The Agent Skills Directory

[COMMAND_EXECUTION]: The skill leverages the Bash tool to perform repository operations using git (branching, committing, pushing) and the GitHub CLI gh (creating issues, viewing PRs, creating PRs). These actions are central to its purpose of automating the software development lifecycle.
[PROMPT_INJECTION]: The skill exhibits a surface for indirect prompt injection (Category 8) as it parses and acts upon instructions contained in external files such as SPEC.md, TASKS.md, progress.md, and ACCEPTANCE.md. These files are ingested into the agent's context to guide implementation steps. The skill attempts to mitigate this through a mandatory 'Turn contract' and strict schema validation for handoff records.
[COMMAND_EXECUTION]: A specific security best practice is documented in references/FORGE_BODY.md, where the skill instructs the agent to use the --body-file flag with the gh CLI rather than passing Markdown content inline. This prevents shell injection vulnerabilities and ensures proper rendering of Markdown characters.
[COMMAND_EXECUTION]: The skill defines a 'Unit-loop' that can automatically retry failed commands (up to a budget) and execute a sequence of phases without manual intervention, which increases the autonomy of the agent in the shell environment.

execute-phase