The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/evaluate_assertions.py uses subprocess.run with shell=True to execute arbitrary shell commands defined in test assertions. This is a primary mechanism for the skill's 'oracle' to verify artifact quality.
[REMOTE_CODE_EXECUTION]: The core logic of the skill involves an autonomous loop that modifies source code and subsequently executes it via the shell (as documented in references/artifact-guide.md and SKILL.md). This allows for the execution of dynamically generated code.
[DYNAMIC_EXECUTION]: The skill is designed to perform runtime modification of scripts and configuration files, followed by immediate execution to capture metrics. While scripts/structural_check.py attempts to mitigate risk by scanning for dangerous patterns (like rm -rf or eval), the execution of modified code is the skill's primary function.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted artifacts provided by the user or external files. These artifacts could contain instructions designed to manipulate the 'evolution controller' behavior during the mutation or evaluation phases.
Ingestion points: The skill reads the 'artifact' from the local filesystem to start the evolution loop.
Boundary markers: The system uses an evolve_plan.md to define 'Editable scope' and 'Forbidden scope', aiming to protect the benchmark harness from modification.
Capability inventory: The skill utilizes subprocess.run for code execution, git for persistence/reverts, and standard file I/O for mutations.
Sanitization: The skill executes scripts/structural_check.py as an 'L1 Quick Gate' to detect secrets and common malicious shell patterns before higher-level evaluation.

self-evolution