Repository navigation
Releases
Recaloom XS 0.2.0
Recaloom XS 0.2.0
换一个会话,也能从项目上次保存的位置继续。
这是 Recaloom XS 的首个正式版本。稳定范围是本地项目级 CLI:保存与恢复经确认的目标、限制、决定和下一步,检查引用变化,进行有明确授权的交接。一个助手即可使用,不需要 Harness、不需要给 Recaloom 配模型。
从 Alpha.7 升级后有什么变化
- 程序和使用说明一起安装。
doctor返回匹配当前程序的 Skill 路径和身份,减少拿错说明、接错项目的情况。 - 项目入口更完整。 在支持的平台上,先预览规则变更,再接入;支持暂停、启用、升级和移除,保留原有规则与项目资料。
- 记忆不只是一段摘要。 项目偏好与可复用方法、引用审阅、保存授权及交接状态分别管理。文件变化时先说明问题,不自动冒认任务完成。
- 失败时给下一步线索。 包括安装运行时检查、受管宿主诊断,以及恢复内容过长时返回所需字符数。不会为了塞进上下文而静默截断限制条件。
- 首次使用入口重写。 提供 macOS/Linux、Windows PowerShell 安装命令、虚构项目练习和脱敏反馈入口。
安装与开始
需要 Python 3.10+(包含 venv 和 pip)。使用本页的 glom_continuity-0.2.0-py3-none-any.whl,不要把旧 Alpha 包当成 XS。
已有项目不要重新初始化。先保留旧安装和项目资料,再按升级流程核对程序身份。网页更新不会自动升级电脑中的程序。
本次验证
最终候选源码提交:b1191088c5d5f513393c49bcdf397ce2e44e54b4。
固定候选 CI 的五个作业通过:
| 环境 | 观察结果 |
|---|---|
| Ubuntu 24.04 / Python 3.10、3.12 | 每条车道 358 项,0 失败、0 错误、0 跳过 |
| macOS 14 / Python 3.12 | 358 项,0 失败、0 错误、0 跳过 |
| Windows 2022 / Python 3.12 | 358 项,其中 42 项为已声明的平台跳过,0 失败、0 错误 |
| macOS / Harness 0.1.5-rc.1 / Node 26.6.0 | 35 项 Node 检查及 1 项受管宿主生命周期检查通过 |
CI 测试的合并树与上述候选树一致。另在 macOS arm64 / Python 3.14 上,对本页同哈希 wheel 做了独立新安装、首存、跨进程恢复、引用变化拒绝、项目入口暂停/启用/移除、旧候选入口升级及卸载保留资料检查。测试脚本自身的一次字段读取错误已保留记录并更正,不作为产品通过项。
源码附件逐文件对齐候选提交;wheel 的程序与随包指南对齐同一源码。SHA256SUMS 用于检查附件字节;它不单独构成发布者身份认证。
使用范围
基础 CLI 本地运行,无第三方运行时依赖、无模型调用、无遥测、无账号要求。安装下载会联网;云端助手读取记录仍受其自身隐私和计费规则约束。建议团队先用虚构资料练习。
MCP、宿主规则加载、自动保存及 Harness 是分层可选能力,不能由 CLI 安装成功推断它们都已接通。Windows 的安全规则自动写入和受管 Harness 不在本次支持承诺内;Harness 验证仅对应表中的指定版本,不代表新版或其他宿主。具体范围见支持矩阵。
本版本不是跨电脑同步服务,不会自动读取所有聊天。测试通过不等于任务语义正确,也不代表已有外部团队采用、达到零错误或优于所有竞品。
遇到问题请提交脱敏试用反馈:系统和 Python/助手版本、失败步骤、错误码、预期与实际结果。不要上传密钥、公司源码或完整记忆数据库。
Recaloom Alpha.7 — Harness 自动恢复 / Opt-in project recovery
Recaloom Alpha.7 — 项目恢复进入 Harness 工作过程
这个版本新增 DeepSeek Harness 的可选自动恢复:明确绑定一个项目并启用后,在 Agent 开始下一步前读回进度,检查引用是否变化。换会话不用每次再贴同一段恢复指令。第一次接入默认暂停,可用 /recaloom status、/recaloom resume、/recaloom pause 控制。
它使用独立的本机 Web 入口,不改动日常 Harness 配置。基础 CLI 和 MCP 仍可供有相应权限的其他助手使用。自动恢复目前只验过指定 Harness/macOS 路径,不是所有 Agent 安装即自动记忆,也不会自动保存普通聊天。
这次还修了什么
- 退出宿主前重新核对暂停回执,避免等待期间状态改变后仍执行旧的退出决定。
- 数据库异常中断后,只读恢复明确返回
STORAGE_RECOVERY_REQUIRED;先备份,再单独执行 CLIrecover-storage。不暗中写库、不重新初始化、不把修复权限加进 MCP。 - README、安装命令和下载版本统一。新环境安装新版;旧附件不覆盖,旧项目不要再次初始化。
安装
需要 Python 3.10+。Mac/Linux 在工具目录创建新的专用环境:
python3 -m venv .recaloom-alpha7
.recaloom-alpha7/bin/python -m pip install --no-index --no-deps "https://github.com/fugui6688661/glom-continuity/releases/download/v0.1.0-alpha.7/glom_continuity-0.1.0a7-py3-none-any.whl"
.recaloom-alpha7/bin/glom-continuity --version版本应显示 0.1.0-alpha.7。完整安装、Windows、备份与卸载。wheel 是程序;需要 Skill、文档和适配说明时,再取同版本具名 ZIP,不必重复安装程序。Harness 安装。
验证范围
发行提交:5d3e888039ff8d7db3b397c8f301fa515dedecb7。最终 PR head:fa9ed1103aba8c429ff622bf6a55e996c00fb788。五项 CI:Linux Python3.10/3.12、macOS Python3.12 各 158 通过;Windows 136 通过、22 项预声明跳过;独立 Harness/macOS 检查包含 35 项 Node 测试及一个真实宿主生命周期复合测试,均通过。
具名 ZIP 的本机固定源另通过 158 项;wheel 在全新环境安装后,版本、无模型演示以及从无关目录运行的真实宿主生命周期复验通过。两项 P2 问题经独立审查、失败复现和修复复验关闭。
这是开发者预览,不是稳定版或模型能力对比结论。未验证其他 Agent 的自动生命周期、Windows Harness 控制、跨设备实时同步或自动保存。真实宿主不是模型推理;这轮没有付费模型调用。已知重开后的中央工作区选择器显示不一致仍见验证记录。
发布后又匿名下载三个附件:哈希与封存文件一致,ZIP 内 86 项清单全部相符。下载所得 wheel 在新环境安装、显示正确版本、完成九次调用的合成演示;doctor 和 resume 成功且未改变项目数据库,卸载测试环境中的工具后项目记录仍保留。此测试只触及虚构项目,不代表你的客户端已经升级。
使用本页三个具名附件,按 SHA256SUMS.txt 核对;GitHub 自动生成的 Source code 不等于发行 ZIP。包内 published: false 是封存时的构建记录,不会为发布改写既定字节,公开状态以本 Release 为准。文件哈希只校验字节,不替代来源信任或任务验收。
English summary
Opt-in, lifecycle-native project recovery for DeepSeek Harness. Bind a project, then explicitly resume; the adapter rechecks session, revision and references before delivering memory. First attachment is paused. Ordinary chat is not autosaved, and other assistants do not acquire automatic recovery merely by installing the CLI or MCP tools.
This developer preview adds a separate local Web home, native status/resume/pause controls, guarded shutdown and explicit CLI-only SQLite crash recovery. It does not migrate your normal profile or provide a sandbox for all Harness tools.
The final five-job CI covers portable Linux/macOS/Windows checks plus a required official-SDK macOS lane. Windows has 22 declared platform skips. Named ZIP/wheel assets received fixed-source and fresh-install checks; no paid-model quality benchmark is claimed. Use the matching assets and checksums above, back up projects before upgrading, and do not reinitialize existing project memory.
Recaloom alpha.6 · 项目记忆与接续
Recaloom alpha.6
把项目留在本地,换会话后接着做。一个助手就能用,需要协作时再交接给第二个助手。
这一版将此前只在开发源码中的能力装进新的下载包。原 glom-continuity 包名、命令和项目目录保留,已有项目不要重新初始化。
相比 alpha.5,多了什么
- 项目恢复
resume:一次读回目标、限制、进度、引用检查和下一步;默认只读。 - 习惯与流程记忆:显式保存经你确认的方法,按任务关键词读取。不会自动挖掘聊天或训练模型。
- 安装诊断
doctor:看清实际调用的程序、版本及存储状态,避免新版说明配着旧程序。 - 成果回存
return-work:把接手方的文件关联回原交接,原助手能查到产物与修订。回存不是业务验收。 - 更直接的首用流程:中英文 README 有固定安装命令和可运行演示;补齐独立环境、匹配 Skill、旧项目升级和卸载保留数据的说明。
- 权限标注回归:真实 stdio
tools/list验证默认 6 个只读工具、启用写入后的 12 个工具及四类 annotations。标注不能代替权限控制。
安装
在工具目录中新建专用环境(Python 3.10+,不要覆盖已有环境):
python3 -m venv .recaloom-alpha6
.recaloom-alpha6/bin/python -m pip install --no-index --no-deps "https://github.com/fugui6688661/glom-continuity/releases/download/v0.1.0-alpha.6/glom_continuity-0.1.0a6-py3-none-any.whl"
.recaloom-alpha6/bin/glom-continuity --versionWindows、便携版、可选 MCP 和升级方式见安装说明。wheel 安装程序;完整 ZIP 另含 Skill、文档和测试。需要助手读取 Skill 时,取得同版本 ZIP 并保留目录结构。旧项目先停写备份,保留原项目路径,不删库、不再 init。
本次验证
- 固定 ZIP:52 个白名单源文件,清单与字节校验一致,111 个本地文档链接可解析。
- 本机固定包:111 项测试通过、零跳过;离线 wheel 构建及独立安装/运行/卸载保留数据通过。
- GitHub CI 35981668797:Linux Python 3.10 / 3.12、macOS Python 3.12 各 111 项通过;Windows Python 3.12 为 102 项通过、9 项明确的 POSIX 专用跳过。四组运行代码、测试与构建输入哈希均与本地候选一致。
- 独立合成升级:alpha.5 保存并交接,alpha.6 恢复、接手、回存、新进程读回;输入和项目身份保留。旧版可以读普通检查点,但不显示新版成果关联;旧版写回和降级未验证。
- 原两次本机 pip-wheel 超时保留。未改代码、断言或超时限制,固定包及独立复测随后通过;原超时根因仍未确认。升级脚本最初因比较动态检查时间而失败,后续只读复核单独记录,不抹去原失败。
发布提交:f77c3f98c0923b42173313ea927066c71647f39f,来自 PR #4。固定包与该提交内容一致。
**发布后复核:**三份具名附件均已匿名重新下载,SHA-256 与固定候选一致。下载的 wheel 在全新独立环境安装成功,版本/帮助正确;9 次调用的协议演示、7 次调用的项目记忆演示及只读恢复通过,随后卸载并确认项目数据库字节不变。全部为本机合成资料,模型调用为 0;不是宿主或真人首用验收。
附件
| 文件 | SHA-256 |
|---|---|
glom-continuity-0.1.0-alpha.6.zip |
0e0bcac77fcc2e65901f8688ea2f625838743e12e0fd5b1e30226a595ec0ac7d |
glom_continuity-0.1.0a6-py3-none-any.whl |
3136512478459376c8a08ea02d53ec7a1b41d010c324d07b0ff0426c98493826 |
附 SHA256SUMS 供校验。使用上述具名附件;GitHub 自动生成的 Source code 压缩包不是同一个构建包。包内 manifest 的 published:false 保留构建时状态,实际发行状态以本 Release 为准。
适用范围
这是开发者预览,不是稳定版。CLI 不调用模型、不上传项目;云端助手仍有自身数据处理与计费规则。MCP 仅本地 stdio,默认只读;没有自动调度、远端 HTTP 或跨设备同步。未新做付费模型实验,历史 alpha.5 模型接力不算本版本的宿主认证。完整效果对照、外部首次使用及更多宿主仍待验证。
遇到问题请提交脱敏试用记录,写明版本、系统、助手、卡在哪一步即可。不要上传真实项目库、密钥或完整聊天。
English: Alpha.6 packages the previously source-only project memory, one-call read-only resume, runtime diagnosis and linked result return. It retains existing storage and compatibility identifiers, adds a first-use checklist and version-specific installation, and verifies raw MCP safety annotations. Frozen local tests passed 111/111. Hosted Linux/macOS passed 111 each; Windows passed 102 with nine declared POSIX-only skips. This remains a developer preview, not universal host compatibility or a stable-production guarantee. See the English README and version-bound verification record; preserve existing projects rather than reinitializing them.
0.1.0-alpha.5 · Developer preview
glom-continuity 0.1.0-alpha.5
Local project checkpoints and reviewed handoffs for AI assistants.
换个助手,不必丢掉项目已经走过的路。保存目标、决定、限制、未决问题与下一步;
新会话接手前,重新检查输入文件。项目记录留在你选择的文件夹里。
Early developer preview — not a stable release. The files below are the
unchanged alpha.5 candidate. Package documentation contains dated pre-release
snapshots; this note records the later verification. Publication is not proof
that the remaining validation has passed.
This is a preview-distribution decision, not a claim that the original full
pre-public research plan was completed. External first-use feedback and the
matched-model comparison remain open; stable-release readiness is not granted
by publishing this preview. Current scope is recorded in
PROVENANCE.
Try it / 先试一个小项目
Python 3.10+ required. Download and extract the ZIP, open its folder, then run:
python3 -B scripts/smoke_demo.py --output ./my-first-demoOn Windows use py -3 instead of python3. Open the generated Chinese
my-first-demo/演示结果.md. This first demo is a local protocol replay,
not a live-model conversation. The CLI itself uses only the standard library;
it does not call a model or upload your files.
For installation, local MCP configuration and removal, see the
current installation guide
or the dated INSTALL.md in the ZIP. The optional MCP adapter requires the separately installed official
Python MCP SDK. Start with synthetic data, not company documents or secrets.
Cloud assistants may send recovered context to their own model providers.
What has been verified
- A real synthetic Codex → DeepSeek Harness → fresh Codex recovery on macOS
using this CLI/MCP: task state restored, an order summary written and saved,
then read back in a fresh Codex session, as observed by the maintainer.
A separate reviewer recomputed the business result. Original inputs stayed unchanged and missing
quantity/currency/tax were preserved as unknown. Earlier failed attempts were
retained; this was a successful recovery, not an uninterrupted first run. - Maintainer-recorded clients/configuration: Codex 0.154.0-alpha.6.2 /
gpt-6-astra low; DSH 0.1.5-rc.1,
official provider aliasdeepseek-flash, low. This does not attest the
provider's hidden model revision or support for every client. - Four GitHub-hosted CI jobs passed at commit
c3b9e4f75b530b9e061415672c503fbdf0a53253: Linux Python 3.10/3.12 and
macOS Python 3.12 each 67 tests; Windows Python 3.12 59 passed and 8
explicitly POSIX-only cases skipped. Not Windows ACL or physical power-loss
certification. CI run - Independent fixed-ZIP review checked all 40 shipped files and reachable
source history against the allowlist; 41 standard-library tests passed in
a fresh local environment. Wheel RECORD entries were separately checked.
Known limits / 尚未验证
- No external human first-use study or complete matched-model comparative
trial yet. Offline mechanical comparison is not evidence of better model
performance, less re-explanation, or token savings. - One live handoff preserved stale prose in
next_action; the receiving
assistant correctly used the current receipt and review protocol. The tool
does not automatically correct the truth or freshness of every sentence. - Assistants must have authorized access to the same project. No automatic
cross-computer synchronization; no automatic access to all chat products. - Recipient labels are not authenticated identities. A matching file hash
is not a semantic quality verdict or permission for external actions. - The current CI workflow is limited to reviewed private-repository runs;
publishing these files does not by itself enable ongoing public CI.
这版可以用于非敏感项目试验,不宣称全助手兼容、成熟 AGI、全面领先或稳定生产保障。
试用反馈请附系统/Python/助手版本、失败步骤、错误码与虚构复现;不要附密钥或原始聊天。
Integrity
MIT licensed. Only this independent utility is distributed; no private project
memory, credentials, conversation logs or GLOM application source is included.
c7e78af576d4ec2a14bb669d6e4f66f76b3a80b50f0996537595461f72656dfa glom-continuity-0.1.0-alpha.5.zip
5a63380c81df229428db5d8d5bdcf384b62358f33a372957eb643705ff117997 glom_continuity-0.1.0a5-py3-none-any.whl