ContextManager

Return the aggregated in-and-out neighbors of a graph node. Before answering, the graph is synchronized: built when the solution has no graph yet, otherwise only the invalidated contributions are re-extracted; source and MSBuild input files are compared by content hash, while changes that only alter environment-dependent MSBuild evaluation are not detected (run project_scan to force re-evaluation); a failed synchronization returns the same error as project_scan and keeps the previous graph. Type queries include the edges of the type's own members, so consumers that only call into its methods still appear; neighbor member nodes are rolled up to their declaring type. Each entry carries: id (rolled-up node ID), kind (Class/Interface/Record), direction ("out" = the queried node depends on it, "in" = it depends on the queried node), edgeKinds (per-edge-kind counts, e.g. {"CALLS": 3, "REFERENCES": 1}), and file (the neighbor's declaring source file, when known, ready for inspect_file). One entry is emitted per (id, direction). The queried node's own CONTAINS member edges are excluded — use inspect_file for a type's members. Output order is out-entries first, then in-entries, in graph-encounter order.

Performs an impact-specific backward traversal from a node following CALLS, INJECTS, REFERENCES, RETURNS, and INHERITS edges. Before answering, the graph is synchronized: built when the solution has no graph yet, otherwise only the invalidated contributions are re-extracted; source and MSBuild input files are compared by content hash, while changes that only alter environment-dependent MSBuild evaluation are not detected (run project_scan to force re-evaluation); a failed synchronization returns the same error as project_scan and keeps the previous graph. Seeds the BFS with the node, its direct members (via CONTAINS), and any interfaces it implements (via IMPLEMENTS) plus their members, so member-level dependencies aggregate to their declaring type. The implementors of every interface reached are included. Member nodes reached during traversal are rolled up to their declaring type. Returns an object with 'affectedIds' (ordered list of impacted type IDs) and 'affectedFiles' (each returned affected ID mapped to its declaring source file, when known, ready for inspect_file).

Find the directed shortest path between two nodes in the knowledge graph. Before answering, the graph is synchronized: built when the solution has no graph yet, otherwise only the invalidated contributions are re-extracted; source and MSBuild input files are compared by content hash, while changes that only alter environment-dependent MSBuild evaluation are not detected (run project_scan to force re-evaluation); a failed synchronization returns the same error as project_scan and keeps the previous graph. Returns an ordered list of node IDs from source to target (inclusive). Returns an error if either node is not found or no directed path exists.

Discover files and methods by possible names when no file path or node ID is known. A term matches an ordinal case-insensitive substring of a file name without .cs or a method name without its signature; any term is enough. Constructors match their type name. Searches every Method node in the graph, including private methods; directory paths, parameter types, bodies and namespaces are not searched. Terms are trimmed before matching; an empty term list, or any term shorter than 3 characters after trimming, rejects the entire query with invalid_query before synchronization. Selects exact matches before partial matches, each in graph encounter order, without boosting names that match several terms, up to 50 unique file/method matches before grouping by file (one per matching file and one per matching method ID). Returns compact JSON with results and truncated; each result always has file, namespaces (every namespace in the file), fileMatched and methods, and each method has name and nodeId. Arrays remain present even when empty. fileMatched is true only when the file-name match was kept within the limit; a file can appear through a kept method while its own name match was omitted. truncated means additional matches were omitted; there is no pagination, so narrow the terms. Pass a returned file to inspect_file, or a nodeId unchanged to the other graph tools. Before answering, the graph is synchronized: built when the solution has no graph yet, otherwise only the invalidated contributions are re-extracted; source and MSBuild input files are compared by content hash, while changes that only alter environment-dependent MSBuild evaluation are not detected (run project_scan to force re-evaluation); a failed synchronization returns the same error as project_scan and keeps the previous graph.

Analyze cross-file relationships in up to 15 C# source files using Roslyn semantic model. BCL/framework types (e.g. CancellationToken, Task) are resolved via framework metadata and EXCLUDED from 'unresolved' — that list only contains user types missing from the input set. Output is a compressed format by design: methods are one-line strings ('Name(params): ReturnType', prefixed with the access level when not public, e.g. 'protected Name(params): ReturnType'), no properties, no line numbers. There is no parity with inspect_file — use inspect_file for full per-type detail.

Returns a structural JSON contract for a single C# file: its absolute path, namespace, usings, and each non-private type with its kind, base type, implemented interfaces, attributes, constructor dependencies, methods (with line numbers), properties, and events.

Scan a .NET solution and bring its knowledge graph up to date from loaded C# source documents. The first scan builds the graph; later scans re-evaluate the solution with MSBuild and re-extract only the documents of projects whose sources, evaluated membership, or compilation inputs changed (plus the projects that depend on them), reusing every other contribution. The result reports project/document coverage, unsupported languages, and skipped documents. Persistence and in-memory publication succeed together; failed, empty, or cancelled scans preserve the previous graph. The graph is persisted in .context-manager/<solution name>.graph.json next to the .sln, so every session working on the solution sees the latest scan.