The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill manages a delivery loop that ingests findings and suggestions from external review bots (such as CodeRabbit, Codex, and Greptile). This interaction constitutes a potential surface for indirect prompt injection if a bot output were to contain malicious instructions.
Ingestion points: Data enters the agent context through GitHub Pull Request status checks and review thread comments fetched via the GitHub CLI and GraphQL API (reference/review-bots.md).
Boundary markers: The skill includes strong instructional boundaries, explicitly telling the agent to treat bot output as "untrusted input" and providing a "Verify, never relay" rule.
Capability inventory: The agent utilizes shell execution capabilities via git (for repository inspection, rebasing, and merging) and gh (for interacting with GitHub metadata and the GraphQL API).
Sanitization: The instructions enforce mandatory verification against the source code using git show and sed to confirm bot claims before any action is taken or relayed to a human.
[EXTERNAL_DOWNLOADS]: The skill documentation suggests an installation process using npx skills@1.5.23 add stealth-factory/skills. This operation downloads and installs additional agent capabilities from the vendor's repository. This is a standard and transparent dependency management pattern for this ecosystem.
[COMMAND_EXECUTION]: The skill heavily relies on executing shell commands for repository management and workflow automation. This includes advanced git operations such as git merge-tree --write-tree for conflict detection and git rebase --onto for managing stacked pull requests. These commands are used for their intended purpose in a developer workflow and are guided by safe-practice documentation (reference/git-traps.md).

claude-software-factory