car-analyzer

Reference: https://harn.app/kb/evals.html — "Deterministic verifiers over LLM-as-judge"

CAR Framework Analyzer

Audit the current project against the three layers of a robust harness.

CAR Scorecard (check each item, tally points)

Control Layer (40 points)

• AGENTS.md or CLAUDE.md exists and < 100 lines (10 pts)
• Type checker configured and runs without errors (10 pts)
• Linter configured (eslint/clippy/ruff/golangci-lint) (5 pts)
• Test directory exists with actual test files (10 pts)
• Architectural constraints documented in AGENTS.md (5 pts)

Agency Layer (25 points)

• PreToolUse security guard hook active (10 pts)
• Tool permissions explicitly scoped (5 pts)
• Sub-agent delegation patterns documented (5 pts)
• File write boundaries defined (5 pts)

Runtime Layer (35 points)

• Stop hook (quality gate) active and verified (15 pts)
• PostToolUse hooks configured (formatter and/or trace) (5 pts)
• Infinite-loop prevention in Stop hook verified (10 pts)
• Session recovery artifacts exist (CHECKPOINT.json or LEARNED.md) (5 pts)

Scoring

• 80-100: Production-grade harness
• 60-79: Functional with gaps — run specific /harn: commands to fill
• 40-59: Needs hardening — run /harn:init to scaffold missing components
• <40: Minimal harness — strongly recommend full /harn:init

Remediation Guide

For each missing item, suggest the specific /harn:* command:

• Missing security guard → /harn:guard
• Missing quality gate → /harn:gate
• Missing formatter → /harn:format
• Missing trace logging → /harn:trace
• Missing checkpoint → /harn:checkpoint
• Missing AGENTS.md → /harn:agents-md
• Missing context management → /harn:context