The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes untrusted external data which creates an indirect prompt injection surface.
Ingestion points: The skill reads project source files, READMEs, and conversation logs located in .agent/memory/conversation-*.md.
Boundary markers: No boundary markers or 'ignore' instructions are used when interpolating ingested data into the agent's context.
Capability inventory: The skill possesses the Write tool to create or modify files like AGENTS.md and Makefile, and the Bash tool to execute git commands.
Sanitization: There is no evidence of sanitization, escaping, or validation of the content extracted from conversation logs or source code before it is used to generate documentation or build scripts.
[COMMAND_EXECUTION]: The skill performs dynamic generation of build scripts which could execute unintended commands.
In Research Mode, the agent is instructed to 'Bootstrap' a Makefile by inferring build and test targets from the codebase.
If a malicious actor places deceptive instructions in a source file or conversation log, the agent might be tricked into including malicious shell commands as targets within the generated Makefile.

distill-context