用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt;运行前请先检查来源。
npx skills add https://github.com/majiayu000/claude-skill-registry --skill converge命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地?可下载 SkillsMP 当前能够提供的文件。
LLM token logprobs and calibration. Per-decision confidence, ECE, Brier, reliability diagrams, low-confidence triage.
Analyze LLM token logprobs and calibration. Use for per-decision confidence, ECE, Brier scores, reliability diagrams, and low-confidence triage.
回顾最近 N 天的 Claude Code 使用记录——扫描原始会话数据,按主题分组汇总"我都做了什么",并从个人操作系统视角输出模式、风险与增删建议。当用户说 /recap、"看看我这几天做了什么"、"回顾一下我最近的会话"、"这两天我用 claude 干了啥"、"活动回顾" 时使用。
基于 SOC 职业分类
正在显示 SKILL.md
| name | converge |
| description | Run converge. |
Quick Ref: Thin memo. The implementation is the Go command
ao converge(cli/cmd/ao/converge.go). Do not reimplement the loop — invoke the binary.
ao converge runs a bounded fix → re-run-judge-panel loop until the judges
agree or it blocks:
--max-rounds elapses.<dir>/.agents/rpi/KILL appears at a round boundary.The independence axis is fresh CONTEXT, not model family. --require-cross-family
is an optional strengthener.
ao converge --max-rounds 5 --min-contexts 2
ao converge --require-cross-family
The FIX step is yours (the orchestrating agent). The dispatched judge leg is non-mutating — verdict + evidence only, never edits the repo.
In a Codex context, the Codex → Claude judge leg has NO headless transport:
delegate it to the codex-approval skill / an NTM Claude pane. Never a headless
claude print-mode call. The Claude → Codex leg uses ao codex dispatch.
Before any dispatch, ao converge runs a two-sided canary entry gate — it proves
the gate rejects a planted self-judge verdict and accepts a known-good one. A
failed canary aborts the run. An empty/PASS result is a lie until proven to bite.
ao converge emits a CLAIM + evidence; MTO is the sole writer of binding verdicts.