Releases: reshuibuduo/tmcra-plugin-deepseek-harness
Release list
v1.0.0-rc.1
TMCRA unified 1.0.0-rc.1
- Visual memory workspace with local Writer/organizer configuration, knowledge-base/graph inspection, task continuity, session privacy modes and recall budgets.
- Source corrections require interactive user confirmation in chat; declined or unsupported confirmation does not mutate memory.
- Account-free Windows x64 local-installation preview: after loading the plugin, run
dsh-tmcra-memory local-install, choose a profile, and confirm installation. The package carries the checked backend; the installer prepares private Python, verifies/downloads models, creates a local identity and starts the service. Restart DSH after switching to local mode. - Local memory requires no TMCRA server/account. First-time downloads require internet; memory execution then uses the user's computer. Stale cloud requests and saved cloud Writer/organizer execution are blocked while local mode is active. Explicit advanced
TMCRA_CONFIG_FILEoverrides remain authoritative. - The setup page has no fake credential and cannot read/write memory or change session controls. User memory and existing credentials are preserved.
Install and verification
Use the attached dsh-tmcra-memory-1.0.0-rc.1.tgz and SHA256SUMS; no npm publication is claimed. The existing DSH installation instructions accept this release URL.
Type checks, 20 tests plus provider contracts, real DSH 0.1.1-rc.2 package install/profile/Web boot, and the account-free setup boundary passed. Release workflow.
Full-local model acceptance remains partial: synthetic CPU ingest/raw recall passed; complex compilation timed out at 600 seconds. Organizer/full-service restart and balanced/high-tier hardware acceptance are pending. See runtime instructions and hardware estimates. The installer currently supports Windows x64. A cloud-hosted Agent main model may still receive recalled content; a fully offline Agent also needs a local host model.
本次为统一大版本预览：免账号、本机运行、首次下载需联网。完整离线模型链路仍有待验收项，详见上方说明；生产服务器和真实用户记忆未改动。
v0.1.7
Full Changelog: v0.1.6...v0.1.7
v0.1.6
v0.1.5
v0.1.4
v0.1.3
dsh-tmcra-memory v0.1.2 - DeepSeek Harness 0.1.1-rc.2 Compatibility (Technical Preview)
Compatibility release for the latest official DeepSeek Harness preview, @deepseek-ai/dsh 0.1.1-rc.2.
What changed
- Updated the complete DSH development/runtime contract set from
0.1.0-rc.6to0.1.1-rc.2. - Added an end-to-end compatibility probe that packages the plugin, installs it through the real
dsh plugincommand into a clean profile, verifies bundle composition, and boots DSH Web to HTTP readiness. - Added a reviewed pnpm native build-script allowlist required by the latest DSH package.
- Bumped the TMCRA plugin client identity to
0.1.2.
Install
dsh plugin --profile web add https://github.com/reshuibuduo/tmcra-plugin-deepseek-harness/releases/download/v0.1.2/dsh-tmcra-memory-0.1.2.tgz
dsh plugin --profile web exec dsh-tmcra-memory login
dsh --profile web --dump-config
dsh webOn Windows, local archive paths containing spaces or non-ASCII characters are still mis-parsed by DSH rc.2. The release URL above is unaffected; for local installation, copy the archive to a short ASCII-only path first.
SHA-256: 2858c4d9ceb58a4e79c79b8944fb05fa5aa82ba5e305e3c2550a2e59e2c44a6e
dsh-tmcra-memory v0.1.1 — Cross-App, Cross-Conversation Project Memory (Technical Preview)
dsh-tmcra-memory 0.1.1
跨软件、跨会话继续同一个项目。切换到 Codex、DeepSeek Harness 或其他已接入 TMCRA 的工具时，无需重新介绍项目背景。
DeepSeek Harness 的 TMCRA 长期记忆插件技术预览版。
跨软件、跨会话衔接
同一个 TMCRA 账号下，每个项目都有独立且稳定的项目记忆。两端识别到同一个项目身份时，Harness 与 Codex 会读取和更新同一份项目记忆：回答前召回相关的用户要求、决策、偏好、Agent 进度、实现结果、测试结论和下一步；回答完成后，再把新的用户陈述与 Agent 结果分角色写回并保留来源。
跨软件衔接只使用已经进入 TMCRA 的记忆。.tmcra/project.json、Git 项目标识和账号下发的项目 Scope 前缀用于维持项目边界，不相关的项目不会被合并；session_id 只承担项目内部的来源追踪。
本版更新
- 公开包、仓库与 CLI 命令统一更名为
dsh-tmcra-memory。 - 增加 TMCRA 账号连接流程：
login、status、logout。 - 登录采用 PKCE 设备授权，用户无需手工复制 API key。
- 自动召回用户全局记忆与当前项目记忆，并注入 Harness 会话。
- 用户内容与 Agent 回答分角色写回，保留项目、会话和 Agent 来源。
- 主 Agent 与子 Agent 可以共享项目记忆，同时保留各自身份与委派来源。
- 将项目协作记录持续整理为可检索的记忆库、知识图谱和可复用工作经验。
- 召回异常默认不阻断回答；写入失败进入本地持久化待写队列。
最终验收补充（2026-08-14）
- 生产 API 全流程通过：两个完整回合写成四条分角色记忆，第二个独立 Harness 会话成功召回用户检查点与 Agent 进度。
- 服务端账本记录两次写入事件（共 53 个估算写入 Token）和两次召回请求。第一次面向全新 Scope，按设计容错放行；第二次完成有效的跨会话召回。
- TMCRA 记忆处理链路记录 7,894 个模型 Token，已知模型 API 成本为 ¥0.00。回答端采用记录型测试 Adapter，没有消耗外部回答模型的 Token 或费用。
- 验收结束后，两个一次性 Session 均已删除；相关记忆记录、服务消息、Session 记录、活动 Base/Delta 索引指针和外键错误均为 0。
- 测试设备连接、上游范围令牌与本地 Harness 凭据均已注销或移除。
使用说明
需要注册 TMCRA 账号。发布初期免费使用，暂不设置记忆写入与召回额度上限；政策如有调整会提前公告。
安装与账号连接命令请查看 中文 README。
本次仅更新公开说明，v0.1.1 Tag 与安装包保持不变。
SHA-256: f4cb0126289a3e73b3c88284f9dba66c91517e216f54254780d65cbb23c93337
兼容性：DeepSeek Harness 0.1.0-rc.6，TMCRA API 0.2.2。
v0.1.0 — Technical Preview
TMCRA Memory for DeepSeek Harness
Technical preview for DeepSeek Harness 0.1.0-rc.6 and TMCRA API 0.2.2.
Included
- account-global and current-project recall before the first model request;
- durable evidence injection into the Harness session log;
- role-separated user and assistant writeback;
- project, session, agent, parent-agent, and delegation provenance;
- credential-pattern redaction and a crash-safe writeback outbox;
deepseek_harnessusage attribution.
Verification
- clean tarball installation into an isolated Harness Web profile;
- composed configuration contains the TMCRA bundle;
dsh webstarts and returns HTTP 200;- typecheck, unit tests, build, package inspection, and production dependency audit pass;
- two-conversation TMCRA acceptance verifies cross-conversation recall and role-separated writeback.
The acceptance run used a recording model adapter, so this preview validates the Harness/TMCRA lifecycle without claiming a live DeepSeek-provider answer.
See the English and Chinese README files for configuration, security boundaries, and current limits.