The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/write_outputs.py implements dynamic module loading based on computed file paths.
The function load_shared_module resolves the location of a script named write_explore_bundle.py by traversing parent directories (parents[3] and parents[2]).
It uses importlib.util to load and execute the script found at the calculated path.
This mechanism executes code from locations outside the skill's immediate directory, which can be risky if the environment's directory structure is not strictly controlled.
[INDIRECT_PROMPT_INJECTION]: The skill's planning script ingests untrusted data from multiple sources.
Ingestion points: scripts/plan_code_changes.py reads the contents of the target repository's source code and processes external JSON files provided via command-line arguments (--idea-card-json, --analysis-json, --variant-spec-json).
Boundary markers: There are no visible delimiters or guardrail instructions used when processing these external files to prevent embedded instructions from influencing the agent.
Capability inventory: The skill identifies specific code blocks for modification and handles the execution of output-writing scripts.
Sanitization: Data extracted from JSON payloads is interpolated into strings for rationale and summary reporting without escaping or validation against malicious instruction patterns.

explore-code