Cross-agent six-phase development loop engine (MCP server). Externalizes working memory to disk and serves phase playbooks just-in-time to cut token usage.
Server defines 6 tools with schemas and descriptions in Chinese. All tools have input schemas (Zod-based) and descriptions. However, descriptions are domain-specific and moderately detailed (avg ~120 chars), parameter descriptions are present but terse, and output schemas are not formally documented. Error handling returns structured TextResult with isError flag, but recovery guidance is minimal. Tool naming follows verb_noun pattern (loop_start, loop_status, loop_record, loop_budget, loop_advance, loop_resume) and is clear. No tool annotations (readOnlyHint/destructiveHint) despite clear risk levels (WRITE vs READ_ONLY). Composition is sound, each tool has one responsibility and outputs chain properly (e.g., loop_start returns state, loop_status reads it). Missing: output schema documentation, tool annotations, and richer error recovery hints.
申请进入目标阶段。引擎校验闸门(如未获批的 plan 不能进 IMPLEMENT),通过则即时下发下一阶段的 playbook。向后回退(如 VERIFY->IMPLEMENT)总是允许。
上报一次工具用量(用于 INVESTIGATE 阶段的探索预算)。如 loop_budget read 3。超预算会在 status/advance 中预警。
把工作记忆写入磁盘产物(context-map / plan / progress),而不是堆在对话里——这是省 token 的核心。默认追加。
恢复活动 loop:从 .active 载入状态,把已批准的 plan + context-map 正文连同当前阶段 playbook 一次性内联返回——用于 /clear 或 subagent 后在干净上下文续跑。只读,不推进阶段。无参。
开始一个新的开发循环(或恢复同名任务)。初始化 .agent-loop/<task>/ 与状态,返回当前阶段的 playbook。每个需求/任务开始时调用一次。
返回当前 loop 的阶段、预算用量、产物清单与下一步动作。只读,随时可调。
Output schemas not documented. Tools return TextResult with untyped content array; LLMs cannot predict structure or plan downstream calls. loop_status returns formatted text, not structured fields (phase, budget_used, artifacts, next_phase).
Tool annotations missing. loop_start, loop_record, loop_advance are destructive (WRITE risk); loop_status, loop_budget, loop_resume are read-only. No readOnlyHint or destructiveHint in tool definitions. LLMs cannot distinguish safe from unsafe calls.
Error recovery guidance minimal. err() function returns '⚠️ ' prefix but no actionable next steps. E.g., loop_budget returns 'stopped散读' (Chinese) without suggesting loop_advance PLAN or context-map write. Errors should guide LLM to recovery.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 57 | 2026-07-28+ | v2 |
Parameter descriptions lack format/constraint details. loop_budget 'tool' param accepts free-form string with hint 'read / grep / glob / semanticSearch / fetch', should be enum. loop_advance 'to' param is enum (good), but 'evidence' param is free-form string with no length/format guidance.
Descriptions in Chinese only. While domain-appropriate, descriptions lack English fallback for non-Chinese-speaking LLMs and developers. Reduces accessibility and maintainability.