Loop Verifier
Independent checking agent that verifies loop-produced changes and rejects unless evidence is strong.
The skill explicitly requires the verifier to default to REJECT, check that tests actually ran, and escalate to a human if unable to run them. This shows safety orientation, but it doesn't explicitly require least privilege or provide rollback mechanisms. Data-flow transparency and sensitive-data handling are not addressed within the skill itself. Dependency security is not discussed, and source attribution is absent. Deductions for incomplete safety details despite clear intent.
The skill provides clear inputs and output format, but it is not verifiable that the tests actually run because this is a static review. No automated tests or scripts are provided. Key-path reproduction is untested, and failure-feedback on abnormal input is not detailed. Deductions for lack of executable evidence and error-handling specifics.
The skill clearly defines its use case (verification of implementer output in loop engineering) and provides inputs and checklist. However, boundaries (e.g., when NOT to use) are minimally stated, and environment fit (including Chinese-language support and mainland reachability) is not discussed. Core functions may depend on overseas services but not disclosed. Deductions for vague boundaries and unclear environment fit.
The document is well-structured with clear sections and output format, but lacks versioning, changelog, and maintenance responsibility details. License is present in repository but not referenced in skill. No FAQs or troubleshooting. Deductions for incomplete governance information despite readability.
The skill describes the verification process but lacks verifiable evidence that it effectively completes the task. No example outputs or success rates. Static review prevents confirming actual effectiveness. Deductions for lack of effectiveness evidence.
The skill provides no reproducible tests or third-party evidence. Claims of safety and effectiveness cannot be independently verified. Deductions for lack of verifiable evidence in static review.
- The skill's core function (running tests, checking scope) depends on the implementer's environment and may involve executing arbitrary code, posing security risks.
- The skill does not address data-flow transparency or sensitive-data handling; users should be cautious.
- For users in mainland China, the overseas services this skill relies on (e.g., GitHub) may not be directly accessible; consider alternatives.
What does this skill do, and when should you use it?
The Loop Verifier is an independent verification agent within the Loop Engineering repository, acting as the checker in a maker/checker split. It reviews diffs produced by implementer sub-agents, enforcing checks on scope, intent, tests, and absence of cheating. The skill adopts a suspicious default, rejecting changes unless proven otherwise, ensuring only well-verified changes get approved.
Reads the implementer's proposal summary, diff, target issue, and test/lint commands. Checks that file changes match allowed scope, confirms the change addresses the intended problem, runs tests or equivalent and reports pass/fail with outputs, checks for disabled tests or skipped assertions, and recommends human review for medium+ risk. Outputs a verdict of APPROVE, REJECT, or ESCALATE_HUMAN with evidence for tests and scope.
- Advanced users enforcing maker/checker separation in loop engineering workflows after an implementer sub-agent produces changes.
- Automating test and scope verification to prevent unverified code from being merged.
- Running CI or regression tests with a second opinion that doesn't trust the implementer's claims.
- Auditing loop quality by tracking which changes get rejected and why, to improve loop performance.
What are this skill's strengths and limitations?
- Enforces strict verification with a default-to-reject stance.
- Provides clear, structured output for loops and humans.
- Prevents cheating by requiring actual test execution.
- Requires detailed inputs from the implementer.
- Assumes a test environment is set up; not documented.
- Only runs tests based on provided commands; no environment setup.
How do you install this skill?
As part of the Loop Engineering repository, this skill resides in skills/loop-verifier/. For standalone use, copy this folder into your project's .skills directory or a location accessible to your agent.
How do you use this skill?
Invoke the skill with the implementer's summary, diff, target issue, and test command. It will output a verdict. Example prompt: 'Verify this diff, run tests, and confirm they pass before approval.'
How does this skill compare with similar options?
Unlike other skills in the collection that implement or estimate costs, this one focuses solely on verification, complementing the maker/checker pattern.