buzzert/Sybil-2
1
0
Fork 0
Code Issues 2 Pull Requests Actions Packages Projects Releases Wiki Activity

Add per-chat settings UI in web app for additional system prompt and tool checkboxes

Browse Source
This commit is contained in:
Agent
2026-05-24 22:04:05 +00:00
committed by James Magahern
parent 0bf0f95a67
commit 4a2493c421
11 changed files with 321 additions and 32 deletions
Download Patch File Download Diff File Expand all files Collapse all files

88
server/src/llm/chat-tools.ts
View File

@@ -192,7 +192,43 @@ const CHAT_TOOLS: any[] = [
...(env.CHAT_SHELL_TOOL_ENABLED ? [SHELL_EXEC_TOOL] : []),
];
const RESPONSES_CHAT_TOOLS: any[] = CHAT_TOOLS.map((tool) => {
function getToolName(tool: any) {
return typeof tool?.function?.name === "string" ? tool.function.name : null;
}
export function getAvailableChatTools() {
return CHAT_TOOLS.map((tool) => {
const name = getToolName(tool);
if (!name) return null;
return {
name,
description: typeof tool?.function?.description === "string" ? tool.function.description : "",
};
}).filter((tool): tool is { name: string; description: string } => tool !== null);
}
export function normalizeEnabledChatTools(value: unknown) {
if (!Array.isArray(value)) return getAvailableChatTools().map((tool) => tool.name);
const available = new Set(getAvailableChatTools().map((tool) => tool.name));
return [...new Set(value.filter((item): item is string => typeof item === "string").map((item) => item.trim()).filter(Boolean))].filter((name) =>
available.has(name)
);
}
function getEnabledToolSet(params: Pick<ToolAwareCompletionParams, "enabledTools">) {
return new Set(normalizeEnabledChatTools(params.enabledTools));
}
function getEnabledChatTools(params: Pick<ToolAwareCompletionParams, "enabledTools">) {
const enabled = getEnabledToolSet(params);
return CHAT_TOOLS.filter((tool) => {
const name = getToolName(tool);
return name ? enabled.has(name) : false;
});
}
function toResponsesChatTools(tools: any[]) {
return tools.map((tool) => {
if (tool?.type !== "function") return tool;
return {
type: "function",
@@ -201,7 +237,8 @@ const RESPONSES_CHAT_TOOLS: any[] = CHAT_TOOLS.map((tool) => {
parameters: tool.function.parameters,
strict: false,
};
});
});
}
export const CHAT_TOOL_SYSTEM_PROMPT =
"You can use tools to gather up-to-date web information when needed. " +
@@ -243,6 +280,7 @@ type ToolAwareCompletionParams = {
client: OpenAI;
model: string;
messages: ChatMessage[];
enabledTools?: string[];
userLocation?: string;
temperature?: number;
maxTokens?: number;
@@ -384,20 +422,38 @@ function extractHtmlTitle(html: string) {
);
}
function normalizeIncomingMessages(messages: ChatMessage[], userLocation?: string) {
function buildChatToolSystemPrompt(params: Pick<ToolAwareCompletionParams, "enabledTools">) {
const enabled = getEnabledToolSet(params);
return (
"You can use tools to gather up-to-date web information when needed. " +
(enabled.has("web_search") ? "Use web_search for discovery and recent facts. " : "") +
(enabled.has("fetch_url") ? "Use fetch_url to read the full content of a specific page. " : "") +
"Prefer tools when the user asks for current events, verification, sources, or details you do not already have. " +
"When you decide tool use is needed, call the tool immediately in the same response; do not say you are running a tool unless you actually call it. " +
(enabled.has("codex_exec")
? "Use codex_exec when a request needs substantial coding work, repository inspection, shell commands, tests, debugging, or another complex task suited to a persistent Codex workspace. Provide codex_exec a complete prompt with the goal, constraints, assumptions, and expected report-back format. Never ask codex_exec to wait for user input or run interactive commands. "
: "") +
(enabled.has("shell_exec")
? "Use shell_exec for direct non-interactive command-line work on the remote devbox, including quick Python programs, calculations, file inspection, running tests, and small scripts. "
: "") +
"Do not fabricate tool outputs; reason only from provided tool results."
);
}
function normalizeIncomingMessages(messages: ChatMessage[], userLocation?: string, params: Pick<ToolAwareCompletionParams, "enabledTools"> = {}) {
const normalized = messages.map((message) => buildOpenAIConversationMessage(message));
return [{ role: "system", content: CHAT_TOOL_SYSTEM_PROMPT }, buildSystemPromptAugmentationMessage(userLocation), ...normalized];
return [{ role: "system", content: buildChatToolSystemPrompt(params) }, buildSystemPromptAugmentationMessage(userLocation), ...normalized];
}
function normalizePlainIncomingMessages(messages: ChatMessage[], userLocation?: string) {
return [buildSystemPromptAugmentationMessage(userLocation), ...messages.map((message) => buildOpenAIConversationMessage(message))];
}
function normalizeIncomingResponsesInput(messages: ChatMessage[], userLocation?: string) {
function normalizeIncomingResponsesInput(messages: ChatMessage[], userLocation?: string, params: Pick<ToolAwareCompletionParams, "enabledTools"> = {}) {
const normalized = messages.map((message) => buildOpenAIResponsesInputMessage(message));
return [{ role: "system", content: CHAT_TOOL_SYSTEM_PROMPT }, buildSystemPromptAugmentationMessage(userLocation), ...normalized];
return [{ role: "system", content: buildChatToolSystemPrompt(params) }, buildSystemPromptAugmentationMessage(userLocation), ...normalized];
}
async function runExaWebSearchTool(args: WebSearchArgs): Promise<ToolRunOutcome> {
@@ -962,7 +1018,8 @@ async function executeToolCallAndBuildEvent(
}
export async function runToolAwareOpenAIChat(params: ToolAwareCompletionParams): Promise<ToolAwareCompletionResult> {
const input: any[] = normalizeIncomingResponsesInput(params.messages, params.userLocation);
const enabledTools = getEnabledChatTools(params);
const input: any[] = normalizeIncomingResponsesInput(params.messages, params.userLocation, params);
const rawResponses: unknown[] = [];
const toolEvents: ToolExecutionEvent[] = [];
const usageAcc: Required<ToolAwareUsage> = { inputTokens: 0, outputTokens: 0, totalTokens: 0 };
@@ -976,7 +1033,7 @@ export async function runToolAwareOpenAIChat(params: ToolAwareCompletionParams):
input,
temperature: params.temperature,
max_output_tokens: params.maxTokens,
tools: RESPONSES_CHAT_TOOLS,
tools: toResponsesChatTools(enabledTools),
tool_choice: "auto",
parallel_tool_calls: true,
// Tool loops pass response output items back as input; reasoning items need persistence.
@@ -1031,7 +1088,8 @@ export async function runToolAwareOpenAIChat(params: ToolAwareCompletionParams):
}
export async function runToolAwareChatCompletions(params: ToolAwareCompletionParams): Promise<ToolAwareCompletionResult> {
const conversation: any[] = normalizeIncomingMessages(params.messages, params.userLocation);
const enabledTools = getEnabledChatTools(params);
const conversation: any[] = normalizeIncomingMessages(params.messages, params.userLocation, params);
const rawResponses: unknown[] = [];
const toolEvents: ToolExecutionEvent[] = [];
const usageAcc: Required<ToolAwareUsage> = { inputTokens: 0, outputTokens: 0, totalTokens: 0 };
@@ -1045,7 +1103,7 @@ export async function runToolAwareChatCompletions(params: ToolAwareCompletionPar
messages: conversation,
temperature: params.temperature,
max_tokens: params.maxTokens,
tools: CHAT_TOOLS,
tools: enabledTools,
tool_choice: "auto",
} as any);
rawResponses.push(completion);
@@ -1139,7 +1197,8 @@ export async function runPlainChatCompletions(params: ToolAwareCompletionParams)
export async function* runToolAwareOpenAIChatStream(
params: ToolAwareCompletionParams
): AsyncGenerator<ToolAwareStreamingEvent> {
const input: any[] = normalizeIncomingResponsesInput(params.messages, params.userLocation);
const enabledTools = getEnabledChatTools(params);
const input: any[] = normalizeIncomingResponsesInput(params.messages, params.userLocation, params);
const rawResponses: unknown[] = [];
const toolEvents: ToolExecutionEvent[] = [];
const usageAcc: Required<ToolAwareUsage> = { inputTokens: 0, outputTokens: 0, totalTokens: 0 };
@@ -1153,7 +1212,7 @@ export async function* runToolAwareOpenAIChatStream(
input,
temperature: params.temperature,
max_output_tokens: params.maxTokens,
tools: RESPONSES_CHAT_TOOLS,
tools: toResponsesChatTools(enabledTools),
tool_choice: "auto",
parallel_tool_calls: true,
// Tool loops pass response output items back as input; reasoning items need persistence.
@@ -1265,7 +1324,8 @@ export async function* runToolAwareOpenAIChatStream(
export async function* runToolAwareChatCompletionsStream(
params: ToolAwareCompletionParams
): AsyncGenerator<ToolAwareStreamingEvent> {
const conversation: any[] = normalizeIncomingMessages(params.messages, params.userLocation);
const enabledTools = getEnabledChatTools(params);
const conversation: any[] = normalizeIncomingMessages(params.messages, params.userLocation, params);
const rawResponses: unknown[] = [];
const toolEvents: ToolExecutionEvent[] = [];
const usageAcc: Required<ToolAwareUsage> = { inputTokens: 0, outputTokens: 0, totalTokens: 0 };
@@ -1279,7 +1339,7 @@ export async function* runToolAwareChatCompletionsStream(
messages: conversation,
temperature: params.temperature,
max_tokens: params.maxTokens,
tools: CHAT_TOOLS,
tools: enabledTools,
tool_choice: "auto",
stream: true,
stream_options: { include_usage: true },

13
server/src/llm/multiplexer.ts
View File

@@ -1,7 +1,7 @@
import { performance } from "node:perf_hooks";
import { prisma } from "../db.js";
import { anthropicClient, hermesAgentClient, openaiClient, xaiClient } from "./providers.js";
import { buildToolLogMessageData, runPlainChatCompletions, runToolAwareChatCompletions, runToolAwareOpenAIChat } from "./chat-tools.js";
import { buildToolLogMessageData, normalizeEnabledChatTools, runPlainChatCompletions, runToolAwareChatCompletions, runToolAwareOpenAIChat } from "./chat-tools.js";
import { buildAnthropicConversationMessage, getAnthropicSystemPrompt } from "./message-content.js";
import { toPrismaProvider } from "./provider-ids.js";
import type { MultiplexRequest, MultiplexResponse, Provider } from "./types.js";
@@ -47,13 +47,15 @@ export async function runMultiplex(req: MultiplexRequest): Promise<MultiplexResp
let usage: MultiplexResponse["usage"] | undefined;
let raw: unknown;
let toolMessages: ReturnType<typeof buildToolLogMessageData>[] = [];
const enabledTools = normalizeEnabledChatTools(req.enabledTools);
if (req.provider === "openai") {
if (req.provider === "openai" && enabledTools.length > 0) {
const client = openaiClient();
const r = await runToolAwareOpenAIChat({
client,
model: req.model,
messages: req.messages,
enabledTools,
userLocation: req.userLocation,
temperature: req.temperature,
maxTokens: req.maxTokens,
@@ -67,12 +69,13 @@ export async function runMultiplex(req: MultiplexRequest): Promise<MultiplexResp
outText = r.text;
usage = r.usage;
toolMessages = r.toolEvents.map((event) => buildToolLogMessageData(call.chatId, event));
} else if (req.provider === "xai") {
} else if (req.provider === "xai" && enabledTools.length > 0) {
const client = xaiClient();
const r = await runToolAwareChatCompletions({
client,
model: req.model,
messages: req.messages,
enabledTools,
userLocation: req.userLocation,
temperature: req.temperature,
maxTokens: req.maxTokens,
@@ -86,8 +89,8 @@ export async function runMultiplex(req: MultiplexRequest): Promise<MultiplexResp
outText = r.text;
usage = r.usage;
toolMessages = r.toolEvents.map((event) => buildToolLogMessageData(call.chatId, event));
} else if (req.provider === "hermes-agent") {
const client = hermesAgentClient();
} else if (req.provider === "openai" || req.provider === "xai" || req.provider === "hermes-agent") {
const client = req.provider === "openai" ? openaiClient() : req.provider === "xai" ? xaiClient() : hermesAgentClient();
const r = await runPlainChatCompletions({
client,
model: req.model,

8
server/src/llm/streaming.ts
View File

@@ -3,6 +3,7 @@ import { prisma } from "../db.js";
import { anthropicClient, hermesAgentClient, openaiClient, xaiClient } from "./providers.js";
import {
buildToolLogMessageData,
normalizeEnabledChatTools,
runPlainChatCompletionsStream,
runToolAwareChatCompletionsStream,
runToolAwareOpenAIChatStream,
@@ -76,12 +77,14 @@ export async function* runMultiplexStream(req: MultiplexRequest): AsyncGenerator
try {
if (req.provider === "openai" || req.provider === "xai" || req.provider === "hermes-agent") {
const client = req.provider === "openai" ? openaiClient() : req.provider === "xai" ? xaiClient() : hermesAgentClient();
const enabledTools = normalizeEnabledChatTools(req.enabledTools);
const streamEvents =
req.provider === "openai"
req.provider === "openai" && enabledTools.length > 0
? runToolAwareOpenAIChatStream({
client,
model: req.model,
messages: req.messages,
enabledTools,
userLocation: req.userLocation,
temperature: req.temperature,
maxTokens: req.maxTokens,
@@ -91,7 +94,7 @@ export async function* runMultiplexStream(req: MultiplexRequest): AsyncGenerator
chatId: chatId ?? undefined,
},
})
: req.provider === "hermes-agent"
: req.provider === "hermes-agent" || enabledTools.length === 0
? runPlainChatCompletionsStream({
client,
model: req.model,
@@ -109,6 +112,7 @@ export async function* runMultiplexStream(req: MultiplexRequest): AsyncGenerator
client,
model: req.model,
messages: req.messages,
enabledTools,
userLocation: req.userLocation,
temperature: req.temperature,
maxTokens: req.maxTokens,

2
server/src/llm/types.ts
View File

@@ -36,6 +36,8 @@ export type MultiplexRequest = {
provider: Provider;
model: string;
messages: ChatMessage[];
additionalSystemPrompt?: string;
enabledTools?: string[];
userLocation?: string;
temperature?: number;
maxTokens?: number;