AgentSkillsCN

Ralph Loop

拉尔夫循环

SKILL.md

Ralph Loop

Regression eval workflow for MLAIX. Run the Ralph Loop to verify code changes have not regressed UI, concurrency, build, appearance, platform, or integration behavior.

When to Use

  • •Proactively after code changes to catch regressions before they ship
  • •Before releases or PR merges
  • •When verifying build, test, or concurrency stability

Quick Run

Invoke the ralph-loop subagent (.cursor/agents/ralph-loop.md) to run evals interactively.

Manifest

Eval prompts live in docs/evals/ralph/prompts.json. Each entry has:

  • •id, category, severity, prompt, expected_outcome
  • •command — the exact shell command to execute
  • •related_tests — Swift test names for reference

Execution

  1. •Read docs/evals/ralph/prompts.json
  2. •For each prompt, run its command field verbatim
  3. •Exit code 0 = pass, non-zero = fail
  4. •Record results in docs/evals/ralph/runs/<timestamp>/
  5. •Update docs/evals/ralph/summary.md and results.scored.json

Verdict

  • •PASS: All critical and high severity checks pass
  • •FAIL: Any critical or high severity check fails — block release

Prerequisites

  • •jq (brew install jq)
  • •Swift toolchain with SwiftPM