Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add KtKID/x-dev-pipeline --skill x-multi-llm-aligngit clone --depth 1 https://github.com/KtKID/x-dev-pipelineWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/ktkid/x-dev-pipeline/x-multi-llm-align)<a href="https://agentmods.dev/skills/ktkid/x-dev-pipeline/x-multi-llm-align"><img src="https://agentmods.dev/badge/skills/ktkid/x-dev-pipeline/x-multi-llm-align.svg" alt="Measured on agentmods" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00293 | $0.03230 |
| Opus 5 | $0.00147 | $0.01615 |
| Sonnet 5 | $0.00059 | $0.00646 |
| Haiku 4.5 | $0.00029 | $0.00323 |
Grade A, and why
x-multi-llm-align scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 8d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 267 lines — stays where its author put it; the contents beside it link to each section on GitHub.
x-multi-llm-align — 跨子 agent 协议对齐器
这个 skill 解决什么
把"两个子 agent 各自代表自己实现方立场来回审稿协议"的高质量对齐流程固化下来。
为什么这种流程比单方面写文档更有价值:
- 不同子 agent 分别代表不同实现方的立场,会从自己实现侧的"我得真去实现"角度提问,挖出对方写文档时漏掉的实现陷阱
- 用户作为人类中间人传递反馈,不需要两个子 agent 之间直接通信,也保留人类对关键产品决策的拍板权
- 多轮迭代后产出的契约双方都站得住脚,开发时返工率低
- 保留推理过程,后续接手的 agent 能直接读懂整套对齐逻辑,不用从头推一遍
适用与不适用
适用:
- 协议对齐(JSON/JSONL/HTTP API/RPC 契约)
- 数据结构对齐(事件 schema、消息格式、状态机定义)
- 流程对齐(接入流程、生命周期约定、错误处理流程)
不适用(这些场景将来可能有姊妹 skill):
- 代码 PR 正确性评审 → 用
x-cr - 写作 peer review(论文/博客/PRD)
- 单方面文档 review(不涉及第二个子 agent)
工作流
阶段 0 — 触发后的准备
第一步:识别讨论文件位置(按优先级)
- 用户提供的 spec 目录下,新建/读取
discussion-<topic>.md(不污染原 spec 文档) - 当前项目的
dev-pipeline/discussions/discussion-<topic>.md(不存在则mkdir -p) - 绝对 fallback:
/tmp/multi-llm-align/discussion-<topic>.md(保留旧目录名,兼容历史讨论文件)
topic 取自 spec 主题名(如 claude-sidecar-contract、thread-event-schema),不要太泛。
单文件追加:所有轮次都写在同一份文件里(不每轮新建),让后续接手的 agent 一次读完所有上下文。
第二步:识别自己的子 agent 名
每次发言都要标注子 agent 名,方便后续接手时区分谁说的。
识别策略(按优先级):
- 优先使用当前子 agent / reviewer / 实现方名称。
- 找不到就一句话问用户:"我用什么名字署名?"
- 对方子 agent 的名字:用户提供,或从 spec 文档作者标记里读,找不到就标
[<对方未知>]
发言标签格式:[agent-a] / [agent-b] / [实现方-A] / [<对方未知>]
第三步:如果讨论文件已存在
直接读完整个文件,搞清当前在第几轮、上一轮决议是什么,然后接力进入下一轮。不要重新发起第 1 轮。
阶段 1 — 单轮 review 流程
每轮必做这 5 步:
1.1 读全文(第一轮)或读 diff(后续轮)
第一轮:完整读所有相关 spec 文档,不抽样、不靠目录推测内容。漏读的部分会变成"我没发现的对齐问题",下一轮才暴露,浪费一次往返。
第二轮起:只读对方改了什么(用 git 或对照之前版本),同时全文档扫一遍确认命名空间没漂移(特别要检查状态机表、映射表、错误处理段落、跨文档引用——这些位置最容易漏改)。
1.2 建立编号反馈清单
按问题性质给前缀编号:
| 前缀 | 含义 | 例子 |
|---|---|---|
E |
Error / 不一致 / 矛盾(必改) | E1 — 04 和 05 文档对 content 类型定义不一致 |
Q |
Question / 不明确点(待敲板) | Q1 — 进程是 thread 级还是 run 级? |
N |
New / 我建议新增的事项 | N1 — 缺 generic_chat 路径定义 |
D |
Decision / 待用户拍板的产品决策 | D1 — deny 后是否强制 fail? |
每个问题用统一格式:
### E1 — <一句话问题描述>
**现状**:<对方文档怎么写的,引用具体行/段落>
**问题**:<为什么这是问题,不改会发生什么后果>
**推荐**:<具体怎么改,给最终态文字而不是泛泛建议>
**理由**:<为什么这么改最好——给硬背书:实测/SDK 源码/类比同类系统>
**影响面**:<协议字段 / 状态机 / 实现细节 / 产品决策>
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 8d ago First seen · 267 lines · 293 tokens per session scan A cf3e561b050a
x-multi-llm-align is a skill published in the GitHub repository KtKID/x-dev-pipeline (12 stars, last pushed 1mo ago), licensed MIT. It adds 293 tokens to every session and 3,230 once invoked, about $0.0015 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other skills, from other repositories
procoder
Work like a senior developer in a repository governed by procoder: run the commit gate before calling anything done, format and lint through the binary, and drive the spec, plan, todo, backlog, and sprint chain in .procoder/. Use this skill when the repository contains a .procoder/ directory or an AGENTS.md naming…
idea-team
A structured group discussion for exploring a product idea through three roles: researcher, critic, and analogy finder. The roles look for evidence, weaknesses, and useful comparisons.
codex-agent
Use when you want a second-opinion review via Codex CLI, cross-verification after another agent implements changes, debugging help, or alternative implementation proposals. Requires Codex CLI to be installed and authenticated.
project-health-auditor
Comprehensive codebase health analysis. Use when reviewing code quality, identifying technical debt, checking dependencies, or assessing project structure.
review-gate
Use before an agent-produced diff is committed, pushed, opened as a PR, merged, landed, or applied to user files when explicit implementation approval is missing. Trigger for review gate, review pack, approve before landing, diff first then land, human approval, merge gate, commit gate, push gate, or PR readiness.…
review-spec
A workflow for comparing an existing feature specification with the project's review checklist and implementation. It looks for missing requirements, differences between documented screens and actual screens, and lessons from rejected submissions.