The Agent Skills Directory

[DYNAMIC_EXECUTION]: The scripts/show_snapshot.py script utilizes pickle.loads() to deserialize population data. Because Python's pickle module can execute arbitrary code during deserialization, this poses a risk if a malicious snapshot file is provided. The script includes a protective --i-trust-this-file flag and a warning message to prevent accidental execution of untrusted data.
[EXTERNAL_DOWNLOADS]: The skill's setup process involves cloning a remote Git repository (https://github.com/imbue-ai/darwinian_evolver.git) and installing its environment via uv sync. While the source is a known research entity, the process involves downloading and running external code.
[INDIRECT_PROMPT_INJECTION]: The core functionality of the skill involves an evolutionary loop that passes previously generated LLM artifacts (prompts, code, etc.) back into the LLM context to generate improvements. This creates a surface for indirect prompt injection.
Ingestion points: Artifacts and failure cases are ingested in the mutate methods of scripts/parrot_openrouter.py and templates/custom_problem_template.py.
Boundary markers: No delimiters or isolation instructions are present in the mutator prompts to distinguish between instructions and evolved content.
Capability inventory: The skill possesses the ability to write files, perform network requests to LLM APIs, and execute shell commands via uv run.
Sanitization: Evolved organisms are interpolated into prompts as raw strings without escaping or filtering.
[COMMAND_EXECUTION]: The skill uses uv run and shell commands to drive the underlying evolution tool, manage iterations, and output results to the filesystem.

darwinian-evolver