Asking Questions

/c4a:context-inspect-search answers questions from approved knowledge and traces answers back to their sources. It reads the knowledge first, because that is where conclusions are already worked out; only when the knowledge leaves a gap that could change the answer does it read the original documents or code. Every answer cites what was actually read.

flowchart LR
  q["Question"] --> k["Search approved knowledge<br/>or a package"]
  k -->|enough| a["Answer with citations"]
  k -->|gap that matters| s["Trace recorded sources<br/>at their pinned versions"]
  s --> a
  a -.->|knowledge missing or stale| u["Suggest an update<br/>starts only if you accept"]

Ask

/c4a:context-inspect-search Why does a refund stay pending after the
provider confirms it? Cite the code.

A knowledge workspace is optional: the Skill can answer from installed or supplied knowledge packages alone. Querying is read-only — it never changes source registrations, snapshots, tasks, approvals or your checkouts, and it never builds just to answer.

How it searches

• It searches approved knowledge/ with business terms, symbols and synonyms, chooses articles from match excerpts rather than filenames, and reads the relevant sections with their conditions.
• Zero hits lead to relaxed or equivalent terms; noisy results lead to a narrower path. A miss alone does not establish that something is absent.
• When the knowledge leaves a gap that could change the conclusion, it traces the recorded sources directly, at the commit or snapshot the knowledge cites — never the default branch in its place. One failed source affects only the claims that depend on it.
• If the knowledge already answers the question, it does not restore code or check every associated source.

Retrieval order and access

The host or Bot configuration can set two independent policies:

Setting	Values	Effect
CONTEXT_QUERY_SOURCE_MODE	repo-first (default), package-first, dual	Read the knowledge workspace's approved knowledge first, a package first, or both at once and reconcile differences
CONTEXT_QUERY_ACCESS_MODE	auto (default), remote, local	Whether knowledge and code are read through a remote read-only code service or local checkouts

They change the order and access path, not what counts as evidence. With remote, see Reading Code Evidence Remotely.

What an answer looks like

The answer comes first, with the conditions that change it right beside it. Complex questions also get the mechanism behind the result, with a short excerpt when it helps you verify. Implementation, declared contracts, tests, runtime observations and inferences are kept distinct.

Citations link to what was read: source files at the inspected commit, knowledge site articles and original documents. When no reliable link exists, you get checkable paths, symbols and short excerpts instead. Site links come from the package's context-site-map.json; a link is not a claim that the site has deployed that version.

When the knowledge is missing

The Skill scopes any gap to what it inspected — an unanswered question in one article is not absence across the knowledge base. With supporting evidence it may suggest an update: the affected page or gap, the proposed change and the source scope. Nothing starts until you accept; then it hands over to the production entry in the selected knowledge workspace. See Capturing from Conversations.

提问

/c4a:context-inspect-search 基于已批准的知识回答问题，并把答案追溯到来源。它会先读知识，因为结论已经在那里梳理好了；只有当知识留下的缺口可能改变答案时，它才会去读原始文档或代码。每个回答都会引用实际读过的内容。

flowchart LR
  q["问题"] --> k["搜索已批准的知识<br/>或知识包"]
  k -->|足够| a["附带引用的回答"]
  k -->|关键缺口| s["按固定版本<br/>追溯已记录的来源"]
  s --> a
  a -.->|知识缺失或过时| u["建议更新<br/>仅在你接受后开始"]

提问方式

/c4a:context-inspect-search Why does a refund stay pending after the
provider confirms it? Cite the code.

知识工作区不是必需的：该 Skill 可以只基于已安装或提供的知识包作答。查询是只读的——它从不修改来源登记、快照、任务、批准或你的代码检出，也从不为了回答而执行构建。

如何搜索

• 它用业务术语、符号和同义词搜索已批准的 knowledge/，依据匹配摘录而不是文件名来选择文章，并连同适用条件阅读相关章节。
• 没有命中时会放宽或换用等价词；结果太杂时会缩小路径。单次未命中并不能证明某样东西不存在。
• 当知识留下的缺口可能改变结论时，它会直接追溯已记录的来源，使用知识所引用的 commit 或快照——绝不用默认分支代替。某个来源读取失败，只影响依赖它的那些论断。
• 如果知识已经回答了问题，它不会恢复代码，也不会逐一检查所有关联来源。

检索顺序与访问方式

宿主或 Bot 配置可以设置两个相互独立的策略：

设置	取值	效果
CONTEXT_QUERY_SOURCE_MODE	repo-first（默认）、package-first、dual	先读知识工作区中已批准的知识、先读知识包，或同时读取两者并调和差异
CONTEXT_QUERY_ACCESS_MODE	auto（默认）、remote、local	通过远程只读代码服务还是本地代码检出来读取知识和代码

它们改变的是顺序和访问路径，而不是什么算作证据。使用 remote 时，参见远程读取代码证据。

回答是什么样的

答案放在最前面，会改变答案的条件紧随其旁。复杂问题还会给出结果背后的机制，并在有助于你核实时附上简短摘录。实现、声明的契约、测试、运行时观察和推断会被明确区分。

引用会链接到实际读过的内容：所检查 commit 下的源文件、知识站点文章和原始文档。没有可靠链接时，会给出可核查的路径、符号和简短摘录。站点链接来自知识包的 context-site-map.json；链接并不代表站点已经部署了该版本。

知识缺失时

该 Skill 会把任何缺口限定在它所检查的范围内——某篇文章中未回答的问题，并不等于整个知识库都没有。有支持证据时，它可能会建议更新：受影响的页面或缺口、拟议的修改以及来源范围。在你接受之前什么都不会开始；接受后，它会移交给所选知识工作区的生产入口。参见从对话中采集。