The Agent Skills Directory

[COMMAND_EXECUTION] (MEDIUM): The script scripts/check-file-size.sh executes shell commands using a user-provided file path argument ($1). While variables are quoted, the lack of input sanitization at the script level means that if the AI agent is tricked into passing a malicious path or a path containing shell metacharacters, it could lead to unintended command execution.
Evidence (scripts/check-file-size.sh): FILE=$1, LINE_COUNT=$(wc -l < "$FILE"), and FILE_TYPE=$(echo "$FILE" | sed 's/.*\.//').
[COMMAND_EXECUTION] (MEDIUM): The script scripts/validate-increment.sh executes arbitrary code defined in the local project's package.json via npm run. While standard for development, this presents a risk if the agent is working within an untrusted repository where the package.json scripts have been tampered with.
Evidence (scripts/validate-increment.sh): Calls to npm run lint, npm run type-check, and npm test.
[PROMPT_INJECTION] (MEDIUM): The template resources/gap-analysis-template.md suggests shell command patterns like grep "[keyword]". If the agent populates these placeholders with untrusted data from a user (Indirect Prompt Injection), it could result in command injection on the agent's host system.
Ingestion points: Template placeholders like [keyword] and [file-path] in resources/gap-analysis-template.md and resources/increment-checklist-template.md.
Boundary markers: None present in templates to distinguish between instructions and data.
Capability inventory: Subprocess execution via wc, sed, grep, and npm.
Sanitization: None detected. The scripts and templates do not include logic to escape or validate user-provided strings before they are used in a shell context.

feature-orchestrator