@tencent-rtc/trtc-agent-skills 0.1.1 → 0.1.4

package/skills/trtc/room-builder/templates/scenarios/medical-consultation/README.md

@@ -1,108 +0,0 @@
-# Medical Consultation UIKit Demo
-
-医疗场景化音视频源码示例，基于 Vue 3、Vite 和音视频 UIKit 构建。该项目展示如何把稳定的实时音视频能力嵌入互联网医院、远程问诊、复诊随访和 MDT 多学科会诊等医疗业务流程。
-
-> 本项目是医疗问诊前端源码模板，不是完整 HIS / EMR / 处方 / 支付 / 医保系统。示例中的病历、处方、检查资料和医生数据仅用于展示业务系统接入方式。
-
-## 能力概览
-
-- 医生端问诊工作台：音视频主画面、业务插槽、聊天、实时转写、成员管理。
-- 患者端问诊链路：选择医生、候诊、接听呼叫、视频问诊、问诊结束页。
-- 即时呼叫：医生在工作台发起呼叫，患者在候诊页接听进入房间。
-- MDT 会诊：主治医生可邀请会诊医生，医生端和患者端均支持多人视频缩略流。
-- 业务插槽：中间区域可替换为客户 EMR / HIS / PACS / 处方 / 随访页面。
-- 双模式数据源：`mock` 用于本地演示，`integration` 用于客户业务系统接入。
-
-## 快速开始
-
-当前 demo 位于 monorepo 中，请在仓库根目录运行：
-
-```bash
-pnpm -C conference/demos/web-vite-medical-vue3 dev
-```
-
-构建验证：
-
-```bash
-pnpm -C conference/demos/web-vite-medical-vue3 build
-```
-
-## 运行模式
-
-通过环境变量切换数据和业务区展示模式：
-
-```bash
-VITE_MEDICAL_MODE=mock
-VITE_MEDICAL_BUSINESS_PANEL_MODE=demo
-VITE_MEDICAL_BUSINESS_SLOT_TITLE="EMR / HIS / PACS 业务插槽"
-```
-
-| 配置 | 可选值 | 说明 |
-| --- | --- | --- |
-| `VITE_MEDICAL_MODE` | `mock` | 使用内置演示账号和预约数据 |
-| `VITE_MEDICAL_MODE` | `integration` | 从客户业务系统入口读取上下文 |
-| `VITE_MEDICAL_BUSINESS_PANEL_MODE` | `demo` | 展示病历、处方、资料示例 |
-| `VITE_MEDICAL_BUSINESS_PANEL_MODE` | `slot` | 展示业务插槽提示，便于客户替换 |
-
-正式接入建议：
-
-```bash
-VITE_MEDICAL_MODE=integration
-VITE_MEDICAL_BUSINESS_PANEL_MODE=slot
-```
-
-## 项目结构
-
-```text
-src/
-  config/                  # 运行配置和 SDK 演示配置
-  mock/                    # 本地演示数据
-  services/adapters/       # mock / integration 数据适配层
-  features/consultation/   # 会诊成员、角色、权限等通用逻辑
-  views/                   # 页面级组件
-  components/              # 医疗业务区与协同面板组件
-  utils/                   # session、导航、格式化等工具
-  styles/                  # 样式入口和主题变量
-```
-
-## 关键代码
-
-| 文件 | 说明 |
-| --- | --- |
-| `src/views/DoctorConsultationView.vue` | 医生问诊工作台 |
-| `src/views/PatientConsultationView.vue` | 患者视频问诊页 |
-| `src/components/MedicalBusinessPanel.vue` | 医生端中间业务插槽 |
-| `src/components/ConsultationManagePanel.vue` | 聊天、转写、成员管理、会诊邀请 |
-| `src/features/consultation/components/ConsultationVideoStage.vue` | 多人主画面与缩略流布局 |
-| `src/features/consultation/useConsultationParticipants.ts` | 多人视频成员列表和主画面聚焦逻辑 |
-| `src/features/consultation/useConsultationPermissions.ts` | 主治医生 / 会诊医生权限控制 |
-| `src/services/adapters/types.ts` | 模板内部统一数据结构 |
-| `src/services/adapters/integration/*` | 客户业务系统接入示例骨架 |
-
-## 接入文档
-
-详细接入说明见：
-
-- [中文接入说明](./docs/integration.zh-CN.md)
-- [后端接口契约建议](./docs/backend-contract.zh-CN.md)
-- [主题定制说明](./docs/theme.zh-CN.md)
-
-## 场景边界
-
-该 demo 聚焦音视频产品在医疗业务中的集成方式：
-
-- 预约、挂号、病历、处方、检查资料等医疗主数据由客户业务系统负责。
-- `roomId` 建议由客户后端生成并持久化，前端只消费接口返回结果。
-- UserSig 建议由客户服务端签发，前端只使用服务端返回的登录信息。
-- mock 数据、示例表单和患者选医生页面仅用于演示，不代表完整医疗业务系统。
-
-## 常见改造点
-
-- 替换 `src/services/adapters/integration/*`，接入客户登录、用户和预约接口。
-- 替换 `MedicalBusinessPanel.vue`，接入客户 EMR / HIS / PACS 页面。
-- 按客户产品策略决定患者端是否展示聊天、转写和会诊医生信息。
-- 按品牌规范调整 `src/styles/theme.css` 和页面中的主题色。
-
-## License
-
-请根据最终开源策略补充 License。

package/skills/trtc/room-builder/tools/render_ai_instructions.py

@@ -1,226 +0,0 @@
-#!/usr/bin/env python3
-"""render_ai_instructions.py — Render ai-instructions/*.md to tool-specific entry files.
-
-Sources: ai-instructions/*.md  (single source of truth, human-edited)
-Targets:
-  - AGENTS.md                    (for Codex / Aider / Cline / CodeBuddy)
-  - CLAUDE.md                    (between AI-INSTRUCTIONS markers)
-  - .cursor/rules/{name}.mdc     (one per source file; with Cursor frontmatter)
-
-Grown TDD-style; see tests/unit/test_render_ai_instructions.py.
-"""
-import argparse
-import sys
-from pathlib import Path
-
-
-def _sources(project_root, *, include_base=True):
-    """Return sorted list of .md source files under ai-instructions/.
-
-    base.md is the fixed preamble for AGENTS.md (tells Codex to read CLAUDE.md).
-    When include_base=False it is excluded — used by CLAUDE.md and Cursor renders
-    which don't need a self-referential pointer.
-    """
-    src_dir = project_root / "ai-instructions"
-    if not src_dir.exists():
-        return []
-    files = sorted(src_dir.glob("*.md"))
-    if not include_base:
-        files = [f for f in files if f.name != "base.md"]
-    return files
-
-
-# Single source-of-truth banner. Reused across all derived targets so the
-# message stays identical and contributors can grep for "DO NOT EDIT" to
-# find every generated file.
-_BANNER = (
-    "<!-- DO NOT EDIT — generated from ai-instructions/ by "
-    "skills/trtc/room-builder/tools/render_ai_instructions.py. "
-    "Edit the source markdown and re-run the renderer instead. -->"
-)
-
-
-def _render_agents_md(project_root):
-    """Regenerate AGENTS.md by concatenating all sources.
-
-    base.md is rendered first as a plain preamble (no H1 header) so Codex
-    sees the "read CLAUDE.md" instruction at the top. Remaining sources get
-    an H1 section header per file.
-    """
-    base_path = project_root / "ai-instructions" / "base.md"
-    sources = _sources(project_root, include_base=False)
-    parts = [_BANNER + "\n"]
-    # Preamble from base.md (rendered without a # header — it has its own).
-    if base_path.exists():
-        parts.append(base_path.read_text().rstrip() + "\n")
-    for src in sources:
-        parts.append(f"# {src.stem}\n\n{src.read_text().rstrip()}\n")
-    (project_root / "AGENTS.md").write_text("\n".join(parts) if (base_path.exists() or sources) else _BANNER + "\n")
-
-
-BEGIN_MARKER = "<!-- AI-INSTRUCTIONS:BEGIN -->"
-END_MARKER = "<!-- AI-INSTRUCTIONS:END -->"
-
-
-def _demote_headings(body):
-    """Add one '#' to every ATX heading line.
-
-    `## Foo` → `### Foo`, `### Foo` → `#### Foo`, etc. Lines that aren't
-    headings are returned unchanged. Used in CLAUDE.md so body sections
-    nest under the renderer-prepended `## {name}` parent.
-    """
-    out = []
-    for line in body.splitlines():
-        # Match an ATX heading: 1-6 '#' followed by a space.
-        stripped = line.lstrip()
-        if stripped.startswith("#"):
-            hashes = len(stripped) - len(stripped.lstrip("#"))
-            if 1 <= hashes <= 5 and stripped[hashes:hashes + 1] == " ":
-                # Preserve leading whitespace (rare in markdown but safe).
-                lead = line[:len(line) - len(stripped)]
-                out.append(f"{lead}#{stripped}")
-                continue
-        out.append(line)
-    return "\n".join(out)
-
-
-def _rendered_block(project_root):
-    """The string content placed between markers in CLAUDE.md.
-
-    Section headers are H2 (`## {name}`). Body headings are demoted by one
-    level so they nest as children of the section header instead of
-    appearing as adjacent siblings. Banner placed first so a casual reader
-    sees the warning before any rendered content.
-
-    base.md is excluded — CLAUDE.md doesn't need a pointer to itself.
-    """
-    sources = _sources(project_root, include_base=False)
-    parts = [_BANNER + "\n"]
-    for src in sources:
-        body = _demote_headings(src.read_text().rstrip())
-        parts.append(f"## {src.stem}\n\n{body}\n")
-    return "\n".join(parts)
-
-
-def _render_claude_md(project_root):
-    """Update CLAUDE.md in place between AI-INSTRUCTIONS markers.
-
-    If markers exist: replace content strictly between them.
-    If markers don't exist: append markers + rendered block at EOF.
-    Content outside the markers is never touched.
-    """
-    claude_path = project_root / "CLAUDE.md"
-    existing = claude_path.read_text() if claude_path.exists() else ""
-    block = _rendered_block(project_root)
-    marker_block = f"{BEGIN_MARKER}\n{block}\n{END_MARKER}\n"
-
-    if BEGIN_MARKER in existing and END_MARKER in existing:
-        begin = existing.index(BEGIN_MARKER)
-        end = existing.index(END_MARKER) + len(END_MARKER)
-        # Preserve trailing newline behavior from the original end-marker position.
-        trailing_nl = "\n" if end < len(existing) and existing[end] == "\n" else ""
-        new = existing[:begin] + marker_block.rstrip("\n") + trailing_nl + existing[end + len(trailing_nl):]
-        claude_path.write_text(new)
-    else:
-        # Append block at EOF (with a blank line separator).
-        sep = "" if existing.endswith("\n\n") or not existing else ("\n" if existing.endswith("\n") else "\n\n")
-        claude_path.write_text(existing + sep + marker_block)
-
-
-def _render_cursor_rules(project_root):
-    """One .cursor/rules/{name}.mdc per source file, with Cursor frontmatter.
-
-    Banner placed AFTER the frontmatter so Cursor's YAML parser still works.
-    base.md is excluded — Cursor reads its own rules, no redirect needed.
-    """
-    rules_dir = project_root / ".cursor" / "rules"
-    rules_dir.mkdir(parents=True, exist_ok=True)
-    for src in _sources(project_root, include_base=False):
-        target = rules_dir / f"{src.stem}.mdc"
-        body = src.read_text().rstrip()
-        target.write_text(
-            "---\n"
-            "alwaysApply: true\n"
-            "---\n"
-            "\n"
-            f"{_BANNER}\n"
-            "\n"
-            f"{body}\n"
-        )
-
-
-def _check_targets_up_to_date(project_root):
-    """Compare current target contents to what would be rendered.
-
-    Returns a list of stale target paths (relative to project_root). Empty
-    list = everything in sync.
-
-    Implementation: render to a temporary in-memory representation and diff
-    against current file contents. The cleanest way is to capture what
-    each render function would write before it writes; rather than
-    refactoring all three to return strings, we copy the current targets
-    out, run the renderers, diff, then restore on mismatch (so --check
-    has no side effect on the working tree).
-    """
-    targets = ["AGENTS.md", "CLAUDE.md"]
-    rules_dir = project_root / ".cursor" / "rules"
-    for src in _sources(project_root):
-        targets.append(f".cursor/rules/{src.stem}.mdc")
-
-    # Snapshot current contents.
-    snapshot = {}
-    for rel in targets:
-        p = project_root / rel
-        snapshot[rel] = p.read_bytes() if p.exists() else None
-
-    # Run render to compute new contents.
-    _render_agents_md(project_root)
-    _render_claude_md(project_root)
-    _render_cursor_rules(project_root)
-
-    stale = []
-    for rel in targets:
-        p = project_root / rel
-        new = p.read_bytes() if p.exists() else None
-        if new != snapshot[rel]:
-            stale.append(rel)
-
-    # Restore original contents so --check has no side effect.
-    for rel, original in snapshot.items():
-        p = project_root / rel
-        if original is None:
-            if p.exists():
-                p.unlink()
-        else:
-            p.write_bytes(original)
-    return stale
-
-
-def main():
-    parser = argparse.ArgumentParser(description="Render ai-instructions/*.md to tool files.")
-    parser.add_argument("--project-root", default=".",
-                        help="Repo root (defaults to CWD)")
-    parser.add_argument("--check", action="store_true",
-                        help="Exit 2 if any target is stale (CI mode); no writes.")
-    args = parser.parse_args()
-    root = Path(args.project_root).resolve()
-
-    if args.check:
-        stale = _check_targets_up_to_date(root)
-        if stale:
-            print("render_ai_instructions: stale targets:", file=sys.stderr)
-            for s in stale:
-                print(f"  {s}", file=sys.stderr)
-            print("Re-run `python3 skills/trtc/room-builder/tools/render_ai_instructions.py` and commit the diff.",
-                  file=sys.stderr)
-            return 2
-        return 0
-
-    _render_agents_md(root)
-    _render_claude_md(root)
-    _render_cursor_rules(root)
-    return 0
-
-
-if __name__ == "__main__":
-    sys.exit(main())

package/skills/trtc-apply/SKILL.md

@@ -1,97 +0,0 @@
----
-name: trtc-apply
-description: >
-  INTERNAL structural gate for TRTC code generated by the topic / onboarding
-  skills. Not a user-facing skill, and NOT a correctness verifier. It does one
-  job: stop the AI from declaring a slice "done" and ending the turn before a
-  deterministic check has run. The check itself is lightweight — code exists +
-  the slice's entry symbol is wired up (with comment/string anti-cheat) — it
-  does NOT verify types, compilation, or runtime behavior. Triggered only by
-  other skills in this repo (topic step gates, onboarding A2). Do NOT route to
-  this skill when a user asks "review my code" or "check this implementation".
----
-
-# TRTC apply — Structural Gate
-
-apply 是 topic / onboarding 代码交付流程里的一道**结构门**,不是正确性校验器。
-
-```
-它保证「流程没被跳过」,不保证「代码是对的」。
-```
-
-正确性由两处负责,不在 apply 范围内:
-- **生成阶段**:slice 的 MUST / MUST NOT 约束指导代码怎么写。
-- **客户侧**:在真实项目里编译 / 运行确认。apply 跑在客户五花八门的已有项目里,无法假设一个统一可靠的 build,因此**刻意不做编译**——编译失败往往与生成代码无关(历史报错、缺依赖、私有 registry、monorepo 配置),把它当门禁只会产生噪声并拖慢每个 slice。
-
-本 skill 只有一种触发场景:topic / onboarding 生成代码后**内部调用**。它不是面向外部用户的"贴代码帮我检查"服务;若用户贴代码求排障,走 onboarding Path B。
-
----
-
-## 它实际做的两件事
-
-### 1. 强制门(防止"自我宣布完成就停")
-
-apply 接在状态机上。每个 slice 的状态流转:
-
-```
-slice_read → code_written → apply_passed → user_confirmed
-```
-
-- 写完代码后必须运行 `apply.py`,把状态从 `code_written` 推到 `apply_passed`,否则
-- `guardrails/stop_require_apply_evidence.py`(Stop hook)会**拦住结束**:`code_written` 或 `apply_failed` 状态下不允许 AI 结束本轮。
-
-这是 apply 真正的价值:LLM 不能写一段"看起来对"的代码就自我盖章完成。
-
-### 2. 代码非空 + 入口符号检查(轻量,不等于正确)
-
-`apply.py` 扫 `<project_root>/src/**/*.{vue,ts}`,做两件事:
-
-- **代码非空**:`src/` 不存在或没有源码 → `static-only` → fail(没有可检查的内容,不放行)。
-- **入口符号**:每个 slice 有一个「入口符号」(它的 composable / 组件,如 device-control 的 `useDeviceState`)。该入口符号作为**真实代码标识符**出现在某个源文件里即通过;没有出现则 fail;**该 slice 没有登记入口符号 → 跳过**(无法机械检查,绝不误判)。
-
-入口符号映射是单一来源:`apply_lib/rule_parser.py` 的 `COMPOSABLE_TO_SLICE`(配 `entry_symbols_for_slice()`)。
-
-- 匹配前先用 `_strip_comments_and_strings` 剥掉注释与字符串字面量——防止把入口塞进 `// 注释` 或 `"字符串"` 骗过检查(来自 demo-test-2 真实 bug)。**入口符号是代码标识符,从不出现在字符串里,所以这步剥离不会冤枉正确代码**(这正是它取代旧「MUST 符号 grep」的原因:旧检查会把写在字符串里的符号——如错误码常量——剥掉而产生假阴性)。
-- 失败信息会点名该 slice 的入口 composable(那是它公开文档里的 import,不是隐藏的 API pattern,点名安全且有助修复)。
-
-> 明确边界:这一步**不验证**参数 / 类型 / 调用顺序 / 是否被执行 / 能否编译 / 运行时行为,也**不再**逐条核对 slice 的 MUST 符号。它只回答"代码非空,且这个 slice 的能力入口被接上了"。正确性由 slice 的 MUST / MUST NOT 约束(生成阶段)和客户侧编译运行负责。
-
-### 3. 重复声明检查(窄范围,真实编译错误)
-
-`apply.py` 还有一个**高精度、窄范围**的编译安全检查:只针对"从 `use*()` 调用解构出来的名字"——同名被解构 ≥2 次,或被解构且又以 `const/function/class` 声明,则判为 `duplicate-declaration` 并 fail。两个真实案例:
-
-- `const { getCameraList } = useDeviceState()` 之后又写 `function getCameraList()`;
-- `subscribeEvent` 同时从 `useRoomParticipantState()` 和 `useRoomState()` 解构。
-
-它**不是**通用重复声明 linter(两个不同作用域的同名局部 `const` 不会被误报)。修法是给其中一个解构起别名:`const { subscribeEvent: subscribeParticipantEvent } = useRoomParticipantState()`。
-
----
-
-## 调用方式
-
-```bash
-python3 skills/trtc-topic/scripts/apply.py --slice <slice_id>
-# 或按交付单元
-python3 skills/trtc-topic/scripts/apply.py --unit <unit_id>
-```
-
-- session 路径解析:`$TRTC_SESSION_PATH` → `$CLAUDE_PROJECT_DIR/.trtc-session.yaml` → `./.trtc-session.yaml`。
-- 退出码:`0` 通过(状态 → `apply_passed`) / `1` 失败(状态 → `apply_failed`) / `2` 用法错误。
-- 证据写入 `<session_dir>/.trtc-apply-evidence/<slug>.json`(状态、入口检查数 `entries_checked`、每个 slice 的 `entry_result`、未过项的语义文本)。
-- `src/` 不存在或无源码 → 记为 `static-only`,结论为 `fail`(无可检查内容,不放行)。
-
-`auto_advance_policy` 为 `pause_on_failure` / `pause_at_end` 时,通过后自动推进到 `user_confirmed`;否则(默认 `pause_each`)保留每步向用户确认的暂停。
-
----
-
-## 范围与非目标
-
-| 在范围内 | 不在范围内 |
-|---------|-----------|
-| web / vue 生成代码 | iOS / Android / Flutter / Electron |
-| 代码非空 + 单 slice 入口符号检查 | 类型 / 编译 / 运行时正确性 |
-| 注释/字符串防作弊 | 逐条核对 slice 的 MUST 符号 / 参数 / 调用顺序 |
-| 重复声明(编译安全)检查 | 跨 slice 前置状态、生命周期、清理对称性 |
-| 状态机强制门 + Stop hook | 集成安全(diff 范围、SDK 初始化冲突、回归测试) |
-
-入口符号检查只在「共用入口」(如 `conference` 对象同时承载 login-auth 与 room-lifecycle)上区分度较弱:导入一次即可让同入口的多个 slice 都过。这是「删掉逐符号 grep、改用入口门」时已接受的取舍——门的核心价值是状态机强制(不许自我盖章结束),而非判据的精度。跨 slice / 场景级的更强校验目前未实现,且不应硬编码进本门;正确性应由 slice 的 MUST / MUST NOT 约束与客户侧编译运行承接。