The Agent Skills Directory

[COMMAND_EXECUTION]: The SKILL.md file contains a dynamic context execution block (!) that runs a shell command with direct interpolation of the $ARGUMENTS variable. This allows for arbitrary command injection at skill load time if a user provides a maliciously crafted argument.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface. Untrusted data enters the context via git diff output processed by build-manifest.sh and precompute.sh. The skill lacks boundary markers or instructions to treat this data as untrusted. Its capability inventory includes Bash for subprocess execution, and the Read/Write tools for file system interaction. No sanitization or validation of the external content is performed before interpolation into prompts for the reasoning agents in Phase 2.
[COMMAND_EXECUTION]: The orchestration script precompute.sh uses eval to ingest output from other scripts (detect-mode.sh and detect-features.sh). While these are local scripts, the use of eval on dynamically generated strings is an unsafe coding practice that could be exploited if those scripts' outputs were influenced by external data.

merge-checks