The Agent Skills Directory

[SAFE]: The skill follows security best practices by implementing a 'HARD-GATE' that prohibits editing local files or starting the requested work, ensuring it remains a non-destructive planning tool. It relies on standard development tools (npm, git, grep) for environment probing without introducing external dependencies or remote execution patterns.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted repository data, such as package scripts and code comments, to identify existing benchmarks and contracts. While this is an ingestion surface for potentially malicious data, the skill mitigates this by enforcing a strictly structured output format ('command' → 'expected output') and by requiring the agent to prove every check runs in a read-only state before finalization. The clear boundary between reading the territory and performing the work prevents embedded instructions from being executed during the skill's own turn.
[COMMAND_EXECUTION]: The skill uses local shell commands like git diff, grep, and npm run bench to verify the state of the repository. These operations are restricted to non-modifying 'probes' and do not involve dynamic command construction from unvalidated user input, as the agent is instructed to use existing repository scripts.

forge-goal