Ten source-checked AI updates for August 17

This roundup uses first-party release pages and repositories because the two upstream brief bodies were not reachable from the current adapter. That is a retrieval limitation, not a claim that the scheduled products were late. Prereleases remain labeled and project-reported scores are not treated as independent benchmarks.

1. Qwen Code 0.21.13 增加会话分叉与更可追踪的代码审查

Qwen Code 0.21.13 adds persistent conversation branches and more traceable code-review sessions. The practical point is narrower than a product-launch headline: read the cited release or repository before upgrading, and do not infer performance gains where maintainers published none. Primary source

2. OpenAI Agents SDK Python/JavaScript 同日补上调用超时与沙箱隔离

OpenAI Agents SDKs add model-call timeouts, run-scoped sandboxes, and an option to disable Docker sandbox networking. The practical point is narrower than a product-launch headline: read the cited release or repository before upgrading, and do not infer performance gains where maintainers published none. Primary source

3. browser-use 0.13.8 默认切换轻量模型，并改善 MCP 错误回传

browser-use 0.13.8 switches its default to a preview mini model and correctly surfaces MCP isError results as failures. The practical point is narrower than a product-launch headline: read the cited release or repository before upgrading, and do not infer performance gains where maintainers published none. Primary source

4. llama.cpp 连发 b10452/b10453 两个维护构建

llama.cpp b10452/b10453 are maintenance builds for message normalization and graph internals—not a new model. The practical point is narrower than a product-launch headline: read the cited release or repository before upgrading, and do not infer performance gains where maintainers published none. Primary source

5. Gemini CLI 0.56.0 nightly 修正 SSR Agent TypeScript 配置

Gemini CLI 0.56.0 nightly fixes an SSR Agent TypeScript build setting; it is a prerelease, not a model upgrade. The practical point is narrower than a product-launch headline: read the cited release or repository before upgrading, and do not infer performance gains where maintainers published none. Primary source

6. LobeHub 2.2.15 canary 连续试验成长画像、推送与任务布局

LobeHub 2.2.15 canaries iterate on goals, agent pushes, growth portraits, and task layout; production use is not advised. The practical point is narrower than a product-launch headline: read the cited release or repository before upgrading, and do not infer performance gains where maintainers published none. Primary source

7. jailbreak-kit 开源一套 LLM 对齐鲁棒性回归测试工具包

jailbreak-kit packages authorized LLM red-team cases into a repeatable regression harness; author scores are not an independent benchmark. The practical point is narrower than a product-launch headline: read the cited release or repository before upgrading, and do not infer performance gains where maintainers published none. Primary source

8. agentleaks 面向 AI 编码 Agent 会话日志做密钥扫描与脱敏

agentleaks scans local coding-agent sessions for secrets and can scrub them with backups and active-session protection. The practical point is narrower than a product-launch headline: read the cited release or repository before upgrading, and do not infer performance gains where maintainers published none. Primary source

9. dsh-crew 用插件把 DeepSeek Harness Agent 接入 Claude Code/Codex 工作流

dsh-crew is an RC plugin for dispatching DeepSeek Harness agents from Claude Code or Codex; multimodality comes from external bridges. The practical point is narrower than a product-launch headline: read the cited release or repository before upgrading, and do not infer performance gains where maintainers published none. Primary source

10. memory-recall-rerank 发布可复现的 Agent 记忆召回重排流水线

memory-recall-rerank publishes a reproducible gate/retrieve/rerank/rule/route pipeline on 18 synthetic Chinese memory cases. The practical point is narrower than a product-launch headline: read the cited release or repository before upgrading, and do not infer performance gains where maintainers published none. Primary source

Two entries have an additional first-party release page: OpenAI Agents SDK Python v0.21.1 and llama.cpp b10453. AI assisted with translation and editing; claims were checked against the linked sources, and none of these tools was independently benchmarked for this roundup.