Your agent can now debug Workers with local tracing

The change adds built-in OpenTelemetry instrumentation to the workerd runtime used by Miniflare, so every fetch, binding, and handler call generates a span during a local wrangler dev or vite dev session. These spans are collected into a SQLite-backed store and served through a local REST endpoint that also provides correlated console logs. The endpoint advertises an OpenAPI schema, allowing tooling to discover its capabilities at runtime.

For AI-driven coding agents, the workflow simplifies dramatically: after reproducing a failing request, the agent can POST a query to the Local Explorer API and receive a full trace that pinpoints the failing operation, such as a missing database column. The agent can then use the same API to inspect bindings or apply a migration, re-run the request, and verify the fix, all without redeploying or sprinkling manual log statements.

Adopting the feature requires only updating the development tooling (wrangler or the Cloudflare Vite plugin) to the latest version; no changes to application code or configuration are needed. The local server must be started in a mode that recognises a supported coding-agent session, which it signals with a console hint. Apart from the update, the cost is limited to the usual resources consumed by the local dev server.

The observability capabilities are confined to the local development environment. They are unavailable for Workers running in the Cloudflare edge, and the API is read-only, so agents cannot modify trace data or affect production behaviour. If a session is not identified as an AI agent, the automatic hint and API exposure are not triggered.

Practically, engineers gain a faster, more deterministic debugging loop for integration points like KV, D1, R2, and Queues, reducing the need for ad-hoc console logs. However, they must still perform production-level testing to capture real-world latency, scaling effects, and edge-specific failures that the local tracer does not emulate.