The Agent Skills Directory

[COMMAND_EXECUTION]: The skill makes extensive use of the Bash tool across almost all modules (e.g., forge-eng, forge-bugfix, forge-ship) to manage Git repositories, detect environment configurations, and execute custom maintenance scripts. Tasks include creating worktrees, performing merges, and running project-specific build or test commands.
[PRIVILEGE_ESCALATION]: forge-ship/SKILL.md includes procedures for production deployment which may involve high-privilege operations such as ssh root@HOST or modifying server-side environment files. These actions are managed through a 'production write-surface authorization gate' (第 0.5 步) that lists all external write operations for user approval before execution.
[DYNAMIC_EXECUTION]: forge-qa dynamically generates and executes E2E testing scripts using the Playwright framework to validate UI and logic against specifications. Additionally, tools/fupan-workbench/launcher.py uses subprocess.Popen to dynamically launch a local FastAPI application.
[DATA_EXFILTRATION]: The generate_image2.py script sends user-generated prompts to the OpenAI API for image creation. The get-session-id.sh script reads internal platform metadata from ~/.claude/sessions/ to identify and log the active session ID for cross-session coordination.
[EXTERNAL_DOWNLOADS]: The skill fetches accessibility testing libraries from Cloudflare's CDN and retrieves interface guidelines from public repositories on GitHub (e.g., Vercel's official guidelines). These fetches are integrated into the QA and documentation synchronization processes.
[INDIRECT_PROMPT_INJECTION]: The skill's architecture involves processing user-provided documentation (PRDs, specs, and bug backlogs) through various agents and sub-agents. It employs 'adversarial review' patterns in forge-brainstorm, where a sub-agent critiques content, creating a potential surface where instructions embedded in those documents could influence the sub-agent's behavior.

forge