forgeproof или forge

ForgeProof turns GitHub issues into working, AI-generated code backed by a tamper-evident receipt. Claude Code reads the issue, extracts requirements, writes code and tests, then seals everything into a signed .rpack provenance bundle. The bundle proves what was done, why, and that nothing was altered after signing — a cryptographic receipt that travels with the code.

Description: Adversarial code hardening for Claude Code. Forge spawns two agents — a Builder and a Breaker — that iterate against each other until the code is unbreakable. The Builder implements a feature, the Breaker writes adversarial tests to prove bugs, and the Builder must fix the implementation without deleting or weakening the tests. The loop continues until the Breaker can't produce a failing test. The result is battle-tested code with a permanent adversarial test suite. Key features: - Evidence-based, not opinion-based. Every issue the Breaker raises must have a failing test that proves the bug. No test, no break. This eliminates false positives and subjective feedback. - Iterative adversarial loop. Builder and Breaker are separate agents that go back and forth for up to N rounds. The Builder cannot dismiss feedback; it must make the failing tests pass. - Adversarial tests are permanent. The Breaker's tests stay in the repo under tests/adversarial/ as a regression suite. Future changes are protected