Loading ToolBench
Read-only, provenance-first repository navigation and progressive disclosure for agents — a locator with pre-registered evidence behind its defaults (BM25-primary per doi:10.5281/zenodo.21969901). CLI + MCP server.
mentu-navigator provides 8 well-named, read-only tools with complete JSON schemas and detailed descriptions. Tool names follow verb_noun convention (navigator, locate, read_range). Descriptions are substantive (100 - 250 chars), explaining WHAT each tool does and WHEN to use it. All parameters have type definitions and descriptions. However, output schemas are not documented, LLMs cannot predict response structure for downstream chaining. Error handling is minimal; no recovery guidance or actionable error messages. Tool annotations (readOnlyHint) are absent despite all tools being read-only. Composition is strong: tools chain naturally (locate → read_range), and parameters accept human-friendly identifiers (repo paths, symbol names). No security issues detected (read-only, no secrets). Baseline: 549 production tools average 18-char names (these average 16 - 24 chars, within range); 94% of A+ tools have docstrings (8/8 present here); 100% of A+ tools document params (8/8 present). Missing output documentation and error guidance prevent a higher score.
Progressive disclosure entrypoint: fuse a ranked lexical leg and an exact leg into k=8 source ranges, each with path, line, heading-anchored range, a snippet of at most 240 characters, retriever attribution, and its reason. Read a range with read_range before making a claim.
Preferred entrypoint for agents. Auto-routes one repository question to a map, ranked query, frontmatter handles, symbol context, or Git change-impact report and returns compact source-backed evidence.
Specialist tool: relate an explicit Git change range to tests, contracts, ticket references, and risk signals.
Specialist tool: discover typed docs and relationships from bounded YAML frontmatter. Results are routing pointers and require reading the source body before claims.
Specialist tool: map repository structure, languages, contracts, manifests, and typed docs without reading secrets.
Specialist tool: find and rank source-backed repository evidence for a natural-language or exact query.
Output schemas not documented. LLMs cannot predict response structure (fields, types, pagination) for downstream tool chaining or data extraction. Baseline: 100% of A+ tools document return types.
No error handling guidance. Tools return errors but do not tell LLMs what to do next (retry, ask user, try alternative tool). Baseline: pattern:recovery-guide requires actionable error messages.
Tool annotations missing. All 8 tools are read-only but lack readOnlyHint annotation. This prevents clients from optimizing caching, retry logic, or permission checks. Baseline: current MCP spec (2026-07-28) rewards tool annotations.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 63 | 2026-07-28+ | v2 |
Specialist tool: collect definitions, references, tests, docs, and config surrounding a known symbol.
Return the source lines a locate hit points at. Widening is ±20 lines per step and stops at heading boundaries; frontmatter comes back in its own field and is never inlined into the body slice.
Parameter 'limit' lacks upper-bound justification in descriptions. Tools accept limit up to 200 but do not explain why or what happens if exceeded. Baseline: pattern:constrained-input requires format/range in descriptions.
Pagination not documented. 'locate' and 'navigator_query' accept 'k' and 'limit' but do not specify if results are truncated, whether a next_cursor exists, or how to fetch more. Baseline: pattern:paginated-result requires pagination metadata.