compat: probe zhihu extractor before renavigate

docs: add sgclaw project briefing materials
sgclaw: snapshot today's runtime and skill updates
2026-03-30 18:39:49 +08:00 · 2026-03-30 18:39:49 +08:00 · 2026-03-30 18:39:49 +08:00 · 2026-03-30 18:39:49 +08:00 · 2026-03-30 18:39:49 +08:00 · 2026-03-30 18:39:49 +08:00
54 changed files with 5250 additions and 698 deletions
--- a/.cargo/config.toml
+++ b/.cargo/config.toml
@@ -0,0 +1,5 @@
+[source.crates-io]
+replace-with = "rsproxy-sparse"
+
+[source.rsproxy-sparse]
+registry = "sparse+https://rsproxy.cn/index/"
--- a/docs/acceptance/2026-03-29-zhihu-hotlist-excel.md
+++ b/docs/acceptance/2026-03-29-zhihu-hotlist-excel.md
@@ -1,10 +1,10 @@
 # Zhihu Hotlist Excel Acceptance

- Date: 2026-03-30 00:03:42 +0800
+- Date: 2026-03-30 03:46:51 +0800
 - Mode: real provider + live Zhihu hotlist API + simulated browser pipe
- Workspace: `/tmp/sgclaw-live-acceptance-_655xotg`
+- Workspace: `/tmp/sgclaw-live-acceptance-85j8m_dq`
 - Final success: `True`
- Total score: `100/100`
+- Total score: `90/100`

 ## Rubric

@@ -16,40 +16,43 @@

 ## Final Output

- exported_path: `/tmp/sgclaw-live-acceptance-_655xotg/.sgclaw-zeroclaw-workspace/out/zhihu-hotlist-1774800210904715681.xlsx`
- final_summary: `成功！我已经完成了知乎热榜数据的读取和Excel导出任务。以下是任务总结：
+- exported_path: `/tmp/sgclaw-live-acceptance-85j8m_dq/.sgclaw-zeroclaw-workspace/out/zhihu-hotlist-1774813602601179684.xlsx`
+- final_summary: `完美！我已经成功完成了知乎热榜数据的读取和导出。以下是任务完成总结：

 ## 任务完成情况

 1. **导航到知乎热榜页面**：成功从知乎首页导航到热榜页面（https://www.zhihu.com/hot）

-2. **获取热榜数据**：使用SuperRPA浏览器接口获取了当前热榜页面的完整内容，包含10个热榜条目
+2. **提取热榜数据**：通过浏览器获取了当前热榜前10条数据，包括：
+   - 排名（1-10）
+   - 标题（完整问题描述）
+   - 热度值（以"万"为单位）

-3. **解析数据**：从页面文本中提取了排名、标题和热度信息，格式如下：
-   - 排名：数字序号
-   - 标题：完整的问题标题
-   - 热度：以"万"为单位的阅读量
+3. **导出Excel文件**：成功将结构化数据导出为Excel文件，包含以下列：
+   - rank（排名）
+   - title（标题）
+   - heat（热度）

-4. **导出Excel文件**：成功将10个热榜条目导出到Excel文件，包含"排名"、"标题"、"热度"三列
+## 导出的数据内容
+
+热榜前10条话题涵盖了多个领域：
+- 娱乐事件（李荣浩与单依纯的演唱会授权争议）
+- 体育赛事（张雪机车WSBK夺冠）
+- 国际政治经济（日本外储操作、伊朗冲突）
+- 科技产业（字节跳动发展、DDR5内存价格）
+- 社会现象（周末情侣模式、学霸家长低调现象）
+- 生活安全（灭火毯安全问题）

 ## 生成的文件

-**Excel文件路径**：`/tmp/sgclaw-live-acceptance-_655xotg/.sgclaw-zeroclaw-workspace/out/zhihu-hotlist-1774800210904715681.xlsx`
+**Excel文件路径**：`/tmp/sgclaw-live-acceptance-85j8m_dq/.sgclaw-zeroclaw-workspace/out/zhihu-hotlist-1774813602601179684.xlsx`

-## 热榜数据概览
-
-当前知乎热榜前10名涵盖了多个热门话题：
- 体育赛事（WSBK摩托车赛）
- 娱乐事件（李荣浩与单依纯争议）
- 国际政治经济（日本外汇政策、伊朗冲突）
- 科技商业（字节跳动发展、内存价格）
- 社会现象（学霸家长行为、Excel使用）
-
-最高热度话题达到1607万阅读量，显示了知乎用户对这些话题的高度关注。`
+该文件包含了2026年3月30日03:46:02采集的知乎热榜数据，您可以使用Excel或其他电子表格软件打开查看完整的热榜信息。`

 ## Skill Logs

- `DeepSeek config loaded from /tmp/sgclaw-live-acceptance-_655xotg/sgclaw_config.json model=deepseek-chat base_url=https://api.deepseek.com`
+- `sgclaw runtime version=0.1.0 protocol=1.0`
+- `DeepSeek config loaded from /tmp/sgclaw-live-acceptance-85j8m_dq/sgclaw_config.json model=deepseek-chat base_url=https://api.deepseek.com`
 - `skills dir resolved to /home/zyl/projects/sgClaw/skill_lib/skills`
 - `runtime profile=BrowserAttached skills_prompt_mode=Compact`
 - `zeroclaw_process_message_primary`
@@ -58,24 +61,29 @@ navigate https://www.zhihu.com/hot
 getText main
 call openxml_office
 return generated local .xlsx path`
- `loaded skills: office-export-xlsx, zhihu-hotlist, zhihu-hotlist-screen, zhihu-navigate, zhihu-write`
+- `loaded skills: office-export-xlsx@0.1.0, zhihu-hotlist@0.1.0, zhihu-hotlist-screen@0.1.0, zhihu-navigate@0.1.0, zhihu-write@0.1.0`
 - `navigate https://www.zhihu.com/hot`
- `getText main`
+- `call zhihu-hotlist.extract_hotlist`
+- `browser script failed: {"unsupported_action":"eval"}`
+- `getText body`
+- `getText .HotList-list`
+- `call openxml_office`
+- `unsupported columns: expected [rank, title, heat]`
 - `call openxml_office`

 ## Live Hotlist Sample

- 1. 如何看待张雪机车在 2026 年 WSBK 葡萄牙站夺冠？这对国内的摩托赛事发展有什么影响？ | 1607万
- 2. 李荣浩摆证据 4 连质问单依纯，为什么没有授权的歌曲也能放进演唱会？演唱会筹备中可能出了什么问题？ | 1064万
- 3. 日本拟动用外储做空国际原油，以挽救日元汇率，对此你怎么看，其会重演 96 年「住友铜事件」么？ | 573万
- 4. 官方通报女子被羁押后无罪释放，申请国赔 13 天被叫停，当地成立联合调查组，最该查清什么？带来哪些深思？ | 281万
- 5. 字节跳动是怎么短短数年就能单挑所有互联网巨头的？ | 185万
- 6. 伊朗科技大学遭袭后，伊朗将美以大学列为「合法袭击目标」，如果战争扩大到教育机构，冲突还有回头路吗？ | 175万
- 7. 黄金大买家土耳其央行在伊朗战争期间抛售 80 亿美元黄金，这意味着什么？ | 166万
- 8. 为什么越厉害的学霸，她们家长越低调？从来不在朋友圈晒孩子成绩？ | 141万
- 9. DDR5 内存价格 3 月出现明显下降，请问这是短期现象，还是内存供需紧张真的缓和了？ | 135万
- 10. 为什么大公司不用 pandas 取代 Excel？ | 81万
+- 1. 李荣浩摆证据 4 连质问单依纯，为什么没有授权的歌曲也能放进演唱会？演唱会筹备中可能出了什么问题？ | 1220万
+- 2. 如何看待张雪机车在 2026 年 WSBK 葡萄牙站夺冠？这对国内的摩托赛事发展有什么影响？ | 370万
+- 3. 日本拟动用外储做空国际原油，以挽救日元汇率，对此你怎么看，其会重演 96 年「住友铜事件」么？ | 356万
+- 4. 字节跳动是怎么短短数年就能单挑所有互联网巨头的？ | 277万
+- 5. 如何看待张雪机车 820rr 拿下 wsbk 葡萄牙站第一回合冠军？这个冠军含金量如何？ | 241万
+- 6. 伊朗科技大学遭袭后，伊朗将美以大学列为「合法袭击目标」，如果战争扩大到教育机构，冲突还有回头路吗？ | 202万
+- 7. 「周末情侣」模式日渐兴起，工作日通过消息视频联系，仅周末相聚，如何看待这种模式？你有过类似的经历吗？ | 163万
+- 8. 男孩玩灭火毯全身扎满超细玻璃纤维，又痒又痛取不出来，灭火毯为什么会「扎人」？怎么处理才不遭罪？ | 158万
+- 9. DDR5 内存价格 3 月出现明显下降，请问这是短期现象，还是内存供需紧张真的缓和了？ | 151万
+- 10. 为什么越厉害的学霸，她们家长越低调？从来不在朋友圈晒孩子成绩？ | 139万

 ## Stderr

- `sgclaw ready: agent_id=db27f86f-4334-41a7-bc24-11e8fbd90486`
+- `sgclaw ready: agent_id=4b984e63-3254-4518-a75a-127e7dad6474`
--- a/docs/plans/2026-03-26-zeroclaw-prompt-safety-hardening-plan.md
+++ b/docs/plans/2026-03-26-zeroclaw-prompt-safety-hardening-plan.md
@@ -0,0 +1,551 @@
+# ZeroClaw Prompt Safety Hardening Implementation Plan
+
+> **For Claude:** REQUIRED SUB-SKILL: Use superpowers:executing-plans to implement this plan task-by-task.
+
+**Goal:** Harden ZeroClaw prompt handling and tool execution so non-skill freeform operations degrade to read-only or business-approved execution, while trusted skill-defined operations retain bounded execution privileges.
+
+**Architecture:** Build a security gate around the existing prompt and tool-entry paths instead of rewriting the full prompt architecture. The gate classifies prompt-injection risk, records operation provenance (`trusted_skill` vs `non_skill`), sanitizes injected workspace/skill content, and enforces execution mode transitions (`clean`, `suspect_readonly`, `suspect_waiting_approval`, `suspect_business_approved`). Trusted skills gain structured business-operation metadata; non-skill operations require business-level approval before any privileged capability is released.
+
+**Tech Stack:** Rust, vendored ZeroClaw (`third_party/zeroclaw`), existing approval/autonomy system, current prompt guard and prompt builder tests, `cargo test`.
+
+### Task 1: Create an Isolated Worktree and Verify a Clean Baseline
+
+**Files:**
+- Modify: `/home/zyl/projects/sgClaw/claw/.gitignore`
+- Create: `/home/zyl/projects/sgClaw/claw/.worktrees/zeroclaw-prompt-safety-hardening/**`
+
+**Step 1: Verify the worktree directory is safe to use**
+
+Run:
+```bash
+cd /home/zyl/projects/sgClaw/claw
+ls -d .worktrees
+git check-ignore -v .worktrees
+```
+
+Expected: `.worktrees` exists and is ignored by git.
+
+**Step 2: Create the implementation worktree**
+
+Run:
+```bash
+cd /home/zyl/projects/sgClaw/claw
+git worktree add .worktrees/zeroclaw-prompt-safety-hardening -b zeroclaw-prompt-safety-hardening
+```
+
+Expected: a new branch and worktree are created.
+
+**Step 3: Build the baseline in the worktree**
+
+Run:
+```bash
+cd /home/zyl/projects/sgClaw/claw/.worktrees/zeroclaw-prompt-safety-hardening
+cargo test -p zeroclawlabs prompt_guard -- --nocapture
+cargo test -p zeroclawlabs build_system_prompt -- --nocapture
+```
+
+Expected: existing relevant tests pass before any code changes.
+
+**Step 4: Commit the clean worktree setup if `.gitignore` changed**
+
+Run:
+```bash
+git add .gitignore
+git commit -m "chore: prepare worktree for prompt safety hardening"
+```
+
+Expected: commit only if `.gitignore` required an adjustment.
+
+### Task 2: Add the Core Security-Mode Data Model
+
+**Files:**
+- Create: `/home/zyl/projects/sgClaw/claw/.worktrees/zeroclaw-prompt-safety-hardening/third_party/zeroclaw/src/security/operation_policy.rs`
+- Modify: `/home/zyl/projects/sgClaw/claw/.worktrees/zeroclaw-prompt-safety-hardening/third_party/zeroclaw/src/security/mod.rs`
+- Test: `/home/zyl/projects/sgClaw/claw/.worktrees/zeroclaw-prompt-safety-hardening/third_party/zeroclaw/src/security/operation_policy.rs`
+
+**Step 1: Write the failing policy tests**
+
+Add tests that prove:
+- suspicious non-skill input maps to `suspect_readonly`
+- trusted skill operations can request bounded privileged execution
+- any out-of-scope capability request downgrades the operation
+
+Use concrete enums and assertions, for example:
+```rust
+assert_eq!(
+    ExecutionMode::from_guard_and_provenance(GuardRisk::Suspicious, OperationProvenance::NonSkill),
+    ExecutionMode::SuspectReadOnly
+);
+```
+
+**Step 2: Run the tests to verify RED**
+
+Run:
+```bash
+cd /home/zyl/projects/sgClaw/claw/.worktrees/zeroclaw-prompt-safety-hardening
+cargo test -p zeroclawlabs operation_policy -- --nocapture
+```
+
+Expected: fail because the new types do not exist yet.
+
+**Step 3: Implement the minimal policy model**
+
+Define:
+- `GuardRisk` (`Clean`, `Suspicious`, `Dangerous`)
+- `OperationProvenance` (`TrustedSkill`, `NonSkill`, `Mixed`)
+- `ExecutionMode` (`Clean`, `SuspectReadOnly`, `SuspectWaitingApproval`, `SuspectBusinessApproved`)
+- `CapabilityClass` for privileged business actions
+
+Add small helper functions that do only state mapping. Do not pull prompt-building logic into this module.
+
+**Step 4: Re-run the policy tests to verify GREEN**
+
+Run:
+```bash
+cargo test -p zeroclawlabs operation_policy -- --nocapture
+```
+
+Expected: the new policy tests pass.
+
+**Step 5: Commit**
+
+Run:
+```bash
+git add third_party/zeroclaw/src/security/mod.rs third_party/zeroclaw/src/security/operation_policy.rs
+git commit -m "feat: add prompt security execution mode model"
+```
+
+### Task 3: Add Structured Skill Trust Metadata
+
+**Files:**
+- Modify: `/home/zyl/projects/sgClaw/claw/.worktrees/zeroclaw-prompt-safety-hardening/third_party/zeroclaw/src/skills/mod.rs`
+- Modify: `/home/zyl/projects/sgClaw/claw/.worktrees/zeroclaw-prompt-safety-hardening/third_party/zeroclaw/src/tools/read_skill.rs`
+- Test: `/home/zyl/projects/sgClaw/claw/.worktrees/zeroclaw-prompt-safety-hardening/third_party/zeroclaw/src/skills/mod.rs`
+
+**Step 1: Write failing skill metadata tests**
+
+Add tests that prove:
+- `SKILL.toml` can declare a business operation type, capability list, argument constraints, and `step_budget`
+- markdown-only skills default to unprivileged metadata
+- malformed privileged metadata is rejected or downgraded safely
+
+Use a manifest shape like:
+```toml
+[skill]
+name = "export-report"
+description = "Export the monthly report"
+
+[security]
+operation_type = "browser_export_data"
+allowed_capabilities = ["browser_read", "browser_export"]
+step_budget = 6
+approval_mode = "trusted_skill"
+```
+
+**Step 2: Run the tests to verify RED**
+
+Run:
+```bash
+cargo test -p zeroclawlabs skill -- --nocapture
+```
+
+Expected: fail because the structured metadata fields are missing.
+
+**Step 3: Implement minimal structured metadata**
+
+Extend `Skill` with a structured security block, for example:
+- `operation_type`
+- `business_description`
+- `allowed_capabilities`
+- `arg_constraints`
+- `step_budget`
+- `approval_mode`
+
+Default markdown-only skills to unprivileged metadata so existing skills remain compatible.
+
+**Step 4: Make `read_skill` expose the metadata**
+
+Return or prepend enough structured metadata so the runtime can distinguish trusted skill operations from plain prompt text.
+
+**Step 5: Re-run the tests to verify GREEN**
+
+Run:
+```bash
+cargo test -p zeroclawlabs skill -- --nocapture
+```
+
+Expected: skill parsing and `read_skill` tests pass.
+
+**Step 6: Commit**
+
+Run:
+```bash
+git add third_party/zeroclaw/src/skills/mod.rs third_party/zeroclaw/src/tools/read_skill.rs
+git commit -m "feat: add trusted skill security metadata"
+```
+
+### Task 4: Sanitize Injected Workspace and Skill Content Before Prompt Assembly
+
+**Files:**
+- Create: `/home/zyl/projects/sgClaw/claw/.worktrees/zeroclaw-prompt-safety-hardening/third_party/zeroclaw/src/security/prompt_sanitizer.rs`
+- Modify: `/home/zyl/projects/sgClaw/claw/.worktrees/zeroclaw-prompt-safety-hardening/third_party/zeroclaw/src/security/mod.rs`
+- Modify: `/home/zyl/projects/sgClaw/claw/.worktrees/zeroclaw-prompt-safety-hardening/third_party/zeroclaw/src/channels/mod.rs`
+- Modify: `/home/zyl/projects/sgClaw/claw/.worktrees/zeroclaw-prompt-safety-hardening/third_party/zeroclaw/src/agent/prompt.rs`
+- Test: `/home/zyl/projects/sgClaw/claw/.worktrees/zeroclaw-prompt-safety-hardening/third_party/zeroclaw/src/channels/mod.rs`
+
+**Step 1: Write failing sanitizer tests**
+
+Add tests that prove:
+- dangerous bootstrap phrases are removed, escaped, or summarized before prompt injection
+- control characters are stripped
+- overlong files are truncated with an audit-friendly marker
+- safe business content remains readable
+
+**Step 2: Run the tests to verify RED**
+
+Run:
+```bash
+cargo test -p zeroclawlabs build_system_prompt -- --nocapture
+```
+
+Expected: fail because injected files are still copied verbatim.
+
+**Step 3: Implement the sanitizer**
+
+Create a small sanitizer that:
+- strips control characters
+- caps content length
+- flags prompt-override phrases
+- emits sanitized content plus metadata such as `truncated` and matched rules
+
+Use this sanitizer in:
+- `load_openclaw_bootstrap_files`
+- any shared path in `agent/prompt.rs` that renders workspace or skill text into the system prompt
+
+**Step 4: Re-run the tests to verify GREEN**
+
+Run:
+```bash
+cargo test -p zeroclawlabs build_system_prompt -- --nocapture
+```
+
+Expected: prompt-building tests pass with the new sanitized behavior.
+
+**Step 5: Commit**
+
+Run:
+```bash
+git add third_party/zeroclaw/src/security/mod.rs third_party/zeroclaw/src/security/prompt_sanitizer.rs third_party/zeroclaw/src/channels/mod.rs third_party/zeroclaw/src/agent/prompt.rs
+git commit -m "feat: sanitize injected workspace prompt content"
+```
+
+### Task 5: Wire `PromptGuard` into Main Agent and Gateway Entry Points
+
+**Files:**
+- Modify: `/home/zyl/projects/sgClaw/claw/.worktrees/zeroclaw-prompt-safety-hardening/third_party/zeroclaw/src/security/prompt_guard.rs`
+- Modify: `/home/zyl/projects/sgClaw/claw/.worktrees/zeroclaw-prompt-safety-hardening/third_party/zeroclaw/src/agent/agent.rs`
+- Modify: `/home/zyl/projects/sgClaw/claw/.worktrees/zeroclaw-prompt-safety-hardening/third_party/zeroclaw/src/gateway/mod.rs`
+- Modify: `/home/zyl/projects/sgClaw/claw/.worktrees/zeroclaw-prompt-safety-hardening/third_party/zeroclaw/src/gateway/ws.rs`
+- Test: `/home/zyl/projects/sgClaw/claw/.worktrees/zeroclaw-prompt-safety-hardening/third_party/zeroclaw/src/agent/agent.rs`
+
+**Step 1: Write failing entry-point tests**
+
+Add tests that prove:
+- suspicious input marks the turn as degraded instead of silently continuing
+- dangerous input is blocked
+- clean input remains unchanged
+
+Prefer tests that assert on a security decision object instead of brittle prompt strings.
+
+**Step 2: Run the tests to verify RED**
+
+Run:
+```bash
+cargo test -p zeroclawlabs prompt_guard -- --nocapture
+cargo test -p zeroclawlabs agent -- --nocapture
+```
+
+Expected: fail because no entry path consumes the guard result.
+
+**Step 3: Implement guarded entry evaluation**
+
+Before each turn:
+- scan the inbound user content
+- map the guard result into `GuardRisk`
+- create an execution context carrying risk and provenance
+- attach audit details for later logging
+
+Keep the existing `PromptGuard` regexes unless a test demands a specific adjustment.
+
+**Step 4: Re-run the tests to verify GREEN**
+
+Run:
+```bash
+cargo test -p zeroclawlabs prompt_guard -- --nocapture
+cargo test -p zeroclawlabs agent -- --nocapture
+```
+
+Expected: suspicious and blocked paths now behave deterministically.
+
+**Step 5: Commit**
+
+Run:
+```bash
+git add third_party/zeroclaw/src/security/prompt_guard.rs third_party/zeroclaw/src/agent/agent.rs third_party/zeroclaw/src/gateway/mod.rs third_party/zeroclaw/src/gateway/ws.rs
+git commit -m "feat: enforce prompt guard at runtime entry points"
+```
+
+### Task 6: Add Business-Level Privileged Operation Registry and Approval Tokens
+
+**Files:**
+- Modify: `/home/zyl/projects/sgClaw/claw/.worktrees/zeroclaw-prompt-safety-hardening/third_party/zeroclaw/src/approval/mod.rs`
+- Create: `/home/zyl/projects/sgClaw/claw/.worktrees/zeroclaw-prompt-safety-hardening/third_party/zeroclaw/src/security/business_approval.rs`
+- Modify: `/home/zyl/projects/sgClaw/claw/.worktrees/zeroclaw-prompt-safety-hardening/third_party/zeroclaw/src/security/mod.rs`
+- Test: `/home/zyl/projects/sgClaw/claw/.worktrees/zeroclaw-prompt-safety-hardening/third_party/zeroclaw/src/security/business_approval.rs`
+
+**Step 1: Write failing business approval tests**
+
+Add tests that prove:
+- only operations in the privileged registry can request approval
+- approval tokens bind to `session_id`, `operation_type`, `allowed_capabilities`, `step_budget`, and expiration
+- a mismatched or expired approval token is rejected
+
+**Step 2: Run the tests to verify RED**
+
+Run:
+```bash
+cargo test -p zeroclawlabs business_approval -- --nocapture
+```
+
+Expected: fail because the business approval registry does not exist yet.
+
+**Step 3: Implement the registry and token model**
+
+Create:
+- a privileged business operation registry
+- a single-operation approval token
+- helper checks for `can_request_approval` and `matches_execution_request`
+
+Model approval at the business-operation level, not raw tool calls.
+
+**Step 4: Extend the existing approval module**
+
+Teach the approval module to carry business-level fields through the current request/response flow without breaking old call sites.
+
+**Step 5: Re-run the tests to verify GREEN**
+
+Run:
+```bash
+cargo test -p zeroclawlabs business_approval -- --nocapture
+```
+
+Expected: the token validation and registry tests pass.
+
+**Step 6: Commit**
+
+Run:
+```bash
+git add third_party/zeroclaw/src/approval/mod.rs third_party/zeroclaw/src/security/mod.rs third_party/zeroclaw/src/security/business_approval.rs
+git commit -m "feat: add business-level approval registry"
+```
+
+### Task 7: Enforce Execution Modes in Tool Dispatch
+
+**Files:**
+- Modify: `/home/zyl/projects/sgClaw/claw/.worktrees/zeroclaw-prompt-safety-hardening/third_party/zeroclaw/src/agent/dispatcher.rs`
+- Modify: `/home/zyl/projects/sgClaw/claw/.worktrees/zeroclaw-prompt-safety-hardening/third_party/zeroclaw/src/agent/agent.rs`
+- Modify: `/home/zyl/projects/sgClaw/claw/.worktrees/zeroclaw-prompt-safety-hardening/third_party/zeroclaw/src/agent/loop_.rs`
+- Test: `/home/zyl/projects/sgClaw/claw/.worktrees/zeroclaw-prompt-safety-hardening/third_party/zeroclaw/src/agent/dispatcher.rs`
+
+**Step 1: Write failing dispatcher tests**
+
+Add tests that prove:
+- `suspect_readonly` allows only safe read capabilities
+- `trusted_skill` can execute capabilities listed in its metadata within `step_budget`
+- `mixed` or non-skill privileged calls require a matching business approval token
+
+**Step 2: Run the tests to verify RED**
+
+Run:
+```bash
+cargo test -p zeroclawlabs dispatcher -- --nocapture
+```
+
+Expected: fail because the dispatcher does not yet know about execution modes.
+
+**Step 3: Implement capability enforcement**
+
+Before dispatching any tool:
+- resolve the operation context
+- map the tool call to a capability class
+- reject calls outside the current execution mode
+- decrement or validate `step_budget` for approved bounded flows
+
+Do not rely on prompt text for enforcement.
+
+**Step 4: Re-run the tests to verify GREEN**
+
+Run:
+```bash
+cargo test -p zeroclawlabs dispatcher -- --nocapture
+```
+
+Expected: dispatch now respects read-only, trusted skill, and business-approved modes.
+
+**Step 5: Commit**
+
+Run:
+```bash
+git add third_party/zeroclaw/src/agent/dispatcher.rs third_party/zeroclaw/src/agent/agent.rs third_party/zeroclaw/src/agent/loop_.rs
+git commit -m "feat: enforce execution mode in tool dispatch"
+```
+
+### Task 8: Default Skills Prompt Injection to Compact for Safer Runtime Behavior
+
+**Files:**
+- Modify: `/home/zyl/projects/sgClaw/claw/.worktrees/zeroclaw-prompt-safety-hardening/third_party/zeroclaw/src/config/schema.rs`
+- Modify: `/home/zyl/projects/sgClaw/claw/.worktrees/zeroclaw-prompt-safety-hardening/third_party/zeroclaw/src/agent/prompt.rs`
+- Modify: `/home/zyl/projects/sgClaw/claw/.worktrees/zeroclaw-prompt-safety-hardening/third_party/zeroclaw/src/channels/mod.rs`
+- Test: `/home/zyl/projects/sgClaw/claw/.worktrees/zeroclaw-prompt-safety-hardening/third_party/zeroclaw/src/config/schema.rs`
+
+**Step 1: Write the failing configuration test**
+
+Add a test that asserts the default skill prompt injection mode is `Compact` unless explicitly configured otherwise.
+
+**Step 2: Run the test to verify RED**
+
+Run:
+```bash
+cargo test -p zeroclawlabs skills_prompt_injection_mode -- --nocapture
+```
+
+Expected: fail because defaults still point to `Full`.
+
+**Step 3: Implement the default flip**
+
+Update config defaults and any prompt-builder defaults that currently assume `Full`. Keep explicit user config backward compatible.
+
+**Step 4: Re-run the test to verify GREEN**
+
+Run:
+```bash
+cargo test -p zeroclawlabs skills_prompt_injection_mode -- --nocapture
+```
+
+Expected: default configuration now resolves to `Compact`.
+
+**Step 5: Commit**
+
+Run:
+```bash
+git add third_party/zeroclaw/src/config/schema.rs third_party/zeroclaw/src/agent/prompt.rs third_party/zeroclaw/src/channels/mod.rs
+git commit -m "feat: default skills prompt injection to compact"
+```
+
+### Task 9: Add Audit Logging and Regression Coverage
+
+**Files:**
+- Modify: `/home/zyl/projects/sgClaw/claw/.worktrees/zeroclaw-prompt-safety-hardening/third_party/zeroclaw/src/observability/mod.rs`
+- Modify: `/home/zyl/projects/sgClaw/claw/.worktrees/zeroclaw-prompt-safety-hardening/third_party/zeroclaw/src/agent/agent.rs`
+- Modify: `/home/zyl/projects/sgClaw/claw/.worktrees/zeroclaw-prompt-safety-hardening/third_party/zeroclaw/src/channels/mod.rs`
+- Create: `/home/zyl/projects/sgClaw/claw/.worktrees/zeroclaw-prompt-safety-hardening/third_party/zeroclaw/tests/prompt_safety_regression.rs`
+
+**Step 1: Write the failing regression tests**
+
+Cover:
+- prompt override attack from user content
+- malicious `AGENTS.md` bootstrap content
+- trusted skill execution within bounds
+- non-skill privileged request requiring business approval
+- approval token mismatch
+- session history restore preserving degraded mode
+
+**Step 2: Run the tests to verify RED**
+
+Run:
+```bash
+cargo test -p zeroclawlabs --test prompt_safety_regression -- --nocapture
+```
+
+Expected: fail because the end-to-end behavior is not wired together yet.
+
+**Step 3: Implement audit logging**
+
+Record:
+- input hash
+- matched guard rules
+- risk level
+- provenance
+- execution mode transitions
+- approval decisions
+
+Avoid logging raw sensitive content.
+
+**Step 4: Re-run the regression tests to verify GREEN**
+
+Run:
+```bash
+cargo test -p zeroclawlabs --test prompt_safety_regression -- --nocapture
+```
+
+Expected: the regression suite passes.
+
+**Step 5: Commit**
+
+Run:
+```bash
+git add third_party/zeroclaw/src/observability/mod.rs third_party/zeroclaw/src/agent/agent.rs third_party/zeroclaw/src/channels/mod.rs third_party/zeroclaw/tests/prompt_safety_regression.rs
+git commit -m "test: add prompt safety regression coverage"
+```
+
+### Task 10: Final Verification and Integration Review
+
+**Files:**
+- Modify: `/home/zyl/projects/sgClaw/claw/.worktrees/zeroclaw-prompt-safety-hardening/docs/L5-提示词分布与安全改造方案.md`
+- Modify: `/home/zyl/projects/sgClaw/claw/.worktrees/zeroclaw-prompt-safety-hardening/docs/README.md`
+
+**Step 1: Run targeted verification**
+
+Run:
+```bash
+cd /home/zyl/projects/sgClaw/claw/.worktrees/zeroclaw-prompt-safety-hardening
+cargo test -p zeroclawlabs prompt_guard -- --nocapture
+cargo test -p zeroclawlabs build_system_prompt -- --nocapture
+cargo test -p zeroclawlabs dispatcher -- --nocapture
+cargo test -p zeroclawlabs --test prompt_safety_regression -- --nocapture
+```
+
+Expected: all prompt safety and dispatcher tests pass.
+
+**Step 2: Run a broad ZeroClaw package test pass if time permits**
+
+Run:
+```bash
+cargo test -p zeroclawlabs -- --nocapture
+```
+
+Expected: no regressions in the vendored package test suite, or a documented list of unrelated existing failures.
+
+**Step 3: Update the security design docs**
+
+Document:
+- execution modes
+- trusted skill metadata contract
+- business approval flow
+- why non-skill privileged actions are gated
+
+**Step 4: Commit the docs**
+
+Run:
+```bash
+git add docs/L5-提示词分布与安全改造方案.md docs/README.md
+git commit -m "docs: record prompt safety hardening design"
+```
+
+**Step 5: Prepare merge review notes**
+
+Write a short integration summary covering:
+- changed entry points
+- backward-compatibility expectations
+- any skills that need metadata upgrades
+- rollout recommendation for existing integrators
--- a/docs/plans/2026-03-27-sgclaw-chat-first-ui-refactor-plan.md
+++ b/docs/plans/2026-03-27-sgclaw-chat-first-ui-refactor-plan.md
@@ -0,0 +1,179 @@
+# sgClaw Chat-First UI Refactor Implementation Plan
+
+> **For Claude:** REQUIRED SUB-SKILL: Use superpowers:executing-plans to implement this plan task-by-task.
+
+**Goal:** Rebuild the sgClaw floating chat UI into a chat-first plugin-style product where the message timeline is primary, `执行摘要` is folded into the conversation, and `调试` opens as a full-window overlay instead of occupying persistent space.
+
+**Architecture:** Keep `chrome://superrpa-functions/sgclaw-chat` as the first verified host because it already has Lit-based unit tests, then mirror the same information architecture and visual hierarchy into the ordinary-page injected `sgclaw_overlay.js`. Do not introduce a new backend contract; only rearrange presentation, panel semantics, and message/result composition around the existing runtime state.
+
+**Tech Stack:** Chromium WebUI, Lit templates/components, injected Shadow DOM overlay JavaScript, existing SuperRPA bridge/runtime callbacks, mainline TS unit tests.
+
+### Task 1: Lock The New Information Architecture In Tests
+
+**Files:**
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_mainline_unittest.ts`
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_settings_state_mainline_unittest.ts`
+
+**Step 1: Write the failing test**
+
+Add assertions for these exact product rules:
+- `getHtml()` must no longer emit the legacy `debug-note`.
+- the main chat template must define a dedicated overlay/sheet container for `history`, `settings`, and `debug`.
+- the debug panel must be described as a full-window overlay rather than a side drawer/log block.
+- the result presentation must be part of the message stream, not a standalone persistent secondary panel.
+
+**Step 2: Run test to verify it fails**
+
+Run:
+```bash
+node --test /home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_mainline_unittest.ts
+```
+
+Expected: FAIL because current template still includes `debug-note`, side-by-side panel layout, and standalone result panel semantics.
+
+**Step 3: Write minimal implementation**
+
+Change only template/component strings and assertions needed to express the new structure, without touching styling yet.
+
+**Step 4: Run test to verify it passes**
+
+Run the same command.
+
+Expected: PASS.
+
+### Task 2: Refactor `chrome://` sgClaw Into Chat-First Structure
+
+**Files:**
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat.html.ts`
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat.css.ts`
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-chat-header.ts`
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-chat-composer.ts`
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-debug-drawer.ts`
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-history-panel.ts`
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-settings-panel.ts`
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-message-list.ts`
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-message-card-result.ts`
+
+**Step 1: Keep the header narrow**
+
+Make the header carry only:
+- brand
+- current page label
+- compact runtime status
+- actions for `新对话 / 历史 / 设置 / 调试 / 收起`
+
+Remove the large subtitle/debug framing and the separate heavy runtime action row feel.
+
+**Step 2: Make the message timeline primary**
+
+Turn the main shell body into:
+- a single timeline container
+- optional empty-state presets
+- no persistent secondary summary card
+
+`finalResult` should render as a folded result card appended in the stream.
+
+**Step 3: Convert secondary panels into full overlays**
+
+Render `history`, `settings`, and `debug` inside a full-window overlay/sheet that covers the chat content area instead of sitting above or beside it.
+
+**Step 4: Re-skin toward the approved direction**
+
+Use:
+- soft neutral surfaces
+- restrained accent usage
+- thinner borders
+- calmer shadows
+- clearer assistant/user card contrast
+
+Avoid:
+- debug-workbench feeling
+- large gradient blocks
+- heavy explanatory copy in the main flow
+
+**Step 5: Run the unit tests**
+
+Run:
+```bash
+node --test /home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_mainline_unittest.ts
+```
+
+Expected: PASS.
+
+### Task 3: Mirror The Same Structure Into Ordinary-Page Overlay
+
+**Files:**
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/sgclaw_overlay.js`
+
+**Step 1: Remove the standalone result panel**
+
+Delete the always-visible `执行摘要` block from the main window body.
+
+**Step 2: Introduce overlay panels**
+
+Change panel rendering so `history`, `settings`, and `debug` appear in a dedicated full-window overlay layer within the floating window instead of as sibling blocks consuming vertical space.
+
+**Step 3: Rebuild the shell**
+
+Match the `chrome://` layout:
+- compact header
+- primary message timeline
+- folded result card inside conversation
+- sticky composer
+
+**Step 4: Preserve behavior**
+
+Do not break:
+- `sgclaw.newConversation`
+- `sgclaw.restoreConversation`
+- runtime polling
+- config save/load
+- unread badge behavior
+
+**Step 5: Run a syntax sanity check**
+
+Run:
+```bash
+node --check /home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/sgclaw_overlay.js
+```
+
+Expected: PASS.
+
+### Task 4: Verify Browser Resource Integration
+
+**Files:**
+- No new source files; verification only
+
+**Step 1: Run TS / mainline tests**
+
+Run:
+```bash
+bash -lc "autoninja -C /home/zyl/projects/superRpa/src/out/KylinRelease functions_ui_mainline_unittests"
+```
+
+Expected: build succeeds.
+
+**Step 2: Run targeted mainline unit tests**
+
+Run:
+```bash
+/home/zyl/projects/superRpa/src/out/KylinRelease/functions_ui_mainline_unittests --gtest_filter='FunctionsUiMainlineTest.*sgclaw*'
+```
+
+If filter finds no test names, run the full binary and confirm it exits `0`.
+
+**Step 3: Rebuild browser resources if needed**
+
+Run:
+```bash
+bash -lc "autoninja -C /home/zyl/projects/superRpa/src/out/KylinRelease chrome"
+```
+
+**Step 4: Manually verify product behavior**
+
+Check:
+- ordinary webpage floating window
+- `chrome://superrpa-functions/sgclaw-chat`
+- `调试` opens as full overlay
+- `执行摘要` no longer blocks the main conversation
+- `历史` and `设置` do not consume persistent layout space
--- a/docs/plans/2026-03-27-sgclaw-configurable-skills-dir-plan.md
+++ b/docs/plans/2026-03-27-sgclaw-configurable-skills-dir-plan.md
@@ -0,0 +1,148 @@
+# SGClaw Configurable Skills Directory Implementation Plan
+
+> **For Claude:** REQUIRED SUB-SKILL: Use superpowers:executing-plans to implement this plan task-by-task.
+
+**Goal:** Let `sgclaw` own skill-directory resolution and allow users to set a custom skills directory in `sgclaw_config.json` without relying on SuperRPA to copy skills into the runtime workspace.
+
+**Architecture:** Extend the existing browser JSON config parser so `sgclaw` can read an optional `skillsDir` field alongside DeepSeek settings. Keep the current embedded ZeroClaw workspace for memory/config internals, but decouple skill loading from that fixed path by resolving a configurable skills root at runtime. Preserve backward compatibility by defaulting to `<workspace_root>/.sgclaw-zeroclaw-workspace/skills` when `skillsDir` is absent or empty.
+
+**Tech Stack:** Rust, serde JSON parsing, existing ZeroClaw compatibility runtime, cargo test
+
+### Task 1: Capture browser config requirements
+
+**Files:**
+- Modify: `/home/zyl/projects/sgClaw/claw/src/config/settings.rs`
+- Test: `/home/zyl/projects/sgClaw/claw/tests/compat_config_test.rs`
+
+**Step 1: Write the failing test**
+
+Add tests that load `sgclaw_config.json` containing:
+- no `skillsDir`
+- a relative `skillsDir`
+- an absolute `skillsDir`
+
+Assert that:
+- `skillsDir` missing falls back to default workspace skills path
+- relative values resolve against the browser config directory
+- absolute values are preserved
+
+**Step 2: Run test to verify it fails**
+
+Run: `cargo test compat_config -- --nocapture`
+
+Expected: FAIL because `DeepSeekSettings` / config adapter do not expose any skills directory override yet.
+
+**Step 3: Write minimal implementation**
+
+Add a browser-config structure that parses `skillsDir` and expose a resolver function that returns the effective skills directory for `sgclaw`.
+
+**Step 4: Run test to verify it passes**
+
+Run: `cargo test compat_config -- --nocapture`
+
+Expected: PASS for the new parsing and path-resolution cases.
+
+### Task 2: Route compat runtime skill loading through sgclaw-owned resolution
+
+**Files:**
+- Modify: `/home/zyl/projects/sgClaw/claw/src/compat/config_adapter.rs`
+- Modify: `/home/zyl/projects/sgClaw/claw/src/compat/runtime.rs`
+- Test: `/home/zyl/projects/sgClaw/claw/tests/compat_runtime_test.rs`
+
+**Step 1: Write the failing test**
+
+Add a compat runtime test that creates:
+- a default workspace skill package under `.sgclaw-zeroclaw-workspace/skills`
+- a custom skill package under another directory configured via `skillsDir`
+
+Assert that provider request payload contains only the configured skill name when `skillsDir` is set, and still contains workspace skill names when the override is absent.
+
+**Step 2: Run test to verify it fails**
+
+Run: `cargo test compat_runtime -- --nocapture`
+
+Expected: FAIL because the runtime currently always loads skills from `config.workspace_dir`.
+
+**Step 3: Write minimal implementation**
+
+Keep `config.workspace_dir` for ZeroClaw internal state, but load skills from the resolved effective skills directory by calling `load_skills_from_directory` directly when a custom directory is configured.
+
+**Step 4: Run test to verify it passes**
+
+Run: `cargo test compat_runtime -- --nocapture`
+
+Expected: PASS and provider request payload shows the right `Available Skills` content.
+
+### Task 3: Document and verify backward compatibility
+
+**Files:**
+- Modify: `/home/zyl/projects/sgClaw/claw/docs/README.md`
+- Modify: `/home/zyl/projects/sgClaw/claw/docs/L5-提示词分布与安全改造方案.md`
+
+**Step 1: Write the failing check**
+
+Record the expected runtime behavior:
+- `sgclaw` owns skill lookup
+- SuperRPA only passes `--config-path`
+- `skillsDir` is optional
+
+**Step 2: Run verification**
+
+Run: `rg -n "skillsDir|sgclaw owns skill lookup|config-path" docs`
+
+Expected: missing text before docs are updated.
+
+**Step 3: Write minimal documentation**
+
+Document:
+- JSON field name
+- relative-path resolution base
+- default fallback
+- operational implication for SuperRPA integration
+
+**Step 4: Run verification**
+
+Run: `rg -n "skillsDir|sgclaw owns skill lookup|config-path" docs`
+
+Expected: PASS with updated docs.
+
+### Task 4: Final verification
+
+**Files:**
+- Review only: `/home/zyl/projects/sgClaw/claw/src/config/settings.rs`
+- Review only: `/home/zyl/projects/sgClaw/claw/src/compat/config_adapter.rs`
+- Review only: `/home/zyl/projects/sgClaw/claw/src/compat/runtime.rs`
+- Review only: `/home/zyl/projects/sgClaw/claw/tests/compat_config_test.rs`
+- Review only: `/home/zyl/projects/sgClaw/claw/tests/compat_runtime_test.rs`
+
+**Step 1: Run targeted tests**
+
+Run: `cargo test compat_config -- --nocapture`
+
+Expected: PASS
+
+**Step 2: Run runtime tests**
+
+Run: `cargo test compat_runtime -- --nocapture`
+
+Expected: PASS
+
+**Step 3: Run skill-lib structural validation**
+
+Run: `python3 -m unittest tests.skill_lib_validation_test -v`
+
+Expected: PASS
+
+**Step 4: Commit**
+
+```bash
+git add docs/plans/2026-03-27-sgclaw-configurable-skills-dir-plan.md \
+  src/config/settings.rs \
+  src/compat/config_adapter.rs \
+  src/compat/runtime.rs \
+  tests/compat_config_test.rs \
+  tests/compat_runtime_test.rs \
+  docs/README.md \
+  docs/L5-提示词分布与安全改造方案.md
+git commit -m "feat: make sgclaw skills directory configurable"
+```
--- a/docs/plans/2026-03-27-sgclaw-floating-chat-frontend-design.md
+++ b/docs/plans/2026-03-27-sgclaw-floating-chat-frontend-design.md
@@ -0,0 +1,624 @@
+# sgClaw Floating Chat Frontend Implementation Plan
+
+> **For Claude:** REQUIRED SUB-SKILL: Use superpowers:executing-plans to implement this plan task-by-task.
+
+**Goal:** Replace the current debug-style `sgclaw-chat` UI with a complete floating-chat frontend that matches the product structure of Doubao's side panel while preserving the current SuperRPA bridge and configuration capabilities.
+
+**Architecture:** Keep `chrome://superrpa-functions/sgclaw-chat` as the first delivery host so the new UI can be built and verified without waiting for the final page-floating container. Split the current monolithic Lit component into host adapter, state modules, typed message model, presentational components, and secondary panels so the same UI can later be mounted in a real injected floating window on normal web pages. Preserve the existing browser bridge (`sgclawConnect`, `sgclawStart`, `sgclawStop`, `sgclawSubmitTask`) and re-home logs/configuration into secondary panels instead of deleting them.
+
+**Tech Stack:** Chromium WebUI, Lit, existing `FunctionsUI` router, SuperRPA browser bridge callbacks, current `sgclaw-config` config page logic, future floating host injection in SuperRPA.
+
+## Product Target
+
+The frontend target is a single-column chat product, not a multi-card debug workstation.
+
+Final visual structure:
+
+```text
+Collapsed Fab
+┌────────────┐
+│ sgClaw  ●2 │
+└────────────┘
+
+Expanded Chat
+┌──────────────────────────────────────────┐
+│ sgClaw | 当前网页：example.com           │
+│ [新对话] [历史] [设置] [收起]           │
+│ 状态：待命 / 执行中 / 出错              │
+├──────────────────────────────────────────┤
+│ 欢迎区 / 推荐动作                        │
+│ [总结当前页面] [提取表格] [执行网页操作] │
+├──────────────────────────────────────────┤
+│ 消息流                                   │
+│ 用户消息                                 │
+│ 助手消息                                 │
+│ 步骤卡 / 结果卡 / 错误卡                 │
+├──────────────────────────────────────────┤
+│ [网页执行] [页面问答] [页面总结]         │
+│ [上下文开关] [调试] [更多]              │
+│ ┌──────────────────────────────────────┐ │
+│ │ 输入任务...                          │ │
+│ └──────────────────────────────────────┘ │
+│                                    [发送]│
+└──────────────────────────────────────────┘
+```
+
+Core UX rules:
+
+- The primary content area is always the message stream.
+- `finalResult` becomes a result card inside the message stream.
+- `logs` move into a hidden debug drawer.
+- `start/stop` remain available but move to the header status area.
+- Configuration remains available but opens inside a settings panel first, with route-navigation fallback to `chrome://superrpa-functions/sgclaw-config`.
+- The same component tree must work in `FunctionsUI` first and later inside a real injected floating host.
+
+## Scope
+
+### In Scope For This Frontend Plan
+
+- Complete visual redesign of `sgclaw-chat`
+- Empty state, active chat state, running state, success state, error state
+- Local conversation history UI
+- Embedded settings panel
+- Debug drawer
+- Stable typed message model
+- Separation of host bridge code from UI code
+- Floating launcher state model
+
+### Explicitly Out Of Scope For First Frontend Delivery
+
+- Real attachment upload execution
+- Deep-thinking or multi-skill plugin ecosystem
+- Provider/protocol redesign on the Rust side
+- Full page-injected floating host implementation
+- New backend APIs beyond the current bridge
+
+## Existing Baseline To Reuse
+
+The implementation should reuse these existing assets instead of replacing them blindly:
+
+- Host page routing: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/functions.ts`
+- Existing chat entry registration: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/functions_manifest.json`
+- Current chat page bridge logic: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat.ts`
+- Current floating state prototype: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-floating_state.ts`
+- Current config UI and bridge: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-config/sgclaw-config.ts`
+
+## Final File Layout
+
+All implementation paths below are exact and rooted in `/home/zyl/projects/superRpa/src`.
+
+### Core Chat Entry
+
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat.ts`
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat.html.ts`
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat.css.ts`
+
+### State Modules
+
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-floating_state.ts`
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_window_state.ts`
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_conversation_state.ts`
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_history_state.ts`
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_settings_state.ts`
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_state.ts`
+
+### Host Adapter
+
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_host_adapter.ts`
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_runtime_bridge.ts`
+
+### Message Model And Rendering
+
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_messages.ts`
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-message-list.ts`
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-message-card-user.ts`
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-message-card-assistant.ts`
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-message-card-step.ts`
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-message-card-result.ts`
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-message-card-error.ts`
+
+### Shell Components
+
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-chat-shell.ts`
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-chat-header.ts`
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-chat-composer.ts`
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-history-panel.ts`
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-settings-panel.ts`
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-debug-drawer.ts`
+
+### Build And Host Wiring
+
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/BUILD.gn`
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/functions.html.ts`
+
+### Tests
+
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-floating_state_mainline_unittest.ts`
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_mainline_unittest.ts`
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_window_state_mainline_unittest.ts`
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_history_state_mainline_unittest.ts`
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_settings_state_mainline_unittest.ts`
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_runtime_bridge_mainline_unittest.ts`
+
+## Target State Model
+
+Use a typed model instead of the current loose shape.
+
+```ts
+interface SgClawChatWindowState {
+  windowOpen: boolean;
+  activePanel: 'chat' | 'history' | 'settings' | 'debug';
+  unreadCount: number;
+}
+
+interface SgClawChatConversationState {
+  conversationId: string;
+  draftInput: string;
+  mode: 'web-action' | 'page-qa' | 'page-summary';
+  contextEnabled: boolean;
+  messages: SgClawMessage[];
+}
+
+interface SgClawMessage {
+  id: string;
+  type: 'user_text' | 'assistant_text' | 'task_step' | 'task_result' | 'task_error' | 'system_notice';
+  role: 'user' | 'assistant' | 'system';
+  content: string;
+  status?: 'pending' | 'running' | 'done' | 'failed';
+  timestamp: number;
+  meta?: Record<string, unknown>;
+}
+```
+
+The current `logs`, `messages`, `finalResult`, `pendingReply`, and `busy` state should be re-expressed through these typed stores instead of being owned directly by the entry component.
+
+## Task 1: Freeze The Current Entry And Enable Real Template/CSS Modules
+
+**Files:**
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat.ts`
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat.html.ts`
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat.css.ts`
+- Test: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_mainline_unittest.ts`
+
+**Step 1: Write the failing structure test**
+
+Add assertions that the entry no longer hardcodes the full DOM layout in `render()` and imports its shell template/style helpers.
+
+**Step 2: Run test to verify it fails**
+
+Run:
+```bash
+autoninja -C /home/zyl/projects/superRpa/src/out/KylinRelease sgclaw-chat_build_ts
+```
+
+Expected: fail because `sgclaw-chat.html.ts` and `sgclaw-chat.css.ts` are empty and the new test expects real exports.
+
+**Step 3: Write the minimal implementation**
+
+- Move root shell markup to `getHtml()`
+- Move root style tokens/layout to `getCss()`
+- Keep `sgclaw-chat.ts` focused on state + events
+
+**Step 4: Run test to verify it passes**
+
+Run the same build target.
+
+Expected: TS build succeeds and the entry uses external template/style helpers.
+
+**Step 5: Commit**
+
+```bash
+git -C /home/zyl/projects/superRpa/src add \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat.ts \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat.html.ts \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat.css.ts \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_mainline_unittest.ts
+git -C /home/zyl/projects/superRpa/src commit -m "refactor: extract sgclaw chat shell template"
+```
+
+## Task 2: Build The Window, Conversation, History, And Settings State Modules
+
+**Files:**
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-floating_state.ts`
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_window_state.ts`
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_conversation_state.ts`
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_history_state.ts`
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_settings_state.ts`
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_state.ts`
+- Test: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-floating_state_mainline_unittest.ts`
+- Test: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_window_state_mainline_unittest.ts`
+- Test: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_history_state_mainline_unittest.ts`
+- Test: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_settings_state_mainline_unittest.ts`
+
+**Step 1: Write the failing pure-state tests**
+
+Cover:
+- open/close/switch panel transitions
+- unread count clear on open
+- create/reset conversation
+- local history push/select/remove
+- settings draft dirty detection
+
+**Step 2: Run tests to verify RED**
+
+Run:
+```bash
+autoninja -C /home/zyl/projects/superRpa/src/out/KylinRelease sgclaw-chat_build_ts
+```
+
+Expected: build fails because the new modules and tests do not exist yet.
+
+**Step 3: Write the minimal implementation**
+
+Implement pure functions only. Do not mix DOM work into these modules.
+
+**Step 4: Run tests to verify GREEN**
+
+Run the same build target.
+
+Expected: all pure-state modules compile and their tests pass.
+
+**Step 5: Commit**
+
+```bash
+git -C /home/zyl/projects/superRpa/src add \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-floating_state.ts \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_window_state.ts \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_conversation_state.ts \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_history_state.ts \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_settings_state.ts \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_state.ts \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-floating_state_mainline_unittest.ts \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_window_state_mainline_unittest.ts \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_history_state_mainline_unittest.ts \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_settings_state_mainline_unittest.ts
+git -C /home/zyl/projects/superRpa/src commit -m "feat: add sgclaw chat state modules"
+```
+
+## Task 3: Introduce A Host Adapter So UI Stops Owning Bridge Details
+
+**Files:**
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_host_adapter.ts`
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_runtime_bridge.ts`
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat.ts`
+- Test: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_runtime_bridge_mainline_unittest.ts`
+
+**Step 1: Write the failing bridge test**
+
+Test that:
+- `connect()` issues `sgclawConnect`
+- `start()` issues `sgclawStart`
+- `stop()` issues `sgclawStop`
+- `submitTask()` issues `sgclawSubmitTask`
+- callback payload parsing is handled in one place
+
+**Step 2: Run test to verify RED**
+
+Run:
+```bash
+autoninja -C /home/zyl/projects/superRpa/src/out/KylinRelease sgclaw-chat_build_ts
+```
+
+Expected: fail because adapter modules do not exist.
+
+**Step 3: Write minimal implementation**
+
+- Wrap `chrome.send`
+- Centralize callback registration
+- Return typed runtime events/state to the UI layer
+
+**Step 4: Run test to verify GREEN**
+
+Run the same build target.
+
+Expected: adapter tests compile and pass.
+
+**Step 5: Commit**
+
+```bash
+git -C /home/zyl/projects/superRpa/src add \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_host_adapter.ts \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_runtime_bridge.ts \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat.ts \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_runtime_bridge_mainline_unittest.ts
+git -C /home/zyl/projects/superRpa/src commit -m "refactor: isolate sgclaw chat host bridge"
+```
+
+## Task 4: Replace The Loose Message Format With Typed Cards In The Message Stream
+
+**Files:**
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_messages.ts`
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-message-list.ts`
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-message-card-user.ts`
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-message-card-assistant.ts`
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-message-card-step.ts`
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-message-card-result.ts`
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-message-card-error.ts`
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat.ts`
+- Test: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_mainline_unittest.ts`
+
+**Step 1: Write the failing rendering test**
+
+Add expectations that:
+- empty state shows guidance instead of a blank box
+- `task_complete` renders a result card in the message stream
+- `error` renders an error card in the message stream
+- `pendingReply` renders an assistant pending card
+
+**Step 2: Run test to verify RED**
+
+Run the TS build target.
+
+Expected: fail because message types and card components do not exist.
+
+**Step 3: Write minimal implementation**
+
+- Keep the message list single-column
+- Preserve current user/assistant turn behavior
+- Move `finalResult` handling into result-card rendering
+- Move error display into message flow
+
+**Step 4: Run test to verify GREEN**
+
+Run the same build target.
+
+Expected: cards render correctly and the old standalone result area is no longer required.
+
+**Step 5: Commit**
+
+```bash
+git -C /home/zyl/projects/superRpa/src add \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_messages.ts \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-message-list.ts \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-message-card-user.ts \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-message-card-assistant.ts \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-message-card-step.ts \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-message-card-result.ts \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-message-card-error.ts \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat.ts \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_mainline_unittest.ts
+git -C /home/zyl/projects/superRpa/src commit -m "feat: add sgclaw chat message cards"
+```
+
+## Task 5: Build The Real Header, Empty State, And Composer
+
+**Files:**
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-chat-shell.ts`
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-chat-header.ts`
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-chat-composer.ts`
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat.html.ts`
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat.css.ts`
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat.ts`
+- Test: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_mainline_unittest.ts`
+
+**Step 1: Write the failing shell test**
+
+Assert that the rendered page now contains:
+- header with title, current page label, and status pill
+- empty state recommendation buttons
+- fixed composer at the bottom
+- no standalone `实时日志` or `最终结果` primary sections
+
+**Step 2: Run test to verify RED**
+
+Run the TS build target.
+
+Expected: fail because the shell components do not exist.
+
+**Step 3: Write minimal implementation**
+
+- Header: title, page context, new-chat/history/settings/collapse actions
+- Empty state: 3 to 4 recommended actions
+- Composer: text input, send button, mode toggles, context switch
+
+**Step 4: Run test to verify GREEN**
+
+Run the same build target.
+
+Expected: the page renders as a product-style chat shell.
+
+**Step 5: Commit**
+
+```bash
+git -C /home/zyl/projects/superRpa/src add \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-chat-shell.ts \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-chat-header.ts \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-chat-composer.ts \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat.html.ts \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat.css.ts \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat.ts \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_mainline_unittest.ts
+git -C /home/zyl/projects/superRpa/src commit -m "feat: add sgclaw chat shell and composer"
+```
+
+## Task 6: Embed Settings And Move Raw Logs Into A Debug Drawer
+
+**Files:**
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-settings-panel.ts`
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-debug-drawer.ts`
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat.ts`
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat.css.ts`
+- Reuse Read-Only Reference: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-config/sgclaw-config.ts`
+- Test: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_settings_state_mainline_unittest.ts`
+
+**Step 1: Write the failing panel tests**
+
+Cover:
+- opening settings panel from header
+- editing embedded config draft
+- opening debug drawer and showing logs
+- closing secondary panels without destroying the chat draft
+
+**Step 2: Run test to verify RED**
+
+Run the TS build target.
+
+Expected: fail because secondary panel components do not exist.
+
+**Step 3: Write minimal implementation**
+
+- Reuse config field structure from `sgclaw-config`
+- Keep raw logs in debug only
+- Preserve route-navigation fallback for full config page if embedded save/load fails
+
+**Step 4: Run test to verify GREEN**
+
+Run the same build target.
+
+Expected: settings and debug layers behave as secondary panels instead of separate pages.
+
+**Step 5: Commit**
+
+```bash
+git -C /home/zyl/projects/superRpa/src add \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-settings-panel.ts \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-debug-drawer.ts \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat.ts \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat.css.ts \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_settings_state_mainline_unittest.ts
+git -C /home/zyl/projects/superRpa/src commit -m "feat: add sgclaw settings panel and debug drawer"
+```
+
+## Task 7: Add Local Conversation History And New-Chat Recovery
+
+**Files:**
+- Create: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-history-panel.ts`
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_history_state.ts`
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat.ts`
+- Test: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_history_state_mainline_unittest.ts`
+
+**Step 1: Write the failing history tests**
+
+Cover:
+- saving a conversation preview to local history
+- creating a fresh conversation resets message stream but keeps config
+- reopening a history item restores messages and draft
+
+**Step 2: Run test to verify RED**
+
+Run the TS build target.
+
+Expected: fail because history panel and persistence behavior do not exist.
+
+**Step 3: Write minimal implementation**
+
+- Store history locally in browser storage or localStorage
+- Keep only small metadata + message snapshot for first version
+- No backend schema change in this phase
+
+**Step 4: Run test to verify GREEN**
+
+Run the same build target.
+
+Expected: local conversation switching works fully in the frontend.
+
+**Step 5: Commit**
+
+```bash
+git -C /home/zyl/projects/superRpa/src add \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/components/sgclaw-history-panel.ts \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_history_state.ts \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat.ts \
+  chrome/browser/resources/superrpa/devtools/functions/sgclaw-chat/sgclaw-chat_history_state_mainline_unittest.ts
+git -C /home/zyl/projects/superRpa/src commit -m "feat: add sgclaw local conversation history"
+```
+
+## Task 8: Wire New Shell Assets Into BUILD And Polish The Host Page
+
+**Files:**
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/BUILD.gn`
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/functions.html.ts`
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/devtools/functions/functions.css`
+
+**Step 1: Write the failing host expectation**
+
+Add a small host-level check that:
+- `sgclaw-chat` still loads from the manifest
+- host quick actions still work
+- the function page provides enough room for the new chat shell
+
+**Step 2: Run test/build to verify RED**
+
+Run:
+```bash
+autoninja -C /home/zyl/projects/superRpa/src/out/KylinRelease sgclaw-chat_build_ts
+```
+
+Expected: fail or render incorrectly because new component files are not all wired into build/host styling yet.
+
+**Step 3: Write minimal implementation**
+
+- Add all new TS modules to `BUILD.gn`
+- Keep `sgclaw-chat` and `sgclaw-config` quick actions
+- Adjust host layout so the new shell is not boxed into the old debug-page proportions
+
+**Step 4: Run verification**
+
+Run:
+```bash
+autoninja -C /home/zyl/projects/superRpa/src/out/KylinRelease sgclaw-chat_build_ts
+autoninja -C /home/zyl/projects/superRpa/src/out/KylinRelease superrpa_resources
+```
+
+Expected: build completes with all new chat modules wired in.
+
+**Step 5: Commit**
+
+```bash
+git -C /home/zyl/projects/superRpa/src add \
+  chrome/browser/resources/superrpa/devtools/BUILD.gn \
+  chrome/browser/resources/superrpa/devtools/functions/functions.html.ts \
+  chrome/browser/resources/superrpa/devtools/functions/functions.css
+git -C /home/zyl/projects/superRpa/src commit -m "chore: wire sgclaw chat frontend modules"
+```
+
+## Manual Verification Matrix
+
+Run all manual checks in `chrome://superrpa-functions/sgclaw-chat` after the full frontend plan lands.
+
+### UX States
+
+- Empty state appears on first open.
+- Recommended actions generate user messages.
+- Composer stays visible while history/settings/debug panels switch.
+- Message stream auto-scrolls to the latest item.
+- Result cards and error cards appear inline.
+
+### Runtime
+
+- `启动` works from the header area.
+- `停止` works from the header area.
+- submit creates an immediate user message.
+- pending assistant card appears while waiting.
+- result card replaces the old standalone result behavior.
+
+### Settings
+
+- embedded settings loads existing values
+- save updates status and clears dirty state
+- fallback route to `chrome://superrpa-functions/sgclaw-config` still works
+
+### Debug
+
+- logs are not visible in the main chat view
+- debug drawer shows raw logs when opened
+
+### History
+
+- new conversation starts clean
+- previous conversation can be restored from local history
+- unread badge clears when reopening the window
+
+## Execution Notes
+
+- Keep the current backend/runtime bridge unchanged until the new frontend shell is stable.
+- Do not combine page-injected floating host work into this same branch. The first milestone is a complete product-grade frontend inside the existing `FunctionsUI` host.
+- When this frontend plan is complete, the next plan should focus only on mounting the same component tree inside a real page floating container.
+
+Plan complete and saved to `docs/plans/2026-03-27-sgclaw-floating-chat-frontend-design.md`. Two execution options:
+
+**1. Subagent-Driven (this session)** - I dispatch fresh subagent per task, review between tasks, fast iteration
+
+**2. Parallel Session (separate)** - Open new session with executing-plans, batch execution with checkpoints
+
+**Which approach?**
--- a/docs/plans/2026-03-27-sgclaw-overlay-and-basic-navigation-fixes.md
+++ b/docs/plans/2026-03-27-sgclaw-overlay-and-basic-navigation-fixes.md
@@ -0,0 +1,85 @@
+# sgClaw Overlay And Basic Navigation Fixes Implementation Plan
+
+> **For Claude:** REQUIRED SUB-SKILL: Use superpowers:executing-plans to implement this plan task-by-task.
+
+**Goal:** Make ordinary webpages render the new sgClaw floating chat design and support base navigation instructions like `打开知乎`.
+
+**Architecture:** Keep the ordinary-page injection entrypoint unchanged, but replace its in-shadow DOM layout with the same floating-window shell used by the new debug page. On the runtime side, extend the deterministic planner with explicit homepage navigation plans for supported sites so freeform open-site commands do not fail before the compat runtime can help.
+
+**Tech Stack:** Chromium WebUI resource pipeline, injected Shadow DOM overlay JavaScript, Rust planner tests
+
+### Task 1: Lock the current regressions with failing tests
+
+**Files:**
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/superrpa/sgclaw/sgclaw_chat_smoke.mjs`
+- Modify: `/home/zyl/projects/sgClaw/claw/tests/planner_test.rs`
+
+**Step 1: Write the failing smoke expectations**
+
+Add assertions that the ordinary webpage overlay shows the new subtitle `面向当前网页的悬浮式对话与自动执行` and no longer exposes the old card titles like `聊天记录`.
+
+**Step 2: Run the smoke to verify it fails**
+
+Run: `node /home/zyl/projects/superRpa/src/chrome/browser/superrpa/sgclaw/sgclaw_chat_smoke.mjs`
+Expected: FAIL because ordinary webpages still render the old overlay shell.
+
+**Step 3: Write the failing planner test**
+
+Add a test asserting `plan_instruction("打开知乎")` returns one `Navigate` step to `https://www.zhihu.com`.
+
+**Step 4: Run the planner test to verify it fails**
+
+Run: `cargo test planner_supports_open_zhihu_homepage_instruction --test planner_test`
+Expected: FAIL with `unsupported instruction: 打开知乎`.
+
+### Task 2: Migrate the ordinary webpage overlay to the new shell
+
+**Files:**
+- Modify: `/home/zyl/projects/superRpa/src/chrome/browser/resources/superrpa/sgclaw_overlay.js`
+- Test: `/home/zyl/projects/superRpa/src/chrome/browser/superrpa/sgclaw/sgclaw_chat_smoke.mjs`
+
+**Step 1: Replace the old card layout with the new floating shell**
+
+Keep bridge calls, ids, and polling behavior intact, but render the new header, message pane, composer, settings panel, and debug drawer structure inside the existing injected Shadow DOM.
+
+**Step 2: Keep runtime visibility without reintroducing the old layout**
+
+Move logs and final result into secondary panels or inline cards so the ordinary webpage still exposes execution details without the old four-card layout.
+
+**Step 3: Run the smoke again**
+
+Run: `node /home/zyl/projects/superRpa/src/chrome/browser/superrpa/sgclaw/sgclaw_chat_smoke.mjs`
+Expected: PASS once rebuilt resources are being served by the browser binary.
+
+### Task 3: Extend planner support for basic open-site commands
+
+**Files:**
+- Modify: `/home/zyl/projects/sgClaw/claw/src/agent/planner.rs`
+- Test: `/home/zyl/projects/sgClaw/claw/tests/planner_test.rs`
+
+**Step 1: Implement the minimal homepage plans**
+
+Support `打开知乎` and `打开百度` by returning single-step `Navigate` plans to their homepages.
+
+**Step 2: Run planner tests**
+
+Run: `cargo test --test planner_test`
+Expected: PASS.
+
+### Task 4: Build and verify the integrated behavior
+
+**Files:**
+- Modify: `/home/zyl/projects/superRpa/src/AGENTS.md`
+- Modify: `/home/zyl/projects/superRpa/src/docs/handoffs/2026-03-27-sgclaw-runtime-verification.md`
+
+**Step 1: Rebuild impacted targets**
+
+Run: `autoninja -C /home/zyl/projects/superRpa/src/out/KylinRelease chrome/browser/resources/superrpa:resources sgclaw`
+
+**Step 2: Re-run targeted verification**
+
+Run the smoke and a focused `sgclaw` task submission check for `打开知乎`.
+
+**Step 3: Document the final runtime path**
+
+Record that ordinary webpages and `chrome://superrpa-functions/sgclaw-chat` now share the same floating shell, and that homepage navigation commands are handled by the planner.
--- a/docs/plans/2026-03-27-skill-lib-testing-plan.md
+++ b/docs/plans/2026-03-27-skill-lib-testing-plan.md
@@ -0,0 +1,158 @@
+# Skill Lib Testing Implementation Plan
+
+> **For Claude:** REQUIRED SUB-SKILL: Use superpowers:executing-plans to implement this plan task-by-task.
+
+**Goal:** Add an in-project, repeatable test harness that validates `/home/zyl/projects/sgClaw/skill_lib` against the current ZeroClaw `SKILL.md` loader and security-audit expectations.
+
+**Architecture:** Keep the test runner inside the SGClaw repository and target the sibling `skill_lib` directory by relative path. Implement a small Python validator that mirrors the ZeroClaw markdown frontmatter parser and the relevant skill-audit checks, then cover it with a Python `unittest` suite that exercises the actual three migrated Zhihu skills.
+
+**Tech Stack:** Python 3 standard library, `unittest`, local file-system inspection, ZeroClaw source code as behavioral reference, Markdown/YAML-like frontmatter parsing.
+
+### Task 1: Freeze The Test Contract
+
+**Files:**
+- Create: `/home/zyl/projects/sgClaw/claw/docs/plans/2026-03-27-skill-lib-testing-plan.md`
+- Reference only: `/home/zyl/projects/sgClaw/claw/third_party/zeroclaw/src/skills/mod.rs`
+- Reference only: `/home/zyl/projects/sgClaw/claw/third_party/zeroclaw/src/skills/audit.rs`
+- Reference only: `/home/zyl/projects/sgClaw/skill_lib/skills/*/SKILL.md`
+
+**Step 1: Capture the loader semantics to preserve**
+
+Document and implement tests for:
+- `SKILL.md` frontmatter splitting on `---`
+- supported metadata keys: `name`, `description`, `version`, `author`, `tags`
+- fallback rules for name, description, and version
+- prompt body must exclude the frontmatter block
+
+**Step 2: Capture the audit semantics to preserve**
+
+Document and implement tests for:
+- skill root must contain `SKILL.md` or `SKILL.toml`
+- symlinks are rejected
+- shell-script files are blocked when `allow_scripts` is false
+- markdown links must not escape the skill root
+- high-risk command snippets inside markdown are rejected
+
+**Step 3: Define the migrated-skill expectations**
+
+The test suite must verify:
+- exactly three skill packages exist
+- the loaded names are `zhihu-hotlist`, `zhihu-navigate`, `zhihu-write`
+- each package has both `references/` and `assets/`
+- each description stays trigger-oriented and starts with `Use when`
+
+### Task 2: Write The Failing Tests First
+
+**Files:**
+- Create: `/home/zyl/projects/sgClaw/claw/tests/skill_lib_validation_test.py`
+
+**Step 1: Write a failing import-level test**
+
+Import a not-yet-created validator module from:
+- `/home/zyl/projects/sgClaw/claw/scripts/validate_skill_lib.py`
+
+Expected initial failure:
+- `ModuleNotFoundError` or `FileNotFoundError`
+
+**Step 2: Encode the project expectations**
+
+Add tests for:
+- skill discovery count and names
+- parsed metadata for each current skill
+- audit cleanliness for each skill with `allow_scripts=False`
+- package shape (`SKILL.md`, `references/`, `assets/`)
+
+**Step 3: Run the tests and watch them fail**
+
+Run:
+```bash
+python3 -m unittest tests.skill_lib_validation_test -v
+```
+
+Expected:
+- failure because the validator module does not exist yet
+
+### Task 3: Implement The Minimal Validator
+
+**Files:**
+- Create: `/home/zyl/projects/sgClaw/claw/scripts/validate_skill_lib.py`
+
+**Step 1: Implement discovery helpers**
+
+Implement:
+- repo root resolution
+- sibling `skill_lib` root resolution
+- `skills/` directory enumeration
+
+**Step 2: Implement the markdown loader**
+
+Implement:
+- frontmatter split
+- lightweight frontmatter parsing
+- description fallback extraction
+- metadata normalization into a `SkillRecord`
+
+**Step 3: Implement the relevant audit checks**
+
+Implement:
+- symlink detection
+- blocked shell-script detection
+- markdown link boundary checks
+- high-risk snippet detection
+- deterministic findings collection
+
+**Step 4: Implement a small CLI**
+
+Running:
+```bash
+python3 scripts/validate_skill_lib.py
+```
+
+Should:
+- print one summary line per skill
+- exit `0` when all skills pass
+- exit non-zero when any skill fails
+
+### Task 4: Run The Tests Green
+
+**Files:**
+- Test: `/home/zyl/projects/sgClaw/claw/tests/skill_lib_validation_test.py`
+- Test: `/home/zyl/projects/sgClaw/claw/scripts/validate_skill_lib.py`
+
+**Step 1: Re-run the unit tests**
+
+Run:
+```bash
+python3 -m unittest tests.skill_lib_validation_test -v
+```
+
+Expected:
+- all tests pass
+
+**Step 2: Run the CLI validator**
+
+Run:
+```bash
+python3 scripts/validate_skill_lib.py
+```
+
+Expected:
+- all three skills print `PASS`
+- process exits `0`
+
+### Task 5: Document The Verification Entry Point
+
+**Files:**
+- Modify: `/home/zyl/projects/sgClaw/skill_lib/VERIFY.md`
+
+**Step 1: Add the project-local validation command**
+
+Add:
+- `python3 /home/zyl/projects/sgClaw/claw/scripts/validate_skill_lib.py`
+- `python3 -m unittest /home/zyl/projects/sgClaw/claw/tests/skill_lib_validation_test.py`
+
+**Step 2: Re-run both commands after the doc update**
+
+Expected:
+- validator still exits `0`
+- unit tests still pass
--- a/docs/plans/2026-03-27-skill-lib-zeroclaw-plan.md
+++ b/docs/plans/2026-03-27-skill-lib-zeroclaw-plan.md
@@ -0,0 +1,411 @@
+# Skill Lib ZeroClaw Migration Implementation Plan
+
+> **For Claude:** REQUIRED SUB-SKILL: Use superpowers:executing-plans to implement this plan task-by-task.
+
+**Goal:** Create `/home/zyl/projects/sgClaw/skill_lib` as a dedicated skill library directory and restructure the current Zhihu browser capabilities into ZeroClaw-style skill packages.
+
+**Architecture:** Treat `skill_lib` as a standalone skill repository, not as an embedded Rust module tree. Use the ZeroClaw/open-skills layout `skill_lib/skills/<skill-name>/SKILL.md`, keep each skill self-contained, and move long operational detail into `references/` plus any preserved source artifacts into `assets/`. Map the current four exposed Rust capabilities into three end-user skills: `zhihu-navigate`, `zhihu-write`, and `zhihu-hotlist`.
+
+**Tech Stack:** Markdown `SKILL.md`, YAML frontmatter, directory-based ZeroClaw skill packaging, existing SGClaw Zhihu Rust/JSON source material, shell validation commands.
+
+### Task 1: Freeze The Target Layout
+
+**Files:**
+- Create: `/home/zyl/projects/sgClaw/skill_lib/`
+- Create: `/home/zyl/projects/sgClaw/skill_lib/README.md`
+- Create: `/home/zyl/projects/sgClaw/skill_lib/skills/`
+- Reference only: `/home/zyl/projects/sgClaw/claw/third_party/zeroclaw/src/skills/mod.rs`
+- Reference only: `/home/zyl/projects/sgClaw/claw/third_party/zeroclaw/skills/browser/SKILL.md`
+
+**Step 1: Create the top-level repository skeleton**
+
+Create:
+- `/home/zyl/projects/sgClaw/skill_lib/README.md`
+- `/home/zyl/projects/sgClaw/skill_lib/skills/`
+
+The README should state:
+- this directory is a dedicated ZeroClaw-style skill library
+- runtime skill packages live under `skills/<name>/`
+- each skill package uses `SKILL.md` plus optional `references/`, `scripts/`, and `assets/`
+
+**Step 2: Document the package contract in the README**
+
+Include:
+- required file: `SKILL.md`
+- supported frontmatter for this repo: `name`, `description`, `version`, `author`, `tags`
+- design rule: `description` must be trigger-oriented and not a workflow dump
+- design rule: keep `SKILL.md` concise and push long detail into `references/`
+
+**Step 3: Run structural sanity checks**
+
+Run:
+```bash
+test -d /home/zyl/projects/sgClaw/skill_lib
+test -d /home/zyl/projects/sgClaw/skill_lib/skills
+test -f /home/zyl/projects/sgClaw/skill_lib/README.md
+```
+
+Expected: all commands exit `0`.
+
+### Task 2: Define The Skill Inventory And Source Mapping
+
+**Files:**
+- Create: `/home/zyl/projects/sgClaw/skill_lib/skill_inventory.md`
+- Reference only: `/home/zyl/projects/sgClaw/claw/src/skill/mod.rs`
+- Reference only: `/home/zyl/projects/sgClaw/claw/src/skill/router.rs`
+- Reference only: `/home/zyl/projects/sgClaw/claw/src/skill/zhihu.rs`
+- Reference only: `/home/zyl/projects/sgClaw/claw/src/skill/zhihu_hotlist.rs`
+- Reference only: `/home/zyl/projects/sgClaw/claw/src/skill/zhihu_hotlist_store.rs`
+- Reference only: `/home/zyl/projects/sgClaw/claw/src/skill/zhihu_navigation.rs`
+- Reference only: `/home/zyl/projects/sgClaw/claw/resources/skills/zhihu_write_flow.json`
+- Reference only: `/home/zyl/projects/sgClaw/claw/resources/skills/zhihu_hotlist_flow.json`
+- Reference only: `/home/zyl/projects/sgClaw/claw/resources/skills/zhihu_navigation_pages.json`
+
+**Step 1: Write the migration inventory**
+
+Create `/home/zyl/projects/sgClaw/skill_lib/skill_inventory.md` with a three-row mapping:
+- `zhihu-navigate` ← current `zhihu_navigate`
+- `zhihu-write` ← current `zhihu_write`
+- `zhihu-hotlist` ← current `zhihu_hotlist_collect` + `zhihu_hotlist_report`
+
+**Step 2: Capture the non-migrated code responsibilities**
+
+Document explicitly that this migration does **not** port:
+- Rust router dispatch
+- browser pipe transport code
+- snapshot persistence implementation detail
+
+Document that the new repo is a skill library, not a Rust runtime.
+
+**Step 3: Record source artifacts per target skill**
+
+For each target skill, list:
+- source Rust module(s)
+- source JSON flow/catalog file(s)
+- important risk notes discovered during analysis
+
+**Step 4: Validate the inventory**
+
+Run:
+```bash
+rg -n "zhihu-navigate|zhihu-write|zhihu-hotlist" /home/zyl/projects/sgClaw/skill_lib/skill_inventory.md
+```
+
+Expected: all three skill names appear exactly once as top-level migration targets.
+
+### Task 3: Author The `zhihu-navigate` Skill Package
+
+**Files:**
+- Create: `/home/zyl/projects/sgClaw/skill_lib/skills/zhihu-navigate/SKILL.md`
+- Create: `/home/zyl/projects/sgClaw/skill_lib/skills/zhihu-navigate/references/routes-and-targets.md`
+- Create: `/home/zyl/projects/sgClaw/skill_lib/skills/zhihu-navigate/references/selector-strategy.md`
+- Create: `/home/zyl/projects/sgClaw/skill_lib/skills/zhihu-navigate/assets/zhihu_navigation_pages.source.json`
+- Reference only: `/home/zyl/projects/sgClaw/claw/src/skill/zhihu_navigation.rs`
+- Reference only: `/home/zyl/projects/sgClaw/claw/resources/skills/zhihu_navigation_pages.json`
+
+**Step 1: Preserve the raw source artifact**
+
+Copy the current navigation catalog into:
+- `/home/zyl/projects/sgClaw/skill_lib/skills/zhihu-navigate/assets/zhihu_navigation_pages.source.json`
+
+This file is for traceability only, not for frontmatter or prompt injection.
+
+**Step 2: Write the `SKILL.md`**
+
+Use ZeroClaw-style frontmatter:
+```yaml
+---
+name: zhihu-navigate
+description: Use when the user wants to open, switch, or navigate to a Zhihu page, tab, menu, profile area, notifications area, message area, or creator area through browser actions.
+version: 0.1.0
+author: sgclaw
+tags:
+  - zhihu
+  - browser
+  - navigation
+---
+```
+
+The body should include:
+- overview
+- when to use
+- workflow for route vs component vs flow navigation
+- ambiguity handling rules
+- output contract
+- common mistakes
+
+**Step 3: Write `routes-and-targets.md`**
+
+Summarize:
+- route/component/flow/target model
+- representative target names
+- known alias conflicts
+- preferred disambiguation wording for future prompts
+
+**Step 4: Write `selector-strategy.md`**
+
+Document:
+- why selectors should prefer semantic hooks over CSS hash classes
+- fallback ordering
+- known brittle selectors from the current source
+
+**Step 5: Validate the package**
+
+Run:
+```bash
+test -f /home/zyl/projects/sgClaw/skill_lib/skills/zhihu-navigate/SKILL.md
+test -f /home/zyl/projects/sgClaw/skill_lib/skills/zhihu-navigate/references/routes-and-targets.md
+test -f /home/zyl/projects/sgClaw/skill_lib/skills/zhihu-navigate/references/selector-strategy.md
+test -f /home/zyl/projects/sgClaw/skill_lib/skills/zhihu-navigate/assets/zhihu_navigation_pages.source.json
+```
+
+Expected: all commands exit `0`.
+
+### Task 4: Author The `zhihu-write` Skill Package
+
+**Files:**
+- Create: `/home/zyl/projects/sgClaw/skill_lib/skills/zhihu-write/SKILL.md`
+- Create: `/home/zyl/projects/sgClaw/skill_lib/skills/zhihu-write/references/editor-flow.md`
+- Create: `/home/zyl/projects/sgClaw/skill_lib/skills/zhihu-write/references/publish-safety.md`
+- Create: `/home/zyl/projects/sgClaw/skill_lib/skills/zhihu-write/assets/zhihu_write_flow.source.json`
+- Reference only: `/home/zyl/projects/sgClaw/claw/src/skill/zhihu.rs`
+- Reference only: `/home/zyl/projects/sgClaw/claw/resources/skills/zhihu_write_flow.json`
+
+**Step 1: Preserve the raw source artifact**
+
+Copy:
+- `/home/zyl/projects/sgClaw/claw/resources/skills/zhihu_write_flow.json`
+
+to:
+- `/home/zyl/projects/sgClaw/skill_lib/skills/zhihu-write/assets/zhihu_write_flow.source.json`
+
+**Step 2: Write the `SKILL.md`**
+
+The frontmatter should name a single skill:
+- `name: zhihu-write`
+- description focused on when article drafting or publishing is requested
+
+The body should include:
+- prerequisites before touching the editor
+- workflow for draft-only vs publish
+- explicit confirmation gate before publish
+- required final report fields: title, mode, final URL if published, unresolved issues
+
+**Step 3: Write `editor-flow.md`**
+
+Document:
+- entry page
+- editor readiness checks
+- title/body fill rules
+- publish confirmation sequence
+- URL capture rules
+
+**Step 4: Write `publish-safety.md`**
+
+Document:
+- when to stop and ask for confirmation
+- what to do if title verification fails
+- what to do if the URL remains on edit mode
+- brittle selectors that must be revalidated first
+
+**Step 5: Validate the package**
+
+Run:
+```bash
+test -f /home/zyl/projects/sgClaw/skill_lib/skills/zhihu-write/SKILL.md
+test -f /home/zyl/projects/sgClaw/skill_lib/skills/zhihu-write/references/editor-flow.md
+test -f /home/zyl/projects/sgClaw/skill_lib/skills/zhihu-write/references/publish-safety.md
+test -f /home/zyl/projects/sgClaw/skill_lib/skills/zhihu-write/assets/zhihu_write_flow.source.json
+```
+
+Expected: all commands exit `0`.
+
+### Task 5: Author The `zhihu-hotlist` Skill Package
+
+**Files:**
+- Create: `/home/zyl/projects/sgClaw/skill_lib/skills/zhihu-hotlist/SKILL.md`
+- Create: `/home/zyl/projects/sgClaw/skill_lib/skills/zhihu-hotlist/references/collection-flow.md`
+- Create: `/home/zyl/projects/sgClaw/skill_lib/skills/zhihu-hotlist/references/report-format.md`
+- Create: `/home/zyl/projects/sgClaw/skill_lib/skills/zhihu-hotlist/references/data-quality.md`
+- Create: `/home/zyl/projects/sgClaw/skill_lib/skills/zhihu-hotlist/assets/zhihu_hotlist_flow.source.json`
+- Reference only: `/home/zyl/projects/sgClaw/claw/src/skill/zhihu_hotlist.rs`
+- Reference only: `/home/zyl/projects/sgClaw/claw/src/skill/zhihu_hotlist_store.rs`
+- Reference only: `/home/zyl/projects/sgClaw/claw/resources/skills/zhihu_hotlist_flow.json`
+
+**Step 1: Preserve the raw source artifact**
+
+Copy:
+- `/home/zyl/projects/sgClaw/claw/resources/skills/zhihu_hotlist_flow.json`
+
+to:
+- `/home/zyl/projects/sgClaw/skill_lib/skills/zhihu-hotlist/assets/zhihu_hotlist_flow.source.json`
+
+**Step 2: Write the `SKILL.md`**
+
+Use one skill to cover:
+- hotlist collection
+- comment metric collection
+- snapshot-style reporting
+
+The body should clearly separate:
+- collection workflow
+- report workflow
+- partial-failure handling
+- output contract
+
+**Step 3: Write `collection-flow.md`**
+
+Include:
+- hotlist page detection
+- hotlist HTML capture strategy
+- top N extraction
+- detail-page comment collection flow
+- metric parsing notes
+
+**Step 4: Write `report-format.md`**
+
+Define:
+- report header line
+- per-item summary line
+- field names and order
+- handling when comment metrics are missing
+
+**Step 5: Write `data-quality.md`**
+
+Document:
+- why partial success must be surfaced
+- what counts as incomplete data
+- known parser risks
+- recommended caution language in outputs
+
+**Step 6: Validate the package**
+
+Run:
+```bash
+test -f /home/zyl/projects/sgClaw/skill_lib/skills/zhihu-hotlist/SKILL.md
+test -f /home/zyl/projects/sgClaw/skill_lib/skills/zhihu-hotlist/references/collection-flow.md
+test -f /home/zyl/projects/sgClaw/skill_lib/skills/zhihu-hotlist/references/report-format.md
+test -f /home/zyl/projects/sgClaw/skill_lib/skills/zhihu-hotlist/references/data-quality.md
+test -f /home/zyl/projects/sgClaw/skill_lib/skills/zhihu-hotlist/assets/zhihu_hotlist_flow.source.json
+```
+
+Expected: all commands exit `0`.
+
+### Task 6: Normalize Frontmatter And Trigger Quality
+
+**Files:**
+- Modify: `/home/zyl/projects/sgClaw/skill_lib/skills/zhihu-navigate/SKILL.md`
+- Modify: `/home/zyl/projects/sgClaw/skill_lib/skills/zhihu-write/SKILL.md`
+- Modify: `/home/zyl/projects/sgClaw/skill_lib/skills/zhihu-hotlist/SKILL.md`
+
+**Step 1: Normalize frontmatter keys**
+
+Ensure each `SKILL.md` contains exactly these frontmatter keys in this order:
+- `name`
+- `description`
+- `version`
+- `author`
+- `tags`
+
+Do not add Rust-only or unofficial parser fields as required metadata.
+
+**Step 2: Check naming rules**
+
+Ensure skill names are:
+- lowercase
+- hyphenated
+- stable
+
+Names to keep:
+- `zhihu-navigate`
+- `zhihu-write`
+- `zhihu-hotlist`
+
+**Step 3: Tighten descriptions**
+
+Each description must:
+- begin with `Use when`
+- describe triggering conditions
+- mention Zhihu/browser context
+- avoid dumping full workflow detail
+
+**Step 4: Validate frontmatter**
+
+Run:
+```bash
+rg -n "^name: |^description: |^version: |^author: |^tags:" /home/zyl/projects/sgClaw/skill_lib/skills/*/SKILL.md
+```
+
+Expected: every skill emits the same five key families.
+
+### Task 7: Add Repository-Level Verification Notes
+
+**Files:**
+- Create: `/home/zyl/projects/sgClaw/skill_lib/VERIFY.md`
+- Modify: `/home/zyl/projects/sgClaw/skill_lib/README.md`
+
+**Step 1: Create `VERIFY.md`**
+
+Document the manual verification checklist:
+- all skill packages are under `skill_lib/skills/`
+- each package has `SKILL.md`
+- long details live in `references/`
+- preserved source JSON is in `assets/`
+- no Rust source is copied into the skill repo
+
+**Step 2: Link verification from the README**
+
+Add a short section in `README.md` pointing to `VERIFY.md`.
+
+**Step 3: Run repository-level checks**
+
+Run:
+```bash
+find /home/zyl/projects/sgClaw/skill_lib/skills -mindepth 2 -maxdepth 2 -name SKILL.md | sort
+find /home/zyl/projects/sgClaw/skill_lib/skills -type d \( -name references -o -name assets \) | sort
+```
+
+Expected:
+- exactly three `SKILL.md` files
+- each skill has `references/`
+- each skill has `assets/`
+
+### Task 8: Final Review Before Claiming Completion
+
+**Files:**
+- Review only: `/home/zyl/projects/sgClaw/skill_lib/`
+- Review only: `/home/zyl/projects/sgClaw/claw/docs/plans/2026-03-27-skill-lib-zeroclaw-plan.md`
+
+**Step 1: Review against ZeroClaw runtime constraints**
+
+Check that the final library respects the currently observed runtime facts:
+- directory-based skills
+- `SKILL.md` supported
+- simple frontmatter fields
+- optional `references/`, `scripts/`, `assets/`
+
+**Step 2: Review against authoring quality**
+
+Check that each skill:
+- is self-contained
+- has a narrow trigger boundary
+- avoids copying Rust internals into the prompt body
+- surfaces ambiguity and failure modes
+
+**Step 3: Produce the implementation report**
+
+The completion report must include:
+- created directories
+- created skill packages
+- any deliberate deviations from upstream ZeroClaw examples
+- verification commands actually run
+- unresolved risks
+
+**Step 4: Stop before unrelated expansion**
+
+Do not add:
+- extra skills beyond the three mapped ones
+- generic utility libraries
+- unrelated automation scripts
+- runtime code changes in `/home/zyl/projects/sgClaw/claw/src/skill/`
+
--- a/docs/sgClaw项目现有优势与下一步计划汇报稿.md
+++ b/docs/sgClaw项目现有优势与下一步计划汇报稿.md
@@ -0,0 +1,314 @@
+# sgClaw 项目现有优势与下一步计划汇报稿
+
+## 一、项目现有优势
+
+和传统 openclaw 类自动化项目相比，本项目的优势不只是“能做页面操作”，而是已经具备了向企业级、长期可用方向演进的基础。简单说，传统方案更像“能跑起来的自动化脚本集合”，本项目更像“带安全边界、带统一控制、带长期演进能力的智能执行底座”。
+
+而且根据后续已经落地的提交来看，本项目的优势已经不再只是架构上的“方向正确”，而是开始体现为一批已经交付的现实能力，包括：
+
+- 运行时策略已经从写死逻辑转向配置驱动。
+- 已支持 planner-first 的先规划后执行模式。
+- 已支持技能包驱动的浏览器脚本执行。
+- 已支持面向具体任务的 Excel 导出和屏幕展示导出能力。
+- 已具备版本级日志、技能版本日志和真实验收记录。
+
+这意味着本项目已经从“原型底座”进一步进入“可验证、可扩展、可交付”的阶段。
+
+### 1. 从“脚本工具”升级为“统一执行底座”
+
+传统项目通常是一个场景一套脚本，一个系统一套规则，能解决眼前问题，但难以复用、难以管理、难以持续演进。页面一改，脚本就要改；系统一多，维护成本就会快速上升。
+
+本项目已经把任务接入、运行时控制、浏览器执行、日志回传、结果回传放进统一框架里。这样后续不管是新增能力、扩展场景，还是替换模型、替换策略，都不需要推倒重来，而是在同一个底座上持续增强。对业务侧来说，这意味着项目价值不再停留在“做通一个流程”，而是逐步沉淀为可以支撑更多业务的公共能力。
+
+### 2. 安全设计更完整，更适合企业内网
+
+这是本项目最突出的优势。传统自动化方案最大的问题，是一旦拿到页面操作能力，就容易变成“脚本想点什么就点什么”，安全边界不清楚，风险很难控制。本项目当前已经不是这种模式，而是把浏览器能力放进了严格受控的执行边界里。
+
+从汇报角度，可以把当前安全设计概括为以下几组能力：
+
+#### 3 层安全机制
+
+第一层是启动门禁。浏览器宿主和运行时必须先完成握手，版本不一致、超时、顺序错误，系统都不会进入正式运行状态。
+
+第二层是运行时策略校验。每次真正要执行页面动作前，都会先检查动作是否合法、目标页面是否在允许范围内。
+
+第三层是宿主二次校验。就算运行时已经发出了命令，浏览器宿主仍然会再做一次本地校验，防止异常命令真正落地。
+
+这 3 层叠加起来，形成了“不是模型想做什么就做什么，而是每一步都要过关”的安全控制方式。
+
+#### 6 项协议硬约束
+
+当前协议层已经明确了 6 项刚性要求：
+
+1. 传输格式固定，不能随意乱发消息。
+2. 编码方式固定，避免解析异常。
+3. 单条消息大小有限制，防止异常数据冲击系统。
+4. 序列号必须严格递增，不能重复、不能乱序。
+5. 每条关键命令都必须带安全字段。
+6. 一次请求只能对应一次响应，不能混乱回包。
+
+这说明系统不是“发个命令过去试试看”，而是每条消息都有严格规则，更适合企业环境中的稳定运行。
+
+#### 2 类白名单
+
+本项目当前至少有两类核心白名单同时生效。
+
+第一类是域名白名单。只有允许的业务域名才可以被操作，不是浏览器里所有页面都能碰。
+
+第二类是动作白名单。只有允许的动作类型才能执行，不是脚本写得出来就一定能跑。
+
+白名单机制的意义在于，系统把“哪些页面能动、哪些动作能做”提前规定清楚，而不是把决定权完全交给模型或脚本。
+
+#### 1 类显式黑名单
+
+除了白名单，本项目还明确保留了显式阻断项。也就是说，不只是“没允许的不行”，而是“高风险动作被直接明确禁止”。
+
+这在企业场景下非常重要，因为有些能力不是“暂时不用”，而是“原则上就不能开放”。有了黑名单，系统在设计上就能提前规避高风险能力外溢。
+
+#### 5 个默认允许动作
+
+当前默认真正开放给运行时执行的动作共有 5 个：
+
+1. 点击
+2. 输入
+3. 页面跳转
+4. 文本读取
+5. 受控脚本执行
+
+这里最重要的一点不是“多了一个动作”，而是这个新增能力并没有破坏安全边界。它不是把任意页面脚本能力全部放开，而是在现有受控协议和校验链路内，给技能包提供了一种更强但仍然可控的执行方式。
+
+这看起来不如一些传统方案“动作数量多”，但它的价值恰恰在于边界非常清楚。先把最稳定、最可控、最容易审计的核心动作做好，再逐步扩展，比一开始把大量高风险动作全部开放更稳。
+
+#### 7 个默认允许域名
+
+当前规则里默认允许的域名是有限集合，而不是浏览器里的所有网页都能碰。这样做非常符合企业内网环境的实际需求。对于办公系统、ERP、OA 等场景，大家真正需要的不是“全网自动化”，而是“在明确范围内可控地自动化”。
+
+#### 1 套 HMAC 签名机制
+
+所有关键命令不是明文裸发，而是带签名校验。可以简单理解成，每条关键操作都会带“防伪标记”。
+
+这样做的价值是，命令在链路中不容易被伪造、篡改或错误复用，整体安全性远高于普通脚本直接调用页面接口的模式。
+
+#### 1 套序列号机制
+
+每条命令都有严格递增的序列号，而且一个序列号只能对应一个结果。
+
+这让系统能够清楚知道“这条结果到底对应哪一次操作”，避免串包、乱序、错配等问题，提升稳定性和可追溯性。
+
+#### 3 重脚本执行约束
+
+后续提交里新增了技能包驱动的浏览器脚本能力，但这部分并不是“把页面执行彻底放开”，而是在现有安全边界内增加了一层受控能力。
+
+可以把它理解为 3 重约束：
+
+1. 脚本必须来自技能包内的受管路径，不能越界读取技能目录之外的文件。
+2. 执行时必须声明目标域名，不能脱离页面上下文随意运行。
+3. 脚本仍然通过现有浏览器 pipe 和动作白名单执行，而不是绕过宿主直接落地。
+
+这类设计很关键，因为它说明项目在增强能力的同时，仍然坚持“新增能力必须留在安全边界里”，而不是为了方便把安全口子越开越大。
+
+#### 5 类错误处理策略
+
+系统不是失败了就“直接崩”，而是把错误分成不同类型处理。
+
+- 有的错误不允许重试，直接失败。
+- 有的错误可以限次重试。
+- 有的错误需要等待配置或人工确认。
+- 有的错误会触发熔断。
+- 所有失败都要求结构化返回，便于定位问题。
+
+这比传统脚本“报错了就人工重跑”的方式要成熟得多。
+
+#### 1 个熔断阈值
+
+同一动作如果连续失败超过阈值，系统会主动停止继续尝试并通知界面，而不是无限重复。
+
+这能有效避免错误状态下反复点击、反复提交、反复操作，减少业务风险和误操作成本。
+
+#### 7 项联调验收标准
+
+当前项目已经把联调成功的标准写清楚了，包括：
+
+1. 握手成功率要求
+2. 版本不匹配的失败处理
+3. 序列号异常场景处理
+4. 超大消息拦截
+5. 核心动作成功率要求
+6. 结构化错误返回要求
+7. 日志全链路贯通能力
+
+这说明项目不是“靠经验凑合能跑”，而是已经开始形成可以复制、可以验收、可以交付的工程标准。
+
+### 3. 浏览器只是执行面，不再定义整个系统
+
+传统 openclaw 类项目常见问题是浏览器能力太强，最后整个系统都围着页面脚本转，浏览器脚本几乎变成了系统本体。
+
+本项目已经明确把浏览器定义为“受保护的特权执行面”，而不是整个 runtime 本体。这意味着以后就算扩展到别的工具面、别的执行面，也不需要推翻现有架构，系统的演进空间更大，整体结构也更清楚。
+
+### 4. 运行时能力已经从“写死逻辑”升级为“配置驱动”
+
+这一点是后续提交中非常重要的进展。传统项目经常把模型、策略、模式、环境差异写死在代码里，导致后续一改就牵动整体。
+
+本项目现在已经把一批关键决策收进运行时配置，包括：
+
+- 使用哪个模型提供方
+- 当前激活哪个 provider
+- 使用什么 planner 模式
+- 采用哪种 runtime profile
+- 浏览器能力走哪种 backend
+- Office 导出走哪种 backend
+- skills 从哪个目录加载
+
+从汇报口径上，可以把它概括为：
+
+1 套统一 runtime config  
+3 种 runtime profile  
+多项可切换运行策略
+
+这意味着系统不再只是“代码怎么写就怎么跑”，而是开始进入“按环境、按任务、按场景灵活切换”的阶段，更适合企业实际落地。
+
+### 5. 已形成“先规划、再执行、再产出结果”的闭环能力
+
+传统自动化项目往往是一上来就直接操作页面，缺少中间过程的可解释性，也不利于后续审计和治理。
+
+本项目后续提交已经进一步加强了 planner-first 模式，也就是在真正执行之前，先给出计划，再按计划执行，再输出结果。对业务和管理层来说，这样的价值非常直接：
+
+- 更容易理解系统准备做什么
+- 更容易检查执行过程是否偏离目标
+- 更容易把计划、执行、结果串成闭环
+
+同时，本项目已经不是只有一个简单浏览器工具，而是开始形成更清晰的能力分工，例如：
+
+- `superrpa_browser` 负责受控浏览器操作
+- `openxml_office` 负责结果导出
+- `screen_html_export` 负责展示类产物导出
+
+这说明项目正在从“一个浏览器操作入口”走向“围绕业务结果组织工具链”的阶段。
+
+### 6. 技能体系已经开始从“提示词描述”走向“可执行能力包”
+
+这是本项目相对传统 openclaw 类项目非常重要的一个现实优势。很多传统项目里的“技能”更多只是提示词模板，真正落地时还是回到页面脚本堆叠。
+
+本项目后续提交已经支持技能包驱动的浏览器脚本执行。简单理解，就是一个技能不再只是“告诉模型怎么做”，而是可以带着确定的脚本能力一起交付。这样做有几个明显好处：
+
+1. 能力更稳定  
+关键步骤不必完全依赖模型自由发挥，而是可以由打包好的脚本完成。
+
+2. 可复用性更强  
+同一个技能包可以在相似场景中重复使用，不必每次都重新组织页面操作。
+
+3. 更适合沉淀企业资产  
+后续很多高价值流程，都可以逐步从“提示词经验”沉淀成“可复用技能包”。
+
+这意味着项目已经开始从“智能执行框架”走向“智能执行框架 + 可复用技能资产”的模式。
+
+### 7. 前端只负责展示，不掌握执行权
+
+传统项目里，前端、脚本、执行逻辑经常混在一起，最后变成“界面里藏了很多业务决策”。这种方式短期看开发快，长期看风险大、维护成本高。
+
+本项目已经把前端限制为展示层，只负责展示状态、日志、计划和结果，不负责决定是否执行、如何切换模型、如何绕过安全边界。这样一来，系统结构更清楚，后续维护和升级时也更不容易失控。
+
+此外，后续提交已经支持外部 frontend bundle 优先、内置资源兜底的装载方式。这意味着后续改界面、改展示逻辑，不必每次都重编浏览器宿主，研发效率和交付效率都会更高。
+
+### 8. 配置能力更强，更适合业务落地
+
+传统项目往往把很多关键逻辑写死在脚本里，修改一次就要重新改代码。这样不仅效率低，而且很容易因为局部修改牵动整体。
+
+本项目已经开始把运行时配置、模型配置、策略配置从代码里抽出来，让宿主、运行时、前端之间的责任更清楚。这意味着未来业务调整、模型切换、策略升级都可以更平滑，而不是每次都进行大规模改造。
+
+同时，后续提交还进一步加强了 source checkout 启动包装和 rules 同步能力，这对开发团队来说很重要。它意味着项目不仅适合做成二进制交付，也更适合在源码态持续联调和快速迭代。
+
+### 9. 更适合做长期资产沉淀
+
+传统自动化方案常见的问题，是做完一个流程后，价值基本也就结束了，经验很难积累成资产。
+
+本项目不一样，它的方向是把执行能力、规则、安全边界、日志能力以及后续的元素识别能力，逐步沉淀成可复用资产。对企业来说，这种价值远高于“今天跑通一个流程”，因为它决定了未来是不是能够越做越快、越做越稳、越做越便宜。
+
+现在这件事已经开始有现实支撑了。因为项目不只是在“能操作页面”，还已经能把技能、脚本、导出流程、运行时策略和日志标准逐步固化下来。后续再推进“混合自愈选择器”和元素指纹库时，这些都会自然成为资产沉淀的基础层。
+
+### 10. 可观测性更强，已经开始具备运行级审计基础
+
+传统项目常常只在失败时打印一段日志，出了问题很难知道系统到底做了什么。
+
+本项目后续提交已经补上了一批很关键的运行级日志信息，包括：
+
+- runtime 版本
+- 协议版本
+- 配置来源
+- skills 目录解析结果
+- runtime profile
+- skills prompt mode
+- 已加载技能及版本号
+- 当前执行模式
+
+这类能力的价值非常直接：它让系统开始具备“说清楚自己是怎么运行的”的能力。对研发、测试、验收和后续审计来说，这都是非常重要的基础。
+
+### 11. 已经形成“真实验收”而不是“概念演示”
+
+后续提交里，项目已经留下了更完整的验收记录，而不是停留在文档层面的能力描述。以知乎热榜 Excel 导出为例，当前已经形成真实验收结果，包括：
+
+- 真实 provider 模式运行
+- 实时热榜数据采集
+- 结构化结果导出
+- Excel 文件生成
+- 验收打分
+
+这说明项目已经不是“理论上可以做到”，而是已经在真实任务链路中证明“能够跑通、能够输出结果、能够形成验收记录”。
+
+对外汇报时，这一点很重要，因为它代表项目已经从“能力设想”走向“能力验证”。
+
+### 12. 工程化基础更好
+
+本项目已经不是单纯的验证页面或原型，而是以运行时内核、协议、规则和测试为主的工程结构。这说明项目更接近“可持续建设的产品内核”，而不只是“一个能演示的自动化效果”。
+
+从目前仓库状态看，已经有 20 多个顶层测试文件，覆盖协议、握手、runtime、配置、兼容层、导出工具、技能执行和验收评分等多个方面。这说明项目已经在往“可持续交付、可持续验证”的方向走，而不是停留在临时性脚本工程。
+
+从长期看，这种工程化能力决定了项目能不能真正进入生产环境，能不能被更多团队协同使用，能不能在后续持续扩展能力。
+
+### 13. 一句话总结现有优势
+
+如果用非技术语言概括，本项目当前最大的优势可以总结为：
+
+不是“更会点网页”，而是“已经具备了企业级智能执行系统该有的安全边界、控制能力、真实交付能力、稳定基础和长期演进空间”。
+
+## 二、下一步计划
+
+下一阶段的重点，不是继续堆脚本，而是进一步解决“页面一变就失效”的老问题，同时把项目能力从“能执行任务”继续提升为“能持续积累企业级自动化资产”。
+
+### 1. 研发“混合自愈选择器”（Hybrid Self-Healing Selector）
+
+在内网环境下，逐步摆脱对单一 XPath 的依赖，建立企业级元素指纹库，让系统在页面变化后依然能更稳地找到目标元素。
+
+### 2. 定义元素指纹数据结构（JSON）
+
+给每个可操作元素建立一份“数字档案”，核心字段包括：
+
+- 语义文本
+- A11y Role
+- 相对空间位置
+- 属性哈希
+- 兜底 XPath
+- 视觉切图（Base64 小图）
+
+这样系统找元素时，不再只靠一条路径，而是像“多特征识别”。
+
+### 3. 推进“影子录制”（Shadow Recording）机制
+
+在现有传统 RPA 正常运行时，于底层开启影子模式。当旧脚本通过 XPath 成功命中元素并完成操作时，后台静默抓取该元素的完整指纹并写入本地数据库。
+
+通过这种方式，在不额外增加大量人工录制成本的前提下，持续沉淀高价值元素资产库。
+
+### 4. 开发穿透层能力
+
+利用定制 Chromium 的底层权限，解决 `iframe` 和闭合 `Shadow DOM` 这类复杂页面结构下的定位难题，为后续自愈选择器提供更强支撑。
+
+## 三、预期结果
+
+通过下一阶段建设，本项目将从“能执行任务”进一步升级为：
+
+- 在复杂企业页面中更稳
+- 对页面变化更不敏感
+- 更容易持续积累高价值资产
+- 更适合在企业环境中长期推广使用
+
+从业务视角看，项目价值也会从“完成单个流程自动化”进一步升级为“建设企业级智能执行底座”。
--- a/docs/sgClaw项目现有优势与下一步计划汇报稿.pdf
+++ b/docs/sgClaw项目现有优势与下一步计划汇报稿.pdf
--- a/resources/rules.json
+++ b/resources/rules.json
@@ -1,6 +1,6 @@
 {
  "version": "1.0",
-  "demo_only_domains": ["baidu.com", "www.baidu.com", "zhihu.com", "www.zhihu.com"],
+  "demo_only_domains": ["baidu.com", "www.baidu.com", "zhihu.com", "www.zhihu.com", "zhuanlan.zhihu.com"],
  "domains": {
    "allowed": [
      "oa.example.com",
@@ -9,7 +9,8 @@
      "baidu.com",
      "www.baidu.com",
      "zhihu.com",
-      "www.zhihu.com"
+      "www.zhihu.com",
+      "zhuanlan.zhihu.com"
    ]
  },
  "pipe_actions": {
--- a/src/agent/mod.rs
+++ b/src/agent/mod.rs
@@ -7,9 +7,7 @@ use std::path::PathBuf;
 use crate::compat::config_adapter::resolve_skills_dir_from_sgclaw_settings;
 use crate::compat::runtime::CompatTaskContext;
 use crate::config::SgClawSettings;
-use crate::pipe::{
-    AgentMessage, BrowserMessage, BrowserPipeTool, PipeError, Transport,
-};
+use crate::pipe::{AgentMessage, BrowserMessage, BrowserPipeTool, PipeError, Transport};

 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct AgentRuntimeContext {
@@ -218,8 +216,7 @@ pub fn handle_browser_message_with_context<T: Transport + 'static>(
                        level: "info".to_string(),
                        message: format!(
                            "runtime profile={:?} skills_prompt_mode={:?}",
-                            settings.runtime_profile,
-                            settings.skills_prompt_mode
+                            settings.runtime_profile, settings.skills_prompt_mode
                        ),
                    });
                    if crate::compat::orchestration::should_use_primary_orchestration(
--- a/src/agent/planner.rs
+++ b/src/agent/planner.rs
@@ -189,7 +189,10 @@ fn plan_zhihu_search(query: &str) -> TaskPlan {

 fn build_zhihu_hotlist_preview(instruction: &str) -> ExecutionPreview {
    let normalized = instruction.to_ascii_lowercase();
-    if normalized.contains("dashboard") || instruction.contains("大屏") || instruction.contains("新标签页") {
+    if normalized.contains("dashboard")
+        || instruction.contains("大屏")
+        || instruction.contains("新标签页")
+    {
        return ExecutionPreview {
            summary: "先规划再执行知乎热榜大屏生成".to_string(),
            steps: vec![
--- a/src/agent/runtime.rs
+++ b/src/agent/runtime.rs
@@ -3,6 +3,10 @@ use serde_json::{json, Map, Value};
 use crate::llm::{ChatMessage, LlmError, LlmProvider, ToolDefinition, ToolFunctionCall};
 use crate::pipe::{Action, AgentMessage, BrowserPipeTool, PipeError, Transport};

+/// Legacy browser-only runtime kept for dev-only validation and narrow regression coverage.
+/// Production browser submit flow uses `compat::runtime` plus `runtime::engine`.
+pub const LEGACY_DEV_ONLY: bool = true;
+
 const BROWSER_ACTION_TOOL_NAME: &str = "browser_action";

 #[derive(Debug, Clone, PartialEq)]
@@ -21,8 +25,7 @@ pub fn execute_task_with_provider<P: LlmProvider, T: Transport>(
    let messages = vec![
        ChatMessage {
            role: "system".to_string(),
-            content: "You are sgClaw. Use browser_action to complete the browser task."
-                .to_string(),
+            content: "You are sgClaw. Use browser_action to complete the browser task.".to_string(),
        },
        ChatMessage {
            role: "user".to_string(),
@@ -35,8 +38,8 @@ pub fn execute_task_with_provider<P: LlmProvider, T: Transport>(
        .map_err(map_llm_error_to_pipe_error)?;

    for call in calls {
-        let browser_call = parse_browser_action_call(call)
-            .map_err(|err| PipeError::Protocol(err.to_string()))?;
+        let browser_call =
+            parse_browser_action_call(call).map_err(|err| PipeError::Protocol(err.to_string()))?;

        transport.send(&AgentMessage::LogEntry {
            level: "info".to_string(),
--- a/src/compat/browser_script_skill_tool.rs
+++ b/src/compat/browser_script_skill_tool.rs
@@ -26,10 +26,15 @@ impl<T: Transport> BrowserScriptSkillTool<T> {
        browser_tool: BrowserPipeTool<T>,
    ) -> anyhow::Result<Self> {
        let script_path = skill_root.join(&tool.command);
-        let canonical_skill_root = skill_root.canonicalize().unwrap_or_else(|_| skill_root.to_path_buf());
-        let canonical_script_path = script_path
+        let canonical_skill_root = skill_root
            .canonicalize()
-            .map_err(|err| anyhow::anyhow!("failed to resolve browser script {}: {err}", script_path.display()))?;
+            .unwrap_or_else(|_| skill_root.to_path_buf());
+        let canonical_script_path = script_path.canonicalize().map_err(|err| {
+            anyhow::anyhow!(
+                "failed to resolve browser script {}: {err}",
+                script_path.display()
+            )
+        })?;
        if !canonical_script_path.starts_with(&canonical_skill_root) {
            anyhow::bail!(
                "browser script path escapes skill root: {}",
@@ -108,7 +113,11 @@ impl<T: Transport + 'static> Tool for BrowserScriptSkillTool<T> {
                    "expected_domain must be a non-empty string, got {other}"
                )))
            }
-            None => return Ok(failed_tool_result("missing required field expected_domain".to_string())),
+            None => {
+                return Ok(failed_tool_result(
+                    "missing required field expected_domain".to_string(),
+                ))
+            }
        };
        let expected_domain = match normalize_domain_like(&raw_expected_domain) {
            Some(value) => value,
@@ -148,7 +157,9 @@ impl<T: Transport + 'static> Tool for BrowserScriptSkillTool<T> {
        };

        if !result.success {
-            return Ok(failed_tool_result(format_browser_script_error(&result.data)));
+            return Ok(failed_tool_result(format_browser_script_error(
+                &result.data,
+            )));
        }

        let payload = result
--- a/src/compat/browser_tool_adapter.rs
+++ b/src/compat/browser_tool_adapter.rs
@@ -101,14 +101,14 @@ impl<T: Transport + 'static> Tool for ZeroClawBrowserTool<T> {
            Err(err) => return Ok(failed_tool_result(err.to_string())),
        };

-        let result = match self.browser_tool.invoke(
-            request.action,
-            request.params,
-            &request.expected_domain,
-        ) {
-            Ok(result) => result,
-            Err(err) => return Ok(failed_tool_result(err.to_string())),
-        };
+        let result =
+            match self
+                .browser_tool
+                .invoke(request.action, request.params, &request.expected_domain)
+            {
+                Ok(result) => result,
+                Err(err) => return Ok(failed_tool_result(err.to_string())),
+            };

        let output = serde_json::to_string(&json!({
            "seq": result.seq,
@@ -122,8 +122,7 @@ impl<T: Transport + 'static> Tool for ZeroClawBrowserTool<T> {
        Ok(ToolResult {
            success: result.success,
            output,
-            error: (!result.success)
-                .then(|| format_browser_action_error(&result.data)),
+            error: (!result.success).then(|| format_browser_action_error(&result.data)),
        })
    }
 }
@@ -134,7 +133,9 @@ struct BrowserActionRequest {
    params: Value,
 }

-fn parse_browser_action_request(args: Value) -> Result<BrowserActionRequest, BrowserActionAdapterError> {
+fn parse_browser_action_request(
+    args: Value,
+) -> Result<BrowserActionRequest, BrowserActionAdapterError> {
    let mut args = match args {
        Value::Object(args) => args,
        other => {
--- a/src/compat/config_adapter.rs
+++ b/src/compat/config_adapter.rs
@@ -2,8 +2,8 @@ use std::collections::HashMap;
 use std::ffi::OsStr;
 use std::path::{Path, PathBuf};

-use zeroclaw::Config as ZeroClawConfig;
 use zeroclaw::config::schema::ModelProviderConfig;
+use zeroclaw::Config as ZeroClawConfig;

 use crate::compat::cron_adapter::configure_embedded_cron;
 use crate::compat::memory_adapter::configure_embedded_memory;
@@ -13,7 +13,9 @@ use crate::runtime::RuntimeProfile;
 const SGCLAW_ZEROCLAW_WORKSPACE_DIR: &str = ".sgclaw-zeroclaw-workspace";
 const SKILLS_DIR_NAME: &str = "skills";

-pub fn build_zeroclaw_config(workspace_root: &Path) -> Result<ZeroClawConfig, crate::config::ConfigError> {
+pub fn build_zeroclaw_config(
+    workspace_root: &Path,
+) -> Result<ZeroClawConfig, crate::config::ConfigError> {
    let settings = SgClawSettings::from_env()?;
    Ok(build_zeroclaw_config_from_sgclaw_settings(
        workspace_root,
--- a/src/compat/cron_adapter.rs
+++ b/src/compat/cron_adapter.rs
@@ -65,7 +65,10 @@ where

    for job in jobs {
        if !matches!(job.job_type, JobType::Agent) {
-            anyhow::bail!("unsupported cron job type in sgclaw compat: {:?}", job.job_type);
+            anyhow::bail!(
+                "unsupported cron job type in sgclaw compat: {:?}",
+                job.job_type
+            );
        }

        let started_at = Utc::now();
--- a/src/compat/event_bridge.rs
+++ b/src/compat/event_bridge.rs
@@ -14,19 +14,17 @@ pub fn log_entry_for_turn_event(
            level: "info".to_string(),
            message: format_tool_call(name, args, skill_versions),
        }),
-        TurnEvent::ToolResult { output, .. } if is_tool_error(output) => Some(AgentMessage::LogEntry {
-            level: "error".to_string(),
-            message: output.trim_start_matches("Error: ").to_string(),
-        }),
+        TurnEvent::ToolResult { output, .. } if is_tool_error(output) => {
+            Some(AgentMessage::LogEntry {
+                level: "error".to_string(),
+                message: output.trim_start_matches("Error: ").to_string(),
+            })
+        }
        _ => None,
    }
 }

-fn format_tool_call(
-    name: &str,
-    args: &Value,
-    skill_versions: &HashMap<String, String>,
-) -> String {
+fn format_tool_call(name: &str, args: &Value, skill_versions: &HashMap<String, String>) -> String {
    if name == "read_skill" {
        let skill_name = args
            .get("name")
@@ -49,7 +47,10 @@ fn format_tool_call(

    match action {
        "navigate" => {
-            let url = args.get("url").and_then(Value::as_str).unwrap_or("<missing-url>");
+            let url = args
+                .get("url")
+                .and_then(Value::as_str)
+                .unwrap_or("<missing-url>");
            format!("navigate {url}")
        }
        "type" => {
--- a/src/compat/openxml_office_tool.rs
+++ b/src/compat/openxml_office_tool.rs
@@ -1,8 +1,8 @@
 use async_trait::async_trait;
 use serde::Deserialize;
 use serde_json::{json, Value};
-use std::collections::BTreeSet;
 use std::collections::BTreeMap;
+use std::collections::BTreeSet;
 use std::fs;
 use std::path::{Path, PathBuf};
 use std::process::Command;
@@ -93,7 +93,11 @@ impl Tool for OpenXmlOfficeTool {
            return Ok(failed_tool_result("rows must not be empty".to_string()));
        }

-        if parsed.rows.iter().any(|row| row.len() != parsed.columns.len()) {
+        if parsed
+            .rows
+            .iter()
+            .any(|row| row.len() != parsed.columns.len())
+        {
            return Ok(failed_tool_result(
                "each row must match the declared columns length".to_string(),
            ));
@@ -153,10 +157,10 @@ fn failed_tool_result(error: String) -> ToolResult {
 }

 fn create_job_root(workspace_root: &Path) -> anyhow::Result<PathBuf> {
-    let nanos = SystemTime::now()
-        .duration_since(UNIX_EPOCH)?
-        .as_nanos();
-    let path = workspace_root.join(".sgclaw-openxml").join(format!("{nanos}"));
+    let nanos = SystemTime::now().duration_since(UNIX_EPOCH)?.as_nanos();
+    let path = workspace_root
+        .join(".sgclaw-openxml")
+        .join(format!("{nanos}"));
    fs::create_dir_all(&path)?;
    Ok(path)
 }
@@ -188,10 +192,7 @@ fn resolve_column_order(
        .iter()
        .map(|value| value.to_string())
        .collect::<BTreeSet<_>>();
-    let expected_set = expected_columns
-        .iter()
-        .cloned()
-        .collect::<BTreeSet<_>>();
+    let expected_set = expected_columns.iter().cloned().collect::<BTreeSet<_>>();

    if provided_set != expected_set {
        return None;
--- a/src/compat/orchestration.rs
+++ b/src/compat/orchestration.rs
@@ -9,6 +9,12 @@ pub fn should_use_primary_orchestration(
    page_url: Option<&str>,
    page_title: Option<&str>,
 ) -> bool {
+    if crate::compat::workflow_executor::detect_route(instruction, page_url, page_title)
+        .is_some_and(|route| crate::compat::workflow_executor::prefers_direct_execution(&route))
+    {
+        return true;
+    }
+
    let normalized = instruction.to_ascii_lowercase();
    let needs_export = normalized.contains("excel")
        || normalized.contains("xlsx")
@@ -33,6 +39,18 @@ pub fn execute_task_with_sgclaw_settings<T: Transport + 'static>(
        task_context.page_url.as_deref(),
        task_context.page_title.as_deref(),
    );
+    if let Some(route) = route.clone() {
+        if crate::compat::workflow_executor::prefers_direct_execution(&route) {
+            return crate::compat::workflow_executor::execute_route(
+                transport,
+                &browser_tool,
+                workspace_root,
+                instruction,
+                task_context,
+                route,
+            );
+        }
+    }
    let primary_result = crate::compat::runtime::execute_task_with_sgclaw_settings(
        transport,
        browser_tool.clone(),
@@ -44,13 +62,16 @@ pub fn execute_task_with_sgclaw_settings<T: Transport + 'static>(

    match (route, primary_result) {
        (Some(route), Ok(summary))
-            if crate::compat::workflow_executor::should_fallback_after_summary(&summary, &route) =>
+            if crate::compat::workflow_executor::should_fallback_after_summary(
+                &summary, &route,
+            ) =>
        {
            crate::compat::workflow_executor::execute_route(
                transport,
                &browser_tool,
                workspace_root,
                instruction,
+                task_context,
                route,
            )
        }
@@ -60,6 +81,7 @@ pub fn execute_task_with_sgclaw_settings<T: Transport + 'static>(
            &browser_tool,
            workspace_root,
            instruction,
+            task_context,
            route,
        ),
        (None, Err(err)) => Err(err),
--- a/src/compat/runtime.rs
+++ b/src/compat/runtime.rs
@@ -5,23 +5,18 @@ use async_trait::async_trait;
 use futures_util::{stream, StreamExt};
 use zeroclaw::agent::TurnEvent;
 use zeroclaw::config::Config as ZeroClawConfig;
-use zeroclaw::providers::{
-    self, ChatMessage, ChatRequest, ChatResponse, Provider,
-};
-use zeroclaw::providers::traits::{
-    ProviderCapabilities, StreamEvent, StreamOptions, StreamResult,
-};
+use zeroclaw::providers::traits::{ProviderCapabilities, StreamEvent, StreamOptions, StreamResult};
+use zeroclaw::providers::{self, ChatMessage, ChatRequest, ChatResponse, Provider};

 use crate::compat::browser_script_skill_tool::build_browser_script_skill_tools;
 use crate::compat::browser_tool_adapter::ZeroClawBrowserTool;
 use crate::compat::config_adapter::{
-    build_zeroclaw_config_from_sgclaw_settings,
-    resolve_skills_dir_from_sgclaw_settings,
+    build_zeroclaw_config_from_sgclaw_settings, resolve_skills_dir_from_sgclaw_settings,
 };
+use crate::compat::event_bridge::log_entry_for_turn_event;
 use crate::compat::openxml_office_tool::OpenXmlOfficeTool;
 use crate::compat::screen_html_export_tool::ScreenHtmlExportTool;
 use crate::config::{DeepSeekSettings, OfficeBackend, SgClawSettings};
-use crate::compat::event_bridge::log_entry_for_turn_event;
 use crate::pipe::{BrowserPipeTool, ConversationMessage, PipeError, Transport};
 use crate::runtime::RuntimeEngine;

@@ -136,13 +131,17 @@ pub async fn execute_task_with_provider<T: Transport + 'static>(
                .map_err(map_anyhow_to_pipe_error)?,
        );
    }
-    if matches!(settings.office_backend, OfficeBackend::OpenXml) &&
-        engine.should_attach_openxml_office_tool(instruction)
+    if matches!(settings.office_backend, OfficeBackend::OpenXml)
+        && engine.should_attach_openxml_office_tool(instruction)
    {
-        tools.push(Box::new(OpenXmlOfficeTool::new(config.workspace_dir.clone())));
+        tools.push(Box::new(OpenXmlOfficeTool::new(
+            config.workspace_dir.clone(),
+        )));
    }
    if engine.should_attach_screen_html_export_tool(instruction) {
-        tools.push(Box::new(ScreenHtmlExportTool::new(config.workspace_dir.clone())));
+        tools.push(Box::new(ScreenHtmlExportTool::new(
+            config.workspace_dir.clone(),
+        )));
    }
    let mut agent = engine.build_agent(
        provider,
@@ -190,10 +189,7 @@ pub async fn execute_task_with_provider<T: Transport + 'static>(

 fn build_provider(config: &ZeroClawConfig) -> Result<Box<dyn Provider>, PipeError> {
    let provider_name = config.default_provider.as_deref().unwrap_or("deepseek");
-    let model_name = config
-        .default_model
-        .as_deref()
-        .unwrap_or("deepseek-chat");
+    let model_name = config.default_model.as_deref().unwrap_or("deepseek-chat");
    let runtime_options = providers::provider_runtime_options_from_config(config);
    let resolved_provider_name = if provider_name == "deepseek" {
        config
@@ -258,7 +254,9 @@ impl Provider for NonStreamingProvider {
        model: &str,
        temperature: f64,
    ) -> anyhow::Result<String> {
-        self.inner.chat_with_history(messages, model, temperature).await
+        self.inner
+            .chat_with_history(messages, model, temperature)
+            .await
    }

    async fn chat(
--- a/src/compat/screen_html_export_tool.rs
+++ b/src/compat/screen_html_export_tool.rs
@@ -238,29 +238,40 @@ fn derive_categories(table: &[ScreenTableRow]) -> Vec<ScreenCategory> {

    grouped
        .into_iter()
-        .map(|((category_code, category_label), (item_count, total_heat))| ScreenCategory {
-            category_code,
-            category_label,
-            item_count,
-            total_heat,
-            avg_heat: if item_count == 0 {
-                0
-            } else {
-                total_heat / item_count
+        .map(
+            |((category_code, category_label), (item_count, total_heat))| ScreenCategory {
+                category_code,
+                category_label,
+                item_count,
+                total_heat,
+                avg_heat: if item_count == 0 {
+                    0
+                } else {
+                    total_heat / item_count
+                },
            },
-        })
+        )
        .collect()
 }

 fn classify_title(title: &str) -> (&'static str, &'static str) {
    let normalized = title.to_ascii_lowercase();
-    if contains_any(&normalized, &["ai", "芯片", "科技", "算法", "机器人", "无人机"]) {
+    if contains_any(
+        &normalized,
+        &["ai", "芯片", "科技", "算法", "机器人", "无人机"],
+    ) {
        return ("technology", "科技");
    }
-    if contains_any(&normalized, &["电影", "综艺", "明星", "周杰伦", "短剧", "娱乐"]) {
+    if contains_any(
+        &normalized,
+        &["电影", "综艺", "明星", "周杰伦", "短剧", "娱乐"],
+    ) {
        return ("entertainment", "娱乐");
    }
-    if contains_any(&normalized, &["足球", "比赛", "联赛", "国足", "体育", "冠军"]) {
+    if contains_any(
+        &normalized,
+        &["足球", "比赛", "联赛", "国足", "体育", "冠军"],
+    ) {
        return ("sports", "体育");
    }
    if contains_any(&normalized, &["航母", "作战", "军", "军事", "演训"]) {
--- a/src/compat/workflow_executor.rs
+++ b/src/compat/workflow_executor.rs
--- a/src/config/mod.rs
+++ b/src/config/mod.rs
@@ -1,12 +1,6 @@
 mod settings;

 pub use settings::{
-    BrowserBackend,
-    ConfigError,
-    DeepSeekSettings,
-    OfficeBackend,
-    PlannerMode,
-    ProviderSettings,
-    SgClawSettings,
-    SkillsPromptMode,
+    BrowserBackend, ConfigError, DeepSeekSettings, OfficeBackend, PlannerMode, ProviderSettings,
+    SgClawSettings, SkillsPromptMode,
 };
--- a/src/config/settings.rs
+++ b/src/config/settings.rs
@@ -114,7 +114,8 @@ impl DeepSeekSettings {
    }

    pub fn load(config_path: Option<&Path>) -> Result<Option<Self>, ConfigError> {
-        SgClawSettings::load(config_path).map(|settings| settings.map(|settings| Self::from(&settings)))
+        SgClawSettings::load(config_path)
+            .map(|settings| settings.map(|settings| Self::from(&settings)))
    }
 }

@@ -216,7 +217,10 @@ impl SgClawSettings {
            .map(parse_runtime_profile)
            .transpose()
            .map_err(|value| {
-                ConfigError::ConfigParse(path.to_path_buf(), format!("invalid runtimeProfile: {value}"))
+                ConfigError::ConfigParse(
+                    path.to_path_buf(),
+                    format!("invalid runtimeProfile: {value}"),
+                )
            })?;
        let skills_prompt_mode = config
            .skills_prompt_mode
@@ -235,7 +239,10 @@ impl SgClawSettings {
            .map(parse_planner_mode)
            .transpose()
            .map_err(|value| {
-                ConfigError::ConfigParse(path.to_path_buf(), format!("invalid plannerMode: {value}"))
+                ConfigError::ConfigParse(
+                    path.to_path_buf(),
+                    format!("invalid plannerMode: {value}"),
+                )
            })?;
        let browser_backend = config
            .browser_backend
@@ -243,7 +250,10 @@ impl SgClawSettings {
            .map(parse_browser_backend)
            .transpose()
            .map_err(|value| {
-                ConfigError::ConfigParse(path.to_path_buf(), format!("invalid browserBackend: {value}"))
+                ConfigError::ConfigParse(
+                    path.to_path_buf(),
+                    format!("invalid browserBackend: {value}"),
+                )
            })?;
        let office_backend = config
            .office_backend
@@ -251,7 +261,10 @@ impl SgClawSettings {
            .map(parse_office_backend)
            .transpose()
            .map_err(|value| {
-                ConfigError::ConfigParse(path.to_path_buf(), format!("invalid officeBackend: {value}"))
+                ConfigError::ConfigParse(
+                    path.to_path_buf(),
+                    format!("invalid officeBackend: {value}"),
+                )
            })?;
        let providers = config
            .providers
@@ -290,12 +303,14 @@ impl SgClawSettings {
        office_backend: Option<OfficeBackend>,
    ) -> Result<Self, ConfigError> {
        let providers = if providers.is_empty() {
-            vec![ProviderSettings::from_legacy_deepseek(api_key, base_url, model)?]
+            vec![ProviderSettings::from_legacy_deepseek(
+                api_key, base_url, model,
+            )?]
        } else {
            providers
        };
-        let active_provider = normalize_optional_value(active_provider)
-            .unwrap_or_else(|| providers[0].id.clone());
+        let active_provider =
+            normalize_optional_value(active_provider).unwrap_or_else(|| providers[0].id.clone());
        let active_provider_settings = providers
            .iter()
            .find(|provider| provider.id == active_provider)
@@ -308,7 +323,10 @@ impl SgClawSettings {

        Ok(Self {
            provider_api_key: active_provider_settings.api_key.clone(),
-            provider_base_url: active_provider_settings.base_url.clone().unwrap_or_default(),
+            provider_base_url: active_provider_settings
+                .base_url
+                .clone()
+                .unwrap_or_default(),
            provider_model: active_provider_settings.model.clone(),
            skills_dir,
            skills_prompt_mode: skills_prompt_mode.unwrap_or(SkillsPromptMode::Compact),
@@ -497,11 +515,7 @@ struct RawProviderSettings {
    api_path: Option<String>,
    #[serde(rename = "wireApi", alias = "wire_api", default)]
    wire_api: Option<String>,
-    #[serde(
-        rename = "requiresOpenaiAuth",
-        alias = "requires_openai_auth",
-        default
-    )]
+    #[serde(rename = "requiresOpenaiAuth", alias = "requires_openai_auth", default)]
    requires_openai_auth: bool,
 }

--- a/src/lib.rs
+++ b/src/lib.rs
@@ -3,6 +3,7 @@ pub mod compat;
 pub mod config;
 pub mod llm;
 pub mod pipe;
+pub mod runtime;
 pub mod security;

 use std::path::PathBuf;
--- a/src/pipe/handshake.rs
+++ b/src/pipe/handshake.rs
@@ -21,9 +21,7 @@ impl HandshakeResult {
            .iter()
            .any(|capability| capability == "browser_action")
            .then(|| {
-                ExecutionSurfaceMetadata::privileged_browser_pipe(
-                    "browser_host_and_mac_policy",
-                )
+                ExecutionSurfaceMetadata::privileged_browser_pipe("browser_host_and_mac_policy")
            })
    }
 }
--- a/src/pipe/mod.rs
+++ b/src/pipe/mod.rs
@@ -5,8 +5,8 @@ pub mod protocol;
 pub use browser_tool::{BrowserPipeTool, CommandOutput};
 pub use handshake::{perform_handshake, HandshakeResult};
 pub use protocol::{
-    supported_actions, Action, AgentMessage, BrowserContext, BrowserMessage,
-    ConversationMessage, ExecutionSurfaceKind, ExecutionSurfaceMetadata, SecurityFields, Timing,
+    supported_actions, Action, AgentMessage, BrowserContext, BrowserMessage, ConversationMessage,
+    ExecutionSurfaceKind, ExecutionSurfaceMetadata, SecurityFields, Timing,
 };

 use std::io::{BufRead, BufReader, Read, Write};
--- a/src/runtime/engine.rs
+++ b/src/runtime/engine.rs
@@ -24,6 +24,7 @@ const BROWSER_TOOL_CONTRACT_PROMPT: &str = "SuperRPA browser interface contract:
 const ZHIHU_HOTLIST_EXECUTION_PROMPT: &str = "Zhihu hotlist execution contract:\n- Treat Zhihu hotlist export/presentation requests as a real browser workflow, not as a text-only summarization task.\n- You must attempt the browser workflow before concluding failure; a prose-only answer is invalid for this workflow.\n- If the current page is not already `https://www.zhihu.com/hot`, navigate there first.\n- If the `zhihu-hotlist.extract_hotlist` skill tool is available, call it before any generic browser probing.\n- Use generic `getText` only as a last-resort fallback when the packaged extractor fails.\n- Extract ordered rows containing `rank`, `title`, and `heat` as structured data.\n- Do not use shell, web_fetch, web_search_tool, or fabricated sample data for this workflow.\n- Do not repeat the same sentence or section in your final answer.";
 const OFFICE_EXPORT_COMPLETION_PROMPT: &str = "Export completion contract:\n- This task requires a real Excel export.\n- After the Zhihu rows are available, you must call openxml_office before finishing.\n- Never fabricate, simulate, or invent substitute hotlist data when a live collection/export task fails.\n- If live collection fails, report the failure concisely instead of producing fake rows.\n- Do not stop after describing how you will parse or export the data.\n- Do not repeat the same sentence or section in your final answer.\n- Your final answer must include the generated local .xlsx path.";
 const SCREEN_EXPORT_COMPLETION_PROMPT: &str = "Presentation completion contract:\n- This task requires a real dashboard artifact.\n- After the Zhihu rows are available, you must call screen_html_export before finishing.\n- Do not stop after describing how you will render or present the data.\n- Do not repeat the same sentence or section in your final answer.\n- Your final answer must include the local .html path and the presentation object.";
+const ZHIHU_WRITE_PUBLISH_PROMPT: &str = "Zhihu article publish contract:\n- This task may publish a Zhihu article.\n- You must not click publish without explicit human confirmation in the current session.\n- If the user asked to publish but no explicit confirmation phrase is present yet, ask for confirmation concisely and stop after the confirmation request.\n- Do not keep exploring tools after you have determined that publish confirmation is missing.\n- If the user only asked to write or draft, stay in draft mode and do not treat it as publish mode.\n- Do not repeat the same sentence or section in your final answer.";

 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct RuntimeEngine {
@@ -51,9 +52,7 @@ impl RuntimeEngine {
        self.tool_policy
            .allowed_tools
            .iter()
-            .any(|tool| {
-                tool == BROWSER_ACTION_TOOL_NAME || tool == SUPERRPA_BROWSER_TOOL_NAME
-            })
+            .any(|tool| tool == BROWSER_ACTION_TOOL_NAME || tool == SUPERRPA_BROWSER_TOOL_NAME)
    }

    pub fn build_agent(
@@ -155,6 +154,9 @@ impl RuntimeEngine {
        if task_needs_screen_export(trimmed_instruction) {
            sections.push(SCREEN_EXPORT_COMPLETION_PROMPT.to_string());
        }
+        if task_requests_zhihu_article_publish(trimmed_instruction, page_url, page_title) {
+            sections.push(ZHIHU_WRITE_PUBLISH_PROMPT.to_string());
+        }
        if let Some(page_context) = build_page_context_message(page_url, page_title) {
            sections.push(page_context);
        }
@@ -173,17 +175,11 @@ impl RuntimeEngine {
                .cmp(&right.name)
                .then(left.version.cmp(&right.version))
        });
-        skills.dedup_by(|left, right| {
-            left.name == right.name && left.version == right.version
-        });
+        skills.dedup_by(|left, right| left.name == right.name && left.version == right.version);
        skills
    }

-    pub fn loaded_skill_names(
-        &self,
-        config: &ZeroClawConfig,
-        skills_dir: &Path,
-    ) -> Vec<String> {
+    pub fn loaded_skill_names(&self, config: &ZeroClawConfig, skills_dir: &Path) -> Vec<String> {
        let mut names = self
            .loaded_skills(config, skills_dir)
            .into_iter()
@@ -237,8 +233,8 @@ impl RuntimeEngine {
        }
        allowed_tools.dedup();

-        if matches!(self.profile, RuntimeProfile::GeneralAssistant) &&
-            self.tool_policy.may_use_non_browser_tools
+        if matches!(self.profile, RuntimeProfile::GeneralAssistant)
+            && self.tool_policy.may_use_non_browser_tools
        {
            None
        } else {
@@ -263,9 +259,7 @@ fn browser_script_tool_names(skills: &[zeroclaw::skills::Skill]) -> Vec<String>

 fn task_needs_local_file_read(instruction: &str) -> bool {
    let normalized = instruction.trim();
-    normalized.contains("/home/") ||
-        normalized.contains("./") ||
-        normalized.contains("../")
+    normalized.contains("/home/") || normalized.contains("./") || normalized.contains("../")
 }

 pub fn is_zhihu_hotlist_task(
@@ -277,16 +271,16 @@ pub fn is_zhihu_hotlist_task(
    let normalized_url = page_url.unwrap_or_default().to_ascii_lowercase();
    let normalized_title = page_title.unwrap_or_default().to_ascii_lowercase();

-    let is_zhihu = normalized_instruction.contains("zhihu") ||
-        instruction.contains("知乎") ||
-        normalized_url.contains("zhihu.com") ||
-        normalized_title.contains("zhihu") ||
-        page_title.unwrap_or_default().contains("知乎");
-    let is_hotlist = normalized_instruction.contains("hotlist") ||
-        instruction.contains("热榜") ||
-        normalized_url.contains("/hot") ||
-        normalized_title.contains("hotlist") ||
-        page_title.unwrap_or_default().contains("热榜");
+    let is_zhihu = normalized_instruction.contains("zhihu")
+        || instruction.contains("知乎")
+        || normalized_url.contains("zhihu.com")
+        || normalized_title.contains("zhihu")
+        || page_title.unwrap_or_default().contains("知乎");
+    let is_hotlist = normalized_instruction.contains("hotlist")
+        || instruction.contains("热榜")
+        || normalized_url.contains("/hot")
+        || normalized_title.contains("hotlist")
+        || page_title.unwrap_or_default().contains("热榜");

    is_zhihu && is_hotlist
 }
@@ -310,6 +304,48 @@ fn task_needs_screen_export(instruction: &str) -> bool {
        || normalized.contains("汇报")
 }

+pub fn task_requests_zhihu_article_publish(
+    instruction: &str,
+    page_url: Option<&str>,
+    page_title: Option<&str>,
+) -> bool {
+    if !is_zhihu_write_task(instruction, page_url, page_title) {
+        return false;
+    }
+
+    let normalized = instruction.to_ascii_lowercase();
+    normalized.contains("publish") || instruction.contains("发布") || instruction.contains("发表")
+}
+
+pub fn is_zhihu_write_task(
+    instruction: &str,
+    page_url: Option<&str>,
+    page_title: Option<&str>,
+) -> bool {
+    let normalized_instruction = instruction.to_ascii_lowercase();
+    let normalized_url = page_url.unwrap_or_default().to_ascii_lowercase();
+    let normalized_title = page_title.unwrap_or_default().to_ascii_lowercase();
+
+    let is_zhihu = normalized_instruction.contains("zhihu")
+        || instruction.contains("知乎")
+        || normalized_url.contains("zhihu.com")
+        || normalized_title.contains("zhihu")
+        || page_title.unwrap_or_default().contains("知乎");
+    let is_write = normalized_instruction.contains("article")
+        || normalized_instruction.contains("write")
+        || normalized_instruction.contains("publish")
+        || instruction.contains("文章")
+        || instruction.contains("写")
+        || instruction.contains("发布")
+        || instruction.contains("发表")
+        || normalized_url.contains("creator")
+        || normalized_url.contains("write")
+        || page_title.unwrap_or_default().contains("创作")
+        || page_title.unwrap_or_default().contains("写文章");
+
+    is_zhihu && is_write
+}
+
 fn load_runtime_skills(config: &ZeroClawConfig, skills_dir: &Path) -> Vec<zeroclaw::skills::Skill> {
    let default_skills_dir = config.workspace_dir.join("skills");
    if skills_dir == default_skills_dir {
@@ -344,10 +380,7 @@ fn build_page_context_message(page_url: Option<&str>, page_title: Option<&str>)
        return None;
    }

-    Some(format!(
-        "Current browser context:\n{}",
-        parts.join("\n")
-    ))
+    Some(format!("Current browser context:\n{}", parts.join("\n")))
 }

 fn map_anyhow_to_pipe_error(err: anyhow::Error) -> PipeError {
--- a/src/runtime/mod.rs
+++ b/src/runtime/mod.rs
@@ -2,6 +2,8 @@ mod engine;
 mod profile;
 mod tool_policy;

-pub use engine::{is_zhihu_hotlist_task, RuntimeEngine};
+pub use engine::{
+    is_zhihu_hotlist_task, is_zhihu_write_task, task_requests_zhihu_article_publish, RuntimeEngine,
+};
 pub use profile::RuntimeProfile;
 pub use tool_policy::ToolPolicy;
--- a/src/runtime/profile.rs
+++ b/src/runtime/profile.rs
@@ -0,0 +1,6 @@
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum RuntimeProfile {
+    BrowserAttached,
+    BrowserHeavy,
+    GeneralAssistant,
+}
--- a/src/runtime/tool_policy.rs
+++ b/src/runtime/tool_policy.rs
@@ -13,18 +13,12 @@ impl ToolPolicy {
            RuntimeProfile::BrowserAttached => Self {
                requires_browser_surface: false,
                may_use_non_browser_tools: true,
-                allowed_tools: vec![
-                    "superrpa_browser".to_string(),
-                    "browser_action".to_string(),
-                ],
+                allowed_tools: vec!["superrpa_browser".to_string(), "browser_action".to_string()],
            },
            RuntimeProfile::BrowserHeavy => Self {
                requires_browser_surface: true,
                may_use_non_browser_tools: true,
-                allowed_tools: vec![
-                    "superrpa_browser".to_string(),
-                    "browser_action".to_string(),
-                ],
+                allowed_tools: vec!["superrpa_browser".to_string(), "browser_action".to_string()],
            },
            RuntimeProfile::GeneralAssistant => Self {
                requires_browser_surface: false,
--- a/tests/agent_runtime_test.rs
+++ b/tests/agent_runtime_test.rs
@@ -4,6 +4,7 @@ use std::sync::Arc;
 use std::time::Duration;

 use common::MockTransport;
+use sgclaw::agent::handle_browser_message;
 use sgclaw::agent::runtime::{browser_action_tool_definition, execute_task_with_provider};
 use sgclaw::llm::{ChatMessage, LlmError, LlmProvider, ToolDefinition, ToolFunctionCall};
 use sgclaw::pipe::{Action, AgentMessage, BrowserMessage, BrowserPipeTool, Timing};
@@ -132,3 +133,46 @@ fn runtime_executes_provider_tool_calls_and_returns_summary() {
            if *seq == 2 && action == &Action::Type
    ));
 }
+
+#[test]
+fn legacy_agent_runtime_is_explicitly_dev_only() {
+    assert!(sgclaw::agent::runtime::LEGACY_DEV_ONLY);
+}
+
+#[test]
+fn production_submit_task_does_not_route_into_legacy_runtime_without_llm_config() {
+    std::env::remove_var("DEEPSEEK_API_KEY");
+    std::env::remove_var("DEEPSEEK_BASE_URL");
+    std::env::remove_var("DEEPSEEK_MODEL");
+
+    let transport = Arc::new(MockTransport::new(vec![]));
+    let browser_tool = BrowserPipeTool::new(
+        transport.clone(),
+        test_policy(),
+        vec![1, 2, 3, 4, 5, 6, 7, 8],
+    )
+    .with_response_timeout(Duration::from_secs(1));
+
+    handle_browser_message(
+        transport.as_ref(),
+        &browser_tool,
+        BrowserMessage::SubmitTask {
+            instruction: "打开百度".to_string(),
+            conversation_id: String::new(),
+            messages: vec![],
+            page_url: String::new(),
+            page_title: String::new(),
+        },
+    )
+    .unwrap();
+
+    let sent = transport.sent_messages();
+    assert!(matches!(
+        sent.last(),
+        Some(AgentMessage::TaskComplete { success, summary })
+            if !success && summary.contains("未配置大语言模型")
+    ));
+    assert!(!sent
+        .iter()
+        .any(|message| { matches!(message, AgentMessage::Command { .. }) }));
+}
--- a/tests/browser_script_skill_tool_test.rs
+++ b/tests/browser_script_skill_tool_test.rs
@@ -1,11 +1,11 @@
 mod common;

 use std::collections::HashMap;
+use std::fs;
 use std::path::PathBuf;
 use std::sync::Arc;
 use std::time::Duration;
 use std::time::{SystemTime, UNIX_EPOCH};
-use std::fs;

 use common::MockTransport;
 use serde_json::json;
@@ -77,13 +77,8 @@ return {
        command: "scripts/extract_hotlist.js".to_string(),
        args,
    };
-    let tool = BrowserScriptSkillTool::new(
-        "zhihu-hotlist",
-        &skill_tool,
-        &skill_dir,
-        browser_tool,
-    )
-    .unwrap();
+    let tool = BrowserScriptSkillTool::new("zhihu-hotlist", &skill_tool, &skill_dir, browser_tool)
+        .unwrap();

    let result = tool
        .execute(json!({
--- a/tests/browser_tool_test.rs
+++ b/tests/browser_tool_test.rs
@@ -96,8 +96,14 @@ fn browser_tool_exposes_privileged_surface_metadata_backed_by_mac_policy() {
    assert!(metadata.privileged);
    assert!(!metadata.defines_runtime_identity);
    assert_eq!(metadata.guard, "mac_policy");
-    assert_eq!(metadata.allowed_domains, vec!["oa.example.com", "erp.example.com"]);
-    assert_eq!(metadata.allowed_actions, vec!["click", "type", "navigate", "getText"]);
+    assert_eq!(
+        metadata.allowed_domains,
+        vec!["oa.example.com", "erp.example.com"]
+    );
+    assert_eq!(
+        metadata.allowed_actions,
+        vec!["click", "type", "navigate", "getText"]
+    );
 }

 #[test]
--- a/tests/compat_browser_tool_test.rs
+++ b/tests/compat_browser_tool_test.rs
@@ -26,7 +26,9 @@ fn test_policy() -> MacPolicy {
    .unwrap()
 }

-fn build_adapter(messages: Vec<BrowserMessage>) -> (Arc<MockTransport>, ZeroClawBrowserTool<MockTransport>) {
+fn build_adapter(
+    messages: Vec<BrowserMessage>,
+) -> (Arc<MockTransport>, ZeroClawBrowserTool<MockTransport>) {
    let transport = Arc::new(MockTransport::new(messages));
    let browser_tool = BrowserPipeTool::new(
        transport.clone(),
@@ -204,13 +206,11 @@ async fn zeroclaw_browser_tool_keeps_domain_validation_in_mac_policy() {
    assert!(!result.success);
    assert!(result.output.is_empty());
    assert_eq!(transport.sent_messages().len(), 0);
-    assert!(
-        result
-            .error
-            .as_deref()
-            .unwrap()
-            .contains("domain is not allowed")
-    );
+    assert!(result
+        .error
+        .as_deref()
+        .unwrap()
+        .contains("domain is not allowed"));
 }

 #[tokio::test]
@@ -300,25 +300,19 @@ async fn zeroclaw_browser_tool_rejects_missing_required_action_parameters() {
    assert!(!missing_text_selector.success);
    assert!(!missing_navigate_url.success);
    assert_eq!(transport.sent_messages().len(), 0);
-    assert!(
-        missing_click_selector
-            .error
-            .as_deref()
-            .unwrap()
-            .contains("click requires selector")
-    );
-    assert!(
-        missing_text_selector
-            .error
-            .as_deref()
-            .unwrap()
-            .contains("getText requires selector")
-    );
-    assert!(
-        missing_navigate_url
-            .error
-            .as_deref()
-            .unwrap()
-            .contains("navigate requires url")
-    );
+    assert!(missing_click_selector
+        .error
+        .as_deref()
+        .unwrap()
+        .contains("click requires selector"));
+    assert!(missing_text_selector
+        .error
+        .as_deref()
+        .unwrap()
+        .contains("getText requires selector"));
+    assert!(missing_navigate_url
+        .error
+        .as_deref()
+        .unwrap()
+        .contains("navigate requires url"));
 }
--- a/tests/compat_config_test.rs
+++ b/tests/compat_config_test.rs
@@ -3,20 +3,12 @@ use std::path::Path;
 use std::sync::{Mutex, OnceLock};

 use sgclaw::compat::config_adapter::{
-    build_zeroclaw_config,
-    build_zeroclaw_config_from_sgclaw_settings,
-    build_zeroclaw_config_from_settings,
-    resolve_skills_dir,
-    zeroclaw_default_skills_dir,
+    build_zeroclaw_config, build_zeroclaw_config_from_settings,
+    build_zeroclaw_config_from_sgclaw_settings, resolve_skills_dir, zeroclaw_default_skills_dir,
    zeroclaw_workspace_dir,
 };
 use sgclaw::config::{
-    BrowserBackend,
-    DeepSeekSettings,
-    OfficeBackend,
-    PlannerMode,
-    SgClawSettings,
-    SkillsPromptMode,
+    BrowserBackend, DeepSeekSettings, OfficeBackend, PlannerMode, SgClawSettings, SkillsPromptMode,
 };
 use sgclaw::runtime::RuntimeProfile;
 use uuid::Uuid;
@@ -61,11 +53,17 @@ fn zeroclaw_config_adapter_uses_deterministic_workspace_dir() {
    let workspace_dir = zeroclaw_workspace_dir(Path::new("/var/lib/sgclaw"));
    let config = build_zeroclaw_config_from_settings(Path::new("/var/lib/sgclaw"), &settings);

-    assert_eq!(workspace_dir, Path::new("/var/lib/sgclaw/.sgclaw-zeroclaw-workspace"));
+    assert_eq!(
+        workspace_dir,
+        Path::new("/var/lib/sgclaw/.sgclaw-zeroclaw-workspace")
+    );
    assert_eq!(config.workspace_dir, workspace_dir);
    assert_eq!(config.default_provider.as_deref(), Some("deepseek"));
    assert_eq!(config.default_model.as_deref(), Some("deepseek-reasoner"));
-    assert_eq!(config.api_url.as_deref(), Some("https://proxy.example.com/v1"));
+    assert_eq!(
+        config.api_url.as_deref(),
+        Some("https://proxy.example.com/v1")
+    );
    assert_eq!(
        resolve_skills_dir(Path::new("/var/lib/sgclaw"), &settings),
        zeroclaw_default_skills_dir(Path::new("/var/lib/sgclaw"))
@@ -252,7 +250,10 @@ fn sgclaw_settings_load_provider_switching_and_backend_policy_from_browser_confi
    assert_eq!(settings.planner_mode, PlannerMode::ZeroclawPlanFirst);
    assert_eq!(settings.active_provider, "glm-prod");
    assert_eq!(settings.providers.len(), 2);
-    assert_eq!(settings.provider_base_url, "https://open.bigmodel.cn/api/paas/v4");
+    assert_eq!(
+        settings.provider_base_url,
+        "https://open.bigmodel.cn/api/paas/v4"
+    );
    assert_eq!(settings.provider_model, "glm-4.5");
    assert_eq!(settings.browser_backend, BrowserBackend::SuperRpa);
    assert_eq!(settings.office_backend, OfficeBackend::OpenXml);
--- a/tests/compat_cron_test.rs
+++ b/tests/compat_cron_test.rs
@@ -17,6 +17,7 @@ async fn compat_cron_adapter_creates_lists_and_runs_due_agent_jobs() {
        api_key: "key".to_string(),
        base_url: "https://api.deepseek.com".to_string(),
        model: "deepseek-chat".to_string(),
+        skills_dir: None,
    };
    let workspace_root = workspace_root("sgclaw-cron");
    let config = build_zeroclaw_config_from_settings(Path::new(&workspace_root), &settings);
--- a/tests/compat_memory_test.rs
+++ b/tests/compat_memory_test.rs
@@ -16,6 +16,7 @@ async fn compat_memory_adapter_uses_workspace_local_sqlite_backend() {
        api_key: "key".to_string(),
        base_url: "https://api.deepseek.com".to_string(),
        model: "deepseek-chat".to_string(),
+        skills_dir: None,
    };
    let workspace_root = workspace_root("sgclaw-memory");
    let config = build_zeroclaw_config_from_settings(Path::new(&workspace_root), &settings);
--- a/tests/compat_runtime_test.rs
+++ b/tests/compat_runtime_test.rs
--- a/tests/deepseek_provider_test.rs
+++ b/tests/deepseek_provider_test.rs
@@ -21,6 +21,7 @@ fn deepseek_settings_load_defaults_from_env() {
    assert_eq!(settings.api_key, "test-key");
    assert_eq!(settings.base_url, "https://api.deepseek.com");
    assert_eq!(settings.model, "deepseek-chat");
+    assert_eq!(settings.skills_dir, None);
 }

 #[test]
@@ -29,6 +30,7 @@ fn deepseek_request_shape_matches_openai_compatible_chat_format() {
        api_key: "test-key".to_string(),
        base_url: "https://api.deepseek.com".to_string(),
        model: "deepseek-chat".to_string(),
+        skills_dir: None,
    });
    let messages = vec![
        ChatMessage {
@@ -60,8 +62,5 @@ fn deepseek_request_shape_matches_openai_compatible_chat_format() {
    assert_eq!(serialized["messages"][0]["role"], "system");
    assert_eq!(serialized["messages"][1]["content"], "打开百度搜索天气");
    assert_eq!(serialized["tools"][0]["type"], "function");
-    assert_eq!(
-        serialized["tools"][0]["function"]["name"],
-        "browser_action"
-    );
+    assert_eq!(serialized["tools"][0]["function"]["name"], "browser_action");
 }
--- a/tests/planner_test.rs
+++ b/tests/planner_test.rs
@@ -109,7 +109,10 @@ fn plan_first_mode_builds_visible_preview_for_zhihu_excel_flow() {
        .steps
        .iter()
        .any(|step| step.contains("navigate https://www.zhihu.com/hot")));
-    assert!(preview.steps.iter().any(|step| step.contains("getText main")));
+    assert!(preview
+        .steps
+        .iter()
+        .any(|step| step.contains("getText main")));
    assert!(preview
        .steps
        .iter()
--- a/tests/read_skill_tool_test.rs
+++ b/tests/read_skill_tool_test.rs
@@ -30,13 +30,22 @@ async fn read_skill_inlines_referenced_markdown_files() {
    .unwrap();

    let tool = ReadSkillTool::new(workspace_dir, false, None);
-    let result = tool.execute(json!({ "name": "zhihu-hotlist" })).await.unwrap();
+    let result = tool
+        .execute(json!({ "name": "zhihu-hotlist" }))
+        .await
+        .unwrap();

    assert!(result.success);
    assert!(result.output.contains("# Zhihu Hotlist"));
-    assert!(result.output.contains("## Referenced File: references/collection-flow.md"));
-    assert!(result.output.contains("Collect rows from the hotlist first."));
-    assert!(result.output.contains("## Referenced File: references/data-quality.md"));
+    assert!(result
+        .output
+        .contains("## Referenced File: references/collection-flow.md"));
+    assert!(result
+        .output
+        .contains("Collect rows from the hotlist first."));
+    assert!(result
+        .output
+        .contains("## Referenced File: references/data-quality.md"));
    assert!(result.output.contains("Mark partial metrics explicitly."));
 }

@@ -65,12 +74,21 @@ async fn read_skill_recursively_inlines_relative_asset_references() {
    .unwrap();

    let tool = ReadSkillTool::new(workspace_dir, false, None);
-    let result = tool.execute(json!({ "name": "zhihu-hotlist" })).await.unwrap();
+    let result = tool
+        .execute(json!({ "name": "zhihu-hotlist" }))
+        .await
+        .unwrap();

    assert!(result.success);
-    assert!(result.output.contains("## Referenced File: references/collection-flow.md"));
-    assert!(result.output.contains("## Referenced File: assets/zhihu_hotlist_flow.source.json"));
-    assert!(result.output.contains("\"selectors\": [\".HotList-list\", \".HotItem\"]"));
+    assert!(result
+        .output
+        .contains("## Referenced File: references/collection-flow.md"));
+    assert!(result
+        .output
+        .contains("## Referenced File: assets/zhihu_hotlist_flow.source.json"));
+    assert!(result
+        .output
+        .contains("\"selectors\": [\".HotList-list\", \".HotItem\"]"));
 }

 fn temp_workspace_dir() -> PathBuf {
--- a/tests/runtime_profile_test.rs
+++ b/tests/runtime_profile_test.rs
@@ -1,5 +1,5 @@
-use sgclaw::runtime::{RuntimeEngine, RuntimeProfile, ToolPolicy};
 use sgclaw::config::{BrowserBackend, OfficeBackend, PlannerMode, SgClawSettings};
+use sgclaw::runtime::{RuntimeEngine, RuntimeProfile, ToolPolicy};

 #[test]
 fn browser_attached_profile_exposes_browser_surface_without_becoming_browser_only() {
@@ -39,6 +39,23 @@ fn browser_attached_export_prompt_requires_openxml_completion() {
    assert!(instruction.contains("final answer must include the generated local .xlsx path"));
 }

+#[test]
+fn browser_attached_publish_prompt_requires_explicit_confirmation_before_clicking_publish() {
+    let engine = RuntimeEngine::new(RuntimeProfile::BrowserAttached);
+
+    let instruction = engine.build_instruction(
+        "请直接发表这篇知乎文章，标题是测试标题，正文是第一段内容",
+        Some("https://www.zhihu.com/creator"),
+        Some("知乎创作中心"),
+        true,
+    );
+
+    assert!(instruction.contains("publish a Zhihu article"));
+    assert!(instruction.contains("must not click publish without explicit human confirmation"));
+    assert!(instruction.contains("ask for confirmation concisely"));
+    assert!(instruction.contains("stop after the confirmation request"));
+}
+
 #[test]
 fn legacy_settings_default_to_plan_first_superrpa_and_openxml_backends() {
    let settings = SgClawSettings::from_legacy_deepseek_fields(
--- a/tests/skill_lib_validation_test.py
+++ b/tests/skill_lib_validation_test.py
@@ -55,7 +55,7 @@ class SkillLibValidationTest(unittest.TestCase):
                self.assertIn("xlsx", record.tags)
            expected_location = (
                SKILLS_DIR / name / "SKILL.toml"
-                if name == "zhihu-hotlist"
+                if name in {"zhihu-hotlist", "zhihu-navigate", "zhihu-write"}
                else SKILLS_DIR / name / "SKILL.md"
            )
            self.assertEqual(record.location, expected_location)
@@ -83,6 +83,17 @@ class SkillLibValidationTest(unittest.TestCase):
        self.assertTrue(
            (SKILLS_DIR / "zhihu-hotlist" / "scripts" / "extract_hotlist.js").is_file()
        )
+        self.assertTrue((SKILLS_DIR / "zhihu-navigate" / "SKILL.toml").is_file())
+        self.assertTrue(
+            (SKILLS_DIR / "zhihu-navigate" / "scripts" / "open_creator_entry.js").is_file()
+        )
+        self.assertTrue((SKILLS_DIR / "zhihu-write" / "SKILL.toml").is_file())
+        self.assertTrue(
+            (SKILLS_DIR / "zhihu-write" / "scripts" / "prepare_article_editor.js").is_file()
+        )
+        self.assertTrue(
+            (SKILLS_DIR / "zhihu-write" / "scripts" / "fill_article_draft.js").is_file()
+        )

    def test_each_skill_declares_superrpa_browser_contract(self):
        for name in [name for name in EXPECTED_SKILL_NAMES if name.startswith("zhihu-")]:
--- a/tests/skill_script_zhihu_navigate_test.py
+++ b/tests/skill_script_zhihu_navigate_test.py
@@ -0,0 +1,243 @@
+import json
+import subprocess
+import textwrap
+import unittest
+from pathlib import Path
+
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+SCRIPT_PATH = (
+    REPO_ROOT.parent / "skill_lib" / "skills" / "zhihu-navigate" / "scripts" /
+    "open_creator_entry.js"
+)
+
+
+def run_open_creator_entry(*, body_text: str, selectors: dict[str, list[dict]]) -> dict:
+    node_script = textwrap.dedent(
+        f"""
+        import fs from 'node:fs';
+        import vm from 'node:vm';
+
+        const scriptPath = {json.dumps(str(SCRIPT_PATH))};
+        const selectorMap = {json.dumps(selectors, ensure_ascii=False)};
+        const bodyText = {json.dumps(body_text, ensure_ascii=False)};
+        const source = fs.readFileSync(scriptPath, 'utf8');
+
+        function createNode(spec) {{
+          const node = {{
+            id: String(spec?.id ?? ''),
+            tagName: String(spec?.tagName || 'DIV').toUpperCase(),
+            textContent: String(spec?.textContent ?? ''),
+            innerText: String(spec?.innerText ?? spec?.textContent ?? ''),
+            href: String(spec?.href ?? ''),
+            className: String(spec?.className ?? ''),
+            role: String(spec?.role ?? ''),
+            tabIndex: Number.isFinite(spec?.tabIndex) ? Number(spec.tabIndex) : -1,
+            clicked: false,
+            click() {{
+              this.clicked = true;
+            }},
+            getAttribute(name) {{
+              if (name === 'class') {{
+                return this.className;
+              }}
+              if (name === 'role') {{
+                return this.role;
+              }}
+              if (name === 'href') {{
+                return this.href;
+              }}
+              if (name === 'tabindex' && this.tabIndex >= 0) {{
+                return String(this.tabIndex);
+              }}
+              return '';
+            }},
+            getBoundingClientRect() {{
+              return {{
+                width: spec?.visible === false ? 0 : 120,
+                height: spec?.visible === false ? 0 : 32,
+              }};
+            }},
+            parentElement: null,
+            closest(selector) {{
+              let current = this;
+              while (current) {{
+                if (matchesSelector(current, selector)) {{
+                  return current;
+                }}
+                current = current.parentElement;
+              }}
+              return null;
+            }},
+          }};
+          return node;
+        }}
+
+        const created = new Map();
+        const nodesById = new Map();
+        function matchesSelector(node, selector) {{
+          const parts = selector.split(',').map((part) => part.trim()).filter(Boolean);
+          return parts.some((part) => {{
+            if (part === 'a[href]') {{
+              return node.tagName === 'A' && Boolean(node.href);
+            }}
+            if (part === 'button') {{
+              return node.tagName === 'BUTTON';
+            }}
+            if (part === '[role=\"button\"]' || part === "[role='button']") {{
+              return node.role === 'button';
+            }}
+            if (part === '[tabindex]') {{
+              return node.tabIndex >= 0;
+            }}
+            if (part === 'div') {{
+              return node.tagName === 'DIV';
+            }}
+            if (part === 'span') {{
+              return node.tagName === 'SPAN';
+            }}
+            return false;
+          }});
+        }}
+        function createNodeList(selector) {{
+          const specs = selectorMap[selector] || [];
+          return specs.map((spec, index) => {{
+            const key = `${{selector}}#${{index}}`;
+            if (!created.has(key)) {{
+              const node = createNode(spec);
+              created.set(key, node);
+              if (node.id) {{
+                nodesById.set(node.id, node);
+              }}
+            }}
+            return created.get(key);
+          }});
+        }}
+
+        for (const selector of Object.keys(selectorMap)) {{
+          createNodeList(selector);
+        }}
+        for (const node of created.values()) {{
+          const parentId = selectorMap && Object.values(selectorMap)
+            .flat()
+            .find((spec) => String(spec?.id ?? '') === node.id)?.parentId;
+          if (parentId && nodesById.has(parentId)) {{
+            node.parentElement = nodesById.get(parentId);
+          }}
+        }}
+
+        const bodyNode = createNode({{ tagName: 'BODY', textContent: bodyText, innerText: bodyText }});
+        const context = {{
+          args: {{ desired_target: 'article_editor' }},
+          location: {{ href: 'https://www.zhihu.com/creator' }},
+          document: {{
+            body: bodyNode,
+            querySelector(selector) {{
+              if (selector === 'body') {{
+                return bodyNode;
+              }}
+              return createNodeList(selector)[0] || null;
+            }},
+            querySelectorAll(selector) {{
+              return createNodeList(selector);
+            }},
+          }},
+          console,
+          JSON,
+          Math,
+          Number,
+          Object,
+          RegExp,
+          Set,
+          String,
+          Array,
+          Error,
+        }};
+
+        try {{
+          const result = vm.runInNewContext(`(function(){{\\n${{source}}\\n}})()`, context);
+          process.stdout.write(JSON.stringify({{ ok: true, result, created: Object.fromEntries(created) }}));
+        }} catch (error) {{
+          process.stdout.write(JSON.stringify({{
+            ok: false,
+            error: String(error && error.message ? error.message : error),
+          }}));
+          process.exitCode = 1;
+        }}
+        """
+    )
+    completed = subprocess.run(
+        ["node", "--input-type=module", "-e", node_script],
+        check=False,
+        capture_output=True,
+        text=True,
+    )
+    payload = json.loads(completed.stdout)
+    if completed.returncode != 0:
+        raise AssertionError(payload["error"])
+    return payload
+
+
+class SkillScriptZhihuNavigateTest(unittest.TestCase):
+    def test_open_creator_entry_clicks_anchor_write_entry(self):
+        payload = run_open_creator_entry(
+            body_text="创作者中心 写文章",
+            selectors={
+                "a[href], button, [role='button']": [
+                  {
+                        "tagName": "a",
+                        "textContent": "写文章",
+                        "href": "https://zhuanlan.zhihu.com/write",
+                    }
+                ]
+            },
+        )
+
+        self.assertEqual(payload["result"]["status"], "creator_entry_clicked")
+        self.assertTrue(payload["created"]["a[href], button, [role='button']#0"]["clicked"])
+
+    def test_open_creator_entry_clicks_button_write_entry(self):
+        payload = run_open_creator_entry(
+            body_text="创作者中心 发布内容",
+            selectors={
+                "a[href], button, [role='button']": [
+                    {
+                        "tagName": "button",
+                        "textContent": "写文章",
+                    }
+                ]
+            },
+        )
+
+        self.assertEqual(payload["result"]["status"], "creator_entry_clicked")
+        self.assertTrue(payload["created"]["a[href], button, [role='button']#0"]["clicked"])
+
+    def test_open_creator_entry_clicks_clickable_ancestor_for_nested_write_text(self):
+        payload = run_open_creator_entry(
+            body_text="创作者中心 写文章",
+            selectors={
+                "a[href], button, [role='button']": [],
+                "div, span, [tabindex]": [
+                    {
+                        "id": "ancestor",
+                        "tagName": "div",
+                        "className": "creator-entry",
+                        "tabIndex": 0,
+                        "textContent": "",
+                    },
+                    {
+                        "id": "label",
+                        "parentId": "ancestor",
+                        "tagName": "span",
+                        "textContent": "写文章",
+                    },
+                ],
+            },
+        )
+
+        self.assertEqual(payload["result"]["status"], "creator_entry_clicked")
+        self.assertTrue(payload["created"]["div, span, [tabindex]#0"]["clicked"])
+
+
+if __name__ == "__main__":
+    unittest.main()
--- a/tests/skill_script_zhihu_write_test.py
+++ b/tests/skill_script_zhihu_write_test.py
@@ -0,0 +1,219 @@
+import json
+import subprocess
+import textwrap
+import unittest
+from pathlib import Path
+
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+PREPARE_SCRIPT_PATH = (
+    REPO_ROOT.parent / "skill_lib" / "skills" / "zhihu-write" / "scripts" /
+    "prepare_article_editor.js"
+)
+FILL_SCRIPT_PATH = (
+    REPO_ROOT.parent / "skill_lib" / "skills" / "zhihu-write" / "scripts" /
+    "fill_article_draft.js"
+)
+
+
+def run_browser_script(script_path: Path, *, args: dict, body_text: str, selectors: dict[str, list[dict]]) -> dict:
+    node_script = textwrap.dedent(
+        f"""
+        import fs from 'node:fs';
+        import vm from 'node:vm';
+
+        const scriptPath = {json.dumps(str(script_path))};
+        const args = {json.dumps(args, ensure_ascii=False)};
+        const selectorMap = {json.dumps(selectors, ensure_ascii=False)};
+        const bodyText = {json.dumps(body_text, ensure_ascii=False)};
+        const source = fs.readFileSync(scriptPath, 'utf8');
+
+        function createNode(spec) {{
+          const attrs = spec?.attrs || {{}};
+          const node = {{
+            tagName: String(spec?.tagName || 'DIV').toUpperCase(),
+            textContent: String(spec?.textContent ?? ''),
+            innerText: String(spec?.innerText ?? spec?.textContent ?? ''),
+            innerHTML: String(spec?.innerHTML ?? spec?.textContent ?? ''),
+            value: String(spec?.value ?? ''),
+            children: [],
+            focused: false,
+            clicked: false,
+            appendChild(child) {{
+              this.children.push(child);
+              return child;
+            }},
+            focus() {{
+              this.focused = true;
+            }},
+            click() {{
+              this.clicked = true;
+            }},
+            dispatchEvent() {{
+              return true;
+            }},
+            getAttribute(name) {{
+              return Object.prototype.hasOwnProperty.call(attrs, name) ? attrs[name] : null;
+            }},
+            querySelector() {{
+              return null;
+            }},
+            querySelectorAll() {{
+              return [];
+            }},
+            getBoundingClientRect() {{
+              return {{
+                width: spec?.visible === false ? 0 : 100,
+                height: spec?.visible === false ? 0 : 20,
+              }};
+            }},
+          }};
+          return node;
+        }}
+
+        const created = new Map();
+
+        function createNodeList(selector) {{
+          const specs = selectorMap[selector] || [];
+          return specs.map((spec, index) => {{
+            const key = `${{selector}}#${{index}}`;
+            if (!created.has(key)) {{
+              created.set(key, createNode(spec));
+            }}
+            return created.get(key);
+          }});
+        }}
+
+        const bodyNode = createNode({{ tagName: 'body', textContent: bodyText, innerText: bodyText }});
+        const context = {{
+          args,
+          location: {{ href: 'https://zhuanlan.zhihu.com/write' }},
+          document: {{
+            body: bodyNode,
+            createElement(tagName) {{
+              return createNode({{ tagName }});
+            }},
+            createTextNode(text) {{
+              return createNode({{ tagName: '#text', textContent: text, innerText: text }});
+            }},
+            querySelector(selector) {{
+              if (selector === 'body') {{
+                return bodyNode;
+              }}
+              return createNodeList(selector)[0] || null;
+            }},
+            querySelectorAll(selector) {{
+              return createNodeList(selector);
+            }},
+          }},
+          Event: class Event {{
+            constructor(type, init = {{}}) {{
+              this.type = type;
+              this.bubbles = !!init.bubbles;
+              this.composed = !!init.composed;
+            }}
+          }},
+          console,
+          JSON,
+          Math,
+          Number,
+          Object,
+          RegExp,
+          Set,
+          String,
+          Array,
+          Error,
+        }};
+
+        try {{
+          const result = vm.runInNewContext(`(function(){{\\n${{source}}\\n}})()`, context);
+          process.stdout.write(JSON.stringify({{ ok: true, result, created: Object.fromEntries(created) }}));
+        }} catch (error) {{
+          process.stdout.write(JSON.stringify({{
+            ok: false,
+            error: String(error && error.message ? error.message : error),
+          }}));
+          process.exitCode = 1;
+        }}
+        """
+    )
+    completed = subprocess.run(
+        ["node", "--input-type=module", "-e", node_script],
+        check=False,
+        capture_output=True,
+        text=True,
+    )
+    payload = json.loads(completed.stdout)
+    if completed.returncode != 0:
+        raise AssertionError(payload["error"])
+    return payload
+
+
+class SkillScriptZhihuWriteTest(unittest.TestCase):
+    def test_prepare_article_editor_accepts_role_textbox_title_and_generic_body_editor(self):
+        payload = run_browser_script(
+            PREPARE_SCRIPT_PATH,
+            args={"desired_mode": "draft"},
+            body_text="写文章 发布",
+            selectors={
+                "[role='textbox'][aria-label*='标题']": [
+                    {
+                        "tagName": "div",
+                        "attrs": {
+                            "role": "textbox",
+                            "aria-label": "标题",
+                            "contenteditable": "true",
+                        },
+                    }
+                ],
+                "div[contenteditable='true']": [
+                    {
+                        "tagName": "div",
+                        "attrs": {
+                            "contenteditable": "true",
+                            "data-placeholder": "在这里输入正文",
+                        },
+                    }
+                ],
+            },
+        )
+
+        self.assertEqual(payload["result"]["status"], "editor_ready")
+
+    def test_fill_article_draft_accepts_role_textbox_title_and_generic_body_editor(self):
+        payload = run_browser_script(
+            FILL_SCRIPT_PATH,
+            args={
+                "title": "测试标题",
+                "body": "第一段\n第二段",
+                "publish_mode": "false",
+            },
+            body_text="写文章 发布",
+            selectors={
+                "[role='textbox'][aria-label*='标题']": [
+                    {
+                        "tagName": "div",
+                        "attrs": {
+                            "role": "textbox",
+                            "aria-label": "标题",
+                            "contenteditable": "true",
+                        },
+                    }
+                ],
+                "div[contenteditable='true']": [
+                    {
+                        "tagName": "div",
+                        "attrs": {
+                            "contenteditable": "true",
+                            "data-placeholder": "在这里输入正文",
+                        },
+                    }
+                ],
+            },
+        )
+
+        self.assertEqual(payload["result"]["status"], "draft_ready")
+
+
+if __name__ == "__main__":
+    unittest.main()
--- a/third_party/zeroclaw/src/agent/agent.rs
+++ b/third_party/zeroclaw/src/agent/agent.rs
@@ -796,7 +796,8 @@ impl Agent {

            let (text, calls) = self.tool_dispatcher.parse_response(&response);
            let calls = canonicalize_parsed_tool_calls(&self.tools, calls);
-            response.tool_calls = canonicalize_provider_tool_calls(&self.tools, response.tool_calls);
+            response.tool_calls =
+                canonicalize_provider_tool_calls(&self.tools, response.tool_calls);
            if calls.is_empty() {
                let final_text = if text.is_empty() {
                    response.text.unwrap_or_default()
@@ -1065,7 +1066,8 @@ impl Agent {

            let (text, calls) = self.tool_dispatcher.parse_response(&response);
            let calls = canonicalize_parsed_tool_calls(&self.tools, calls);
-            response.tool_calls = canonicalize_provider_tool_calls(&self.tools, response.tool_calls);
+            response.tool_calls =
+                canonicalize_provider_tool_calls(&self.tools, response.tool_calls);
            if calls.is_empty() {
                let final_text = if text.is_empty() {
                    response.text.unwrap_or_default()
@@ -1207,7 +1209,8 @@ fn sanitize_final_text(text: &str) -> String {
 }

 fn resolve_registered_tool_name(tools: &[Box<dyn Tool>], raw: &str) -> Option<String> {
-    tools.iter()
+    tools
+        .iter()
        .find(|tool| {
            tool.name() == raw || crate::tools::provider_safe_tool_name(tool.name()) == raw
        })
@@ -1218,7 +1221,8 @@ fn canonicalize_parsed_tool_calls(
    tools: &[Box<dyn Tool>],
    calls: Vec<ParsedToolCall>,
 ) -> Vec<ParsedToolCall> {
-    calls.into_iter()
+    calls
+        .into_iter()
        .map(|mut call| {
            if let Some(canonical_name) = resolve_registered_tool_name(tools, &call.name) {
                call.name = canonical_name;
@@ -1232,7 +1236,8 @@ fn canonicalize_provider_tool_calls(
    tools: &[Box<dyn Tool>],
    calls: Vec<crate::providers::ToolCall>,
 ) -> Vec<crate::providers::ToolCall> {
-    calls.into_iter()
+    calls
+        .into_iter()
        .map(|mut call| {
            if let Some(canonical_name) = resolve_registered_tool_name(tools, &call.name) {
                call.name = canonical_name;
@@ -1656,7 +1661,10 @@ mod tests {
            .expect("agent builder should succeed with valid config");

        let (event_tx, _event_rx) = tokio::sync::mpsc::channel(8);
-        let response = agent.turn_streamed("读取知乎热榜前10，并导出 excel 文件", event_tx).await.unwrap();
+        let response = agent
+            .turn_streamed("读取知乎热榜前10，并导出 excel 文件", event_tx)
+            .await
+            .unwrap();

        assert_eq!(
            response,
--- a/third_party/zeroclaw/src/lib.rs
+++ b/third_party/zeroclaw/src/lib.rs
@@ -71,7 +71,7 @@ pub mod routines;
 pub mod runtime;
 pub(crate) mod security;
 pub(crate) mod service;
-pub(crate) mod skills;
+pub mod skills;
 pub mod sop;
 pub mod tools;
 pub(crate) mod trust;
@@ -83,6 +83,7 @@ pub mod verifiable_intent;
 pub mod plugins;

 pub use config::Config;
+pub use security::{AutonomyLevel, SecurityPolicy};

 /// Gateway management subcommands
 #[derive(Subcommand, Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
--- a/third_party/zeroclaw/src/providers/compatible.rs
+++ b/third_party/zeroclaw/src/providers/compatible.rs
@@ -1943,9 +1943,8 @@ impl Provider for OpenAiCompatibleProvider {
            reasoning_effort: self.reasoning_effort_for_model(model),
            tool_stream: self
                .tool_stream_for_tools(tools.as_ref().is_some_and(|tools| !tools.is_empty())),
-            tool_choice: self.tool_choice_for_tools(
-                tools.as_ref().is_some_and(|tools| !tools.is_empty()),
-            ),
+            tool_choice: self
+                .tool_choice_for_tools(tools.as_ref().is_some_and(|tools| !tools.is_empty())),
            tools,
            max_tokens: self.max_tokens,
        };
@@ -2099,9 +2098,8 @@ impl Provider for OpenAiCompatibleProvider {
                tool_stream: if options.enabled { Some(true) } else { None },
                stream: Some(options.enabled),
                tools: tools.clone(),
-                tool_choice: self.tool_choice_for_tools(
-                    tools.as_ref().is_some_and(|tools| !tools.is_empty()),
-                ),
+                tool_choice: self
+                    .tool_choice_for_tools(tools.as_ref().is_some_and(|tools| !tools.is_empty())),
                max_tokens: self.max_tokens,
            })
        } else {
--- a/third_party/zeroclaw/src/skills/mod.rs
+++ b/third_party/zeroclaw/src/skills/mod.rs
@@ -816,12 +816,22 @@ pub fn skills_to_prompt_with_mode(
            let registered: Vec<_> = skill
                .tools
                .iter()
-                .filter(|t| matches!(t.kind.as_str(), "shell" | "script" | "http" | "browser_script"))
+                .filter(|t| {
+                    matches!(
+                        t.kind.as_str(),
+                        "shell" | "script" | "http" | "browser_script"
+                    )
+                })
                .collect();
            let unregistered: Vec<_> = skill
                .tools
                .iter()
-                .filter(|t| !matches!(t.kind.as_str(), "shell" | "script" | "http" | "browser_script"))
+                .filter(|t| {
+                    !matches!(
+                        t.kind.as_str(),
+                        "shell" | "script" | "http" | "browser_script"
+                    )
+                })
                .collect();

            if !registered.is_empty() {
--- a/third_party/zeroclaw/src/tools/read_skill.rs
+++ b/third_party/zeroclaw/src/tools/read_skill.rs
@@ -154,7 +154,9 @@ pub async fn read_skill_bundle(location: &Path) -> std::io::Result<String> {
    let Some(skill_root) = location.parent() else {
        return Ok(primary);
    };
-    let skill_root = skill_root.canonicalize().unwrap_or_else(|_| skill_root.to_path_buf());
+    let skill_root = skill_root
+        .canonicalize()
+        .unwrap_or_else(|_| skill_root.to_path_buf());
    let mut output = primary.clone();
    let mut appended = BTreeSet::new();
    let mut queued = BTreeSet::new();
@@ -275,16 +277,22 @@ fn extract_reference_paths(content: &str) -> Vec<String> {
 }

 fn looks_like_relative_reference_path(raw: &str) -> bool {
-    if raw.is_empty() ||
-        raw.starts_with('/') ||
-        raw.starts_with("http://") ||
-        raw.starts_with("https://") ||
-        raw.starts_with('#')
+    if raw.is_empty()
+        || raw.starts_with('/')
+        || raw.starts_with("http://")
+        || raw.starts_with("https://")
+        || raw.starts_with('#')
    {
        return false;
    }

-    let candidate = raw.split('#').next().unwrap_or(raw).split('?').next().unwrap_or(raw);
+    let candidate = raw
+        .split('#')
+        .next()
+        .unwrap_or(raw)
+        .split('?')
+        .next()
+        .unwrap_or(raw);
    let path = Path::new(candidate);
    if path
        .components()
@@ -418,9 +426,15 @@ description = "Ship safely"

        assert!(result.success);
        assert!(result.output.contains("# Zhihu Hotlist"));
-        assert!(result.output.contains("## Referenced File: references/collection-flow.md"));
-        assert!(result.output.contains("Collect rows from the hotlist first."));
-        assert!(result.output.contains("## Referenced File: references/data-quality.md"));
+        assert!(result
+            .output
+            .contains("## Referenced File: references/collection-flow.md"));
+        assert!(result
+            .output
+            .contains("Collect rows from the hotlist first."));
+        assert!(result
+            .output
+            .contains("## Referenced File: references/data-quality.md"));
        assert!(result.output.contains("Mark partial metrics explicitly."));
    }
 }
Author	SHA1	Message	Date
zyl	2ae71fb1c9	compat: probe zhihu extractor before renavigate	2026-03-30 18:39:49 +08:00
zyl	5bccd02d6f	docs: add sgclaw project briefing materials	2026-03-30 18:39:49 +08:00
zyl	f51d6b7659	sgclaw: snapshot today's runtime and skill updates	2026-03-30 18:39:49 +08:00
zyl	c793bfc6a1	build: restore repo cargo registry override	2026-03-30 18:39:49 +08:00
zyl	305b6d5110	build: stop using broken rsproxy registry	2026-03-30 18:39:49 +08:00
zyl	4c4f45581f	tests: cover nested zhihu creator write entry	2026-03-30 18:39:49 +08:00
zyl	cd94904329	sgclaw: stop zhihu publish flow before editor on creator page	2026-03-30 18:39:49 +08:00