The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary autonomous loop (Mode 3) is guided by instructions read from a local file (program.md). This creates an attack surface where instructions embedded in project files or the program file itself could influence the agent's behavior during execution.
Ingestion points: Reads program.md and iterates over source files in src/ (as defined in files.txt).
Boundary markers: No explicit delimiters or safety headers are defined to distinguish instructions from data.
Capability inventory: The agent has the ability to modify source code, execute local bash scripts (measure.sh), and perform git operations (commit, checkout).
Sanitization: No sanitization is performed on the contents of the files before processing.
[COMMAND_EXECUTION]: The skill uses bash to run a local evaluation harness (measure.sh) and utilizes standard CLI tools like grep, wc, and git to measure codebase state and manage changes.
[DYNAMIC_EXECUTION]: The skill scaffolds and subsequently executes shell scripts (measure.sh) at runtime. While these scripts are generated based on a local template, they represent dynamic execution of generated content.

autoresearch