---

name: challenge description: Presents adaptive codebase challenge questions with multiple-choice and trace exercises. Use when testing contributor knowledge of the codebase. model_hint: standard

Run Gauntlet Challenge

Present challenges from the knowledge base and evaluate answers.

When NOT To Use

• The knowledge base does not exist yet (use gauntlet:extract)
• A staged path for a new contributor (use gauntlet:onboard)

In-Loop Provider Setup

Before generating a challenge, register the in-loop variation provider so we do not call out to the Anthropic API just to spawn a sibling Claude (issue #464). Outside Claude Code this is a no-op and the default Anthropic provider remains active.

from gauntlet.providers.in_loop import (
    register_in_loop_provider_if_inside_claude_code,
)
register_in_loop_provider_if_inside_claude_code()

Steps

1. Load state: read .gauntlet/knowledge.json and developer progress

2. Check for pending challenge: if .gauntlet/state/pending_challenge.json exists, evaluate the developer's most recent message as an answer before generating a new one

3. Generate challenge: use adaptive weighting to select a knowledge entry and challenge type

4. Present challenge: show the question with context

5. Evaluate answer: score the response (pass/partial/fail)

6. Record result: update developer progress and streak

7. On pass: write pass token if from pre-commit gate. Show next challenge if in session.

8. On fail: show correct answer with explanation. Present a new challenge.

Scoring

Result	Score	Streak
Pass	1.0	+1
Partial	0.5	reset
Fail	0.0	reset

Exit Criteria

• .gauntlet/knowledge.json exists and is readable before a challenge is generated; if missing, the skill surfaces the error and suggests running gauntlet:extract
• Each challenge attempt results in a score (1.0/0.5/0.0) written to the developer's progress store and a streak update applied
• .gauntlet/state/pending_challenge.json is evaluated before generating a new challenge when it exists; the file is removed or updated after evaluation
• On fail, the correct answer and explanation are shown before the next challenge is presented (not skipped silently)