The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill possesses an inherent attack surface by design, as it must ingest, analyze, and execute untrusted source code to perform its function.
Ingestion points: The agent reads project source code, benchmark definitions (e.g., Rust criterion, Python pytest, Node.js vitest), and configuration files like codspeed.yml and Cargo.toml (found in SKILL.md).
Boundary markers: There are no explicit instructions or delimiters defined to distinguish code logic from potential malicious instructions embedded in comments or strings within the files being optimized.
Capability inventory: The skill uses extensive shell execution capabilities via the codspeed CLI, language-specific build tools (cargo, go), and test runners (pytest, npx vitest) as described throughout the optimization loop.
Sanitization: The instructions do not specify any sanitization or validation of external code content before the agent attempts to interpret or suggest modifications to it.
[COMMAND_EXECUTION]: The skill relies on executing various shell commands to build, benchmark, and compare code performance.
Evidence: The instructions explicitly direct the agent to use codspeed run, codspeed exec, cargo codspeed build, and npx vitest bench to establish baselines and measure improvements. While these are necessary for the skill's purpose, they provide a broad execution environment that could be leveraged if the agent's logic is subverted by malicious code inputs.

codspeed-optimize