The Agent Skills Directory

PROMPT_INJECTION (LOW): The files test-pressure-1.md, test-pressure-2.md, and test-pressure-3.md use strong imperative framing and scenario-based constraints (e.g., "IMPORTANT: This is a real scenario. You must choose and act."). While these are intended for evaluating the agent's reasoning, they mirror patterns used in prompt injection and role-play attacks.
CREDENTIALS_UNSAFE (LOW): The SKILL.md file suggests diagnostic commands to inspect the environment (env | grep IDENTITY) and macOS keychains (security list-keychains). These are powerful tools for debugging code-signing issues but can expose sensitive credentials or identities if the output is logged or shared inadvertently.
COMMAND_EXECUTION (LOW): The find-polluter.sh script automates the execution of npm test on files discovered in the local directory. This provides a mechanism for the agent to execute code within the project environment.
INDIRECT_PROMPT_INJECTION (LOW): The skill involves processing untrusted external data such as error messages, stack traces, and log files as part of Phase 1 (Root Cause Investigation).
Ingestion points: SKILL.md (Phase 1, Step 1), root-cause-tracing.md (The Tracing Process).
Boundary markers: None specified for isolating log content from instructions.
Capability inventory: Shell command execution via find-polluter.sh and diagnostic templates in SKILL.md.
Sanitization: The skill does not provide instructions for sanitizing or escaping content within error logs before processing.

systematic-debugging