diff --git a/.agents/skills/unidesk-oa/references/project-planning.md b/.agents/skills/unidesk-oa/references/project-planning.md
index 30576aeb..d684fd33 100644
--- a/.agents/skills/unidesk-oa/references/project-planning.md
+++ b/.agents/skills/unidesk-oa/references/project-planning.md
@@ -126,18 +126,18 @@ Primary L1 rule:
 
 - L1 is acceptance ownership, not a loose tag. Every non-trivial L2/L3/L4 issue must name exactly one primary L1.
 - Related L1s may be listed for handoff, but the primary L1 is the direction that defines the done criteria.
-- If ownership is unclear, ask which completion evidence would close the issue: AI development workflow, hardware evidence, user/billing ledger, execution runtime, reusable capability, or user-facing product surface.
+- If ownership is unclear, ask which completion evidence would close the issue: hardware resource readiness, Agent execution lifecycle, Harness/RL evidence, client/API behavior, user/admin policy, or platform delivery/operations.
 
 Typical HWLAB Cloud M1 L1 responsibility matrix:
 
 | Code | Short name | Primary responsibility | Does not own | Handoff examples |
 | --- | --- | --- | --- | --- |
-| `PJ2026-0101` | 云智研 | AI hardware-development workflow semantics: request, workspace, Code Agent coding/review/delivery loop | Web shell UX, billing, hardware pass/fail, runtime scheduling | Product UI to 产品协同; runtime/session to 执行编排; hardware evidence to 硬验闭环 |
-| `PJ2026-0102` | 硬验闭环 | Real hardware validation and evidence standards: target device, probe, CaseRun evidence, pass/fail signal | General runner infrastructure, Code Agent planning, UI display, billing | Runtime assembly to 执行编排; case learning to 能力沉淀 |
-| `PJ2026-0103` | 用户计量 | Multi-user service and commercial metering: auth, API keys, credit, usage ledger, billing, admin operations | Task UX, Code Agent behavior, hardware evidence, training quality | User-facing display to 产品协同; resource enforcement hooks to 执行编排 |
-| `PJ2026-0104` | 执行编排 | Execution substrate and resource orchestration: AgentRun/HWPOD/Code Agent runtime assembly, sessions, provider profiles, bundles, recovery | Hardware acceptance criteria, product UX, billing policy, capability taxonomy | Hardware validation to 硬验闭环; user interaction to 产品协同 |
-| `PJ2026-0105` | 能力沉淀 | Reusable capability accumulation: case library, failure taxonomy, review rules, templates, training feedback, model/tool improvement | Immediate feature delivery, runtime operations, billing, UI ownership | Consumes evidence from all L1s and returns standards/templates |
-| `PJ2026-0106` | 产品协同 | User-facing product surface and collaboration experience: Web/CLI/API workflows, history, multi-session UX, admin-facing interaction | Core Code Agent semantics, hardware pass/fail, billing ledger truth, runtime scheduling | Workflow semantics to 云智研; billing truth to 用户计量; runtime state to 执行编排 |
+| `PJ2026-0101` | 硬件池 | Hardware resource pool: boards, probes, HWPOD nodes, device registration, occupancy, health, availability | CaseRun pass/fail criteria, Agent scheduling, UI/API presentation, CI/CD | Harness evidence to HarnessRL; resource allocation hooks to Agent编排 |
+| `PJ2026-0102` | Agent编排 | Agent execution lifecycle: Code Agent, AgentRun, workspace, sessions, provider profiles, task recovery | Hardware inventory, Harness scoring, client UX/API contracts, CI/CD | Device allocation to 硬件池; evidence to HarnessRL; client status to 客户端 |
+| `PJ2026-0103` | HarnessRL | Harness/RL loop: harness, CaseRun, ioProbe, evidence, evaluation, replay, training feedback, RL closure | Hardware capacity, Agent scheduling, user accounts, rollout pipelines | Device facts from 硬件池; run lifecycle from Agent编排 |
+| `PJ2026-0104` | 客户端 | User entry surfaces: Web, CLI, HTTP API, SDK/IDE plugin, webhook, public docs, compatibility | User/billing truth, Agent scheduling internals, hardware inventory, CI/CD | Auth/account policy to 用户管理; runtime status to Agent编排 |
+| `PJ2026-0105` | 用户管理 | User and admin domain: accounts, registration/login, permissions, API keys, credit, usage, billing, admin, tenant isolation | Client UX, Agent execution, hardware evidence, deployment mechanics | User-facing flows to 客户端; usage sources from Agent编排/HarnessRL |
+| `PJ2026-0106` | 平台运维 | Platform delivery and operations: CI/CD, git mirror, YAML-first, Secret sync, rollout, observability, GC, platform release | Product behavior, user policy, Agent semantics, hardware validation criteria | Deployment needs from all L1s; health signals back to owners |
 
 ### L2 Topic
 
diff --git a/.agents/skills/unidesk-oa/references/templates.md b/.agents/skills/unidesk-oa/references/templates.md
index 366b7ca9..586d0032 100644
--- a/.agents/skills/unidesk-oa/references/templates.md
+++ b/.agents/skills/unidesk-oa/references/templates.md
@@ -26,12 +26,12 @@ One paragraph describing the product and research objective.
 
 | Level | Title | GitHub issue | Spec section | Current status |
 | --- | --- | --- | --- | --- |
-| L1 | PJ2026-0101 云智研 | TBD | ... | ... |
-| L1 | PJ2026-0102 硬验闭环 | TBD | ... | ... |
-| L1 | PJ2026-0103 用户计量 | TBD | ... | ... |
-| L1 | PJ2026-0104 执行编排 | TBD | ... | ... |
-| L1 | PJ2026-0105 能力沉淀 | TBD | ... | ... |
-| L1 | PJ2026-0106 产品协同 | TBD | ... | ... |
+| L1 | PJ2026-0101 硬件池 | TBD | ... | ... |
+| L1 | PJ2026-0102 Agent编排 | TBD | ... | ... |
+| L1 | PJ2026-0103 HarnessRL | TBD | ... | ... |
+| L1 | PJ2026-0104 客户端 | TBD | ... | ... |
+| L1 | PJ2026-0105 用户管理 | TBD | ... | ... |
+| L1 | PJ2026-0106 平台运维 | TBD | ... | ... |
 
 ## 5. Global Acceptance Criteria
 
@@ -135,6 +135,13 @@ One paragraph describing the product and research objective.
 - 不负责:
 - 交接方向:
 
+## 交叉引用
+
+- 上游依赖:
+- 下游支撑:
+- 接口/证据边界:
+- 需同步回写:
+
 ## 方向目标
 
 ...