The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/tests/test_bootstrap_goldset.py utilizes exec() and compile() to dynamically execute Python code from a local asset template (assets/templates/rag/corpus_hash.py.template) during internal testing. This capability allows for the runtime execution of code strings.
[DYNAMIC_EXECUTION]: Automation scripts such as scripts/audit_checks.py and scripts/scaffold_eval.py execute shell commands via subprocess.run(). These are primarily used for Git operations (e.g., git status, git init) to track and manage evaluation source files. While the arguments are passed in a structured format, shell command execution remains a sensitive capability.
[INDIRECT_PROMPT_INJECTION]: The skill ingests documentation and metadata (e.g., SKILL.md, README.md) from target repositories to generate eval profiles. This creates a surface for indirect prompt injection if the target content contains malicious instructions.
Ingestion points: Reads files from the target directory including SKILL.md, README.md, and package manifests.
Boundary markers: The assessment logic does not explicitly document the use of delimiters or sanitization for interpolated target data.
Capability inventory: The skill has file-write and system-command (Git) capabilities.
Sanitization: No specific sanitization or filtering of the ingested external content is described.
[DATA_EXPOSURE_EXFILTRATION]: Evaluation harness templates (assets/templates/deepeval/.env.example) point to a remote vendor-owned proxy service at litellm-ai.accelint.dev. While associated with the skill author, this configuration involves transmitting data to an external service for LLM judgment.

accelint-eval-architect