AI Gateway traffic splitting returns 400 when multiple model versions configured

[zjacobson1016]

Summary

Configuring AI Gateway traffic splitting between multiple model versions of the same family (e.g. databricks-claude-opus-4-7 and databricks-claude-opus-4-6) causes requests from ucode to fail with HTTP 400.

Expected behavior

When AI Gateway traffic splitting is configured across model versions, ucode-launched coding agents should route requests successfully according to the gateway endpoint configuration.

Actual behavior

• AI Gateway endpoint has traffic splitting enabled between model versions (e.g. opus-4-7 and opus-4-6)
• Launching an agent via ucode (e.g. ucode claude) results in a 400 error
• Agent fails to connect or complete requests through the gateway

Example scenario

• Workspace AI Gateway configured with traffic split between:
  • databricks-claude-opus-4-7
  • databricks-claude-opus-4-6
• ucode claude fails with 400

Possible causes

• ucode may be sending a specific model ID in headers/env that conflicts with traffic-split endpoint routing
• Claude Code default model env vars (ANTHROPIC_DEFAULT_OPUS_MODEL, etc.) may not align with split endpoint expectations
• Gateway may require a different model identifier when splitting is enabled

Suggested investigation

• Capture the exact 400 response body from AI Gateway
• Compare request headers/model IDs sent by ucode vs. a working direct gateway call
• Document supported traffic-split configurations for coding agent integration

Environment

• ucode (latest)
• Databricks AI Gateway v2 with traffic splitting enabled
• Claude (opus family); may affect other agents as well

[rohita5l]

We dont support user created AI Gateway endpoints yet. Working on adding support for it