Code intelligence

The hub indexes each repo into an AST graph (GitNexus) and a vector index (Qdrant). Agents query both through MCP, so they can find code in any repo of their organization without cloning it.

Pick the tool by what you already know

You already know	Start with	Why
A symbol name (buildHybridQuery)	cortex_code_context(name)	Exact graph lookup. Callers, callees and imports in one call
Only the behaviour ("where do we verify API keys?")	cortex_code_search(query, limit: 10)	Ranked hybrid search over the index
An exact literal: env var, config key, error string	rg / grep	Not a ranking problem; the string is there or it is not
A relationship ("who calls X across packages")	cortex_cypher	One graph query instead of several searches
Whether someone else is in this file	cortex_changes	Before you edit anything shared

This table is also what /cs tells the agent at the start of each session.

Search once, then read all ten

On this repo's own index (benchmarks/retrieval_bench.ts, a hand-written set of 15 queries) the target file is in the top 10 for 15 of 15 queries, in the top 3 for 11 and at rank 1 for 8. So:

• Read the whole list before you choose. Acting on hit #1 alone is wrong about four times in ten.
• Don't reword the same question. With recall@10 at 1.000, a paraphrase returns the same set. If nothing fits, ask a different question or switch tool.

cortex_code_search defaults to limit: 5; pass limit: 10. Its response says which retrieval it used: hybrid (vector and BM25 fused with reciprocal rank fusion) or vector. A repo gets hybrid search the next time it is re-indexed. The numbers behind the choice are in Benchmarks.

The tools

Tool	Use it for
cortex_code_search	Find code by behaviour. repo, branch, limit
cortex_code_context	A symbol's definition, callers, callees and imports. name, optional file
cortex_code_impact	Blast radius before you change something exported. direction: downstream (default) or upstream
cortex_detect_changes	Symbols a diff touches, and a risk level
cortex_cypher	Raw Cypher over the graph. Nodes have id, name, filePath, startLine, endLine; use labels(n) for the kind
cortex_list_repos	The repos a cross-repo search covers
cortex_code_reindex	Re-index a repo and branch after a push

Parameters for each are in the MCP tools reference.

Across repos

Projects belong to an organization. Put a product's client, server and tools in the same one; that is the boundary of a cross-repo search.

cortex_code_search(query: "login packet")                  # every repo in the organization
cortex_code_search(query: "user auth", repo: "my-backend")  # one repo
cortex_code_context(name: "validateToken", repo: "my-backend")
cortex_list_repos()                                         # what "every repo" covers

The organization comes from org if you pass it, else from the repo you named, else from your latest cortex_session_start (its project.orgId). If you belong to several organizations and none of those applies, the search is refused rather than widened.

Before a commit, after a push

The hub cannot see your working tree, so cortex_detect_changes needs the diff:

git diff --staged
# for a large change, only the headers:
git diff --staged -U0 | grep -E '^(diff |--- |\+\+\+ |@@ )'

risk_level: "unknown" means a lookup failed, not that the change is safe.

After you push, cortex_code_reindex(repo, branch) keeps the graph current. Branch indexes are dropped after BRANCH_RETENTION_DAYS (default 14) without a re-index, or when the branch is merged or deleted.