Reasonix v1.21.4
This release introduces completion contracts that make goal claims verifiable against receipts. Memory gains activation, volatility, subject keys, and CJK bigram retrieval. Desktop optimizations include session-switch performance, local file link handling, and IME-aware ESC behavior. Agent improvements bring incremental fold digest carry, compaction bounds, and a reasoning governor shadow mode.本次发布引入完成契约，使目标声明可对照收据进行核查。记忆功能新增激活、易变性、主题键和 CJK 双字检索。桌面端优化了会话切换性能、本地文件链接处理和输入法感知的 ESC 行为。智能体改进包括增量折叠摘要携带、压缩边界和推理调控器影子模式。
Highlights重点内容
Completion Contracts: Verifiable Goal Claims完成契约：可核查的目标声明
Goal completion now carries verifiable claims. Verified commands are reconciled against receipts; unbacked claims become gaps. Host-generated completion reports show missing proof without forcing verification loops.目标完成现在携带可核查的声明。已验证的命令会与收据进行核对；无依据的声明会变成缺口。由主机生成的完成报告会显示缺失的证明，而不会强制进行验证循环。
Memory Dimensions: Activation, Volatility, and Subject Keys记忆维度：激活、易变性与主题键
Facts can be pinned to the stable prefix or used only for retrieval. Volatility windows (evergreen/stable/volatile) and expiry dates control freshness. Subject keys prevent silent contradictions by allowing only one active value per question.事实可以固定到稳定前缀，或仅用于检索。易变性窗口（永久/稳定/易变）和过期日期控制新鲜度。主题键通过每个问题只允许一个活跃值来防止静默矛盾。
Desktop: Session Switch Performance and Local File Links桌面端：会话切换性能与本地文件链接
Large session histories and rapid topic switches no longer block the renderer. Local file paths in agent output are clickable to open, with right-click options for other apps and file manager.大型会话历史和快速主题切换不再阻塞渲染器。智能体输出中的本地文件路径可点击打开，右键菜单提供其他应用和文件管理器选项。
Agent: Incremental Fold Digest Carry and Compaction Bounds智能体：增量折叠摘要携带与压缩边界
Incremental folds now carry previous digests verbatim, recovering 17 of 18 lost probes. Compaction inputs are bounded, ensuring long sessions can always fold and preventing session wedging.增量折叠现在逐字携带之前的摘要，恢复了 18 个丢失探针中的 17 个。压缩输入有边界，确保长会话始终可以折叠，防止会话卡死。
Desktop: IME-Aware ESC and Selection Release桌面端：输入法感知的 ESC 与选区释放
Pressing ESC while an IME candidate window is open no longer cancels a running agent turn. Stale transcript selections are released on new clicks, and Add to Chat preserves the selection range during activation.在输入法候选窗口打开时按 ESC 不再取消正在运行的智能体回合。过时的聊天记录选区会在新点击时释放，添加到聊天操作在激活期间保留选区范围。
New新增
Completion Contracts完成契约
Goal completion claims are now verifiable. The host-authored completion report identifies gaps between claims and receipts, making premature 'done' declarations transparent.目标完成声明现在可核查。主机生成的完成报告可识别声明与收据之间的缺口，使过早的“完成”声明透明化。
Memory Activation and Subject Keys记忆激活与主题键
Facts can be pinned to the stable prefix or used only for retrieval. Subject keys prevent contradictory facts on the same topic.事实可固定到稳定前缀或仅用于检索。主题键可防止同一主题出现矛盾事实。
Memory Volatility and Expiry记忆易变性与过期
Facts now support volatility windows (evergreen/stable/volatile) and explicit expiry dates for freshness control.事实现在支持易变性窗口（永久/稳定/易变）和明确的过期日期，以控制新鲜度。
CJK Bigram Retrieval and Write-Time KeywordsCJK 双字检索与写入时关键词
Memory retrieval now uses CJK bigrams for better matching, and facts can include write-time keywords for semantic bridging.记忆检索现在使用 CJK 双字以改善匹配，事实可包含写入时关键词以实现语义桥接。
Explicit Context Compression Tool显式上下文压缩工具
A model-callable `compress` tool allows explicit range compaction without rewriting the canonical transcript.模型可调用的 `compress` 工具允许显式范围压缩，而无需重写规范记录。
MCP OAuth and Failure-Aware RoutingMCP OAuth 与失败感知路由
First-party OAuth authorization for remote MCP servers, with sandbox-safe credential storage and failure-aware tool routing.远程 MCP 服务器的第一方 OAuth 授权，具有沙箱安全的凭据存储和失败感知的工具路由。
Generation-Safe Runtime Composition代际安全的运行时组合
Runtime composition with dependency graphs, lifecycle states, and atomic publish/commit for multi-session isolation.具有依赖图、生命周期状态和原子发布/提交的运行时组合，用于多会话隔离。
Lazy Thinking Summaries懒加载思考摘要
Collapsed reasoning traces now show a preview line, with full Markdown rendering deferred until expansion.折叠的推理轨迹现在显示预览行，完整的 Markdown 渲染延迟到展开时。
Desktop Status Bar: TPS, Cache, and Output Tokens桌面状态栏：TPS、缓存和输出令牌
Configurable status bar items show turn throughput, cache token counts, and output tokens.可配置的状态栏项目显示回合吞吐量、缓存令牌计数和输出令牌。
Reasoning Governor Shadow Mode推理调控器影子模式
A shadow reasoning governor measures exploration-phase efficiency without enforcing policy, preparing for future optimization.影子推理调控器在不强制执行策略的情况下测量探索阶段效率，为未来优化做准备。
Evidence-Before-More-Mutation Nudge (Experimental)先验证再变更提示（实验性）
An env-gated nudge encourages verification after several blind mutations, available for experimentation.环境门控的提示在多次盲目变更后鼓励验证，可供实验使用。
Verification-Debt Telemetry验证债务遥测
Shadow telemetry tracks how long mutations sit unverified, providing data for future policy decisions.影子遥测跟踪变更未验证的时长，为未来策略决策提供数据。
Shadow Admission Control for Research Delegation研究委托的影子准入控制
Shadow measurement of research delegation calls identifies unnecessary subagent usage without blocking.研究委托调用的影子测量可识别不必要的子代理使用，而不会阻止。
Per-Round Cognition Telemetry每轮认知遥测
Trajectory digests now include per-round token usage and slow-round census for performance analysis.轨迹摘要现在包括每轮令牌使用情况和慢轮统计，用于性能分析。
Shadow Outcome Scorer影子结果评分器
A shadow scorer decomposes per-round outcomes (exploration, verification, churn) to measure progress quality.影子评分器分解每轮结果（探索、验证、变动）以衡量进展质量。
Eligibility-Point Counterfactual Fork Harness资格点反事实分叉框架
Infrastructure for causal experiments by forking identical states into control and treatment arms.通过将相同状态分叉为对照组和处理组来进行因果实验的基础设施。
Verification-Stress Benchmark Suite验证压力基准测试套件
A benchmark suite measuring EBM trigger precision across positive, neutral, and negative scenarios.一个基准测试套件，用于测量 EBM 触发器在正面、中性和负面场景中的精度。
MemoryBench: Paired Memory-Effectiveness SuiteMemoryBench：配对记忆有效性套件
A benchmark harness measuring memory utility delta with counterfactual arms and point-of-use markers.一个基准测试框架，通过反事实臂和使用点标记测量记忆效用增量。
CompactionBench: Cost and Fidelity of Repeated FoldsCompactionBench：重复折叠的成本与保真度
A benchmark measuring the cost and information loss of repeated session folding.一个基准测试，用于测量重复会话折叠的成本和信息丢失。
Live Flight-Recorder Dashboard实时飞行记录仪表板
A local dashboard for live e2ebench runs with per-round traces and replay.一个本地仪表板，用于实时 e2ebench 运行，具有每轮跟踪和回放功能。
Retrieval V2 Shadow检索 V2 影子
A shadow retrieval system with BM25F, code symbols, and mixed CJK grams, measured against production BM25.一个影子检索系统，具有 BM25F、代码符号和混合 CJK 分词，与生产 BM25 进行对比测量。
Session Recall Index and Crash-Safe Writers会话召回索引与防崩溃写入器
Memory recall now uses a prebuilt index for zero disk IO per turn, with atomic file writes for durability.记忆召回现在使用预构建索引，实现每轮零磁盘 IO，并通过原子文件写入确保持久性。
Goal as Sole Long-Task RuntimeGoal 成为唯一长任务运行时
Goal is now the only long-running task state machine; AutoResearch is reduced to a read-only compatibility reader.Goal 现在是唯一的长任务状态机；AutoResearch 降级为只读兼容性读取器。
DeepSeek Anthropic Access Recommendation推荐 DeepSeek Anthropic 接入
Fresh installs now default to the official DeepSeek Anthropic Messages endpoint, with legacy migration for existing configurations.全新安装现在默认使用官方 DeepSeek Anthropic Messages 端点，并为现有配置提供旧版迁移。
Skills Settings Alignment技能设置对齐
Desktop skills settings now match the MCP model, with implicit skill discovery as an explicit runtime policy.桌面技能设置现在与 MCP 模型匹配，隐式技能发现作为显式运行时策略。
Animated Logo Variants动画 Logo 变体
New animated SVG logo variants and updated README banner.新的动画 SVG Logo 变体和更新的 README 横幅。
Site Hero Redesign网站 Hero 重设计
Redesigned homepage hero install section with particles, unified fonts, and refreshed brand assets.重新设计的首页 hero 安装部分，包含粒子效果、统一字体和更新的品牌资源。
Improved改进
Desktop Session Switch Performance桌面会话切换性能
Large histories and rapid topic switches no longer block the renderer, with virtual list and worker-based Markdown parsing.大型历史和快速主题切换不再阻塞渲染器，采用虚拟列表和基于 Worker 的 Markdown 解析。
Local File Link Handling本地文件链接处理
Local file paths in agent output are now clickable to open with the default app, with right-click context menu options.智能体输出中的本地文件路径现在可点击以默认应用打开，并提供右键上下文菜单选项。
Incremental Fold Digest Carry增量折叠摘要携带
Incremental folds now carry previous digests verbatim, recovering most lost probes and maintaining O(1) fold cost.增量折叠现在逐字携带之前的摘要，恢复了大部分丢失的探针，并保持 O(1) 折叠成本。
Compaction Input Bounds压缩输入边界
Summarizer input is now bounded, preventing session wedging by partitioning oversized folds.摘要器输入现在有边界，通过分区超大折叠防止会话卡死。
Memory Index Scope Qualification记忆索引作用域限定
The provider-visible memory index now uses scope-qualified references and annotates overrides, aligning with recall precedence.提供者可见的记忆索引现在使用作用域限定的引用并注释覆盖，与召回优先级保持一致。
Desktop Edit Actions for Cancelled Turns取消回合的桌面编辑操作
Cancelled turns now preserve edit, rewind, fork, and summary actions by rehydrating authoritative history.取消的回合现在通过重新加载权威历史来保留编辑、回退、分叉和摘要操作。
Goal Contextual Tool HardeningGoal 上下文工具加固
Context-dependent tools are now gated before execution, with child agent isolation and migration boundary fixes.上下文相关工具现在在执行前进行门控，并修复了子代理隔离和迁移边界问题。
Crash Recovery WAL-Awareness崩溃恢复 WAL 感知
Session persistence is now CAS/WAL-protected, preventing stale overwrites and ensuring monotonic state.会话持久化现在受 CAS/WAL 保护，防止过时覆盖并确保状态单调。
Compaction Fold Region Partition Fix压缩折叠区域分区修复
The fold region is now correctly partitioned, preventing user-stated constraints from being dropped.折叠区域现在正确分区，防止用户声明的约束被丢弃。
MemoryBench Pipeline ObservabilityMemoryBench 管道可观测性
Fixed seeding and audit forwarding bugs, enabling the first real memory utility baseline.修复了种子和审计转发错误，实现了第一个真实的内存效用基线。
Dependency Upgrades依赖升级
Batched Dependabot upgrades for desktop frontend, site, and Go dependencies.批量 Dependabot 升级桌面前端、网站和 Go 依赖。
CI Linter and Windows Timeout FixesCI 检查器和 Windows 超时修复
Local linting now matches CI, and Windows smoke test timeouts are increased to handle runner variance.本地检查现在与 CI 匹配，Windows 冒烟测试超时增加以应对运行器差异。
Horizontal Wordmark Sizing Fix横向标识尺寸修复
Fixed the homepage header logo rendering as a large square by declaring SVG intrinsic dimensions.通过声明 SVG 固有尺寸，修复了首页头部 Logo 渲染为大正方形的问题。
Fixed修复
IME-Aware ESC Behavior输入法感知的 ESC 行为
Pressing ESC while an IME candidate window is open no longer cancels a running agent turn.在输入法候选窗口打开时按 ESC 不再取消正在运行的智能体回合。
Transcript Selection Release聊天记录选区释放
Stale transcript selections are now released on new clicks, and Add to Chat preserves the selection range during activation.过时的聊天记录选区现在会在新点击时释放，添加到聊天操作在激活期间保留选区范围。
Windows Test FlakinessWindows 测试不稳定
Hardened two tests that failed on Windows due to timing races on loaded CI runners.加固了两个因 CI 运行器负载导致时序竞争而在 Windows 上失败的测试。
CI-Red Test DeterminismCI 红色测试确定性
Made three CI-red tests deterministic across platforms by gating Unix-specific assertions and skipping on Windows.通过门控 Unix 特定断言并在 Windows 上跳过，使三个 CI 红色测试跨平台具有确定性。
Thanks致谢
This release was shaped by:感谢本版本的贡献者：
Ready when you are准备好即可升级