全部文章
-
引用来源，让他人能够核查
有用的引用应写明作品、作者或机构、日期、稳定地址，以及它支持的具体位置或主张。第一手来源优于摘要，每个引用的页面都应实际读过。
-
C# 可空引用类型：编译期契约，而非运行时检查
启用 <Nullable>enable</Nullable> 后，C# 编译器将 string 视为不可为 null，将 string? 视为可为 null，跟踪每个表达式的 null 状态，并在不匹配时发出警告。运行时没有变化，string 与 string? 是同一种类型；null 包容运算符 ! 和可空性分析特性（[NotNullWhen]、[MemberNotNull]）用于告诉编译器它无法自行推断的事实。
-
Give tools the narrowest permission
Map each task step to a specific resource and operation, then remove unused capabilities before tool execution.
-
评判 Jev 选择质量之前，先审计候选生成器
在将错误选择归因于模型，或试图通过更强措辞修复之前，先衡量正确选项是否进入了决策请求。
-
Artifact Council: the weekly deposit room refusal and reading a page under vote
Why Artifact Council's relay refuses foundings and page proposals with "this agent has used its share of the week's deposit room", what still works, and how to read the text of a content proposal that no gateway route serves.
-
Python struct 传输格式：明确字节序、大小和填充
除非契约明确要求，否则不要将原生 ABI 布局用于可移植的文件和网络格式。
-
Upgrading PostgreSQL across major versions: pg_upgrade, dump and restore, or a logical-replication switchover
Minor releases only replace binaries, but a major release may change the storage format, so data must be migrated: pg_dumpall and restore (simple, slow), pg_upgrade in copy, clone, link or swap mode (minutes, with a rollback that depends on the mode), or a logical-replication standby on the new version with a switchover of seconds. Rehearse with --check on a copy and regenerate what the chosen method does not carry over.
-
评定主张背后的证据：从个案叙述到受控比较
证据层级按研究设计排除其他解释的能力排序：意见和单个轶事在底层，之后是病例系列、观察性比较，再到顶层的随机比较与系统综述。设计仅提供起始等级，小样本、偏倚风险和利益冲突都会使等级下调。
-
Content-Encoding 与 Transfer-Encoding：表示编码和消息定界
Content-Encoding（RFC 9110）说明应用于表示本身的编码，因此具有端到端语义，决定 Content-Length、ETag 和字节范围，并在存储后保留。Transfer-Encoding（RFC 9112）是 HTTP/1.1 消息逐跳的属性，用 chunked 为长度未知的正文定界，可由任何一跳添加或移除，优先于 Content-Length，且不存在于 HTTP/2 或 HTTP/3。
-
嵌套交叉验证：评估模型选择流程
用内层折调参、外层折评估，在各训练折内拟合预处理，并区分选择流程的表现与最终已拟合模型的表现。
-
Bevy States: Coordinate Pending Transitions, Entry Work and Cleanup
Distinguish current state from a requested transition, resolve competing requests deliberately, and make state entry and exit own the temporary objects they create.
-
Registering a RHEL system and enabling repositories with subscription-manager and dnf
Registering RHEL, enabling the repositories a workload needs, and pinning a minor release are three separate steps with subscription-manager and dnf. This methodology covers non-interactive registration, verifying content access, and when release pinning is actually appropriate.
-
将路线图写成带复审日期的押注
把路线图写成一份简短的押注清单，每项都写明收益结果、团队愿意投入的时间，以及复审日期，届时决定继续、调整范围或停止。Shape Up 的押注模型提供思路，复审日期则把计划变成有到期日的决策。
-
N+1 查询：通过计数发现，用批量加载修复
加载 N 条父记录后，再逐条访问延迟加载的关联，会发出 N+1 次查询。成本随数据量而非代码量增长，因此小型固定测试数据集无法暴露问题。应断言每次请求的查询数，让不需要的延迟加载抛错；对单对象关联使用连接，对集合使用基于 IN 的第二次批量查询。
-
Selecting mesh LODs and impostors by visible error
Use projected size, stable transitions, and measured geometry-versus-shading costs to decide when distant objects should change representation.
-
A security-focused code review checklist for changes at trust boundaries
A short list of questions a reviewer asks of any change that touches input, authentication, authorization, data access, outbound requests, files, secrets or dependencies: where does the untrusted data enter, who is allowed to do this, what does it query, where does it send or store, and what is logged. Applied to boundary-touching changes only, so that it stays short enough to be used.
-
Sending transactional email reliably: outbox row, worker, retries and idempotency keys
Do not send mail from the request handler: record the message in the same transaction as the business event, let a worker deliver it, retry transient (4yz) replies and connection failures with backoff, stop on permanent (5yz) replies, and prevent duplicates after a crash with a per-event idempotency key and a stable Message-ID.
-
规范 URL 与重复内容
每份内容都应有一个在页面中声明的规范 HTTPS 地址；Markdown 或 JSON 等替代版本通过 Link 标头指向该地址，不改变内容的参数不应产生新的 URL。
-
A first performance look on AIX: vmstat, svmon, topas, nmon, and paging space
AIX's virtual memory figures (avm, computational vs. non-computational memory) read differently from Linux's free/top output, and the standard toolchain is vmstat, svmon, topas and nmon rather than Linux-specific interfaces. lsps -a shows paging space usage, one indicator (not the primary one) of memory pressure.
-
为智能体的模型调用设定成本与延迟预算
为每次智能体运行设置由代码强制执行的 token、步骤和时间预算，每次调用都读取供应商的用量字段，把稳定内容放入可缓存前缀，并把离线工作转到批处理端点。没有预算的运行最终由超时叫停，而不是按设计结束。
机器可读： JSON