The incremental sync used `find -type f` which misses symlinks (type l),
so dataset symlinks were never detected and synced to the sandbox.
Additionally, `tar.add()` without `dereference=True` would store broken
symlinks pointing to host-only paths.
- _list_local_changed_files: match both regular files and symlinks
- _tar_workspace_entries: dereference symlinks to pack actual content
- Unify dataset path to `datasets/` (plural) in prompts and SKILL.md
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When find_device_by_area matches an area successfully but the area lacks the requested device type (e.g., DCU/light), the AI will now proactively suggest nearby controllable areas.
Related Issue: #2766
* add page number
* feat: add skill feature memory
添加 skill 功能的 feature memory，记录技能包管理服务和 Hook 系统的核心信息。
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* feat(skill): add feature memory with changelog and decisions
添加 skill 功能的完整记忆文档：
Changelog:
- 2025-Q4: 初始实现 (GRPC 层 + 内置 skills)
- 2026-Q1: API 完善 (REST API + Hook 系统)
Design Decisions:
- 001: Skill 架构设计 (目录结构、Hook 系统)
- 002: 上传安全措施 (ZipSlip、路径遍历防护)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* soffice sharp 支持
* shell_env support
* feat: move enable_thinking control from docker-compose to request body
Remove DEFAULT_THINKING_ENABLE environment variable from docker-compose
and settings.py. The enable_thinking flag is now solely controlled via
request body (default: false), as felo-mygpt already passes this config
from RobotConfig database.
Closessparticleinc/felo-mygpt#2473
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: 朱潮 <zhuchaowe@users.noreply.github.com>
Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
Co-authored-by: zhuchao <zhuchaowe@163.com>
将 fact extraction prompt 中的技术性格式改为自然语言：
- "Contact: [name] (relationship, referred as [nick])" → "[name] is a [relationship], also called [nick]"
- 移除 "DEFAULT when user says" 等技术标记
- 添加 Plain Language Rule 明确要求输出通俗易懂的文本
- 更新所有示例为自然语言风格
Co-authored-by: zhuchao <zhuchaowe@163.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
EB API 始终返回人员最后已知坐标，不区分在馆/离馆。通过检查 last_communication_time
与当前时刻的差值（阈值5分钟/24小时），实现三级在離判定并提供对应的日语回答模板。
Closessparticleinc/mygpt-frontend#1963
Co-authored-by: zhuchao <zhuchaowe@163.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
