Khwand

Every commit to your repo triggers a Khwand webhook. The platform intercepts the push, generates tests against edge cases your prompt never considered, runs those tests across model versions to catch drift, and attempts to auto-patch any failures — all before the PR merges. The vendor describes this as a self-healing deployment pipeline rather than a passive test runner. Output surfaces in a dashboard with a Stability Score, confidence ratings on auto-fixes, and alerts when a model update introduces logic regression.

The cross-model prompt regression check is the feature most CI tools skip entirely. When OpenAI or Anthropic ships a model update, Khwand runs your existing agent logic against the new model version and flags behavioral drift before it reaches production. The vendor’s own framing — ‘deployment blocked’ on detected logic regression — suggests this is a hard gate, not an advisory warning. That distinction matters when you’re the team that woke up to customer complaints after a silent model rollover.

The AST-based security scanner adds a layer specifically designed for agents that call external tools: it detects prompt injection attempts, insecure tool access patterns, and data exfiltration signatures at the code structure level, not just at runtime. Input sanitization covers XSS, SQL injection, and path traversal. For teams that skipped these checks because they shipped fast, this is the assurance layer that catches what vibe-coded speed leaves behind.

Khwand is a hosted GitHub App with no self-host option and no open-source repository or license visible on the product page — the vendor describes it as open-source, but no source code or self-host instructions appear in the published documentation. The failure-pattern dataset it queries is in beta. The platform is in early access, which means the auto-healing pattern library is growing but not yet deep. Teams with strict data-residency requirements or air-gapped CI environments will hit the hosted-only ceiling immediately.