Agent Consumption¶

A bundle exists to be read by an agent. The read path is code.context, available as a CLI command and an MCP tool.

code.context¶

kl4a --use codekb context ./my-code-bundle --task explain --query "batch processing"

--task is a free-form task id describing what the agent is doing (explain, plan-change, find). --query is the term to retrieve around; when omitted, the task id itself is used as the query. --language narrows the search.

What comes back¶

Key	Contains
`task`	The task id and the query actually used
`symbols`	Matched symbols
`evidence`	The source excerpt behind each matched symbol
`relations`	Edges touching the matched symbols
`code_knowledge`	Usable claims about them
`tests`	Tests related to the selected symbols
`warnings`	What to be careful about in this result
`context_rules`	How the result is meant to be used

Retrieval widens by one hop: symbols matching the query, plus symbols reachable from them by a relation. Claims are returned when they concern any symbol in that set.

If the query matches nothing, retrieval falls back to a small sample of symbols rather than returning empty — so check task.query and the symbol list before concluding the bundle lacks something.

Warnings are part of the answer¶

Unresolved relations and review-required claims are reported alongside the result. An agent that ignores the warnings is reasoning past the edge of what the bundle actually knows.

What Is Filtered Out¶

Claims are excluded from code.context when rejected, when no longer active, or when Tier 4 and flagged for review. See Knowledge & Review.

Approving a Tier 4 claim makes it retrievable; deferring or commenting does not.

MCP¶

kl4a --use codekb mcp serve ./my-code-bundle

Serves the bundle over stdio as read-only MCP tools. The code surface:

Tool	Purpose
`code.repo.describe`	Describe the bundle
`code.files.search`	Search source files
`code.symbols.search`	Search symbols
`code.symbols.get`	One symbol, with evidence and relations
`code.relations.search`	Search edges by subject, predicate, object or status
`code.callgraph.neighborhood`	Relations around a node, to a given depth
`code.tests.for_symbol`	Tests related to a symbol
`code.change_impact`	Direct impact of changing a symbol
`code.context`	Task-ready context, as above

That is the whole surface, and all of it is read-only. There is no tool that writes to the bundle, so an agent cannot approve a claim, change review state, or alter the repository record — review remains human-controlled.

(sopkb's MCP server has an optional review.note writer, off unless explicitly enabled. codekb has no equivalent.)

Change Impact¶

Before editing a symbol, ask what depends on it:

kl4a --use codekb change-impact ./my-code-bundle symbol-app-answer-function
kl4a --use codekb tests for-symbol ./my-code-bundle symbol-app-answer-function

Impact is direct — callers and the relations touching the symbol — not a transitive closure. It is a starting point for a blast radius, not the whole one.