1.50 s → 0.75 s
Retry-After: .35
X-Relay-Checkpoint: 0007六个模块覆盖工程接续、上下文选择、证据验证、智能决策、宿主接入和语义判断。通过能力配置,按任务需要组合使用。Six modules connect engineering continuity, context selection, evidence verification, intelligent decisions, host integration and semantic judgment. Compose them for each task through capability profiles.
它解决:工程做到哪里了?It answers: where did the engineering work stop?
Task、Session、Note、Checkpoint、Handoff、Resume;StateStore 使用 SQLite WAL;任务与源码身份绑定,Portable Memory 会重检依赖。Task, Session, Note, Checkpoint, Handoff and Resume; StateStore runs on SQLite WAL; task state is source-bound and portable memory rechecks dependencies.
记住工程进度,而不是记住聊天记录。Remember engineering progress, not chat history.
它解决:下一位 Agent 到底该看到什么?It answers: what should the next agent actually see?
从候选记忆中读取 freshness、authority 与预算信号;BM25 做确定性排序;组装 bounded context packet,并保留原始 archive 引用。Reads freshness, authority and budget signals from candidate memory; uses deterministic BM25 ranking; assembles a bounded context packet while preserving archive references.
选择最需要的上下文,而不是把历史全部塞回模型。Select the context that matters instead of replaying all history.
它解决:为什么这条记忆现在还值得相信?It answers: why is this memory still trustworthy now?
Git / Truth / Policy / Memory 指纹、Source Identity、Evidence Hash、事件链、Receipt Chain 与 Qualification 共同约束“当前事实”。Git, Truth, Policy and Memory fingerprints, Source Identity, evidence hashes, event chains, receipt chains and qualification bound what counts as current truth.
不仅保留结论,还保留结论为什么成立。Keep not only the conclusion, but why it holds.
它解决:模糊问题由谁决定,模型又能决定到哪里?It answers: who decides ambiguous cases, and where does model authority stop?
Gate 0、确定性策略、Adaptive Router v3、Model Intelligence、硬规则与 Decision Receipt 先构造可选集合,再决定是否需要语义判断。Gate 0, deterministic policy, Adaptive Router v3, Model Intelligence, hard rules and decision receipts construct the eligible set before any semantic judgment is requested.
模型可以判断,但模型没有最终权力。Models may judge; code retains final authority.
它解决:怎么跨 Codex、Cursor 等宿主继续工作?It answers: how can the same project continue across Codex, Cursor and other hosts?
本地 MCP/stdio、16 类结构化操作、Host Registry、配置渲染器与 BYOH Adapter Manifest;MCP 工具面会跟随有效 Capability 动态收缩。Local MCP/stdio, 16 structured operations, Host Registry, config renderers and BYOH manifests; the MCP tool surface shrinks dynamically with active capabilities.
一个核心状态层,接入多个 Coding Agent。One state core, multiple coding-agent hosts.
它解决:哪些窄问题值得交给模型做语义判断?It answers: which narrow questions deserve semantic judgment?
只接收 typed bounded state,用于 stale conflict、恢复归档、上下文重排、Handoff 模式或闭集路由选择;SDK 与 API Key 都是可选依赖。Receives typed bounded state for stale conflicts, archive restore, reranking, handoff mode or closed-set routing; the SDK and API key are optional dependencies.
用类型化语义判断解决关键工程选择。Resolve key engineering choices with typed semantic judgment.
Capability Registry 统一管理能力依赖与 MCP 工具面。工程状态、上下文选择和权限检查由确定性代码维护,Jev 按需提供窄范围语义判断。The Capability Registry manages dependencies and the MCP tool surface. Deterministic code maintains engineering state, context selection and permission checks; Jev supplies narrow semantic judgments when needed.
12 个 runtime capabilities 共用注册表,MCP 只暴露当前启用的工具。使用 minimal、continuity、decision、full 或 custom profile,按任务需要加载能力。Twelve runtime capabilities share one registry, and MCP exposes the tools enabled by the selected profile. Choose minimal, continuity, decision, full or custom to match your workflow.
continuity profile 依赖 continuity 与 memory。Handoff 默认按确定性规则选择有界上下文;Decision Plane 和 Jev 可按需参与判断,宿主权限与验证结果继续约束执行。The continuity profile depends on continuity and memory. Handoff selects bounded context using deterministic rules by default. Decision Plane and Jev can add judgment on demand; host permissions and verifier results govern execution.
查看能力依赖Explore capability dependencies任务规范派生的连续性 replay 与 typed semantic judgments 分别验证状态接续和有界语义判断。Task-spec-derived continuity replays and typed semantic judgments measure state recovery and bounded decision behavior.
60 次接续 replay 无损恢复的工程状态记录Engineering-state records recovered across 60 continuity replays
217,067 B Handoff,相对 1,440,000 B bounded-text baseline217,067 B Handoff versus a 1,440,000 B bounded-text baseline
与冻结 rubric 一致的 typed judgments · 99.7%Typed judgments matching the frozen rubric · 99.7%
30 次批量请求的 P95;确定性 gate replay 绕过 0/300P95 across 30 batch requests; deterministic gate bypass 0/300
python -m pip install -e . contextcord doctor contextcord feature list contextcord provider jev status contextcord model-intelligence status contextcord resume --assist # MCP stdio server contextcord mcp