The Agent Skills Directory

[COMMAND_EXECUTION]: The skill performs shell-based repository management using git commands and executes project-specific scripts found in the root directory.
[REMOTE_CODE_EXECUTION]: In Phase 4, the skill clones the repository and runs its bootstrap command (e.g., install.sh, make setup). This executes code defined within the repository being analyzed.
[DATA_EXFILTRATION]: The skill can push repository content to remote servers. It mitigates risk by scanning for secrets in tracked files and requiring user confirmation before the network operation occurs.
[PROMPT_INJECTION]: The skill creates an indirect prompt injection surface in Phase 3b by reading repository files and passing them to a fresh subagent for takeover verification.
Ingestion points: CLAUDE.md, README.md, AGENTS.md, and DECISIONS.md (discovered in Step 0).
Boundary markers: Absent. The subagent receives raw repository content as the primary context for the takeover probe.
Capability inventory: Shell command execution (bootstrap), git operations (rm, push, tag), and subagent spawning (Task tool / codex exec).
Sanitization: Absent. Project documentation is processed without filtering or escaping before being passed to the subagent.

iteration-close