Loading ToolBench
Multi-service stack operations console for local development environments. Provides tools for service management, build orchestration, Git/CI status monitoring, and documentation management.
Critical gaps across all dimensions. Tool definitions exist but lack proper JSON Schema structures, only string hints visible in StackToolSpec records. No input schemas with type definitions found in source. Descriptions are present but many are under 20 chars or lack actionable detail. Parameters documented only as space-separated hints ('mode path'), not as structured schema objects. No output schemas documented. Error handling is minimal, policy rejection returns JSON but lacks recovery guidance. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk levels (READ_ONLY, WRITE, DESTRUCTIVE). Tool names are generally good (verb_noun pattern), but parameter documentation is severely deficient.
編譯。mode=stale（預設）／services／projects／one。one 時必填 path。
與 list_projects 相同：編譯徽章與需重編數字。fresh／最新＝無需重編。
遠端 GitHub Actions 最近執行：通過／失敗／進行中。只摘要，不含 log。
文件體系狀態：docs/ 是否存在、骨架、待補頁數、DocFX／Pages workflow。
環境體檢（git、依專案偵測的語言環境／Agent 後端／MCP 政策）。
值班一眼摘要：就緒、離線、需重編。政策擋下的 MCP 呼叫不列入警報。晨會第一個工具。
讀取服務 Log 尾端。
No JSON Schema input definitions visible. Tools define parameters only as space-separated hints (e.g., 'mode path') in ArgsHint field, not as structured JSON Schema with type, description, enum, or constraints.
No output schemas documented. Tools return JSON strings but no schema describes the response structure, field types, or required fields. LLMs cannot plan downstream calls or extract data reliably.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 44 | 2026-07-28+ | v2 |
目前分支、未提交、領先／落後。
讀取最近的 MCP 審計紀錄（.ai_project/mcp-audit.jsonl）。
列出 docs/ 內的 Markdown 與 DocFX 設定檔。
列出各專案編譯徽章、一句用途，與需重編數字。needsRebuild=0 或 badge=最新 表示無需重編；不要把「已與 DLL 一致」理解成要重編。
列出 ai-project 服務 id、port、宿主與一句用途。
目前分支的 PR 與檢查是否通過。只回答能不能請人審，不含審查內容。
讀取 docs/ 內一份檔案。path 為相對 docs/ 的路徑，例如 user/getting-started.md。
堆疊總覽：就緒數、需重編、各服務線上／離線。開發管理者晨會或值班第一個工具。
啟動尚未在線的服務（依 startOrder，並先起 dependsOn）。
啟動一個服務。預設先起 dependsOn 並等到就緒。skipDepends=true 只起自己；skipOptional=true 略過可選相依。
停止全部服務。破壞性操作：必須先向使用者確認，再帶 confirm=true。
停止一個服務。可在政策 confirm 列入後要求 confirm=true。
Tool annotations missing. Risk levels are declared (READ_ONLY, WRITE, DESTRUCTIVE, REVERSIBLE) but not exposed as readOnlyHint, destructiveHint, or idempotentHint in tool metadata. Agents cannot distinguish safe from dangerous operations.
Parameter descriptions incomplete or missing. Tools with parameters (build, start_service, stop_service, get_log, list_audit, read_doc) document params only as hints; no descriptions explain format, constraints, or valid values.
Error handling lacks recovery guidance. Policy rejection returns {ok: false, error, needConfirm} but does not suggest next steps (e.g., 'Try calling doctor() to check permissions'). Errors do not categorize as retryable, user-fixable, or fatal.