The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill workflow involves processing potentially untrusted data that could contain malicious instructions designed to influence the agent's behavior.
Ingestion points: SKILL.md (Phase 1 & 2) instructs the agent to read and analyze HAR files, log dumps, core dumps, and captured network traces.
Boundary markers: While the skill mandates redacting secrets (credentials), it does not specify markers or instructions to treat data content as untrusted to prevent indirect prompt injection.
Capability inventory: The skill allows the agent to execute shell scripts, run local test suites, and perform network operations via curl.
Sanitization: The primary sanitization mentioned is privacy-focused (redacting auth headers and secrets) rather than security-focused (escaping or filtering embedded instructions).
[DYNAMIC_EXECUTION]: The debugging process requires the agent to generate and run transient code to reproduce and minimize failures.
Evidence: SKILL.md outlines the creation of 'throwaway harnesses', 'bisection harnesses', and scripts for Playwright/Puppeteer to exercise code paths.
[COMMAND_EXECUTION]: The skill utilizes various local system commands to execute the feedback loop and diagnostics.
Evidence: The instructions include the use of curl for HTTP requests, git bisect for regression testing, and the execution of the provided scripts/hitl-loop.template.sh bash script.

diagnosing-bugs