The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill analyzes repository configuration files (such as justfile, package.json, and cargo.toml) to identify authoritative local commands for benchmarking. This creates a surface for indirect prompt injection where malicious instructions embedded in these files could be interpreted as legitimate commands by the agent.
Ingestion points: Target repository build instructions, script definitions, and metadata within project files.
Boundary markers: The skill defines a strict "Iron Law" requiring the agent to obtain explicit approval for a run plan and follow a specific workflow that stops if red flags (e.g., changes to checks or workloads) are detected.
Capability inventory: The agent can execute arbitrary shell commands through the hyperfine benchmarking tool and process performance data using a Python helper script.
Sanitization: The skill relies on human-in-the-loop validation of the "run plan" and the final "edit plan" before any execution or source code modification occurs.
[REMOTE_CODE_EXECUTION]: The skill references and executes a Python script located at ../ci-optimize/scripts/ci_optimize.py. This script is part of a sibling directory structure and is not included in the skill's own package files, meaning its internal logic is not verified by the current analysis.
Evidence: The workflow explicitly instructs the agent to Resolve symlinks before appending ../ci-optimize/scripts/ci_optimize.py and subsequently execute it using python3 "$CI_OPTIMIZE_HELPER".
[COMMAND_EXECUTION]: The core functionality of the skill involves the execution of shell commands to measure build times. While the skill uses the hyperfine tool for measurement and requires an approved run plan, it still facilitates the execution of project-specific commands that could potentially perform unintended actions if the project environment is compromised.

build-optimize