feat(providers): Gemini tool-calling end-to-end on /v1beta (#6222) (#6394)

Gemini tool-calling end-to-end on /v1beta (#6222) (net +1/-0, tests OK). Integrated into release/v3.8.46.

CHANGELOG.md

@@ -8,6 +8,7 @@
 
 ### ✨ New Features
 
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
 - **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
 - **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
 - **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)

open-sse/translator/response/openai-to-gemini-sse.ts

@@ -37,10 +37,18 @@ export const OPENAI_TO_GEMINI_FINISH_REASON: Record<string, string> = {
   content_filter: "SAFETY",
 };
 
+interface OpenAIToolCallDelta {
+  index?: number;
+  id?: string;
+  type?: string;
+  function?: { name?: string; arguments?: string };
+}
+
 interface OpenAIChoiceDelta {
   content?: string | null;
   reasoning_content?: string | null;
   role?: string;
+  tool_calls?: OpenAIToolCallDelta[];
 }
 
 interface OpenAIChoice {
@@ -63,9 +71,31 @@ interface OpenAIStreamChunk {
   model?: string;
 }
 
+interface GeminiFunctionCall {
+  name: string;
+  args: Record<string, unknown>;
+}
+
+interface GeminiFunctionResponse {
+  name: string;
+  response: Record<string, unknown>;
+}
+
 interface GeminiPart {
-  text: string;
+  text?: string;
   thought?: boolean;
+  functionCall?: GeminiFunctionCall;
+  functionResponse?: GeminiFunctionResponse;
+}
+
+/**
+ * Per-stream mutable accumulator for OpenAI streamed tool calls. OpenAI emits
+ * a tool call's `arguments` as partial JSON fragments across several delta
+ * chunks, keyed by `index`; we buffer them here and flush a complete
+ * `functionCall` part once `finish_reason` arrives.
+ */
+export interface GeminiToolCallState {
+  toolCallAccum?: Record<number, { id: string; name: string; arguments: string }>;
 }
 
 interface GeminiCandidate {
@@ -97,7 +127,8 @@ interface GeminiStreamChunk {
  */
 export function openAIChunkToGeminiChunk(
   parsed: OpenAIStreamChunk,
-  fallbackModel: string
+  fallbackModel: string,
+  state?: GeminiToolCallState
 ): GeminiStreamChunk | null {
   const choice = parsed.choices?.[0];
   if (!choice) return null;
@@ -112,6 +143,35 @@ export function openAIChunkToGeminiChunk(
     parts.push({ text: String(delta.content) });
   }
 
+  // Accumulate streamed tool-call fragments (OpenAI streams partial JSON
+  // `arguments` across chunks, keyed by index). Requires a caller-supplied
+  // per-stream `state`; nothing is emitted until finish_reason arrives.
+  if (state && Array.isArray(delta.tool_calls)) {
+    const accum = (state.toolCallAccum ??= {});
+    for (const tc of delta.tool_calls) {
+      const idx = tc.index ?? 0;
+      const entry = (accum[idx] ??= { id: "", name: "", arguments: "" });
+      if (tc.id) entry.id = tc.id;
+      if (tc.function?.name) entry.name += tc.function.name;
+      if (tc.function?.arguments) entry.arguments += tc.function.arguments;
+    }
+  }
+
+  // On finish, flush accumulated tool calls as complete functionCall parts.
+  if (choice.finish_reason && state?.toolCallAccum) {
+    for (const key of Object.keys(state.toolCallAccum)) {
+      const entry = state.toolCallAccum[Number(key)];
+      if (!entry.name) continue;
+      let args: Record<string, unknown> = {};
+      try {
+        args = JSON.parse(entry.arguments || "{}") as Record<string, unknown>;
+      } catch {
+        args = {};
+      }
+      parts.push({ functionCall: { name: entry.name, args } });
+    }
+  }
+
   // Skip pure role-only deltas with no content and no finish signal.
   if (parts.length === 0 && !choice.finish_reason) return null;
 
@@ -168,6 +228,9 @@ export function transformOpenAISSEToGeminiSSE(upstreamResponse: Response, model:
   // chunk so we never JSON.parse a half-event.
   let buffer = "";
 
+  // Per-stream tool-call accumulator (shared across transform + flush).
+  const toolCallState: GeminiToolCallState = {};
+
   const transform = new TransformStream<Uint8Array, Uint8Array>({
     transform(chunk, controller) {
       buffer += decoder.decode(chunk, { stream: true });
@@ -192,7 +255,7 @@ export function transformOpenAISSEToGeminiSSE(upstreamResponse: Response, model:
           continue;
         }
 
-        const geminiChunk = openAIChunkToGeminiChunk(parsed, model);
+        const geminiChunk = openAIChunkToGeminiChunk(parsed, model, toolCallState);
         if (!geminiChunk) continue;
 
         controller.enqueue(encoder.encode("data: " + JSON.stringify(geminiChunk) + "\r\n\r\n"));
@@ -212,7 +275,7 @@ export function transformOpenAISSEToGeminiSSE(upstreamResponse: Response, model:
       } catch {
         return;
       }
-      const geminiChunk = openAIChunkToGeminiChunk(parsed, model);
+      const geminiChunk = openAIChunkToGeminiChunk(parsed, model, toolCallState);
       if (!geminiChunk) return;
       controller.enqueue(encoder.encode("data: " + JSON.stringify(geminiChunk) + "\r\n\r\n"));
     },
@@ -228,10 +291,17 @@ export function transformOpenAISSEToGeminiSSE(upstreamResponse: Response, model:
   });
 }
 
+interface OpenAIToolCall {
+  id?: string;
+  type?: string;
+  function?: { name?: string; arguments?: string };
+}
+
 interface OpenAIMessage {
   content?: string | null;
   reasoning_content?: string | null;
   role?: string;
+  tool_calls?: OpenAIToolCall[];
 }
 
 interface OpenAINonStreamChoice {
@@ -311,7 +381,22 @@ export async function convertOpenAIResponseToGemini(
   if (message.reasoning_content) {
     parts.push({ text: String(message.reasoning_content), thought: true });
   }
-  parts.push({ text: String(message.content ?? "") });
+  // Text content — only emit a text part when there is actual text, so a
+  // pure tool-call response isn't padded with an empty-string part.
+  const toolCalls = Array.isArray(message.tool_calls) ? message.tool_calls : [];
+  if (message.content || toolCalls.length === 0) {
+    parts.push({ text: String(message.content ?? "") });
+  }
+  // Tool calls → functionCall parts (arguments JSON string → object).
+  for (const tc of toolCalls) {
+    let args: Record<string, unknown> = {};
+    try {
+      args = JSON.parse(tc.function?.arguments || "{}") as Record<string, unknown>;
+    } catch {
+      args = {};
+    }
+    parts.push({ functionCall: { name: tc.function?.name || "", args } });
+  }
 
   const finishReason = OPENAI_TO_GEMINI_FINISH_REASON[finish_reason ?? "stop"] ?? "STOP";
 

src/app/api/v1beta/models/[...path]/convertGeminiToInternal.ts

@@ -0,0 +1,211 @@
+/**
+ * Convert a native Gemini `generateContent` request body into the internal
+ * OpenAI Chat Completions shape consumed by `handleChat`.
+ *
+ * Extracted from the route handler so the (pure) conversion can be unit-tested
+ * without importing the full chat-handler graph (which keeps timers alive and
+ * hangs the node:test runner). See feature #6222.
+ *
+ * Tool/function calling is preserved in the request direction:
+ *   - `tools[].functionDeclarations` → OpenAI `tools[{type:"function",...}]`
+ *   - prior `functionCall` parts       → assistant `tool_calls`
+ *   - `functionResponse` parts         → `tool`-role messages
+ *
+ * Mirrors the shapes already used by the request translator
+ * `open-sse/translator/request/gemini-to-openai.ts`.
+ */
+
+interface GeminiFunctionCall {
+  name?: string;
+  args?: Record<string, unknown>;
+  id?: string;
+}
+
+interface GeminiFunctionResponse {
+  name?: string;
+  id?: string;
+  response?: { result?: unknown } & Record<string, unknown>;
+}
+
+interface GeminiPart {
+  text?: string;
+  functionCall?: GeminiFunctionCall;
+  functionResponse?: GeminiFunctionResponse;
+  [key: string]: unknown;
+}
+
+interface GeminiContent {
+  role?: string;
+  parts?: GeminiPart[];
+}
+
+interface GeminiTool {
+  functionDeclarations?: Array<{
+    name?: string;
+    description?: string;
+    parameters?: unknown;
+  }>;
+}
+
+interface GeminiGenerateBody {
+  systemInstruction?: { parts?: GeminiPart[] };
+  contents?: GeminiContent[];
+  tools?: GeminiTool[];
+  generationConfig?: {
+    maxOutputTokens?: number;
+    temperature?: number;
+    topP?: number;
+  };
+}
+
+interface InternalMessage {
+  role: string;
+  content?: string | null;
+  tool_calls?: Array<{
+    id: string;
+    type: "function";
+    function: { name: string; arguments: string };
+  }>;
+  tool_call_id?: string;
+}
+
+interface InternalTool {
+  type: "function";
+  function: { name: string; description: string; parameters: unknown };
+}
+
+export interface InternalChatBody {
+  model: string;
+  messages: InternalMessage[];
+  stream: boolean;
+  max_tokens?: number;
+  temperature?: number;
+  top_p?: number;
+  tools?: InternalTool[];
+}
+
+let toolCallSeq = 0;
+
+function newToolCallId(): string {
+  toolCallSeq += 1;
+  return `call_${Date.now()}_${toolCallSeq}_${Math.random().toString(36).slice(2, 8)}`;
+}
+
+/**
+ * Convert a single Gemini `content` entry into one internal message.
+ *
+ * `functionResponse` parts become a `tool` message; `functionCall` parts become
+ * an assistant message carrying `tool_calls`; otherwise a plain text message.
+ * Returns `null` when the content has nothing to contribute.
+ */
+function convertContent(content: GeminiContent): InternalMessage | null {
+  const parts = content.parts;
+  if (!parts || !Array.isArray(parts)) return null;
+
+  // A functionResponse turn maps to a `tool` role message.
+  for (const part of parts) {
+    if (part.functionResponse) {
+      const fr = part.functionResponse;
+      const payload =
+        fr.response && "result" in fr.response ? fr.response.result : fr.response ?? {};
+      return {
+        role: "tool",
+        tool_call_id: fr.id || fr.name || "",
+        content: JSON.stringify(payload ?? {}),
+      };
+    }
+  }
+
+  const textSegments: string[] = [];
+  const toolCalls: InternalMessage["tool_calls"] = [];
+
+  for (const part of parts) {
+    if (typeof part.text === "string") {
+      textSegments.push(part.text);
+    }
+    if (part.functionCall) {
+      toolCalls.push({
+        id: part.functionCall.id || newToolCallId(),
+        type: "function",
+        function: {
+          name: part.functionCall.name || "",
+          arguments: JSON.stringify(part.functionCall.args || {}),
+        },
+      });
+    }
+  }
+
+  const text = textSegments.join("\n");
+
+  if (toolCalls.length > 0) {
+    const msg: InternalMessage = { role: "assistant" };
+    if (text) msg.content = text;
+    msg.tool_calls = toolCalls;
+    return msg;
+  }
+
+  const role = content.role === "model" ? "assistant" : "user";
+  return { role, content: text };
+}
+
+/**
+ * Convert Gemini request format to OpenAI/internal format.
+ *
+ * @param geminiBody parsed Gemini request body
+ * @param model      resolved model string (e.g. "gemini/gemini-pro")
+ * @param stream     whether to stream (derived from URL action suffix)
+ */
+export function convertGeminiToInternal(
+  geminiBody: GeminiGenerateBody,
+  model: string,
+  stream: boolean
+): InternalChatBody {
+  const messages: InternalMessage[] = [];
+
+  // Convert system instruction
+  if (geminiBody.systemInstruction) {
+    const systemText =
+      geminiBody.systemInstruction.parts?.map((p) => p.text ?? "").join("\n") || "";
+    if (systemText) {
+      messages.push({ role: "system", content: systemText });
+    }
+  }
+
+  // Convert contents to messages (text + tool calls + tool responses)
+  if (geminiBody.contents) {
+    for (const content of geminiBody.contents) {
+      const converted = convertContent(content);
+      if (converted) messages.push(converted);
+    }
+  }
+
+  const result: InternalChatBody = {
+    model,
+    messages,
+    stream,
+    max_tokens: geminiBody.generationConfig?.maxOutputTokens,
+    temperature: geminiBody.generationConfig?.temperature,
+    top_p: geminiBody.generationConfig?.topP,
+  };
+
+  // Convert tool declarations → OpenAI tools.
+  if (Array.isArray(geminiBody.tools)) {
+    const tools: InternalTool[] = [];
+    for (const tool of geminiBody.tools) {
+      if (!tool.functionDeclarations) continue;
+      for (const func of tool.functionDeclarations) {
+        tools.push({
+          type: "function",
+          function: {
+            name: func.name || "",
+            description: func.description || "",
+            parameters: func.parameters || { type: "object", properties: {} },
+          },
+        });
+      }
+    }
+    if (tools.length > 0) result.tools = tools;
+  }
+
+  return result;
+}

src/app/api/v1beta/models/[...path]/route.ts

@@ -7,6 +7,7 @@ import {
 import { sanitizeErrorMessage } from "@omniroute/open-sse/utils/error";
 import { v1betaGeminiGenerateSchema } from "@/shared/validation/schemas";
 import { isValidationFailure, validateBody } from "@/shared/validation/helpers";
+import { convertGeminiToInternal } from "./convertGeminiToInternal";
 
 let initialized = false;
 
@@ -132,40 +133,3 @@ export async function POST(request, { params }) {
     );
   }
 }
-
-/**
- * Convert Gemini request format to OpenAI/internal format.
- *
- * @param geminiBody parsed Gemini request body
- * @param model      resolved model string (e.g. "gemini/gemini-pro")
- * @param stream     whether to stream (derived from URL action suffix)
- */
-function convertGeminiToInternal(geminiBody, model, stream) {
-  const messages = [];
-
-  // Convert system instruction
-  if (geminiBody.systemInstruction) {
-    const systemText = geminiBody.systemInstruction.parts?.map((p) => p.text).join("\n") || "";
-    if (systemText) {
-      messages.push({ role: "system", content: systemText });
-    }
-  }
-
-  // Convert contents to messages
-  if (geminiBody.contents) {
-    for (const content of geminiBody.contents) {
-      const role = content.role === "model" ? "assistant" : "user";
-      const text = content.parts?.map((p) => p.text).join("\n") || "";
-      messages.push({ role, content: text });
-    }
-  }
-
-  return {
-    model,
-    messages,
-    stream,
-    max_tokens: geminiBody.generationConfig?.maxOutputTokens,
-    temperature: geminiBody.generationConfig?.temperature,
-    top_p: geminiBody.generationConfig?.topP,
-  };
-}

tests/unit/v1beta-gemini-tool-calling-6222.test.ts

@@ -0,0 +1,292 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+
+// Feature #6222 — Gemini tool/function calling end-to-end on /v1beta.
+// Covers the three converters that previously dropped tool calls:
+//   1. Request:  convertGeminiToInternal (sibling of route.ts)
+//   2. Non-stream response: convertOpenAIResponseToGemini
+//   3. Stream response: openAIChunkToGeminiChunk / transformOpenAISSEToGeminiSSE
+//
+// The request converter lives in its own module (not route.ts) so it can be
+// unit-tested without importing the chat-handler graph, which keeps timers
+// alive and hangs the node:test runner.
+
+const { convertGeminiToInternal } = await import(
+  "../../src/app/api/v1beta/models/[...path]/convertGeminiToInternal.ts"
+);
+const {
+  openAIChunkToGeminiChunk,
+  transformOpenAISSEToGeminiSSE,
+  convertOpenAIResponseToGemini,
+} = await import("../../open-sse/translator/response/openai-to-gemini-sse.ts");
+
+// ---------------------------------------------------------------------------
+// 1. Request converter
+// ---------------------------------------------------------------------------
+
+test("request: tools[].functionDeclarations → OpenAI tools", () => {
+  const geminiBody = {
+    contents: [{ role: "user", parts: [{ text: "What is the weather in Paris?" }] }],
+    tools: [
+      {
+        functionDeclarations: [
+          {
+            name: "get_weather",
+            description: "Get the current weather for a city",
+            parameters: {
+              type: "object",
+              properties: { city: { type: "string" } },
+              required: ["city"],
+            },
+          },
+        ],
+      },
+    ],
+  };
+
+  const out = convertGeminiToInternal(geminiBody, "gemini/gemini-pro", false);
+
+  assert.ok(Array.isArray(out.tools), "tools should be an array");
+  assert.equal(out.tools.length, 1);
+  assert.deepEqual(out.tools[0], {
+    type: "function",
+    function: {
+      name: "get_weather",
+      description: "Get the current weather for a city",
+      parameters: {
+        type: "object",
+        properties: { city: { type: "string" } },
+        required: ["city"],
+      },
+    },
+  });
+  // Existing text mapping preserved.
+  const userMsg = out.messages.find((m) => m.role === "user");
+  assert.ok(userMsg);
+  assert.equal(userMsg.content, "What is the weather in Paris?");
+});
+
+test("request: prior functionCall part → assistant tool_calls", () => {
+  const geminiBody = {
+    contents: [
+      { role: "user", parts: [{ text: "Weather in Paris?" }] },
+      {
+        role: "model",
+        parts: [{ functionCall: { name: "get_weather", args: { city: "Paris" } } }],
+      },
+    ],
+  };
+
+  const out = convertGeminiToInternal(geminiBody, "gemini/gemini-pro", false);
+
+  const assistantMsg = out.messages.find((m) => m.role === "assistant");
+  assert.ok(assistantMsg, "assistant message should exist");
+  assert.ok(Array.isArray(assistantMsg.tool_calls), "assistant should carry tool_calls");
+  assert.equal(assistantMsg.tool_calls.length, 1);
+  assert.equal(assistantMsg.tool_calls[0].type, "function");
+  assert.equal(assistantMsg.tool_calls[0].function.name, "get_weather");
+  assert.deepEqual(
+    JSON.parse(assistantMsg.tool_calls[0].function.arguments),
+    { city: "Paris" }
+  );
+});
+
+test("request: functionResponse part → tool role message", () => {
+  const geminiBody = {
+    contents: [
+      { role: "user", parts: [{ text: "Weather in Paris?" }] },
+      {
+        role: "model",
+        parts: [{ functionCall: { name: "get_weather", args: { city: "Paris" } } }],
+      },
+      {
+        role: "user",
+        parts: [
+          {
+            functionResponse: {
+              name: "get_weather",
+              response: { result: { tempC: 18 } },
+            },
+          },
+        ],
+      },
+    ],
+  };
+
+  const out = convertGeminiToInternal(geminiBody, "gemini/gemini-pro", false);
+
+  const toolMsg = out.messages.find((m) => m.role === "tool");
+  assert.ok(toolMsg, "tool message should exist");
+  assert.equal(toolMsg.tool_call_id, "get_weather");
+  assert.deepEqual(JSON.parse(toolMsg.content), { tempC: 18 });
+});
+
+// ---------------------------------------------------------------------------
+// 2. Non-stream response converter
+// ---------------------------------------------------------------------------
+
+function makeJsonResponse(obj: unknown): Response {
+  return new Response(JSON.stringify(obj), {
+    status: 200,
+    headers: { "Content-Type": "application/json" },
+  });
+}
+
+test("non-stream: message.tool_calls → parts[].functionCall {name,args}", async () => {
+  const openaiResponse = {
+    model: "gemini-pro",
+    choices: [
+      {
+        message: {
+          role: "assistant",
+          content: null,
+          tool_calls: [
+            {
+              id: "call_1",
+              type: "function",
+              function: {
+                name: "get_weather",
+                arguments: '{"city":"Paris"}',
+              },
+            },
+          ],
+        },
+        finish_reason: "tool_calls",
+      },
+    ],
+    usage: { prompt_tokens: 10, completion_tokens: 5, total_tokens: 15 },
+  };
+
+  const geminiResp = await convertOpenAIResponseToGemini(
+    makeJsonResponse(openaiResponse),
+    "gemini/gemini-pro"
+  );
+  const body = (await geminiResp.json()) as {
+    candidates: Array<{
+      content: { parts: Array<Record<string, unknown>> };
+      finishReason: string;
+    }>;
+  };
+
+  const parts = body.candidates[0].content.parts;
+  const fcPart = parts.find((p) => "functionCall" in p) as
+    | { functionCall: { name: string; args: Record<string, unknown> } }
+    | undefined;
+  assert.ok(fcPart, "should emit a functionCall part");
+  assert.equal(fcPart.functionCall.name, "get_weather");
+  // args must be parsed to an object, NOT left as a JSON string.
+  assert.deepEqual(fcPart.functionCall.args, { city: "Paris" });
+  assert.equal(body.candidates[0].finishReason, "STOP");
+});
+
+// ---------------------------------------------------------------------------
+// 3. Stream converter — fragmented tool_calls accumulate
+// ---------------------------------------------------------------------------
+
+test("stream (unit): fragmented tool_calls accumulate into one functionCall", () => {
+  const state = {} as Record<string, unknown>;
+
+  // Chunk 1: opens the tool call with name + partial args.
+  const c1 = openAIChunkToGeminiChunk(
+    {
+      choices: [
+        {
+          delta: {
+            tool_calls: [
+              {
+                index: 0,
+                id: "call_1",
+                function: { name: "get_weather", arguments: '{"ci' },
+              },
+            ],
+          },
+          finish_reason: null,
+        },
+      ],
+    },
+    "gemini/gemini-pro",
+    state
+  );
+  assert.equal(c1, null, "intermediate tool-call chunk emits nothing");
+
+  // Chunk 2: continuation of args.
+  const c2 = openAIChunkToGeminiChunk(
+    {
+      choices: [
+        { delta: { tool_calls: [{ index: 0, function: { arguments: 'ty":"Paris"}' } }] }, finish_reason: null },
+      ],
+    },
+    "gemini/gemini-pro",
+    state
+  );
+  assert.equal(c2, null, "second fragment still emits nothing");
+
+  // Final chunk with finish_reason — emit the accumulated functionCall.
+  const c3 = openAIChunkToGeminiChunk(
+    { choices: [{ delta: {}, finish_reason: "tool_calls" }] },
+    "gemini/gemini-pro",
+    state
+  );
+  assert.ok(c3, "final chunk should emit");
+  const parts = c3!.candidates[0].content.parts as Array<Record<string, unknown>>;
+  const fcPart = parts.find((p) => "functionCall" in p) as
+    | { functionCall: { name: string; args: Record<string, unknown> } }
+    | undefined;
+  assert.ok(fcPart, "final chunk carries the functionCall part");
+  assert.equal(fcPart.functionCall.name, "get_weather");
+  assert.deepEqual(fcPart.functionCall.args, { city: "Paris" });
+  assert.equal(c3!.candidates[0].finishReason, "STOP");
+});
+
+test("stream (e2e): SSE with fragmented tool_calls → Gemini functionCall", async () => {
+  const events = [
+    'data: {"choices":[{"delta":{"role":"assistant","tool_calls":[{"index":0,"id":"call_1","function":{"name":"get_weather","arguments":"{\\"ci"}}]},"finish_reason":null}]}',
+    'data: {"choices":[{"delta":{"tool_calls":[{"index":0,"function":{"arguments":"ty\\":\\"Paris\\"}"}}]},"finish_reason":null}]}',
+    'data: {"choices":[{"delta":{},"finish_reason":"tool_calls"}],"usage":{"prompt_tokens":10,"completion_tokens":5,"total_tokens":15},"model":"gemini-pro"}',
+    "data: [DONE]",
+  ];
+  const body = events.map((e) => e + "\n\n").join("");
+  const upstream = new Response(
+    new ReadableStream<Uint8Array>({
+      start(controller) {
+        controller.enqueue(new TextEncoder().encode(body));
+        controller.close();
+      },
+    }),
+    { status: 200, headers: { "Content-Type": "text/event-stream" } }
+  );
+
+  const geminiResp = transformOpenAISSEToGeminiSSE(upstream, "gemini/gemini-pro");
+  const reader = geminiResp.body!.getReader();
+  const decoder = new TextDecoder();
+  let raw = "";
+  for (;;) {
+    const { done, value } = await reader.read();
+    if (done) break;
+    raw += decoder.decode(value, { stream: true });
+  }
+  raw += decoder.decode();
+
+  const chunks: Array<Record<string, unknown>> = [];
+  for (const line of raw.split("\n")) {
+    const trimmed = line.endsWith("\r") ? line.slice(0, -1) : line;
+    if (!trimmed.startsWith("data:")) continue;
+    const data = trimmed.slice(5).trim();
+    if (!data) continue;
+    chunks.push(JSON.parse(data));
+  }
+
+  // Find the functionCall part anywhere in the emitted stream.
+  let fc: { name: string; args: Record<string, unknown> } | undefined;
+  for (const ch of chunks) {
+    const parts =
+      (ch.candidates as Array<{ content: { parts: Array<Record<string, unknown>> } }>)?.[0]?.content
+        ?.parts ?? [];
+    for (const p of parts) {
+      if ("functionCall" in p) fc = (p as { functionCall: typeof fc }).functionCall;
+    }
+  }
+  assert.ok(fc, "stream should emit a functionCall part");
+  assert.equal(fc!.name, "get_weather");
+  assert.deepEqual(fc!.args, { city: "Paris" });
+});