The Agent Skills Directory

[SAFE]: The skill implements a robust security model that treats project files and compiler output as untrusted evidence. It explicitly defines that raw filenames and compiler messages should not be echoed back to the user, mitigating risks where repository content might attempt to manipulate agent behavior.
[COMMAND_EXECUTION]: Helper scripts (inspect_typescript.py, run_typecheck.py, trace_perf.py) use a localized binary resolution mechanism defined in local_tools.py. This ensures only project-local tools within the repository boundary are executed, preventing the execution of arbitrary system-wide or potentially malicious binaries outside the project scope.
[SAFE]: The run_bounded_compiler function in inspect_typescript.py implements critical resource constraints, including output size limits, line length limits, and execution timeouts. These protections prevent the agent from hanging or being overwhelmed by malicious or excessively large project data.
[INDIRECT_PROMPT_INJECTION]: The skill mitigates indirect prompt injection risks by summarizing findings into category counts (e.g., production vs test files) and utilizing a stable diagnostic reporting system rather than outputting raw filenames or compiler messages that could contain malicious payloads.
[SAFE]: The included regression test suite (test_helpers.py) explicitly validates that the scripts correctly filter 'hostile' markers (such as instructions to ignore previous rules) found in filenames or compiler output, ensuring they do not leak into the tool's reports or the agent's context.

typescript