The Agent Skills Directory

[COMMAND_EXECUTION]: The skill performs extensive shell command execution using standard development tools such as git and the GitHub CLI (gh). It executes commands for branch management, diffing, and merging, and explicitly invokes a coding-agent CLI for multi-file implementations. Additionally, it executes the target repository's 'quality gates', which may involve arbitrary shell commands defined within the project's own build or test configuration.
[REMOTE_CODE_EXECUTION]: The skill is designed to fetch and execute scripts from a target repository during the 'Act' phase. By running 'the target's own quality gates', the agent executes local code that is determined by the content of the repository it is managing. While this is necessary for the skill's purpose, it represents a potential execution path for malicious code if the target repository is compromised.
[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests and processes untrusted data from repositories to decide on and implement code improvements.
Ingestion points: The agent reads current repository state, in-flight work, commit diffs, and human/bot feedback (review comments and issue comments) as detailed in SKILL.md and references/loop-cycle.md.
Boundary markers: There are no explicit boundary markers or instructions to ignore embedded directives when repository content or feedback is passed to the coding-agent CLI.
Capability inventory: The skill has the capability to write to the file system, perform network operations via git/GitHub, and execute shell commands via quality gates.
Sanitization: The instructions do not define any sanitization, filtering, or validation steps for the data ingested from the repository before it influences the agent's decision-making process.

autonomous-improvement-loop