dsh-maxcompute

Working with MaxCompute (ODPS) inside DSH used to mean hand-typing CLI commands, writing throwaway scripts, or hard-coding table names into your prompts. dsh-maxcompute wraps metadata browsing, SQL execution, async job management, and result export into tools a model can call directly, so you can drive the full pipeline from list_tables through run_sql to job_result in plain language.

Safety and cost control sit at the core of the design. Every SQL statement is dry-run first; if the estimated scan crosses the configured GB threshold the call is refused outright. A monotonic read-only guard denies DDL/DML by default and cannot be re-opened by later listeners. Credentials never touch the plugin layer — they stay inside maxc's own config. Scan-size caps, timeouts, preview row limits, and the readonly flag are all overridable per patch layer, making it easy to tighten permissions in shared environments.

It is aimed at data engineers and analysts who query MaxCompute regularly from within a DSH workflow — anyone who wants an AI assistant to safely inspect schemas, run analytical SQL, and pull async job results without bouncing back and forth to a terminal. The plugin follows a swappable-provider architecture: if you have an internal gateway, you replace a single service module and the tools layer keeps working unchanged.