AI-powered character roleplay theater with agent runtime, coding tools, and extensible skill system
Liyuan exposes 5 tools with minimal documentation. Tool names lack action verbs (read, bash, edit, write, echo_session are acceptable but generic). Descriptions are extremely brief (10-50 chars), below the 194-char baseline for production tools. Input schemas are present but sparse: parameters lack type constraints, enums, ranges, or validation guidance. The 'bash' tool accepts arbitrary commands with only a timeout parameter, no safety constraints, no dry-run option, no permission gates. The 'edit' tool schema is incomplete (missing content/operation details). No output schemas documented. Error handling is absent, no recovery guidance, no categorization. Security is critical: bash execution without sandboxing or audit trails is a severe risk. Overall, this reads as an early prototype, not production-grade.
Execute bash commands with optional timeout
Echo a message with prefix: [session]
Edit file contents with diff tracking
Read file contents with optional offset and limit parameters
Write content to a file
bash tool executes arbitrary shell commands with no sandboxing, permission checks, or audit trail. Poses critical security and safety risk.
All tool descriptions are under 50 characters, far below the 194-char production baseline. Descriptions lack WHAT/WHEN/WHY context needed for LLM tool selection.
Input schemas lack type constraints, enums, ranges, or validation rules. Parameters like 'timeout' and 'offset' have no min/max bounds; 'command' has no injection guards.
No output schemas documented. LLMs cannot predict return structure, forcing them to guess field names and types for downstream tool chaining.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 42 | 2026-07-28+ | v2 |
No error handling guidance. Tools do not categorize errors as retryable/user-fixable/fatal or provide recovery hints. LLMs cannot self-correct on failure.