The Agent Skills Directory

[PROMPT_INJECTION]: The instructions establish a rigorous workflow for the agent to follow, requiring re-reading of documentation and mandatory planning steps. These are designed to ensure correctness in a complex refactoring task and do not contain patterns aimed at bypassing agent safety filters or ignoring system instructions.
[COMMAND_EXECUTION]: The skill guides the agent to execute shell commands for running project tests and compiler-generated crash reproduction scripts. These actions are governed by explicit instructions to confirm commands with the user and perform operations only within the project context for debugging and verification.
[DATA_EXPOSURE]: During compiler crash handling, the skill directs the agent to gather crash logs from '~/Library/Logs/DiagnosticReports/'. This access is strictly scoped to the purpose of generating bug reports for the Apple Clang toolchain and includes a zip-and-review process with the user.
[INDIRECT_PROMPT_INJECTION]: The skill analyzes user-provided C source code. While the code is an untrusted ingestion point, the risk of the agent obeying instructions embedded in code comments is mitigated by the skill's mandatory multi-step 'plan mode' and the requirement for explicit human approval for every file modification and git commit.
Ingestion points: C source files (.c) and header files (.h) provided by the user.
Boundary markers: Not explicitly defined for code comments, but the workflow uses task-based isolation.
Capability inventory: File system writes, shell execution for testing/building, and sub-agent spawning for isolated analysis.
Sanitization: Not explicitly implemented, relying instead on human-in-the-loop validation.

c-bounds-safety