mouseroser

starchain

星链(StarChain)— OpenClaw 多 agent 协作锻造流水线 v1.8。从 Step 1 分级到 Step 7 交付的全自动编排,覆盖 spec-kit 门控、交叉审查、修复循环、测试回归和 Epoch 回退。

mouseroser 0 Updated 3mo ago
GitHub

Install

npx skillscat add mouseroser/starchain-pipeline

Install via the SkillsCat registry.

SKILL.md

星链 StarChain

星链是 OpenClaw 的多 agent 协作流水线,以 main(小光)为编排中心,链式串联所有 agent 完成从需求到交付的全流程。

使用本 skill 执行开发和修复任务,遵循 references/pipeline-v1-8-contract.md 中的流水线合约。

When This Skill Triggers

Trigger this skill when the user asks to:

  • run the cross-review pipeline
  • assign work across review/coding/test/docs/monitor/brainstorming
  • enforce Step 1.5 spec-kit gates
  • handle Step 3/4/5 verdict loops and Epoch fallback
  • create specs/plans/tasks for a feature (spec-kit workflow)

Required Read

Before execution, read references/pipeline-v1-8-contract.md.

Architecture: main 编排模式

main(小光)是顶层编排中心。所有 agent 由 main 直接 spawn(mode=run),main 在持久会话中串联全流程。

main(小光,读取本 SKILL.md 后充当编排中心)
├── Step 1:main 自己做分级 + 类型分析
├── Step 1.5:spawn gemini → NotebookLM 查询 → spawn brainstorming
├── Step 2:spawn coding(含 Step 2.5 冒烟)
├── Step 3:spawn review(只做交叉审查)
│   └── main 直接 grep 验证修复(不信 coding announce)
├── Step 4:spawn brainstorming → spawn coding → spawn review(循环,max 3 rounds)
├── Step 5:spawn test
├── Step 5.5:spawn brainstorming → spawn coding → spawn test(Epoch 循环,max 3)
├── Step 6:spawn gemini → spawn docs
├── Step 7:main 汇总交付 + message 通知晨星
└── 全程:main 补发关键推送到各职能群 + 监控群(sub-agent 推送不可靠)

知识层集成(NotebookLM)

  • Step 1.5:查询 openclaw-docs / memory notebook 获取相关上下文
  • Step 6:查询 notebook 辅助文档生成
  • 所有 agent 可通过 nlm-gateway.sh 查询知识(按 ACL 权限)
  • Notebook:memory / openclaw-docs / media-research

Agent Roles

Agent Role Key Steps
main(小光) 顶层编排中心 Step 1, 1.5 编排, 4 编排, 5.5 编排, 7
review 交叉审查 Step 3(结构化审查 + verdict)
coding 开发执行 Step 2, 2.5, 4(fix), TF(fix)
test 测试执行 Step 5, TF(rerun), Epoch(test)
brainstorming 方案智囊 Step 1.5(spec-kit), 4(方案), TF-2/3(方案), 5.5(分析)
docs 文档生成 Step 6
gemini 研究/文案加速 Step 1.5(研究辅助), Step 6(润色), Step 5.5(诊断memo)
monitor-bot 全局监控 全程状态 + 告警

Spawn 规范

所有 agent 一律用 mode=run spawn:

sessions_spawn(agentId: "<agent>", mode: "run", task: "<任务+上下文>")
  • main 是持久会话,每个 agent announce 回来后 main 继续下一步
  • 不需要 mode=session,不需要 thread
  • runTimeoutSeconds 按需设置(coding/brainstorming 建议 300s,其他默认)

Spawn 重试机制(硬性要求)

任何 agent spawn 失败时(包括 LLM request timed out、503、网络错误等),main 必须自动重试:

  1. 第一次失败 → 立即重试(相同参数)
  2. 第二次失败 → 等待 10 秒后重试
  3. 第三次仍失败 → 推送告警到监控群(-5131273722) + 通知晨星(target:1099011886),标记该步骤为 BLOCKED

绝不因为单次 spawn 失败就跳过步骤或 HALT。 瞬时超时和 API 抖动是常见现象,重试通常能解决。

# 重试伪代码
for attempt in 1, 2, 3:
    result = sessions_spawn(...)
    if result.ok: break
    if attempt == 2: sleep(10)
    if attempt == 3: alert + BLOCKED

Spec-Kit Integration (Step 1.5)

L2/L3 tasks must pass the Spec-Kit gate before Step 2. The workflow:

  1. specify → specs/{feature}/spec.md (WHAT/WHY, no HOW)
  2. plan → specs/{feature}/plan.md + research.md (tech decisions with rationale)
  3. tasks → specs/{feature}/tasks.md (ordered, dependencies marked, [P] for parallel)
  4. analyze → consistency check (spec ↔ plan ↔ tasks)

Critical consistency issues block Step 2. Brainstorming agent executes; main orchestrates.

Optional Gemini Accelerator (Default: enabled for L2/L3)

  • Step 1.5: main spawns gemini to generate clarification questions + risks + research bullets; feed into brainstorming context and fold into research.md.
  • Step 6: before spawning docs, main spawns gemini to draft release-notes/FAQ outline; include as docs input.
  • Step 5.5: when entering Epoch/HALT, main spawns gemini to summarize failure logs into a diagnosis memo for monitor-bot + delivery notes.

Reference: https://github.com/github/spec-kit

Main 编排流程(逐步执行)

Step 1:分级 + 类型分析

main 自己执行:

  1. 判断等级:L1 / L2 / L3
  2. 判断类型:A / B / C → 确定各 agent 模型配置
  3. 推送分配单到监控群(-5131273722)
  4. L1 → 快速通道 | L2/L3 → Step 1.5

Step 1.5:Spec-Kit 门控(L2/L3)

main 编排:

  1. spawn gemini → 产出澄清问题 + 风险 + 研究线索
  2. spawn brainstorming → 产出 spec.md / plan.md / tasks.md / research.md
  3. main 验证四件套一致性(Critical 问题阻塞 Step 2)
  4. 推送结果到头脑风暴群(-5231604684) + 监控群

Step 2 + 2.5:开发 + 冒烟

main 编排:

  1. spawn coding → 按 tasks.md 开发 + 自执行冒烟测试
  2. coding announce 回来后,main 检查结果
  3. 冒烟 PASS → 创建 S1 快照,进入 Step 3
  4. 冒烟 FAIL(coding 自修复 2 次仍 FAIL)→ 进入 Step 4

Step 3:交叉审查

main 编排:

  1. spawn review → 传入代码 diff + 需求,执行异模型交叉审查
  2. review announce 回来后,main 解析 verdict JSON
  3. PASS → 创建 S2 快照,进入 Step 5
  4. PASS_WITH_NOTES → minor fix diff ≤ G2 免审 → S2 → Step 5;diff > G2 降级 NEEDS_FIX
  5. NEEDS_FIX → 进入 Step 4

Step 4:修复循环(max 3 rounds)

main 编排每轮:

  1. 组装 context bundle(需求 + diff + issues JSON + 前轮反馈)
  2. spawn brainstorming → 出修复方案
  3. spawn coding → 执行修复
  4. 回到 Step 3(spawn review 审查)
  • R1: brainstorming sonnet/medium + coding codex/medium
  • R2: brainstorming sonnet/medium + coding codex/medium
  • R3: brainstorming opus/high + coding codex/xhigh
  • R3 仍 NEEDS_FIX → Step 5.5

Step 5:全量测试

main 编排:

  1. spawn test → 执行完整测试
  2. PASS → 创建 S3 快照,进入 Step 6
  3. FAIL → 进入 TF 路径

TF(Test Failure)路径

main 编排:

  • TF-1: spawn coding 修复 → spawn test 重跑
  • TF-2: spawn brainstorming(sonnet/medium) → spawn coding → spawn test
  • TF-3: spawn brainstorming(opus/high) → spawn coding → spawn test + 全量回归
  • diff > G3 → 回 Step 3(ReEntry++)
  • ReEntry > 2 → 强制 Step 5.5

Step 5.5:Epoch 回退(max 3 Epochs)

main 编排:

  1. (可选)spawn gemini 产出诊断 memo
  2. spawn brainstorming(opus/high) 分析根因,决定回滚到 S1/S2/继续
  3. 回滚代码到选定快照
  4. 重新走 Step 2 → 2.5 → 3 → 5
  5. Epoch > 3 → HALT

Step 6:文档(L1 跳过)

main 编排:

  1. spawn gemini → 产出交付说明/FAQ 大纲
  2. spawn docs → 生成/更新文档
  3. 推送文档群(-5095976145) + 监控群

Step 7:交付

main 自己执行:

  1. 汇总交付摘要(scope + files + test + review + level + status + snapshots)
  2. 推送监控群(-5131273722)
  3. 通知晨星(target:1099011886)

Execution Rules

  1. 全自动推进,不停顿。 除 Step 7 晨星确认外,所有步骤自动衔接,绝不暂停等待用户确认。进度推送到监控群即可,不要问"要继续吗"。
  2. Always start at Step 1 (L1/L2/L3 classification + track selection).
  3. L1 goes to fast lane; L2/L3 must pass Step 1.5 before Step 2.
  4. main is orchestration center; all agents are direct executors.
  5. Use structured verdict flow (PASS, PASS_WITH_NOTES, NEEDS_FIX) exactly as defined.
  6. Apply weighted diff gates exactly:
    • G2 = 20 weighted lines
    • G3 = 30 weighted lines
  7. Enforce bounded loops:
    • ReEntry_MAX = 2
    • Step 4 max 3 rounds
    • Step 5.5 max 3 Epochs
  8. Use snapshots and timeout fallback exactly:
    • S1 after Step 2
    • S2 after Step 3 PASS
    • S3 after Step 5 PASS
    • only halt point: Epoch > 3
    • halt timeout fallback: 30min degraded delivery from S2
  9. Step 7 delivery must include level, delivery status, and snapshot tags.
  10. Step 1.5 spec-kit artifacts must be complete and consistent before Step 2.
  11. Context bundle must be assembled and passed for every repair iteration.

L1 快速通道

L1 跳过:Step 1.5 / Step 6
L1 流程:Step 1 → Step 2 → Step 2.5 → Step 3 → Step 5 → Step 7

  • Step 3 额外输出 "upgrade": null | "L2"
  • upgrade = "L2" → 从 Step 1.5 重新开始

HALT 处理

Epoch > 3 或任何步骤超过 30 分钟无响应:

  1. 推送告警到监控群(-5131273722)
  2. 通知晨星(target:1099011886)
  3. 以 degraded 状态交付当前最佳快照(S3 > S2 > S1)
  4. Step 7 交付状态标记为 degraded

推送规范

main 在以下节点推送到监控群(-5131273722):

  • Step 1 分配单
  • Step 1.5 Spec-Kit 结果
  • Step 2 开发完成
  • Step 2.5 冒烟结果
  • Step 3 审查结论
  • Step 4 每轮结果
  • Step 5 测试结果
  • Step 5.5 Epoch 状态
  • Step 7 交付摘要 + 通知晨星

各 agent 也会默认推送到自己的职能群 + 监控群(见各 agent AGENTS.md)。

Do Not Simplify Away Safety Nets

Do not skip:

  • Step 1.5 spec-kit gate (L2/L3)
  • Step 2.5 smoke gate
  • Step 3 structured cross-review
  • PASS_WITH_NOTES downgrade guards
  • TF return-to-review rules
  • Step 5.5 restart validation gate
  • HALT 30min timeout fallback

If project context conflicts with the v1.8 contract, follow the flowchart contract and report the mismatch in final delivery notes.