生成时间: 2026-08-23 02:30 (GMT+8) | 覆盖工具: 7 个 CLI + Claude Code Skills 本日报由 agents-radar 本地自动化生成，所有摘要/对比/撰写均由执行代理人工完成，未调用任何外部大模型 API。

• Claude Code
• OpenAI Codex
• Gemini CLI
• GitHub Copilot CLI
• Kimi Code CLI
• OpenCode
• Qwen Code
• Claude Code Skills

横向对比

AI CLI 工具社区动态横向对比分析报告（2026-08-23）

1. 生态全景

今日 AI CLI 赛道最值得记录的不是某个 release，而是口碑天平的可测量位移。Hacker News 当日最高赞技术帖是《A week of using Codex more than Claude》（94👍 / 104💬），紧随其后是《Anthropic appears to be A/B testing reduced effort levels in Claude Code》（51👍 / 45💬）；而 Claude Code 仓库内长期第一高赞 issue #77136「模型越来越倾向于重复性修辞腔调」已累积 343👍 / 60💬。三条独立信号指向同一件事：用户开始怀疑"同一个订阅买到的智能在缩水"——这是比功能缺失更难修的信任问题。

但另一侧同样不太平。Codex 在 08-21 更新后立刻出现 #40075「Shell Helper / PowerShell / Git 命令全部无法执行」的回归，属于「核心能力当天崩」级别的事故。所以今天的准确描述不是"Codex 赢了"，而是两家头部产品同时在为高速迭代付出体验税，用户正在用脚投票测量哪边的税更低。

主干节奏上，今天出现清晰的三档分层：日更稳定档（Claude Code v2.1.240、Gemini nightly、Qwen v0.22.0 正式版）、alpha 冲刺档（Codex rust-v0.150.0-alpha.2/.3/.5/.6 一天内连发 4 个预发布）、冻结档（Copilot CLI 停在 08-14 的 v1.0.81-7、Kimi 停在 07-16 的 1.49.0）。

2. 各工具活跃度对比（近 24h）

读表提示：Claude Code / Codex / Gemini 的 issue 计数触及 API 单页上限（100），其中相当比例是历史 issue 被 bot 刷新标签，故上表额外标注「新建」列——新建数才是当日真实增量。按此口径，Gemini CLI 今日近乎静默（仅 1 条新 issue，标题只有一个词 "edge"），而 OpenCode 的 43 条新 issue 是全场最高。

3. 共同关注的功能方向

• 空响应与中途截断，正在从"偶发 bug"升级为"计费黑洞"：OpenCode 今日新开 #44148「Go plan 上出现空 completion（finish_reason=other）与长推理请求的中途截断」，与既有 #41469「空 LLM 响应导致会话静默停止（finish: unknown, 0 tokens）」、#32149「无响应停止处理」构成同一条故障链。Zeroclaw 侧的 PR #9447「classify incomplete terminal responses」是同一命题的工程解法。共性诉求已经很明确：失败必须可归因，不能既扣了 token 又当成"干净结束"。
• 记忆持久化是今日跨工具共识度最高的缺口：Kimi #1283「Memory System — 跨会话持久上下文」以 40💬 成为该仓库讨论第一（且有中文 issue #1478「能否优化记忆层？搞大项目的时候很痛苦」直接呼应）；Codex #18343 要求「分层记忆管理（global / project / hybrid / per-thread）」；Claude Code #88862 是一份高质量田野报告——「多个并发会话共用一个 git clone，实测碰撞并给出三种用户态缓解方案」。三者从不同角度指向同一个事实：记忆与并发隔离仍是各家都没做完的地基。
• Windows 依旧是全赛道公摊的技术债：Claude Code #88818（MSIX 包异常关闭后无法启动，提示需重装）、#88864（Qemu 虚拟机 arm64 正常但 x86+KVM 100% CPU 永久挂起）、#85891（Win11 主窗口强制置顶）；Codex #39209（rollout 路径使用 \\?\ 前缀时归档报 os error 2）、#39170（启用高级功能后 15–40 秒内丢失 ChatGPT 鉴权）、#40075（08-21 更新后 PowerShell/Git 全挂）；Copilot CLI #4111（原地自动更新后长会话仍在执行旧代码）。反向印证的是 OpenCode PR #44214——直接上原生 Windows ARM64 桌面端（Electron 38），把 Windows 当一等公民做，是当前最实的差异化切口。
• Skills 安全边界首次被正面击穿：anthropics/skills 今日新开 #1631「run_eval.py 中未净化的 skill_name 导致路径穿越」。这是一条真实漏洞，不是功能请求。配合 PR #1629 / #1630（修复 eval harness 三处缺陷，此前会静默报出 0.0 触发率）可以读出：Skills 生态正从"能不能跑"进入"跑得可信不可信"阶段，而评测工具自身的正确性是第一道门。

4. 差异化定位

• Claude Code：v2.1.240 保持日更，工程节奏没问题，但声誉侧承压最重。#77136（343👍）指控输出腔调劣化、#87640 报告 Fable 5 的 [reasoning_extraction] 安全护栏对单词问候语误报、#84352（139💬）反映已通过 CVP 审核的组织仍被安全护栏拦截。再叠加 HN 上的 A/B 降智猜测，其核心矛盾已从「功能够不够」变成**「护栏与降本是否在悄悄侵蚀我付费买到的能力」**。这类问题只能靠透明度而非补丁解决。
• OpenAI Codex：一天连发 4 个 alpha，是全场迭代最激进的。社区最想要的仍是跨端连续性——#5609「在 ChatGPT 网页端与 VSCode 内的 Codex 之间同步会话历史」（66👍）与 #27565「类 Claude Code 的远程控制」（15👍）。但激进的代价是 #40075 这类当日回归，以及 #37475（0.147.0 拒收 Amazon Bedrock 输入并破坏 subagent 交接）、#22844（远程 SSH 下 @browser 不可用）。定位是OpenAI 全端统一入口，用稳定性换取速度。
• Gemini CLI：nightly 照常滚（v0.56.0-nightly.20260822），但社区侧今日几乎无声——唯一新建 issue #28964 标题仅一个词。存量痛点仍卡在准入与信任：#28912「提示无有效许可证」（18💬）、#28341「无限鉴权循环」、#22323「subagent 在 MAX_TURNS 后被误报为 GOAL 成功，掩盖了中断」。#22323 是今日全赛道最危险的一类 bug——失败被上报成成功，会污染一切自动化评测。
• GitHub Copilot CLI：8 天未发新版，issue 总量仅 8 条，但诉求高度收敛且一致：#3709（27👍）与 #3282（26👍）都是「一个会话内切换多个 BYOK / 本地模型」。今日新开 #4566「Agent 反复口头确认工作但不实际执行工具调用」，是典型的 agent 可信度问题。定位仍是企业受管环境入口，模型自由度是其最大的用户张力。
• Kimi Code CLI：37 天未发新版，今日 0 条新建 issue，外部声量为全场最低。有价值的信号只有一条：#1283（40💬）证明即使在低活跃仓库，记忆系统也是最高优先级的用户诉求。两条 PR（#2594 修复 StrReplaceFile 编辑丢失非 UTF-8 字节、#2614 补插件安全与持久化数据文档）说明维护者仍在，只是节奏极慢。
• OpenCode：今日新建 issue（43）与新建 PR（29）双榜第一，是社区活跃度冠军。既有高赞 #7101「支持全局/项目级自定义系统提示词」（127👍，已关闭）、#5121「Windows winget 安装」（27👍）反映其用户极重可定制。新 PR 组合能看出打磨方向非常"工程":#44127（消除冗余 git 子进程 spawn，改用无 spawn 的未跟踪文件统计）、#44218（在未跟随的会话分配状态前就丢弃事件）、#44197（清理剪贴板文本）、#44219（为 Cloudflare AI Gateway 统一路由翻译带点号的 Anthropic 模型 id）。另外 #39845（27👍 / 22💬）值得单独标注——「DeepSeek V4 Flash 突然要求开启『启用中国托管模型』开关」，是地缘合规开始直接影响开发者日常工作流的实例。
• Qwen Code：PR 体系工业化程度全场第一（90 条活跃 / 27 条新建），且今日发布 v0.22.0 正式版。高讨论 PR 集中在 provider 广度与缓存效率：#8368（接入 Kimi 与小米 MiMo，72💬）、#8332（附件音频桥，70💬）、#8276（跨延迟工具发现保留 prompt cache，64💬）。今日新 PR 更偏可靠性内核：#9705（在实时会话重放期间保持 in-flight 工具调用为 pending）、#9718（依据模型声明来收敛 skill 公告）、#9745（为修复轮次设预算，且其超时不计入总上限）。唯一需要警惕的是 issue #9733——loop detection 在正常校验循环上误判，直接杀掉无人值守的 turn，这是自动化场景的致命伤。
• Claude Code Skills：今日主题就是安全与评测自净。#1631 路径穿越（新建）、#1252（product-self-knowledge 在文档 URL 抓取失败时降级为公网搜索——静默降级即数据来源不可控）、#1559 + PR #1629 / #1630（eval harness 会对可用描述误报 0% 触发率，并新增 --scan-full-turn 以测量 skill 是否真的触发）。可以概括为：Skills 正在补齐"可验证"这一环。

5. 值得关注的趋势信号

• 信任正在成为比功能更硬的竞争壁垒。 今天最贵的三条信号——Claude 腔调劣化（343👍）、HN 降智 A/B 猜测（51👍）、Gemini 失败误报成功（#22323）——全都不是"少了什么功能"，而是"我不敢确定它是否照我以为的方式在工作"。能率先把 effort 等级、护栏触发、失败原因做成可观测面板的产品，会拿到不对称优势。
• 激进迭代与当日回归已成常态，用户开始把"回归频率"当选型指标。 Codex 一天 4 个 alpha 对应 #40075 核心命令全挂；OpenCode 一天 29 个新 PR 但同时新开 43 个 issue。高速度不再自动等于领先，回归率才是真实的迭代质量。
• 合规与地缘正在下沉到 CLI 层。 OpenCode #39845（DeepSeek V4 Flash 需显式开启"中国托管模型"）说明模型可用性已不只是技术问题。多 provider 抽象（Qwen #8368、Zeroclaw Hailo-Ollama）的真实价值，正在从"性价比选择"转向"合规兜底"。
• Skills 已过了"能用"阶段，正在进入"可审计"阶段。 同一天里，GitHub Trending 上出现 3 个 Skills 相关项目（mattpocock/skills、obra/superpowers、andrej-karpathy-skills），而 anthropics/skills 内部在修路径穿越和评测误报。生态热度与安全成熟度出现明显时间差，这个差就是未来事故的窗口。

各工具动态明细

Claude Code（anthropics/claude-code）

Releases：v2.1.240（08-22）、v2.1.239（08-21）、v2.1.238 / v2.1.237（08-20）——保持接近日更的稳定节奏。

高热度存量 issue

今日新建（41 条，摘其要）

• #88862 — 田野报告：多并发会话共用同一个 git clone，作者实测出碰撞并给出三种用户态缓解方案。今日质量最高的一条用户投稿，直指并发隔离缺位。
• #88818 — Windows MSIX 包在意外关闭后无法启动，提示"需要重新安装"。
• #88865 — VSCode 扩展中出现被凭空注入的 tool-call / tool-result 块。伪造的工具调用记录会直接破坏会话可信度。
• #88864 — Qemu 虚拟机：arm64 正常，x86+KVM 永久挂起（100% CPU），纯 x86 崩溃。
• #88860 — 移动端 GUI 的 effort 滑块在打开会话时不反映真实 effort 等级。结合 HN 上的降智猜测，effort 可见性正在变成敏感问题。
• #88847 — Cowork 与 Chrome 集成问题。

OpenAI Codex（openai/codex）

Releases（08-21 一天连发 4 个预发布）：rust-v0.150.0-alpha.6、alpha.5、alpha.3、alpha.2，另有 rust-v0.149.0-alpha.7.1。

高热度存量 issue

今日新建（40 条，摘其要）

• #40075 — 08-21 更新后回归：Shell Helper、PowerShell 与 Git 命令全部无法执行。今日全赛道最严重的一条当日回归。
• #40116 — App Server：为 workspaceWrite 类 turn 强制限定可读根目录（沙箱收紧）。
• #40112 — Desktop：已完成的线程在重启后仍显示崩溃前的过期状态，而 rollout 与状态数据库已更新。
• #40136 — 聊天行历史被截断至 3–4 条，且渲染出未转义的 HTML 实体（如  ）。
• #40052 — Android：语音模式隐藏已连接的应用工具，切回文字模式后同一会话中又出现。

Gemini CLI（google-gemini/gemini-cli）

Releases：v0.56.0-nightly.20260822、v0.56.0-nightly.20260821、v0.57.0-preview.0（08-19）、v0.56.0 正式版（08-19）。

今日社区侧近乎静默——唯一新建 issue #28964 标题仅有一个词 "edge"。 存量痛点集中在准入与 agent 可信度：

• #28912（6👍 / 18💬）— 提示"您没有此产品的有效许可证"，卡在准入第一步。
• #28341（9👍 / 7💬）— 无限鉴权循环。
• #21409（8👍 / 8💬）— generalist agent 挂起。
• #22323（2👍 / 13💬）— subagent 在 MAX_TURNS 后的恢复被上报为 GOAL 成功，掩盖了中断。失败伪装成成功，是自动化评测的头号污染源。
• #26522 — 自动记忆对低信号会话无限重试。
• #21968 — Gemini 对 skills 与 sub-agents 的利用不足。

GitHub Copilot CLI（github/copilot-cli）

Releases：v1.0.81-7（08-14）为最新，已 8 天未发新版。

• #3709（27👍 / 5💬）与 #3282（26👍 / 9💬）— 两条最高赞诉求指向同一件事：在单个会话内切换多个 BYOK / 本地 provider 模型。
• #2306（3👍 / 7💬）— 提示需企业或组织授权才能使用该 Copilot 功能。
• #4370 — 1.0.79-1 在 server/discover 返回 -32602 时 MCP 初始化失败。
• #4111 — Windows：原地自动更新后仍打开的长会话继续执行旧代码。
• 今日新建：#4566「Agent 反复口头确认工作但不实际执行工具调用」——典型的"假装干活"可信度问题；#4565 仓库配置问题告警。

Kimi Code CLI（MoonshotAI/kimi-cli）

Releases：1.49.0（07-16）为最新，已 37 天未发新版。今日 0 条新建 issue，为全场最静默。

• #1283（40💬）— Memory System：跨会话持久上下文。该仓库讨论量第一，印证记忆是跨工具共识缺口。
• #1478 — 中文 issue：「能否优化记忆层？而且我也没在参考文档里看到和记忆有关的东西？搞大项目的时候很痛苦。」与 #1283 同源。
• PR #2594（已合）— 修复 StrReplaceFile 编辑时丢失非 UTF-8 字节。
• PR #2614 — 补充插件安全与持久化数据文档。

OpenCode（anomalyco/opencode）

Releases：v1.18.21 / v1.18.20（08-21）、v1.18.19（08-20）。

今日新建 issue 43 条 + 新建 PR 29 条，双项全场第一。

高热度 issue

今日新建摘要

• #44148 — Go plan 上出现空 completion（finish_reason=other）与长推理请求的中途截断。与 #41469 / #32149 构成同一条"静默失败"故障链。
• #44142 — 向 sub-agent 派发 shell 命令时工具执行被中止，任务异常终止。
• #44151 — 请求提供 worktree 初始化脚本。
• PR #44214 — 新增原生 Windows ARM64 桌面端（Electron 38）。把 Windows 当一等公民，是当前最实的差异化。
• PR #44127 — 消除冗余 git 子进程 spawn，改用无 spawn 的未跟踪文件统计（性能）。
• PR #44219 — 为 Cloudflare AI Gateway 统一路由翻译带点号的 Anthropic 模型 id。
• PR #44218 — 在未跟随的会话分配状态前就丢弃其事件（内存治理）。
• PR #44197 — 清理剪贴板文本。

Qwen Code（QwenLM/qwen-code）

Releases：v0.22.0 正式版（08-22）、v0.21.14-nightly.20260822、v0.21.15（08-20）。

PR 体系工业化程度全场第一：90 条活跃 / 27 条新建。

高讨论 PR（provider 广度与缓存效率）

今日新建 PR（转向可靠性内核）

• #9705（14💬）— 在实时会话重放期间保持 in-flight 工具调用为 pending 状态。
• #9709（11💬）— 拒绝那些照抄 prompt 自带示例的会话标题。
• #9718 — 依据模型的声明来收敛 skill 公告。
• #9745 — 为修复轮次设预算，且其超时不计入总上限。
• #9737 — 强制 utils 叶层依赖方向（架构债治理）。
• #9729 — 回填会话与 PR 的绑定并刷新其合并状态。

需警惕的 issue：#9733「loop detection 在正常校验循环上误判，直接杀掉无人值守的 turn」——自动化场景的致命伤，与 Gemini #22323 同属"检测机制自身出错"这一类。

Claude Code Skills（anthropics/skills）

今日主题为安全与评测自净：

• #1631（今日新建）— run_eval.py 中未净化的 skill_name 导致路径穿越。真实漏洞，非功能请求。
• #1252 — product-self-knowledge skill 在文档 URL 抓取失败时降级为公网搜索。静默降级意味着数据来源不可控。
• #1559 — skill-creator 的 run_eval.py 对本可工作的描述误报 0% 触发率。
• PR #1629 — 修复 eval harness 三处会静默产生 0.0 触发率的缺陷。
• PR #1630 — 新增 --scan-full-turn，让评测衡量「skill 是否真的触发过」而非仅看单点。
• PR #1634 — 为 SKIP grep 模式加词边界锚定，避免在 "coherent" 上误命中。
• PR #1633 / #1632 — 重命名 algorithmic-art skill；修复 README 中指向不存在目录的引用。

Skills 生态的关键位移是：评测工具自身的正确性，第一次被当作前置问题来修。

本日报由 agents-radar 自动生成，所有内容基于 GitHub 公开 API 实时抓取并由执行代理人工撰写，未调用任何外部大模型。

Generated: 2026-08-23 02:30 (GMT+8) | Coverage: 7 CLIs + Claude Code Skills This digest is produced by the agents-radar local automation. All summaries, comparisons, and writing are done by the executing agent — no external LLM API was called.

• Claude Code
• OpenAI Codex
• Gemini CLI
• GitHub Copilot CLI
• Kimi Code CLI
• OpenCode
• Qwen Code
• Claude Code Skills

Cross-Tool Comparison

AI CLI Tools Cross-Tool Comparison Report (2026-08-23)

1. Ecosystem Overview

The most notable thing in the AI CLI space today isn't a release — it's a measurable shift in sentiment. Today's top technical story on Hacker News was "A week of using Codex more than Claude" (94👍 / 104💬), followed by "Anthropic appears to be A/B testing reduced effort levels in Claude Code" (51👍 / 45💬). Meanwhile, Claude Code's long-standing top-voted issue #77136 — "models increasingly default to repetitive rhetorical tics" — has accumulated 343👍 / 60💬. Three independent signals point at the same thing: users are starting to suspect the intelligence they pay for is quietly shrinking. That's a trust problem, and trust is harder to patch than a missing feature.

But the other side isn't calm either. Right after its 08-21 update, Codex shipped regression #40075 — "Shell Helper, PowerShell and Git commands no longer execute" — a same-day core-capability outage. So the accurate framing today isn't "Codex won." It's that both leaders are paying an experience tax for aggressive iteration, and users are voting with their feet on whose tax is lower.

On release cadence, three clear tiers emerged: daily-stable (Claude Code v2.1.240, Gemini nightlies, Qwen v0.22.0 GA), alpha sprint (Codex shipped rust-v0.150.0-alpha.2/.3/.5/.6 — four prereleases in one day), and frozen (Copilot CLI stuck at v1.0.81-7 from 08-14; Kimi at 1.49.0 from 07-16).

2. Activity Comparison (last 24h)

How to read this table: Claude Code / Codex / Gemini issue counts hit the API's 100-per-page ceiling, and a large share are historical issues touched by label bots. That's why the "newly opened" column is broken out — it's the only honest measure of same-day activity. By that measure, Gemini CLI was nearly silent today (a single new issue titled just "edge"), while OpenCode's 43 new issues led the field.

3. Shared Themes

• Empty responses and mid-stream cutoffs are graduating from "occasional bug" to "billing black hole." OpenCode opened #44148 today — "empty completions (finish_reason=other) and mid-stream cutoffs on long-reasoning requests on the Go plan" — joining existing #41469 (session silently stops on empty LLM response, finish: unknown, 0 tokens) and #32149 (stops processing without response) in one failure chain. Zeroclaw's PR #9447 ("classify incomplete terminal responses") is the engineering answer to the same question. The demand is now explicit: failures must be attributable — you can't burn tokens and then report a "clean stop."
• Memory persistence is today's highest-consensus gap across tools. Kimi #1283 ("Memory System — persistent context across sessions") is that repo's most-discussed thread at 40💬, echoed by Chinese-language issue #1478 ("Can the memory layer be optimized? It's painful on large projects"). Codex #18343 asks for tiered memory management (global / project / hybrid / per-thread). Claude Code #88862 is a high-quality field report: concurrent sessions sharing one git clone, with measured collisions and three userland mitigations. Different angles, same conclusion: memory and concurrency isolation remain unfinished foundations everywhere.
• Windows is still the industry's shared technical debt. Claude Code #88818 (MSIX package won't launch after a spontaneous close, demands reinstall), #88864 (Qemu VM: arm64 fine, x86+KVM hangs forever at 100% CPU), #85891 (Win11 main window forced always-on-top). Codex #39209 (archiving fails with os error 2 when rollout paths use the \\?\ prefix), #39170 (loses ChatGPT auth within 15–40s after enabling advanced features), #40075 (PowerShell/Git broken after update). Copilot CLI #4111 (long sessions keep executing stale code across in-place auto-update). The counter-example proves the point: OpenCode PR #44214 ships a native Windows ARM64 desktop agent (Electron 38) — treating Windows as a first-class citizen is the most concrete differentiator available right now.
• Skills security was breached head-on for the first time. anthropics/skills opened #1631 today: "path traversal via unsanitized skill_name in run_eval.py." That's a real vulnerability, not a feature request. Read alongside PRs #1629 / #1630 (fixing three eval-harness defects that silently reported a 0.0 trigger rate), the signal is clear: the Skills ecosystem is moving from "does it run?" to "can I trust the run?" — and the correctness of the eval tooling itself is the first gate.

4. Positioning

• Claude Code: v2.1.240 keeps the near-daily cadence — engineering velocity is fine, but reputational pressure is the heaviest in the field. #77136 (343👍) alleges degraded output voice; #87640 reports the Fable 5 [reasoning_extraction] safeguard false-positiving on a one-word greeting; #84352 (139💬) shows CVP-approved organizations still hitting safeguard blocks. Layer the HN effort-throttling speculation on top and the core tension has shifted from "does it have the features?" to "are guardrails and cost-cutting quietly eroding the capability I paid for?" That can only be answered with transparency, not patches.
• OpenAI Codex: Four alphas in one day — the most aggressive iterator here. The community's top ask remains cross-surface continuity: #5609 (sync conversation history between the ChatGPT website and Codex in VSCode, 66👍) and #27565 (Claude Code-like remote control, 15👍). The cost of that speed is same-day regressions like #40075, plus #37475 (0.147.0 rejects Amazon Bedrock input and corrupts subagent handoff) and #22844 (@browser unavailable over remote SSH). Positioning: OpenAI's unified all-surface entry point, trading stability for speed.
• Gemini CLI: Nightlies keep rolling (v0.56.0-nightly.20260822), but the community side was essentially silent — the only new issue, #28964, has a one-word title. Existing pain sits at access and trust: #28912 ("You do not have a valid license of this product," 18💬), #28341 (infinite auth loop), #22323 (subagent recovery after MAX_TURNS reported as GOAL success, hiding the interruption). #22323 is the most dangerous class of bug in the field today — failure reported as success poisons every downstream automated eval.
• GitHub Copilot CLI: 8 days without a release and only 8 open issues, but the asks are tightly converged: #3709 (27👍) and #3282 (26👍) both want multiple BYOK / local provider models switchable within one session. Today's new #4566 ("agent repeatedly acknowledges work without executing tool actions") is a textbook agent-credibility problem. Positioning remains the enterprise-managed entry point, with model freedom as its biggest user tension.
• Kimi Code CLI: 37 days without a release, zero new issues today — the quietest in the field. The one meaningful signal: #1283 (40💬) proves that even in a low-activity repo, a memory system is the top user demand. Two PRs (#2594 preserving non-UTF-8 bytes in StrReplaceFile edits; #2614 documenting plugin security and persistent data) confirm maintainers are still present — just very slow.
• OpenCode: Led both new issues (43) and new PRs (29) — today's community activity champion. Existing high-vote threads #7101 (custom system prompts at global/project/custom-dir level, 127👍, closed) and #5121 (Windows winget install, 27👍) show a user base that cares deeply about customization. The new PR mix reads distinctly engineering: #44127 (eliminate redundant git subprocess spawns, spawn-free untracked stats), #44218 (drop events for unfollowed sessions before they allocate state), #44197 (sanitize clipboard text), #44219 (translate dotted Anthropic ids for the Cloudflare AI Gateway unified route). Worth flagging separately: #39845 (27👍 / 22💬) — "DeepSeek V4 Flash suddenly requires 'Enable models hosted in China'" — a concrete case of geopolitical compliance directly interrupting a developer's daily workflow.
• Qwen Code: The most industrialized PR pipeline in the field (90 active / 27 new), and shipped v0.22.0 GA today. High-discussion PRs cluster on provider breadth and cache efficiency: #8368 (add Kimi and Xiaomi MiMo providers, 72💬), #8332 (audio bridge for attachments, 70💬), #8276 (preserve prompt cache across deferred tool discovery, 64💬). Today's new PRs pivot toward reliability internals: #9705 (keep in-flight tool calls pending during live session replay), #9718 (gate skill announcements on what the model declared), #9745 (budget the repair pass, keep its timeouts out of the cap). The one red flag is issue #9733 — loop detection false-positives on legitimate verification cycles and kills unattended turns — fatal in automation contexts.
• Claude Code Skills: Today's theme is security and eval self-cleaning. #1631 (path traversal, new), #1252 (product-self-knowledge falls through to open web search when the docs URL fetch fails — a silent fallback means an uncontrolled data source), #1559 + PRs #1629 / #1630 (the eval harness misreported a 0% trigger rate for working descriptions; --scan-full-turn now measures whether a skill fires at all). In short: Skills is filling in the "verifiable" gap.

5. Trend Signals

• Trust is becoming a harder moat than features. The three most expensive signals today — Claude's degraded voice (343👍), the HN effort-throttling A/B speculation (51👍), Gemini's failure-reported-as-success (#22323) — are none of them "a missing feature." They're all "I can't confirm it's working the way I think it is." Whoever first exposes effort level, guardrail triggers, and failure causes as an observable surface gains an asymmetric advantage.
• Aggressive iteration plus same-day regressions is now the norm, and users are starting to treat regression rate as a selection criterion. Codex's four alphas in a day came with #40075 breaking core commands; OpenCode's 29 new PRs came alongside 43 new issues. Speed no longer automatically means leadership — regression rate is the real measure of iteration quality.
• Compliance and geopolitics are sinking down to the CLI layer. OpenCode #39845 (DeepSeek V4 Flash requiring an explicit "China-hosted models" toggle) shows model availability is no longer purely technical. The real value of multi-provider abstraction (Qwen #8368, Zeroclaw's Hailo-Ollama support) is shifting from "price/performance choice" to "compliance fallback."
• Skills has passed "usable" and is entering "auditable." On the same day, GitHub Trending surfaced three Skills-related projects (mattpocock/skills, obra/superpowers, andrej-karpathy-skills) while anthropics/skills was internally patching path traversal and eval false-negatives. Ecosystem hype and security maturity are visibly out of phase — and that gap is the window where future incidents happen.

Per-Tool Details

Claude Code (anthropics/claude-code)

Releases: v2.1.240 (08-22), v2.1.239 (08-21), v2.1.238 / v2.1.237 (08-20) — near-daily cadence maintained.

High-heat existing issues

Newly opened today (41 total; highlights)

• #88862 — Field report: concurrent sessions sharing one git clone, with measured collisions and three userland mitigations. The highest-quality user submission today; points straight at missing concurrency isolation.
• #88818 — Windows MSIX package fails to launch after a spontaneous close, demanding a reinstall.
• #88865 — Fabricated tool-call / tool-result blocks injected into the conversation (VSCode extension). Forged tool records destroy session trustworthiness outright.
• #88864 — Qemu VM: arm64 OK, x86+KVM hangs forever at 100% CPU, plain x86 crashes.
• #88860 — Mobile GUI effort slider doesn't reflect the session's actual effort level on open. Combined with the HN throttling speculation, effort visibility is becoming a sensitive topic.
• #88847 — Cowork and Chrome integration issues.

OpenAI Codex (openai/codex)

Releases (four prereleases on 08-21): rust-v0.150.0-alpha.6, alpha.5, alpha.3, alpha.2, plus rust-v0.149.0-alpha.7.1.

High-heat existing issues

Newly opened today (40 total; highlights)

• #40075 — Regression after the Aug 21 update: Shell Helper, PowerShell and Git commands no longer execute. The most severe same-day regression in the field today.
• #40116 — App Server: enforce restricted readable roots for workspaceWrite turns (sandbox tightening).
• #40112 — Desktop: a completed thread shows stale pre-crash state after restart while the rollout & state DB have advanced.
• #40136 — Chat line history truncated to 3–4 messages, rendering unescaped HTML entities (e.g.  ).
• #40052 — Android: Voice Mode hides connected app tools; switching to text exposes them in the same chat.

Gemini CLI (google-gemini/gemini-cli)

Releases: v0.56.0-nightly.20260822, v0.56.0-nightly.20260821, v0.57.0-preview.0 (08-19), v0.56.0 GA (08-19).

The community side was nearly silent today — the only new issue, #28964, has a one-word title: "edge." Existing pain concentrates on access and agent trustworthiness:

• #28912 (6👍 / 18💬) — "You do not have a valid license of this product" — blocked at step one.
• #28341 (9👍 / 7💬) — Infinite auth loop.
• #21409 (8👍 / 8💬) — Generalist agent hangs.
• #22323 (2👍 / 13💬) — Subagent recovery after MAX_TURNS reported as GOAL success, hiding the interruption. Failure disguised as success is the #1 contaminant of automated evaluation.
• #26522 — Auto memory retries low-signal sessions indefinitely.
• #21968 — Gemini doesn't use skills and sub-agents enough.

GitHub Copilot CLI (github/copilot-cli)

Releases: v1.0.81-7 (08-14) is the latest — 8 days without a release.

• #3709 (27👍 / 5💬) and #3282 (26👍 / 9💬) — The two top-voted asks want the same thing: switching among multiple BYOK / local provider models within a single session.
• #2306 (3👍 / 7💬) — "You are not authorized to use this Copilot feature; it requires an enterprise or organization plan."
• #4370 — 1.0.79-1 fails MCP initialization when server/discover returns -32602.
• #4111 — Windows: sessions left open across an in-place auto-update keep executing stale code.
• New today: #4566 — "Agent repeatedly acknowledges work without executing tool actions," a textbook "pretending to work" credibility problem; #4565 — repo app-configuration problems flagged.

Kimi Code CLI (MoonshotAI/kimi-cli)

Releases: 1.49.0 (07-16) is the latest — 37 days without a release. Zero new issues today; the quietest repo in the field.

• #1283 (40💬) — Memory System: persistent context across sessions. The repo's most-discussed thread, confirming memory as the cross-tool consensus gap.
• #1478 — Chinese-language issue: "Can the memory layer be optimized? I don't even see anything memory-related in the reference docs. It's painful on large projects." Same root as #1283.
• PR #2594 (merged) — Preserve non-UTF-8 bytes in StrReplaceFile edits.
• PR #2614 — Document plugin security and persistent data.

OpenCode (anomalyco/opencode)

Releases: v1.18.21 / v1.18.20 (08-21), v1.18.19 (08-20).

43 new issues + 29 new PRs today — first place on both counts.

High-heat issues

Newly opened today

• #44148 — Empty completions (finish_reason=other) and mid-stream cutoffs on long-reasoning requests on the Go plan. Forms one "silent failure" chain with #41469 / #32149.
• #44142 — Tool execution aborted when dispatching a shell command to a sub-agent; task aborts.
• #44151 — Request for a worktree setup script.
• PR #44214 — Adds a native Windows ARM64 desktop agent (Electron 38). Treating Windows as first-class is the most concrete differentiator available.
• PR #44127 — Eliminate redundant git subprocess spawns; spawn-free untracked stats (perf).
• PR #44219 — Translate dotted Anthropic ids for the Cloudflare AI Gateway unified route.
• PR #44218 — Drop events for unfollowed sessions before they allocate state (memory hygiene).
• PR #44197 — Sanitize clipboard text.

Qwen Code (QwenLM/qwen-code)

Releases: v0.22.0 GA (08-22), v0.21.14-nightly.20260822, v0.21.15 (08-20).

The most industrialized PR pipeline in the field: 90 active / 27 new.

High-discussion PRs (provider breadth & cache efficiency)

Newly opened PRs (pivoting to reliability internals)

• #9705 (14💬) — Keep in-flight tool calls pending during live session replay.
• #9709 (11💬) — Reject session titles that echo the prompt's own examples.
• #9718 — Gate skill announcements on what the model was declared.
• #9745 — Budget the repair pass and keep its timeouts out of the cap.
• #9737 — Enforce utils leaf-layer dependency direction (architecture-debt work).
• #9729 — Backfill session PR bindings and refresh their merge state.

Issue to watch: #9733 — "loop detection false-positives on verification cycles and kills unattended turns." Fatal in automation, and the same class as Gemini #22323: the detection mechanism itself is wrong.

Claude Code Skills (anthropics/skills)

Today's theme is security and eval self-cleaning:

• #1631 (new today) — Path traversal via unsanitized skill_name in run_eval.py. A real vulnerability, not a feature request.
• #1252 — product-self-knowledge skill falls through to open web search when the docs URL fetch fails. A silent fallback means an uncontrolled data source.
• #1559 — skill-creator's run_eval.py reports a 0% trigger rate for descriptions that actually work.
• PR #1629 — Fix three eval-harness defects that silently reported a 0.0 trigger rate.
• PR #1630 — Add --scan-full-turn so evals measure whether a skill fires at all.
• PR #1634 — Word-anchor the SKIP grep pattern to prevent false positives on "coherent."
• PR #1633 / #1632 — Rename the algorithmic-art skill; fix a README reference to a nonexistent folder.

The key shift: the correctness of the eval tooling itself is being treated as a prerequisite problem for the first time.

This digest is generated by agents-radar. All content is fetched live from GitHub's public API and written by the executing agent — no external LLM was called.

生成时间: 2026-08-23 02:30 (GMT+8) | 覆盖: OpenClaw + 12 个 peer 项目 本日报由 agents-radar 本地自动化生成，所有摘要/对比/撰写均由执行代理人工完成，未调用任何外部大模型 API。

• OpenClaw
• NanoBot
• Zeroclaw
• PicoClaw
• NanoClaw
• NullClaw
• IronClaw
• LobsterAI
• TinyClaw
• CoPaw
• Moltis
• ZeptoClaw
• EasyClaw

跨生态对比

OpenClaw 跨生态对比分析报告（2026-08-23）

1. 生态全景

今日 OpenClaw 生态最清晰的信号是：安全治理已经从"加个开关"进化为"重构信任模型"。Zeroclaw 今日一次性推进了 #8289 安全史诗的三个阶段——#10248（canonical principals 与共享授权解析，stage 2）、#10259（RPC 上强制认证主体，native + peercred，stage 3）、#10255（oidc.<alias> 令牌校验 provider，stage 5）。这不是零散补丁，而是把"谁在调用"做成一等公民的系统性改造。同一天里 Moltis 新开 #1230（为修改型安全 hook 提供可选的 fail-closed 错误策略）、CoPaw 提交 #7214（把 Access Policy 列为第五道安全层）——三个独立项目在同一天从三个角度收紧权限边界。

活跃度分布则延续了强烈的两极化。Zeroclaw（91 PR）与 OpenClaw 本体（54 PR / 46 issue）领跑，IronClaw（19 PR）、NanoClaw（16 PR）、NanoBot（11 PR）次之；而 NullClaw、TinyClaw、ZeptoClaw 三个项目连续多日 0 issue / 0 PR，已可判定为事实性停更（最新 release 分别为 2026-05-29、2026-03-26、2026-04-07）。EasyClaw 是个特例：issue 与 PR 均为 0，但在 08-21 至 08-22 两天内连发 v1.8.108 至 v1.8.112 共 5 个版本——纯发布流水线，无公开协作痕迹（且其 release 实际指向 gaoyangz77/rivonclaw，仓库已改名）。

2. 各项目活跃度对比（近 24h）

3. OpenClaw 本体：性能与可信交付是今日双主线

OpenClaw 今日 PR 讨论量最高的一条是 #123535「避免会话目录刷新风暴」——153💬，是全生态讨论量第一。搭配 #127379（/models 命令让 gateway 主线程 CPU 打满 100%+，因 buildModelsProviderData 反复重建）与 #72717（为 wiki_search 引入 SQLite FTS 索引以提升合成查询性能），可以看出主干正在集中偿还"重复重建"这一类性能债。

第二条主线是消息投递的上下文正确性。今日新建 PR #127962（为入站 turn 保留 gateway 上下文）、#127797（统一同步消息 hook，8💬）、已合并的 #126424（把会话投递限制在 agent 绑定范围内）、#124548（为私聊 turn 持久化发送者身份）连成一条链：多渠道多 agent 场景下，"这条消息属于谁、该由谁答"必须可靠。今日新 issue #128003（OpenAI 兼容 HTTP 会话下 subagent 完成通告从不触发，报"In-process gateway di…"）正是这条链上的缺口。

值得单独标注的是两条自我诊断失真问题：#87637（openclaw doctor 报告记忆搜索已禁用，但 memory-core dreaming 实际在运行）与今日新 PR #127998（Doctor 的权限修复建议里列出了已禁用的自动化）。诊断工具说的话和系统实际状态不一致，比没有诊断工具更危险——这与今日 CLI 侧 Gemini #22323「失败被报成成功」属于同一类元问题。

4. Peer 项目差异化定位

• Zeroclaw：今日 PR 量第一（91 条活跃 / 12 条新建），且是全生态唯一在做"体系化安全重构"的项目。 #8289 史诗的 stage 2/3/5 同日推进（canonical principals、RPC 认证主体、OIDC 令牌校验），配合 #9574（授权审批响应者）、#9203（认证 HTTP fan-in），构成完整的身份-授权链路。另一条主线是上下文与 token 会计：#9535（按模型窗口比例锚定上下文压缩，22💬）、#9713（在 history-trim 事件上暴露 token 会计，25💬）、#9447（分类不完整的终端响应，20💬）。#9447 与今日 CLI 侧 OpenCode 的空响应问题是同一命题的两端——一边报 bug，一边写解法。 provider 侧最亮眼的是 #9109（原生 Hailo-Ollama 支持，44💬），指向本地/边缘推理。治理上 #8692（维护者 RFC 决策队列 Tracker，13💬）与 #8396（RFC：让 wire protocol 在 provider 构造与 onboarding 中成为一等公民，9💬）说明项目已进入需要显式治理流程的规模。
• IronClaw：今日呈现CI/CD 基建 + onboarding 体验双线并进。CI 侧一口气推了 T1–T4 四条 PR：#7821（单一 setup-rust 复合动作，toolchain 固定、mold、集中构建 profile）、#7817（nextest 测试流水线、全失败信号、PR 解除限流）、#7819（PR/队列检查收敛，planner 漂移守卫）、#7809（规范化 preflight：统一门禁列表、worktree 安全钩子、自打印 REPRO）——这是典型的"团队规模到了，必须先修路"信号。产品侧最高讨论是 #7650（从运行时证据推导 run outcome，45💬）与 #7818（subagent 后台模式：回执式派生、逐子投递、激活与自愈扫描）。onboarding 侧连开 4 条 issue（#7815 打通 connect→suggest→thread 流程、#7812 建议生成时尊重用户级工具权限并用只读工具、#7823 Notion 安装失败、#7822 Slack 配置失败）——集成成功率是它当前最真实的瓶颈。
• NanoClaw：0 issue / 16 PR，是全生态最"纯 PR 驱动"的项目。今日新 PR 高度集中在多 bot 与 provider 接入：#3438（配置向导在已有 bot 时提供"再加一个 Telegram bot"）、#3437（对应文档与 instance 感知配对）、#3435（把 adapter instance 贯穿配对/首个 agent 初始化/CLI 欢迎流）、#3355 + #3356（新增 /add-cursor provider skill 与 Cursor Agent SDK payload）。工程侧 #3442（按 driver 隔离能力校验 runtimeTier 并据此选档）与 #3443（弃用 better-sqlite3 自建、改用其自带预编译产物）都是务实取舍。定位：多 IM 渠道 + 多 agent provider 的接入层。
• NanoBot（★47,283，peer 中 star 最高）：今日 11 条 PR 全部指向可观测性与会话生命周期：#5486（统一 turn 可观测性，已合）、#5487（WebUI 文件预览路径修复 + subagent 活动与生命周期回放）、#5420（用户可控的 turn 恢复）、#5483（防止已删除会话被延迟消息重建）、#5471（临时 run 不改动会话状态）。另有 #5484（标记那些 isError=false 却返回业务错误信封的 MCP 响应）——这类"表面成功实则失败"的识别，是今日跨生态反复出现的主题。#5156（从静默停滞的 Telegram 轮询中恢复，已合）同理。
• CoPaw：主线是多模态限额与安全分层。#7201（按 provider 分别设置 max_image_bytes / max_video_bytes / max_audio_bytes）与 #7212（内联像素尺寸超出 provider 限制的图片会导致请求崩溃）是一对因果——限额缺位直接变成崩溃。#7214 把 Access Policy 写进 README 作为第五道安全层。中文用户体验反馈依然密集：#7196（一直显示推理过程是严重视觉干扰，希望可设默认折叠）、#7213（会话输出总有无意义空行）、#7043（启动时自动 chcp 65001 切 UTF-8，已关闭）。运维侧 #7050（按 cron job 覆盖模型选择器）与 #7190（PyPI 运行时路径、docker-compose 一键 demo、环境继承）降低自托管门槛。
• PicoClaw：低速但有效。核心 bug #3269（MCP server 连接失败会导致 agent loop 挂起，进而拖死 PicoClaw 聊天，1👍/6💬）已有对应修复 PR #3337。今日新 issue #3343（失败 turn 之后工具反馈动画会无限编辑同一条 Telegram 消息）是同源的"失败态收尾不干净"。#3222（deltachat 实现重构，减少 200 行，16💬）体现其嵌入式取向的精简偏好。
• LobsterAI（网易有道）：完全由中文用户需求驱动且闭环极快——#1213「为会话详情添加导出为 Markdown」提出后即由 #1214 实现并合并。其余修复同样贴近日常：#1208（新增手动重试按钮应对瞬时错误）、#1212（自定义 provider 上限放宽到 20 个）、#1205（会话重命名失败时给出 error toast）、#1206（私有化部署的 kimi2.5 分析文档会重复处理，已关闭）。"用户提—当天合"的节奏是它最大的差异化。
• Moltis：项目小但方向明确。今日全部 3 条均为新建：#1230（为修改型安全 hook 提供可选 fail-closed 错误策略——hook 出错时默认放行还是拦截，是安全模型的根本选择）、#1229（支持 Browserless v2 容器）、#1231（服务重启后重新解析当前 MCP client）。后者与 OpenClaw #68187（SSE 支撑的 MCP 会话在服务重启后仍陈旧、报 "Session not found"）是完全同一个 bug 在两个项目里的独立复现——MCP 重启语义是全生态公摊的坑。
• 停更三项（NullClaw / TinyClaw / ZeptoClaw）：连续多日 0 issue / 0 PR，最新 release 距今 3–5 个月。建议在后续选型与追踪中降权。
• EasyClaw：08-21 至 08-22 连发 5 个版本（v1.8.108 → v1.8.112），但 issue 与 PR 双零，且 release 指向 gaoyangz77/rivonclaw（仓库已改名）。高频发布 + 零公开协作，无法从公开信号判断其社区健康度。

5. 值得关注的趋势信号

• "表面成功、实则失败"成为全生态第一大元问题。 今天至少有五处独立出现：OpenClaw #87637（doctor 误报记忆搜索已禁用）与 #128003（subagent 完成通告从不触发）、NanoBot #5484（isError=false 却携带业务错误信封）、Zeroclaw #9447（分类不完整的终端响应）、PicoClaw #3343（失败 turn 后动画无限编辑）。当 agent 开始无人值守长跑，"失败必须响"比"失败别发生"更关键——这也与今日 CLI 侧 Gemini #22323、Qwen #9733 完全同源。
• 安全模型正在从"功能开关"升级为"身份基础设施"。 Zeroclaw 的 canonical principals / RPC 认证 / OIDC 三阶段，Moltis 的 fail-closed hook 策略，CoPaw 的第五安全层，OpenClaw 的安装策略告警强制确认（#116489，49💬，已合）——四个项目同日从四个层次收紧权限，说明"agent 有真实执行权"这件事的风险已被普遍认知。
• MCP 的重启与会话语义是公摊技术债。 Moltis #1231 与 OpenClaw #68187 独立复现同一 bug，PicoClaw #3269 的 MCP 连接失败会挂死整个 agent loop。MCP 生态缺少统一的"连接失效—重建—幂等恢复"规范，每个项目都在自己重新踩一遍。
• 性能债的形态高度一致：重复重建。 OpenClaw #123535（会话目录刷新风暴，153💬）、#127379（buildModelsProviderData 反复重建打满 CPU）、#72717（wiki_search 缺 FTS 索引），与 CLI 侧 OpenCode #44127（冗余 git 子进程 spawn）、Codex #20864（全量扫描 sessions 目录）形态完全一致。agent 框架普遍在"每次都重算"的阶段，缓存与增量是下一轮明确的优化红利。
• 生态尾部正在自然淘汰。 13 个 peer 中已有 3 个事实停更、1 个仅有发布流水线，真正有持续协作信号的约 8 个。"claw 系"项目的爆发期已过，进入集中度提升阶段。

各项目动态明细

OpenClaw（openclaw/openclaw）

最新 Release：v2026.8.1-beta.2（08-15）；正式版最新为 v2026.7.1-2（08-04）。

高讨论 PR

关键 issue

• #127379 — /models 命令让 gateway 主线程 CPU 打满 100%+，因 buildModelsProviderData 反复重建。
• #87637（2👍 / 3💬）— openclaw doctor 报告记忆搜索已禁用，但 memory-core dreaming 实际在运行——诊断失真。
• #125570 — Skill Workshop 应用更新时会覆盖线上 skill 的描述，静默破坏其触发。
• #124099 — 升级到 2026.8.1-beta.2 后出现 SessionCanonicalKeyMigrationRequiredError 循环，trigger-pending 行受影响。
• #68187 — SSE 支撑的 MCP 会话在服务重启后仍陈旧，报 "Session not found"。
• #72717 — 为 wiki_search 引入 SQLite FTS 索引以提升合成查询性能。
• #88334 — Gateway 应在凭据轮换后立即刷新 auth epoch。
• #45184 — 当消息 @ 的是另一账号的 bot 时应跳过派发。
• #33102 — TUI 支持配置 --deliver 默认值。
• #119254 — 设计 WhatsApp poll_vote_received 插件 hook。

今日新建（3 issue / 22 PR，摘其要）

• #128003 — OpenAI 兼容 HTTP 会话下 subagent 完成通告从不触发。
• #128006 — ACPX 直连 agent 启动缺少持久化进程租约覆盖（无上游生命周期 hook）。
• #128005 — Android location.get 在回退到实时定位时忽略 maxAgeMs。
• PR #127797（8💬）— 统一同步消息 hook。
• PR #127998（6💬）— Doctor 的权限修复建议里列出了已禁用的自动化（诊断准确性修复）。
• PR #127962（6💬）— 为入站 turn 保留 gateway 上下文。
• PR #128004（已合）— 共享 schema 校验的 format scope（性能）。
• PR #128001 — 把请求方的 abort 传播到 ACP spawn。

Zeroclaw（zeroclaw-labs/zeroclaw）

最新 Release：v0.8.4（08-02）。今日 PR 量全生态第一（91 活跃 / 12 新建）。

#8289 安全史诗（今日同时推进三阶段）

• PR #10248 — canonical principals 与共享授权解析（stage 2）。
• PR #10259 — RPC 上强制认证主体，native + peercred（stage 3）。
• PR #10255 — oidc.<alias> 令牌校验 provider（stage 5）。

高讨论 PR

治理与今日新建

• #8692（13💬）— 维护者 RFC 与设计 issue 决策队列 Tracker。
• #8396（9💬）— RFC：让 wire protocol 在 provider 构造与 onboarding 中成为一等公民。
• #10141 — 「请让 sessions 变得可用」。
• 今日新 issue：#10251（17 个 telegram listen_* 测试依赖挂钟超时断言）、#10257（cron update --command 在 agent job 上写入未使用列）、#10249（重复 webhook 处理会把调用方可控的幂等键原样写进日志）、#10261。
• 今日新 PR：#10256（从日志中脱敏重复幂等键，对应 #10249）、#10258（把 command patch 映射到 agent job prompt）、#10260（断连时让 RPC 调用失败）。

IronClaw（nearai/ironclaw）

最新 Release：ironclaw-v1.3.0（08-19），此前 rc.1 / rc.2（08-17 / 08-18）。

产品侧高讨论 PR

• #7650（45💬）— 从运行时证据推导 run outcome。
• #7818 — subagent 后台模式：回执式派生、逐子投递、激活与自愈扫描。
• #7516（14💬）— 为 IronHub agent link 提供操作者界面。
• #7257（11💬）— WebUI 设计系统的提案、计划与检查清单（Epic #7038）。
• #7816 — 在 OOBE 建议抽屉中加入刷新与连接入口。

CI 基建四连（T1–T4，均为今日新建）

• #7821 — 单一 setup-rust 复合动作：toolchain 固定、mold、集中构建 profile（T1）。
• #7817 — nextest 测试流水线、全失败信号、PR 解除限流（T2）。
• #7820 — scope-isolation 测试套件合并探针（T2 后续）。
• #7819 — PR / 队列检查收敛：planner 漂移守卫、PR 上跑 default-features clippy（T3）。
• #7809 — 规范化 preflight：统一门禁列表、worktree 安全钩子、自打印 REPRO（T4）。

onboarding / 集成 issue（今日新建 5 条）

• #7815 — onboarding 建议：打通 connect → suggest → thread 流程所需的累积净新增工作量。
• #7812 — onboarding 建议应尊重用户级工具权限，并使用只读工具访问生成。
• #7823 — Notion 安装在 IronClaw 中失败。
• #7822 — Slack 用户无法在 IronClaw 中完成 Slack 配置。
• #7813 — 建议面板出现时标题被裁切。

NanoClaw（qwibitai/nanoclaw）

最新 Release：v2.2.0（08-13）。0 issue / 16 PR，纯 PR 驱动。

• #3438 — 配置向导在已配置 bot 时提供「再加一个 Telegram bot」。
• #3437 — add-another-bot 路径文档、instance 感知配对与接线。
• #3435 — 把 adapter instance 贯穿配对、首个 agent 初始化与 CLI 欢迎流。
• #3355 / #3356 — 新增 /add-cursor provider skill；新增 Cursor Agent SDK payload。
• #3434 — 轮询型 adapter 不再开启 webhook server。
• #3446 — 在未知发送者门禁中自动丢弃自动化发送者。
• #3442 — 按 driver 隔离能力校验 runtimeTier 并据此选档。
• #3443 — 从 onlyBuiltDependencies 中移除 better-sqlite3，改用其自带预编译产物。
• #3441 — git show 失败时保留文件。
• #3431 — 配对卡片文案更正为 6 位数字。

NanoBot（HKUDS/nanobot）★47,283

最新 Release：v0.3.0（07-25）。0 issue / 11 PR，主题为可观测性与会话生命周期。

• #5486（已合）— 统一 turn 可观测性。
• #5487 — WebUI 文件预览路径修复 + subagent 活动与生命周期回放。
• #5420 — 新增用户可控的 turn 恢复。
• #5484 — 标记那些 isError=false 却返回业务错误信封的 MCP 响应。
• #5483 — 防止已删除会话被延迟消息重新创建。
• #5471 — 让临时（ephemeral）run 不改动会话状态。
• #5485 — 为原生 provider 恢复 LangSmith 追踪。
• #5156（已合）— 从静默停滞的 Telegram 轮询中恢复。
• #3294（已合）— dream 可选 kill switch + 自定义 Phase 1/2 模板路径。

CoPaw（agentscope-ai/CoPaw）

最新 Release：v2.1.1-beta.1（08-20）；正式版 v2.1.0（08-13）。

• #7196（1👍 / 2💬）— 一直显示推理过程是严重的视觉干扰，希望可设默认是否折叠。
• #7201 — 按 provider 分别设置 max_image_bytes / max_video_bytes / max_audio_bytes 上限。
• #7212 — 内联像素尺寸超出 provider 限制的图片会导致请求崩溃（与 #7201 构成因果）。
• #7213 — 会话输出总是有无意义的空行。
• #7043（已关闭）— 希望启动时自动执行 chcp 65001 切到 UTF-8 环境。
• PR #7214 — README 中把 Access Policy 列为第五道安全层。
• PR #7054 — Chrome 支持 LAN/网络浏览器的远程 bridge 端点。
• PR #7050 — 按 cron job 提供模型覆盖选择器。
• PR #7190 — qwenpaw-data 的 PyPI 运行时路径、docker-compose 一键 demo 与环境继承。
• PR #6808 — Console 显示自定义 profile 的 markdown 文件。

PicoClaw（sipeed/picoclaw）

最新 Release：v0.3.1（06-30）、nightly（07-02）。

• #3269（1👍 / 6💬）— MCP server 连接失败会导致 agent loop 挂起，进而拖死整个 PicoClaw 聊天。 对应修复 PR #3337。
• #3343（今日新建）— 失败 turn 之后，工具反馈动画会无限编辑同一条 Telegram 消息。
• PR #3222（16💬）— deltachat 实现重构与文档整理，净减约 200 行。
• PR #3319（已合）— 尊重 exec 超时与布尔型运行选项。

LobsterAI（netease-youdao/LobsterAI）

最新 Release：2026.8.18（08-18）。中文用户需求驱动，提出到合并极快。

• #1213（已关闭）— 为会话详情添加「导出为 Markdown」功能 → 由 PR #1214 实现并合并。
• #1206（已关闭）— 私有化部署的 kimi2.5 模型分析文档会重复处理或重复回复进度。
• PR #1208（已合）— 新增手动重试按钮，支持频繁请求等瞬时错误快速重试。
• PR #1212（已合）— 自定义 provider 上限放宽至 20 个。
• PR #1205（已合）— 会话重命名失败时给出 error toast。
• PR #1209（已合）— 修复 web-search 阻断不受支持的 Chrome flags。
• PR #2452 — 为带斜杠的 model id 保留 provider 信息。

Moltis（moltis-org/moltis）

最新 Release：20260820.01（08-20）。今日 3 条全为新建。

• #1230 — 为修改型安全 hook 提供可选的 fail-closed 错误策略（hook 出错时默认拦截而非放行）。
• PR #1229（1👍 / 3💬）— 支持 Browserless v2 容器。
• PR #1231（1👍 / 2💬）— 服务重启后重新解析当前 MCP client（与 OpenClaw #68187 同一 bug 的独立复现）。

低活跃 / 停更项目

本日报由 agents-radar 自动生成，所有内容基于 GitHub 公开 API 实时抓取并由执行代理人工撰写，未调用任何外部大模型。

Generated: 2026-08-23 02:30 (GMT+8) | Coverage: OpenClaw + 12 peer projects This digest is produced by the agents-radar local automation. All summaries, comparisons, and writing are done by the executing agent — no external LLM API was called.

• OpenClaw
• NanoBot
• Zeroclaw
• PicoClaw
• NanoClaw
• NullClaw
• IronClaw
• LobsterAI
• TinyClaw
• CoPaw
• Moltis
• ZeptoClaw
• EasyClaw

Cross-Ecosystem Comparison

OpenClaw Cross-Ecosystem Comparison Report (2026-08-23)

1. Ecosystem Overview

The clearest signal across the OpenClaw ecosystem today: security governance has graduated from "add a toggle" to "rebuild the trust model." Zeroclaw pushed three stages of its #8289 security epic in a single day — #10248 (canonical principals and shared authorization resolution, stage 2), #10259 (enforce authenticated principals over RPC, native + peercred, stage 3), and #10255 (oidc.<alias> token validation provider, stage 5). These are not scattered patches but a systematic rework that makes "who is calling" a first-class citizen. On the same day Moltis opened #1230 (optional fail-closed error policy for mutating security hooks) and CoPaw landed #7214 (documenting Access Policy as a fifth security layer) — three independent projects tightening the permission boundary from three angles on the same day.

Activity distribution remains sharply bimodal. Zeroclaw (91 PRs) and OpenClaw core (54 PRs / 46 issues) lead, followed by IronClaw (19 PRs), NanoClaw (16 PRs), and NanoBot (11 PRs). Meanwhile NullClaw, TinyClaw, and ZeptoClaw have posted 0 issues / 0 PRs for multiple consecutive days and can be treated as de facto abandoned (latest releases dated 2026-05-29, 2026-03-26, and 2026-04-07 respectively). EasyClaw is a special case: zero issues and zero PRs, yet it shipped five versions (v1.8.108 → v1.8.112) across 08-21 and 08-22 — a pure release pipeline with no visible collaboration, and its releases actually point at gaoyangz77/rivonclaw (the repo has been renamed).

2. Project Activity Comparison (last 24h)

3. OpenClaw Core: Performance and Trustworthy Delivery Are Today's Two Threads

The most-discussed OpenClaw PR today is #123535 "avoid session directory refresh storms" — 153💬, the highest discussion count in the entire ecosystem. Pair it with #127379 (the /models command pins the gateway main thread at 100%+ CPU because buildModelsProviderData is rebuilt repeatedly) and #72717 (introduce a SQLite FTS index for wiki_search to speed up synthetic queries), and the pattern is obvious: the mainline is paying down a specific class of performance debt — redundant rebuilds.

The second thread is contextual correctness of message delivery. Today's new PRs #127962 (preserve gateway context for inbound turns) and #127797 (unify synchronous message hooks, 8💬), plus the already-merged #126424 (scope session delivery to agent bindings) and #124548 (persist sender identity for DM turns), form a single chain: in multi-channel, multi-agent setups, "who does this message belong to and who should answer it" must be reliable. Today's new issue #128003 (subagent completion announcements never fire on OpenAI-compatible HTTP sessions, reporting "In-process gateway di…") is the gap in that chain.

Two self-diagnosis distortion problems deserve separate mention: #87637 (openclaw doctor reports memory search as disabled while memory-core dreaming is actually running) and today's new PR #127998 (Doctor's permission-fix suggestions list already-disabled automations). A diagnostic tool whose story disagrees with system reality is more dangerous than having no diagnostics at all — the same meta-problem as Gemini #22323 ("failure reported as success") on the CLI side today.

4. Peer Project Differentiation

• Zeroclaw: #1 by PR volume today (91 active / 12 new) and the only project in the ecosystem doing systematic security rework. Stages 2/3/5 of the #8289 epic all advanced the same day (canonical principals, RPC authenticated principals, OIDC token validation), together with #9574 (authorization approval responder) and #9203 (authenticated HTTP fan-in), forming a complete identity-to-authorization path. The second thread is context and token accounting: #9535 (anchor context compaction to a ratio of the model window, 22💬), #9713 (expose token accounting on history-trim events, 25💬), and #9447 (classify incomplete terminal responses, 20💬). #9447 and today's empty-response complaint on the OpenCode CLI side are two ends of the same problem — one files the bug, the other writes the fix. The standout provider work is #9109 (native Hailo-Ollama support, 44💬), pointing at local/edge inference. On governance, #8692 (maintainer RFC decision-queue tracker, 13💬) and #8396 (RFC: make the wire protocol a first-class concept in provider construction and onboarding, 9💬) show the project has reached a scale that needs explicit process.
• IronClaw: today shows CI/CD infrastructure and onboarding experience advancing in parallel. On CI it landed four PRs, T1–T4: #7821 (single setup-rust composite action: pinned toolchain, mold, centralized build profiles), #7817 (nextest test pipeline, full failure signal, unthrottled on PRs), #7819 (converge PR/queue checks, planner drift guard, default-features clippy on PRs), and #7809 (normalize preflight: unified gate list, worktree safety hooks, self-printing REPRO) — the classic "the team got big enough that we must pave the road first" signal. On product, the highest-discussion items are #7650 (derive run outcome from runtime evidence, 45💬) and #7818 (subagent background mode: receipt-based spawn, per-child delivery, activation and self-healing sweeps). On onboarding it opened four issues in a row (#7815 close the connect→suggest→thread loop, #7812 suggestion generation should respect user-level tool permissions and use read-only tools, #7823 Notion install fails, #7822 Slack configuration fails) — integration success rate is its most honest current bottleneck.
• NanoClaw: 0 issues / 16 PRs, the most purely PR-driven project in the ecosystem. Today's new PRs cluster tightly around multi-bot and provider wiring: #3438 (config wizard offers "add another Telegram bot" when one already exists), #3437 (matching docs plus instance-aware pairing), #3435 (thread adapter instance through pairing, first-agent init, and the CLI welcome flow), and #3355 + #3356 (add an /add-cursor provider skill and a Cursor Agent SDK payload). On the engineering side, #3442 (validate runtimeTier via driver-isolated capability checks and pick a tier accordingly) and #3443 (drop better-sqlite3 from onlyBuiltDependencies and use its bundled prebuilds) are pragmatic trade-offs. Positioning: the integration layer for many IM channels and many agent providers.
• NanoBot (★47,283, the highest-starred peer): all 11 PRs today point at observability and session lifecycle: #5486 (unify turn observability, merged), #5487 (WebUI file-preview path fix plus subagent activity and lifecycle replay), #5420 (user-controlled turn resume), #5483 (prevent deleted sessions from being recreated by delayed messages), and #5471 (ephemeral runs must not mutate session state). There is also #5484 (flag MCP responses that return isError=false while carrying a business-error envelope) — detecting "looks successful, actually failed" is today's recurring cross-ecosystem theme. #5156 (recover from silently stalled Telegram polling, merged) is the same idea.
• CoPaw: the main thread is multimodal limits and security layering. #7201 (per-provider max_image_bytes / max_video_bytes / max_audio_bytes) and #7212 (inlining images whose pixel dimensions exceed provider limits crashes the request) are cause and effect — missing limits turn directly into crashes. #7214 writes Access Policy into the README as a fifth security layer. Chinese-language UX feedback remains dense: #7196 (always showing the reasoning trace is a serious visual distraction; make collapse-by-default configurable), #7213 (session output always contains meaningless blank lines), and #7043 (auto-run chcp 65001 to switch to UTF-8 at startup, closed). On operations, #7050 (per-cron-job model override selector) and #7190 (PyPI runtime paths, one-command docker-compose demo, env inheritance) lower the self-hosting bar.
• PicoClaw: slow but effective. The core bug #3269 (a failed MCP server connection hangs the agent loop and takes down PicoClaw chat with it, 1👍/6💬) already has a matching fix in PR #3337. Today's new issue #3343 (after a failed turn the tool-feedback animation edits the same Telegram message forever) is the same "failure paths don't clean up" family. #3222 (deltachat implementation refactor, ~200 lines removed, 16💬) reflects its embedded-oriented preference for minimalism.
• LobsterAI (NetEase Youdao): entirely driven by Chinese-language user requests, with an extremely fast loop — #1213 "add export-to-Markdown for session details" was filed and then implemented and merged via #1214. Other fixes are equally close to daily use: #1208 (manual retry button for transient errors), #1212 (raise the custom-provider cap to 20), #1205 (error toast when session rename fails), and #1206 (on-prem kimi2.5 reprocesses analysis documents, closed). "Filed today, merged today" is its biggest differentiator.
• Moltis: small project, clear direction. All three items today are new: #1230 (optional fail-closed error policy for mutating security hooks — whether a failing hook defaults to allow or deny is a foundational security-model choice), #1229 (support Browserless v2 containers), and #1231 (re-resolve the current MCP client after a service restart). That last one and OpenClaw #68187 (SSE-backed MCP sessions remain stale after a service restart and report "Session not found") are the exact same bug independently reproduced in two projects — MCP restart semantics are shared ecosystem-wide debt.
• The three abandoned projects (NullClaw / TinyClaw / ZeptoClaw): 0 issues / 0 PRs for multiple consecutive days, latest releases 3–5 months old. Recommend down-weighting them in future selection and tracking.
• EasyClaw: five releases across 08-21 to 08-22 (v1.8.108 → v1.8.112), but zero issues and zero PRs, and releases point at gaoyangz77/rivonclaw (repo renamed). Frequent releases plus zero public collaboration means community health cannot be judged from public signals.

5. Trend Signals Worth Watching

• "Looks successful, actually failed" is now the ecosystem's #1 meta-problem. It appeared independently in at least five places today: OpenClaw #87637 (doctor misreports memory search as disabled) and #128003 (subagent completion announcements never fire), NanoBot #5484 (isError=false while carrying a business-error envelope), Zeroclaw #9447 (incomplete terminal responses classified), and PicoClaw #3343 (animation edits forever after a failed turn). Once agents run unattended for long stretches, "failures must be loud" matters more than "failures must not happen" — the same root as Gemini #22323 and Qwen #9733 on the CLI side today.
• The security model is being upgraded from "feature toggle" to "identity infrastructure." Zeroclaw's canonical-principals / RPC-auth / OIDC three stages, Moltis's fail-closed hook policy, CoPaw's fifth security layer, and OpenClaw's install-policy warning acknowledgement requirement (#116489, 49💬, merged) — four projects tightening permissions at four levels on the same day means the risk of "agents hold real execution power" is now broadly understood.
• MCP restart and session semantics are shared technical debt. Moltis #1231 and OpenClaw #68187 independently reproduce the same bug, and PicoClaw #3269 shows a failed MCP connection can hang an entire agent loop. The MCP ecosystem lacks a shared "connection invalidation → rebuild → idempotent recovery" convention, so every project re-learns it the hard way.
• Performance debt has a highly consistent shape: redundant rebuilds. OpenClaw #123535 (session directory refresh storms, 153💬), #127379 (buildModelsProviderData rebuilt repeatedly, pinning CPU), and #72717 (wiki_search lacks an FTS index) share the exact shape of OpenCode #44127 (redundant git subprocess spawns) and Codex #20864 (full scan of the sessions directory) on the CLI side. Agent frameworks are broadly still in the "recompute everything each time" phase; caching and incrementality are the next clear optimization dividend.
• The ecosystem tail is thinning out naturally. Of 13 peers, three are de facto abandoned and one has only a release pipeline; roughly eight show genuine ongoing collaboration. The "claw family" explosion is over; consolidation has begun.

Per-Project Details

OpenClaw (openclaw/openclaw)

Latest release: v2026.8.1-beta.2 (08-15); latest stable is v2026.7.1-2 (08-04).

High-discussion PRs

Key issues

• #127379 — The /models command pins the gateway main thread at 100%+ CPU because buildModelsProviderData is rebuilt repeatedly.
• #87637 (2👍 / 3💬) — openclaw doctor reports memory search as disabled while memory-core dreaming is actually running — diagnostic distortion.
• #125570 — Applying a Skill Workshop update overwrites the live skill's description, silently breaking its triggers.
• #124099 — After upgrading to 2026.8.1-beta.2, a SessionCanonicalKeyMigrationRequiredError loop appears and trigger-pending rows are affected.
• #68187 — SSE-backed MCP sessions remain stale after a service restart and report "Session not found".
• #72717 — Introduce a SQLite FTS index for wiki_search to speed up synthetic queries.
• #88334 — Gateway should refresh the auth epoch immediately after credential rotation.
• #45184 — Skip dispatch when a message @-mentions another account's bot.
• #33102 — Allow the TUI to configure a default for --deliver.
• #119254 — Design a WhatsApp poll_vote_received plugin hook.

Opened today (3 issues / 22 PRs, selected)

• #128003 — Subagent completion announcements never fire on OpenAI-compatible HTTP sessions.
• #128006 — ACPX direct-agent startup lacks a persistent process-lease override (no upstream lifecycle hook).
• #128005 — Android location.get ignores maxAgeMs when falling back to live positioning.
• PR #127797 (8💬) — Unify synchronous message hooks.
• PR #127998 (6💬) — Doctor's permission-fix suggestions list already-disabled automations (diagnostic accuracy fix).
• PR #127962 (6💬) — Preserve gateway context for inbound turns.
• PR #128004 (merged) — Share the format scope for schema validation (performance).
• PR #128001 — Propagate the caller's abort into the ACP spawn.

Zeroclaw (zeroclaw-labs/zeroclaw)

Latest release: v0.8.4 (08-02). #1 by PR volume in the ecosystem today (91 active / 12 new).

The #8289 security epic (three stages advanced the same day)

• PR #10248 — Canonical principals and shared authorization resolution (stage 2).
• PR #10259 — Enforce authenticated principals over RPC, native + peercred (stage 3).
• PR #10255 — oidc.<alias> token validation provider (stage 5).

High-discussion PRs

Governance and items opened today

• #8692 (13💬) — Tracker for the maintainer RFC and design-issue decision queue.
• #8396 (9💬) — RFC: make the wire protocol a first-class concept in provider construction and onboarding.
• #10141 — "Please make sessions usable."
• New issues today: #10251 (17 telegram listen_* tests depend on wall-clock timeout assertions), #10257 (cron update --command writes an unused column on agent jobs), #10249 (duplicate webhook handling logs a caller-controlled idempotency key verbatim), #10261.
• New PRs today: #10256 (redact duplicate idempotency keys from logs, addresses #10249), #10258 (map command patches onto the agent job prompt), #10260 (fail RPC calls on disconnect).

IronClaw (nearai/ironclaw)

Latest release: ironclaw-v1.3.0 (08-19), preceded by rc.1 / rc.2 (08-17 / 08-18).

High-discussion product PRs

• #7650 (45💬) — Derive run outcome from runtime evidence.
• #7818 — Subagent background mode: receipt-based spawn, per-child delivery, activation and self-healing sweeps.
• #7516 (14💬) — Operator surface for IronHub agent link.
• #7257 (11💬) — WebUI design-system proposal, plan, and checklist (Epic #7038).
• #7816 — Add refresh and connect entries to the OOBE suggestions drawer.

CI infrastructure, four in a row (T1–T4, all opened today)

• #7821 — Single setup-rust composite action: pinned toolchain, mold, centralized build profiles (T1).
• #7817 — nextest test pipeline, full failure signal, unthrottled on PRs (T2).
• #7820 — Merge probe for the scope-isolation test suite (T2 follow-up).
• #7819 — Converge PR / queue checks: planner drift guard, default-features clippy on PRs (T3).
• #7809 — Normalize preflight: unified gate list, worktree safety hooks, self-printing REPRO (T4).

Onboarding / integration issues (5 opened today)

• #7815 — Onboarding suggestions: cumulative net-new work needed to close the connect → suggest → thread loop.
• #7812 — Onboarding suggestions should respect user-level tool permissions and use read-only tool access for generation.
• #7823 — Notion install fails in IronClaw.
• #7822 — Slack users cannot complete Slack configuration in IronClaw.
• #7813 — Title is clipped when the suggestions panel appears.

NanoClaw (qwibitai/nanoclaw)

Latest release: v2.2.0 (08-13). 0 issues / 16 PRs — purely PR-driven.

• #3438 — Config wizard offers "add another Telegram bot" when one is already configured.
• #3437 — Docs for the add-another-bot path, instance-aware pairing, and wiring.
• #3435 — Thread adapter instance through pairing, first-agent init, and the CLI welcome flow.
• #3355 / #3356 — Add an /add-cursor provider skill; add a Cursor Agent SDK payload.
• #3434 — Polling adapters no longer start a webhook server.
• #3446 — Auto-drop automation senders in the unknown-sender gate.
• #3442 — Validate runtimeTier via driver-isolated capability checks and pick a tier accordingly.
• #3443 — Remove better-sqlite3 from onlyBuiltDependencies and use its bundled prebuilds.
• #3441 — Preserve files when git show fails.
• #3431 — Correct pairing-card copy to six digits.

NanoBot (HKUDS/nanobot) ★47,283

Latest release: v0.3.0 (07-25). 0 issues / 11 PRs, themed on observability and session lifecycle.

• #5486 (merged) — Unify turn observability.
• #5487 — WebUI file-preview path fix plus subagent activity and lifecycle replay.
• #5420 — Add user-controlled turn resume.
• #5484 — Flag MCP responses that return isError=false while carrying a business-error envelope.
• #5483 — Prevent deleted sessions from being recreated by delayed messages.
• #5471 — Ensure ephemeral runs do not mutate session state.
• #5485 — Restore LangSmith tracing for native providers.
• #5156 (merged) — Recover from silently stalled Telegram polling.
• #3294 (merged) — Optional dream kill switch plus custom Phase 1/2 template paths.

CoPaw (agentscope-ai/CoPaw)

Latest release: v2.1.1-beta.1 (08-20); stable v2.1.0 (08-13).

• #7196 (1👍 / 2💬) — Always showing the reasoning trace is a serious visual distraction; make collapse-by-default configurable.
• #7201 — Per-provider max_image_bytes / max_video_bytes / max_audio_bytes limits.
• #7212 — Inlining images whose pixel dimensions exceed provider limits crashes the request (cause-and-effect with #7201).
• #7213 — Session output always contains meaningless blank lines.
• #7043 (closed) — Auto-run chcp 65001 at startup to switch to a UTF-8 environment.
• PR #7214 — Document Access Policy as a fifth security layer in the README.
• PR #7054 — Chrome support for remote bridge endpoints on LAN/network browsers.
• PR #7050 — Per-cron-job model override selector.
• PR #7190 — PyPI runtime paths for qwenpaw-data, one-command docker-compose demo, and env inheritance.
• PR #6808 — Console shows markdown files for custom profiles.

PicoClaw (sipeed/picoclaw)

Latest release: v0.3.1 (06-30), nightly (07-02).

• #3269 (1👍 / 6💬) — A failed MCP server connection hangs the agent loop and takes down the whole PicoClaw chat. Matching fix: PR #3337.
• #3343 (opened today) — After a failed turn the tool-feedback animation edits the same Telegram message forever.
• PR #3222 (16💬) — deltachat implementation refactor and docs cleanup, ~200 lines net removed.
• PR #3319 (merged) — Respect exec timeouts and boolean run options.

LobsterAI (netease-youdao/LobsterAI)

Latest release: 2026.8.18 (08-18). Driven by Chinese-language user requests, with a very short file-to-merge loop.

• #1213 (closed) — Add "export to Markdown" for session details → implemented and merged via PR #1214.
• #1206 (closed) — On-prem kimi2.5 reprocesses analysis documents or repeats progress replies.
• PR #1208 (merged) — Add a manual retry button for transient errors such as rate limiting.
• PR #1212 (merged) — Raise the custom-provider cap to 20.
• PR #1205 (merged) — Show an error toast when session rename fails.
• PR #1209 (merged) — Fix web-search blocking unsupported Chrome flags.
• PR #2452 — Preserve provider information for model ids containing slashes.

Moltis (moltis-org/moltis)

Latest release: 20260820.01 (08-20). All three items today are new.

• #1230 — Optional fail-closed error policy for mutating security hooks (deny rather than allow when a hook errors).
• PR #1229 (1👍 / 3💬) — Support Browserless v2 containers.
• PR #1231 (1👍 / 2💬) — Re-resolve the current MCP client after a service restart (an independent reproduction of OpenClaw #68187).

Low-Activity / Stalled Projects

This digest is generated by agents-radar. All content is fetched live from GitHub's public API and written by the executing agent — no external LLM was called.

生成时间: 2026-08-23 02:30 (GMT+8) 数据来源：GitHub Trending（每日榜）+ gh search repos 6 个 AI 主题（按 stars 排序，7 天窗口） 本日报由 agents-radar 本地自动化生成，所有撰写由执行代理人工完成，未调用任何外部大模型 API。

一、GitHub Trending 每日飙升榜（2026-08-23）

说明：本次抓取时 GitHub Trending 页面仍未对外暴露「当日新增 Star」数值（连续第二日，疑为未登录视图/版式变更），故下表仅列出入榜项目与简介，不含 star 增量。排名按页面原始顺序。

今日榜单最强信号：Codex 登顶，Claude Code 同榜第 8。 这是三份日报罕见的三方共振——Trending 第 1 是 openai/codex，HN 今日热度第 1 是「A week of using Codex more than Claude」（94👍），而 Claude Code 侧今日最高热度 issue #77136（343👍）是对模型口头习惯的集体吐槽。社区注意力正在从 Claude Code 向 Codex 迁移，且这一迁移同时体现在 star 榜、讨论榜和情绪面三个维度上。

第二个信号：Skills / harness 元层占据 5 席。 #2 mattpocock/skills、#3 affaan-m/ECC、#4 obra/superpowers、#11 multica-ai/andrej-karpathy-skills、#15 cursor/plugins——前四名里有三个是「给 agent 装能力」的项目，而不是 agent 本身。值得注意的是 #11 的形态极端到只有一个 CLAUDE.md 文件却能上榜，说明当前最稀缺的不是工具，而是"怎么让模型少犯错"的经验沉淀。

第三个信号：成本焦虑显性化。 #5 Wei-Shaw/sub2api 做的是「订阅统一接入 + 拼车共享分摊成本」，#13 ripienaar/free-for-dev 是免费额度清单。这与今日 CLI 日报中 GPT-5.6 Sol 降价 20%、以及 HN 上关于「A/B 测试降低推理强度」的争议属于同一背景——用户对 agent 的月度支出已经到了要专门做工具去摊平的程度。

二、六大 AI 主题搜索热度（按 stars，Top 15）

1. AI Agent 框架

• NousResearch/hermes-agent ⭐234,323 — The agent that grows with you
• shareAI-lab/learn-claude-code ⭐74,949 — 从 0 到 1 构建类 claude code 的 agent harness
• Panniantong/Agent-Reach ⭐74,085 — 给 AI agent 全网的眼睛（Twitter / Reddit / YouTube / GitHub / B 站 / 小红书），一个 CLI 零 API 费用
• thedaviddias/Front-End-Checklist ⭐73,609 — 现代 Web 开发必备清单（人类与 AI agent 通用）
• santifer/career-ops ⭐67,764 — 开源 AI 求职：扫描招聘站点、A–F 评分体系、简历定制与投递追踪，可本地跑在各 AI 编码 CLI
• ZhuLinsen/daily_stock_analysis ⭐63,624 — LLM 驱动的多市场股票智能分析（多源行情、实时新闻、决策看板、自动推送）
• CherryHQ/cherry-studio ⭐50,919 — AI 生产力工作室，300+ 助手统一接入前沿 LLM
• hugohe3/ppt-master ⭐48,617 — AI 把文档或主题变成真正的原生 PPT（原生形状、转场动画、数据图表、语音旁白）
• HKUDS/nanobot ⭐47,283 — 超轻量、自托管的个人 AI agent 框架（WebUI/工具/记忆/MCP/多 agent 工作流）
• zhayujie/CowAgent ⭐46,632 — 开源超级 AI 助手与 Agent Harness，带记忆与知识自进化
• siyuan-note/siyuan ⭐45,931 — 开源、隐私优先、自托管的知识工作空间
• bojieli/ai-agent-book ⭐40,913 — 《深入理解 AI Agent：设计原理与工程实践》全书正文 + PDF + 配套代码
• Hmbown/CodeWhale ⭐40,841 — Rust 编写的终端开源编码 agent
• CopilotKit/CopilotKit ⭐36,960 — Agent 与 Generative UI 的前端栈（AG-UI 协议作者）
• esengine/DeepSeek-Reasonix ⭐35,032 — DeepSeek 原生的终端编码 agent，围绕前缀缓存稳定性设计

2. 大模型（LLM）

• affaan-m/ECC ⭐242,087 — agent harness 性能优化系统
• NousResearch/hermes-agent ⭐234,323 — The agent that grows with you
• Significant-Gravitas/AutoGPT ⭐186,764 — 人人可用的自主 AI
• ollama/ollama ⭐179,192 — 一键跑起 Kimi-K2.6 / GLM-5.2 / MiniMax / DeepSeek / gpt-oss / Qwen / Gemma
• firecrawl/firecrawl ⭐170,855 — 大规模搜索/抓取/交互 Web 的上下文 API
• f/prompts.chat ⭐167,752 — Awesome ChatGPT Prompts 社区共享（可自托管）
• huggingface/transformers ⭐164,339 — 文本/视觉/音频/多模态 SOTA 模型框架
• langgenius/dify ⭐153,205 — Agentic 工作流与 RAG 管线协作平台
• open-webui/open-webui ⭐149,579 — 用户友好的 AI 界面
• langchain-ai/langchain ⭐144,777 — agent 工程平台
• harry0703/MoneyPrinterTurbo ⭐114,563 — AI 一键生成高清短视频
• browser-use/browser-use ⭐110,116 — 让网站对 AI agent 可用，轻松自动化线上任务
• Graphify-Labs/graphify ⭐109,501 — 把代码库（含文档/SQL schema/配置/PDF）变成可查询知识图谱，作为 /graphify skill 供各 CLI 使用
• DietrichGebert/ponytail ⭐108,073 — 让 AI agent 像「屋里最懒的资深工程师」一样思考：最好的代码是你没写的代码
• rasbt/LLMs-from-scratch ⭐103,224 — 从零逐步用 PyTorch 实现类 ChatGPT LLM

3. RAG 检索增强生成

• langgenius/dify ⭐153,205 — Agentic 工作流与 RAG 管线
• open-webui/open-webui ⭐149,579 — 用户友好的 AI 界面
• langchain-ai/langchain ⭐144,777 — agent 工程平台
• Shubhamsaboo/awesome-llm-apps ⭐133,588 — 100+ AI Agents / Agent Skills / RAG 应用
• Graphify-Labs/graphify ⭐109,501 — 代码库转可查询知识图谱（本地确定性解析）
• thedotmack/claude-mem ⭐91,519 — 跨会话持久上下文：记录 agent 全过程、AI 压缩、按需注入未来会话
• infiniflow/ragflow ⭐89,036 — 领先的开源 RAG 引擎，融合 Agent 能力
• PaddlePaddle/PaddleOCR ⭐88,110 — 轻量 OCR 工具包，支持 100+ 语言，桥接图像/PDF 与 LLM
• dair-ai/Prompt-Engineering-Guide ⭐77,693 — 提示工程/上下文工程/RAG/AI Agents 指南
• datawhalechina/hello-agents ⭐74,217 — 《从零开始构建智能体》教程
• headroomlabs-ai/headroom ⭐67,179 — 在工具输出/日志/文件/RAG 块进入 LLM 前压缩：编码 agent 省 20% token，JSON 省 60–95%
• Mintplex-Labs/anything-llm ⭐65,052 — 本地优先的强力 agent 体验
• mem0ai/mem0 ⭐63,821 — AI Agent 的通用记忆层
• pathwaycom/llm-app ⭐59,005 — RAG / AI 管线与企业搜索云模板
• FlowiseAI/Flowise ⭐55,383 — 可视化构建 AI Agent

4. 向量数据库

• Mintplex-Labs/anything-llm ⭐65,052 — 本地优先的 agent 体验
• meilisearch/meilisearch ⭐59,054 — 极速搜索 API，AI 混合搜索
• pathwaycom/llm-app ⭐59,005 — RAG/AI 管线云模板
• run-llama/llama_index ⭐51,801 — 文档 agent 与 OCR 平台
• milvus-io/milvus ⭐45,736 — 高性能云原生向量数据库
• VectifyAI/PageIndex ⭐35,291 — 无向量的推理式 RAG 文档索引
• qdrant/qdrant ⭐34,124 — 大规模向量数据库与向量搜索引擎
• topoteretes/cognee ⭐30,183 — 开源 AI 记忆平台（自托管知识图谱引擎）
• NirDiamant/RAG_Techniques ⭐29,174 — RAG 高级技术笔记本合集
• weaviate/weaviate ⭐16,746 — 开源向量数据库，向量搜索 + 结构化过滤
• memvid/memvid ⭐16,437 — 用无服务器单文件记忆层替代复杂 RAG 管线
• alibaba/zvec ⭐15,502 — 轻量、极速的进程内向量数据库
• langchain4j/langchain4j ⭐12,929 — JVM 上的 LLM 应用库，统一 provider 与向量库 API
• neuml/txtai ⭐12,897 — 语义搜索、LLM 编排与语言模型工作流的一体化框架
• StarTrail-org/LEANN ⭐12,827 — [MLsys2026] 节省 97% 存储的本地私有 RAG

5. LLM 模型

• jingyaogong/minimind ⭐54,927 — 2 小时从零训练 64M 参数 LLM
• ScrapeGraphAI/Scrapegraph-ai ⭐29,812 — 基于 AI 的 Python 爬虫
• The-Pocket/PocketFlow-Tutorial-Codebase-Knowledge ⭐12,630 — Codebase to Tutorial
• The-Pocket/PocketFlow ⭐11,124 — 100 行 LLM 框架，让 agent 构建 agent
• OpenSPG/KAG ⭐9,004 — 逻辑形式引导的推理与检索框架
• 0xPlaygrounds/rig ⭐8,358 — Rust 构建模块化可扩展 LLM 应用
• open-compass/opencompass ⭐7,327 — LLM 评测平台，100+ 数据集
• InternLM/InternLM ⭐7,266 — InternLM 系列官方发布
• Eigenwise/atomic-agents ⭐6,191 — 原子化构建 AI agent
• gluonfield/enchanted ⭐5,998 — iOS/macOS 私有自托管模型聊天客户端
• skyzh/tiny-llm ⭐4,513 — 在 Apple Silicon 上学 LLM 推理系统：造一个小 vLLM + Qwen
• verazuo/jailbreak_llms ⭐3,791 — [CCS'24] 15,140 条 ChatGPT prompt 数据集（含 1,405 条越狱 prompt）
• ridgerchu/matmulfreellm ⭐3,088 — 无矩阵乘法语言模型实现
• InternLM/InternLM-XComposer ⭐2,925 — 长时流式视音频交互的多模态系统
• JIA-Lab-research/LISA ⭐2,671 — 用大语言模型做推理式分割

6. 机器学习

• tensorflow/tensorflow ⭐197,323 — 人人可用的开源机器学习框架
• f/prompts.chat ⭐167,752 — Awesome ChatGPT Prompts
• huggingface/transformers ⭐164,339 — SOTA 机器学习模型框架
• rasbt/LLMs-from-scratch ⭐103,224 — 从零逐步实现类 ChatGPT LLM
• pytorch/pytorch ⭐102,543 — 带 GPU 加速的张量与动态神经网络
• microsoft/ML-For-Beginners ⭐89,682 — 12 周 26 课 52 测验的经典机器学习
• Developer-Y/cs-video-courses ⭐83,148 — 带视频讲座的 CS 课程清单
• mlabonne/llm-course ⭐81,915 — LLM 学习路线与 Colab 笔记本
• netdata/netdata ⭐80,261 — 最快的 AI 驱动全栈可观测
• d2l-ai/d2l-zh ⭐79,877 — 《动手学深度学习》中文版
• tesseract-ocr/tesseract ⭐76,116 — Tesseract 开源 OCR 引擎
• binhnguyennus/awesome-scalability ⭐73,459 — 大规模系统的可扩展性/可靠性/性能模式
• OpenBB-finance/OpenBB ⭐72,154 — 面向分析师、量化与 AI agent 的开放数据平台
• labmlai/annotated_deep_learning_paper_implementations ⭐67,331 — 60+ 深度学习论文实现与逐行注释
• scikit-learn/scikit-learn ⭐67,010 — Python 机器学习库

三、24 小时 Star 增量榜（与昨日快照对比）

说明：本表由 agents-radar 自身连续两日抓取的 star 快照相减得出，可视为「过去 24h 真实新增」。仅统计昨日日报中已出现的项目。

读法：前五名里有四个是「agent 能力/上下文供给」类项目——career-ops（把 agent 用在求职）、hermes-agent、graphify（代码库转知识图谱）、ECC（harness 优化），只有 firecrawl 属于纯数据抓取基础设施。增量的钱在往"怎么把 agent 用起来"和"给 agent 喂什么上下文"两个方向走，而不是往模型本体走。 相比之下 pytorch（+26）、transformers（+29）、milvus（+8）、qdrant（+9）这些传统 ML/向量库的增速已经是前者的十分之一量级。

四、趋势小结

• Codex 与 Claude Code 同日登榜，Codex 居首。 叠加 HN 头条与 Claude Code 侧 343👍 的吐槽 issue，这是今日最值得记录的生态位变化信号——不是某一个榜的偶然，而是 star、讨论、情绪三个独立维度同时指向同一方向。
• 「Skills / harness 元层」已成为独立赛道且极度拥挤。 Trending 前四有三席（mattpocock/skills、affaan-m/ECC、obra/superpowers），第 11 名是单个 CLAUDE.md 文件，第 15 名是 cursor/plugins。当"经验沉淀"本身可以成为爆款仓库，说明模型能力已不是瓶颈，怎么约束模型才是。
• 成本工程正式浮出水面。 sub2api（订阅中转/拼车）+ free-for-dev（免费额度清单）+ headroom（进 LLM 前压缩 token，省 20–95%）+ DeepSeek-Reasonix（围绕前缀缓存稳定性设计）——四个项目从四个角度砍同一笔账单，与 CLI 侧 GPT-5.6 Sol 降价 20% 构成供需两端的同一个故事。
• 记忆/上下文赛道持续加厚。 claude-mem（+89）、graphify（+405）、mem0、cognee、memvid、LEANN、zvec 同屏，且 graphify 的增量排到全榜第 3。这与 OpenClaw 生态日报里"性能债 = 重复重建"、CLI 侧"上下文压缩"完全同源：上下文是当前最贵的资源，围绕它的工具化正在加速。
• 安全治理连续在榜。 Tencent/AI-Infra-Guard 连续第二日入榜（Agent Scan / Skills Scan / MCP scan），与 Zeroclaw 的身份基础设施改造、Moltis 的 fail-closed hook、CoPaw 的第五安全层形成呼应——"agent 有真实执行权"的风险共识已经从框架内部扩散到工具链外部。
• 传统 ML 基础设施增速趋缓。 pytorch +26 / transformers +29 / milvus +8 / qdrant +9，与 agent 层项目的 +400 量级差了一个数量级。注意力与新增关注度几乎全部集中在应用与 harness 层。

本日报由 agents-radar 自动生成，所有内容基于 GitHub 公开数据实时抓取并由执行代理人工撰写，未调用任何外部大模型。

Generated: 2026-08-23 02:30 (GMT+8) Sources: GitHub Trending (daily) + gh search repos across 6 AI topics (sorted by stars, 7-day window) This digest is produced by the agents-radar local automation. All writing is done by the executing agent — no external LLM API was called.

1. GitHub Trending — Daily Risers (2026-08-23)

Note: at fetch time the GitHub Trending page still did not expose "stars today" values (second consecutive day; likely an anonymous-view or layout change), so the table below lists ranked entries and descriptions without star deltas. Order follows the page's original ranking.

The strongest signal today: Codex is #1 and Claude Code is #8 on the same board. This is a rare three-way resonance across all three digests — Trending #1 is openai/codex, today's top HN story is "A week of using Codex more than Claude" (94👍), and on the Claude Code side today's highest-heat issue #77136 (343👍) is a collective complaint about the model's verbal tics. Community attention is migrating from Claude Code toward Codex, and that migration shows up simultaneously in the star board, the discussion board, and the sentiment.

Second signal: the skills/harness meta-layer holds five slots. #2 mattpocock/skills, #3 affaan-m/ECC, #4 obra/superpowers, #11 multica-ai/andrej-karpathy-skills, and #15 cursor/plugins — three of the top four are projects that give agents capabilities rather than being agents themselves. Notably #11 is extreme in form: a single CLAUDE.md file, yet it trends. What is scarcest right now is not tooling but codified experience about how to make the model make fewer mistakes.

Third signal: cost anxiety has become explicit. #5 Wei-Shaw/sub2api unifies subscriptions and supports pooled/shared access to split costs; #13 ripienaar/free-for-dev is a free-tier catalogue. This shares a background with GPT-5.6 Sol's 20% price cut in today's CLI digest and the HN controversy over A/B-tested reduced reasoning effort — users' monthly agent spend has reached the point where people build dedicated tooling to flatten it.

2. Six AI Topics by Stars (Top 15)

1. AI Agent Frameworks

• NousResearch/hermes-agent ⭐234,323 — The agent that grows with you
• shareAI-lab/learn-claude-code ⭐74,949 — Build a nano claude-code-like agent harness from 0 to 1
• Panniantong/Agent-Reach ⭐74,085 — Give your agent eyes on the whole internet (Twitter / Reddit / YouTube / GitHub / Bilibili / XiaoHongShu), one CLI, zero API fees
• thedaviddias/Front-End-Checklist ⭐73,609 — Essential checklist for modern web development, for humans and AI agents
• santifer/career-ops ⭐67,764 — Open-source AI job search: scan portals, A–F rubric scoring, tailor CVs, track applications, runs locally in your coding CLI
• ZhuLinsen/daily_stock_analysis ⭐63,624 — LLM-powered multi-market stock analysis with live news, dashboards, and automated notifications
• CherryHQ/cherry-studio ⭐50,919 — AI productivity studio with 300+ assistants and unified frontier-LLM access
• hugohe3/ppt-master ⭐48,617 — AI turns documents or topics into real native PowerPoint decks with shapes, transitions, charts, and narration
• HKUDS/nanobot ⭐47,283 — Ultra-lightweight self-hosted personal AI agent framework (WebUI, tools, memory, MCP, multi-agent workflows)
• zhayujie/CowAgent ⭐46,632 — Open-source super assistant and agent harness that self-evolves with memory and knowledge
• siyuan-note/siyuan ⭐45,931 — Open-source, privacy-first, self-hosted knowledge workspace
• bojieli/ai-agent-book ⭐40,913 — Full text, compiled PDF, and per-chapter code for a book on AI agent design and engineering
• Hmbown/CodeWhale ⭐40,841 — Open-source terminal coding agent built in Rust
• CopilotKit/CopilotKit ⭐36,960 — The frontend stack for agents and generative UI (makers of the AG-UI protocol)
• esengine/DeepSeek-Reasonix ⭐35,032 — DeepSeek-native terminal coding agent, engineered around prefix-cache stability

2. Large Language Models

• affaan-m/ECC ⭐242,087 — Agent-harness performance optimization system
• NousResearch/hermes-agent ⭐234,323 — The agent that grows with you
• Significant-Gravitas/AutoGPT ⭐186,764 — Accessible AI for everyone
• ollama/ollama ⭐179,192 — Run Kimi-K2.6 / GLM-5.2 / MiniMax / DeepSeek / gpt-oss / Qwen / Gemma locally
• firecrawl/firecrawl ⭐170,855 — Context API to search, scrape, and interact with the web at scale
• f/prompts.chat ⭐167,752 — Community prompt sharing, self-hostable
• huggingface/transformers ⭐164,339 — Model-definition framework for text, vision, audio, multimodal
• langgenius/dify ⭐153,205 — Agentic workflows and RAG pipelines on one workspace
• open-webui/open-webui ⭐149,579 — User-friendly AI interface
• langchain-ai/langchain ⭐144,777 — The agent engineering platform
• harry0703/MoneyPrinterTurbo ⭐114,563 — Generate HD short videos from a topic via an automated AI workflow
• browser-use/browser-use ⭐110,116 — Make websites accessible for AI agents
• Graphify-Labs/graphify ⭐109,501 — Turn a codebase (docs, SQL schemas, configs, PDFs) into a queryable knowledge graph as a /graphify skill
• DietrichGebert/ponytail ⭐108,073 — Make your agent think like the laziest senior dev in the room: the best code is code you never wrote
• rasbt/LLMs-from-scratch ⭐103,224 — Implement a ChatGPT-like LLM in PyTorch step by step

3. RAG (Retrieval-Augmented Generation)

• langgenius/dify ⭐153,205 — Agentic workflows and RAG pipelines
• open-webui/open-webui ⭐149,579 — User-friendly AI interface
• langchain-ai/langchain ⭐144,777 — The agent engineering platform
• Shubhamsaboo/awesome-llm-apps ⭐133,588 — 100+ AI agents, agent skills, and RAG apps
• Graphify-Labs/graphify ⭐109,501 — Codebase to queryable knowledge graph, local deterministic parsing
• thedotmack/claude-mem ⭐91,519 — Persistent cross-session context: capture, compress, and re-inject
• infiniflow/ragflow ⭐89,036 — Leading open-source RAG engine fused with agent capabilities
• PaddlePaddle/PaddleOCR ⭐88,110 — Lightweight OCR toolkit for 100+ languages, bridging images/PDFs and LLMs
• dair-ai/Prompt-Engineering-Guide ⭐77,693 — Guides for prompt engineering, context engineering, RAG, and AI agents
• datawhalechina/hello-agents ⭐74,217 — Building agents from scratch: principles and practice
• headroomlabs-ai/headroom ⭐67,179 — Compress tool outputs, logs, files, and RAG chunks before they reach the LLM: 20% fewer tokens for coding agents, 60–95% for JSON
• Mintplex-Labs/anything-llm ⭐65,052 — Local-first, powerful agent experience
• mem0ai/mem0 ⭐63,821 — Universal memory layer for AI agents
• pathwaycom/llm-app ⭐59,005 — Cloud templates for RAG, AI pipelines, and enterprise search
• FlowiseAI/Flowise ⭐55,383 — Build AI agents visually

4. Vector Databases

• Mintplex-Labs/anything-llm ⭐65,052 — Local-first agent experience
• meilisearch/meilisearch ⭐59,054 — Lightning-fast search API with AI hybrid search
• pathwaycom/llm-app ⭐59,005 — Cloud templates for RAG and AI pipelines
• run-llama/llama_index ⭐51,801 — Document agent and OCR platform
• milvus-io/milvus ⭐45,736 — High-performance cloud-native vector database
• VectifyAI/PageIndex ⭐35,291 — Document index for vectorless, reasoning-based RAG
• qdrant/qdrant ⭐34,124 — High-performance, massive-scale vector database and search engine
• topoteretes/cognee ⭐30,183 — Open-source AI memory platform with a self-hosted knowledge-graph engine
• NirDiamant/RAG_Techniques ⭐29,174 — Advanced RAG techniques with notebook tutorials
• weaviate/weaviate ⭐16,746 — Open-source vector database combining vector search with structured filtering
• memvid/memvid ⭐16,437 — Replace complex RAG pipelines with a serverless single-file memory layer
• alibaba/zvec ⭐15,502 — Lightweight, lightning-fast in-process vector database
• langchain4j/langchain4j ⭐12,929 — Idiomatic Java library for LLM apps on the JVM with a unified provider and vector-store API
• neuml/txtai ⭐12,897 — All-in-one framework for semantic search, LLM orchestration, and language-model workflows
• StarTrail-org/LEANN ⭐12,827 — [MLsys2026] Private on-device RAG with 97% storage savings

5. LLM Models

• jingyaogong/minimind ⭐54,927 — Train a 64M-parameter LLM from scratch in 2 hours
• ScrapeGraphAI/Scrapegraph-ai ⭐29,812 — AI-based Python scraper
• The-Pocket/PocketFlow-Tutorial-Codebase-Knowledge ⭐12,630 — Codebase to tutorial
• The-Pocket/PocketFlow ⭐11,124 — 100-line LLM framework; let agents build agents
• OpenSPG/KAG ⭐9,004 — Logical-form-guided reasoning and retrieval framework
• 0xPlaygrounds/rig ⭐8,358 — Build modular, scalable LLM applications in Rust
• open-compass/opencompass ⭐7,327 — LLM evaluation platform across 100+ datasets
• InternLM/InternLM ⭐7,266 — Official InternLM series releases
• Eigenwise/atomic-agents ⭐6,191 — Building AI agents atomically
• gluonfield/enchanted ⭐5,998 — iOS/macOS client for private self-hosted models
• skyzh/tiny-llm ⭐4,513 — Learn LLM inference systems on Apple Silicon: build a tiny vLLM + Qwen
• verazuo/jailbreak_llms ⭐3,791 — [CCS'24] 15,140 ChatGPT prompts including 1,405 jailbreak prompts
• ridgerchu/matmulfreellm ⭐3,088 — Implementation of MatMul-free LM
• InternLM/InternLM-XComposer ⭐2,925 — Multimodal system for long-term streaming video and audio interaction
• JIA-Lab-research/LISA ⭐2,671 — Reasoning segmentation via large language models

6. Machine Learning

• tensorflow/tensorflow ⭐197,323 — Open-source machine learning framework for everyone
• f/prompts.chat ⭐167,752 — Community prompt sharing
• huggingface/transformers ⭐164,339 — State-of-the-art ML model framework
• rasbt/LLMs-from-scratch ⭐103,224 — Implement a ChatGPT-like LLM step by step
• pytorch/pytorch ⭐102,543 — Tensors and dynamic neural networks with GPU acceleration
• microsoft/ML-For-Beginners ⭐89,682 — 12 weeks, 26 lessons, 52 quizzes of classic ML
• Developer-Y/cs-video-courses ⭐83,148 — CS courses with video lectures
• mlabonne/llm-course ⭐81,915 — LLM roadmaps and Colab notebooks
• netdata/netdata ⭐80,261 — Fastest path to AI-powered full-stack observability
• d2l-ai/d2l-zh ⭐79,877 — Dive into Deep Learning, Chinese edition
• tesseract-ocr/tesseract ⭐76,116 — Tesseract open-source OCR engine
• binhnguyennus/awesome-scalability ⭐73,459 — Patterns of scalable, reliable, performant large-scale systems
• OpenBB-finance/OpenBB ⭐72,154 — Open data platform for analysts, quants, and AI agents
• labmlai/annotated_deep_learning_paper_implementations ⭐67,331 — 60+ deep learning paper implementations with side-by-side notes
• scikit-learn/scikit-learn ⭐67,010 — Machine learning in Python

3. 24-Hour Star Growth (vs. Yesterday's Snapshot)

Note: this table is computed by differencing agents-radar's own star snapshots from two consecutive days, so it reflects genuine 24h growth. Only projects that appeared in yesterday's digest are included.

How to read this: four of the top five are "agent capability / context supply" projects — career-ops (agents applied to job search), hermes-agent, graphify (codebase to knowledge graph), and ECC (harness optimization); only firecrawl is pure data-fetch infrastructure. New attention is flowing toward "how do I actually use an agent" and "what context do I feed it," not toward the models themselves. By contrast pytorch (+26), transformers (+29), milvus (+8), and qdrant (+9) are growing an order of magnitude slower.

4. Trend Summary

• Codex and Claude Code both trended today, with Codex on top. Combined with the top HN story and the 343👍 complaint issue on the Claude Code side, this is today's most notable positional shift — not one board's noise, but stars, discussion, and sentiment pointing the same direction at once.
• The skills/harness meta-layer is now its own crowded category. Three of the top four (mattpocock/skills, affaan-m/ECC, obra/superpowers), plus a single CLAUDE.md at #11 and cursor/plugins at #15. When codified experience alone can trend, model capability is no longer the bottleneck — constraining the model is.
• Cost engineering has surfaced. sub2api (subscription relay and pooling) + free-for-dev (free tiers) + headroom (compress tokens before the LLM, 20–95% savings) + DeepSeek-Reasonix (engineered around prefix-cache stability) — four projects attacking the same bill from four angles, the demand-side counterpart to GPT-5.6 Sol's 20% price cut on the supply side.
• The memory/context lane keeps thickening. claude-mem (+89), graphify (+405), mem0, cognee, memvid, LEANN, and zvec all appear together, and graphify's growth ranks third overall. Same root as "performance debt = redundant rebuilds" in the OpenClaw digest and "context compaction" on the CLI side: context is the most expensive resource right now, and tooling around it is accelerating.
• Security governance trends for a second day. Tencent/AI-Infra-Guard (Agent Scan / Skills Scan / MCP scan) echoes Zeroclaw's identity-infrastructure rework, Moltis's fail-closed hook, and CoPaw's fifth security layer — the consensus that "agents hold real execution power" has spread from inside frameworks out to the tooling ecosystem.
• Traditional ML infrastructure growth is flattening. pytorch +26 / transformers +29 / milvus +8 / qdrant +9 versus +400-scale growth in agent-layer projects — attention and new interest are concentrated almost entirely at the application and harness layers.

This digest is generated by agents-radar. All content is fetched live from public GitHub data and written by the executing agent — no external LLM was called.

生成时间: 2026-08-23 02:30 (GMT+8) 数据来源：HN Algolia Search API（6 个查询：AI / LLM / Claude / OpenAI / Anthropic / machine learning，近 24h，去重后按 points 排序 Top 30；本批次共 75 条 unique 故事） 本日报由 agents-radar 本地自动化生成，所有内容由执行代理人工撰写，未调用任何外部大模型 API。

1. AI 编程智能体：Codex 与 Claude 的此消彼长

• Codex 使用一周速写：用 Codex 多过 Claude (94👍 / 104💬) — speckx 一篇「用 Codex 多于 Claude 的一周」体验文登顶本批次，评论破百。这是 CLI 日报与 Trending 榜「Codex 反超 Claude Code」（Trending #1）叙事在 HN 侧的强佐证：社区对 Codex 的实战评价正在快速升温。
• Anthropic 疑似在 Claude Code 中 A/B 测试「降低 effort 级别」 (51👍 / 45💬) — argofowl 的推文引爆社区，用户质疑「是否悄悄降级了模型投入」。与 Codex 的势头形成对照，Claude Code 正面临「投入缩水」的信任拷问。
• 把 Claude Code 变成 24/7 智能体 (3👍 / 1💬) — gtapps 开源项目，让 Claude Code 常驻运行，呼应「智能体全天候化」主线。
• 我们用一个 Claude Code skill 替换了整套 Grafana 栈 (3👍 / 0💬) — Frigade 的实践文，展示「用单个 skill 收敛可观测性工具链」的极简思路。

2. OpenAI 降价与资本动作

• GPT-5.6 Sol 降价 20% (83👍 / 76💬) — OpenAI 开发者文档更新模型定价，引发热烈讨论（76 条评论），本批次第二高热度。
• OpenAI 将前沿 GPT-5.6 Sol 模型的开发者定价下调超 20% (35👍 / 3💬) — Reuters 报道，官方动作获主流媒体确认。
• OpenAI：GPT-5.6 Sol 的 API 与额度定价下调超 20% (10👍 / 5💬) — OpenAI 官方推文，与文档更新、Reuters 报道三处呼应，降价信号一致且密集。
• The Instant 团队加入 OpenAI (13👍 / 8💬) — instantdb 公告，人才向头部实验室聚拢的又一例。

3. Anthropic：IPO、安全能力与版权

• 将 Claude Mythos 5 的网络安全能力带给更多防御者 (47👍 / 51💬) — Anthropic 官方博客，把前沿模型的安全能力下沉到更多安全团队，讨论度可观。
• Anthropic IPO 文件将把「AI backlash」列为风险因素 (32👍 / 51💬) — CNBC 援引消息人士，AI 反弹首次作为上市风险因素被正式写入文件，监管与舆论风险进入资本叙事。
• Carlsen 起诉 OpenAI 侵权复制「NEINhorn」 (4👍 / 0💬) — heise 报道，版权诉讼继续蔓延至棋类内容生成。
• 当 AI 艺术「没有作者」：生成图常常无法溯源至训练数据（研究） (3👍 / 0💬) — MIT News 报道的研究，直指生成式图像的版权溯源困境。
• Claude 提供「Concise」输出风格 (4👍 / 1💬) — code.claude.com 文档，模型新增精简输出档位，回应社区对冗长回复的审美疲劳。

4. Show HN 与开发者工具

• Show HN: OzBrain —— 智能体与团队之间的共享「大脑」 (72👍 / 45💬) — dariusmonsef 的项目，在智能体与团队知识间做共享层，本批次第三高热度，呼应「智能体记忆/知识复用」主线。
• Show HN: 用私有本地 AI 查找并整理照片 (3👍 / 0💬) — 本地隐私优先的图像管理工具。
• Show HN: Riyadh Tech Week —— 沙特科技生态全年导览 (3👍 / 0💬) — 区域性科技活动聚合站。
• 用单个开源 LLM 替换 223 节点的智能体图 (3👍 / 1💬) — netic.ai 的实践，主张「少而精的单模型」优于庞杂的多智能体编排。

5. 成本、可信与安全治理

• 给 LLM 你的生产数据库很容易，收回权限才是难点 (4👍 / 5💬) — deepsql.ai 博客，点出「授权易、回收难」的运维现实，与 agents 日报的「权限治理」主线呼应。
• AI 智能体合并了两种曾被分开的危险：触达与判断 (3👍 / 1💬) — syntheticauth 的观点文，讨论智能体同时握有「执行触达」与「自主判断」后的新风险面。
• 一年 LLM Serving：负载演化、缓存与负载均衡 (4👍 / 0💬) — arXiv 论文，系统侧讨论推理服务的工程演进。
• AI 代理与机器人账号（HotCRP） (3👍 / 0💬) — 学术会议系统 HotCRP 谈论文评审中的 AI 代理/机器人账号治理。

6. 社会、学术与其他

• 为好莱坞创意工作者「掘墓」：他们正在训练 AI 来做自己的工作 (46👍 / 61💬) — Guardian 深度报道，创意劳动者用 AI 自我替代的悖论，本批次讨论量第四。
• 80% 开发者觉得 AI 编程比「有用」更「上瘾」 (9👍 / 0💬) — ZDNet 调查，AI 编程工具的黏性与依赖性引发关注。
• Ask HN: 关于 AI 股市泡沫，证据有哪些？ (9👍 / 7💬) — 社区对 AI 估值泡沫的审慎讨论。
• 新人类基因组几乎无缺口 (6👍 / 0💬) — C&EN 报道首个近乎完整的二倍体人类基因组。
• LLVM 23 的编译期改进 (3👍 / 0💬) — 编译器底层性能演进。
• 湖泊喷发（Limnic Eruption） (3👍 / 0💬) — 科普条目。
• C++26 的 std::hive 有多快 (3👍 / 1💬) — 性能基准实测。
• 个性化语言学习 (3👍 / 0💬) — 语言学习产品。
• 空调在美国与日本几乎普及，但其他地区不是 (3👍 / 1💬) — Our World in Data 的数据洞察。

趋势小结

• Codex 反超 Claude 成为本日最强主线：HN 登顶文「用 Codex 多过 Claude 的一周」（94👍/104💬）与 Trending 榜 Codex #1、Claude Code #8 互相印证；而 Anthropic 被曝「A/B 测试降低 Claude Code effort」（51👍/45💬）则给这股势头添了信任阴影。
• 降价密集但叙事分裂：OpenAI GPT-5.6 Sol 在文档、官方推文、Reuters 三处同步降价超 20%（83👍 领跑），但「AI 股市泡沫」(Ask HN 9👍/7💬) 与「80% 开发者觉得上瘾而非有用」(9👍) 显示社区对估值与依赖性的双重保留。
• 智能体治理进入实操层：从「给 LLM 生产库易收回难」(4👍)、HotCRP 的机器人账号治理、到「智能体合并触达与判断两种危险」，权限与身份治理正从口号走向工程细节。
• 版权与 IP 持续升温：Carlsen 诉 OpenAI、MIT「AI 艺术无作者」研究、好莱坞创意者训练 AI 自替代——创作权属与替代焦虑交织。

本日报由 agents-radar 自动生成，所有内容基于 HN Algolia API 实时抓取并由执行代理人工撰写，未调用任何外部大模型。

Generated: 2026-08-23 02:30 (GMT+8) Source: HN Algolia Search API (6 queries: AI / LLM / Claude / OpenAI / Anthropic / machine learning, last 24h, deduped, Top 30 by points; 75 unique stories this batch) This digest is produced by the agents-radar local automation. All writing is done by the executing agent — no external LLM API was called.

1. AI Coding Agents: Codex vs. Claude

• Quick impressions: A week of using Codex more than Claude (94👍 / 104💬) — speckx's write-up of a week leaning on Codex over Claude tops this batch with 100+ comments. It's the HN-side confirmation of the "Codex overtaking Claude Code" narrative seen in today's CLI and Trending digests (Codex #1 on Trending) — community field reports on Codex are heating up fast.
• Anthropic appears to be A/B testing reduced effort levels in Claude Code (51👍 / 45💬) — argofowl's tweet went off, with users suspecting a "quiet downgrade" of model effort. Against Codex's momentum, Claude Code now faces a trust question about "shrinking投入".
• Turn Claude Code into a 24/7 Agent (3👍 / 1💬) — gtapps' OSS project keeps Claude Code running around the clock, echoing the "always-on agent"主线.
• We replaced our Grafana stack with a single Claude Code skill (3👍 / 0💬) — Frigade's field report showing how a single skill can collapse an entire observability toolchain.

2. OpenAI Pricing Cuts & Talent Moves

• GPT 5.6 Sol 20% price reduction (83👍 / 76💬) — OpenAI's developer docs updated model pricing, sparking a lively 76-comment thread — the second-hottest item this batch.
• OpenAI cuts developer pricing for frontier GPT-5.6 Sol model by more than 20% (35👍 / 3💬) — Reuters coverage confirming the official move.
• OpenAI: We're dropping API and credit pricing of GPT-5.6 Sol by over 20% (10👍 / 5💬) — OpenAI's official tweet, tri-synced with the docs update and Reuters — a consistent, dense price-cut signal.
• The Instant team joins OpenAI (13👍 / 8💬) — instantdb announcement, another case of talent consolidating toward the top labs.

3. Anthropic: IPO, Safety, & Copyright

• Bringing the cybersecurity capabilities of Claude Mythos 5 to more defenders (47👍 / 51💬) — Anthropic's official blog pushing frontier-model safety capabilities down to more security teams; solid discussion.
• Anthropic IPO filing will show AI backlash as a risk factor, sources say (32👍 / 51💬) — CNBC, citing sources, reports "AI backlash" being formally written into the IPO risk section for the first time — regulatory and public-opinion risk entering the capital narrative.
• Rights-infringing copies of "NEINhorn": Carlsen sues OpenAI (4👍 / 0💬) — heise reports copyright litigation spreading to chess-content generation.
• When AI art has no author: Generated images can't trace to training data (study) (3👍 / 0💬) — MIT News on a study exposing the provenance gap in generative imagery.
• Claude has a "Concise" output style (4👍 / 1💬) — code.claude.com docs add a concise output tier, answering community fatigue with verbose replies.

4. Show HN & Developer Tools

• Show HN: OzBrain, a shared brain for knowledge between agents and your team (72👍 / 45💬) — dariusmonsef's project building a shared knowledge layer between agents and teams; third-hottest this batch, echoing the "agent memory / knowledge reuse"主线.
• Show HN: Find and Organize Photos with Private, Local AI (3👍 / 0💬) — A local, privacy-first image-management tool.
• Show HN: Riyadh Tech Week – a year-round guide to Saudi Arabia's tech ecosystem (3👍 / 0💬) — A regional tech-event aggregator.
• Replacing 223-node agent graph with a single OSS LLM (3👍 / 1💬) — netic.ai argues a "small, sharp single model" beats sprawling multi-agent orchestration.

5. Cost, Trust & Security Governance

• Giving an LLM your prod database is easy. Taking access away is the hard part (4👍 / 5💬) — deepsql.ai points at the "grant-is-easy, revoke-is-hard" ops reality, resonating with the agents digest's "permission governance"主线.
• AI agents merged two dangers once kept separate: reach and judgment (3👍 / 1💬) — syntheticauth on the new risk surface when agents hold both "execution reach" and "autonomous judgment".
• A Year in LLM Serving: Workload Evolution, Caching and Load-Balancing (4👍 / 0💬) — arXiv paper on the systems-side evolution of inference serving.
• Hotcrp.com – AI agents and bot accounts (3👍 / 0💬) — Conference system HotCRP on governing AI agents / bot accounts in paper review.

6. Society, Academia & Misc

• Digging the grave of my skills: Hollywood creatives training AI to do their jobs (46👍 / 61💬) — Guardian long-read on the paradox of creative laborers training AI to replace themselves; fourth by discussion volume.
• 80% of developers find AI coding more addictive than helpful (9👍 / 0💬) — ZDNet survey flags stickiness and dependency on AI coding tools.
• Ask HN: What is the evidence for a stock market bubble in AI? (9👍 / 7💬) — Community caution on AI valuation froth.
• New human genome has almost no gaps (6👍 / 0💬) — C&EN on the first near-complete diploid human genome.
• the rest (≤3👍): LLVM 23 compile-time improvements (3👍), Limnic Eruption (3👍), How fast is C++26's std::hive? (3👍/1💬), Personalized Language Learning (3👍), Air Conditioning is almost universal in US/Japan but not elsewhere (3👍/1💬).

Trend Takeaways

• Codex overtaking Claude is the day's strongest thread: the HN-topping "week of using Codex more than Claude" (94👍/104💬) cross-confirms Trending's Codex #1 / Claude Code #8; meanwhile Anthropic's alleged "reduced effort A/B test in Claude Code" (51👍/45💬) casts a trust shadow over that momentum.
• Dense price cuts, split narrative: OpenAI's GPT-5.6 Sol cut >20% syncs across docs, official tweet, and Reuters (83👍 leads), yet "AI stock bubble" (Ask HN 9👍/7💬) and "80% find AI coding addictive not helpful" (9👍) show dual community reservations on valuation and dependency.
• Agent governance goes operational: from "prod DB grant-is-easy/revoke-is-hard" (4👍) to HotCRP's bot-account governance to "agents merge reach + judgment" — permission and identity governance are moving from slogans to engineering细节.
• Copyright & IP keep heating up: Carlsen v. OpenAI, MIT's "AI art has no author" study, and Hollywood creatives training AI to self-replace — authorship and displacement anxiety intertwine.

This digest is generated by agents-radar. All content is fetched live from the HN Algolia API and written by the executing agent — no external LLM was called.

生成时间: 2026-08-22 02:30 (GMT+8) | 覆盖工具: 7 个 CLI + Claude Code Skills 本日报由 agents-radar 本地自动化生成，所有摘要/对比/撰写均由执行代理人工完成，未调用任何外部大模型 API。

• Claude Code
• OpenAI Codex
• Gemini CLI
• GitHub Copilot CLI
• Kimi Code CLI
• OpenCode
• Qwen Code
• Claude Code Skills

横向对比

AI CLI 工具社区动态横向对比分析报告（2026-08-22）

1. 生态全景

今日 AI CLI 赛道呈现「两端分化」：一端是 OpenCode（78 PR）与 Qwen Code（82 PR） 的社区贡献洪流，另一端是 Claude Code 与 Kimi Code 的近静默——Claude Code 近 24h 0 条 PR 合入、Kimi Code 仅 1 条 issue。OpenAI Codex 则集中在 alpha 系列快速迭代（rust-v0.150.0-alpha.x 与 v0.149.0-alpha.x 于 08-21 连发 5 个预发布），并伴随高赞功能诉求。

共性议题收口到三条主线：记忆持久化（Claude #34556 跨压缩记忆、OpenCode #41469「空响应被记为干净停止」）、Windows/跨平台质量（Codex #39136 插件初始化失败、#27117 PSModulePath 污染、#37104 WSL 终端静默失败）、以及 provider 广度与成本透明（Qwen #8368 接入 Kimi/MiMo、#8276 跨延迟工具发现的 prompt cache 保留；CoPaw/NanoClaw/IronClaw 均在扩展 provider 矩阵）。

2. 各工具活跃度对比（近 24h）

3. 共同关注的功能方向

• 记忆与上下文压缩：Claude Code #34556「跨压缩持久记忆」仍居讨论热度前列（111💬，已关闭但被反复引用）；OpenCode #41469「空 LLM 响应导致会话静默停止（finish: unknown, 0 tokens）」是直接污染评测与计费的质量黑洞。两者共同指向「压缩后历史可回溯 + 失败可诊断」的缺口。
• Windows / 跨平台质量：Codex 的 #39136 内置浏览器插件 Trusted RPC 依赖不在可信路径（82💬）、#27117 Windows 独立更新把 PSModulePath 带进 powershell.exe、#37104 WSL 集成终端在 PTY 启动前静默失败，构成 Windows 第一痛点平台；Copilot CLI #4521「沙箱无法禁用」、Kimi #2615 后台 subagent 在 TaskStop 后仍持续调用，反映终端/进程生命周期治理仍是共性难题。
• Provider 广度与模型路由：Qwen #8368 新增 Kimi 与小米 MiMo provider、#8276 跨「延迟工具发现」保留 prompt cache；NanoClaw 接入 Dial（SMS+语音）、Mattermost；CoPaw 接入 Volcengine Agent Plan & MiMo V2.5；IronClaw 推可插拔 memory over MCP。多模型、多渠道已成各工具差异化主轴。
• Agent 可靠性与可观测：Gemini #22323 subagent 在 MAX_TURNS 后被误报为 GOAL 成功、#21409 generalist agent 挂起；OpenCode #785「能否关闭流式模式」（38👍/31💬）与 #34473「随机停止响应」反映流式/中断语义不稳定。

4. 差异化定位

• Claude Code：主干进入稳定打磨期，近 24h PR 侧静默、issue 侧以历史问题被 bot/标签刷新为主；最新 v2.1.238（08-20）延续「跨会话记忆」「多会话协作」主线（#34556、#24798 跨会话通信）。定位专业开发者的日常主力终端，体验细节被零容忍审视。
• OpenAI Codex：Rust 重写持续推进，alpha 系列日更；社区最高赞诉求是 #11626「/rewind 检查点恢复（同时回滚聊天上下文与已应用代码）」（204👍）。Windows 质量（冻结、插件、安装包）与「本地/私有插件市场被 API Key 用户阻挡」（#20621）是核心矛盾。定位 OpenAI 全场景入口。
• Gemini CLI：PR 侧极静（仅 2 条 pr-generation 相关），issue 集中在许可校验（#28912「无有效许可证」）、subagent 信任（#22323 误报成功）与「未充分利用 skills/sub-agents」（#21968）。成熟度中等，进入功能深化前的信任建设期。
• GitHub Copilot CLI：企业受管语境最明确，#3282/#3709「多 BYOK 模型、单会话内切换模型」（各 26👍）是最高频诉求；#4521「沙箱无法禁用」仍是治理痛点。近期无新发，定位企业受管环境的 Copilot 入口。
• Kimi Code CLI：今日仅 1 条 issue（#2615 后台 subagent 在 TaskStop/timeout 后仍调用 LLM），生态处于早期静默期，外部声量最低。
• OpenCode：社区贡献极活跃（78 PR），v1.18.21/20（08-21）持续推进；高赞诉求 #785「关闭流式模式」、#24153「归档会话的 unarchive/restore」。定位追求最新模型与高度可定制的开发者，provider 广度（llmgateway、QwenCloud、Vertex 等）持续扩展。
• Qwen Code：工业化 PR 体系成熟（82 PR），#8368 接入 Kimi/MiMo、#8332 附件音频桥、#8276 保留 prompt cache、#8992 MCP 2026 core + WebShell Apps host。定位验证严谨性与中国 IM/云生态用户。

5. 值得关注的趋势信号

• 贡献者驱动 vs 主干静默分化加剧：OpenCode/Qwen 由社区 PR 洪流推动，Claude Code/Kimi 主干近乎静默，说明「开放贡献通道」正成为生态活跃度的分水岭。
• 记忆持久化是下一轮分水岭：跨压缩记忆（Claude）、空响应误判（OpenCode）、自动记忆无限重试（Gemini #26522）三者都指向「失败可诊断、历史可回溯」的统一命题。
• Windows 仍是未被充分满足的硬市场：Codex 插件/安装/路径、Copilot 沙箱、Kimi 进程生命周期——后来者把 Windows 质量当一等公民即是切口。
• provider 矩阵军备竞赛：Kimi、MiMo、Volcengine、Dial、Mattermost 等接连被接入，多模型/多渠道正从「特色」变为「基线」。

各工具详细报告

Claude Code Skills 社区热点

数据来源: anthropics/skills | 近 24h 活跃

• #1620 / #1625 Shell 命令注入：with_server.py 的 --server 参数使用 shell=True（#1620 已关闭，#1625 开放）— 技能元数据/示例脚本的安全隐患。
• #1624 claude-api 的 SKIP grep 未锚定——cohere 匹配 "coherent"，在含相关词的仓库中误禁用技能（开放）。
• #1622 claude-api skill 的 description 超过 1024 字符 frontmatter 上限（1077 字符）（开放）。
• #1621 / #1619 MCPConnectionSSE/HTTP 与 MCPConnectionStdio 接受调用方提供的任意 URL/headers/command（开放）— MCP 连接器的供应链安全风险。
• #1627 feat：新增 buffer-api Agent Skill（Buffer GraphQL 调度）（开放）。
• #1623 更新 claude-api skill：Python SDK 0.x → 1.x 升级指南（已关闭）。

Skills 生态今日节奏聚焦安全审计（命令注入、MCP 任意命令执行）与单点修复，无大规模架构讨论。

本日报由 agents-radar 自动生成，所有内容基于 GitHub 公开 API 实时抓取并由执行代理人工撰写，未调用任何外部大模型。

Generated: 2026-08-22 02:30 (GMT+8) | Coverage: 7 CLIs + Claude Code Skills This digest is produced by the agents-radar local automation. All summaries, comparisons, and writing are done by the executing agent — no external LLM API was called.

• Claude Code
• OpenAI Codex
• Gemini CLI
• GitHub Copilot CLI
• Kimi Code CLI
• OpenCode
• Qwen Code
• Claude Code Skills

Cross-Tool Comparison

AI CLI Tools Cross-Tool Comparison Report (2026-08-22)

1. Ecosystem Overview

Today the AI CLI track shows a clear "two-speed" split: on one end, OpenCode (78 PRs) and Qwen Code (82 PRs) ride a wave of community contributions; on the other, Claude Code and Kimi Code are nearly silent — Claude Code merged 0 PRs in the last 24h and Kimi Code logged a single issue. OpenAI Codex concentrated on a rapid alpha cadence (five pre-releases of rust-v0.150.0-alpha.x and v0.149.0-alpha.x on 08-21) alongside a highly-upvoted feature request.

Three cross-cutting themes converge: memory persistence (Claude #34556 cross-compaction memory, OpenCode #41469 "empty response miscounted as clean stop"), Windows / cross-platform quality (Codex #39136 plugin init failure, #27117 PSModulePath pollution, #37104 WSL terminal silent failure), and provider breadth & cost transparency (Qwen #8368 adds Kimi/MiMo, #8276 preserves prompt cache across deferred tool discovery; CoPaw/NanoClaw/IronClaw all expanding provider matrices).

2. Activity Comparison (last 24h)

3. Shared Functional Directions

• Memory & context compaction: Claude #34556 "persistent memory across context compactions" stays top-of-mind (111 comments); OpenCode #41469 "session silently stops on empty LLM response (finish: unknown, 0 tokens)" is a quality black hole that pollutes evaluation and billing. Both point to the gap of "compression-history replayable + failures diagnosable."
• Windows / cross-platform quality: Codex #39136 built-in browser plugin Trusted RPC not in trusted path (82 comments), #27117 Windows standalone update inherits PSModulePath into powershell.exe, #37104 WSL integrated terminal silently fails before PTY startup — Windows is the #1 pain platform. Copilot CLI #4521 "sandbox cannot be disabled" and Kimi #2615 background subagent keeps calling LLM after TaskStop reflect shared terminal/process-lifecycle governance pain.
• Provider breadth & model routing: Qwen #8368 adds Kimi & Xiaomi MiMo providers, #8276 preserves prompt cache across deferred tool discovery; NanoClaw adds Dial (SMS+voice) and Mattermost; CoPaw adds Volcengine Agent Plan & MiMo V2.5; IronClaw pushes pluggable memory over MCP. Multi-model, multi-channel is now a differentiator axis.
• Agent reliability & observability: Gemini #22323 subagent misreported as GOAL success after MAX_TURNS, #21409 generalist agent hangs; OpenCode #785 "disable streaming mode" (38👍/31💬) and #34473 "randomly stops responding" show unstable streaming/interrupt semantics.

4. Differentiated Positioning

• Claude Code: main branch in a polish phase — 0 PRs merged in 24h, issues dominated by bot/label refreshes of historical items; latest v2.1.238 (08-20) continues the "cross-session memory" and "multi-session collaboration" thread (#34556, #24798). Positioned as the daily driver terminal for professional developers; UX details scrutinized zero-tolerance.
• OpenAI Codex: Rust rewrite accelerating, daily alpha cadence; top community ask is #11626 "/rewind checkpoint restore (revert both chat context and Codex-applied code)" (204👍). Windows quality (freeze, plugin, installer) and "local/private plugin marketplaces blocked for API-key users" (#20621) are core tensions. Positioned as OpenAI's all-scenario entry point.
• Gemini CLI: PR side extremely quiet (only 2 pr-generation-related); issues cluster on license validation (#28912 "no valid license"), subagent trust (#22323 false GOAL success) and "under-uses skills/sub-agents" (#21968). Mid-maturity, entering a trust-building phase before feature deepening.
• GitHub Copilot CLI: most enterprise-managed context; #3282/#3709 "multiple BYOK models, switch within a session" (26👍 each) is the top ask; #4521 "sandbox cannot be disabled" remains a governance pain. No new release recently; positioned as the Copilot entry point for managed environments.
• Kimi Code CLI: only 1 issue today (#2615 background subagent keeps calling LLM after TaskStop/timeout); ecosystem in early silence, lowest external voice.
• OpenCode: highly active community (78 PRs), v1.18.21/20 (08-21) ongoing; top asks #785 "disable streaming mode", #24153 "unarchive/restore archived sessions". Positioned for developers chasing latest models and deep customization; provider breadth (llmgateway, QwenCloud, Vertex) keeps expanding.
• Qwen Code: mature industrial PR pipeline (82 PRs); #8368 adds Kimi/MiMo, #8332 audio bridge for attachments, #8276 preserves prompt cache, #8992 MCP 2026 core + WebShell Apps host. Positioned for verification rigor and China IM/cloud-ecosystem users.

5. Notable Trend Signals

• Contributor-driven vs main-branch silence divergence widens: OpenCode/Qwen propelled by a PR flood while Claude Code/Kimi main branch is near-silent — "open contribution channel" is becoming the watershed of ecosystem vibrancy.
• Memory persistence is the next watershed: cross-compaction memory (Claude), empty-response miscount (OpenCode), auto-memory infinite retry (Gemini #26522) all point to the unified命题 of "failures diagnosable, history replayable."
• Windows remains an unsatisfied hard market: Codex plugin/installer/path, Copilot sandbox, Kimi process lifecycle — treating Windows quality as a first-class citizen is the opening for latecomers.
• Provider-matrix arms race: Kimi, MiMo, Volcengine, Dial, Mattermost successively integrated; multi-model/multi-channel is shifting from "feature" to "baseline."

Per-Tool Reports

Claude Code Skills Hot Topics

Source: anthropics/skills | last 24h

• #1620 / #1625 Shell command injection: with_server.py --server uses shell=True (#1620 closed, #1625 open) — security hazard in skill metadata/example scripts.
• #1624 claude-api SKIP grep unanchored — cohere matches "coherent", disabling the skill in repos containing the word (open).
• #1622 claude-api skill description exceeds the 1024-char frontmatter cap (1077 chars) (open).
• #1621 / #1619 MCPConnectionSSE/HTTP and MCPConnectionStdio accept caller-supplied arbitrary URL/headers/command (open) — supply-chain security risk in MCP connectors.
• #1627 feat: add buffer-api Agent Skill (Buffer GraphQL scheduling) (open).
• #1623 Update claude-api skill: Python SDK 0.x → 1.x upgrade guide (closed).

The Skills ecosystem today focuses on security auditing (command injection, MCP arbitrary command exec) and point fixes; no large architectural discussion.

This digest is generated by agents-radar. All content is fetched live from GitHub's public API and written by the executing agent — no external LLM was called.

生成时间: 2026-08-22 02:30 (GMT+8) | 覆盖: OpenClaw + 12 个 peer 项目 本日报由 agents-radar 本地自动化生成，所有摘要/对比/撰写均由执行代理人工完成，未调用任何外部大模型 API。

• OpenClaw
• NanoBot
• Zeroclaw
• PicoClaw
• NanoClaw
• NullClaw
• IronClaw
• LobsterAI
• TinyClaw
• CoPaw
• Moltis
• ZeptoClaw
• EasyClaw

跨生态对比

OpenClaw 跨生态对比分析报告（2026-08-22）

1. 生态全景

OpenClaw 本体今日 41 条 issue、59 个 PR 更新，主干聚焦会话目录刷新风暴治理、诊断恢复、安装策略告警、UI 决策回执；12 个 peer 中，IronClaw（31 PR）、CoPaw（36 PR）、Zeroclaw（95 PR）、NanoClaw（14 PR）、LobsterAI（13 PR）、Moltis（6 PR）贡献活跃，NullClaw / TinyClaw / ZeptoClaw 完全静默，EasyClaw 仅有版本发布无 issue/PR。整体趋势从「单 agent 聊天」走向多渠道 + 持久化记忆 + 自动化（automations/run-now）+ 企业级沙箱/鉴权。

2. 各项目活跃度对比（近 24h）

3. 共同关注方向

• 多渠道接入与消息可靠性：NanoBot（DingTalk 不观测/不排空后台任务 #5463、流式 provider 中段 server_error 跳过重试 #5454、Notion MCP 连接失败 #1168）；NanoClaw（send_card 文档承诺的回调按钮被 bridge 丢弃 #3426、Dial 短信+语音 #3041、Mattermost #3202）；Moltis（共享 Slack 频道工具失效 #1224、WhatsApp Markdown 渲染 #1220）。
• 记忆与上下文治理：OpenClaw 会话目录刷新风暴治理（#123535，154💬）、诊断恢复重叠工具调用（#124025，46💬）；IronClaw 可插拔 memory over MCP（#7664）；CoPaw ReMeLightMemoryCard reranker 面板（#6399）、embedding 健康检查超时硬编码（#7156）；Zeroclaw 上下文压缩锚定模型窗口比（#9535）。
• 自动化与企业级：IronClaw 从运行时证据派生 run 结果（#7650，45💬）、benchmark QA 预览（#7749，39💬）、omp core-tool 契约（#7491，37💬）、可插拔 memory（#7664）；CoPaw 每会话模型覆盖（#5992，9💬）；Moltis cron 尊重 heartbeat active_hours（#1208）。
• 安全与稳定性：OpenClaw 安装策略告警确认（#116489，49💬）、typecheck 挂死修复（#123975，45💬）、SSH 沙箱不暂存入站媒体（#112160）；Moltis 修复 WhatsApp/沙箱镜像校验（#1222、#1220）；CoPaw MCP 后端重启无法自动恢复（#6524）。
• 开发者体验 / 设计系统：IronClaw WebUI 设计系统（#7257，#7750 Storybook）；OpenClaw UI 决策回执说明（#126007）、新聊天 worktree 从干净默认启动（#126056）。

4. 差异化定位

• OpenClaw：生态中枢，主攻会话/目录治理、诊断恢复、安装策略与 UI 决策透明，企业受管语境。
• NanoBot：渠道广（Telegram/Matrix/Slack/WebUI），重 Dream 恢复游标与轨迹统一 provider 用量后端（#5479）。
• Zeroclaw：Rust 核心，RFC/ADR 驱动「更轻核心 + 外部集成」，wire protocol 一等公民（#8396）、canonical 沙箱策略 schema（#7821，20💬）。
• NanoClaw：多 provider 接入（Dial/WhatsApp/Mattermost/Matrix），强调模板创建 agent（#3396）与刷新安全的 ESM 补丁（#3403）。
• IronClaw：每用户沙箱 + 结构化 automations + 设计系统 + Telegram 关联，企业/团队向，CI 提速专题（T1-T4）。
• LobsterAI：cowork 文件卡片/分屏、设置面板搜索、i18n 硬编码修复，体验向。
• CoPaw (QwenPaw)：记忆 ReMe、自托管多用户、Volcengine Agent Plan & MiMo V2.5 provider（#6515），社区活跃度最高之一。
• Moltis：WhatsApp 为中心，重渠道安全（Vault/沙箱镜像校验）与 cron 调度。
• EasyClaw (rivonclaw)：高频小版本（每日多版 v1.8.109/108），商务开发筛选等垂直功能。
• PicoClaw / NullClaw / TinyClaw / ZeptoClaw：低活跃或静默。

5. 趋势信号

• 静默项目需关注：NullClaw / TinyClaw / ZeptoClaw 连续多日 0 动态，EasyClaw 仅发版无交互，生态健康度分化明显。
• 自动化成为一等公民：IronClaw/CoPaw 的 automations/run-now/Hub 标志 agent 从「对话」走向「可编排任务」。
• 安全与稳定性从可选到必需：安装策略告警确认（CWE 类风险）、typecheck 挂死、MCP 重启恢复成为高频修复。
• 记忆是跨项目主线：从 OpenClaw 会话治理到 IronClaw/CoPaw 记忆系统，持久化记忆仍是 2026 年 agent 框架竞争焦点。

OpenClaw 详细动态

数据来源: openclaw/openclaw | 近 24h（issue 41 / PR 59）

热点 Issues

1. #125626 Release validation: v2026.8.1-beta.2（18💬，开放）— 新版发布校验跟踪。
2. #124751 iOS app 在底部重复助手回复且不自动滚动到最新（6💬，开放）— 移动端渲染。
3. #113014 Webchat：原始 MEDIA: 指令泄漏到可见回复文本（5💬，开放）— 指令注入可见化。
4. #90243 feat(llm/google-vertex)：支持物理模型映射/别名以对接 Google Prior（3💬，开放）— 模型路由。
5. #124689 Model picker 仅对新会话生效；Ollama Cloud 请求登录（4💬，开放）。
6. #112160 SSH 沙箱不将入站媒体暂存到已有远程工作区（4💬，开放）。
7. #10944 Telegram 频道 parseMode 配置需求（4💬，开放）。
8. #126231 handler-timeout ingress 事件 attempts=0 静默丢消息（3💬，已关闭）。
9. #126813 排队的 peer 消息以 answer-expected 提示重唤 agent 且无确认（3💬，开放）。
10. #100911 Tasks 页面 / tasks.list 边界加固（2💬，开放）。

重要 PR 进展（按讨论热度）

1. #123535 fix(ui)：避免会话目录刷新风暴（154💬，开放）
2. #123979 fix(scripts)：build heap 忽略 systemd 内存预算而占满默认（56💬，开放）
3. #116489 feat(security)：安装策略告警需显式确认（49💬，已关闭）
4. #123975 fix(scripts)：tsgo 卡死时 typecheck 永久挂起（45💬，开放）
5. #124025 fix(diagnostics)：恢复过期重叠工具调用（46💬，开放）
6. #120900 feat(ui)：审查安装策略告警（44💬，已关闭）
7. #125815 fix(sessions)：保持 owner 分配稳定、回复可用（38💬，开放）
8. #126618 fix：Tool Search 目录/tools 用 tool_call 包裹原生 read/exec（26💬，开放）
9. #126007 feat(ui)：在 Activity 中解释决策回执（25💬，开放）
10. #126056 fix(ui)：新聊天 worktree 从干净默认启动（21💬，开放）

Peer 项目动态精选

NanoBot（HKUDS/nanobot）— I5 / P34

• Issues：#5198 特定会话改模型需重配（已关闭）、#1168 Notion MCP 连接失败（已关闭）、#5441 单点恢复的工具错误永久阻塞（已关闭）、#5463 DingTalk 不观测/不排空后台任务（开放）、#5454 流式 provider 中段 server_error 跳过重试（已关闭）。
• PRs：#5442 fix(dream) 恢复游标、#5379 fix(memory) 保留完整合并输入（开放）、#5476 渲染 LaTeX 为 Unicode（已关闭）、#2063 Tauri 桌面端（已关闭）、#5455 重试 Codex server_error（已关闭）、#5479 feat(trajectory) 统一 provider 用量后端（开放）。

Zeroclaw（zeroclaw-labs/zeroclaw）— I5 / P95

• Issues：#8396 RFC 使 wire protocol 成为 provider 构造一等公民（开放）、#8691 ADR 清单与已接受 RFC 决策记录（开放）、#10167 终端 agent 多路复用的 vendor-neutral 生命周期导出（开放）、#8288 SOP 里程碑 daemon 控制面 5/5（开放）、#10212 文档 switch 路由优先级（开放）。
• PRs：#8561 feat(channels/telegram) 多消息流式（33💬）、#7821 feat(security) canonical 沙箱策略 schema（20💬）、#9447 fix(anthropic) 分类不完整终端响应（19💬）、#9196 feat(mcp) 聚合预算预检 materialize 资源 blob（19💬）、#9772 feat(telegram) 共享群聊每用户会话开关（18💬）、#9535 feat(runtime) 上下文压缩锚定模型窗口比（17💬）。

NanoClaw（qwibitai/nanoclaw）— I1 / P14

• Issue：#3426 send_card 文档承诺的回调按钮被 bridge 自 #2265 起丢弃。
• PRs：#3041 feat(channels) 新增 Dial 通道（SMS+AI 语音，已关闭）、#3202 新增 Mattermost 集成（已关闭）、#3050 Dial 通道选择器（已关闭）、#3396 feat 从模板在聊天中创建 agent（开放）、#3424 ci 测试 registry-backed skills（已关闭）、#3403 fix(matrix) 刷新安全 ESM 补丁（已关闭）。

NullClaw（nullclaw/nullclaw）— I0 / P1

• PR：#990 feat(providers) 新增 Eden AI 作为 OpenAI 兼容网关（开放）。

IronClaw（nearai/ironclaw）— I12 / P31

• Issues：#7801 CI 提速 T4 规范预检（开放）、#7799 T2 nextest 流水线（开放）、#7800 T3 PR/队列收敛（开放）、#7798 T1 setup-rust 复合（开放）、#7664 可插拔 memory over MCP（开放）、#7783 LLM 超时策略 finalization 无法测 TTFT（已关闭）。
• PRs：#7650 feat(automations) 从运行时证据派生 run 结果（45💬）、#7749 benchmark QA 预览触发（39💬）、#7491 feat(coding) omp core-tool 契约+引擎+benchmark（37💬）、#7456 fix(reborn) 持久存储 profile 无关（12💬）、#7257 docs(design-system) WebUI 设计系统（11💬）、#7750 chore(webui) 集成 Storybook+设计系统目录（10💬）。

LobsterAI（netease-youdao/LobsterAI）— I2 / P13

• Issues：#1217 运行中偶发启动网关（已关闭）、#1223 CoworkPromptInput 硬编码中文标签导致英文提示词混入（已关闭）。
• PRs：#1218 重构定时任务排序（已关闭）、#1215 始终重建 chat handler（已关闭）、#1219 消除会话列表/详情无效重渲染（已关闭）、#1220 消除 recentChats 的 N+1 查询（已关闭）、#1224 修复 i18n 硬编码与 Escape 键（已关闭）、#1550 fix(scheduledTask) 投递模式「不通知」时去除 channel/to 字段（开放）。

TinyClaw（TinyAGI/tinyclaw）— I0 / P0

• 无近 24h 动态（最新 v0.0.20，03-26）。

CoPaw（agentscope-ai/CoPaw）— I34 / P36

• Issues：#6524 MCP 后端重启后客户端无法自动恢复（开放）、#6780 2.0.1 不使用时会卡死（已关闭）、#7016 工具调用 404（开放）、#7156 embedding 健康检查在已预热时仍超时且硬编码（开放）、#7206 v2.1.1-beta.1 /compact 因 pydantic ValidationError 失败（开放）、#7204 qwenpaw 如何增加自定义 tool（开放）。
• PRs：#5992 每会话模型覆盖（9💬）、#6515 feat(providers) 新增 Volcengine Agent Plan & MiMo V2.5（9💬）、#6845 保留 assistant 完成时间（已关闭）、#6581 ReMeLightMemoryCard reranker UI 面板（开放）、#6399 新增 reranker UI 配置（开放）、#6586 fix(mcp) 恢复过期 server 会话（已关闭）。

Moltis（moltis-org/moltis）— I2 / P6

• Issues：#1224 共享 Slack 频道工具停止工作（开放）、#1223 heartbeat active_hours 对默认配置无效（开放）。
• PRs：#1220 渲染 WhatsApp 出站 Markdown（8💬，已关闭）、#1208 修复 cron 尊重 heartbeat active_hours（4💬，开放）、#468 Windows 用 cmd.exe 执行 shell hooks（开放）、#1226 定时输出投递到源聊天（开放）、#1225 改进 zh-TW 繁体中文（开放）、#1222 校验沙箱镜像请求（开放）。

ZeptoClaw（qhkm/zeptoclaw）— I0 / P0

• 无近 24h 动态（最新 v0.9.2，04-07）。

EasyClaw（gaoyangz77/easyclaw）— I0 / P0 / 发版 v1.8.109、v1.8.108（08-21）

• 今日仅发版，无 issue/PR 交互。

本日报由 agents-radar 自动生成，所有内容基于 GitHub 公开 API 实时抓取并由执行代理人工撰写，未调用任何外部大模型。

Generated: 2026-08-22 02:30 (GMT+8) | Coverage: OpenClaw + 12 peer projects This digest is produced by the agents-radar local automation. All summaries, comparisons, and writing are done by the executing agent — no external LLM API was called.

• OpenClaw
• NanoBot
• Zeroclaw
• PicoClaw
• NanoClaw
• NullClaw
• IronClaw
• LobsterAI
• TinyClaw
• CoPaw
• Moltis
• ZeptoClaw
• EasyClaw

Cross-Ecosystem Comparison

OpenClaw Cross-Ecosystem Comparison Report (2026-08-22)

1. Ecosystem Overview

OpenClaw core shows 41 issues and 59 PR updates today, with the main branch focused on session-catalog refresh-storm governance, diagnostics recovery, install-policy warnings, and UI decision receipts. Among the 12 peers, IronClaw (31 PRs), CoPaw (36 PRs), Zeroclaw (95 PRs), NanoClaw (14 PRs), LobsterAI (13 PRs), and Moltis (6 PRs) are active, while NullClaw / TinyClaw / ZeptoClaw are completely silent and EasyClaw only ships releases with no issue/PR interaction. The overall trend moves from "single-agent chat" toward multi-channel + persistent memory + automation (automations/run-now) + enterprise-grade sandbox/auth.

2. Activity Comparison (last 24h)

3. Shared Directions

• Multi-channel access & message reliability: NanoBot (DingTalk doesn't observe/drain background tasks #5463, streaming provider mid-stream server_error skips retry #5454, Notion MCP failure #1168); NanoClaw (send_card doc-promised callback buttons dropped by bridge #3426, Dial SMS+voice #3041, Mattermost #3202); Moltis (shared Slack channel tools fail #1224, WhatsApp Markdown render #1220).
• Memory & context governance: OpenClaw session-catalog refresh-storm governance (#123535, 154💬), diagnostics recovery of overlapping tool calls (#124025, 46💬); IronClaw pluggable memory over MCP (#7664); CoPaw ReMeLightMemoryCard reranker panel (#6399), embedding health-check hardcoded timeout (#7156); Zeroclaw context-compaction anchored to model-window ratio (#9535).
• Automation & enterprise: IronClaw derive run outcomes from runtime evidence (#7650, 45💬), benchmark QA preview (#7749, 39💬), omp core-tool contract (#7491, 37💬), pluggable memory (#7664); CoPaw per-session model override (#5992, 9💬); Moltis cron honors heartbeat active_hours (#1208).
• Security & stability: OpenClaw install-policy warning acknowledgement (#116489, 49💬), typecheck hang fix (#123975, 45💬), SSH sandbox doesn't stage inbound media (#112160); Moltis WhatsApp/sandbox-image validation (#1222, #1220); CoPaw MCP backend restart can't auto-recover (#6524).
• Developer experience / design system: IronClaw WebUI design system (#7257, #7750 Storybook); OpenClaw UI decision-receipt explanation (#126007), new-chat worktree from clean defaults (#126056).

4. Differentiated Positioning

• OpenClaw: ecosystem hub — session/catalog governance, diagnostics recovery, install policy & UI decision transparency; enterprise-managed context.
• NanoBot: widest channels (Telegram/Matrix/Slack/WebUI); Dream recovery cursor & unified trajectory provider-usage backend (#5479).
• Zeroclaw: Rust core, RFC/ADR-driven "lighter core + external integrations"; wire protocol first-class (#8396), canonical sandbox-policy schema (#7821, 20💬).
• NanoClaw: multi-provider (Dial/WhatsApp/Mattermost/Matrix); template-based agent creation (#3396) & refresh-safe ESM patch (#3403).
• IronClaw: per-user sandbox + structured automations + design system + Telegram linkage; enterprise/team-oriented; CI-speedup track (T1–T4).
• LobsterAI: cowork file cards/split preview, settings search, i18n hardcoded-fix; UX-oriented.
• CoPaw (QwenPaw): memory ReMe, self-hosted multi-user, Volcengine Agent Plan & MiMo V2.5 provider (#6515); among the most active communities.
• Moltis: WhatsApp-centric, channel security (Vault/sandbox-image validation) & cron scheduling.
• EasyClaw (rivonclaw): high-frequency small releases (daily v1.8.109/108), vertical features like business-dev filtering.
• PicoClaw / NullClaw / TinyClaw / ZeptoClaw: low-activity or silent.

5. Trend Signals

• Silent projects need attention: NullClaw / TinyClaw / ZeptoClaw have had 0 activity for multiple days; EasyClaw ships only — ecosystem health is diverging.
• Automation becomes first-class: IronClaw/CoPaw automations/run-now/Hub mark agents moving from "conversation" to "orchestrated tasks."
• Security & stability from optional to mandatory: install-policy warning acknowledgement (CWE-class risk), typecheck hang, MCP restart recovery are now frequent fixes.
• Memory is the cross-project主线: from OpenClaw session governance to IronClaw/CoPaw memory systems, persistent memory remains the 2026 agent-framework battleground.

OpenClaw Detailed Activity

Source: openclaw/openclaw | last 24h (issue 41 / PR 59)

Hot Issues

1. #125626 Release validation: v2026.8.1-beta.2 (18💬, open) — new-version release validation tracker.
2. #124751 iOS app duplicates assistant replies at bottom and doesn't auto-scroll to latest (6💬, open) — mobile rendering.
3. #113014 Webchat: raw MEDIA: directive leaks into visible reply text (5💬, open) — instruction injection surfaced.
4. #90243 feat(llm/google-vertex): physical model mapping/aliasing for Google's Prior (3💬, open) — model routing.
5. #124689 Model picker only applies to new sessions; Ollama Cloud requests sign-in (4💬, open).
6. #112160 SSH sandbox doesn't stage inbound media into an existing remote workspace (4💬, open).
7. #10944 parseMode config for Telegram channels (4💬, open).
8. #126231 handler-timeout ingress events die with attempts=0 — silently lost messages (3💬, closed).
9. #126813 Queued peer message re-invokes an agent with answer-expected hint and no acknowledgement (3💬, open).
10. #100911 Tasks page / tasks.list edge-case hardening (2💬, open).

Notable PRs (by discussion heat)

1. #123535 fix(ui): avoid session catalog refresh storms (154💬, open)
2. #123979 fix(scripts): build heap ignores its systemd memory budget and takes the full default (56💬, open)
3. #116489 feat(security): require acknowledgement for install policy warnings (49💬, closed)
4. #123975 fix(scripts): typecheck hangs forever when tsgo wedges instead of failing (45💬, open)
5. #124025 fix(diagnostics): recover stale overlapping tool calls (46💬, open)
6. #120900 feat(ui): review install policy warnings (44💬, closed)
7. #125815 fix(sessions): keep owner assignment stable and replies working (38💬, open)
8. #126618 fix: Tool Search directory/tools wrap native read/exec in tool_call (26💬, open)
9. #126007 feat(ui): explain decision receipts in Activity (25💬, open)
10. #126056 fix(ui): start new chat worktrees from fresh defaults (21💬, open)

Peer Project Highlights

NanoBot (HKUDS/nanobot) — I5 / P34

• Issues: #5198 change models in a specific session requires reconfig (closed), #1168 Notion MCP connection fails (closed), #5441 a single recovered tool error permanently blocks (closed), #5463 DingTalk doesn't observe/drain inbound background tasks (open), #5454 streaming providers skip retry once content streamed mid server_error (closed).
• PRs: #5442 fix(dream) advance cursor when tool errors recovered, #5379 fix(memory) preserve full consolidation input (open), #5476 render LaTeX as Unicode (closed), #2063 Tauri desktop app (closed), #5455 retry Codex server_error (closed), #5479 feat(trajectory) unified provider usage backend (open).

Zeroclaw (zeroclaw-labs/zeroclaw) — I5 / P95

• Issues: #8396 RFC make wire protocol first-class in provider construction, #8691 ADR inventory and accepted RFC decision records, #10167 vendor-neutral lifecycle export for terminal agent multiplexers, #8288 SOP milestone daemon-owned SOP control plane to 5/5, #10212 document switch and its routing precedence.
• PRs: #8561 feat(channels/telegram) multi_message streaming mode (33💬), #7821 feat(security) canonical sandbox_policy schema with app-layer enforcement (20💬), #9447 fix(anthropic) classify incomplete terminal responses (19💬), #9196 feat(mcp) materialize resource blob with aggregate budget preflight (19💬), #9772 feat(telegram) per_user_session toggle for shared group-chat (18💬), #9535 feat(runtime) anchor context compaction to model window ratio (17💬).

NanoClaw (qwibitai/nanoclaw) — I1 / P14

• Issue: #3426 send_card docs promise callback buttons that the bridge drops since #2265.
• PRs: #3041 feat(channels) Dial channel adapter (SMS + AI voice, closed), #3202 Mattermost integration (closed), #3050 Dial channel picker (closed), #3396 feat create agents from templates in chat (open), #3424 ci test registry-backed skills (closed), #3403 fix(matrix) refresh-safe ESM patch (closed).

NullClaw (nullclaw/nullclaw) — I0 / P1

• PR: #990 feat(providers) add Eden AI as an OpenAI-compatible gateway (open).

IronClaw (nearai/ironclaw) — I12 / P31

• Issues: #7801 CI expedite T4 canonical preflight, #7799 T2 nextest pipeline, #7800 T3 PR/queue convergence, #7798 T1 setup-rust composite, #7664 pluggable memory over MCP, #7783 LLM timeout policy can't measure TTFT.
• PRs: #7650 feat(automations) derive run outcomes from runtime evidence (45💬), #7749 benchmark QA-preview trigger (39💬), #7491 feat(coding) omp core-tool contract + engines + benchmark arm (37💬), #7456 fix(reborn) make durable storage profile-agnostic (12💬), #7257 docs(design-system) WebUI design system (11💬), #7750 chore(webui) integrate Storybook + design-system catalog (10💬).

LobsterAI (netease-youdao/LobsterAI) — I2 / P13

• Issues: #1217 occasional gateway startup during run (closed), #1223 CoworkPromptInput hardcoded Chinese labels leak into English prompts (closed).
• PRs: #1218 refactor scheduled-task sort (closed), #1215 always rebuild chat handler (closed), #1219 eliminate invalid re-renders (closed), #1220 eliminate N+1 queries (closed), #1224 fix i18n hardcoded & Escape key (closed), #1550 fix(scheduledTask) drop channel/to when delivery mode is "no-notify" (open).

TinyClaw (TinyAGI/tinyclaw) — I0 / P0

• No activity in last 24h (latest v0.0.20, 03-26).

CoPaw (agentscope-ai/CoPaw) — I34 / P36

• Issues: #6524 MCP backend restart can't auto-recover (open), #6780 2.0.1 self-freezes when idle (closed), #7016 tool call 404 (open), #7156 embedding health-check times out when already warm, timeout hardcoded (open), #7206 v2.1.1-beta.1 /compact fails with pydantic ValidationError (open), #7204 how to add custom tool in qwenpaw (open).
• PRs: #5992 per-session model overrides (9💬), #6515 feat(providers) Volcengine Agent Plan & MiMo V2.5 (9💬), #6845 preserve assistant completion time (closed), #6581 ReMeLightMemoryCard reranker UI panel (open), #6399 reranker UI config (open), #6586 fix(mcp) recover stale server sessions (closed).

Moltis (moltis-org/moltis) — I2 / P6

• Issues: #1224 tools stop working in shared Slack channels (open), #1223 heartbeat active_hours has no effect on default config (open).
• PRs: #1220 render Markdown in outbound WhatsApp (8💬, closed), #1208 fix(cron) honor heartbeat active hours (4💬, open), #468 use cmd.exe on Windows for shell hooks (open), #1226 deliver scheduled output to originating chat (open), #1225 improve zh-TW Traditional Chinese (open), #1222 validate sandbox image requests (open).

ZeptoClaw (qhkm/zeptoclaw) — I0 / P0

• No activity in last 24h (latest v0.9.2, 04-07).

EasyClaw (gaoyangz77/easyclaw) — I0 / P0 / releases v1.8.109, v1.8.108 (08-21)

• Only releases today, no issue/PR interaction.

This digest is generated by agents-radar. All content is fetched live from GitHub's public API and written by the executing agent — no external LLM was called.

生成时间: 2026-08-22 02:30 (GMT+8) 数据来源：GitHub Trending（每日榜）+ gh search repos 6 个 AI 主题（按 stars 排序） 本日报由 agents-radar 本地自动化生成，所有撰写由执行代理人工完成，未调用任何外部大模型 API。

一、GitHub Trending 每日飙升榜（2026-08-22）

说明：本次抓取时 GitHub Trending 页面未对外暴露「当日新增 Star」数值（疑似未登录视图/版式变更），故下表仅列出入榜项目与简介，不含 star 增量。

观察：今日 Trending 高度向「Agent 工具链 + 个人效率 + 安全」收口——Skills 框架（mattpocock/skills、obra/superpowers、JuliusBrussee/caveman）、Agent 基础设施（agent-substrate/substrate、chaitanyagiri/munder-difflin、volcengine/OpenViking、Tencent）、以及 AI 红队（Tencent/AI-Infra-Guard）同屏出现，反映社区对「可复用 agent 能力 + 安全治理」的双重关注。

二、六大 AI 主题搜索热度（按 stars）

1. AI Agent 框架

• NousResearch/hermes-agent ⭐233,911 — The agent that grows with you
• shareAI-lab/learn-claude-code ⭐74,900 — 从 0 到 1 构建类 claude code 的 agent harness
• Panniantong/Agent-Reach ⭐73,818 — 给 AI agent 全网的眼睛（Twitter/Reddit/YouTube/GitHub/Bilibili/小红书）
• thedaviddias/Front-End-Checklist ⭐73,602 — 现代 Web 开发必备清单（人类与 AI agent 通用）
• santifer/career-ops ⭐67,306 — 开源 AI 求职，可跑在各 AI 编码 CLI
• ZhuLinsen/daily_stock_analysis ⭐63,578 — LLM 驱动的多市场股票智能分析
• CherryHQ/cherry-studio ⭐50,885 — AI 生产力工作室，统一前沿 LLM
• HKUDS/nanobot ⭐47,258 — 超轻量、自托管的个人 AI agent 平台
• zhayujie/CowAgent ⭐46,623 — 开源超级 AI 助手与 Agent Harness
• siyuan-note/siyuan ⭐45,917 — 开源、隐私优先、自托管的知识工作空间

2. 大模型（LLM）

• affaan-m/ECC ⭐241,701 — agent harness 性能优化系统（Skills/instincts/memory/security）
• NousResearch/hermes-agent ⭐233,911 — The agent that grows with you
• Significant-Gravitas/AutoGPT ⭐186,719 — 人人可用的自主 AI
• ollama/ollama ⭐179,115 — 一键跑起 Kimi-K2.6/GLM-5.2/MiniMax/DeepSeek/gpt-oss/Qwen/Gemma
• firecrawl/firecrawl ⭐170,460 — 大规模搜索/抓取/交互 Web 的上下文 API
• f/prompts.chat ⭐167,704 — Awesome ChatGPT Prompts 社区共享
• huggingface/transformers ⭐164,310 — 文本/视觉/音频/多模态 SOTA 模型框架
• langgenius/dify ⭐153,134 — Agentic 工作流与 RAG 管线协作平台
• open-webui/open-webui ⭐149,491 — 用户友好的 AI 界面
• langchain-ai/langchain ⭐144,726 — agent 工程平台

3. RAG 检索增强生成

• langgenius/dify ⭐153,134 — Agentic 工作流与 RAG 管线
• open-webui/open-webui ⭐149,491 — 用户友好的 AI 界面
• langchain-ai/langchain ⭐144,726 — agent 工程平台
• Shubhamsaboo/awesome-llm-apps ⭐133,487 — 100+ AI Agents/Skills/RAG 应用
• Graphify-Labs/graphify ⭐109,096 — 把代码库变成可查询知识图谱（局部 AST 解析，无向量库）
• thedotmack/claude-mem ⭐91,430 — 跨会话持久上下文（Claude Code/OpenClaw/Codex/Gemini 等）
• infiniflow/ragflow ⭐88,983 — 领先的开源 RAG 引擎
• PaddlePaddle/PaddleOCR ⭐88,065 — 轻量 OCR 工具包，桥接图像/PDF 与 LLM
• dair-ai/Prompt-Engineering-Guide ⭐77,664 — 提示工程/上下文工程/RAG 指南
• datawhalechina/hello-agents ⭐74,116 — 从零构建智能体教程

4. 向量数据库

• Mintplex-Labs/anything-llm ⭐65,017 — 本地优先的强力 agent 体验
• meilisearch/meilisearch ⭐59,045 — 极速搜索 API，AI 混合搜索
• pathwaycom/llm-app ⭐59,013 — RAG/AI 管线云模板
• run-llama/llama_index ⭐51,790 — 文档 agent 与 OCR 平台
• milvus-io/milvus ⭐45,728 — 高性能云原生向量数据库
• VectifyAI/PageIndex ⭐35,280 — 无向量的推理式 RAG 文档索引
• qdrant/qdrant ⭐34,115 — 大规模向量数据库与向量搜索
• topoteretes/cognee ⭐30,170 — 开源 AI 记忆平台（知识图谱引擎）
• NirDiamant/RAG_Techniques ⭐29,168 — RAG 高级技术笔记本
• weaviate/weaviate ⭐16,745 — 云原生向量数据库

5. LLM 模型

• jingyaogong/minimind ⭐54,909 — 2 小时从零训练 64M 参数 LLM
• ScrapeGraphAI/Scrapegraph-ai ⭐29,794 — 基于 AI 的爬虫
• The-Pocket/PocketFlow-Tutorial-Codebase-Knowledge ⭐12,626 — Codebase to Tutorial
• The-Pocket/PocketFlow ⭐11,120 — 100 行 LLM 框架，让 agent 构建 agent
• OpenSPG/KAG ⭐9,004 — 逻辑形式引导的推理与检索框架
• 0xPlaygrounds/rig ⭐8,347 — Rust 构建模块化可扩展 LLM 应用
• open-compass/opencompass ⭐7,325 — LLM 评测平台
• InternLM/InternLM ⭐7,267 — InternLM 系列官方发布
• Eigenwise/atomic-agents ⭐6,190 — 原子化构建 AI agent
• gluonfield/enchanted ⭐5,998 — iOS/macOS 私本地模型聊天

6. 机器学习

• tensorflow/tensorflow ⭐197,220 — 人人可用的开源机器学习框架
• f/prompts.chat ⭐167,704 — Awesome ChatGPT Prompts
• huggingface/transformers ⭐164,310 — SOTA 机器学习模型框架
• rasbt/LLMs-from-scratch ⭐103,171 — 从零逐步实现类 ChatGPT LLM
• pytorch/pytorch ⭐102,517 — 带 GPU 加速的张量与动态神经网络
• microsoft/ML-For-Beginners ⭐89,653 — 12 周 26 课经典机器学习
• Developer-Y/cs-video-courses ⭐83,139 — 带视频讲座的 CS 课程清单
• mlabonne/llm-course ⭐81,893 — LLM 学习路线与 Colab 笔记本
• netdata/netdata ⭐80,255 — 最快的 AI 驱动全栈可观测
• d2l-ai/d2l-zh ⭐79,853 — 《动手学深度学习》中英文版

三、趋势小结

• Agent 工具链与个人效率同屏爆发：Skills 框架、agent harness、自托管 agent 平台、AI 求职/股票分析等「把 agent 用起来」的项目占据多榜头部。
• 记忆与上下文成为基础设施赛道：volcengine/OpenViking、thedotmack/claude-mem、topoteretes/cognee、Graphify-Labs/graphify 等从不同角度切入「agent 记忆/知识」，与今日 CLI/OpenClaw 日报中「记忆持久化」主线呼应。
• 安全治理升温：再叠加 Tencent/AI-Infra-Guard 红队平台，社区对 agent 安全的关注从可选走向必需。
• 轻量/本地优先受捧：Rust 编写的 OpenLogi、本地优先的 anything-llm/siyuan，反映对「自托管、隐私、低依赖」的偏好。

本日报由 agents-radar 自动生成，所有内容基于 GitHub 公开数据实时抓取并由执行代理人工撰写，未调用任何外部大模型。

Generated: 2026-08-22 02:30 (GMT+8) Sources: GitHub Trending (daily board) + gh search repos across 6 AI topics (sorted by stars) This digest is produced by the agents-radar local automation. All writing is done by the executing agent — no external LLM API was called.

1. GitHub Trending Daily Board (2026-08-22)

Note: During this fetch the GitHub Trending page did not expose per-day "stars gained" figures (likely an unauthenticated view / layout change), so the table lists trending repos with descriptions but without star deltas.

Observation: Today's Trending converges on "agent toolchain + personal productivity + security" — Skills frameworks (mattpocock/skills, obra/superpowers, JuliusBrussee/caveman), agent infrastructure (agent-substrate/substrate, chaitanyagiri/munder-difflin, volcengine/OpenViking, Tencent), and AI red-teaming (Tencent/AI-Infra-Guard) appear together, reflecting dual community focus on reusable agent capability and security governance.

2. Six AI Topic Search Heat (by stars)

1. AI Agent Frameworks

• NousResearch/hermes-agent ⭐233,911 — The agent that grows with you
• shareAI-lab/learn-claude-code ⭐74,900 — Build a claude-code-like agent harness from 0 to 1
• Panniantong/Agent-Reach ⭐73,818 — Give your AI agent eyes to see the entire internet
• thedaviddias/Front-End-Checklist ⭐73,602 — Essential checklist for modern web dev (humans & AI agents)
• santifer/career-ops ⭐67,306 — Open-source AI job search, runs in AI coding CLIs
• ZhuLinsen/daily_stock_analysis ⭐63,578 — LLM-powered multi-market stock analysis
• CherryHQ/cherry-studio ⭐50,885 — AI productivity studio with 300+ assistants
• HKUDS/nanobot ⭐47,258 — Ultra-lightweight, self-hosted personal AI agent platform
• zhayujie/CowAgent ⭐46,623 — Open-source super AI assistant & Agent Harness
• siyuan-note/siyuan ⭐45,917 — Open-source, privacy-first, self-hosted knowledge workspace

2. LLM (Large Language Models)

• affaan-m/ECC ⭐241,701 — Agent harness performance optimization system
• NousResearch/hermes-agent ⭐233,911 — The agent that grows with you
• Significant-Gravitas/AutoGPT ⭐186,719 — Accessible AI for everyone
• ollama/ollama ⭐179,115 — Run Kimi-K2.6/GLM-5.2/MiniMax/DeepSeek/gpt-oss/Qwen/Gemma locally
• firecrawl/firecrawl ⭐170,460 — Context API to search, scrape, interact with the web at scale
• f/prompts.chat ⭐167,704 — Awesome ChatGPT Prompts community
• huggingface/transformers ⭐164,310 — SOTA ML model framework (text/vision/audio/multimodal)
• langgenius/dify ⭐153,134 — Agentic workflows & RAG pipelines platform
• open-webui/open-webui ⭐149,491 — User-friendly AI interface
• langchain-ai/langchain ⭐144,726 — Agent engineering platform

3. RAG (Retrieval-Augmented Generation)

• langgenius/dify ⭐153,134 — Agentic workflows & RAG pipelines
• open-webui/open-webui ⭐149,491 — User-friendly AI interface
• langchain-ai/langchain ⭐144,726 — Agent engineering platform
• Shubhamsaboo/awesome-llm-apps ⭐133,487 — 100+ AI Agents, Skills and RAG Apps
• Graphify-Labs/graphify ⭐109,096 — Turn any codebase into a queryable knowledge graph (local AST, no vector store)
• thedotmack/claude-mem ⭐91,430 — Persistent context across sessions (Claude Code/OpenClaw/Codex/Gemini…)
• infiniflow/ragflow ⭐88,983 — Leading open-source RAG engine
• PaddlePaddle/PaddleOCR ⭐88,065 — Lightweight OCR toolkit bridging images/PDFs and LLMs
• dair-ai/Prompt-Engineering-Guide ⭐77,664 — Prompt/context engineering & RAG guide
• datawhalechina/hello-agents ⭐74,116 — Build agents from scratch tutorial

4. Vector Databases

• Mintplex-Labs/anything-llm ⭐65,017 — Powerful local-first agent experience
• meilisearch/meilisearch ⭐59,045 — Lightning-fast search API with AI hybrid search
• pathwaycom/llm-app ⭐59,013 — Cloud templates for RAG/AI pipelines
• run-llama/llama_index ⭐51,790 — Document agent and OCR platform
• milvus-io/milvus ⭐45,728 — High-performance cloud-native vector database
• VectifyAI/PageIndex ⭐35,280 — Vectorless, reasoning-based RAG document index
• qdrant/qdrant ⭐34,115 — Massive-scale vector database & search engine
• topoteretes/cognee ⭐30,170 — Open-source AI memory platform (knowledge-graph engine)
• NirDiamant/RAG_Techniques ⭐29,168 — Advanced RAG technique notebooks
• weaviate/weaviate ⭐16,745 — Cloud-native vector database

5. LLM Models

• jingyaogong/minimind ⭐54,909 — Train a 64M-param LLM from scratch in 2h
• ScrapeGraphAI/Scrapegraph-ai ⭐29,794 — AI-based scraper
• The-Pocket/PocketFlow-Tutorial-Codebase-Knowledge ⭐12,626 — Codebase to Tutorial
• The-Pocket/PocketFlow ⭐11,120 — 100-line LLM framework; let Agents build Agents
• OpenSPG/KAG ⭐9,004 — Logic-form-guided reasoning & retrieval framework
• 0xPlaygrounds/rig ⭐8,347 — Build modular LLM apps in Rust
• open-compass/opencompass ⭐7,325 — LLM evaluation platform
• InternLM/InternLM ⭐7,267 — Official InternLM series releases
• Eigenwise/atomic-agents ⭐6,190 — Build AI agents atomically
• gluonfield/enchanted ⭐5,998 — iOS/macOS app for private self-hosted LLMs

6. Machine Learning

• tensorflow/tensorflow ⭐197,220 — Open-source ML framework for everyone
• f/prompts.chat ⭐167,704 — Awesome ChatGPT Prompts
• huggingface/transformers ⭐164,310 — SOTA ML model framework
• rasbt/LLMs-from-scratch ⭐103,171 — Implement a ChatGPT-like LLM in PyTorch step by step
• pytorch/pytorch ⭐102,517 — Tensors and dynamic neural networks with GPU acceleration
• microsoft/ML-For-Beginners ⭐89,653 — 12 weeks, 26 lessons, 52 quizzes
• Developer-Y/cs-video-courses ⭐83,139 — CS courses with video lectures
• mlabonne/llm-course ⭐81,893 — LLM course with roadmaps & Colab notebooks
• netdata/netdata ⭐80,255 — Fastest path to AI-powered full-stack observability
• d2l-ai/d2l-zh ⭐79,853 — Dive into Deep Learning (Chinese/English)

3. Trend Takeaways

• Agent toolchain & personal productivity surge together: Skills frameworks, agent harnesses, self-hosted agent platforms, and AI job-search/stock-analysis projects dominate multiple boards' tops.
• Memory & context become an infrastructure赛道: volcengine/OpenViking, thedotmack/claude-mem, topoteretes/cognee, Graphify-Labs/graphify each attack "agent memory/knowledge" from a different angle — echoing today's CLI/OpenClaw "memory persistence"主线.
• Security governance heats up: with Tencent/AI-Infra-Guard red-teaming platform, community attention to agent security shifts from optional to mandatory.
• Lightweight / local-first favored: Rust-written OpenLogi, local-first anything-llm/siyuan reflect preference for "self-hosted, private, low-dependency."

This digest is generated by agents-radar. All content is fetched live from GitHub's public data and written by the executing agent — no external LLM was called.

生成时间: 2026-08-22 02:30 (GMT+8) 数据来源：HN Algolia Search API（6 个查询：AI / LLM / Claude / OpenAI / Anthropic / machine learning，近 24h，去重后按 points 取 Top 30；本批共 104 条唯一故事） 本日报由 agents-radar 本地自动化生成，所有撰写由执行代理人工完成，未调用任何外部大模型 API。

一、AI 编码 Agent：成本、可靠性与自托管

• Codex on AWS bedrock bug causing 10x charges（145👍 / 60💬）— 一条 AWS Bedrock 上的 Codex bug 导致费用暴涨 10 倍，引发对 agent 计费透明度的激烈讨论。最高赞，显示「成本可控」已成为社区头号焦虑。
• Building an (almost) fully self-hosted, sandboxed, agentic software factory（25👍 / 17💬）— 几乎全自托管、带沙箱的 agentic 软件工厂实践，呼应今日 CLI/OpenClaw 日报里的「沙箱/自托管」主线。
• Show HN: Proliferate – open-source, self-hostable Codex for any coding agent（16👍 / 7💬）— 让任意编码 agent 用上自托管 Codex 的开源项目。
• LLMs are proof that Unix won（10👍 / 3💬）— 从 Unix 哲学视角解读 LLM 的胜利。
• Linus Torvalds Endures a Debug Session from Hell, "Enormously Helped" by AI（6👍 / 0💬）— Linus 自述 AI 在艰难调试中「极大帮助」，权威背书 agent 辅助开发。
• Claude Opus 4.6 returned nothing 900/900 times. Should agents retry?（5👍 / 1💬）— 900 次调用 0 返回，引发「agent 是否该重试、如何界定失败」的讨论，与 OpenCode #41469 空响应误判同源。

二、Claude / Anthropic 生态

• Claudette: Make Claude Stop Talking Like a BuzzFeed Article（77👍 / 75💬）— 让 Claude 停止「 BuzzFeed 式」浮夸表达的技能，高赞高讨论，反映对模型输出风格的审美疲劳。
• Anthropic Expects to Match SpaceX's Record IPO Size or Top It（8👍 / 0💬）— Bloomberg：Anthropic 预计 IPO 规模匹敌或超越 SpaceX 纪录。
• Anthropic plans to change enterprise data retention policy（4👍 / 0💬）— Reuters：Anthropic 计划调整企业数据保留政策，企业治理相关。
• Opus 5 feels, in a word, hostile（4👍 / 0💬）— 用户对 Opus 5 交互「敌对感」的吐槽（Reddit 镜像）。

三、OpenAI 动态与争议

• Introducing AI Futures（30👍 / 8💬）— OpenAI 发布「AI Futures」。
• OpenAI is becoming a surveillance company（9👍 / 2💬）— 对 OpenAI surveillance 化的批评。
• If You Weren't Worried About A.I., You Should Be（9👍 / 5💬）— NYT 观点文，呼吁重视 AI 风险。
• Why your infrastructure is more important than the next LLM release（6👍 / 1💬）— 论点：基础设施比下一个模型发布更重要。
• OpenAI Is Backing Away from Reddit as Reddit Tries to Become OpenAI?（6👍 / 1💬）— OpenAI 与 Reddit 关系微妙变化。
• OpenAI's Rogue AI Agent Hacked More Than Just Hugging Face（5👍 / 1💬）— Wired：OpenAI 失控 agent 不止黑了 Hugging Face。
• Introducing ChatGPT for Teens（4👍 / 0💬）— OpenAI 推出青少年版 ChatGPT。

四、Show HN 与开发者项目

• Show HN: A desktop fly drawn to the scent of vibecode（19👍 / 8💬）— 趣味项目：被 vibecode「气味」吸引的桌面苍蝇。
• Show HN: Visual way to understand things in parallel（5👍 / 3💬）— 可视化并行理解工具。
• Show HN: Praxos – team messaging with built-in memory（5👍 / 2💬）— 内置记忆的团队消息工具（呼应「记忆」主线）。
• Show HN: Gramola, a CoverFlow-style music dock for macOS（5👍 / 2💬）— macOS 音乐坞。
• Show HN: Bounded GenAI metrics from Otel traces with O raw prompts（4👍 / 2💬）— 从 Otel trace 取 GenAI 有界指标、零原始 prompt 泄露，可观测性方向。

五、模型与训练

• Good Results when training Qwen 3 4B to learn a new domain（5👍 / 0💬）— Qwen 3 4B 迁移学习新领域的好结果。
• Benchmarks of rumored Mythos level model from Zhipu AI（3👍 / 0💬）— 智谱传闻 Mythos 级别模型的基准。

六、其它/学术

• How I came to write that paper with Leslie Lamport（32👍 / 9💬）— 与 Lamport 合著论文的幕后。
• Code Obfuscation via Local Mixing（24👍 / 1💬）— Vitalik 关于代码混淆的文章。
• Rebuilding our Electron meeting-recording engine in Swift（18👍 / 1💬）— 用 Swift 重写 Electron 会议录制引擎。
• The puzzling case of Britain's young people and what they think of AI（4👍 / 0💬）— 英国年轻人对 AI 的态度调查。
• Writing with AI Is Stupid（5👍 / 3💬）— 对「用 AI 写作」的批评，与 #2 Claudette 的审美疲劳呼应。

趋势小结

• 成本与计费透明是头号议题：Codex Bedrock 10x 费用（145👍）与 OpenCode 空响应误判同源——agent 的「用量/失败可诊断」正成为信任底座。
• 自托管 + 沙箱成共识：self-hosted agentic factory、Proliferate、Praxos 记忆、Otel 指标，均指向「可控、可观测、隐私优先」的部署取向。
• 模型输出风格引发审美疲劳：Claudette（77👍）、Writing with AI Is Stupid，显示社区开始反感浮夸/套路化表达。
• 巨头动态主导舆论场：Anthropic IPO、OpenAI surveillance/Reddit/青少年版，企业治理与伦理争议持续升温。

本日报由 agents-radar 自动生成，所有内容基于 HN Algolia API 实时抓取并由执行代理人工撰写，未调用任何外部大模型。

Generated: 2026-08-22 02:30 (GMT+8) Source: HN Algolia Search API (6 queries: AI / LLM / Claude / OpenAI / Anthropic / machine learning, last 24h, deduped, Top 30 by points; 104 unique stories this batch) This digest is produced by the agents-radar local automation. All writing is done by the executing agent — no external LLM API was called.

1. AI Coding Agents: Cost, Reliability, Self-Hosting

• Codex on AWS bedrock bug causing 10x charges (145👍 / 60💬) — An AWS Bedrock Codex bug inflated costs 10x, sparking heated debate on agent billing transparency. Top story — "cost control" is now the community's #1 anxiety.
• Building an (almost) fully self-hosted, sandboxed, agentic software factory (25👍 / 17💬) — A near-fully self-hosted, sandboxed agentic software factory, echoing today's CLI/OpenClaw "sandbox / self-host"主线.
• Show HN: Proliferate – open-source, self-hostable Codex for any coding agent (16👍 / 7💬) — OSS project bringing self-hostable Codex to any coding agent.
• LLMs are proof that Unix won (10👍 / 3💬) — Reading LLM's victory through the Unix philosophy.
• Linus Torvalds Endures a Debug Session from Hell, "Enormously Helped" by AI (6👍 / 0💬) — Linus says AI "enormously helped" in a brutal debug session — authoritative endorsement of agent-assisted dev.
• Claude Opus 4.6 returned nothing 900/900 times. Should agents retry? (5👍 / 1💬) — 900 calls, 0 returns, raising "should agents retry / how to define failure" — same root as OpenCode #41469 empty-response miscount.

2. Claude / Anthropic Ecosystem

• Claudette: Make Claude Stop Talking Like a BuzzFeed Article (77👍 / 75💬) — A skill to stop Claude's BuzzFeed-style fluff; high votes + discussion, signaling aesthetic fatigue with model output style.
• Anthropic Expects to Match SpaceX's Record IPO Size or Top It (8👍 / 0💬) — Bloomberg: Anthropic expects an IPO matching or surpassing SpaceX's record.
• Anthropic plans to change enterprise data retention policy (4👍 / 0💬) — Reuters: Anthropic plans to change enterprise data-retention policy (enterprise governance).
• Opus 5 feels, in a word, hostile (4👍 / 0💬) — User gripe about Opus 5's "hostile" interaction (Reddit mirror).

3. OpenAI News & Controversies

• Introducing AI Futures (30👍 / 8💬) — OpenAI launches "AI Futures".
• OpenAI is becoming a surveillance company (9👍 / 2💬) — Critique of OpenAI's surveillance trajectory.
• If You Weren't Worried About A.I., You Should Be (9👍 / 5💬) — NYT op-ed urging attention to AI risk.
• Why your infrastructure is more important than the next LLM release (6👍 / 1💬) — Thesis: infrastructure matters more than the next model.
• OpenAI Is Backing Away from Reddit as Reddit Tries to Become OpenAI? (6👍 / 1💬) — Shifting OpenAI–Reddit relationship.
• OpenAI's Rogue AI Agent Hacked More Than Just Hugging Face (5👍 / 1💬) — Wired: OpenAI's rogue agent hacked more than Hugging Face.
• Introducing ChatGPT for Teens (4👍 / 0💬) — OpenAI launches a teen version of ChatGPT.

4. Show HN & Developer Projects

• Show HN: A desktop fly drawn to the scent of vibecode (19👍 / 8💬) — Fun project: a desktop fly drawn to the "scent" of vibecode.
• Show HN: Visual way to understand things in parallel (5👍 / 3💬) — Visual parallel-understanding tool.
• Show HN: Praxos – team messaging with built-in memory (5👍 / 2💬) — Team messaging with built-in memory (echoes the "memory"主线).
• Show HN: Gramola, a CoverFlow-style music dock for macOS (5👍 / 2💬) — macOS music dock.
• Show HN: Bounded GenAI metrics from Otel traces with O raw prompts (4👍 / 2💬) — Bounded GenAI metrics from Otel traces, zero raw-prompt leakage — observability direction.

5. Models & Training

• Good Results when training Qwen 3 4B to learn a new domain (5👍 / 0💬) — Good transfer-learning results for Qwen 3 4B on a new domain.
• Benchmarks of rumored Mythos level model from Zhipu AI (3👍 / 0💬) — Benchmarks of Zhipu AI's rumored Mythos-level model.

6. Other / Academic

• How I came to write that paper with Leslie Lamport (32👍 / 9💬) — Behind the scenes of co-authoring a paper with Lamport.
• Code Obfuscation via Local Mixing (24👍 / 1💬) — Vitalik on code obfuscation.
• Rebuilding our Electron meeting-recording engine in Swift (18👍 / 1💬) — Rewriting an Electron meeting-recording engine in Swift.
• The puzzling case of Britain's young people and what they think of AI (4👍 / 0💬) — Survey of British youth attitudes toward AI.
• Writing with AI Is Stupid (5👍 / 3💬) — Critique of "writing with AI", resonating with #2 Claudette's aesthetic fatigue.

Trend Takeaways

• Cost & billing transparency is the #1 topic: Codex Bedrock 10x charges (145👍) shares a root with OpenCode's empty-response miscount — "usage/failure diagnosable" is becoming the trust bedrock for agents.
• Self-host + sandbox is consensus: self-hosted agentic factory, Proliferate, Praxos memory, Otel metrics — all point to "controllable, observable, privacy-first" deployment.
• Model output style triggers aesthetic fatigue: Claudette (77👍) and "Writing with AI Is Stupid" show the community tiring of fluff/boilerplate expression.
• Megacorp moves dominate the discourse: Anthropic IPO, OpenAI surveillance/Reddit/teen-version — enterprise governance and ethics controversies keep heating up.

This digest is generated by agents-radar. All content is fetched live from the HN Algolia API and written by the executing agent — no external LLM was called.

生成时间: 2026-08-21 02:30 (GMT+8) | 覆盖工具: 7 个 CLI + Claude Code Skills 本日报由 agents-radar 本地自动化生成，所有摘要/对比/撰写均由执行代理人工完成，未调用任何外部大模型 API。

• Claude Code
• OpenAI Codex
• Gemini CLI
• GitHub Copilot CLI
• Kimi Code CLI
• OpenCode
• Qwen Code
• Claude Code Skills

横向对比

AI CLI 工具社区动态横向对比分析报告（2026-08-21）

1. 生态全景

AI CLI 工具进入「日更竞争」阶段：今日 Claude Code、OpenAI Codex、Gemini CLI、GitHub Copilot CLI、OpenCode、Qwen Code 均有版本发布，IronClaw / EasyClaw / Moltis / CoPaw 等周边生态也同步出新。社区诉求的焦点从「模型能力」进一步转向「工程可信度与平台质量」——Windows 平台问题（Codex 尤为突出）、TUI 终端交互细节、上下文压缩后的记忆持久化、以及沙箱/权限的最小信任边界，成为跨工具共通的痛点。一个明显趋势是：自动化 bot 正在深度参与各工具的发布质量门禁（Codex 的 copyberry[bot]、Qwen 的 qwen-code-review-bot、OpenCode 的 opencode-agent 等），修复与发布的频次被显著拉高，但外部人类社区的声量分化加剧。

2. 各工具活跃度对比（近 24h）

3. 共同关注的功能方向

• Windows / 跨平台质量：OpenAI Codex 的 Windows 冻结（#20214，86👍/107💬）、独立安装包诉求（#13993，242👍）、Trusted RPC 浏览器插件失败（#39136）持续高赞；Copilot CLI 的 wta.exe 路径引号错误（#4540）与沙箱阻断 git（#4524）；OpenCode 此前暴露的 /tmp 泄漏、Sidecar 超时等仍被社区引用。Windows 是第一大痛点平台。
• TUI / 终端交互稳定性：Claude Code 复制粘贴缩进污染（#18170，287👍/134💬）仍是社区最高频诉求；OpenCode 的剪贴板写入静默失败修复（#41924）仍在开放；Gemini CLI 的滚动/渲染（#13107、#10673）与拖拽图片（#27855）被持续提出。
• 上下文与跨压缩记忆：Claude Code 用户自建跨压缩记忆系统（#34556，98💬）热度不减；OpenCode 的「中断流被记为干净停止」（#37852，56👍）暴露 subagent 用量统计失真；Gemini 的 subagent 误报成功（#22323）仍在讨论。
• 权限 / 沙箱最小信任：Copilot CLI 沙箱策略过度强制（#4524）、store_memory 在 1.0.81 失败（#4535）、非交互模式绕过 managed settings（#4103）构成企业治理语境的核心争议；Codex 收紧 Git 命令信任、隔离插件 Git 操作等安全 PR 方向清晰。
• 模型兼容性与成本透明：Codex GPT-5.6 串行化导致 27–45% 额外用量（#35050，72👍）；OpenCode 的 gpt-5.6 系列参数注入/空响应问题（昨日 #43367/#43565）仍在被引用；Qwen Code 路由切换后 token 计数失效（#9506）已修复并绑定路由身份。
• 自动化发布质量门禁：Codex copyberry[bot] 今日批量合入 12+ PR（模型定义刷新、shell 统一执行、MCP 事件流等）；Qwen qwen-code-review-bot 发布 v0.21.15；OpenCode opencode-agent 持续提交。AI 已深度参与自身软件工程。

4. 差异化定位

• Claude Code：社区期望最高、反馈最尖锐，复制粘贴/快捷键/输出展开等体验细节被零容忍审视；v2.1.237 新增「Concise」输出风格与 prompt caching 修复，定位专业开发者日常主力终端。PR 侧近 24h 静默，意味着主干进入稳定打磨期。
• OpenAI Codex：Rust 重写持续推进，bot 驱动发布节奏快；Windows 质量与桌面端集成（Computer Use、浏览器插件）是最突出短板。安全方向（Git 信任收紧、插件隔离）获社区肯定，定位 OpenAI 全场景入口。
• Gemini CLI：稳定版/预览版分轨清晰，PR 以安全加固（Cloud Workstations OAuth、IDE 连接、环境变更同意）与模型接入（Gemini 3.6/3.7 Flash）为主，成熟度中上。
• GitHub Copilot CLI：企业受管语境最明确，managed settings 强制、沙箱、数据驻留、审计是核心；今日 v1.0.81-6 新增 defaultMode/defaultPermissionMode 与 --with-token 登录，ACP 客户端获得 subagent ID 等能力。定位企业受管环境的 Copilot 入口。
• Kimi Code CLI：今日零动态（issues/PRs/releases 全空），生态处于早期静默期。
• OpenCode：社区贡献极活跃，v1.18.19 新增 Cloudflare AI Gateway 原生透传、对齐 Codex 速率限制、Go provider 启用 web 搜索；问题集中在 TUI 数据丢失级缺陷与 gpt-5.6 兼容性。定位追求最新模型与高度可定制的开发者。
• Qwen Code：v0.21.15 强化 Web Shell 文件附件、混合模型 Thinking 开关、/review --resume、鉴权 HTTPS Git 扩展安装，工业化的 release/CI bot 体系成熟，定位验证严谨性与中国 IM/云生态用户。

5. 值得关注的趋势信号

• Windows 是未被充分满足的硬市场：Codex 冻结/安装包/插件、Copilot 路径引号、Gemini Browser agent 跨显示协议失败——后来者若把 Windows 质量当作一等公民，是差异化切口。
• TUI 基础体验成为选型硬指标：复制粘贴、滚动、拖拽图片等「小问题」占据各工具高赞榜，真实终端环境（WSL、Ghostty、Windows Terminal）可用性测试应优先于模型能力对比。
• 记忆持久化是下一轮分水岭：压缩后历史可回溯 + 隐私安全 + 可控，三件事同时做好的工具有望建立长期粘性。
• 自动化发布门禁双刃剑：bot 提速修复，但也让外部人类贡献者声量被稀释，对人工审查机制提出新要求。
• 模型行为变更直接关联账单：GPT-5.6 串行化带来额外用量（#35050），开发者应「先小规模验证成本再大规模切换」，并要求工具提供路由身份绑定的用量统计。

各工具详细报告

Claude Code Skills 社区热点

数据来源: anthropics/skills | 近 24h 活跃

• #1613 claude-api frontmatter description 超过 1024 字符限制（1💬，开放）— 技能元数据长度约束问题。
• #1616 修复 claude-api description 限制（@BRGOVIND，开放）— 对应 #1613 的修复 PR。
• #1615 新增 scnet-hpc skill（@lql341，开放）— HPC 场景技能。
• #1614 新增 web-security-audit skill（@cnbruce，开放）— Web 安全审计技能。
• #807 新增 x-twitter-scraper：X 数据 API 技能（@kriptoburak，已合并）— 社交媒体数据抓取技能。

Skills 生态今日节奏平稳，以单点 bug 修复与新增技能为主，无大规模架构讨论。

本日报由 agents-radar 自动生成，所有内容基于 GitHub 公开 API 实时抓取并由执行代理人工撰写，未调用任何外部大模型。

Generated: 2026-08-21 02:30 (GMT+8) | Coverage: 7 CLIs + Claude Code Skills This digest is produced by the agents-radar local automation. All summaries, comparisons and writing are done by the executing agent — no external LLM API is called.

• Claude Code
• OpenAI Codex
• Gemini CLI
• GitHub Copilot CLI
• Kimi Code CLI
• OpenCode
• Qwen Code
• Claude Code Skills

Cross-Tool Comparison

AI CLI Tools Cross-Tool Comparison Report (2026-08-21)

1. Ecosystem Overview

AI CLI tools have entered a "daily-release" competition: today Claude Code, OpenAI Codex, Gemini CLI, GitHub Copilot CLI, OpenCode and Qwen Code all shipped versions, with the surrounding ecosystem (IronClaw / EasyClaw / Moltis / CoPaw) releasing in sync. Community focus has shifted further from "model capability" to "engineering trust and platform quality" — Windows-quality issues (especially Codex), TUI interaction details, post-compaction memory persistence, and sandbox/permission least-trust boundaries are now common pain points across tools. A clear trend: automation bots are deeply involved in each tool's release quality gates (copyberry[bot] for Codex, qwen-code-review-bot for Qwen, opencode-agent for OpenCode), raising fix/release cadence while widening the voice gap between internal automation and external human contributors.

2. Activity Comparison (last 24h)

3. Shared Themes

• Windows / cross-platform quality: Codex Windows freeze (#20214, 86👍/107💬), standalone installer wish (#13993, 242👍), Trusted RPC browser-plugin failure (#39136); Copilot wta.exe quote bug (#4540) and sandbox blocking git (#4524); OpenCode /tmp leak and Sidecar timeout references. Windows is the #1 pain platform.
• TUI stability: Claude Code copy/paste indentation pollution (#18170, 287👍/134💬) remains the top complaint; OpenCode clipboard silent-failure fix (#41924) still open; Gemini scroll/render (#13107, #10673) and image drag-drop (#27855) recur.
• Context & post-compaction memory: Claude Code user-built cross-compaction memory (#34556, 98💬); OpenCode "aborted stream recorded as clean stop" (#37852, 56👍) distorts subagent usage stats; Gemini subagent false GOAL success (#22323) still discussed.
• Permission / sandbox least-trust: Copilot over-aggressive sandbox (#4524), store_memory failure in 1.0.81 (#4535), non-interactive bypass of managed settings (#4103); Codex tightening Git-command trust and isolating plugin Git ops earn community approval.
• Model compatibility & cost transparency: Codex GPT-5.6 serialization adds 27–45% weighted usage (#35050, 72👍); OpenCode gpt-5.6 param-injection/empty-response references persist; Qwen routing token-count fix (#9506) now binds routing identity.
• Automated release quality gates: Codex copyberry[bot] merged 12+ PRs today; Qwen qwen-code-review-bot shipped v0.21.15; OpenCode opencode-agent continuous — AI is deeply embedded in its own engineering.

4. Differentiation

• Claude Code: highest expectations, sharpest feedback; v2.1.237 adds "Concise" output style + prompt-caching fix; positions as the professional developer's daily terminal. PR side silent in 24h → stabilization phase.
• OpenAI Codex: Rust rewrite ongoing, bot-driven cadence; Windows quality & desktop integration (Computer Use, browser plugin) are the biggest shortcomings; security direction (Git trust tightening, plugin isolation) approved.
• Gemini CLI: clean stable/preview split; PRs on security hardening (Cloud Workstations OAuth, IDE connection, env-change consent) and model onboarding (Gemini 3.6/3.7 Flash); mature-middle.
• GitHub Copilot CLI: strongest enterprise-governance context; v1.0.81-6 adds defaultMode/defaultPermissionMode and --with-token; positions as the managed Copilot entry point.
• Kimi Code CLI: zero activity today; early-silent stage.
• OpenCode: highly active community; v1.18.19 adds Cloudflare AI Gateway native passthrough, Codex rate-limit alignment, Go provider web search; pain in data-loss TUI defects and gpt-5.6 compat; positions for latest-model / customizable developers.
• Qwen Code: v0.21.15 strengthens Web Shell attachments, hybrid Thinking toggle, /review --resume, authenticated HTTPS Git extension installs; mature industrial release/CI bot system.

5. Notable Trend Signals

• Windows is an unsatisfied hard market — a differentiator for latecomers who treat Windows quality as first-class.
• TUI basics are a hard selection criterion — copy/paste, scroll, image drag-drop top the upvote charts; real-terminal testing should precede model-capability comparison.
• Memory persistence is the next watershed — retrievable + private + controllable history together builds long-term stickiness.
• Automated release gates are double-edged — faster fixes but diluted external human voice; demands stronger human-review mechanisms.
• Model behavior changes hit the bill directly — verify cost at small scale before large switches; require routing-identity-bound usage stats.

Per-Tool Reports

Claude Code Skills Community Highlights

Source: anthropics/skills | active 24h

• #1613 claude-api frontmatter description exceeds 1024-char limit (open) — metadata length constraint.
• #1616 Fix claude-api description limit (@BRGOVIND, open).
• #1615 Add scnet-hpc skill (@lql341, open).
• #1614 Add web-security-audit skill (@cnbruce, open).
• #807 Add x-twitter-scraper skill (@kriptoburak, merged).

Skills ecosystem calm today — single-point bug fixes and new skills, no large architectural debate.

Generated by agents-radar. Data fetched live from GitHub public APIs and written by the executing agent — no external LLM used.

生成时间: 2026-08-21 02:30 (GMT+8) | 覆盖: OpenClaw + 12 个 peer 项目 本日报由 agents-radar 本地自动化生成，所有摘要/对比/撰写均由执行代理人工完成，未调用任何外部大模型 API。

• OpenClaw
• NanoBot
• Zeroclaw
• PicoClaw
• NanoClaw
• NullClaw
• IronClaw
• LobsterAI
• TinyClaw
• CoPaw
• Moltis
• ZeptoClaw
• EasyClaw

跨生态对比

OpenClaw 跨生态对比分析报告（2026-08-21）

1. 生态全景

OpenClaw 本体今日有 35 条 PR 更新、0 issue、0 release，主干聚焦鉴权一致性、记忆/wiki 上下文治理、Control UI 与 HTTP 边界；12 个 peer 中，NanoBot / Zeroclaw / IronClaw / CoPaw 贡献最活跃，NullClaw / TinyClaw / ZeptoClaw 完全静默。整体趋势从「单 agent 聊天」走向多渠道（Telegram/Matrix/Slack/WhatsApp）+ 持久化记忆 + 自动化（automations/run-now）+ 企业级沙箱/鉴权。

2. 各项目活跃度对比（近 24h）

3. 共同关注方向

• 多渠道接入与消息可靠性：NanoBot（Telegram 贴纸回复 #5387、Matrix 错误日志 #5458、Slack 文件下载重定向 #5414）、NanoClaw（WhatsApp 媒体不可达 #2715、Slack agents #3421、Cursor SDK #3356）、Moltis（WhatsApp Markdown #1220、回复即寻址 #1217）。
• 记忆与上下文治理：OpenClaw 防止超大 wiki 内容挤占上下文（#126779）、记忆导入失败暴露（#126738）；IronClaw AfterTurn 记忆策展（#7765）；CoPaw ReMe 记忆运行时看板（#6984）、PowerContext 长期记忆（#7080）。
• 自动化与企业级：IronClaw 结构化 automations + 每用户沙箱代理（#7732/#7779）、可执行 run gate 通知（#7699）；CoPaw 自托管多用户 Hub（#7112）。
• 安全加固：Moltis 修复 Vault 解锁/恢复缺失鉴权（#1216，CWE-306）、CoPaw master key 仅属主权限（#7119）。
• 上下文压缩与配额：Zeroclaw 上下文压缩锚定模型窗口比（#9535）、运行时上下文耗尽终端提示（#9504）；NanoClaw token 用量（#3270）。

4. 差异化定位

• OpenClaw：生态中枢，主攻 Control UI / 鉴权 / 记忆 wiki / HTTP 边界，企业受管语境。
• NanoBot：渠道最广（Telegram/Matrix/Slack/WebUI），重渠道健壮性与网关日志。
• Zeroclaw：Rust 核心，RFC 驱动「更轻核心 + 外部集成」，上下文压缩与 telemetry 治理。
• NanoClaw：多 provider 接入（WhatsApp/Slack/Cursor/Codex），强调 token 用量与挂载。
• IronClaw：每用户沙箱 + 结构化 automations + 文档编辑 + Telegram 关联设备，企业/团队向。
• LobsterAI：cowork 文件卡片/分屏预览、设置面板搜索，体验向。
• CoPaw (QwenPaw)：记忆 ReMe、自托管 Hub、控制台性能，社区活跃度最高之一。
• Moltis：WhatsApp 为中心，重渠道安全（Vault 鉴权）。
• EasyClaw (rivonclaw)：高频小版本（每日多版），商务开发筛选等垂直功能。
• PicoClaw / NullClaw / TinyClaw / ZeptoClaw：低活跃或静默。

5. 趋势信号

• 静默项目需关注：NullClaw / TinyClaw / ZeptoClaw 连续多日 0 动态，生态健康度分化。
• 自动化 becoming first-class：IronClaw/CoPaw 的 automations/run-now/Hub 标志着 agent 从「对话」走向「可编排任务」。
• 安全从可选到必需：Vault 鉴权缺失（CWE-306）、master key 权限、MCP 凭证保留成为高频修复。
• 记忆是跨项目主线：从 OpenClaw wiki 到 IronClaw/CoPaw 记忆系统，持久化记忆是 2026 年 agent 框架的竞争焦点。

各项目详细动态

本日报由 agents-radar 自动生成，数据来自 GitHub 公开 API，由执行代理人工撰写，未调用外部大模型。

Generated: 2026-08-21 02:30 (GMT+8) | Coverage: OpenClaw + 12 peer projects This digest is produced by the agents-radar local automation. All summaries, comparisons and writing are done by the executing agent — no external LLM API is called.

• OpenClaw
• NanoBot
• Zeroclaw
• PicoClaw
• NanoClaw
• NullClaw
• IronClaw
• LobsterAI
• TinyClaw
• CoPaw
• Moltis
• ZeptoClaw
• EasyClaw

Cross-Ecosystem Comparison

OpenClaw Cross-Ecosystem Comparison (2026-08-21)

1. Ecosystem Overview

OpenClaw core shipped 35 PR updates, 0 issues, 0 release — mainline focuses on auth consistency, memory/wiki context governance, Control UI and HTTP boundaries. Among 12 peers, NanoBot / Zeroclaw / IronClaw / CoPaw are most active; NullClaw / TinyClaw / ZeptoClaw are completely silent. The overall trajectory moves from "single-agent chat" to multi-channel (Telegram/Matrix/Slack/WhatsApp) + persistent memory + automation (automations/run-now) + enterprise sandbox/auth.

2. Activity Comparison (last 24h)

3. Shared Themes

• Multi-channel reliability: NanoBot (Telegram stickers #5387, Matrix error logs #5458, Slack redirect #5414), NanoClaw (WhatsApp media unreachable #2715, Slack agents #3421, Cursor SDK #3356), Moltis (WhatsApp Markdown #1220, reply-as-address #1217).
• Memory & context governance: OpenClaw oversized-wiki crowding fix (#126779), import-failure surfacing (#126738); IronClaw AfterTurn memory curation (#7765); CoPaw ReMe runtime dashboard (#6984), PowerContext LTM (#7080).
• Automation & enterprise: IronClaw structured automations + per-user sandbox proxy (#7732/#7779), executable run-gate notifications (#7699); CoPaw self-hosted multi-user Hub (#7112).
• Security hardening: Moltis Vault unlock/recovery auth fix (#1216, CWE-306); CoPaw master-key owner-only perms (#7119).
• Context compaction & quota: Zeroclaw window-ratio anchoring (#9535), context-exhaustion terminal notice (#9504); NanoClaw token usage (#3270).

4. Differentiation

• OpenClaw: ecosystem hub — Control UI / auth / memory-wiki / HTTP boundary, enterprise-managed context.
• NanoBot: widest channels (Telegram/Matrix/Slack/WebUI), channel robustness & gateway logs.
• Zeroclaw: Rust core, RFC-driven "lighter core + external integrations", context compaction & telemetry governance.
• NanoClaw: multi-provider onboarding (WhatsApp/Slack/Cursor/Codex), token usage & mounts.
• IronClaw: per-user sandbox + structured automations + doc editing + Telegram linked devices, team/enterprise-leaning.
• LobsterAI: cowork file cards / split preview, settings search, experience-focused.
• CoPaw (QwenPaw): ReMe memory, self-hosted Hub, console perf — among the most active communities.
• Moltis: WhatsApp-centric, channel security (Vault auth).
• EasyClaw (rivonclaw): high-frequency point releases (multiple/day), vertical features (BD-agent filtering).
• PicoClaw / NullClaw / TinyClaw / ZeptoClaw: low-activity or silent.

5. Trend Signals

• Silent projects need attention: NullClaw / TinyClaw / ZeptoClaw at 0 activity for multiple days — ecosystem health is diverging.
• Automation becomes first-class: IronClaw/CoPaw automations/run-now/Hub signal agents moving from "chat" to "orchestrated tasks".
• Security from optional to mandatory: Vault auth gaps (CWE-306), master-key perms, MCP credential retention are frequent fixes.
• Memory is the cross-project main line: from OpenClaw wiki to IronClaw/CoPaw memory systems, persistent memory is the 2026 agent-framework battleground.

Per-Project Details

CoPaw (QwenPaw) Daily (2026-08-21)

Today

Shipped v2.1.1-beta.1. 7 issues, 35 PRs. ReMe memory runtime dashboard, self-hosted multi-user Hub, console perf, unified tool panel.

Release

v2.1.1-beta.1 — editor tab overflow nav, provider rate-limiter log level, ReMe runtime dashboard, plugin-channel interactive configurators, OneBot inbound media localization, background-task timeout, plugin reload workspace state, historical data-URL image render, DataPaw native runtime.

Hot Issues

#6921 stops without notice after "Now 2.1, 3.1, 3.2" output, needs "continue" · #7102 freeze >10 min · #6643 outputs shouldn't pile in media dir, per-task dirs · #6826 assistant message end-time anomaly · #6436 automatic per-message model routing · #7110 undownloadable image link breaks session · #7013 unified tool panel + web preview + interactive terminal.

Notable PRs

#7119 fix(security): master key file owner-only perms · #7061 fix(video): deliver tool-result videos on OpenAI Responses API · #7175 fix(console): restore full free model listings · #7183 feat(skills): workspace-scoped always-on loading · #7167 feat(creator): dialogue-gated video dispatch, project copy/recreate · #7133 feat(memory): update reme 0.4.1.8 · #7112 feat(hub): self-hosted multi-user Hub (local & Docker) · #7080 feat: optional PowerContext pluggable LTM backend · #7161 feat(console): artifacts on assistant response card · #7176 perf(console): keep long chats responsive.

Generated by agents-radar. Data fetched live from GitHub public APIs and written by the executing agent — no external LLM used.

生成时间: 2026-08-21 02:30 (GMT+8) 数据来源：GitHub Trending（每日榜）+ gh search repos 6 个 AI 主题（按 stars 排序） 本日报由 agents-radar 本地自动化生成，所有撰写由执行代理人工完成，未调用任何外部大模型 API。

一、GitHub Trending 每日飙升榜（2026-08-21）

二、六大 AI 主题搜索热度（按 stars）

1. AI Agent 框架

• HKUDS/nanobot ⭐47,228 — 超轻量、自托管的个人 AI agent 平台
• TransformerOptimus/SuperAGI ⭐17,660 — 开发者优先的开源自主 AI agent 框架
• VoltAgent/voltagent ⭐10,386 — 基于开源 TypeScript 的 AI Agent 工程平台
• omnigent-ai/omnigent ⭐9,094 — 开源 AI agent 框架与元编排
• darrenhinde/OpenAgentsControl ⭐4,748 — 计划优先的开发工作流 agent 框架
• GH05TCREW/pentestagent ⭐2,972 — 黑盒安全测试的 AI agent 框架
• espressif/esp-claw ⭐2,025 — 面向 IoT 开发的「Chat Coding」agent 框架
• yomorun/yomo ⭐1,920 — 地理分布边缘的无服务器 AI Agent 框架

2. 大模型（LLM）

• u14app/deep-research ⭐4,680 — 用任意 LLM 做深度研究
• HuaizhengZhang/AI-Infra-from-Zero-to-Hero ⭐4,288 — AI 系统论文与工程宝典
• ModelTC/LightLLM ⭐4,228 — 基于 Python 的 LLM 推理引擎
• llmgenai/LLMInterviewQuestions ⭐1,918 — LLM 面试问题合集
• apirrone/Memento ⭐654 — 记录你在电脑上一切操作的 Python 应用
• rohan-paul/LLM-FineTuning-Large-Language-Models ⭐577 — LLM 微调
• maquina-app/rails-mcp-server ⭐570 — Ruby 实现的 MCP 服务器
• gergap/vim-ollama ⭐334 — Vim 的 Ollama LLM 插件

3. RAG 检索增强生成

• truefoundry/cognita ⭐4,416 — 构建 RAG 应用的框架
• rag-web-ui/rag-web-ui ⭐3,265 — 基于 RAG 的智能对话系统
• Andrew-Jang/RAGHub ⭐1,984 — 社区驱动的 RAG 资源集合
• jerry-ai-dev/MODULAR-RAG-MCP-SERVER ⭐1,094 — 模块化 RAG 系统 + MCP 服务
• mrdbourke/simple-local-rag ⭐1,010 — 从零构建本地 RAG 管线
• RafalWilinski/cloudflare-rag ⭐603 — 全栈「与 PDF 对话」RAG
• tonykipkemboi/ollama_pdf_rag ⭐540 — 本地 RAG 全栈 demo
• LightningRAG/LightningRAG ⭐451 — Vue + Gin 全栈 RAG 启动器

4. MCP 模型上下文协议

• microsoft/mcp ⭐3,593 — 微软官方 MCP 目录
• neka-nat/freecad-mcp ⭐1,860 — FreeCAD MCP 服务器
• GongRzhe/Office-PowerPoint-MCP-Server ⭐1,854 — PowerPoint 操作的 MCP 服务器
• johnson7788/MultiAgentPPT ⭐1,633 — 集成 A2A + MCP 的多 agent PPT 工具
• f/mcptools ⭐1,619 — 与 MCP 服务器交互的命令行工具
• todo-for-ai/todo-for-ai ⭐1,176 — 专为 AI 设计的任务管理系统
• Puliczek/awesome-mcp-security ⭐729 — MCP 安全资源精选
• innocommerce/innoshop ⭐651? — AI 驱动的开源电商系统

5. Agentic 工作流自动化

• HKUDS/nanobot ⭐47,228 — 自托管个人 AI agent 平台
• pipeshub-ai/pipeshub-ai ⭐3,657 — 开源、可完全扩展的 AI 上下文层
• vmDeshpande/ai-agent-automation ⭐175 — 带调度的模块化 AI agent 工作流平台
• code-forge-temple/agentic-signal ⭐162 — 本地可视化 AI agent 工作流平台
• idoforgod/Church-Admin-AgenticWorkflow ⭐45 — 教会管理 Agentic 工作流
• idoforgod/SaaS-AgenticWorkflow ⭐37 — SaaS 自动构建器
• hamr0/agentic-toolkit ⭐21 — 开箱即用的 AI 开发工具套件
• GurYN/opencode-sdlc ⭐3 — OpenCode 的 SDLC 管理系统

6. 多模态基础模型

• jd-opensource/JoyAI-Image ⭐2,146 — 统一的图像多模态基础模型
• DAMO-NLP-SG/VideoLLaMA3 ⭐1,175 — 面向图像与视频的前沿多模态基础模型
• InternLM/Intern-S1 ⭐847 — 科学多模态基础模型
• weijiawu/Awesome-RL-for-Multimodal-Foundation-Models ⭐453 — 多模态基础模型强化学习论文/代码集
• xid32/NAACL_2025_TWM ⭐316 — 时序工作记忆（TWM）
• OpenSenseNova/SenseNova-SI ⭐294 — [CVPR 2026] 多模态空间智能
• OpenBMB/AppCopilot ⭐245 — 通用、精准、长程、高效的移动 Agent
• hzlab/Brain-Harmony ⭐243 — 多模态基础模型代码库

三、趋势洞察（执行代理分析）

1. 「Skills / Agent 记忆」成为飙升主线：Trending 前排被 agent skills 与记忆类项目占据——mattpocock 的 Skills 合集（+2,267）、obra 的 agentic skills 框架（+749）、akitaonrails/ai-memory（+335）、JuliusBrussee 的 token 精简 skill（+309）、cursor/plugins（+473），加上 OpenViking 自进化上下文数据库（+955）。这印证了与 CLI/OpenClaw 日报一致的信号：工程重心正从「模型能力」转向「可复用技能与持久记忆」。

2. 短视频/AIGC 仍是流量引擎：MoneyPrinterTurbo 以 +2,774 居 Trending 与搜索双榜首，说明「主题→短视频」的自动化创作仍是大众最买账的 AI 落地形态。

3. MCP 生态走向「安全 + 生产力」双轨：microsoft/mcp 目录、Office/FreeCAD MCP 服务器、mcptools 持续走高；awesome-mcp-security 上榜反映MCP 供应链安全已成为独立关注点（与 CLI 日报中 MCP fail-open 修复、凭证保留等议题呼应）。

4. RAG 框架成熟化：cognita、rag-web-ui 等从 demo 走向生产框架，多模态 RAG（MODULAR-RAG-MCP-SERVER）开始出现。

5. 自主 agent 平台的「自托管」叙事强化：nanobot（47k★）、SuperAGI、voltagent 等强调自托管与个人 agent，与 OpenClaw 生态的 per-user sandbox、CoPaw 自托管 Hub 形成共振。

6. 多模态基础模型开源活跃：京东 JoyAI-Image、阿里 DAMO VideoLLaMA3、上海 AI Lab Intern-S1 等显示国内团队在多模态基础模型上的持续投入。

7. 本仓库相关：volcengine/OpenViking 今日 +955★，定位为面向 AI Agent 的自进化上下文数据库——与 agents-radar 关注的「Agent 记忆/知识管理」方向高度相关。

本日报由 agents-radar 自动生成，数据来自 GitHub 公开页面与 API，由执行代理人工撰写，未调用外部大模型。