Built a stopgap code_evolver.py that extracts Python functions, LLM-mutates them, validates (syntax/import/smoke test), and commits if gates pass. Works but hits the same gaps identified in PLAN.md Phase 4: no integration with batch_runner, no real fitness metric, no PR workflow, no test-suite integration.
The building blocks are there (Darwinian Evolver engine, constraint gates in evolution/core/) but the glue for project-level code evolution isnt. Would love to see Phase 4 bumped up — even a minimal version (target function → mutate → pytest → branch) would let us delete custom glue scripts.
Specific use case: evolving core domain logic (arbitrage detection, EV calculations, Monte Carlo sims) where an LLM can find real logic holes, not just add docstrings. The kind of code where a wrong edge case costs real money.
Happy to contribute patterns from the glue script if that helps de-risk the implementation.
Built a stopgap
code_evolver.pythat extracts Python functions, LLM-mutates them, validates (syntax/import/smoke test), and commits if gates pass. Works but hits the same gaps identified in PLAN.md Phase 4: no integration withbatch_runner, no real fitness metric, no PR workflow, no test-suite integration.The building blocks are there (Darwinian Evolver engine, constraint gates in
evolution/core/) but the glue for project-level code evolution isnt. Would love to see Phase 4 bumped up — even a minimal version (target function → mutate → pytest → branch) would let us delete custom glue scripts.Specific use case: evolving core domain logic (arbitrage detection, EV calculations, Monte Carlo sims) where an LLM can find real logic holes, not just add docstrings. The kind of code where a wrong edge case costs real money.
Happy to contribute patterns from the glue script if that helps de-risk the implementation.