The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted code provided via the TARGET_CODE input, which creates a vulnerability surface where instructions embedded in code comments could influence the agent's behavior.
Ingestion points: TARGET_CODE is ingested and processed by SKILL.md, strict-baseline-mapper.md, strict-rewrite-strategist.md, and strict-rewrite-implementer.md.
Boundary markers: The instructions do not define strict delimiters or "ignore embedded instructions" directives for the subagents when they process the provided code content.
Capability inventory: The skill possesses file system write capabilities via strict-rewrite-implementer.md and shell command execution capabilities via the user-supplied VALIDATION_COMMAND or discovered project-local checks.
Sanitization: No explicit sanitization or filtering of the input code is performed before it is analyzed or modified by the LLM.
[COMMAND_EXECUTION]: The strict-rewrite-implementer subagent executes shell commands to validate code changes. It prioritizes a user-supplied VALIDATION_COMMAND and can also discover and run local project scripts such as go test, mypy, or npm test. This intended functionality allows the execution of arbitrary commands if the input is manipulated.
[EXTERNAL_DOWNLOADS]: The strict-rewrite-strategist subagent is designed to fetch documentation from external URLs defined in references/external-sources.md. These downloads target trusted sources, including official documentation sites for Python, Go, and TypeScript, as well as official GitHub repositories for tools like Pydantic and Zod.
[DYNAMIC_EXECUTION]: The skill acts as a script generator that modifies project code and executes the resulting changes through various validation suites. This behavior is consistent with its primary purpose of hardening and verifying code strictness.

rewriting-code-strictly