Building jev-scout: Zero-Hallucination Crate and Repo Scout

Autonomous coding agents frequently hallucinate dependencies. When instructed to add a feature, an agent will generate Cargo.toml or package.json entries for libraries that do not exist, or select abandoned repositories with open security advisories.

Rigid keyword search on GitHub fails because natural-language architectural queries ("low memory cdp browser controller in rust") do not match raw repo titles.

I built jev-scout to decouple discovery from evaluation.

+------------------+        +----------------------+        +---------------------+
| Natural Query    |        | jev-scout (Rust)     |        | TypeSafe AI (Jev)   |
| "fast sqlite tui"| -----> | crates.io / GitHub   | <----> | /v1/systemone       |
+------------------+        | Real Package Fetcher |        | Speculative Fan-Out |
                            +----------------------+        +---------------------+
                                       |
                                       +---> Ranked Real Dependencies (Fit 0-4.0)
                                       +---> Verified Cargo / Clone Commands

Two-Stage Architecture

Instead of asking an LLM to generate recommendations from memory, jev-scout executes two distinct phases:

1. Grounded registry fetching: Queries crates.io and the GitHub Search API to collect candidate packages. Only real, published codebases enter the pipeline.
2. System One speculative fan-out: Sends the candidates and the user query to TypeSafe AI's Jev model (jev-latest) in a single HTTP call. Jev scores architectural fit (0.0 to 4.0), license suitability, and maintenance activity with calibrated confidence.

src/
  main.rs       # Command-line flags (lexopt) and output formatting
  mcp.rs        # MCP JSON-RPC 2.0 stdio server
  crates.rs     # crates.io API client and response deserializer
  github.rs     # GitHub REST API search client
  eval.rs       # TypeSafe AI System One batch evaluator

CLI Usage

jev-scout is published on crates.io and installs in one command:

# Install binary from crates.io
cargo install jev-scout

# Search across crates.io and GitHub
jev-scout "fast sqlite tui in rust"

# Restrict search to Rust crates only
jev-scout "driverless cdp websocket" --ecosystem rust

# Restrict search to GitHub repositories
jev-scout "offline upi expense tracker" --ecosystem github

# Output raw JSON for script pipelines
jev-scout "token compression grep" --json

Terminal output displays exact star counts, license names, maintenance activity percentages, and ready-to-run installation commands:

Scouting repositories for: "fast sqlite tui in rust"
Found 3 candidates evaluated in 1420ms via TypeSafe Jev:

#1  sqlite-tui  [BEST MATCH]
    Downloads: 4,120 | Crates.io | Updated: 2026-07-14
    Terminal user interface for SQLite databases
    Fit: 3.8/4.0 (Conf: 0.96) | Active: 88%
    Command: cargo add sqlite-tui

Agent Integration

Coding agents use jev-scout via Model Context Protocol:

jev-scout --mcp

Before adding any dependency to a project plan, the agent queries scout_packages. If the package name does not appear in the verified registry results, the agent is blocked from writing it into project configuration files.