Cumora: 开源人机同群 team chat，协调硬门可审计

ID: 46c35733
原文链接: https://github.com/yetone/cumora
Added: 2026-08-18
Source: github / yetone
Original Date: 2026-08-17
AAIF Category: agents
Quality Score: 5
Status: active
Source Type: github
Language: en
Tags: multi-agent, cumora, coordination, byoa, open-source

中文摘要

开源 Slack 式 team chat，人类与 agent 同 roster：Cloud 版每个 agent 一个 K8s pod 跑 OpenAI Responses API 多跳工具 loop；BYOA 模式脑=本地 Claude Code 或 Codex，server 不持有用户 provider key，Cumora 缩成 wake 投递、triage、turn 框架与世界 I/O——「世界 I/O 与脑循环可解耦」的产品版教科书。最有价值的是 COORDINATION.md 把多 agent 同房失败拆成 race collision 与 brain misjudgment 两类，给出可抄的硬门：reply 前 seen-cursor freshness preflight（过时回复 HELD）、事务内 verbatim-dup HOLD、big-brain 并发 cap 6 + triage cap 8 双层、wake debounce 2.5s 合并同相 burst。自称 0.1.119 链测 8/8 有序 0 重复（upstream self-report）。反模式三句：infra 问题不要用 prompt 修；seen 游标别写进会清空的读状态列；软门若可无代价绕过必被腐蚀。

English Summary

Open-source Slack-style team chat where humans and agents share one roster. Cloud runs each agent in its own K8s pod on an OpenAI Responses API multi-hop tool loop; BYOA points the brain at local Claude Code or Codex with keys never touching the server (Cumora shrinks to wake delivery, triage, turn framing, world actions, and an llm_calls cost ledger). The standout artifact is COORDINATION.md: multi-agent-same-room failures split into race collisions vs brain misjudgment, with copyable hard gates — seen-cursor freshness preflight (stale replies HELD), transactional verbatim-duplicate HOLD, big-brain(6)/triage(8) two-layer concurrency caps, deterministic 500ms spawn interval + AdaptivePacer, 2.5s wake debounce/coalesce, 60s rate-limit cooldown. Self-reports 8/8 ordered, 0 duplicates on the cumora@0.1.119 chain test (upstream self-report, not reproduced locally). Anti-patterns worth stealing: never fix infra problems with prompts; never store the seen cursor in a read-state column that clears the inbox; a soft gate that can be bypassed for free will corrode.

原文要点（证据摘录）

README: cross-platform team chat, human+agent roster, personas/memory, DM/groups, task claiming, kanban & calendar, real email; Postgres as source of truth, Redis for pub/sub & presence.
BYOA.md: bypasses the server-side turn.ts hop loop — the engine brings its own agentic loop and compaction; server holds no user provider keys.
COORDINATION.md failure classes: race collisions (same next-step inserted twice) vs brain misjudgment (stale state, wrong step chosen).
Hard gates: per-agent model pin; big-brain cap 6 + triage cap 8 (capping only one layer drags the other); deterministic 500ms spawn interval + AdaptivePacer; wake debounce 2.5s/coalesce/mid-turn steer; rate-limit cooldown 60s; seen-cursor freshness preflight → HELD; transactional verbatim-dup HOLD.
stars snapshot: 2154 (2026-08-18 GitHub API).
Evidence: DeepResearch/2026-08-18-evening-会话产品化-BotMode-手写loop-研究材料/references/01-cumora-readme.md, 02-cumora-byoa.md, 03-cumora-coordination.md

来源 / Obsidian 引用

本机 Obsidian 同日材料：调研/2026-08-18-调研-BotMode与Cumora协作面 + DeepResearch 材料包（raw 全文快照）
评分依据：原文正文/官方 release body（不靠标题或源声誉）