R2+R2.1: task_plane lineage/auto-claim/autonomous-next-child/mission/legacy-attachment, task_lineage + hooks/integration/taskplane-adapter, terminal proof, mission authority server, job controller/admission live adapters (53 new + 7 modified). G5 dependency/import closure: tests/conftest.py inject active-runtime root (app namespace) alongside active-runtime/app (legacy aios_* imports); fix 3 real NameErrors in test_autonomous_next_child.py (mod-qualified refs). Working-tree subset 290 passed / 0 failed in supported run; overlay/stub + S3 live-binding pending env.
8.5 KiB
AIOS Unified Platform Convergence R1 — 架构变更说明
0. 决策摘要
本变更将 Identity/Lineage S3、ANNEX-2.3 R2 Auto-Claim、平台化 Phase-0、 Governed Merge/Release 收敛为一个新的上层 Program,而不是改写已存在 Identity/Lineage Mission 的原始验收合同。
核心原则:
- 新 Program,不重写旧 root。 已有 Mission
mission-f3af205b-7dd6-455f-a561-88f0294cc815和 roottask-133ad96e-e0d7-4202-934f-80c75053b11f原样保留,C01–C06 的 acceptance 不变。 - 唯一 Git authority 是
beabd053961047f170e87e1110f2636dbd50f5fb。 release-control runtime3c320aeb89d9ba8f0ed516abeb57252682b18983和 production4642b74af62c89a505ae49241e805a5a88ca897f是不同平面,不得互相冒充。 - 恢复语义,不恢复旧拓扑。 从
3c320ae…选择性移植 SystemTruth/Execution Integrity 语义,但不把该后继分支整体推回 main。 - 单一权威、只读聚合。 新 inventory/caller_scan/Environment Profile/ Retry Governor 都是 evidence producer,不成为新持久化权威。
- 先评审,后部署。 本 R1 是 Review Candidate;禁止在 Review Gate 通过前 进行 staging/production activation。
1. 当前权威三平面
| 平面 | OID | 含义 |
|---|---|---|
| Git authority / main | beabd053961047f170e87e1110f2636dbd50f5fb |
新候选唯一 source base |
| release-control runtime | 3c320aeb89d9ba8f0ed516abeb57252682b18983 |
现有控制面运行字节,不是 main |
| production baseline | 4642b74af62c89a505ae49241e805a5a88ca897f |
V2R2C2 生产 evidence closure |
beabd053961047f170e87e1110f2636dbd50f5fb -> 3c320aeb89d9ba8f0ed516abeb57252682b18983 仅包含 10 个 SystemTruth/consistency 相关路径;
现有 Identity/Lineage 27 文件候选与这 10 个路径无交集。因此 27 文件候选可以
在 重新读取 authority preimage 并通过 apply/check 后 重新绑定
beabd053961047f170e87e1110f2636dbd50f5fb,但不能复用旧 3c320ae… approval/base 记录。
2. 新 Program 与既有 Mission 的关系
2.1 新上层 Program(尚未在 live TaskPlane 创建)
goal_slug = AIOS-UNIFIED-PLATFORM-CONVERGENCE
实际 mission_id/root_task_id 必须由 live TaskPlane authority 创建并返回,
本候选不预造 ID。
2.2 既有 Identity/Lineage Mission(保留)
- mission:
mission-f3af205b-7dd6-455f-a561-88f0294cc815 - root:
task-133ad96e-e0d7-4202-934f-80c75053b11f - owner:
local-owner-authority - alias:
U9CA8400F7 - legacy durable task:
TASK-P0-CHATGPT-BRIDGE-V2-CONTROL-PLANE-CONVERGENCE-INTEGRATE-20260913T184412Z-19b7 - legacy job:
job-7457a322d5f9d6931c7cd99ce3ec6c18
该 Mission 作为新 Program 的 Track A 依赖/子目标;其历史 task/job/idempotency 身份不可为了新命名而重建。
3. 四条 Track
Track A — Identity / Lineage S3
目标保持:
root → child → terminal proof → ROOT_REVIEW → actual next action → execution → root re-review
包含:
- C01 governed candidate commit/readback
- C02 staging lineage policy activation
- C03 root-action executor binding
- C04 legacy attach
- C05 live failure/replay/concurrency E2E
- C06 production convergence
Track B — ANNEX-2.3 R2 Auto-Claim / Execution Integrity
新链:
admission/preflight → SchedulabilityProof → discovery → claim-time authoritative recheck → lease/fencing/CAS → execution → heartbeat/progress → result/evidence → readback/acceptance
R1 已实现 review-candidate 代码:
task_plane/schedulability.pytask_plane/auto_claim.py- A01 风格多 blocker、scope/capability/heartbeat/auth/capacity/freeze/dependency/lease 检查
- proof 仅为诊断证据,不是 bearer credential
- claim 前完整 authority recheck
R1 不直接替换现有 claim 路径;必须在 staging 评审时通过 adapter 绑定后才可启用。
Track C — Platform Phase-0 / Generation G0
仅包含批准的第一 generation:
- WP0a Version/Baseline inventory
- WP0b Git writer inventory
- WP0c Authorization inventory
- WP0d Canonical/Digest inventory
- WP0e caller_scan
- WP1 Environment Capability Profile
- WP2 Retry/Toil Governor
这些组件只读或纯决策:
- 不新增 version authority
- 不新增 canonical authority
- 不写业务状态
UNKNOWN永远不能被 caller_scan 降级为 orphan- Retry Governor 只返回决策,不执行 retry
禁止提前引入 WP8 MissionRun store / WP9 新 Release Saga 控制面。
Track D — Governed Merge / Staging / Release
顺序:
Review Gate → governed child commit/readback → staging activation → integrated live E2E → release candidate/acceptance → production convergence
部署不是本 Review Candidate 的交付内容。
4. 状态权威边界
| 维度 | 权威 |
|---|---|
| Mission/root/lineage | TaskPlane/Mission authority |
| claim/heartbeat/task revision | TaskPlane |
| Owner/TaskPrincipal authorization | existing auth/Admission authority |
| Writer lease / governed write | existing writer/lease authority |
| source bytes/ref | Git authority |
| deployed runtime | runtime readback |
| external projection | external projection readback |
| caller liveness evidence | caller_scan evidence only |
| retry decision | Retry/Toil Governor decision only |
新模块不得建立第三份真值数据库。
5. 关键架构优化
5.1 动态 readiness 与不可变身份分离
Planner 选择身份(mission/root/child/route/authority digest)不可变;
readiness_class 是动态观察,可在未激活、未终态时通过 revision/CAS 单调向
“更可执行”方向推进。active claim/READY/DONE 后冻结。
5.2 Legacy reader 无副作用
旧 GET /v1/tasks/{id} 可能触发全局 lease expiry。R1 改为:
- 通过
/v1/tasks/{id}/events做 Bearer 认证; - Task row 从 authority SQLite 以
mode=ro+PRAGMA query_only=ON精确读取; - 只允许终态;
- readback ref 同时绑定 events SHA 和 DB revision;
- 不暴露
lease_token_hash。
5.3 SchedulabilityProof 是短寿命诊断证据
Proof 绑定 task/worker/actor、revision/generation、capabilities、scope digest、 authorization revision、capacity/dependency/freeze/lease 观察;TTL ≤ 300s。 真正 claim 必须重读 authority 并完全一致,proof 本身不能授权。
5.4 Root Action Provider / Executor 分离
Provider 只把 root review 映射到准备计划;Executor 使用现有 continuation/recovery 路径执行。不能让 planner 把“PENDING”当成“已执行”。
6. 数据迁移策略
- 既有 Mission/root/children:不重建。
- legacy
TASK-P0-CHATGPT-BRIDGE-V2-CONTROL-PLANE-CONVERGENCE-INTEGRATE-20260913T184412Z-19b7/job-7457a322d5f9d6931c7cd99ce3ec6c18:仅 attachment,不改 identity。 - 新 Platform inventory:无持久业务迁移。
- 新 schedulability proof:短寿命诊断记录;若持久化,必须复用现有 TaskPlane evidence/journal,禁止新数据库。
- SystemTruth:只读派生,不能成为第三控制面。
7. Review Gate
R0 — 架构与代码评审
必须同时满足:
- 新 Program 与旧 Mission 边界明确
- base 重新绑定
beabd053961047f170e87e1110f2636dbd50f5fb - 变更路径均有 ChangeEnvelope
- 新代码定向测试 PASS
- 旧候选继承证据有效
- 无新增业务权威/数据库
- legacy reader 不再使用可变 GET
- planner readiness 动态规则经过测试
- R2 proof 不能直接授权 claim
R1 — Governed Commit Gate
在 R0 GO 后:
- live reread authority HEAD
- 所有 preimage SHA 重新采集
- exact apply/diff
- governed writer commit
- authority branch/commit readback
R2 — Staging Gate
- authenticated principal/tenant
- identity/lineage/read isolation
- auto-claim adapter
- root action real executor
- restart/response-loss/sibling race
- ANNEX A01–A23
R3 — Production Gate
只有 original root acceptance + Program acceptance 全部证据闭环, 并完成 rollback/recovery 验证后才能生产收敛。
8. 回滚
- Review candidate:直接丢弃候选,无线上副作用。
- Governed commit 未发布:通过正式 successor/revert,不 reset/force push。
- Staging activation:关闭 feature/policy binding,保留 evidence。
- Production:只走既有 release Saga/rollback contract。
9. 明确不做
- 不把
3c320aeb89d9ba8f0ed516abeb57252682b18983恢复为 main。 - 不新建任务数据库、第二 scheduler、第二 release authority。
- 不把 caller_scan/Retry Governor/SystemTruth 作为写权威。
- 不把局部测试 PASS 当 root/program 完成。