The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted external data, including network traces, payloads, event logs, and user-provided artifacts during the 'Replay a captured trace' and 'Build a feedback loop' phases. This creates a surface for indirect prompt injection if malicious instructions are embedded in the debug data.
Ingestion points: The agent reads GLOSSARY.md, ADRs, and captured artifacts (HAR files, log dumps, core dumps) described in SKILL.md (Phase 1).
Boundary markers: The skill partially mitigates risks by requiring the agent to 'Redact every secret first' and 'quote only the lines that carry the signal', which serves as a manual filtering step.
Capability inventory: The skill utilizes curl, git bisect, Playwright/Puppeteer, and local shell script execution (scripts/hitl-loop.template.sh) to verify bugs.
Sanitization: The instructions mandate redacting credentials and secrets before displaying any artifacts or command outputs to the user.
[COMMAND_EXECUTION]: The skill facilitates the execution of various developer tools and diagnostic commands.
Evidence: SKILL.md outlines the use of curl, git bisect, and headless browser scripts. It also provides scripts/hitl-loop.template.sh for human-in-the-loop debugging. These actions are strictly within the intended scope of a debugging utility.

diagnosing-bugs