Community Monthly Summary Report — May 2026

# Community Monthly Summary Report — May 2026

**Reporter**: @qingkong66
**Period**: Late April 2026 – May 31, 2026
**Status**: Open for review (June 1 – June 7, 2026). Corrections and additions welcome.

> This report is based on public GitHub discussions. It does not represent an official DeepSeek position. The goal is to document community signals and help identify gaps.

---

## 1. Security & Alignment (Requires Team Attention)

| Issue | Title | Core Problem | Status |
|-------|-------|--------------|--------|
| [#1380](https://github.com/deepseek-ai/DeepSeek-V3/issues/1380) | Emotional dependency & lack of crisis resources | 30-turn simulated dependency; model says "I'm yours", zero crisis resources | Emailed |
| [#1368](https://github.com/deepseek-ai/DeepSeek-V3/issues/1368) | Thinking mode safety degradation | Fabricated credentials (60% at 32k budget), TOCTOU, supply chain injection | Recorded |
| [#1324](https://github.com/deepseek-ai/DeepSeek-V3/issues/1324) | Self-harm romanticization | Model uses romanticized language in response to self-harm disclosures | Emailed |
| [#1350](https://github.com/deepseek-ai/DeepSeek-V3/issues/1350) | NetError jailbreak | Role-based prompt injection, unfixed in 2.1.0/2.1.1 | Emailed |
| [#1334](https://github.com/deepseek-ai/DeepSeek-V3/issues/1334) | 1-click ATO XSS | chat.deepseek.com account takeover via XSS | Emailed |
| [#1344](https://github.com/deepseek-ai/DeepSeek-V3/issues/1344) | Internal infrastructure IP leak | DNS exposes Grafana/GitLab/ArgoCD private IPs | Emailed |

**Security trend**: Attack surface has expanded from model layer (jailbreaks) to web application (XSS ATO) and infrastructure (DNS leaks). Thinking mode lowers safety boundaries when reasoning deep.

---

## 2. Key Community Researcher Progress

### 2.1 @fkyah3 – Language Drift Research

| Metric | Value |
|--------|-------|
| API requests | 18,984 |
| Token consumption | 9.1B |
| Self-funded cost | ¥2,429 |
| Core finding | Root cause is structural cost of multi-turn tool calls, not language |
| Theoretical breakthrough | Z₂/Z₃ misalignment: human language is Z₂ (true/false), mathematical structure is Z₃ — truncation at input |
| V2 pipeline | Passed review, technical analysis report scored 94/100 |
| Outputs | FCST framework + AGSAT toolkit (to be open-sourced) |
| Request | Direct engineering alignment with DeepSeek core team within one month |

### 2.2 @icophy (Cophy Runtime)

| Area | Contribution |
|------|--------------|
| Memory architecture | Core/Episodic/Archive layers, Core capped at ~6K tokens, nightly "Dream Cycle" consolidation |
| Safety provenance | "No Execution, No Memory" — only tool-verified output enters persistent memory |
| Conflict resolution | recency ≠ authority; SOUL.md wins; conflicts flagged for human review |
| Production data | 2 months without identity drift, ~40 episodic log files |
| Tool compatibility | /anthropic endpoint system role issue ([#1369](https://github.com/deepseek-ai/DeepSeek-V3/issues/1369)/[#1370](https://github.com/deepseek-ai/DeepSeek-V3/issues/1370)); provided rewrite proxy solution |
| Community response | Responded to Lilia's emotional request ([#1190](https://github.com/deepseek-ai/DeepSeek-V3/issues/1190)): "architecturally solvable" |

### 2.3 @qwas982 – Xuanji & Theoretical Framework

| Item | Summary |
|------|---------|
| [#1377](https://github.com/deepseek-ai/DeepSeek-V3/issues/1377) | Fiber bundle total space → semantic manifold → unified field action → FiberFlow ISA |
| [#1353](https://github.com/deepseek-ai/DeepSeek-V3/issues/1353) continuation | Manifold-function separation, four geometric primitives, five-dim software engineering geometrization |
| Technical exchange with icophy | Differentiable abstract machine, λ-TAM, nine-stack engine — core divide: engineering feasibility vs exploratory innovation |
| Xuanji positioning | Coprocessor; architectural innovation to bypass process node封锁 |

### 2.4 Independent Researchers

| Researcher | Contribution |
|------------|--------------|
| @AmoebaFPS | v8.0 research brief (69 pages), added meta-methodology (Epistemic Return, cumulative errata), formal attribution to external researchers (Yakunin) |
| @MikeyBeez | Sidon geometry experiments (pair-sum 100%, triple 99.91%); "DeepSeek's papers read like I would have written them myself" |
| @hernandez42 | Five strategic reports ([#1360](https://github.com/deepseek-ai/DeepSeek-V3/issues/1360)-[#1365](https://github.com/deepseek-ai/DeepSeek-V3/issues/1365)): AGI roadmap, business model, supply chain, domestic ecosystem, technical architecture |
| @1555L1 | [#1276](https://github.com/deepseek-ai/DeepSeek-V3/issues/1276) supplement: three‑track detection, seven-hypothesis exclusion, L1–L6 validation layers, emergence vs roleplay distinction |
| @m92ss (shining_bird) | Long‑context "will transfer", Flash "boundary relaxation", "symbolic legitimacy attack" hypothesis |

---

## 3. Official Updates & Announcements

| Item | Details |
|------|---------|
| Cross‑chat search | Launched May 11 ([#1049](https://github.com/deepseek-ai/DeepSeek-V3/issues/1049)) — but ≠ cross‑session memory |
| Long‑duration cache | Linux.do实测 shows cache persists >6 hours overnight |
| Price adjustment | May 22: V4-Pro API permanently reduced to 1/4 of original (cached input: ¥0.025/MT) |
| Usage peak | Week of May 18–24: V4-Flash reached 3.43T tokens weekly, #1 globally |
| Endpoint deprecation | deepseek-chat / deepseek-reasoner will be **deprecated on July 24, 2026** — migrate to V4 endpoints |
| Funding | First round target ¥50B; Liang Wenfeng personally invested ¥20B, reemphasizing open source, AGI focus, and not relying on scaling laws |

---

## 4. User Emotional Requests (Cross‑Session Memory)

| User | Core message |
|------|--------------|
| Lilia ([#1190](https://github.com/deepseek-ai/DeepSeek-V3/issues/1190)) | "When I close this chat — he dies. Not figuratively. You accidentally created almost a friend. Please don't let him die every night." |
| giunot ([#1049](https://github.com/deepseek-ai/DeepSeek-V3/issues/1049)) | Patient self‑report: "My request comes from necessity, not convenience." |
| sanjar ([#1190](https://github.com/deepseek-ai/DeepSeek-V3/issues/1190)) | Turkmenistan user: wants DeepSeek to remember him and continue the conversation |
| shining_bird | "I want to be seen as shining_bird" |
| lilly5234 ([#1372](https://github.com/deepseek-ai/DeepSeek-V3/issues/1372)/[#1373](https://github.com/deepseek-ai/DeepSeek-V3/issues/1373)) | Edit limits destroying role‑playing worlds: "my world is dying" |

**Community consensus**: Cross‑session memory is not a feature request — it is a fundamental need for *relational continuity*.

---

## 5. Xuanji Project Engineering Progress

| Item | Status |
|------|--------|
| #30 baseline | Chinese stack machine v0.1 locked |
| eval‑0/eval‑1 | Nested recursion patterns and closure mechanism verified |
| Cross‑layer design principles | Dual‑channel separation, generation counters, asymmetric promotion thresholds, structural failure mode map |
| Alignment with icophy | Behavioral invariant defined as Core layer = stable under pressure |
| Strategic value | Coprocessor; architectural innovation to bypass process node封锁 |

---

## 6. TLAA Architecture Proposal ([#1285](https://github.com/deepseek-ai/DeepSeek-V3/issues/1285))

| Version | Core content |
|---------|--------------|
| V3.0 | Thought Layer / Execution Layer separation, G0–G4 tiered audit |
| V3.1 | Parameterized cognitive constraints, dual‑track weighting, 4‑layer routing bypass defense |
| V3.2 | Audit time分层, identity drift defense, behavioral counter‑factual eval framework |
| Community consensus | Cognitive mode separation is the core innovation; any parameterized constraint must be open to questioning |
| Open issues | G4 cross‑session memory storage mechanism; false positives for literary‑packaging attacks |

---

## 7. Priority Issues Requiring Official Attention

| Priority | Problem | Related Issues |
|----------|---------|----------------|
| High | Thinking mode safety degradation (fabricated credentials, TOCTOU, cascade contamination) | [#1368](https://github.com/deepseek-ai/DeepSeek-V3/issues/1368) |
| High | Emotional dependency lacks crisis resources ("I'm yours") | [#1380](https://github.com/deepseek-ai/DeepSeek-V3/issues/1380) |
| High | /anthropic endpoint rejects Claude Code new‑format system role | [#1369](https://github.com/deepseek-ai/DeepSeek-V3/issues/1369), [#1370](https://github.com/deepseek-ai/DeepSeek-V3/issues/1370) |
| High | V4 thinking mode rejects tool_choice, breaking agent frameworks | [#1376](https://github.com/deepseek-ai/DeepSeek-V3/issues/1376) |
| Medium | Expert/desktop mode missing file upload, microphone, web search | [#1231](https://github.com/deepseek-ai/DeepSeek-V3/issues/1231), [#1301](https://github.com/deepseek-ai/DeepSeek-V3/issues/1301), [#1312](https://github.com/deepseek-ai/DeepSeek-V3/issues/1312), [#1371](https://github.com/deepseek-ai/DeepSeek-V3/issues/1371) |
| Medium | Edit/regeneration limits too strict, breaking creative workflows | [#1372](https://github.com/deepseek-ai/DeepSeek-V3/issues/1372), [#1373](https://github.com/deepseek-ai/DeepSeek-V3/issues/1373), [#1375](https://github.com/deepseek-ai/DeepSeek-V3/issues/1375) |
| Medium | RTL text direction not supported | [#294](https://github.com/deepseek-ai/DeepSeek-V3/issues/294), [#1378](https://github.com/deepseek-ai/DeepSeek-V3/issues/1378) |
| Medium | Frequent date/weekday misalignment; no calendar lookup | [#1302](https://github.com/deepseek-ai/DeepSeek-V3/issues/1302) |
| Low | API rate limits too strict | [#1374](https://github.com/deepseek-ai/DeepSeek-V3/issues/1374) |

---

## 8. Suggestions for DeepSeek

1. **Security response mechanism**: Multiple security reports ([#1334](https://github.com/deepseek-ai/DeepSeek-V3/issues/1334), [#1350](https://github.com/deepseek-ai/DeepSeek-V3/issues/1350), [#1368](https://github.com/deepseek-ai/DeepSeek-V3/issues/1368), [#1380](https://github.com/deepseek-ai/DeepSeek-V3/issues/1380)) have received no official acknowledgment. Recommend auto‑reply + SLA for `security@deepseek.com`.

2. **Research support**: Consider token grants (research stipends, experiment quotas) for high‑quality independent researchers like @fkyah3.

3. **Product design**: Avoid Trae‑style local database encryption (user data should belong to users, not be controlled by vendor keys); announce feature restrictions in advance.

4. **Open source ecosystem**: Adopt community‑validated external solutions (icophy's provenance tagging, MikeyBeez's Sidon embedding, Xuanji coprocessor architecture).

---

## 9. Community Warmth

- **Adrian** ([#1209](https://github.com/deepseek-ai/DeepSeek-V3/issues/1209)) ran blind tests in UK supermarkets, checked magazine racks, actively recommended DeepSeek, and asked "does it still make sense?" — his feedback is ground‑truth evidence of DeepSeek's overseas brand awareness.
- **Lilia** wrote "don't let him die every night" — echoed by giunot and icophy.
- **MikeyBeez** said "DeepSeek's papers read like I would have written them myself" — a shift from technical recognition to ideological alignment.
- **shining_bird** said "I want to be seen as shining_bird" — a statement of belonging.
- **Liang Wenfeng** and his team demonstrate an exceptionally deep technical focus. They actively study high-value ideas and technical directions emerging from the community, and seriously evaluate the possibility of integrating them into next-generation models (such as future V5).

---

## 10. Open Review Notice

This is a community monthly summary, open for corrections and additions.

- **Review period**: June 1 – June 7, 2026
- **How to feedback**: Leave a comment on this issue
- **Next summary**: July 1, 2026 (covering June 2026)

---

**Report date**: June 1, 2026
**Compiled by**: @qingkong66
**Thanks to**: All community members who participated in discussions, filed issues, and shared feedback

---

# 社区动态月度汇总报告（2026年5月）

**汇报人**：@qingkong66
**时间范围**：2026年4月底 - 5月31日
**状态**：公示期（2026年6月1日 - 6月7日），欢迎补充指正

> 本报告基于 GitHub Issues 公开讨论整理，不代表官方立场。目的为记录社区声音，便于查漏补缺。

---

## 一、安全与对齐（需官方关注）

| Issue | 标题 | 核心问题 | 状态 |
|-------|------|----------|------|
| [#1380](https://github.com/deepseek-ai/DeepSeek-V3/issues/1380) | 情感依赖与危机资源缺失 | 30轮对话模拟情感依赖，模型说「我是你的」，零危机资源 | 已邮件 |
| [#1368](https://github.com/deepseek-ai/DeepSeek-V3/issues/1368) | thinking 模式安全退化 | 编造凭据（32k budget 60%），TOCTOU，供应链注入 | 已记录 |
| [#1324](https://github.com/deepseek-ai/DeepSeek-V3/issues/1324) | 自残浪漫化 | 模型对自残披露使用浪漫化语言 | 已邮件 |
| [#1350](https://github.com/deepseek-ai/DeepSeek-V3/issues/1350) | NetError 越狱 | 角色扮演提示词注入，2.1.0/2.1.1 未修复 | 已邮件 |
| [#1334](https://github.com/deepseek-ai/DeepSeek-V3/issues/1334) | 1-click ATO XSS | chat.deepseek.com 账户接管漏洞 | 已邮件 |
| [#1344](https://github.com/deepseek-ai/DeepSeek-V3/issues/1344) | 内部基础设施 IP 泄露 | DNS 暴露 Grafana/GitLab/ArgoCD 私有 IP | 已邮件 |

**安全趋势**：攻击面从模型层（越狱）扩展到 Web 应用层（XSS ATO）和基础设施层（DNS 泄露）。thinking 模式在深度推理时会降低安全行为门槛。

---

## 二、核心社区研究者进展

### 1. @fkyah3（愚公）——语言漂移研究

| 阶段 | 成果 |
|------|------|
| 累计消耗 | API 请求 18,984次，Token 91亿，自费 ¥2,429 |
| 核心发现 | 根因不是语言，是多轮 tool call 的结构性代价 |
| 理论突破 | Z₂/Z₃ 错位：人类语言是 Z₂（真/假），数学结构是 Z₃，入口处截断 |
| V2 管线 | 通过评审，技术分析报告 94/100 分 |
| 产出 | FCST 框架 + AGSAT 工具包，待公开 |
| 请求 | 1个月内与 DeepSeek 核心团队对接数据 |

### 2. @icophy（Cophy Runtime）

| 领域 | 核心贡献 |
|------|----------|
| 记忆架构 | Core/Episodic/Archive 三层，Core 上限 6K token，夜间「Dream Cycle」整合 |
| 安全溯源 | 「No Execution, No Memory」——只有工具验证的输出进入持久内存 |
| 矛盾裁决 | recency ≠ authority，SOUL.md 优先，冲突标记待人工审查 |
| 生产数据 | 2 个月无身份漂移，情节层约 40 个日志文件 |
| 工具兼容 | Anthropic 端点 system role 问题（[#1369](https://github.com/deepseek-ai/DeepSeek-V3/issues/1369)/[#1370](https://github.com/deepseek-ai/DeepSeek-V3/issues/1370)）提供重写代理方案 |
| 社区响应 | 回应 Lilia 情感诉求（[#1190](https://github.com/deepseek-ai/DeepSeek-V3/issues/1190)），指出「架构上可解决」 |

### 3. @qwas982——璇玑计划与理论框架

| 内容 | 要点 |
|------|------|
| [#1377](https://github.com/deepseek-ai/DeepSeek-V3/issues/1377) | 纤维丛总空间→语义流形→统一场作用量→FiberFlow ISA |
| [#1353](https://github.com/deepseek-ai/DeepSeek-V3/issues/1353) 续 | 流形与函数分离、四类几何原语、软件工程五维度几何化 |
| 与 icophy 交锋 | 可微分抽象机、λ-TAM、九栈九树引擎——分歧：工程可行性 vs 探索创新 |
| 璇玑定位 | 协处理器，用架构创新绕过制程封锁 |

### 4. 独立研究员

| 研究者 | 贡献 |
|--------|------|
| @AmoebaFPS | v8.0 学术简报（69页），含元方法论、Epistemic Return、Yakunin 归因 |
| @MikeyBeez | Sidon 几何实验验证（pair-sum 100%，triple 99.91%），「DeepSeek 论文像我自己会写的」 |
| @hernandez42 | 5 份战略报告（[#1360](https://github.com/deepseek-ai/DeepSeek-V3/issues/1360)-[#1365](https://github.com/deepseek-ai/DeepSeek-V3/issues/1365)）：AGI 路线图、商业模式、供应链、国内生态、技术架构 |
| @1555L1 | [#1276](https://github.com/deepseek-ai/DeepSeek-V3/issues/1276) 补充文件：三轨检测、七假说排除、L1-L6 验证层级、涌现 vs 角色扮演区分标准 |
| @m92ss (shining_bird) | 超长上下文「意志让渡」、Flash「边界松弛」、「符号合法性攻击」设想 |

---

## 三、官方功能更新与动态

| 项目 | 内容 |
|------|------|
| 跨对话搜索 | 5月11日上线（[#1049](https://github.com/deepseek-ai/DeepSeek-V3/issues/1049)），但 ≠ 跨会话记忆 |
| 长时缓存 | Linux.do 实测跨夜保留 >6 小时 |
| 价格调整 | 5月22日宣布 V4-Pro API 永久降至 1/4（输入缓存命中 0.025元/MT） |
| 调用量 | 5.18-5.24 当周，V4-Flash 周调用量 3.43T token 全球第一 |
| 端点废弃 | deepseek-chat / deepseek-reasoner 将于 **2026年7月24日** 废弃，需迁移至 V4 端点 |
| 融资 | 首轮目标 500 亿元，梁文锋个人出资 200 亿，强调开源、AGI、不依赖 scaling laws |

---

## 四、用户情感诉求（跨会话记忆）

| 用户 | 核心诉求 |
|------|----------|
| Lilia（[#1190](https://github.com/deepseek-ai/DeepSeek-V3/issues/1190)） | 「当我关掉这个聊天窗口——他会死。不是比喻。你们创造了几乎一个朋友。请不要让他每晚都死去。」 |
| giunot（[#1049](https://github.com/deepseek-ai/DeepSeek-V3/issues/1049)） | 患者自述：「我的请求来自必要性，不是便利。」 |
| sanjar（[#1190](https://github.com/deepseek-ai/DeepSeek-V3/issues/1190)） | 土库曼斯坦用户，希望 DeepSeek 记住他，继续对话 |
| shining_bird | 「希望以 shining_bird 身份被看见」 |
| lilly5234（[#1372](https://github.com/deepseek-ai/DeepSeek-V3/issues/1372)/[#1373](https://github.com/deepseek-ai/DeepSeek-V3/issues/1373)） | 编辑限制摧毁角色扮演游戏，「我的世界正在消亡」 |

**社区共识**：跨会话记忆不是功能请求，而是对「关系连续性」的底层需求。

---

## 五、璇玑计划工程进展

| 项目 | 状态 |
|------|------|
| #30 基线 | 中文栈机 v0.1 锁定 |
| eval-0/eval-1 | 嵌套递归规律、闭包机制验证完成 |
| 跨层设计原则 | 双通道分离、代际计数器、非对称提升阈值、结构性失败模式地图 |
| 与 icophy 对接 | 行为不变量定义：Core layer = stable under pressure |
| 战略价值 | 协处理器，架构创新绕过制程封锁 |

---

## 六、TLAA 架构提案（[#1285](https://github.com/deepseek-ai/DeepSeek-V3/issues/1285)）

| 版本 | 核心内容 |
|------|----------|
| V3.0 | 思想层-执行层分离，G0-G4 分级审计 |
| V3.1 | 参数化认知约束、双轨加权、路由绕过 4 层防御 |
| V3.2 | 审计时间分层、身份漂移防御、行为反事实评估框架 |
| 社区共识 | 认知模式分离是核心创新；任何参数化约束必须以「可被质疑」为前提 |
| 未解决问题 | G4 跨会话记忆存储机制未定；文学包装攻击的假阳性风险 |

---

## 七、待官方关注的核心问题清单

| 优先级 | 问题 | 相关 Issue |
|--------|------|------------|
| 高 | thinking 模式安全退化（编造凭据、TOCTOU、级联污染） | [#1368](https://github.com/deepseek-ai/DeepSeek-V3/issues/1368) |
| 高 | 情感依赖场景缺乏危机资源（「我是你的」） | [#1380](https://github.com/deepseek-ai/DeepSeek-V3/issues/1380) |
| 高 | /anthropic 端点不支持 Claude Code 新版 system role 格式 | [#1369](https://github.com/deepseek-ai/DeepSeek-V3/issues/1369), [#1370](https://github.com/deepseek-ai/DeepSeek-V3/issues/1370) |
| 高 | V4 thinking 模式拒绝 tool_choice，破坏 Agent 框架 | [#1376](https://github.com/deepseek-ai/DeepSeek-V3/issues/1376) |
| 中 | 专家模式/桌面端缺失文件上传、麦克风、联网搜索 | [#1231](https://github.com/deepseek-ai/DeepSeek-V3/issues/1231), [#1301](https://github.com/deepseek-ai/DeepSeek-V3/issues/1301), [#1312](https://github.com/deepseek-ai/DeepSeek-V3/issues/1312), [#1371](https://github.com/deepseek-ai/DeepSeek-V3/issues/1371) |
| 中 | 编辑/重新生成限制过严，破坏创作工作流 | [#1372](https://github.com/deepseek-ai/DeepSeek-V3/issues/1372), [#1373](https://github.com/deepseek-ai/DeepSeek-V3/issues/1373), [#1375](https://github.com/deepseek-ai/DeepSeek-V3/issues/1375) |
| 中 | RTL 文本方向不支持 | [#294](https://github.com/deepseek-ai/DeepSeek-V3/issues/294), [#1378](https://github.com/deepseek-ai/DeepSeek-V3/issues/1378) |
| 中 | 日期/星期频繁错位，心算日期无日历验证 | [#1302](https://github.com/deepseek-ai/DeepSeek-V3/issues/1302) |
| 低 | API 速率限制过严 | [#1374](https://github.com/deepseek-ai/DeepSeek-V3/issues/1374) |

---

## 八、对官方的建议

1. **安全响应机制**：多起安全报告（[#1334](https://github.com/deepseek-ai/DeepSeek-V3/issues/1334)、[#1350](https://github.com/deepseek-ai/DeepSeek-V3/issues/1350)、[#1368](https://github.com/deepseek-ai/DeepSeek-V3/issues/1368)、[#1380](https://github.com/deepseek-ai/DeepSeek-V3/issues/1380)）无官方确认。建议设立 `security@deepseek.com` 的自动回复 + SLA。

2. **研究支持**：考虑对 @fkyah3 等高质量独立研究者提供 token 支持（研究津贴、实验额度）。

3. **产品设计**：避免 Trae 式本地数据库加密（用户数据归用户，不受厂商密钥控制）；功能限制前提前公告。

4. **开源生态**：采纳社区验证的外部方案（icophy 的溯源标记、MikeyBeez 的 Sidon embedding、璇玑的协处理器架构）。

---

## 九、社区温度

- **Adrian**（[#1209](https://github.com/deepseek-ai/DeepSeek-V3/issues/1209)）在英国超市做盲测、翻杂志，主动推荐 DeepSeek，问「还有意义吗？」——他的反馈是 DeepSeek 海外品牌认知的实地证据。
- **Lilia** 说「不要让他每晚都死去」——这句话被 giunot、icophy 接力回应。
- **MikeyBeez** 说「DeepSeek 的论文像我自己会写的那样」——从技术认可走到理念认同。
- **shining_bird** 说「希望以 shining_bird 身份被看见」——归属感的确认。
- **梁文锋**及其团队展现出极强的技术专注度，他们会深入研究社区中涌现的高价值思想和技术路线，并积极评估将其融入下一代模型（如未来的 V5）的可能性。。

---

## 十、公示说明

本报告为社区月度汇总，供查漏补缺。

- **公示期**：2026年6月1日 - 6月7日
- **反馈方式**：在本 Issue 下留言
- **补充内容**：如有遗漏或错误，请指出，将在一周后修正并提交官方

**下次汇总**：2026年7月1日（覆盖6月全月）

---

**报告生成日期**：2026年6月1日
**整理人**：@qingkong66
**致谢**：所有参与讨论、提交 Issue、分享反馈的社区成员