Loading ToolBench
Generate verification-focused operating prompts for AI agents. MCP server for Layered Cognition prompt composition with model adapters and schema validation.
Server has 4 tools with basic schemas and descriptions, but significant gaps limit production readiness. Tool names lack clear action verbs (build_layer_prompt, build_chain_prompts are vague about what 'build' produces). Descriptions are present but often in Korean with minimal English context, making them inaccessible to most LLM agents. Parameter descriptions exist but lack specificity on constraints, ranges, and error cases. Output schemas are mentioned but not formally documented. No error handling guidance, tools return generic error dicts without recovery hints. Schema completeness varies: build_layer_prompt has typed params but validate_output lacks strict type enforcement. No tool annotations (readOnlyHint, destructiveHint) despite clear read-only semantics.
기존 프롬프트를 ANALYZE_CHECKS 7항목으로 로컬 감사. LLM 호출 없음. 결정론적 휴리스틱.
L0→L5 전체 레이어 체인 프롬프트 배열 반환. 반환값: {"chain": [{"layer_id": ..., "prompt": ...}, ...]} 에이전트가 순서대로 실행하고 prev_output을 다음 레이어에 전달.
특정 레이어 프롬프트 생성. layer_id: L0_scope | L1_evidence | L2_analyze | L3_hypothesize | L4_critique | L5_report model: claude | gemini | codex | hermes | generic task_context: {goal, config_name?, constraints?, ...} prev_output: 이전 레이어의 JSON 출력 (체인 진행 시)
레이어 출력 스키마 검증. LLM 아닌 코드가 게이트. schema_name: evidence_ledger | findings | hypotheses | critique
Tool names lack clear action verbs. 'build_layer_prompt' and 'build_chain_prompts' are ambiguous, does 'build' mean generate, compose, retrieve, or validate? LLMs cannot infer intent from these names alone.
Descriptions are primarily in Korean with minimal English context. LLM agents trained on English cannot reliably parse Korean descriptions. Descriptions must be in the agent's language.
Parameter descriptions lack constraint details. 'layer_id' lists enum values in description text but not as a formal enum constraint. 'model' similarly lacks enum validation. LLMs cannot reliably parse free-text enums.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 47 | 2026-07-28+ | v2 |
No error handling guidance. Tools return {'error': '...'} dicts without recovery hints. When layer_path.exists() fails, the LLM receives no guidance on what to do next (e.g., 'Call list_available_layers() to see valid layer_ids').
Output schemas are mentioned in docstrings but not formally documented in tool definitions. LLMs cannot parse informal schema descriptions. Schemas must be explicit in the tool's output_schema field or response documentation.