The Agent Skills Directory

[PROMPT_INJECTION]: The skill ingests untrusted data from user arguments and the local codebase during the trace and interview phases. It mitigates the risk of indirect prompt injection by using explicit <trace-context> boundary markers and instructing the agent to treat injected content as data rather than instructions. The skill's capabilities, including file writing and sub-agent invocation, are scoped to the project's spec directories.\n- [COMMAND_EXECUTION]: The skill invokes a local validation script (scripts/validate_deep_dive_artifacts.py) to verify artifact structure. This script is part of the skill's own source and implements path validation to ensure it only interacts with authorized spec paths within the .omc, .omx, and .agents directories.\n- [DATA_EXFILTRATION]: The skill accesses local configuration files (e.g., settings.json, config.*) to resolve runtime-specific settings for Claude Code, Codex, and OMA adapters. This data is used locally to facilitate cross-runtime portability and is not transmitted to external domains.

deep-dive