitxml to produce a JUnit report, convert it to a failure catalogue with tools/ingest.py --from junit, then run the loop in tests mode. Each iteration patches one failure class and re-runs the suite via tools/verify.py, keeping the fix only if no previously-passing test breaks.
Convert the red-team failures.jsonl into a catalogue with tools/ingest.py --from red-team, then run the loop in oracle mode with a frozen ground-truth oracle command and a benign holdout file. A case is closed when the target agrees with the oracle, and any newly broken holdout input counts as a regression that reverts the patch.
Oracle mode fixes a guardrail or filter by comparing its verdicts against a frozen ground-truth oracle on each catalogue input, using a holdout set as the regression guard. Tests mode fixes code by re-running a test command that writes JUnit XML, treating any newly failing previously-passing test as a regression.
Git and the gh CLI are needed only for the default branches strategy and pull-request handoff. In snapshots mode it works on non-git trees, and without gh it degrades to writing git format-patch output plus a PR_BODY.md into the sandbox for manual PR creation.
Do not use it to discover new failures — that is the red-team skill's job — or to edit the oracle, tests, or holdout that define ground truth, since those are frozen. It also stops rather than thrashes when a class resists the configured patience of attempts, marking it a residual.
apacheRuns, reviews, and triages Gradle test failures across grails-core modules.
gaasherApply patches to fix code failures from CI/CD test reports and create pull requests.
AI-Driven-Highspeed-DevelopmentOrchestrate end-to-end testing workflows across specialist agents.
irrit-usGenerates and reviews property-based tests across multiple languages and smart contracts.