adds gemini support

server/README.md

@@ -1,7 +1,7 @@
 # Sybil Server
 
 Backend API for:
-- LLM multiplexer (OpenAI Responses / Anthropic / xAI Chat Completions-compatible Grok / Hermes Agent)
+- LLM multiplexer (OpenAI Responses / Anthropic / xAI Chat Completions-compatible Grok / Gemini / Hermes Agent)
 - Personal chat database (chats/messages + LLM call log)
 
 ## Stack
@@ -43,6 +43,7 @@ If `ADMIN_TOKEN` is not set, the server runs in open mode (dev).
 - `OPENAI_API_KEY`
 - `ANTHROPIC_API_KEY`
 - `XAI_API_KEY`
+- `GEMINI_API_KEY`
 - `HERMES_AGENT_API_BASE_URL` (`http://127.0.0.1:8642/v1` by default; include the `/v1` suffix)
 - `HERMES_AGENT_API_KEY` (enables the Hermes Agent provider; set to Hermes `API_SERVER_KEY`, or any non-empty value if that local server does not require auth)
 - `HERMES_AGENT_MODEL` (optional fallback/override model id; defaults client-side to `hermes-agent`)
@@ -50,7 +51,7 @@ If `ADMIN_TOKEN` is not set, the server runs in open mode (dev).
 - `CHAT_WEB_SEARCH_ENGINE` (`exa` by default, or `searxng` for chat tool calls only)
 - `SEARXNG_BASE_URL` (required when `CHAT_WEB_SEARCH_ENGINE=searxng`; instance must allow `format=json`)
 - `CHAT_MAX_TOOL_ROUNDS` (`100` by default; maximum model/tool result cycles per chat completion)
-- `CHAT_CODEX_TOOL_ENABLED` (`false` by default; enables the `codex_exec` chat tool for OpenAI/xAI)
+- `CHAT_CODEX_TOOL_ENABLED` (`false` by default; enables the `codex_exec` chat tool for managed-tool providers)
 - `CHAT_CODEX_REMOTE_HOST` (required when Codex tool is enabled; SSH host/IP or `user@host`)
 - `CHAT_CODEX_REMOTE_USER` (optional SSH user when host does not include one)
 - `CHAT_CODEX_REMOTE_PORT` (`22` by default)
@@ -58,7 +59,7 @@ If `ADMIN_TOKEN` is not set, the server runs in open mode (dev).
 - `CHAT_CODEX_SSH_KEY_PATH` (recommended: path to a read-only mounted private key)
 - `CHAT_CODEX_SSH_PRIVATE_KEY_B64` (optional fallback private key delivery)
 - `CHAT_CODEX_EXEC_TIMEOUT_MS` (`600000` by default)
-- `CHAT_SHELL_TOOL_ENABLED` (`false` by default; enables the `shell_exec` chat tool for OpenAI/xAI on the same devbox)
+- `CHAT_SHELL_TOOL_ENABLED` (`false` by default; enables the `shell_exec` chat tool for managed-tool providers on the same devbox)
 - `CHAT_SHELL_EXEC_TIMEOUT_MS` (`120000` by default)
 
 ## API

server/prisma/schema.prisma

@@ -13,6 +13,7 @@ enum Provider {
   openai
   anthropic
   xai
+  gemini
   hermes_agent @map("hermes-agent")
 }
 

server/src/env.ts

@@ -66,6 +66,7 @@ const EnvSchema = z.object({
   OPENAI_API_KEY: z.string().optional(),
   ANTHROPIC_API_KEY: z.string().optional(),
   XAI_API_KEY: z.string().optional(),
+  GEMINI_API_KEY: z.string().optional(),
   HERMES_AGENT_API_BASE_URL: HermesAgentApiBaseUrlSchema,
   HERMES_AGENT_API_KEY: OptionalTrimmedStringSchema,
   HERMES_AGENT_MODEL: OptionalTrimmedStringSchema,

server/src/llm/protocols/gemini-api.ts

@@ -0,0 +1,501 @@
+import {
+  buildChatToolSystemPrompt,
+  executeToolCallAndBuildEvent,
+  getEnabledChatTools,
+  getUnstreamedText,
+  looksLikeDanglingToolIntent,
+  MAX_DANGLING_TOOL_INTENT_RETRIES,
+  MAX_TOOL_ROUNDS,
+  prepareToolCallExecution,
+  type NormalizedToolCall,
+  type ToolAwareCompletionParams,
+  type ToolAwareCompletionResult,
+  type ToolAwareStreamingEvent,
+  type ToolAwareUsage,
+  type ToolExecutionEvent,
+} from "../chat-tools.js";
+import {
+  buildImageSummaryText,
+  buildTextAttachmentPrompt,
+  buildTopLevelSystemPrompt,
+  getImageAttachments,
+  getTextAttachments,
+  parseImageDataUrl,
+} from "../message-content.js";
+import type { ChatMessage } from "../types.js";
+
+type GeminiClient = {
+  apiKey: string;
+  baseURL: string;
+};
+
+const INTERNAL_CORRECTION =
+  "Internal correction: the previous assistant message claimed it would run a tool, but no tool call was made. If the task needs an available tool, call it now. Otherwise provide the final answer directly without saying you will run a tool.";
+
+function normalizeModelResourceName(model: string) {
+  const trimmed = model.trim().replace(/^\/+/, "");
+  return trimmed.startsWith("models/") || trimmed.startsWith("tunedModels/") ? trimmed : `models/${trimmed}`;
+}
+
+function geminiUrl(client: GeminiClient, model: string, method: "generateContent" | "streamGenerateContent", extraParams: Record<string, string> = {}) {
+  const url = new URL(`${client.baseURL.replace(/\/+$/, "")}/${normalizeModelResourceName(model)}:${method}`);
+  url.searchParams.set("key", client.apiKey);
+  for (const [key, value] of Object.entries(extraParams)) {
+    url.searchParams.set(key, value);
+  }
+  return url;
+}
+
+function generationConfig(params: Pick<ToolAwareCompletionParams, "temperature" | "maxTokens">) {
+  const config: Record<string, unknown> = {};
+  if (params.temperature !== undefined) config.temperature = params.temperature;
+  if (params.maxTokens !== undefined) config.maxOutputTokens = params.maxTokens;
+  return Object.keys(config).length ? config : undefined;
+}
+
+function toGeminiJsonSchema(schema: unknown): Record<string, unknown> | undefined {
+  if (!schema || typeof schema !== "object" || Array.isArray(schema)) return undefined;
+  const input = schema as Record<string, unknown>;
+  const output: Record<string, unknown> = {};
+
+  if (typeof input.type === "string") output.type = input.type;
+  if (typeof input.description === "string") output.description = input.description;
+  if (typeof input.format === "string") output.format = input.format;
+  if (typeof input.nullable === "boolean") output.nullable = input.nullable;
+  if (Array.isArray(input.enum)) output.enum = input.enum.filter((value) => typeof value === "string");
+  if (Array.isArray(input.required)) output.required = input.required.filter((value) => typeof value === "string");
+
+  const items = toGeminiJsonSchema(input.items);
+  if (items) output.items = items;
+
+  if (input.properties && typeof input.properties === "object" && !Array.isArray(input.properties)) {
+    const properties: Record<string, unknown> = {};
+    for (const [key, value] of Object.entries(input.properties)) {
+      const propertySchema = toGeminiJsonSchema(value);
+      if (propertySchema) properties[key] = propertySchema;
+    }
+    if (Object.keys(properties).length) output.properties = properties;
+  }
+
+  return Object.keys(output).length ? output : undefined;
+}
+
+function toGeminiTools(tools: any[]) {
+  const functionDeclarations = tools
+    .map((tool) => {
+      if (tool?.type !== "function") return null;
+      const declaration: Record<string, unknown> = {
+        name: tool.function.name,
+        description: tool.function.description,
+      };
+      const parameters = toGeminiJsonSchema(tool.function.parameters);
+      if (parameters) declaration.parameters = parameters;
+      return declaration;
+    })
+    .filter(Boolean);
+
+  return functionDeclarations.length ? [{ functionDeclarations }] : undefined;
+}
+
+function toContentParts(message: ChatMessage) {
+  const imageAttachments = getImageAttachments(message);
+  const textAttachments = getTextAttachments(message);
+  const parts: Array<Record<string, unknown>> = [];
+
+  for (const attachment of imageAttachments) {
+    const source = parseImageDataUrl(attachment);
+    parts.push({
+      inlineData: {
+        mimeType: source.mediaType,
+        data: source.data,
+      },
+    });
+  }
+
+  const imageSummary = buildImageSummaryText(imageAttachments);
+  if (imageSummary) {
+    parts.push({ text: imageSummary });
+  }
+
+  for (const attachment of textAttachments) {
+    parts.push({ text: buildTextAttachmentPrompt(attachment) });
+  }
+
+  if (message.content.trim()) {
+    parts.push({ text: message.content });
+  }
+
+  return parts.length ? parts : [{ text: "" }];
+}
+
+function buildConversationContent(message: ChatMessage) {
+  if (message.role === "system") {
+    throw new Error("System messages must be handled separately for Gemini.");
+  }
+
+  if (message.role === "tool") {
+    const name = message.name?.trim() || "tool";
+    return {
+      role: "user",
+      parts: [{ text: `Tool output (${name}):\n${message.content}` }],
+    };
+  }
+
+  return {
+    role: message.role === "assistant" ? "model" : "user",
+    parts: toContentParts(message),
+  };
+}
+
+function buildBaseContents(messages: ChatMessage[]) {
+  return messages.filter((message) => message.role !== "system").map((message) => buildConversationContent(message));
+}
+
+function buildSystemInstruction(params: ToolAwareCompletionParams, toolSystemPrompt?: string) {
+  const text = buildTopLevelSystemPrompt(params.messages, params.userLocation, toolSystemPrompt);
+  return text ? { parts: [{ text }] } : undefined;
+}
+
+function mergeUsage(acc: Required<ToolAwareUsage>, usage: any) {
+  const normalized = normalizeUsage(usage);
+  if (!normalized) return false;
+  acc.inputTokens += normalized.inputTokens;
+  acc.outputTokens += normalized.outputTokens;
+  acc.totalTokens += normalized.totalTokens;
+  return true;
+}
+
+function normalizeUsage(usage: any) {
+  if (!usage) return null;
+  const inputTokens = usage.promptTokenCount ?? 0;
+  const outputTokens = usage.candidatesTokenCount ?? 0;
+  const totalTokens = usage.totalTokenCount ?? inputTokens + outputTokens;
+  return { inputTokens, outputTokens, totalTokens };
+}
+
+function getCandidate(response: any) {
+  return Array.isArray(response?.candidates) ? response.candidates[0] : null;
+}
+
+function getParts(response: any) {
+  const parts = getCandidate(response)?.content?.parts;
+  return Array.isArray(parts) ? parts : [];
+}
+
+function extractText(response: any) {
+  return getParts(response)
+    .map((part: any) => (typeof part?.text === "string" ? part.text : ""))
+    .join("");
+}
+
+function stringifyToolArgs(args: unknown) {
+  try {
+    return JSON.stringify(args ?? {});
+  } catch {
+    return "{}";
+  }
+}
+
+function normalizeToolCallsFromParts(parts: any[], round: number): NormalizedToolCall[] {
+  return parts
+    .filter((part) => part?.functionCall)
+    .map((part, index) => ({
+      id: part.functionCall.id ?? `tool_call_${round}_${index}`,
+      name: part.functionCall.name ?? "unknown_tool",
+      arguments: stringifyToolArgs(part.functionCall.args),
+    }));
+}
+
+function buildFunctionResponsePart(call: NormalizedToolCall, toolResult: unknown) {
+  return {
+    functionResponse: {
+      id: call.id,
+      name: call.name,
+      response: toolResult,
+    },
+  };
+}
+
+function appendCorrection(conversation: any[], text: string) {
+  conversation.push({ role: "model", parts: [{ text }] });
+  conversation.push({ role: "user", parts: [{ text: INTERNAL_CORRECTION }] });
+}
+
+async function parseGeminiResponse(response: Response) {
+  const bodyText = await response.text();
+  let body: any = null;
+  try {
+    body = bodyText ? JSON.parse(bodyText) : null;
+  } catch {
+    body = { raw: bodyText };
+  }
+
+  if (!response.ok) {
+    throw new Error(body?.error?.message ?? `Gemini API request failed with status ${response.status}.`);
+  }
+
+  return body;
+}
+
+async function generateContent(params: ToolAwareCompletionParams, body: Record<string, unknown>) {
+  const response = await fetch(geminiUrl(params.client, params.model, "generateContent"), {
+    method: "POST",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify(body),
+  });
+  return parseGeminiResponse(response);
+}
+
+function getFailureMessage(response: any, text: string, toolCallCount: number) {
+  const promptBlockReason = response?.promptFeedback?.blockReason;
+  if (promptBlockReason) return `Gemini prompt blocked: ${promptBlockReason}.`;
+
+  const candidate = getCandidate(response);
+  const finishReason = candidate?.finishReason;
+  if (!finishReason || finishReason === "STOP" || finishReason === "MAX_TOKENS") return null;
+  if (text || toolCallCount > 0) return null;
+  return candidate?.finishMessage ?? `Gemini response stopped: ${finishReason}.`;
+}
+
+function buildRequest(params: ToolAwareCompletionParams, conversation: any[], enabledTools: any[] = []) {
+  const tools = toGeminiTools(enabledTools);
+  return {
+    contents: conversation,
+    systemInstruction: buildSystemInstruction(params, enabledTools.length ? buildChatToolSystemPrompt(params) : undefined),
+    generationConfig: generationConfig(params),
+    tools,
+    toolConfig: tools ? { functionCallingConfig: { mode: "AUTO" } } : undefined,
+  };
+}
+
+export async function completeWithGeminiApi(params: ToolAwareCompletionParams): Promise<ToolAwareCompletionResult> {
+  const enabledTools = getEnabledChatTools(params);
+  const conversation = buildBaseContents(params.messages);
+  const rawResponses: unknown[] = [];
+  const toolEvents: ToolExecutionEvent[] = [];
+  const usageAcc: Required<ToolAwareUsage> = { inputTokens: 0, outputTokens: 0, totalTokens: 0 };
+  let sawUsage = false;
+  let totalToolCalls = 0;
+  let danglingToolIntentRetries = 0;
+
+  for (let round = 0; round < MAX_TOOL_ROUNDS; round += 1) {
+    const response = await generateContent(params, buildRequest(params, conversation, enabledTools));
+    rawResponses.push(response);
+    sawUsage = mergeUsage(usageAcc, response?.usageMetadata) || sawUsage;
+
+    const parts = getParts(response);
+    const text = extractText(response);
+    const normalizedToolCalls = normalizeToolCallsFromParts(parts, round);
+    const failureMessage = getFailureMessage(response, text, normalizedToolCalls.length);
+    if (failureMessage) throw new Error(failureMessage);
+
+    if (!normalizedToolCalls.length) {
+      if (danglingToolIntentRetries < MAX_DANGLING_TOOL_INTENT_RETRIES && looksLikeDanglingToolIntent(text)) {
+        danglingToolIntentRetries += 1;
+        appendCorrection(conversation, text);
+        continue;
+      }
+      return {
+        text,
+        usage: sawUsage ? usageAcc : undefined,
+        raw: { responses: rawResponses, toolCallsUsed: totalToolCalls, api: "gemini.generateContent" },
+        toolEvents,
+      };
+    }
+
+    totalToolCalls += normalizedToolCalls.length;
+    conversation.push({ role: "model", parts });
+
+    const toolResultParts: any[] = [];
+    for (const call of normalizedToolCalls) {
+      const { execution } = prepareToolCallExecution(call);
+      const { event, toolResult } = await executeToolCallAndBuildEvent(call, execution, params);
+      toolEvents.push(event);
+      toolResultParts.push(buildFunctionResponsePart(call, toolResult));
+    }
+
+    conversation.push({ role: "user", parts: toolResultParts });
+  }
+
+  return {
+    text: "I reached the tool-call limit while gathering information. Please narrow the request and try again.",
+    usage: sawUsage ? usageAcc : undefined,
+    raw: { responses: rawResponses, toolCallsUsed: totalToolCalls, toolCallLimitReached: true, api: "gemini.generateContent" },
+    toolEvents,
+  };
+}
+
+function findSseBoundary(buffer: string) {
+  const crlf = buffer.indexOf("\r\n\r\n");
+  const lf = buffer.indexOf("\n\n");
+  if (crlf === -1) return lf === -1 ? null : { index: lf, length: 2 };
+  if (lf === -1) return { index: crlf, length: 4 };
+  return crlf < lf ? { index: crlf, length: 4 } : { index: lf, length: 2 };
+}
+
+function parseSseEvent(rawEvent: string) {
+  const data = rawEvent
+    .split(/\r?\n/)
+    .filter((line) => line.startsWith("data:"))
+    .map((line) => line.slice("data:".length).trimStart())
+    .join("\n")
+    .trim();
+  if (!data || data === "[DONE]") return null;
+  return JSON.parse(data);
+}
+
+async function* streamGeminiResponses(params: ToolAwareCompletionParams, body: Record<string, unknown>) {
+  const response = await fetch(geminiUrl(params.client, params.model, "streamGenerateContent", { alt: "sse" }), {
+    method: "POST",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify(body),
+  });
+
+  if (!response.ok) {
+    await parseGeminiResponse(response);
+    return;
+  }
+
+  if (!response.body) {
+    throw new Error("Gemini stream response did not include a body.");
+  }
+
+  const reader = response.body.getReader();
+  const decoder = new TextDecoder();
+  let buffer = "";
+
+  while (true) {
+    const { value, done } = await reader.read();
+    if (done) break;
+    buffer += decoder.decode(value, { stream: true });
+    let boundary = findSseBoundary(buffer);
+    while (boundary) {
+      const rawEvent = buffer.slice(0, boundary.index);
+      buffer = buffer.slice(boundary.index + boundary.length);
+      const event = parseSseEvent(rawEvent);
+      if (event) yield event;
+      boundary = findSseBoundary(buffer);
+    }
+  }
+
+  buffer += decoder.decode();
+  const tail = buffer.trim();
+  if (tail) {
+    const event = parseSseEvent(tail);
+    if (event) yield event;
+  }
+}
+
+export async function* streamWithGeminiApi(params: ToolAwareCompletionParams): AsyncGenerator<ToolAwareStreamingEvent> {
+  const enabledTools = getEnabledChatTools(params);
+  const conversation = buildBaseContents(params.messages);
+  const rawResponses: unknown[] = [];
+  const toolEvents: ToolExecutionEvent[] = [];
+  const usageAcc: Required<ToolAwareUsage> = { inputTokens: 0, outputTokens: 0, totalTokens: 0 };
+  let sawUsage = false;
+  let totalToolCalls = 0;
+  let danglingToolIntentRetries = 0;
+
+  if (!enabledTools.length) {
+    let text = "";
+    let latestUsage: any = null;
+    for await (const response of streamGeminiResponses(params, buildRequest(params, conversation))) {
+      rawResponses.push(response);
+      if (response?.usageMetadata) latestUsage = response.usageMetadata;
+      const failureMessage = getFailureMessage(response, extractText(response), 0);
+      if (failureMessage) throw new Error(failureMessage);
+      const delta = extractText(response);
+      if (delta) {
+        text += delta;
+        yield { type: "delta", text: delta };
+      }
+    }
+
+    sawUsage = mergeUsage(usageAcc, latestUsage) || sawUsage;
+
+    yield {
+      type: "done",
+      result: {
+        text,
+        usage: sawUsage ? usageAcc : undefined,
+        raw: { streamed: true, responses: rawResponses, toolCallsUsed: 0, api: "gemini.streamGenerateContent" },
+        toolEvents: [],
+      },
+    };
+    return;
+  }
+
+  for (let round = 0; round < MAX_TOOL_ROUNDS; round += 1) {
+    const roundParts: any[] = [];
+    let roundText = "";
+    let latestRoundResponse: any = null;
+    let latestRoundUsage: any = null;
+
+    for await (const response of streamGeminiResponses(params, buildRequest(params, conversation, enabledTools))) {
+      rawResponses.push(response);
+      latestRoundResponse = response;
+      if (response?.usageMetadata) latestRoundUsage = response.usageMetadata;
+      roundParts.push(...getParts(response));
+      roundText += extractText(response);
+    }
+
+    sawUsage = mergeUsage(usageAcc, latestRoundUsage) || sawUsage;
+
+    const normalizedToolCalls = normalizeToolCallsFromParts(roundParts, round);
+    const failureMessage = getFailureMessage(latestRoundResponse ?? { candidates: [{ content: { parts: roundParts } }] }, roundText, normalizedToolCalls.length);
+    if (failureMessage) throw new Error(failureMessage);
+
+    if (!normalizedToolCalls.length) {
+      if (danglingToolIntentRetries < MAX_DANGLING_TOOL_INTENT_RETRIES && looksLikeDanglingToolIntent(roundText)) {
+        danglingToolIntentRetries += 1;
+        appendCorrection(conversation, roundText);
+        continue;
+      }
+      const unstreamedText = getUnstreamedText(roundText, "");
+      if (unstreamedText) {
+        yield { type: "delta", text: unstreamedText };
+      }
+      yield {
+        type: "done",
+        result: {
+          text: roundText,
+          usage: sawUsage ? usageAcc : undefined,
+          raw: { streamed: true, responses: rawResponses, toolCallsUsed: totalToolCalls, api: "gemini.streamGenerateContent" },
+          toolEvents,
+        },
+      };
+      return;
+    }
+
+    totalToolCalls += normalizedToolCalls.length;
+    conversation.push({ role: "model", parts: roundParts });
+
+    const toolResultParts: any[] = [];
+    for (const call of normalizedToolCalls) {
+      const { event: initiatedEvent, execution } = prepareToolCallExecution(call);
+      yield { type: "tool_call", event: initiatedEvent };
+      const { event, toolResult } = await executeToolCallAndBuildEvent(call, execution, params);
+      toolEvents.push(event);
+      yield { type: "tool_call", event };
+      toolResultParts.push(buildFunctionResponsePart(call, toolResult));
+    }
+
+    conversation.push({ role: "user", parts: toolResultParts });
+  }
+
+  yield {
+    type: "done",
+    result: {
+      text: "I reached the tool-call limit while gathering information. Please narrow the request and try again.",
+      usage: sawUsage ? usageAcc : undefined,
+      raw: {
+        streamed: true,
+        responses: rawResponses,
+        toolCallsUsed: totalToolCalls,
+        toolCallLimitReached: true,
+        api: "gemini.streamGenerateContent",
+      },
+      toolEvents,
+    },
+  };
+}

server/src/llm/provider-adapters.ts

@@ -5,10 +5,11 @@ import {
   type ToolAwareStreamingEvent,
 } from "./chat-tools.js";
 import { completeWithChatCompletionsApi, streamWithChatCompletionsApi } from "./protocols/chat-completions-api.js";
+import { completeWithGeminiApi, streamWithGeminiApi } from "./protocols/gemini-api.js";
 import { completeWithMessagesApi, streamWithMessagesApi } from "./protocols/messages-api.js";
 import { completeWithResponsesApi, streamWithResponsesApi } from "./protocols/responses-api.js";
 import { env } from "../env.js";
-import { anthropicClient, hermesAgentClient, isHermesAgentConfigured, openaiClient, xaiClient } from "./providers.js";
+import { anthropicClient, geminiClient, hermesAgentClient, isHermesAgentConfigured, openaiClient, xaiClient } from "./providers.js";
 import type { ChatMessage, Provider } from "./types.js";
 
 type ProviderAdapterParams = {
@@ -27,7 +28,7 @@ export type ProviderChatAdapter = {
   stream(params: ProviderAdapterParams): AsyncGenerator<ToolAwareStreamingEvent>;
 };
 
-type ChatProtocolId = "chat-completions" | "messages" | "responses";
+type ChatProtocolId = "chat-completions" | "gemini" | "messages" | "responses";
 
 type ChatProtocol = {
   id: ChatProtocolId;
@@ -39,6 +40,7 @@ type ModelCatalogSpec = {
   enabled?: () => boolean;
   fetchModels(client: any): Promise<string[]>;
   fallbackModels?: () => string[];
+  sortModels?: (models: string[]) => string[];
 };
 
 type ProviderBackendSpec = {
@@ -61,6 +63,12 @@ const messagesProtocol: ChatProtocol = {
   stream: streamWithMessagesApi,
 };
 
+const geminiProtocol: ChatProtocol = {
+  id: "gemini",
+  complete: completeWithGeminiApi,
+  stream: streamWithGeminiApi,
+};
+
 const responsesProtocol: ChatProtocol = {
   id: "responses",
   complete: completeWithResponsesApi,
@@ -77,6 +85,10 @@ function modelIdsFromListResponse(page: any) {
     : [];
 }
 
+function stripModelResourcePrefix(model: string) {
+  return model.startsWith("models/") ? model.slice("models/".length) : model;
+}
+
 function isLikelyResponsesApiModel(model: string) {
   const id = model.toLowerCase();
   if (id.includes("embedding") || id.includes("moderation")) return false;
@@ -86,6 +98,37 @@ function isLikelyResponsesApiModel(model: string) {
   return /^(gpt-|o\d|chatgpt-)/.test(id);
 }
 
+function isLikelyGeminiChatModel(model: string) {
+  const id = model.toLowerCase();
+  if (!id.startsWith("gemini-")) return false;
+  if (id.includes("embedding") || id.includes("embed")) return false;
+  if (id.includes("image") || id.includes("imagen") || id.includes("veo")) return false;
+  if (id.includes("audio") || id.includes("tts") || id.includes("live")) return false;
+  if (id.includes("computer-use") || id.includes("robotics")) return false;
+  return true;
+}
+
+function preferGeminiModels(models: string[]) {
+  const preferred = [
+    "gemini-3.5-flash",
+    "gemini-flash-latest",
+    "gemini-3.1-flash-lite",
+    "gemini-3-flash-preview",
+    "gemini-pro-latest",
+  ];
+  const modelSet = new Set(models);
+  return [...preferred.filter((model) => modelSet.delete(model)), ...[...modelSet].sort((a, b) => a.localeCompare(b))];
+}
+
+async function fetchJson(url: URL): Promise<any> {
+  const response = await fetch(url);
+  const body: any = await response.json().catch(() => null);
+  if (!response.ok) {
+    throw new Error(body?.error?.message ?? `Gemini model fetch failed with status ${response.status}.`);
+  }
+  return body;
+}
+
 function withClient(params: ProviderAdapterParams, client: any, enabledTools?: string[]): ToolAwareCompletionParams {
   return {
     client,
@@ -160,6 +203,29 @@ const backendSpecs: Record<Provider, ProviderBackendSpec> = {
       },
     },
   },
+  gemini: {
+    createClient: geminiClient,
+    plainProtocol: geminiProtocol,
+    toolProtocol: geminiProtocol,
+    managedTools: true,
+    modelCatalog: {
+      async fetchModels(client) {
+        const url = new URL(`${client.baseURL.replace(/\/+$/, "")}/models`);
+        url.searchParams.set("key", client.apiKey);
+        url.searchParams.set("pageSize", "1000");
+        const page = await fetchJson(url);
+        return Array.isArray(page?.models)
+          ? page.models
+              .filter((model: any) => Array.isArray(model?.supportedGenerationMethods) && model.supportedGenerationMethods.includes("generateContent"))
+              .map((model: any) => model?.name)
+              .filter((id: unknown): id is string => typeof id === "string")
+              .map(stripModelResourcePrefix)
+              .filter(isLikelyGeminiChatModel)
+          : [];
+      },
+      sortModels: preferGeminiModels,
+    },
+  },
   "hermes-agent": {
     createClient: hermesAgentClient,
     plainProtocol: chatCompletionsProtocol,
@@ -209,7 +275,8 @@ export function listModelCatalogProviders(): Provider[] {
 export async function fetchProviderCatalogModels(provider: Provider) {
   const spec = backendSpecs[provider].modelCatalog;
   if (!spec) return [];
-  return uniqSorted(await spec.fetchModels(backendSpecs[provider].createClient()));
+  const models = uniqSorted(await spec.fetchModels(backendSpecs[provider].createClient()));
+  return spec.sortModels ? spec.sortModels(models) : models;
 }
 
 export function getProviderCatalogFallbackModels(provider: Provider) {

server/src/llm/provider-ids.ts

@@ -6,6 +6,7 @@ const apiToPrismaProvider = {
   openai: "openai",
   anthropic: "anthropic",
   xai: "xai",
+  gemini: "gemini",
   "hermes-agent": "hermes_agent",
 } as const satisfies Record<Provider, PrismaProvider>;
 
@@ -13,6 +14,7 @@ const prismaToApiProvider = {
   openai: "openai",
   anthropic: "anthropic",
   xai: "xai",
+  gemini: "gemini",
   hermes_agent: "hermes-agent",
   "hermes-agent": "hermes-agent",
 } as const satisfies Record<PrismaProvider | "hermes-agent", Provider>;

server/src/llm/providers.ts

@@ -1,5 +1,5 @@
-import OpenAI from "openai";
 import Anthropic from "@anthropic-ai/sdk";
+import OpenAI from "openai";
 import { env } from "../env.js";
 
 export function openaiClient() {
@@ -13,6 +13,14 @@ export function xaiClient() {
   return new OpenAI({ apiKey: env.XAI_API_KEY, baseURL: "https://api.x.ai/v1" });
 }
 
+export function geminiClient() {
+  if (!env.GEMINI_API_KEY) throw new Error("GEMINI_API_KEY not set");
+  return {
+    apiKey: env.GEMINI_API_KEY,
+    baseURL: "https://generativelanguage.googleapis.com/v1beta",
+  };
+}
+
 export function isHermesAgentConfigured() {
   return Boolean(env.HERMES_AGENT_API_KEY);
 }

server/src/llm/types.ts

@@ -1,4 +1,4 @@
-export const PROVIDERS = ["openai", "anthropic", "xai", "hermes-agent"] as const;
+export const PROVIDERS = ["openai", "anthropic", "xai", "gemini", "hermes-agent"] as const;
 
 export type Provider = (typeof PROVIDERS)[number];
 

server/src/routes.ts

@@ -16,7 +16,7 @@ import { exaClient } from "./search/exa.js";
 import { isFreshSearchCacheHit, normalizeSearchQuery } from "./search-cache.js";
 import type { ChatAttachment } from "./llm/types.js";
 
-const ProviderSchema = z.enum(["openai", "anthropic", "xai", "hermes-agent"]);
+const ProviderSchema = z.enum(["openai", "anthropic", "xai", "gemini", "hermes-agent"]);
 const MAX_ADDITIONAL_SYSTEM_PROMPT_CHARS = 12_000;
 const EnabledToolsSchema = z.array(z.string().trim().min(1).max(80)).max(20).transform((value) => normalizeEnabledChatTools(value));
 

server/tests/provider-adapters.test.ts

@@ -27,6 +27,12 @@ test("provider backend registry selects chat protocol and managed-tool mode", ()
     managedTools: true,
     enabledTools: ["web_search"],
   });
+  assert.deepEqual(describeProviderChatBackend("gemini", ["web_search"]), {
+    provider: "gemini",
+    protocol: "gemini",
+    managedTools: true,
+    enabledTools: ["web_search"],
+  });
   assert.deepEqual(describeProviderChatBackend("hermes-agent", ["web_search"]), {
     provider: "hermes-agent",
     protocol: "chat-completions",

server/tests/provider-ids.test.ts

@@ -5,8 +5,10 @@ import { fromPrismaProvider, serializeProviderFields, toPrismaProvider } from ".
 test("Hermes Agent provider id maps between API and Prisma enum forms", () => {
   assert.equal(toPrismaProvider("hermes-agent"), "hermes_agent");
   assert.equal(fromPrismaProvider("hermes_agent"), "hermes-agent");
-  assert.deepEqual(serializeProviderFields({ initiatedProvider: "hermes_agent", lastUsedProvider: "xai" }), {
+  assert.equal(toPrismaProvider("gemini"), "gemini");
+  assert.equal(fromPrismaProvider("gemini"), "gemini");
+  assert.deepEqual(serializeProviderFields({ initiatedProvider: "hermes_agent", lastUsedProvider: "gemini" }), {
     initiatedProvider: "hermes-agent",
-    lastUsedProvider: "xai",
+    lastUsedProvider: "gemini",
   });
 });