The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input from git diffs and commit logs. Although it enforces a 'Clean-room' policy (Iron Law 1) and a 'Trace' requirement (Step 2) to ensure messages are strictly backed by code hunks, malicious instructions hidden in code comments within the diff represent a potential attack surface.
Ingestion points: Git diffs are read in Step 1 of SKILL.md; commit message history is read in Step 4.
Boundary markers: The skill requires a 'locked diff hash' and implements a 'shared rejection check' in REFERENCE.md to filter out unproved or external text.
Capability inventory: Executes shell commands (git commit, git hash-object) and spawns Python/Node.js REPL environments.
Sanitization: Implements a strict mapping of setiap commit line back to original hunks, deleting any text that cannot be traced to the diff.
[DYNAMIC_EXECUTION]: The skill instructs the agent to open a Python or Node.js REPL to perform 'Trailer hygiene' (Step 4 and REFERENCE.md). This involves the agent generating and executing script code at runtime to inspect and modify the HEAD commit message. This pattern of using an interactive interpreter to bypass hook-added metadata is a form of dynamic code generation.
[COMMAND_EXECUTION]: The skill relies on shell command execution for its primary functionality, including git commit, git diff, and git hash-object. It specifies security best practices for variable quoting in Step 3 to prevent shell metacharacter injection.

commit