OpenCode App sends periodic background requests to grok-4.3 via custom OpenAI-compatible provider (no such model in config)

OpenCode Desktop 1.18.31 (macOS 15.8) sends a background request with model=grok-4.3 roughly every 15-35 minutes to my custom OpenAI-compatible provider (baseURL pointing to a self-hosted gateway, provider package @ai-sdk/openai-compatible). The requests have an empty body (0 input / 0 output tokens), reasoning_effort=medium, no user session involved.

Facts:

• opencode.jsonc defines only grok-4.6 for this provider block; the string grok-4.3 appears nowhere in my config, app logs (main.log/renderer.log), or active sessions (no new sessions for weeks).
• Root model is a Gemini model, small_model is set to a Gemini model - title/summary paths should not touch Grok.
• Confirmed at network level (Little Snitch traffic log): the only process opening connections to the gateway is OpenCode Helper; requests go over a long-lived keep-alive connection, cadence is clock-precise (same second each tick).
• Restarting OpenCode App does not stop it - ticks resume with the new process, so it looks like a built-in background mechanism (healthcheck/warmup?) using a hardcoded/stale default model name (grok-4.3 was xAI's default in the past), not a stuck worker.
• On the gateway side each tick fails with 402 (empty xAI balance) - pure log noise, but it means the app is polling a model the user never configured.

Expected: background mechanism should use a configured model (e.g. small_model) or not fire against custom providers at all; it should never reference a model ID absent from config/catalog.

No source file or test is named. Start by tracing the background request mechanism and searching for the grok-4.3 model string or provider polling path, then reproduce the periodic request with a custom OpenAI-compatible provider. Done means no request targets a model absent from configuration, or the mechanism skips custom providers.