Skip to content

Mebular v0.3.2

Choose a tag to compare

View all tags
@github-actions github-actions released this 06 Nov 05:46
v0.3.2
118028a

English · 中文

Mebular v0.3.2 — Memory Runtime stability + Fleet orchestration

Mebular v0.3.2 is the stability release for Memory Runtime, with Fleet black-box orchestration included. The public MCP surface stays at 29 tools, and graph/event/sync formats are unchanged. There is no data migration.

Highlights

  • Memory Runtime lifecycle — managed, attached and manual sidecars now have explicit ownership and disable semantics. Attached/manual processes are never killed or replaced by Mebular; an unclean disable is reported as degraded / disable_incomplete, not a fake stopped.
  • Data lifecycle and privacy — staging, details, outbox, logs and remote-model requests use one minimization/redaction policy. Details follow role/kind retention (conversation 30 days, tool_output 7 days, document_slice 90 days, explicit pinned), with bounded idle cleanup and evidence-aware association.
  • Honest state model — accepted, pending, queued, committed, failed, dead and degraded are distinct. HTTP 202 and outbox acceptance are never rendered as graph commits; partial writes are reported as partial.
  • Offline recovery — stable idempotency keys survive restarts; ambiguous timeouts require authoritative readback; offline Compact no longer writes a title-only placeholder; Checkpoint high-water, expectedRevision, pending views and same-domain evidence validation are enforced.
  • Host alignment — DSH, Hermes and OpenCode/OpenChamber share the same pending/queued/committed, retry, mode-mismatch and checkpoint semantics, while keeping their documented host-specific limits.
  • Operations — sanitized JSONL logs, mebular memory-runtime logs, doctor, and outbox list/show/rebuild/purge. Diagnostics contain no tokens, message bodies or real home paths.
  • Fleet orchestration — capability/role cards, immutable plans, deterministic materialization, executable routing, join/MR gates, unified executor/adapter layer and the loopback OpenChamber runner.

Fixes

  • Capability idempotency is now single-flight for concurrent requests with the same composite key and body; this closes the timeout-retry duplicate-write race behind the original 10,023/10,000 10k failure.
  • Recall now queries every canonical detail kind, so tool_output and document_slice details expand instead of incorrectly reporting detail_pending_sync.
  • MCP graceful close no longer force-closes an already-admitted HTTP request before it can answer, so the configuration rollback request can return the contract's 409; long-lived connections are still closed at the bounded deadline.
  • DSH, Hermes and OpenCode no longer turn a 202/pending/queued response into a committed record or checkpoint.

Verification

  • Full Jest: 307 suites passed / 1 skipped; 3,389 tests passed / 1 skipped; 0 failed.
  • Coverage: statements 89.16%, branches 77.92%, functions 90.26%, lines 92.20%.
  • Hosts: DSH 115, Hermes 57, OpenCode 41 and OpenCode checkpoint-recovery 32 host E2E checks; DSH memory runtime 37/37; immersive journey 84/84.
  • Scale: 10,000 MemoryRecords + 10,000 details, 0 extras, 200/200 read samples, 100% sampled detail-expand coverage; 6/6 failure/interruption scenarios passed.
  • Real model: Qwen3-4B Q4_K_M on macOS x64 with a test-only mirror; official size/SHA-256 verified, llama.cpp b12017, non-degraded extract/summarize/planQuery.
  • Release: all eight npm packages are visible at 0.3.2; the release workflow passed verify, DSH, macOS/Windows/Linux packaging and visibility gates.

Compatibility

  • Public MCP tools remain 29; no new public runtime_* tools.
  • No graph/event/sync format change and no migration; 0.2/0.3.0 IDs, namespaces, tags and event history remain valid.
  • The sequential v0.3.0 → v0.3.2 → v0.3.0 single-home upgrade/rollback drill passed.
  • Run the same version on every device. Mixed-version concurrency and a direct 0.2.4 upgrade/rollback are not verified.
  • Desktop packages are unsigned; macOS Gatekeeper / Windows SmartScreen confirmation is required. Auto-update is not included.

Known boundaries

  • The real-model canary covers one pack, one platform and a test-only HuggingFace mirror; it does not prove the default HuggingFace transport, other packs/platforms, or model quality.
  • A real-host + real-model prompt-injection test and recall-quality metrics are not part of this release evidence.
  • DSH managed-cli-missing is environment-blocked on a macOS host with an installed Mebular.app; the CI DSH matrix passes.
  • The 10k run is a single-host scale and correctness test, not an SLA or a concurrency benchmark.

Install

npm install -g @mebular/core@0.3.2
npm install -g @mebular/memory-runtime@0.3.2

Eight packages: core, service, skill, mcp, fleet, memory-runtime, opencode-plugin, dsh-plugin.

Full release notes: CHANGELOG.md.

中文

Mebular v0.3.2 — Memory Runtime 稳定性收口 + Fleet 编排

v0.3.2 是 Memory Runtime 的稳定性版本，同时包含 Fleet 黑盒任务编排。公共 MCP 仍为 29 个工具，图/事件/同步格式不变，无数据迁移。

重点变化

  • Memory Runtime 生命周期：managed / attached / manual sidecar 有明确的所有权和关闭语义；attached/manual 外部进程不会被 Mebular 杀掉或替换；无法确认干净关闭时显示 degraded / disable_incomplete，不会伪装成 stopped。
  • 数据生命周期与隐私：staging、详情、outbox、日志和远程模型请求使用统一的脱敏与数据最小化策略；详情按 conversation 30 天、tool_output 7 天、document_slice 90 天或显式 pinned 保留，并执行有界空闲清理和证据关联。
  • 诚实状态模型：accepted、pending、queued、committed、failed、dead、degraded 严格分开；HTTP 202 和 outbox 接受不再被当成图提交；部分写入明确报告为 partial。
  • 离线恢复：稳定幂等键跨重启保留；ambiguous timeout 要求权威回读；离线 Compact 不再写标题占位；Checkpoint 启用 high-water、expectedRevision、pending 视图和同域证据校验。
  • 三宿主对齐：DSH、Hermes、OpenCode/OpenChamber 在 pending/queued/committed、重试、mode mismatch 和 checkpoint 语义上保持一致，同时保留各自明确的宿主限制。
  • 运维诊断：脱敏 JSONL 日志、mebular memory-runtime logs、doctor 和 outbox list/show/rebuild/purge；诊断不包含 token、消息正文或真实 home 路径。
  • Fleet 编排：capability/role cards、不可变计划、确定性 materialization、可执行路由、join/MR 门禁、统一 executor/adapter 层和 loopback OpenChamber runner。

关键修复

  • Capability 幂等写入现在对同一 composite key + body 的并发请求做 single-flight，关闭了原始 10k 10,023/10,000 失败背后的 timeout-retry 重复写竞态。
  • Recall 现在查询所有 canonical detail kind，tool_output / document_slice 可以正常展开，不再错误报告 detail_pending_sync。
  • MCP 优雅关闭不会再提前切断已接受的 HTTP 请求，配置回滚请求可以返回契约要求的 409；长连接仍会在有界截止时间后关闭。
  • DSH、Hermes、OpenCode 不再把 202/pending/queued 当成 committed 记录或 checkpoint。

验证结果

  • 全量 Jest：307 suites passed / 1 skipped；3,389 tests passed / 1 skipped；0 failed。
  • 覆盖率：statements 89.16% / branches 77.92% / functions 90.26% / lines 92.20%。
  • 宿主：DSH 115、Hermes 57、OpenCode 41、OpenCode checkpoint recovery 32 个 host E2E 检查；DSH memory runtime 37/37；沉浸式旅程 84/84。
  • 规模：10,000 MemoryRecords + 10,000 details，0 extras，200/200 读取样本，采样的 detail-expand 覆盖率 100%；6/6 失败/中断场景通过。
  • 真实模型：Qwen3-4B Q4_K_M / macOS x64 / 测试专用 mirror；官方大小/SHA-256 校验、llama.cpp b12017、extract/summarize/planQuery 非 degraded。
  • 发布：8 个 npm 包均可见 0.3.2；release workflow 已通过 verify、DSH、macOS/Windows/Linux 打包和 visibility 门禁。

兼容性

  • 公共 MCP 工具仍为 29 个，没有新增公共 runtime_* 工具。
  • 图/事件/同步格式不变，无数据迁移；0.2/0.3.0 的 ID、namespace、tag 和事件历史继续有效。
  • 单 home 的顺序 v0.3.0 → v0.3.2 → v0.3.0 升级/回滚演练通过。
  • 所有设备应运行同一版本；mixed-version 并发和 0.2.4 直升/降级未验证。
  • 桌面包仍为 unsigned；macOS Gatekeeper / Windows SmartScreen 需要按文档确认。不包含自动更新。

已知边界

  • 真实模型 canary 只覆盖一个模型包、一个平台和测试专用 HuggingFace mirror；不证明默认 HF transport、其他 pack/平台或模型质量。
  • 真实宿主 + 真实模型 prompt-injection 和 recall 质量指标不在本次发布证据内。
  • DSH managed-cli-missing 在安装了 Mebular.app 的 macOS 本机受环境限制；CI 的 DSH matrix 已通过。
  • 10k 是单机规模/正确性测试，不是 SLA 或并发压测。

安装

npm install -g @mebular/core@0.3.2
npm install -g @mebular/memory-runtime@0.3.2

8 个包：core、service、skill、mcp、fleet、memory-runtime、opencode-plugin、dsh-plugin。

完整 release notes：CHANGELOG.md。

Assets 8
Loading