Codebase Memory

Codebase Memory — Knowledge Graph Tools

Graph tools return precise structural results in ~500 tokens vs ~80K for grep.

Quick Decision Matrix

Exploration Workflow

1. list_projects — check if project is indexed
2. get_graph_schema — understand node/edge types
3. search_graph(label="Function", name_pattern=".*Pattern.*") — find code
4. get_code_snippet(qualified_name="project.path.FuncName") — read source

Tracing Workflow

1. search_graph(name_pattern=".*FuncName.*") — discover exact name
2. trace_path(function_name="FuncName", direction="both", depth=3) — trace
3. detect_changes() — map git diff to affected symbols

Evidence Tiers

• Scout (Tier 1): fast positive lookup with few graph calls and targeted source checks. Treat results as provisional; never make absence, exhaustive, dead-code, or complete-impact claims.
• Verify (Tier 2, default): task-directed searches, relevant trace directions, exact snippets for material claims, and all relevant result pages.
• Auditor (Tier 3): bounded-scope full verification with a current graph generation, complete relevant pagination, both call directions and broader relationships when material, plus explicit unresolved limitations.
• Every tier: after candidate paths are known, call check_index_coverage once with every evidence path. For negative or exhaustive claims also include the relevant scopes. A clean result means no recorded gap, not proof of completeness. For partial, skipped, excluded, stale, pending, or unknown coverage, read/grep the reported ranges or scope before relying on the graph.

Sessions and Subagents

• At session start or after compaction, call list_projects/index_status before structural exploration, then choose Scout, Verify, or Auditor for the task.
• Before delegating, query the graph and coverage in the parent. Pass the tier, exact project, generation/freshness, bounded scope, queries and pagination state, qualified symbols, paths, call-chain findings, coverage ranges/reasons, source fallback already performed, and unresolved questions to the child.
• Runtimes that isolate child context (subagents without a shared conversation) need the graph findings passed explicitly in the delegation payload; do not assume the child inherits MCP access or the parent's conversation.
• A child without MCP tools must not call or claim MCP access. It should work from the supplied evidence and use read/grep on exact source, especially every reported missed-coverage range.

Quality Analysis

• Dead code: search_graph(max_degree=0, exclude_entry_points=true)
• High fan-out: search_graph(min_degree=10, relationship="CALLS", direction="outbound")
• High fan-in: search_graph(min_degree=10, relationship="CALLS", direction="inbound")

15 MCP Tools

index_repository, index_status, list_projects, delete_project, search_graph, search_code, trace_path, detect_changes, query_graph, get_graph_schema, get_code_snippet, get_architecture, check_index_coverage, manage_adr, ingest_traces

Edge Types

CALLS, HTTP_CALLS, ASYNC_CALLS, DATA_FLOWS, IMPORTS, DEFINES, DEFINES_METHOD, HANDLES, IMPLEMENTS, OVERRIDE, USAGE, CALL_REFERENCE, CONFIGURES, FILE_CHANGES_WITH, SIMILAR_TO, SEMANTICALLY_RELATED, CONTAINS_FILE, CONTAINS_FOLDER, CONTAINS_PACKAGE

Cypher Examples (for query_graph)

MATCH (a)-[r:HTTP_CALLS]->(b) RETURN a.name, b.name, r.url_path, r.confidence LIMIT 20
MATCH (f:Function) WHERE f.name =~ '.*Handler.*' RETURN f.name, f.file_path
MATCH (a)-[r:CALLS]->(b) WHERE a.name = 'main' RETURN b.name

Gotchas

1. search_graph(relationship="HTTP_CALLS") filters nodes by degree — use query_graph with Cypher to see actual edges.
2. query_graph has a 100k row ceiling — add a Cypher LIMIT for broad queries or use search_graph pagination.
3. trace_path needs exact names — use search_graph(name_pattern=...) first.
4. direction="outbound" misses cross-service callers — use direction="both".
5. search_graph results default to 50 per page — check has_more and use offset.