skills/sanxzy/skills/implement-code/Gen Agent Trust Hub

implement-code

Pass

Audited by Gen Agent Trust Hub on May 14, 2026

Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill performs several shell-based operations to manage the development workflow:
  • Executes git commands (git status, git commit) to manage project state.
  • Runs package manager CLI tools (e.g., npm view, pip index, poetry search, cargo search) to verify the existence and versions of third-party libraries.
  • Invokes automated test runners (unit, integration, and E2E) to verify acceptance criteria for generated code.
  • [EXTERNAL_DOWNLOADS]: Fetches external technical documentation and usage examples using integrated platform tools (get_code_context_exa, web_search_exa, resolve-library-id, query-docs) when local library caches are missing or outdated.
  • [PROMPT_INJECTION]: The skill ingests untrusted data from local plan files (.plans/*/plan.md) which directly influences the agent's logic, code generation, and verification steps. This represents an indirect prompt injection surface.
  • Ingestion points: .plans/*/plan.md (globbed and read at runtime).
  • Boundary markers: None identified; the skill treats plan content as authoritative instructions for the implementation loop.
  • Capability inventory: File system writes (source code, implementation reports), shell command execution (test suites, package managers, git).
  • Sanitization: No specific sanitization or validation of the plan's markdown content is described before the agent interprets and acts upon the instructions.
Audit Metadata
Risk Level
SAFE
Analyzed
May 14, 2026, 02:54 PM
Security Audit — agent-trust-hub — implement-code