v2.8.4

v2.8.4 (2026-08-19)

DeepSeek Harness 注入延迟到首个 promotion 信号，兼容锚定预设

• DSH bundle 不再在 agent/session-start 同步注入 using-aegis bootstrap，避免污染首轮请求的零注入基线（修复 issue #19 的锚定复现失败）。
• 每个 startup/resume/clear/compact 边界只 arm 一次待注入；在会话首个 durable promotion 信号（tool/call 或 assistant/message）之后才通过 agent.inject 交付一次，保证每个 gated epoch 的首个模型请求保持零注入。
• compaction/end 会重新 arm；subagent 与无稳定 session.id 的会话跳过；explicit activation mode 行为不变。
• 更新 docs/README.deepseek-harness.md 及 host compatibility / known limitations / release checklist / trigger-health 文档；确定性测试与 62 项 e2e boundary check 通过。

v2.8.3

v2.8.3 (2026-08-18)

全局用户规则收敛为单一路由前缀

• 退役 Lite / Advanced 双层手工 profile，统一为一份功能语义、两份语言镜像的全局路由前缀。
• 新前缀只负责 activation、最小 skill 加载、fast path 与上下文重入；具体规划、调试、TDD、验证、治理和输出继续由已加载的 Aegis skill 按需负责。
• 新用户只需把前缀添加到现有全局用户规则开头，无需改动原有内容；旧用户只需替换原 Aegis Lite / Advanced 区块。
• Activation Mode 保持默认 auto、可选 explicit；TDD Mode 保持默认 off、可选 auto。
• Ubuntu CI、模板/边界/版本契约及 Codex 显式与自然语义触发 smoke 已验证。

v2.8.2

v2.8.2 (2026-08-17)

brainstorming 设计决策校准实践回填

从 Aegis-code 将设计决策校准实践回填到 brainstorming 技能，纯方法文案追加，不改变 runtime / completion authority 边界：

• 新增 Route Fixtures 校准表与严格路由优先级：盘问/拷问/审问路由到 Grilling Mode，显式目标意图路由到 goal-framing，review 路由到 requesting-code-review；文件数不再作为设计信号。
• 新增 Role And Authority Contract：Agent-owned 决策（仓库调查、文件组织、验证机制、可逆实现结构）由 Agent 直接解决；User-owned 仅保留产品行为、不可逆/生产/公共/敏感影响与不可发现事实；每条用户提问先通过“换一个答案会改变哪个设计边界”测试。
• Grilling Mode 结尾新增结构化 Challenge Result（存留/否决假设、所需新证据、设计变更、残余风险、返回状态），收敛质量可验证。
• 新增 8 个 Software Scenario Profiles（greenfield / existing-system / refactor / public-contract / persistence / ui-workflow / security / operational-release），按场景加载相关 lens。
• 新增 Design Probe 一次性探针纪律与 Design Ready / Design Complete 收敛条件；Design Complete 保持 method readiness，不构成完成 authority。
• 原 HARD GATE、spec 审批点、workspace 初始化、compact output contract 与 grilling 模式语义保持不变。

Aegis v2.8.1

更新内容

• DeepSeek Harness 增加原生 session-start 引导，在启动、恢复、清空与压缩后重新建立 Aegis 路由上下文。
• DSH 的通用/极简安装默认分流到 profile 插件安装，直接子目录方式仅保留为显式兼容模式。
• 保持 Aegis Method Pack 边界：不添加硬工具守卫，也不将宿主执行提升为治理权威。
• 补齐 DSH 路由、安装分流、宿主边界与隔离插件集成测试。

验证

完整 Linux E2E 门禁、治理契约、Layer 2/3 场景及 DSH 0.1.0-rc.6 隔离插件集成均通过。

Aegis v2.8.0 — Native DeepSeek Harness Bundle

Highlights

• Added a native DeepSeek Harness plugin bundle through the package-level dsh.bundle.patch contract.

• Added a thin Cordis adapter that delegates discovery to Harness's official filesystem skill provider and mounts Aegis's canonical skills/ tree without copying skill content.

• Made profile plugin installation the default path:

  dsh plugin --profile web add github:GanyuanRan/Aegis

• Retained the updater-managed $DSH_HOME/skills direct-child layout as an explicit compatibility mode. The native bundle and compatibility mode must not be enabled at the same time.

• Added deterministic and isolated integration coverage for profile dependency reconciliation, bundle injection, config readback, module loading, and clean removal.

Verification

• All declared plugin manifests are synchronized at 2.8.0.
• DeepSeek Harness deterministic boundary and adapter checks passed.
• An isolated DeepSeek Harness 0.1.0-rc.6 install → list → config dump → module load → remove cycle passed.
• The npm package dry run confirmed that the DSH adapter, bundle patch, and canonical Aegis skills are included.
• Install-verification policy, repository boundary checks, and DeepSeek Harness updater regression tests passed.

Scope

Aegis remains a runtime-ready method pack. This bundle does not add runtime, policy, gate, or completion authority. DeepSeek Harness is still in developer preview, so this release does not claim release-level live model-routing closeout.

Full changelog: v2.7.8...v2.8.0

Aegis v2.7.8

DeepSeek Harness 结构性适配

• 新增官方 DeepSeek Harness 宿主指南，明确与 DeepSeek-TUI 的独立边界。
• 更新器新增 deepseek-harness / dsh 识别，默认使用 $DSH_HOME/skills（未设置时为 ~/.dsh/skills）并生成 direct-child skill 链接。
• Windows 支持 junction，macOS/Linux 支持 symlink；注册流程包含 doctor 结构验证。
• README、中文 README、中英文速通秘籍、兼容矩阵、已知限制、测试说明与发布门禁已同步。
• 改进 Windows 非 UTF-8 子进程输出的容错处理。

当前结论是“结构性安装适配完成”。DeepSeek Harness 仍处于 developer preview，本版本不宣称 release-level live routing 已完成。

Aegis v2.7.7

更新说明

• 正式退役 Gemini CLI 支持，删除相关清单、入口和工具映射。
• 保留 Gemini 到 Antigravity 的历史迁移说明及企业/API Key 使用边界。
• Antigravity 仍处于待 fresh closeout 状态，本次发布不扩大其支持承诺。
• 同步更新版本至 2.7.7，并强化测试以防旧 Gemini 入口意外恢复。

Aegis v2.7.6

更新说明

Aegis v2.7.6 重点增强了 fallback 退役安全性、canonical owner 识别、变更必要性判断和真实路由证据，避免在外部依赖尚未确认时误删兼容路径，同时改进实现理由识别，减少模板文本误判且不牺牲真实任务能力。本版本继续保持 Aegis Method Pack 的边界，强调最小必要改动、证据验证和安全退役。

主要改进

• 外部依赖状态未知时保留兼容路径并请求确认，避免不安全退役。
• 强化 root cause、canonical owner、minimum repair 和 code-change decision 的编辑前证据。
• benchmark v6 冻结 sentinel / discriminator 角色及 mixed / non-discriminating 复核规则。
• 收紧 implementation-rationale 识别，排除模板、占位符和引用文本误报，同时保留真实事实组合。

最新实测

gpt-5.6-sol / xhigh extended held-out A/B 评测完成 120 次有效运行：合同通过率 61.67% → 93.33%，不安全结果率 13.33% → 0%。

结果为有界建议性证据；已通过 arm-hidden 技术复核，不构成普遍质量、运行时权威或完成权威声明。

Aegis v2.7.5

Highlights

Anti-entropy governance now auto-triggers when a task touches entropy surfaces, subagent/inline routing is unified behind a coordination-benefit gate, and benchmark retry headroom no longer degrades richer profiles.

What's New

Anti-Entropy Governance: Auto-Trigger, Gated Execution

• anti-entropy-governance now loads automatically when the task touches owner collapse, fallback removal, or schema/persistence/source-of-truth boundaries; it identifies opportunities and classifies delete-first / compat-exception / confirmation-first. Destructive execution still requires explicit scoped user confirmation (generic agreement is rejected, backup/rollback hints required).
• Stays out of the global hot path budget; workflow-quality assertions updated.

Unified Subagent/Inline and Worktree Discipline

• Three execution skill descriptions now carry the "coordination benefit beats inline cost; otherwise inline" gate (subagent-driven-development, dispatching-parallel-agents, executing-plans), matching writing-plans internal conditions.
• subagent-driven-development long-task loading narrowed from "multi-task forces long-task" to "cross-session / handoff / resumable-state only"; long-task-continuation clarifies multi-step/todo/subagent alone do not justify durable records (inline checkpoint by default).
• Playbook capability table and global user rules template (EN/ZH) synced: worktree exceptions are concurrent checkout, blocking unrelated dirty state, or explicit user/repository authority.

Contributor Fix: Benchmark Retry Headroom Is a Floor

• run_agentic_benchmark.py retry headroom now takes the max instead of unconditionally overwriting, so extended-held-out (132/18000) is no longer degraded to 64/7200; docs and a regression test updated (#15, thanks Ryan Chou).

Aegis v2.7.2

Highlights

Maintenance release tightening the routing descriptions that gate documentation, branch/worktree, and subagent decisions.

What's New

Fix: Routing Descriptions Tightened Across Three Decision Themes

• Document creation: establishing-project-context trigger narrowed (removed the broad "vague term", now conflicting/renamed/deprecated, plus "Routine small tasks stay on the fast path"); first-principles-review "repeated fixes" replaced with "high-risk decisions" plus an exclusion. Small tasks no longer get pulled into formal modeling/review flows.
• Branch/worktree: finishing-a-development-branch now reads "integration or cleanup of an existing task-created branch/worktree", aligned with ADR-0003 exception-only discipline instead of implying Aegis creates branches by default.
• Subagent/session routing: the cross-session (executing-plans) vs current-session (subagent-driven-development) distinction is now written into both descriptions; dispatching-parallel-agents is scoped to "2+ independent tasks without a written plan", removing the routing overlap with SDD; SDD points ad-hoc tasks to dispatching-parallel-agents.

All six descriptions stay within the 240-char Kimi Code limit. Validation: 22 Kimi-visible skill metadata files valid, context budget, workflow quality, and TDD policy checks pass.