The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill utilizes Python's eval() and compile() functions within scripts/ybm/probes.py to execute logic checks defined in the local rules/rules.json file. These checks are used to verify database behavior on specific releases. The implementation includes a security filter using the ast module that restricts execution to a whitelist of safe nodes (arithmetic, comparisons, and attribute access) and blocks access to built-in functions.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a potential attack surface as it ingests untrusted data from user-provided schema dumps and query lists.
Ingestion points: scripts/ybm/inputs.py reads content from schema.sql, queries.sql, and various CSV statistics files.
Boundary markers: The skill processes these inputs through a deterministic Python engine rather than passing them directly as instructions to the LLM.
Capability inventory: The skill can execute shell commands, manage Docker containers, and write files to the local filesystem.
Sanitization: Input data is tokenized and parsed by internal SQL readers to identify structure and identifiers before analysis, reducing the risk of malicious instructions being obeyed.
[EXTERNAL_DOWNLOADS]: The scripts/extract-version-data.py script fetches release metadata and source files from the project's official GitHub repository (raw.githubusercontent.com/yugabyte/yugabyte-db/). This is used to build a local cache of version-specific behavioral data for the analysis engine.
[COMMAND_EXECUTION]: The tool performs legitimate command execution to support its diagnostic functions. It invokes docker or podman to run temporary database instances for plan replay, uses ysqlsh for database interaction, and may use git to query local repository checkouts for version data.

yb-model