Skip to Main Content

Adaptive Change Harness is a self-evolving verification and repair system for real codebases. Its core value is not just that it can call a model to suggest a patch, but that it wraps models inside a reproducible evidence loop: ingest a repo, discover a real latent failure, save the exact repro, repair it with grounded context, rerun deterministic validation, and then convert validated fixes into reusable repair skills. The product gets stronger over time because each successful repair becomes structured runtime knowledge that can be matched, reused, and improved on future failures. The most eye-catching part is that it turns model output into hard operational proof. Instead of stopping at “the AI thinks this bug is fixed,” it shows failure before repair, no failure after repair, and passing baseline tests, all inside an operator-facing run console. It can capture latent bugs that baseline tests missed, replay them deterministically, generate and apply a patch, and end with a safe or unsafe verdict backed by evidence. On top of that, it automatically creates or updates a skill library from validated repairs, making the system visibly self-evolving rather than a one-off patch generator. Standout features: • Real repo intake and profiling from uploaded zip files • Latent failure discovery with saved reproducible failure cases • Deterministic replay-based repair validation • OpenAI-powered grounded diagnosis and patch generation • Evidence-first verdicts instead of confidence-only outputs • Automatic creation, reuse, and revision of repair skills in skill_assets/ • A run console that makes the full discover -> repair -> validate -> learn loop visible end to end

Team

hades

More from The IDE Reimagined: JetBrains Codex Hackathon

We built a self-healing coding agent that searches over compact DSL implementation plans before generating code. The LLM proposes semantic reasoning branches…

thinking project preview

Scopecreep turns your JetBrains IDE into an agentic hardware testbench. Describe a test in plain English; an AI agent writes Python directly into your open…

Scopecreep project preview

Mesh Code: shared session state for coding agents. Install mesh-code-mcp in any IDE and any agent — Claude, Cursor, Codex — sees the same feature card, event…

Periscope is a local-first viewer that shows you exactly what's in your coding agent's context window, turn by turn, and tells you what to do about it. What…

Latent Signal project preview

SecureLoop turns production Sentry alerts into Codex-generated, sandbox-tested, human-approved fixes inside JetBrains.

SecureLoop project preview

Pinpoint is a visual annotation overlay for frontend review. Not just that, it also helps pinpoint things in complex software :)