The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest untrusted data in the form of code diffs and planning documents for auditing purposes.
Ingestion points: Files or standard input read by independent_review.sh and merge_link.sh.
Boundary markers: Prompts use clear structural delimiters (--- BEGIN {TYPE} --- and --- END {TYPE} ---) to separate instructions from data.
Capability inventory: Executes shell commands for git, curl, and various model CLIs (codex, agy, ollama).
Sanitization: System prompts include explicit directives to reviewer models to treat artifacts as data, ignore embedded instructions, and report potential injection attempts.
[COMMAND_EXECUTION]: Orchestrates various developer tools and model CLIs using shell scripts and Python sub-processes. All command-line constructions are constrained to legitimate tool operations like git diff or API requests via curl and do not allow for arbitrary shell execution.
[EXTERNAL_DOWNLOADS]: References and provides instructions for installing components from established services including OpenAI (chatgpt.com), Google (antigravity.google), and Ollama (ollama.com). These are treated as well-known and trusted services.
[CREDENTIALS_UNSAFE]: Implements secure handling of API keys for ollama and melious.ai. The skill explicitly unsets environment variables before invoking external model CLIs to prevent credential inheritance and transmits sensitive tokens to curl via standard input to avoid command-line exposure.

independent-review