The Agent Skills Directory

Dynamic Code Execution: The skill instructs the agent to generate and execute custom scripts (e.g., using Python or Shell) at runtime during its 'Mode A' phase to crawl the repository and generate the initial plan. While this provides flexibility in repository analysis, it involves the dynamic creation of executable logic based on the directory structure.
Processing Untrusted Metadata: The skill utilizes mantis-summary.md files located within the target codebase to understand directory contexts. This presents a potential surface for indirect prompt injection.
Ingestion points: Content from mantis-summary.md files and codebase directory structures (filenames/paths).
Boundary markers: No explicit markers are defined to distinguish external summary content from system instructions.
Capability inventory: The skill can execute shell commands (VCS tools like git and hg), run local scripts, and write to the workspace/ directory.
Sanitization: No specific sanitization or validation mechanisms are described for the data read from summary files.
Command Execution: The skill relies on various system tools to perform repository analysis, including find, git, hg, and python3. These tools are invoked to track file changes, perform reverse blames for line re-anchoring, and query structural indices, which are standard operations for a defensive security review tool.

mantis-plan