agent-verification — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited agent-verification (Agent Skill) and scored it 100/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 0 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 0 flagged
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
Models are optimists about their own work: they declare "task complete" with tests failing, edge cases unhandled, or the goal only half-met — and when asked to grade themselves, they mark up. Generation and evaluation must be separate. This skill is the check that makes an agent's "done" trustworthy: an independent evaluator (ideally a different model) verifies the claim against external done-conditions and can refuse, sending the work back. It's the brake that keeps [[autonomous-loops]] and [[long-running-agents]] from shipping confident-but-wrong output.
This is verification of an agent's completion claim, distinct from [[review-gate]] (human reviewing a finished diff for quality) and [[test-first]] (writing the tests). It composes both; it does not replace them.
([[long-running-agents]])
Skip when a human has already verified the work against its acceptance criteria, or for trivial changes a typecheck/test already gates automatically.
Not a substitute for [[test-first]] (the tests themselves) or [[review-gate]] (human quality review). This skill checks the claim of completion; those produce and review the work.
Take the written done-conditions (from the spec / plan / [[long-running-agents]] files) and check each one against reality. "I think it's done" is not evidence; a passing check against the stated criteria is. If no done-conditions were written, that's the first failure — define them before claiming completion.
The checker must not be the agent that did the work. Use a fresh sub-agent, ideally a different model, with the done-conditions and the diff — not the maker's narrative. Independence is the whole point; a self-review re-imports the same blind spots.
Execute the tests, run the build, exercise the actual behavior ([[browser-checks]] / [[e2e-testing]] where relevant). Confirm acceptance criteria pass against the running system, not by assertion in prose.
Reject any "pass" achieved by deleting or skipping failing tests, weakening assertions, or hard-coding outputs. A test suite that shrank or got more permissive to go green is a red flag, not a success. Compare test count and coverage to the baseline.
Walk the original task list: is every item actually done, or did the agent stop at the easy ones and declare victory? Unfinished sub-tasks, TODOs left in code, and "out of scope" hand-waving on in-scope work all fail verification.
If verification fails, the checker refuses the completion claim and returns specific, reproducible gaps — not vibes. In a loop, that sends the iteration back to the maker; for a human run, it blocks the merge until the gaps close.
Don't trust a "done" inherited across a context reset or handoff. Re-run the check at restart so a premature completion from an earlier turn can't slip through ([[long-running-agents]]).
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.