The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from GitHub issues, pull requests, and comments, which presents a potential indirect prompt injection surface.
Ingestion points: Data is ingested via git fetch and GitHub API/GraphQL calls, as detailed in the Queue Gate and Remote Truth sections of SKILL.md.
Boundary markers: The skill mitigates risks through a structured lane_map and an Output Firewall policy that restricts raw command output and logs from polluting the parent agent's context.
Capability inventory: Capabilities include Git command execution, GitHub API interaction, and local filesystem access for logging purposes.
Sanitization: The scripts/run_log_schema.py script implements a robust redaction system using regex patterns to identify and remove credentials (tokens, keys, passwords) from data before it is written to logs.
[COMMAND_EXECUTION]: The skill performs local command execution to manage repository state and perform verifications.
It utilizes standard Git operations (fetch, worktree, branch) and GitHub CLI tools (gh) to synchronize remote state and audit PR readiness.
Safety is maintained by enforcing disjoint writable file scopes for workers and requiring explicit 'intent contracts' before execution.
[SAFE]: The skill demonstrates high security maturity through its design. Features such as the capability_gate, merge_gate, and local logging with restricted permissions (0o600) represent best practices for managing autonomous agent sub-tasks.

threads