vector-forge

Mutation-driven test vector generation for cryptographic algorithms and protocols. The skill uses mutation testing to systematically identify gaps in test vector coverage, then generates new test vectors that deliberately exercise the uncovered code paths, and measures effectiveness by comparing mutation kill rates before and after. It is intended for generating cryptographic test vectors, measuring Wycheproof-style coverage gaps, finding escaped mutants, and building cross-implementation test suites; it is not for cases where no implementation exists yet, a single trivial implementation with no edge cases, application logic rather than algorithm implementations, or an algorithm with no public test vectors to compare against.

The workflow runs in six phases. Discovery finds implementations of the target algorithm, distinguishing pure implementations in high-level languages (the best mutation targets) from FFI wrapper crates (low value at the binding layer) and reference implementations. Harness writes or adapts a test vector harness per implementation that reads vectors from JSON (Wycheproof format recommended), exercises the API, and asserts both acceptance and rejection plus roundtrip assertions; the harness must live inside the implementation's package so the mutation framework can see it. Baseline runs mutation testing with existing vectors only and records metrics such as total mutants, killed, survived, not covered, timed out, efficacy, and coverage. Escape Analysis classifies escaped mutants by code path using the Trailmark call graph and the genotoxic skill's triage methodology. Vector Gen creates vectors targeting the escapes, and Validation re-runs mutation testing to compare before and after.

Prerequisites include trailmark (installed via `uv pip install trailmark` if `uv run trailmark` fails), at least one implementation in a language with mutation testing support, a test harness, and a mutation testing framework. Framework choices include gremlins for Go, cargo-mutants for Rust, mutmut for Python, and Mull for C/C++, with parallel execution recommended for large codebases. A central principle is that FFI bindings must not be mutated at the wrapper layer; the underlying implementation language must be mutated instead, and a harness that only checks valid vectors will miss permissive mutations such as turning an AND into an OR in validation.