fix(ai): send Gemini tool schemas as parametersJsonSchema - #51009

Problem

The Gemini protocol (Google AI Studio, Vertex Gemini, and Gemini on the Zen gateway) sent tool schemas in the legacy parameters field. That field only accepts Google's OpenAPI-subset schema, so GeminiToolSchema.convert projected every tool schema into it. The projection was lossy and still produced schemas Gemini rejects:

• Rejected requests. Vertex, and Zen (which reaches Gemini through Vertex), require a type on every node, so any const, untyped enum, or untyped field fails with "didn't specify the schema type field". The API also rejects properties on non-object types. Tuples, prefixItems, and items: true fail everywhere. Vercel's MCP toolset (169 tools, sent in 5 chunks) failed in 4 of 5 chunks on Vertex and Zen, and 3 of 5 on the API.
• Wrong arguments when accepted. $ref/$defs, additionalProperties, and oneOf detail are dropped and enum values are stringified, so Gemini fills arguments with the wrong types. In forced tool calls: integer enums came back as "3", referenced objects and oneOf unions came back as JSON strings, and additionalProperties maps came back empty (0/3 correct in each scenario).

Change

• gemini.ts: tool declarations use parametersJsonSchema, which accepts standard JSON Schema.
• New utils/gemini-json-schema.ts: rewrites only the shapes parametersJsonSchema still rejects, each verified live:
  • required entries with no declared property are removed.
  • Draft-04 boolean exclusiveMinimum/exclusiveMaximum become the numeric form (seen in a real Vercel tool).
  • Draft-07 tuple items: [...] (+ additionalItems) become prefixItems (+ items).
  • Gemini only accepts a recursive $ref loop that passes through an optional property or potentially empty array items. Other self-references become {}. The main real-world case is Zod's JSON-value schema, whose loop goes through additionalProperties. Wrapping the ref as nullable does not satisfy Gemini (tested).
  • Data under const, default, enum, examples, and dependentRequired is never rewritten.
• GeminiToolSchema stays for the separate toolSchema: "gemini" compatibility mode on OpenAI-compatible routes, which this PR does not touch.

Verification

Live through the V2 client on Gemini API, Vertex, and Zen (gemini-3.8-flash). The corpus has 45 schemas, including a regression case for every past Gemini schema fix, plus Vercel's full MCP toolset:

The only remaining rejections are synthetic schemas nested 24+ levels deep. parametersJsonSchema allows about 15 nested properties levels, versus about 31 for parameters. No real schema tested comes close; Vercel's deepest is 13.

Known limitation, unchanged by this PR: Gemini cannot reliably fill a free-form "any JSON value" parameter with either field.

Tests: gemini.test.ts covers the normalizer rules and sending standard JSON Schema unchanged. The three cassettes whose request bodies contain tool declarations (google-vertex/calls-a-tool, google-vertex/continues-after-a-tool-result, pdf/gemini-tool-result) were re-recorded. bun test in packages/ai: 1505 pass. bun run check: 35/35.