Auto-discover enterprise architecture artifacts from Java source code
This MCP server has 7 tools with reasonable naming and Chinese/English descriptions, but suffers from critical gaps in parameter schema documentation and error handling guidance. While tool names follow verb_noun conventions (scan_project_tool, generate_app_architecture_tool, etc.), the input schemas visible in the code show only basic type information without proper JSON Schema structure, constraints, or validation rules. Descriptions are present but generic (20-100 characters, often just English translations of Chinese text without actionable LLM-optimized guidance). Output schemas are entirely undocumented, callers do not know what fields to expect beyond JSON summaries in the code comments. Error handling is minimal: only basic path validation via _validate_project_path(), no recovery guidance, no categorization of error types, no per-tool error specifics. The server lacks composition guidance, parameter relationships documentation, and any indication of side effects or idempotency. Tools operate on Java/Maven projects but expose no permission gates, rate limiting, or audit trails despite performing filesystem writes (generate_* tools). No tool annotations (readOnlyHint, destructiveHint) are present despite clear write operations in generate_app_architecture_tool, generate_data_architecture_tool, generate_data_lineage_tool, and export_intermediate_data.
分析某个字段变更的影响范围。追踪指定字段的所有下游受影响字段,用于变更影响评估。可选生成 DA-IMPACT Excel 报告。
将项目扫描和分析结果导出为 JSON 文件,供架构顾问导入做进一步分析。导出内容为结构化元数据(模块、API、集成关系),不含源码原文。
扫描 Java 项目并生成应用架构全套制品(AA-01~AA-08)。生成 Excel(含 AA-01 至 AA-05 五个工作表)和两个 PPTX 架构图。
扫描 Java 项目并生成数据架构全套制品(DA-01~DA-07)。生成 7 个独立 Excel 文件(DA-01 至 DA-07)及一个合并多 Sheet 工作簿。
生成数据血缘制品(DA-08 字段级映射清单、DA-09 数据流转清单)。分析项目中的字段级数据血缘:MapStruct 注解映射、setter/getter 链、BeanUtils.copyProperties、SQL 字段来源。同时检测跨系统数据流转(Feign/MQ/Dubbo 集成)。
生成技术架构(TA 系列)制品。
扫描 Java/Maven 项目结构,返回模块、框架、文件数量等概览信息。
Input schemas lack JSON Schema formalization. Parameters show only inline descriptions without explicit type constraints, enums, min/max bounds, or format declarations. For example, 'locale' accepts 'zh|en' but this is embedded in description text, not enforced as an enum constraint. LLM cannot programmatically validate before sending.
Output schemas completely undocumented. Tools return JSON responses (visible in code as dicts converted to JSON) but callers do not know field names, types, or whether results are paginated. For example, scan_project_tool returns a 'summary' dict but what fields does it contain? Is 'frameworks' a list? Are module objects nested? No schema definition.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 37 | <=2025-11-25 | v2 |
No error recovery guidance. Error handling is limited to _validate_project_path() which raises ValueError with i18n text. Tools do not categorize errors as retryable vs fatal, do not suggest next steps when errors occur, and do not return actionable error details. An LLM receives 'path not found' but does not know whether to ask user for a different path, retry, or abandon the goal.
Destructive operations (generate_app_architecture_tool, generate_data_architecture_tool, generate_data_lineage_tool, export_intermediate_data) lack tool annotations and confirmation mechanisms. No readOnlyHint or destructiveHint markers. No dry-run or confirmation_before_execute pattern. Agents cannot distinguish safe (read-only) tools from risky ones, risking accidental overwrites of output directories.
No audit trail or permission gates. Tools perform filesystem writes and analysis without logging who called them, with what parameters, or what side effects occurred. No permission checks (e.g., can this agent write to output_dir?). No rate limiting despite tools performing expensive I/O (Java project scanning).
Parameter relationships undocumented. 'output_dir' has a default value but tools do not clarify: if output_dir is empty string, does the tool create arch-output/ automatically? What if output_dir points to a non-writable location? No description states the behavior or failure modes.
analyze_field_impact parameter 'field' requires a specific format ('system.entity.field') but description does not validate this at runtime nor provide clear error guidance if format is wrong. LLM could pass 'Order.id' or 'order.Order.id' and receive a cryptic parsing error.
Descriptions are minimal and not LLM-optimized. Most descriptions are 20 - 80 characters (below the 50-200 baseline for A-grade tools). Example: 'generate_tech_architecture' has description '生成技术架构(TA 系列)制品。' (13 chars translated = ~15 in English). Does not state WHAT it generates, WHEN to use it vs other generators, or what artifacts are produced.