darwinian-evolver

Define an Organism holding your prompt template, an Evaluator that scores outputs against expected results, and an LLM-based Mutator that proposes fixes from failure cases. Then run the EvolveProblemLoop for several iterations and inspect the ranked population snapshots.

The upstream CLI is hardcoded to Anthropic, but the skill ships a parrot_openrouter.py driver that routes LLM calls through OpenRouter using the OpenAI SDK. Set OPENROUTER_API_KEY and EVOLVER_MODEL to use any provider model.

Yes, the Organism can hold any artifact string such as a regex_pattern or sql_query, as long as you implement a run method and an Evaluator that scores its behavior. The custom problem template shows the three required pieces: Organism, Evaluator, and Mutator.

Avoid it when the optimization target is differentiable (use gradient descent or DSPy instead), when you only need two or three hand-written variants, or when the fitness signal is purely subjective with no measurable criterion.

Azure-backed OpenRouter models reject prompts containing phrases like 'ignore previous instructions' due to content filters. Wrap the LLM call in try/except and return an error marker string so the evolver scores that organism zero and continues.

Snapshots are nested pickles containing a dict with a population_snapshot key, and the Organism class must be importable under its original dotted path. Use the shipped show_snapshot.py script to rank organisms by score and print their artifact fields.