The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from git repositories, including commit messages and code diffs, which creates a surface for indirect prompt injection attacks where malicious repository content could attempt to influence agent behavior.
Ingestion points: Commit logs, blame metadata, and diff hunks read via git log, git show, and git blame commands.
Boundary markers: The skill does not explicitly define delimiters for external data but relies on natural language instructions to guide the agent.
Capability inventory: Execution of shell commands via git and the git bisect run command.
Sanitization: No technical sanitization of git metadata is mentioned prior to processing.
[DYNAMIC_EXECUTION]: The skill generates and executes local automation scripts to drive the git bisect process.
Evidence: The reference/framework-templates.md file provides a bash script template (trail_bisect.sh) for automated regression testing.
Workflow: The skill encourages the generation of test scripts that are executed dynamically during the binary search process to verify commit states.
[COMMAND_EXECUTION]: The skill relies on extensive execution of shell commands for repository interaction and performance profiling.
Evidence: Use of standard git tools (log, blame, show, bisect) and profilers such as py-spy and async-profiler.
Safety boundaries: The core instructions include explicit 'Ask First' requirements for checking out old commits and running user-provided test commands to mitigate risks.

trail