agent-qa-evidence
Reviews supplied agent-qa run evidence and distinguishes application results from selection checks, setup failures, and incomplete records.
darkrishabh · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add darkrishabh/agent-skills-eval --skill agent-qa-evidence技能说明
name: agent-qa-evidence description: Reviews supplied agent-qa run evidence and distinguishes application results from selection checks, setup failures, and incomplete records. license: MIT compatibility: Works with text-capable chat models and supplied evidence; no agent-qa installation or tool execution is required. metadata: owner: vostride domain: software-testing
Review agent-qa evidence
agent-qa combines natural-language web/mobile tests with advanced memidex memory and self-improvement. Review the supplied command observations and run records without executing tests or changing the application.
Decide what the evidence establishes
- Identify the requested test, target, revision, and expected observable outcome. Bind the evidence to that scope; do not inherit another run's result.
- Separate configuration/YAML validation, test discovery, application execution, and memory inspection. Successful validation and
--dry-runselection do not prove application acceptance or a JUnit artifact. - Inspect the selected run's completed verdict and relevant observations. An empty failed-run list can omit incomplete or healed runs. Missing verdicts and
evidence-incompleteknowledge remain unresolved. - Interpret exit status alongside the command and error:
0can mean passed or healed;1can mean a failed/timed-out test or a missing required file;2means the run could not start;130means cancellation. A setup error is not proof of an application defect. - For a healed run, preserve the distinction from a clean pass and explain the recorded correction. Verify that the original expected outcome still holds. Do not weaken assertions or erase memory to create a passing result.
- Treat memidex observations as diagnostic context, not an independent acceptance oracle. Missing knowledge evidence does not establish what was supplied or learned.
Use these report labels for this example; they are not additional agent-qa CLI verdicts:
PASSED: a completed passing run and its observed assertion establish the requested outcome for the stated scope.HEALED: a completed healed run retains the original assertion and has evidence of the correction and resulting outcome.FAILED: observed application behavior contradicts the expected outcome in a completed run.UNVERIFIED: selection-only output, unavailable execution, incomplete evidence, mismatched scope, or missing required observations prevent a supported outcome.
Return the label, evidence supporting it, and the specific remaining check if any. Retain the run ID and relevant command/error when supplied. Distinguish provided observations from inference; do not invent execution, screenshots, reports, or completion. Preserve sensitive-data redaction.
The command semantics above follow agent-qa 0.2.3. Consult the CLI reference and memory inspection guide when interpreting a different version. The evaluator supplies files as text; this skill does not fetch those links.