Scored across 47 tools
Most tools have distinct anchors and the descriptions help, but the set is dense enough that several pairings blur together: memory_validate vs memory_doctor, commit_search vs commits_touching_query, and the four domain-specific search tools (semantic_search, commit_search, papertrail_issue_search, rationale_search). An agent can usually pick correctly, but with 47 tools misselection risk is real.
Prefix families such as memory_* and papertrail_* are consistent, but the overall naming mixes noun phrases (symbol_lookup, repo_brief), verb phrases (heal_index, find_callers), and odd asymmetries like commit_search vs commits_touching_query. It is readable, but the convention is not uniform.
47 tools is far beyond the typical well-scoped tool surface and pushes the server into awkward density even for a broad code-intelligence domain. Many status, graph, search, and memory subdomains could be collapsed or grouped without losing capability.
The surface covers memory CRUD, code graph queries, semantic/lexical search, git history, papertrail evidence, diagnostics, and maintenance workflows quite thoroughly. The main gaps are that some advanced checks depend on CLI-side bootstrap work like 'oracle run', and source reading is chunk-id-centric rather than offering a direct path-based read fixture.
