.Codex

Keep repository documentation accurate and agent-legible through drift detection and deterministic fixes. Use when asked to scan docs for stale content, broken links, metadata/nav drift, API coverage gaps, or to run recurring documentation gardening in on-demand or scheduled maintenance loops.

Test code built on AI SDK v7 using its real `ai/test` utilities and V4 provider protocol. Use when writing, fixing, reviewing, or migrating tests around AI SDK agents, generation, streaming, structured output, tools, provider registries, model modalities, or simulated streams. Trigger for requests to test an agent, mock a model, fake a stream, simulate a tool call, test retry logic, inspect prompts, or debug hanging streams and empty output. Remove V3 fixtures and custom helpers that duplicate public AI SDK APIs.

Update the Text2SQL system prompt in a generalizable, non-overfit way using failure-pattern driven evaluation loops. Use when SQL generation quality regresses or specific failure classes repeat.