AI Gateway chat-completions forwards OpenAI-format tool-schema extras ($schema, exclusiveMinimum, stream_options) into Gemini's native API — 400s ALL Gemini models for OpenCode agents

[bradaffleck-75]

Summary

The AI Gateway's Gemini translation layer on the standard serving-endpoints chat-completions surface (POST {host}/serving-endpoints/<gemini-model>/invocations, OpenAI wire format) forwards unknown OpenAI-format fields verbatim into Gemini's native API, which rejects them with a hard 400:

Invalid JSON payload received. Unknown name "$schema" at 'tools[0].function_declarations[N].parameters': Cannot find field.
Invalid JSON payload received. Unknown name "exclusiveMinimum" at 'tools[0].function_declarations[N].parameters.properties[1].value': Cannot find field.
Invalid JSON payload received. Unknown name "stream_options" at 'generation_config': Cannot find field.

Any OpenAI-compatible client whose tool JSON-Schemas carry standard JSON-Schema 2020-12 keywords, or that requests usage via stream_options, gets rejected. This makes every Gemini model unusable for tool-carrying agents on this surface. OpenCode (1.18.x) emits all three on every turn, so OpenCode + Gemini is 100% broken.

Verified identical on databricks-gemini-3-7-flash and databricks-gemini-3-8-flash — this is the translation layer, not a model regression.

Environment

• Workspace: dbc-0927d274-c1a8.cloud.databricks.com (ws 2502993781179439)
• Models: databricks-gemini-3-7-flash, databricks-gemini-3-8-flash (AI Gateway foundation models)
• Date: 2026-09-03

Reproduction (3 minimal requests, same token, same model)

1. Clean tools → 200 (the surface itself works fine):

POST /serving-endpoints/databricks-gemini-3-8-flash/invocations
{"messages":[{"role":"user","content":"Say OK"}],
 "tools":[{"type":"function","function":{"name":"read","description":"Read a file",
   "parameters":{"type":"object","properties":{"path":{"type":"string"}},"required":["path"]}}}]}

→ HTTP 200

2. Same request, parameters gain two bog-standard JSON-Schema fields → 400:

"parameters": {
  "$schema": "https://json-schema.org/draft/2020-12/schema",
  "type": "object",
  "properties": {"path": {"type": "string"},
                 "limit": {"type": "number", "exclusiveMinimum": 0}},
  "required": ["path"]
}

→ HTTP 400 — Unknown name "$schema" at 'tools[0].function_declarations[0].parameters' (plus the exclusiveMinimum violation)

3. No tools at all, just stream_options → 400:

{"messages":[{"role":"user","content":"Say OK"}], "stream_options": {"include_usage": true}}

→ HTTP 400 — Unknown name "stream_options" at 'generation_config'

So each of $schema, exclusiveMinimum, and stream_options independently 400s, while identical payloads without them succeed — including image_url vision content (Gemini vision itself is fine on this surface).

Production evidence

A real OpenCode-native session turn failed with the full 40-tool payload (excerpts above). The response carried:

• request_id: fbf02b36-8824-49df-848f-07b18751a2e6
• x-databricks-response-source: CallerError
• "via": "1.1 google" — the request is translated and forwarded to Google's API, which rejects the unknown keys

Expected behavior

The Gemini translation layer should sanitize OpenAI-compat payloads before mapping to Gemini's native schema, the way mature OpenAI-compatibility layers do:

• strip $schema (and other meta-keywords) from tool parameter schemas
• map or strip validation keywords Gemini's Schema proto doesn't model (exclusiveMinimum/exclusiveMaximum, and similar) rather than rejecting the whole request
• map or drop stream_options (e.g., honor usage reporting natively, or ignore it) instead of failing the request

Workaround status / impact

• No other gateway surface serves Gemini: the Responses surface (/ai-gateway/openai/v1/responses) returns "Responses API passthrough is not supported for model databricks-gemini-3-8-flash." — OpenAI-family only. The Anthropic surface doesn't serve Gemini either. So there is no harness path to Gemini with tools today (OpenCode → this surface; openai-agents/codex → Responses; pi → same chat-completions surface).
• We moved our vision-review tier from Gemini Flash to Claude Sonnet 5 to keep it alive; we'd move it back the moment the translation sanitizes these fields.