v4.0.0 — Claude Code-Inspired Evolution Intelligence

12 new features making the evolution loop smarter, cheaper, and more autonomous

New tools (7)

Tool	Purpose
validate_state.py	Validate .evolver.json against LangSmith state before each iteration
iteration_gate.py	Three-gate check: score plateau, cost budget, convergence detection
regression_tracker.py	Track failing→passing transitions, auto-add regression guard examples
consolidate.py	Cross-iteration memory consolidation (orient/gather/consolidate/prune)
synthesize_strategy.py	Generate targeted strategy document for proposers
add_evaluator.py	Programmatically add/remove evaluators from config
adversarial_inject.py	Detect memorization + inject adversarial test variations

New agent

• evolver-consolidator — background agent for cross-iteration memory consolidation (autoDream-inspired)

Upgraded agents

• Critic → Active Critic v3.1: detects gaming AND implements stricter evaluators
• Architect → ULTRAPLAN Mode: runs with model: opus for deep architectural analysis
• Proposer → 16-turn budget, per-strategy tool restrictions, KV cache-optimized shared prompt prefix
• TestGen → adversarial injection mode for anti-distillation

Evolution loop enhancements

• State validation before each iteration (skeptical memory pattern)
• Coordinator synthesis phase generates strategy.md instead of passing raw traces
• Regression tracking with auto-guard example injection
• Three-gate iteration triggers replace blind N-iteration loops
• Self-scheduling: background and cron-based modes for unattended optimization

Stats

• 17 files changed, +4,181 lines
• 7 new Python tools + 1 new agent
• Version: 3.3.1 → 4.0.0

Full changelog: https://github.com/raphaelchristi/harness-evolver/blob/main/CHANGELOG.md