The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted test output (stdout/stderr) and interpolates it into an LLM prompt for failure triage. An attacker who can influence test results or console logs could attempt to manipulate the root cause analysis or suggested fixes.
Ingestion points: output.log containing captured test output from the Bash tool execution.
Boundary markers: The LLM prompt template uses markdown headers (## Parser Tags, ## Compressed Output) as delimiters but lacks robust escaping for the interpolated content.
Capability inventory: The agent has access to Bash, Write, and Agent tools, allowing it to execute system commands and manage background agents.
Sanitization: The skill includes a mandatory secret redaction pipeline to remove sensitive data like API keys and private keys before LLM processing.
[COMMAND_EXECUTION]: The skill uses a fixer catalog to automate environment repairs. This includes executing shell commands for retry, clear_cache, reinstall_deps, and port_cleanup using the Bash tool.
[EXTERNAL_DOWNLOADS]: One of the core fixers, reinstall_deps, executes npm install, which downloads software packages from external registries to the local environment.

test-deep