Kenon
Coding agents lose critical engineering decisions every time a session ends. Kenon writes them back into the repo — as structured, retraceable, machine-readable protocol — so the next session starts where the last one left off.每次会话结束，Coding Agent 都会丢失关键工程决策。Kenon 把它们作为结构化、可追溯、机器可读的协议写回仓库——让下一次会话从上次结束的地方开始。
Decisions, boundaries, evidence — structured into the repo, not left in chat.决策、边界、证据——结构化写进仓库，不留在聊天记录里。
The real problem isn't "not enough context". 真正的问题不是"上下文不够"。
It's that your Agent loses key engineering decisions between sessions — and makes the same mistakes again. 而是你的 Agent 在会话之间丢掉了关键的工程决策——然后犯同样的错。
Agent A explored the config-loader last week and decided "don't touch it — downstream modules depend on its undocumented parsing behavior." Agent A recorded this in the chat, then the session ended.
上周 Agent A 探索了 config-loader，得出结论："不要动它——下游模块依赖它的未文档化解析行为。" Agent A 把这个结论留在了聊天记录里，然后会话结束了。
Agent B in a new session sees the "messy" config-loader, refactors it.Agent B 在新会话里看到"混乱的" config-loader，重构了它。
Downstream modules silently break. Half a day debugging.下游模块静默出错。排查了半天。
Agent B's decisions also stay in chat. Next round, same risk.Agent B 的决策也留在聊天记录里。下一轮，同样的风险。
load src/config --scenario edit
→ Returns→ 返回 excluded: refactor-config-loader with grounds & reconsider_when附带依据与重审条件
check --diff
→ Intercepts: change touches an explicitly excluded path→ 拦截：改动触碰了显式排除路径
finalize
→ Agent B's decisions also write back to .kenon/, visible next session→ Agent B 的决策也写回 .kenon/，下个会话自动可见
Kenon doesn't record everything — only the part that will cause errors if forgotten next time. Kenon 不记录一切——只记录"下次忘了会出错"的那部分。
What Kenon changes. Kenon 改变了什么。
Excluded paths and key decisions stay in chat logs. Next session, they're gone. 排除路径和关键决策留在聊天记录里。下个会话，就没了。
load returns past decisions and excluded paths with grounds and reconsider_when.
load 返回过去的决策和排除路径，附带依据和重审条件。
Changes cross historical design boundaries with no warning. 变更无声无息地越过项目历史设计边界。
check --diff intercepts boundary violations before commit. impact shows the blast radius.
check --diff 在提交前拦截越界。impact 展示爆炸半径。
Agents rebuild state from scratch every round. Logs and diffs, every time. Agent 每轮都从零拼凑状态。每次都是日志和 Diff。
load returns folded state summaries, anomaly signals, and next-step navigation. Zero LLM.
load 返回折叠后的状态摘要、异常信号和下一步导航。零 LLM。
Each Agent runtime builds its own context silo. 每个 Agent runtime 都在搞自己的上下文孤岛。
One protocol layer natively connects 6 coding agent runtimes + MCP Server mode. 一层协议原生接入 6 个 Agent 运行时 + MCP Server 模式。
A short loop for real work. 主导真实工作的极简循环。
The standard lifecycle for agent interaction. 规范化 Agent 交互的标准生命周期。
Every step on this loop is deterministic — zero LLM. load filters, check detects boundaries, impact traverses the graph. Heavy judgment only at batch-end.
循环中每一步都是确定性的——零 LLM。load 过滤，check 检测边界，impact 遍历依赖图。重判断仅在批次末尾。
3-Minute Fast Track 三分钟快速上手
Do it yourself, or just hand the keys to your agent.亲自操作，或者把钥匙直接交给 Agent。
added 1 package in 1s
[kenon] Found 12 main modules, writing to .kenon/space/modules.jsonl
Read-only boundaries enforced for /shared/lib.
Prompt Drop-in注入提示词
Drop this straight into your agent. It will install and configure Kenon.直接粘贴这段提示词发给 Agent 进行原生初始化。
请在这个仓库里安装并初始化 Kenon。 1. 运行 npm install -g @kenon-ai/cli 2. 运行 kenon run onboard 3. 运行 kenon load --global --scenario edit 4. 告诉我这个仓库最值得先看的模块和原因 5. 后续在变更前默认先经过 kenon check --diff
Please install and initialize Kenon in this repository. 1. Run npm install -g @kenon-ai/cli 2. Run kenon run onboard 3. Run kenon load --global --scenario edit 4. Tell me which module is most worth looking at first and why. 5. In future changes, always pass through kenon check --diff by default.
kenon help --json.
人类可以通过简短路径起步，Agents 通过 kenon help --json 导航。
Agent-first.
Inspection-first.
Agent-first。
观察优先。
Kenon looks like a CLI, but for agents it's a supervisory cockpit: facts, signals, grounds, evidence, and next-step suggestions — never verdicts. The agent reads structured state, makes its own decisions, and Kenon keeps the ledger honest. Kenon 外观是 CLI，但对 Agent 来说，它更像一个监督座舱：事实、信号、依据、证据和下一步建议——但不裁决。Agent 读取结构化状态，自行决策，Kenon 只负责让账本保持诚实。
Humans don't need to learn Kenon. Agents navigate via kenon help --json.
人类不需要先学 Kenon。Agent 通过 kenon help --json 导航。
JSON is truthJSON 是真源
Markdown is projectionMarkdown 只是投影
Context is loaded上下文是加载的
Not pasted or injected不是粘贴或注入的
Commands self-describe命令自描述
reads / writes / risk / side effects读/写/风险/副作用
Signals guide信号引导
Never judges or blocks不裁决、不阻断
One protocol layer. Multiple runtimes. 单层协议，多重运行环境组合。
Native integration with 6 coding agent runtimes + MCP Server mode. 原生接入 6 个 Agent 运行时 + MCP Server 模式。
One anchor, four faces.
The dimensional model of agent context.
Paths & Anchors路径与锚点
The foundational map of your repository. Defines where code lives, what it depends on, and the physical limits of an agent's workspace.代码仓库的基础地图。定义了代码位置、依赖关系以及 Agent 能够触碰的物理边界。
Sessions & Continuity. Track history beyond the chat prompt. Know what was tried and why.会话与连贯性。追踪提示词之外的全局历史，记住尝试过什么以及原因。
Verification & Findings. Grounding truth in execution.验证与发现。以真实执行结果为基础的事实对齐。
Decisions & Exclusions. Explicit goals and forbidden paths.决策与排除。明确的目标约束与禁止触碰的红线。
Dependency Analysis. Seeing the ripple effect.影响面分析。洞悉任何一次修改所引发的连锁反应。
Start small. Grow when needed. 从最小开始，按需扩展。
Day 1 needs exactly 3 commands. 50+ commands are there when you're ready. 第一天只需要 3 条命令。50+ 条命令在你需要时随时可用。
Cold Start冷启动
run onboard + load — build space/deps foundation, generate first readable projectionsrun onboard + load — 建立空间/依赖基础，生成首轮可读投影
Daily Cycle日常循环
+ check --diff + finalize — add boundary guardrails and traceable close-outs+ check --diff + finalize — 加入边界守卫和可追溯收口
Change Closure变更收口
+ impact + verify + history --trace — narrow verification radius, add causal chain tracing+ impact + verify + history --trace — 缩小验证半径，补齐因果链追踪
Quality Governance & Concurrency质量治理与并发协作
quality-check, acceptance-gate, postmortem, review, stale, intent-lock, invoke, session, batch runner...quality-check、acceptance-gate、postmortem、review、stale、intent-lock、invoke、session、batch runner...
Humans don't need to learn Kenon first. 人类不需要先学 Kenon。
After install, it's the Agent that learns Kenon — via structured commands, projections, and the supervisory cockpit. 安装后是 Agent 学 Kenon——通过结构化命令、投影层和监督座舱自行导航。
Design Stance 设计立场
Telescope, not judge.望远镜，不是法官。
Kenon makes facts and structure visible. It never passes verdicts.Kenon 让事实和结构可见，但不替现实下结论。
Boundaries are first-class.边界是一等公民。
"Why not" is as important as "why". Exclusions, grounds and reconsider conditions are core data."为什么不做"和"为什么做"同等重要。排除路径、依据和重审条件是核心数据。
Hot ledger, not docs.热层账本，不是文档系统。
Kenon manages structured project knowledge (JSONL). Project docs (Markdown) stay independent.Kenon 管结构化项目知识（JSONL 热层），项目文档（Markdown 冷层）独立存在。
Workflow Tools vs Kenon. Workflow 工具 与 Kenon 区别。
It's a repo-local semantic ledger beneath the workflow layer. Kenon 是一层嵌入在工作流之下的本地工程语义账本。
Built for real codebases that evolve. 为持续演进的真实代码仓库而生。
Kenon shines when agents work on real codebases with evolving decisions, dependencies and quality signals. Of course, if the project is a throwaway script, a solo chat without a repo, or you need cloud-hosted context, Kenon isn't the right fit — it's a repo-local semantic layer, not a workflow replacement or a SaaS. Kenon 适合 Agent 在有真实决策、依赖和质量信号的代码仓库中持续工作的场景。当然，如果项目只是一次性脚本、纯聊天没有仓库、或需要云端管理上下文，Kenon 就不太适合——它是仓库本地语义层，不是工作流替代品，也不是 SaaS。