This commit is contained in:
@@ -12,6 +12,7 @@ services:
|
||||
OPENAI_API_KEY: ${OPENAI_API_KEY:-}
|
||||
ANTHROPIC_API_KEY: ${ANTHROPIC_API_KEY:-}
|
||||
XAI_API_KEY: ${XAI_API_KEY:-}
|
||||
GEMINI_API_KEY: ${GEMINI_API_KEY:-}
|
||||
HERMES_AGENT_API_BASE_URL: ${HERMES_AGENT_API_BASE_URL:-http://127.0.0.1:8642/v1}
|
||||
HERMES_AGENT_API_KEY: ${HERMES_AGENT_API_KEY:-}
|
||||
HERMES_AGENT_MODEL: ${HERMES_AGENT_MODEL:-}
|
||||
|
||||
@@ -34,11 +34,13 @@ Chat upload limits:
|
||||
"openai": { "models": ["gpt-4.1-mini"], "loadedAt": "2026-02-14T00:00:00.000Z", "error": null },
|
||||
"anthropic": { "models": ["claude-3-5-sonnet-latest"], "loadedAt": null, "error": null },
|
||||
"xai": { "models": ["grok-3-mini"], "loadedAt": null, "error": null },
|
||||
"gemini": { "models": ["gemini-3.5-flash"], "loadedAt": null, "error": null },
|
||||
"hermes-agent": { "models": ["hermes-agent"], "loadedAt": null, "error": null }
|
||||
}
|
||||
}
|
||||
```
|
||||
- OpenAI model lists are filtered to models that are expected to work with the backend's Responses API implementation.
|
||||
- Gemini model lists are loaded from Google's native Models API and filtered to Gemini `generateContent` model ids.
|
||||
- `hermes-agent` is included only when `HERMES_AGENT_API_KEY` is configured. Set it to Hermes `API_SERVER_KEY`, or any non-empty value if that local server does not require auth. `HERMES_AGENT_API_BASE_URL` defaults to `http://127.0.0.1:8642/v1`; set `HERMES_AGENT_MODEL` only when you need an additional fallback/override model id.
|
||||
- The backend loads provider model lists at startup and refreshes them about once every 24 hours. If a later provider refresh fails, the response keeps the last loaded model list for that provider and sets `error` to the latest failure message.
|
||||
|
||||
@@ -56,7 +58,7 @@ Chat upload limits:
|
||||
```
|
||||
|
||||
Behavior notes:
|
||||
- Lists Sybil-managed chat tools that can be enabled for `openai`, `anthropic`, and `xai` chat completions.
|
||||
- Lists Sybil-managed chat tools that can be enabled for `openai`, `anthropic`, `xai`, and `gemini` chat completions.
|
||||
- Optional tools such as `codex_exec` and `shell_exec` appear only when enabled by server environment configuration.
|
||||
|
||||
## Active Runs
|
||||
@@ -128,7 +130,7 @@ Behavior notes:
|
||||
```json
|
||||
{
|
||||
"title": "optional title",
|
||||
"provider": "optional openai|anthropic|xai|hermes-agent",
|
||||
"provider": "optional openai|anthropic|xai|gemini|hermes-agent",
|
||||
"model": "optional model id",
|
||||
"additionalSystemPrompt": "optional stored system prompt",
|
||||
"enabledTools": ["web_search", "fetch_url"],
|
||||
@@ -234,7 +236,7 @@ Notes:
|
||||
```json
|
||||
{
|
||||
"chatId": "optional-chat-id",
|
||||
"provider": "openai|anthropic|xai|hermes-agent",
|
||||
"provider": "openai|anthropic|xai|gemini|hermes-agent",
|
||||
"model": "string",
|
||||
"messages": [
|
||||
{
|
||||
@@ -294,12 +296,14 @@ Behavior notes:
|
||||
- For `openai`, backend calls OpenAI's Responses API and enables internal tool use with an internal system instruction.
|
||||
- For `anthropic`, backend calls Anthropic's Messages API and enables internal tool use with Anthropic `tool_use`/`tool_result` content blocks.
|
||||
- For `xai`, backend calls xAI's OpenAI-compatible Chat Completions API and enables internal tool use with the same internal system instruction.
|
||||
- For `gemini`, backend calls Google's native Gemini `generateContent` API and enables internal tool use with Gemini function calling.
|
||||
- For `hermes-agent`, backend calls the configured Hermes Agent OpenAI-compatible Chat Completions API without adding Sybil-managed tool definitions; Hermes Agent handles its own tools server-side.
|
||||
- For `openai`, image attachments are sent as Responses `input_image` items and text attachments are sent as `input_text` items.
|
||||
- For `gemini`, image attachments are sent as native Gemini `inlineData` parts and text attachments are sent as text parts.
|
||||
- For `xai` and `hermes-agent`, image attachments are sent as Chat Completions content parts alongside text.
|
||||
- For `openai`, Responses calls that can enter the server-managed tool loop use `store: true` so reasoning and function-call items can be passed between tool rounds.
|
||||
- For `anthropic`, image attachments are sent as Messages API `image` blocks using base64 source data; text attachments are added as `text` blocks.
|
||||
- Available Sybil-managed tool calls for `openai`, `anthropic`, and `xai`: `web_search` and `fetch_url`. When `CHAT_CODEX_TOOL_ENABLED=true`, `codex_exec` is also available. When `CHAT_SHELL_TOOL_ENABLED=true`, `shell_exec` is also available.
|
||||
- Available Sybil-managed tool calls for `openai`, `anthropic`, `xai`, and `gemini`: `web_search` and `fetch_url`. When `CHAT_CODEX_TOOL_ENABLED=true`, `codex_exec` is also available. When `CHAT_SHELL_TOOL_ENABLED=true`, `shell_exec` is also available.
|
||||
- `web_search` returns ranked results with per-result summaries/snippets. Its backend engine is selected by `CHAT_WEB_SEARCH_ENGINE` (`exa` default, or `searxng` with `SEARXNG_BASE_URL` set). SearXNG mode requires the instance to allow `format=json`.
|
||||
- `fetch_url` fetches a URL with browser-like navigation headers and returns plaintext page content (HTML converted to text server-side).
|
||||
- `codex_exec` delegates coding, shell, repository inspection, and other complex software tasks to a persistent remote Codex CLI workspace over SSH. The server runs `codex exec --dangerously-bypass-approvals-and-sandbox --skip-git-repo-check <non-interactive wrapped prompt>` on the configured devbox inside `CHAT_CODEX_REMOTE_WORKDIR`, with SSH stdin closed.
|
||||
@@ -417,9 +421,9 @@ Behavior notes:
|
||||
"updatedAt": "...",
|
||||
"starred": false,
|
||||
"starredAt": null,
|
||||
"initiatedProvider": "openai|anthropic|xai|hermes-agent|null",
|
||||
"initiatedProvider": "openai|anthropic|xai|gemini|hermes-agent|null",
|
||||
"initiatedModel": "string|null",
|
||||
"lastUsedProvider": "openai|anthropic|xai|hermes-agent|null",
|
||||
"lastUsedProvider": "openai|anthropic|xai|gemini|hermes-agent|null",
|
||||
"lastUsedModel": "string|null",
|
||||
"additionalSystemPrompt": null,
|
||||
"enabledTools": ["web_search", "fetch_url"]
|
||||
@@ -469,9 +473,9 @@ Behavior notes:
|
||||
"updatedAt": "...",
|
||||
"starred": false,
|
||||
"starredAt": null,
|
||||
"initiatedProvider": "openai|anthropic|xai|hermes-agent|null",
|
||||
"initiatedProvider": "openai|anthropic|xai|gemini|hermes-agent|null",
|
||||
"initiatedModel": "string|null",
|
||||
"lastUsedProvider": "openai|anthropic|xai|hermes-agent|null",
|
||||
"lastUsedProvider": "openai|anthropic|xai|gemini|hermes-agent|null",
|
||||
"lastUsedModel": "string|null",
|
||||
"additionalSystemPrompt": null,
|
||||
"enabledTools": ["web_search", "fetch_url"],
|
||||
|
||||
@@ -21,7 +21,7 @@ Authentication:
|
||||
{
|
||||
"chatId": "optional-chat-id",
|
||||
"persist": true,
|
||||
"provider": "openai|anthropic|xai|hermes-agent",
|
||||
"provider": "openai|anthropic|xai|gemini|hermes-agent",
|
||||
"model": "string",
|
||||
"messages": [
|
||||
{
|
||||
@@ -174,9 +174,11 @@ Terminal tool-call event:
|
||||
- `openai`: backend uses OpenAI's Responses API and may execute internal function tool calls (`web_search`, `fetch_url`, optional `codex_exec`, and optional `shell_exec`) before producing final text.
|
||||
- `anthropic`: backend uses Anthropic's Messages API and may execute the same internal tools with `tool_use`/`tool_result` content blocks before producing final text.
|
||||
- `xai`: backend uses xAI's OpenAI-compatible Chat Completions API and may execute the same internal tool calls before producing final text.
|
||||
- `gemini`: backend uses Google's native Gemini `streamGenerateContent` API and may execute the same internal tool calls before producing final text.
|
||||
- `fetch_url` sends browser-like navigation headers for outbound URL requests to reduce false 403s from sites that reject generic server clients.
|
||||
- `hermes-agent`: backend uses the configured Hermes Agent OpenAI-compatible Chat Completions API. Sybil does not add its own tool definitions for this provider; Hermes Agent handles its own tools server-side. Custom Hermes stream events are normalized away unless they produce text deltas in this SSE contract.
|
||||
- `openai`: image attachments are sent as Responses `input_image` items; text attachments are sent as `input_text` items.
|
||||
- `gemini`: image attachments are sent as native Gemini `inlineData` parts; text attachments are inlined as text parts.
|
||||
- `xai` and `hermes-agent`: image attachments are sent as Chat Completions content parts; text attachments are inlined as text parts.
|
||||
- `openai`: Responses calls that can enter the server-managed tool loop use `store: true` so reasoning and function-call items can be passed between tool rounds.
|
||||
- `anthropic`: streamed via event stream; emits `delta` from `content_block_delta` with `text_delta`, and emits normalized `tool_call` SSE events when Anthropic `tool_use` blocks are executed. Image attachments are sent as base64 `image` blocks and text attachments are appended as `text` blocks.
|
||||
@@ -185,7 +187,7 @@ Terminal tool-call event:
|
||||
- `shell_exec` is available only when `CHAT_SHELL_TOOL_ENABLED=true`. It uses the same devbox SSH configuration, starts in `CHAT_CODEX_REMOTE_WORKDIR`, and runs non-interactive shell commands there with SSH stdin closed, not inside the Sybil server container.
|
||||
- `CHAT_MAX_TOOL_ROUNDS` controls how many model/tool result cycles may occur before the backend returns a tool-call limit message; default is 100.
|
||||
|
||||
Tool-enabled streaming notes (`openai`/`anthropic`/`xai`):
|
||||
Tool-enabled streaming notes (`openai`/`anthropic`/`xai`/`gemini`):
|
||||
- Stream still emits standard `meta`, `delta`, `done|error` events.
|
||||
- Stream may emit `tool_call` events while tool calls are executed.
|
||||
- `delta` events carry assistant text and are emitted incrementally for normal text rounds. The backend may buffer model-native text briefly while determining whether a provider round contains tool calls.
|
||||
|
||||
@@ -57,4 +57,5 @@ Instructions for work under `/Users/buzzert/src/sybil-2/ios`.
|
||||
- OpenAI: `gpt-4.1-mini`
|
||||
- Anthropic: `claude-3-5-sonnet-latest`
|
||||
- xAI: `grok-3-mini`
|
||||
- Gemini: `gemini-3.5-flash`
|
||||
- Hermes Agent: `hermes-agent`
|
||||
|
||||
@@ -4,6 +4,7 @@ public enum Provider: String, Codable, CaseIterable, Hashable, Sendable {
|
||||
case openai
|
||||
case anthropic
|
||||
case xai
|
||||
case gemini
|
||||
case hermesAgent = "hermes-agent"
|
||||
|
||||
public var displayName: String {
|
||||
@@ -11,6 +12,7 @@ public enum Provider: String, Codable, CaseIterable, Hashable, Sendable {
|
||||
case .openai: return "OpenAI"
|
||||
case .anthropic: return "Anthropic"
|
||||
case .xai: return "xAI"
|
||||
case .gemini: return "Gemini"
|
||||
case .hermesAgent: return "Hermes Agent"
|
||||
}
|
||||
}
|
||||
|
||||
@@ -11,11 +11,13 @@ final class SybilSettingsStore {
|
||||
static let preferredOpenAIModel = "sybil.ios.preferredOpenAIModel"
|
||||
static let preferredAnthropicModel = "sybil.ios.preferredAnthropicModel"
|
||||
static let preferredXAIModel = "sybil.ios.preferredXAIModel"
|
||||
static let preferredGeminiModel = "sybil.ios.preferredGeminiModel"
|
||||
static let preferredHermesAgentModel = "sybil.ios.preferredHermesAgentModel"
|
||||
static let quickQuestionPreferredProvider = "sybil.ios.quickQuestionPreferredProvider"
|
||||
static let quickQuestionPreferredOpenAIModel = "sybil.ios.quickQuestionPreferredOpenAIModel"
|
||||
static let quickQuestionPreferredAnthropicModel = "sybil.ios.quickQuestionPreferredAnthropicModel"
|
||||
static let quickQuestionPreferredXAIModel = "sybil.ios.quickQuestionPreferredXAIModel"
|
||||
static let quickQuestionPreferredGeminiModel = "sybil.ios.quickQuestionPreferredGeminiModel"
|
||||
static let quickQuestionPreferredHermesAgentModel = "sybil.ios.quickQuestionPreferredHermesAgentModel"
|
||||
}
|
||||
|
||||
@@ -44,6 +46,7 @@ final class SybilSettingsStore {
|
||||
.openai: defaults.string(forKey: Keys.preferredOpenAIModel) ?? "gpt-4.1-mini",
|
||||
.anthropic: defaults.string(forKey: Keys.preferredAnthropicModel) ?? "claude-3-5-sonnet-latest",
|
||||
.xai: defaults.string(forKey: Keys.preferredXAIModel) ?? "grok-3-mini",
|
||||
.gemini: defaults.string(forKey: Keys.preferredGeminiModel) ?? "gemini-3.5-flash",
|
||||
.hermesAgent: defaults.string(forKey: Keys.preferredHermesAgentModel) ?? "hermes-agent"
|
||||
]
|
||||
self.preferredModelByProvider = preferredModels
|
||||
@@ -54,6 +57,7 @@ final class SybilSettingsStore {
|
||||
.openai: defaults.string(forKey: Keys.quickQuestionPreferredOpenAIModel) ?? preferredModels[.openai] ?? "gpt-4.1-mini",
|
||||
.anthropic: defaults.string(forKey: Keys.quickQuestionPreferredAnthropicModel) ?? preferredModels[.anthropic] ?? "claude-3-5-sonnet-latest",
|
||||
.xai: defaults.string(forKey: Keys.quickQuestionPreferredXAIModel) ?? preferredModels[.xai] ?? "grok-3-mini",
|
||||
.gemini: defaults.string(forKey: Keys.quickQuestionPreferredGeminiModel) ?? preferredModels[.gemini] ?? "gemini-3.5-flash",
|
||||
.hermesAgent: defaults.string(forKey: Keys.quickQuestionPreferredHermesAgentModel) ?? preferredModels[.hermesAgent] ?? "hermes-agent"
|
||||
]
|
||||
}
|
||||
@@ -72,12 +76,14 @@ final class SybilSettingsStore {
|
||||
defaults.set(preferredModelByProvider[.openai], forKey: Keys.preferredOpenAIModel)
|
||||
defaults.set(preferredModelByProvider[.anthropic], forKey: Keys.preferredAnthropicModel)
|
||||
defaults.set(preferredModelByProvider[.xai], forKey: Keys.preferredXAIModel)
|
||||
defaults.set(preferredModelByProvider[.gemini], forKey: Keys.preferredGeminiModel)
|
||||
defaults.set(preferredModelByProvider[.hermesAgent], forKey: Keys.preferredHermesAgentModel)
|
||||
|
||||
defaults.set(quickQuestionPreferredProvider.rawValue, forKey: Keys.quickQuestionPreferredProvider)
|
||||
defaults.set(quickQuestionPreferredModelByProvider[.openai], forKey: Keys.quickQuestionPreferredOpenAIModel)
|
||||
defaults.set(quickQuestionPreferredModelByProvider[.anthropic], forKey: Keys.quickQuestionPreferredAnthropicModel)
|
||||
defaults.set(quickQuestionPreferredModelByProvider[.xai], forKey: Keys.quickQuestionPreferredXAIModel)
|
||||
defaults.set(quickQuestionPreferredModelByProvider[.gemini], forKey: Keys.quickQuestionPreferredGeminiModel)
|
||||
defaults.set(quickQuestionPreferredModelByProvider[.hermesAgent], forKey: Keys.quickQuestionPreferredHermesAgentModel)
|
||||
}
|
||||
|
||||
|
||||
@@ -160,6 +160,7 @@ final class SybilViewModel {
|
||||
.openai: ["gpt-4.1-mini"],
|
||||
.anthropic: ["claude-3-5-sonnet-latest"],
|
||||
.xai: ["grok-3-mini"],
|
||||
.gemini: ["gemini-3.5-flash", "gemini-flash-latest"],
|
||||
.hermesAgent: ["hermes-agent"]
|
||||
]
|
||||
|
||||
|
||||
@@ -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
|
||||
|
||||
@@ -13,6 +13,7 @@ enum Provider {
|
||||
openai
|
||||
anthropic
|
||||
xai
|
||||
gemini
|
||||
hermes_agent @map("hermes-agent")
|
||||
}
|
||||
|
||||
|
||||
@@ -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,
|
||||
|
||||
501
server/src/llm/protocols/gemini-api.ts
Normal file
501
server/src/llm/protocols/gemini-api.ts
Normal file
@@ -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,
|
||||
},
|
||||
};
|
||||
}
|
||||
@@ -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) {
|
||||
|
||||
@@ -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>;
|
||||
|
||||
@@ -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);
|
||||
}
|
||||
|
||||
@@ -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];
|
||||
|
||||
|
||||
@@ -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));
|
||||
|
||||
|
||||
@@ -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",
|
||||
|
||||
@@ -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",
|
||||
});
|
||||
});
|
||||
|
||||
@@ -1,6 +1,6 @@
|
||||
import type { Provider } from "./types.js";
|
||||
|
||||
const PROVIDERS: Provider[] = ["openai", "anthropic", "xai", "hermes-agent"];
|
||||
const PROVIDERS: Provider[] = ["openai", "anthropic", "xai", "gemini", "hermes-agent"];
|
||||
|
||||
function normalizeBaseUrl(value: string) {
|
||||
const trimmed = value.trim();
|
||||
|
||||
@@ -42,12 +42,13 @@ type ToolLogMetadata = {
|
||||
resultPreview?: string | null;
|
||||
};
|
||||
|
||||
const BASE_PROVIDERS: Provider[] = ["openai", "anthropic", "xai"];
|
||||
const BASE_PROVIDERS: Provider[] = ["openai", "anthropic", "xai", "gemini"];
|
||||
const PROVIDERS: Provider[] = [...BASE_PROVIDERS, "hermes-agent"];
|
||||
const PROVIDER_FALLBACK_MODELS: Record<Provider, string[]> = {
|
||||
openai: ["gpt-4.1-mini"],
|
||||
anthropic: ["claude-3-5-sonnet-latest"],
|
||||
xai: ["grok-3-mini"],
|
||||
gemini: ["gemini-3.5-flash", "gemini-flash-latest"],
|
||||
"hermes-agent": ["hermes-agent"],
|
||||
};
|
||||
|
||||
@@ -55,6 +56,7 @@ const EMPTY_MODEL_CATALOG: ModelCatalogResponse["providers"] = {
|
||||
openai: { models: [], loadedAt: null, error: null },
|
||||
anthropic: { models: [], loadedAt: null, error: null },
|
||||
xai: { models: [], loadedAt: null, error: null },
|
||||
gemini: { models: [], loadedAt: null, error: null },
|
||||
};
|
||||
|
||||
function escapeTags(value: string) {
|
||||
@@ -79,6 +81,7 @@ function getProviderLabel(provider: Provider | null | undefined) {
|
||||
if (provider === "openai") return "OpenAI";
|
||||
if (provider === "anthropic") return "Anthropic";
|
||||
if (provider === "xai") return "xAI";
|
||||
if (provider === "gemini") return "Gemini";
|
||||
if (provider === "hermes-agent") return "Hermes Agent";
|
||||
return "";
|
||||
}
|
||||
@@ -266,6 +269,7 @@ async function main() {
|
||||
openai: null,
|
||||
anthropic: null,
|
||||
xai: null,
|
||||
gemini: null,
|
||||
"hermes-agent": null,
|
||||
};
|
||||
let model: string = config.defaultModel ?? pickProviderModel(getModelOptions(modelCatalog, provider), null);
|
||||
|
||||
@@ -1,4 +1,4 @@
|
||||
export type Provider = "openai" | "anthropic" | "xai" | "hermes-agent";
|
||||
export type Provider = "openai" | "anthropic" | "xai" | "gemini" | "hermes-agent";
|
||||
|
||||
export type ProviderModelInfo = {
|
||||
models: string[];
|
||||
|
||||
@@ -123,6 +123,7 @@ const PROVIDER_FALLBACK_MODELS: Record<Provider, string[]> = {
|
||||
openai: ["gpt-4.1-mini"],
|
||||
anthropic: ["claude-3-5-sonnet-latest"],
|
||||
xai: ["grok-3-mini"],
|
||||
gemini: ["gemini-3.5-flash", "gemini-flash-latest"],
|
||||
"hermes-agent": ["hermes-agent"],
|
||||
};
|
||||
|
||||
@@ -130,9 +131,10 @@ const EMPTY_MODEL_CATALOG: ModelCatalogResponse["providers"] = {
|
||||
openai: { models: [], loadedAt: null, error: null },
|
||||
anthropic: { models: [], loadedAt: null, error: null },
|
||||
xai: { models: [], loadedAt: null, error: null },
|
||||
gemini: { models: [], loadedAt: null, error: null },
|
||||
};
|
||||
|
||||
const BASE_PROVIDERS: Provider[] = ["openai", "anthropic", "xai"];
|
||||
const BASE_PROVIDERS: Provider[] = ["openai", "anthropic", "xai", "gemini"];
|
||||
const ALL_PROVIDERS: Provider[] = [...BASE_PROVIDERS, "hermes-agent"];
|
||||
|
||||
const MODEL_PREFERENCES_STORAGE_KEY = "sybil:modelPreferencesByProvider";
|
||||
@@ -149,6 +151,7 @@ const EMPTY_MODEL_PREFERENCES: ProviderModelPreferences = {
|
||||
openai: null,
|
||||
anthropic: null,
|
||||
xai: null,
|
||||
gemini: null,
|
||||
"hermes-agent": null,
|
||||
};
|
||||
const EMPTY_ACTIVE_RUNS: ActiveRunsState = {
|
||||
@@ -345,6 +348,7 @@ function loadStoredModelPreferences() {
|
||||
openai: typeof parsed.openai === "string" && parsed.openai.trim() ? parsed.openai.trim() : null,
|
||||
anthropic: typeof parsed.anthropic === "string" && parsed.anthropic.trim() ? parsed.anthropic.trim() : null,
|
||||
xai: typeof parsed.xai === "string" && parsed.xai.trim() ? parsed.xai.trim() : null,
|
||||
gemini: typeof parsed.gemini === "string" && parsed.gemini.trim() ? parsed.gemini.trim() : null,
|
||||
"hermes-agent":
|
||||
typeof parsed["hermes-agent"] === "string" && parsed["hermes-agent"].trim() ? parsed["hermes-agent"].trim() : null,
|
||||
};
|
||||
@@ -354,7 +358,9 @@ function loadStoredModelPreferences() {
|
||||
}
|
||||
|
||||
function normalizeStoredProvider(value: unknown): Provider {
|
||||
return value === "anthropic" || value === "xai" || value === "openai" || value === "hermes-agent" ? value : "openai";
|
||||
return value === "anthropic" || value === "xai" || value === "gemini" || value === "openai" || value === "hermes-agent"
|
||||
? value
|
||||
: "openai";
|
||||
}
|
||||
|
||||
function normalizeStoredModelPreferences(value: unknown): ProviderModelPreferences {
|
||||
@@ -364,6 +370,7 @@ function normalizeStoredModelPreferences(value: unknown): ProviderModelPreferenc
|
||||
openai: typeof parsed.openai === "string" && parsed.openai.trim() ? parsed.openai.trim() : null,
|
||||
anthropic: typeof parsed.anthropic === "string" && parsed.anthropic.trim() ? parsed.anthropic.trim() : null,
|
||||
xai: typeof parsed.xai === "string" && parsed.xai.trim() ? parsed.xai.trim() : null,
|
||||
gemini: typeof parsed.gemini === "string" && parsed.gemini.trim() ? parsed.gemini.trim() : null,
|
||||
"hermes-agent":
|
||||
typeof parsed["hermes-agent"] === "string" && parsed["hermes-agent"].trim() ? parsed["hermes-agent"].trim() : null,
|
||||
};
|
||||
@@ -395,6 +402,7 @@ function getProviderLabel(provider: Provider | null | undefined) {
|
||||
if (provider === "openai") return "OpenAI";
|
||||
if (provider === "anthropic") return "Anthropic";
|
||||
if (provider === "xai") return "xAI";
|
||||
if (provider === "gemini") return "Gemini";
|
||||
if (provider === "hermes-agent") return "Hermes Agent";
|
||||
return "";
|
||||
}
|
||||
|
||||
@@ -149,7 +149,7 @@ export type CompletionRequestMessage = {
|
||||
attachments?: ChatAttachment[];
|
||||
};
|
||||
|
||||
export type Provider = "openai" | "anthropic" | "xai" | "hermes-agent";
|
||||
export type Provider = "openai" | "anthropic" | "xai" | "gemini" | "hermes-agent";
|
||||
|
||||
export type ProviderModelInfo = {
|
||||
models: string[];
|
||||
|
||||
Reference in New Issue
Block a user