The Agent Skills Directory

[DATA_EXPOSURE_&_EXFILTRATION]: The skill implements a robust redaction system to prevent sensitive data exposure. It scans for API keys, tokens, and PII using regex patterns in several scripts (e.g., hooks/extract-state.sh, hooks/opencode.plugin.js, hooks/extract-raw.ps1) and replaces them with a [REDACTED] placeholder before writing to either persistent or working memory files.
[PRIVILEGE_ESCALATION]: The skill operates within standard user permissions. Hook installation for various harnesses (/duck-tape init) focuses on writing local configuration files like .claude/settings.local.json or .github/hooks/duck-tape.json, which are standard practices for these tools and do not involve administrative escalation.
[COMMAND_EXECUTION]: Local command execution is limited to standard development tools such as git (for branch/repo info), jq (for JSON parsing in bash), and standard shell utilities (ls, wc, mkdir). These are used for session state management and do not accept unvalidated user input for arbitrary execution.
[EXTERNAL_DOWNLOADS]: The skill does not download external scripts or packages at runtime. It references its own included hook scripts for session state extraction.
[PROMPT_INJECTION]: Instructions are focused on file schema management and state synchronization. No patterns attempting to bypass LLM safety filters or override system-level constraints were identified.
[DYNAMIC_EXECUTION]: The hooks/opencode.plugin.js is a standard JavaScript plugin for the opencode environment. It uses built-in Node.js/Bun modules (fs, path) for local file operations and does not employ unsafe dynamic execution methods like eval() on untrusted data.

duck-tape