RimeJev 0.5.1 · Daily compression & rolling retention

中文

RimeJev 0.5.1 为可选的训练数据采集增加按天压缩和自动容量管理，持续保留较新的输入记录。

• 默认关闭：采集与输入增强独立，初始为仅采集所选应用、名单为空。升级不改变已有的开关、保存路径、应用范围或模型设置。
• 按天压缩加密：按本机日期分文件，记录先压缩再用 AES-GCM 加密，当天也即时压缩。不可压缩的记录使用加密原始负载，避免体积反增。
• 保留新记录：预计占用达到上限 90% 时，从最旧采集文件开始永久清理至约 80%，然后继续写入。自动清理不进入废纸篓；手动删除仍可移到废纸篓。界面显示清理数量和释放空间。
• 兼容旧记录：0.5.0 文件仍可预览和导出；下次启用写入时会尝试压缩旧文件，只在完整写入成功且更小时原子替换。关闭采集和只读管理不会触发迁移或自动清理。
• 固定版本安装：中英文 README 和独立的 LLM_INSTALL.md 均已同步到 v0.5.1。

完整部署前检查、原生构建/签名、真实 Rime 回归和本机安装后检查已通过。隔离的合成数据验证了压缩还原、跨天滚动、旧格式迁移、90%/80% 清理、最旧优先、异常中断及采集到导出的完整流程；实际安装的设置窗口确认了新规则和关闭状态。未使用真实用户数据触发容量清理。

兼容性边界： 0.5.0 无法读取新压缩格式。若需降级，请保留 0.5.1 或更新版本用于读取，或提前明确导出需要保留的记录。自动清理以完整文件为单位，实际保留量会有差异。目录、密钥或写入异常仍会暂停采集，基础输入保持可用。

仅发布源码，不附带训练记录、训练导出、个人词库、凭据、本机配置、日志、备份、模型权重或二进制。仓库内的评估样例及测试词库为有明确来源说明的人工合成材料。发布内容及公开 Git 历史经过凭据与私有文件检查。没有提供 Apple 公证的即装安装包。

English

RimeJev 0.5.1 adds daily compression and automatic capacity management to optional training capture, retaining newer input records.

• Off by default: capture is independent of enhancement, initially using selected-app scope with an empty list. Upgrades preserve existing capture state, directory, app scope and model settings.
• Daily compressed encryption: files follow the Mac's local calendar day. Records are compressed before AES-GCM encryption, including today's writes. Incompressible records use an encrypted raw payload to avoid growth.
• Keep new records: projected 90% quota usage permanently deletes oldest managed capture files toward 80%, then continues writing. Automatic cleanup does not use Trash; manual deletion remains recoverable. The UI shows removed-file and reclaimed-space counters.
• Read legacy records: 0.5.0 files remain readable/exportable. On the next enabled write, legacy compaction atomically replaces only complete, smaller encrypted files. Off/read-only operations never compact or prune.
• Pinned installation: both README languages and the dedicated LLM_INSTALL.md now select v0.5.1.

The full deployment preflight, native build/signature, real-Rime regressions and installed smoke passed. Isolated synthetic tests exercised compression roundtrips, midnight rotation, legacy migration, 90%/80% retention, oldest-first ordering, interruptions and the complete capture-to-export pipeline. The installed UI confirmed the policy and off state. Real user data was not used to trigger quota cleanup.

Compatibility: 0.5.0 cannot decode the new compressed format. Retain a 0.5.1-or-newer reader or explicitly export needed records before downgrading. Retention removes whole files, so remaining usage varies. Directory, key or write failures can still pause capture; native input remains available.

This release provides source only. No captures, training exports, personal dictionaries, credentials, local configuration, logs, backups, model weights or binaries are attached. Documented hand-authored evaluation/test fixtures remain in the repository. Release content and public Git history were checked for credentials and private artifacts. An Apple-notarized installer is not provided.

Installation and evidence / 安装与验证

• README 中文 / English
• LLM_INSTALL.md
• Training capture / 训练采集说明
• Validation evidence / 验证记录

Runtime code is unchanged from tested and locally installed source 186487664e1c5d974eaed86792c7f4a560b0957a; later publication changes are Markdown only. Existing installations of that 0.5.1 build need no reinstall for this source release. New installations must run preflight for their exact release checkout.

RimeJev 0.5.0 · Encrypted training capture

中文

RimeJev 0.5.0 增加独立、默认关闭的本机训练数据采集，并包含自 0.4.1 发布以来的同行语境补全改进。

• 在“训练数据采集与管理…”中选择保存文件夹、仅采集或排除的应用、前文/模型字段及容量上限。初始为关闭、仅采集所选应用、名单为空；开启增强不会自动开启采集。
• 关联真实拼音、有限前文、原生候选、展示顺序和实际人工选词。原始记录采用 AES-GCM 加密，密钥独立保存在本机钥匙串；采集不额外加载或调用模型。
• 支持本机预览、移到废纸篓及离线 records/sft 导出。Tev prompt/completion 导出需用户提供兼容的本地分词器；导出是明文，应先审阅。数据管理先关闭采集，结束后保持关闭；不自动上传或训练。
• 同行语境可使用已有、粘贴及停顿后继续输入的前文；空语境也可请求模型判断。继续保留原首选、按设置将不同推荐放在第二/三位；失败时原生输入继续可用。
• 中英文 README 和 LLM_INSTALL.md 同步到固定 v0.5.0 安装入口。

验证覆盖采集开关和应用范围、异步加密存储、人工标签与模型结果分离、离线导出，以及现有真实 Rime 候选行为。专用合成验收应用实测了加密写入、预览、导出和可恢复删除，结束后采集关闭。Tev 指令格式仅通过合成分词器验证，未现场训练模型。

本次仅发布源码。源码和公开 Git 历史经过凭据及私有文件检查；不附带采集记录、训练导出、个人词库、凭据、配置、模型权重或本机二进制。仓库保留有明确说明的人工合成评估和测试样例。安装目标为 Apple Silicon / macOS 14+；没有 Apple 公证的即装安装包。模型推荐准确率和应用兼容性仍需分别验证。

English

RimeJev 0.5.0 adds independent, off-by-default local training capture and includes current-line context improvements since the 0.4.1 release.

• Choose the capture directory, included or excluded apps, context/model fields and size limit. Defaults are off, selected-app scope and an empty list. Enabling enhancement never enables capture automatically.
• Link actual pinyin, bounded context, native candidates, displayed order and observed human choices. AES-GCM encrypts raw records with a separate local Keychain key; capture does not load or call extra models.
• Preview, recoverable deletion and offline records/sft exports are available. Tev prompt/completion export requires a compatible user-supplied local tokenizer. Exports are plaintext and need review. Data management disables capture and leaves it off; no automatic upload or training occurs.
• Existing, pasted and paused same-line text can supply context; empty context can also request a model decision. The native first candidate remains stable, distinct recommendations use slots two/three, and native input remains available on failure.
• Both README languages and the dedicated LLM installation guide now select the fixed v0.5.0 tag.

Validation covers opt-in/app scope, asynchronous encrypted storage, separation of human labels from model decisions, offline export and existing real-Rime candidate behavior. A dedicated synthetic app verified encrypted writes, preview, export and recoverable deletion; capture was left off. Tev instruction export was tested with synthetic tokenizers; no live model training was performed.

This is a source-only release. Source and public Git history were checked for credentials and private artifacts. No captures, training exports, personal dictionaries, credentials, local configuration, model weights or local binaries are attached. Documented hand-authored evaluation/test fixtures remain in the repository. The installation target is Apple Silicon / macOS 14+; an Apple-notarized installer is not provided. Model accuracy and application compatibility require separate validation.

Installation and verification / 安装与验证

• README 中文 / English
• LLM_INSTALL.md
• Training capture / 训练采集说明
• Validation evidence / 验证记录

The runtime is unchanged from tested and locally installed source 7cdde9e6ad3cc4375110d241b6bbfc3bec8a131b; publication changes only Markdown documentation. An existing installation of that 0.5.0 build does not need reinstalling for this source release. New installations must run preflight against their exact release checkout.

RimeJev 0.4.1 · All-application defaults

中文

本版本新增所有应用默认设置：在“语境选词设置”中一次选择增强模式、本地模型及云端/本地推荐顺序，未单独配置的应用自动继承，无需逐个开启。

• 支持默认关闭、Jev 云端、本地模型、云端＋本地模型；本地可选 TT、Tev 0.8B 或 Tev 4B。
• 当前应用可单独覆盖；明确关闭也优先于默认。点击“跟随默认”即可移除覆盖并恢复继承。
• 新全局默认初始关闭；升级保留已有应用的模式、模型、顺序、超时及凭据配置，不会把此前某个应用的选择提升为全局设置。
• 修改默认设置会更新已有输入会话；打开设置的原应用关闭后，全局设置仍可保存。安全输入和输入法自身设置不运行增强。
• 中英文 README 和独立 LLM_INSTALL.md 已固定安装版本到 v0.4.1。

**验证：**完整安装前检查、原生构建/签名、安装后检查均通过。回归包含 396 项 Python 测试及 437 组真实 Rime 引擎探针；新增实际设置代码的八组逻辑检查与六种独立加载场景。已安装界面确认全局关闭、现有应用覆盖及“跟随默认”入口。按用户要求，全局激活继承在隔离配置中验证，本机全局默认保持关闭。引擎探针使用受控推荐和合成词库，不代表模型准确率。详见 验证记录。

这是 Apple Silicon macOS 14+ 的源码发布，依赖已正常工作的鼠须管 1.1.2 和万象；不提供 Apple 公证安装包。源码与已验证的本机 0.4.1 运行代码一致，发布准备仅更新文档。仓库及发布包不包含真实凭据、个人词库、本机设置、输入记录、模型权重或本机应用包。

English

This release adds defaults for all applications. Choose enhancement mode, local backend and cloud/local recommendation priority once; apps without overrides inherit the profile automatically.

• Defaults support off, Jev cloud, local, and cloud + local, with TT, Tev 0.8B or Tev 4B as the local backend.
• Per-app overrides take precedence, including explicit off. Follow default removes the complete override and resumes inheritance.
• The new global default starts off. Upgrades preserve existing app modes, model choices, ordering, timeouts and credential configuration; an old app selection never becomes the global default.
• Default changes refresh existing input sessions. Global settings can still be saved after the originating app closes. Secure input and the input method’s own settings remain excluded.
• The bilingual README and dedicated LLM installation guide now pin installation to v0.4.1.

Validation: full preflight, native build/signature and installed smoke passed. Regression includes 396 Python tests and 437 real-Rime probes, plus eight settings logic groups and six isolated loading scenarios using the actual Settings implementation. Installed UI checks confirm global off, preserved app overrides and the follow-default control. Active inheritance was tested in isolated profiles; the live global default stayed off as requested. Controlled engine responses and synthetic dictionaries verify behavior, not model accuracy. See validation evidence.

This is a source release for Apple Silicon macOS 14+, requiring working Squirrel 1.1.2 and Wanxiang installations, rather than an Apple-notarized installer. Runtime code is identical to the tested local 0.4.1 installation; publication preparation only changes documentation. Credentials, personal dictionaries, machine settings, input records, model weights and the personalized app bundle are excluded.

RimeJev 0.4.0 · Dual models, secure keys & Shift feedback

中文

RimeJev 0.4.0 新增云端与本地双模型选词、隐藏密钥设置和 Shift 中英文提示。

• 语境选词支持关闭增强、云端 Jev、本地模型、云端＋本地四种模式；本地可选 TT、Tev 0.8B 或 Tev 4B。
• 双模型同时工作，通过上下排序列表设置优先级。万象首选始终保留；未命中首选的不同推荐按顺序放入第二、第三位。命中首选时标注该模型，另一推荐放第二位；相同推荐合并来源，不重复出现。
• 每个模型独立使用超时设置；无结果、低置信度、迟到或失败时保留原生候选。空格、数字 2/3 与显示顺序保持一致。
• Jev API 密钥使用隐藏输入框，保存到 macOS 钥匙串，保存后不回显。已有凭据服务配置继续兼容；真实密钥不进入配置文件、日志、命令参数或源码。
• Shift 实际切换语言时显示短暂的“中文／英文”独立提示，不抢焦点、不占候选栏。
• 修复快速切换中的凭据重载、本地内存压力误丢弃云端结果、钥匙串授权等待阻塞本地启动，以及首次显示候选前打不开设置的问题。

验证：396 项后端测试、437 项真实 Rime 引擎探针、原生排序/进程/凭据/提示测试通过；本机安装检查和设置交互通过，用户确认真实 Shift 提示。运行代码与已安装验证的版本一致，本次发布只补充正式安装指引和记录。详见 验证记录。

适用于 Apple Silicon、macOS 14+ 的源码版本，不包含 Apple 公证安装包、模型权重、个人词库、本机设置或凭据。请遵循独立的 LLM_INSTALL.md。候选探针使用受控推荐，不代表真实模型准确率；新输入密钥的真实云端认证未纳入本轮验证。

English

RimeJev 0.4.0 adds concurrent cloud/local recommendations, secure API-key settings, and Shift language feedback.

• Four per-app modes: off, Jev cloud, local, or cloud + local. Select TT, Tev 0.8B, or Tev 4B as the one local backend.
• An Up/Down list controls recommendation priority. Preserve the original native first choice; place distinct recommendations in slots 2/3. A first-choice match receives its model badge, with the other recommendation second. Same-word results merge source badges without duplication.
• Independent model deadlines preserve immediate native input. Missing, low-confidence, late, or failed responses do not block selection. Space and numeric selection match the displayed order.
• A secure field stores the Jev API key in macOS Keychain and never redisplays it after saving. Existing credential-service configurations remain compatible. Secrets are not written to settings, logs, process arguments, or source.
• An actual Shift-triggered language change shows a brief Chinese/English indicator without taking focus or replacing the candidate panel.
• Fixes credential reload across rapid switches, local memory pressure discarding cloud results, blocked Keychain authorization delaying local startup, and opening settings before the first candidate panel.

Validation: 396 backend tests, 437 real-Rime probes, native planning/lifecycle/credential/indicator tests, installation checks, and settings interaction passed. The user confirmed actual Shift feedback. Runtime code matches the verified local installation; publication updates documentation only. See validation scope.

Source release for Apple Silicon / macOS 14+, without an Apple-notarized installer. Model weights, personal dictionaries, local settings and credentials are excluded. Follow LLM_INSTALL.md. Controlled engine responses prove placement, not real-model accuracy; live authentication with a newly entered key was not tested.

RimeJev 0.3.0 · Tev & model timeouts

鼠须管 Jev 0.3.0 / RimeJev 0.3.0

中文

新增本地 Tev1 0.8B、4B，与 this-that-model 1.2 和 Jev 云端按应用自由切换。候选来源统一为 ✦ Jev、✦ TT、✦ Tev。

每个模型有独立的超时设置（0.3–10 秒），支持保存和恢复默认。Jev、TT、Tev 0.8B 默认 0.6 秒，Tev 4B 默认 2 秒；计时包含输入防抖及候选采样。原生候选立即可用，超时或模型失败不会阻碍选词。

保留原首选；推荐命中首选时仅加标记，否则移到第二位，其余候选保持相对顺序。模型切换会作废旧请求，等待旧模型释放再加载新模型。下载完成后自动重新加载当前所选模型。

Tev 使用固定版本的独立本机运行组件，不替换系统 Ollama。所有权重、运行组件、凭据、个人词库及本机配置均不随仓库发布。

这是 Apple Silicon / macOS 14+ 的源码版本。安装助手请遵循 LLM_INSTALL.md。模型对中文语境可能给出低置信度或错误判断；详细验证范围见 docs/VALIDATION.md。

English

Adds local Tev1 0.8B and 4B alongside this-that-model 1.2 and Jev cloud, selectable per application. Candidate badges are exactly ✦ Jev, ✦ TT, and ✦ Tev.

Each model now has an independent 0.3–10 second timeout with save/reset controls. Defaults: 0.6s for Jev, TT and Tev 0.8B; 2s for Tev 4B. The budget includes debounce and candidate sampling. Native candidates remain immediately selectable regardless of model latency or failure.

The original first choice remains first. Matching recommendations add only a badge; other genuine candidates move to slot 2 without duplicates. Switching invalidates old responses and waits for model release. Successful downloads reload the still-selected provider.

Tev runs through an isolated pinned local service; the system Ollama installation is untouched. No credentials, personal dictionaries, local settings, weights, or runtime binaries are published.

Source release for Apple Silicon / macOS 14+. Follow LLM_INSTALL.md. Chinese prediction quality varies by model; see docs/VALIDATION.md for measured results and validation limits.

Verification / 验证

341 backend tests, 168 real-Rime probes, native settings/lifecycle tests, build/signature, installed smoke, and actual settings-window save/switch/reset checks passed. Backend line coverage: 89.31%; native coordinator: 91%. Gitleaks and private-metadata checks passed. Fresh manual typing feedback is tracked separately from engine and settings validation.

RimeJev 0.2.4

中文

RimeJev 0.2.4 源码发布，为鼠须管／Rime／万象加入按应用启用的 Jev 云端及 this-that-model 1.2 本地语境推荐。

• 保留原第一候选；模型命中首选时添加来源标记，否则推荐放在第二位，其他候选保持相对顺序。
• 空格选择原首选，数字 2 选择第二项；无结果、低可信或超时继续使用原生输入。
• 提供中英文 README、独立 LLM_INSTALL.md、固定依赖、自包含检查流程和公开合成测试词库。
• 128 项后端／门禁测试、86 组真实 Rime 回归及原生构建通过；公开 Git 历史经 Gitleaks 扫描未发现凭据。

本次只发布源码，不包含个人词库、设置、输入日志、模型权重、API Key 或 Apple 公证安装包。云端需要用户自己的凭据服务配置。请先阅读 LLM_INSTALL.md。

English

Source release of RimeJev 0.2.4: per-application Jev cloud and local this-that-model 1.2 recommendations for Squirrel/Rime/Wanxiang.

• Preserve the native first candidate. Label a matching first choice; otherwise place the recommendation second while preserving the remaining relative order.
• Space selects the original first word; 2 selects the second. No match, low confidence, or timeouts leave native input available.
• Includes bilingual README, dedicated LLM_INSTALL.md, pinned dependencies, a standalone quality gate, and synthetic public Rime fixtures.
• 128 backend/gate tests, 86 real-Rime regression probes, and the native build passed. Gitleaks found no credentials in published Git history.

Source only: no credentials, personal dictionaries/settings/input logs, model weights, or Apple-notarized installer. Cloud mode requires a user-configured credential service. Read LLM_INSTALL.md before installation.