A Screenshot Is Not an Agent Failure Artifact

Описание и применение

Скриншоты не содержат полной доказательной цепочки. Используйте пакет Evidence v2 от AgentInspect, чтобы предоставить проверяемый, редактируемый и защищённый хешами набор доказательств.

Задача

Reviewers rely on screenshots of execution trees, which lack provenance, redaction context, and offline inspectability, leading to incomplete or misleading evidence.

Результат

Using AgentInspect’s Evidence v2 bundle provides a bounded, reviewable evidence artifact that preserves provenance, redaction context, and file integrity, enabling reviewers to inspect failures offline and verify consistency without exposing raw data.

Рабочий процесс

• 1. Run npx agent-inspect bundle <run-id> --dir .agent-inspect --profile share --out ./agent-evidence for a single run or use --session for multiple runs.
• 2. Inspect the generated folder: open evidence.html and summary.md to review the trace and summary.
• 3. Review redaction-report.json and check-results.json to understand what was redacted and any verification issues.
• 4. Run npx agent-inspect bundle verify ./agent-evidence to ensure the bundle’s integrity and that all hashes match the manifest.
• 5. If verification passes, attach the agent-evidence folder to the PR or support ticket; otherwise regenerate or correct the source.
• 6. Apply organizational data‑retention and sharing policies before publishing the bundle.

Доказательства

The source remains private. The reviewer gets a bounded object with enough context to question the result.

Источник

Run the offline verifier: npx agent-inspect bundle verify ./agent-evidence

Источник