The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions in SKILL.md and references/measurement-matrix.md direct the agent to run shell commands, specifically cargo run and repository-resident scripts like benchmarks/run_paired_performance.sh to measure system performance.
[DYNAMIC_EXECUTION]: The skill workflow involves the agent implementing optimization candidates by modifying or writing source code and then executing that code via the project's build and benchmark tools (e.g., cargo run --release). This represents a potential vector for executing malicious code if the implementation logic is influenced by an attacker.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze arbitrary repository source code, which creates a surface for indirect prompt injection.
Ingestion points: The agent reads repository source code, version control history, and project configuration files as defined in the Define the actual outcome and Measure total work sections of SKILL.md.
Boundary markers: The instructions do not define specific delimiters or instructions to ignore embedded malicious content within the analyzed files.
Capability inventory: The agent has the capability to write and modify source files within the repository and execute shell commands via cargo and bash.
Sanitization: The guidelines lack instructions for sanitizing repository content or verifying the integrity of repository-owned benchmarking scripts before they are executed.

optimize-accuracy-first