mentu-navigator

Read-only, provenance-first repository navigation for humans and AI agents.

Point it at a repository and ask a question. It returns ranked file-and-line ranges to read, each carrying the reason it ranked, and it never writes to the repository it reads. It is built for agents that must find the document that answers a question in a documentation-heavy repository, without pulling the whole thing into context, and for the people reviewing what those agents did.

Scope. The ranking default is measured on documentation: markdown corpora, in pre-registered studies (see Evidence). It has not been measured on source code, where published comparisons favour other locators (see Scope of the evidence). The code-structure commands (map, symbol, impact) work on any repository; their usefulness has not been measured.

mentu-navigator is the product name. mentu-nav is its short CLI.

Install

Requires Node.js 20 or newer.

npm install -g mentu-navigator     # CLI: mentu-nav, MCP server: mentu-navigator-mcp

To run it without installing:

npx -p mentu-navigator mentu-nav "where is retry handled?"

The package name and the command name differ, so npx mentu-navigator on its own cannot resolve a binary. Pass -p as above.

License: Apache-2.0. Telemetry is local-only JSONL under ~/.mentu/pd1/ (spec: docs/TELEMETRY-SPEC.md); disable with MENTU_NAV_TELEMETRY=off. Nothing ever leaves the machine.

Where it sits in Mentu

mentu-navigator is the retrieval surface. mentu-pdv validates the frontmatter schema this tool consumes and emits the demotion sets locate applies. The two are designed as a pair, and the schema is published as a spec (SPEC-frontmatter.md).

Why it exists

Search tools find matching text. A repository navigator also preserves the question's intent and returns the nearby contracts, tests, docs, Git lineage, and risk surfaces needed to act safely.

The first release is intentionally deterministic:

• no embeddings;
• no background index;
• no writes to target repositories;
• on-demand, bounded frontmatter handles that route to documents but never replace reading them;
• bounded snippets with file-and-line provenance;
• known secret-bearing paths excluded before content reads.

One-command start

cd /path/to/repository
mentu-nav
mentu-nav "where is DEMO-274 implemented and tested?"

Interactive terminals receive a concise human view. Pipes receive compact JSON; --json requests the full envelope. Agents should use --agent for a compact, token-efficient JSON contract:

mentu-nav --agent "where is DEMO-274 implemented and tested?"

The front door auto-routes to map, query, handles, symbol, or impact. Their explicit commands remain available for scripts and advanced use.

For a docs-as-code network:

mentu-nav handles "catalog lineage"

This returns metadata pointers, typed relationships, and diagnostics. Every pointer carries requiresHydration: true. query keeps these pointers separate from source-body evidence so a summary cannot silently become an answer.

Capabilities

Progressive disclosure: locate and read-range

locate is the agent surface. Its default arm is ranked lexical retrieval (BM25). That default was set by a pre-registered study, not by taste (see Evidence below). Two legs exist:

• a ranked lexical leg: in-memory Okapi BM25 over the same file set the walker already produces, with vendored Snowball stemmers for Spanish and English, routed by each document's lang frontmatter tag (detected as a fallback, and the detection is logged, never written);
• an exact leg: the deterministic query pipeline, unchanged in semantics.

A fused arm (reciprocal rank fusion of the two legs) exists as a measurement arm. It was the original default and was retired from the default path by its own pre-registered ablation rule when the registered bake-off found it trailing plain BM25 by 7.8 points of localization (see Evidence). Every hit says which leg (or both) put it there.

mentu-nav locate "compaction policy" --k 8
mentu-nav read-range docs/adr/ADR-014-ledger-compaction.md 38 62 --widen 1

locate returns {path, line, range, snippet, score, retriever, why}. That is a range to read, not an answer. read-range returns the slice; each --widen step reaches ±20 lines further and stops at the enclosing heading boundary, and frontmatter comes back in its own field so metadata cannot be mistaken for body evidence. Handles remain the pointer layer, unchanged: every pointer still carries requiresHydration: true.

Pinned parameters

These are design parameters, not implementation details. Each is registered as an ablation and measured there; changing one is a dated decision plus a re-measurement.

The legacy query command keeps its own human-facing default of 40 results; the pins above govern locate.

Demotions

--demotions <path> reads a pdv demotions JSON file (resolved against the repository root) and multiplies those documents' scores by 0.5. A demoted document ranks lower and is never removed. An unavailable document is the more expensive error. An unreadable or malformed demotion set is reported in the envelope diagnostics rather than silently ignored.

--retriever is for measurement

--retriever=bm25|exact|fused selects an arm (default bm25). It exists so a registered bake-off's arms are produced by the shipped code path rather than by a harness fork. That is exactly how the defaults here were decided. The flag is not a tuning knob.

Evidence

Every performance-relevant default in this tool traces to a registered, mechanically adjudicated study, and every claim below carries its scope: one 141-document bilingual operational documentation corpus, a fresh 115-question blind set, this tool's k=8 contract. The bake-off (doi:10.5281/zenodo.21969901, companion to doi:10.5281/zenodo.21960138):

• BM25 located the gold document on 93.0% of questions vs hardened exact search's 71.3% (McNemar p < 1e-5). That is why bm25 is the default.
• The fused arm trailed BM25-alone by 7.8 pp (p = 0.0225), failing its frozen "fusion never costs localization" prediction; the pre-registered ablation rule retired it from the default path (docs/build/D3-REVISION-2026-08-16.md).
• An off-the-shelf SQLite FTS5 control (89.6%) was not statistically distinguishable from this implementation. BM25 as such carries the gain.
• Downstream answer accuracy moved +5.2 pp under the better locator, almost entirely through localization.

Nothing here claims generality beyond that corpus class; the study, corpus manifest, question set, and adjudicator are public in the DOIs above for re-running.

Scope of the evidence

Every study behind these defaults used markdown documentation; none used source code, and the difference matters. Finding the file to change from an issue description is a lower-overlap task, and published comparisons rank BM25 below embedding retrieval there. On SWE-bench Lite, BM25 placed a correct file among its top five 61.7% of the time, against 84.7% for a code-embedding retriever (Chen et al., LocAgent, Table 4). So:

• for a question answered by a document (a runbook, an ADR, a specification), use locate, then read the range it returns;
• to find code from an issue or a behaviour description, prefer your harness's native search, or a code-aware locator, until a study on code exists;
• a pre-registered comparison of this tool inside a live agent, against the harness's native tools, is drafted with a code-localization stratum. It has not been run. When it has, this section will report its result either way.

What is additionally asserted by the test suite on every commit: the index writes nothing to a target repository and lives in memory for the life of the process; identical corpus and query produce byte-identical hit lists across runs and index rebuilds; secret-bearing paths are excluded before tokenization, not after; and a search pattern beginning with - is passed after a literal -- so it can never be parsed as an engine flag.

MCP

mentu-navigator-mcp exposes:

• navigator: preferred compact, auto-routing entrypoint
• locate: BM25-ranked ranges, with retriever and demotions
• read_range: the disclosure step locate hands off to
• navigator_map
• navigator_query
• navigator_handles
• navigator_symbol_context
• navigator_change_impact

An MCP client configuration:

{
  "mcpServers": {
    "mentu-navigator": {
      "command": "npx",
      "args": ["-y", "-p", "mentu-navigator", "mentu-navigator-mcp"]
    }
  }
}

or launch the installed binary mentu-navigator-mcp directly.

Agent setup

mentu-nav setup --target all
mentu-nav doctor --human

Setup links the bundled skill into Codex and Claude without copying its logic. It refuses to replace an existing path. Repository navigation itself remains read-only.

Adoption

The executable remains centralized. Repositories adopt only a short operating contract; they do not copy the implementation. See docs/adoption.md.