The Agent Skills Directory

[Indirect Prompt Injection]: The skill processes data from the local environment, such as file names, package metadata, and tool outputs. This creates a surface where instructions could be inadvertently or intentionally embedded in repository data.
Ingestion points: The agent reads outputs from git diff and git status commands, package names from package.json files, and diagnostic logs from pnpm and fluid-build as described in SKILL.md and ci-readiness-check.sh.
Boundary markers: The instructions do not employ specific delimiters or guidance to distinguish between data and potentially executable instructions within these inputs.
Capability inventory: The skill possesses the ability to execute shell commands, run build processes via pnpm, and interact with the file system.
Sanitization: Tool outputs and file contents are analyzed directly without explicit sanitization or filtering of instruction-like patterns.
[Command Execution]: The skill executes various system commands and local scripts to perform its functions.
Evidence: The skill invokes a bundled shell script (ci-readiness-check.sh) and uses package management tools (pnpm) for building, linting, and testing (e.g., pnpm exec fluid-build). These operations are consistent with a developer-focused utility.
[External Downloads]: The skill performs dependency installation using standard package managers.
Evidence: The ci-readiness-check.sh script runs pnpm install --frozen-lockfile if node_modules are missing. This is a routine operation for managing project dependencies.
[Dynamic Execution]: The skill uses runtime evaluation to retrieve project metadata.
Evidence: The ci-readiness-check.sh script executes a node -p command to evaluate a snippet of JavaScript that reads the package name from package.json.

ci-readiness-check