Scored across 47 tools
Most tools target a distinct anchor or data source, and the search family is cleanly partitioned across source/docs, commits, tracker issues, rationale, and memories. A few near-overlaps exist—commit_search vs commits_touching_query and dream vs memory_doctor both touch stale-memory work—but explicit descriptions make misselection unlikely for most calls.
Names are uniformly snake_case and generally follow a readable pattern: verb_noun for operations like memory_create and find_callers, and domain_for_target for retrievals like papertrail_for_symbol and memory_for_path. Odd phrasings like commits_touching_query, compare_graph_to_text, and repo_clusters deviate from a strict verb_noun rhythm but remain predictable.
47 tools is a very large MCP surface, well beyond the 25-tool threshold the rubric flags as too many. The count is inflated by narrow query variants—papertrail_for_chunk/symbol/commit/path and memory_for_symbol/path/call_path could reasonably be parameterized anchors—so the set feels over-fragmented rather than tightly scoped.
The server covers the domain well: source/symbol/graph/history/tracker retrieval, a full memory lifecycle with edges, index maintenance, and a reviewable maintenance worklist. Minor gaps exist where required supporting actions are CLI-only, such as oracle run for SCIP comparisons and dream --verify/--compact, but core workflows have workarounds and do not dead-end.
