Search what the workspace knows

Knowledge search answers “have we already worked this out?” from your own work: huddles, tasks, summaries, artifacts, research findings, and indexed repository files. An agent searches the project it is working in, and looks wider when the question calls for it.

Where it appears

There is no search box of its own. Knowledge search is a tool agents use, on three surfaces:

• With Eva. Ask her a question about your own work and she searches the corpus to answer it, quoting what she read.
• In huddles. An agent that needs to know what was decided before, or whether something already exists, searches rather than guessing or asking you.
• In coding runs. The coding agent on a runner reaches the same corpus through a small local shim, so it can check for an existing implementation before writing a new one. See knowledge in runs.

Because it is a tool rather than a page, the way to use it is to ask a question that needs it: “what did we decide about offline caching”, “has anyone looked at trail conditions staleness”, “where is the tile eviction rule written down”.

What it can see

Everything your workspace has produced, indexed as work happens rather than on a schedule:

• huddle messages and huddle summaries,
• tasks and their specs,
• workspace items, artifacts, and research findings,
• content assets,
• repository files, for repositories that have been indexed.

Results are ranked by relevance rather than recency, and each one carries what it came from, so an answer can be traced back to the huddle or task that said it.

How far it reaches

A search covers the project the agent is working in: the huddle’s project, the task’s project, or the project a conversation with Eva is about.

An agent reaches further only on purpose, and there are two ways it does:

• Projects you name. “Check how we positioned Summitline App” points the search at that project instead.
• Every project. A question that is plainly about the company rather than one product, asked in a huddle whose title and objective say so, is searched across the whole workspace.

A conversation with Eva that belongs to no project has nothing to narrow to. It searches everything, and the answer says so rather than leaving you to assume.

Whatever comes back from outside the project you are in is labelled with the project it came from, so you can tell a claim about this product from a claim about another one. See knowledge for why the default is narrow.

What it cannot see

• Anything outside your workspace. A search never crosses into another workspace’s corpus. For questions about the outside world, use a research job.
• Other projects, unless it is asked. Work in another project is there to be found, but not by default. Name the project, or say the question is about all of them.
• Anything deleted. Removing the source removes its passages.
• Secrets and environment values. Those are never indexed.

When a search comes back empty

An empty result means nothing in the project being searched answers the question, which is a real answer: it usually means the thing was decided in a channel outside CommandChain. It does not mean the search is broken, and agents are told to treat it that way rather than inventing an answer.

If you believe the answer exists but in another piece of work, say which project, or say the question is about every product. The search runs again with that reach.

Next steps

• Run a research job: for questions the workspace cannot answer.
• Knowledge: what the corpus is, in full.
• Knowledge in runs: how a coding agent reaches the same corpus.