Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

gauntlet

Codebase learning through knowledge extraction, challenges, and spaced repetition.

Overview

Gauntlet prevents knowledge atrophy for experienced developers and accelerates onboarding for new ones. It extracts knowledge from the codebase and tests understanding through adaptive challenges.

Installation

/plugin install gauntlet@claude-night-market

Skills

  • extract - Analyze codebase and build a knowledge base
  • challenge - Adaptive difficulty challenge session
  • onboard - Guided five-stage onboarding path
  • curate - Add or edit knowledge annotations

Commands

  • /gauntlet - Run an ad-hoc challenge session
  • /gauntlet-extract - Rebuild the knowledge base
  • /gauntlet-progress - Show accuracy stats and streak
  • /gauntlet-onboard - Start or resume onboarding
  • /gauntlet-curate - Add or edit a knowledge annotation

ML Scoring

Gauntlet uses a pluggable Scorer protocol to evaluate answers. Two implementations ship by default:

  • YamlScorer (default): heuristic scoring based on YAML rule files. Always available, no external dependencies.
  • OnnxSidecarScorer: upgrades scoring quality by calling the oracle sidecar daemon for ONNX model inference. Activates automatically when oracle is running.

The scorer selection is automatic. When oracle’s port file exists and the health check passes, gauntlet uses the sidecar scorer with configurable blend weights. When the sidecar is unavailable, it falls back to YamlScorer with no user intervention.

See oracle for daemon setup and ADR-0009 for the discovery pattern.

Variation Provider

Gauntlet varies challenge wording through a VariationProvider abstraction. Two implementations ship: an Anthropic-SDK provider (default outside Claude Code) and an in-loop provider (default inside Claude Code, issue #464) that perturbs integers and small numeric tokens deterministically. The in-loop provider avoids spawning a sibling Claude through the API when the parent LLM is already in-loop, and imports nothing from anthropic.

The challenge skill calls register_in_loop_provider_if_inside_claude_code() to switch providers based on runtime. Selection is automatic.

Code Knowledge Graph

The graph module builds a SQLite-backed knowledge graph using Tree-sitter parsing. GraphStore supports context manager usage for safe resource cleanup. Community detection groups related nodes, and blast radius analysis scores the risk of code changes using security keywords from constants.py.

Problem Bank

Curated algorithm problems in data/problems/*.yaml cover arrays, graphs, trees, dynamic programming, and 15 other categories. Each entry includes difficulty level and pattern metadata. The challenge engine draws from this bank for targeted practice sessions.

Agents

  • extractor - Autonomous knowledge extraction agent