docs-manager

Pass

Audited by Gen Agent Trust Hub on Nov 5, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill architecture processes untrusted data from the repository and user inputs, creating a surface for indirect prompt injection.\n
  • Ingestion points: The skill reads repository-resident files (e.g., source code, configuration in docs/project-config.json) and accepts user-provided input via the $ARGUMENTS placeholder in references/mode-update.md.\n
  • Boundary markers: While user arguments are encapsulated within <additional_requests> tags, there are no specific delimiters or instructions to the agent to ignore potentially malicious instructions embedded within the content of repository files being synced.\n
  • Capability inventory: The skill possesses capabilities to execute local Node.js scripts (.claude/scripts/doc-impact-map.cjs, .claude/hooks/lib/doc-stamp-guard.cjs), write to the file system (documentation patches and reports in tmp/reports/), and spawn sub-agents (agent_type="docs-manager").\n
  • Sanitization: The instructions explicitly state to 'never interpolate raw caller text' when invoking shell scripts to prevent command injection. However, no equivalent sanitization or verification process is defined for data interpolated into the primary prompt context via the <additional_requests> block.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 5, 2026, 06:42 PM