ai-core/middleware

You add observability to TanStack AI chat hooks by applying middleware arrays that inspect ChatMiddlewareContext to automatically log request IDs, track usage, and report errors across init, streaming, and terminal phases.

Yes, onBeforeToolCall and onAfterToolCall middleware hooks intercept tool invocations to enforce guards, transform arguments, and apply tool-caching decisions before the conversation stream continues.

Use ctx.defer within your middleware implementations to execute resilient analytics logging and deferred side effects asynchronously. This ensures tracking pipelines run without disrupting or aborting the main chat stream.

Yes, middleware implementations can return short-circuit decisions from lifecycle hooks to guard tools and control execution flow, maintaining strict observability and behavioral enforcement throughout the chat invocation.

The best way is using middleware entries that provide explicit error handling and lifecycle hook coverage at every chat phase. This surfaces errors and emits telemetry without requiring manual instrumentation woven into each run.