Use configured inference model for MCP assists

This commit is contained in:
Codex
2026-04-29 16:03:13 +00:00
parent 50e8775276
commit 524fb29ed5
6 changed files with 78 additions and 26 deletions
+6 -5
View File
@@ -1,8 +1,7 @@
import { NextResponse } from "next/server"
import { resolveOpenAiCompatibleBaseUrl, resolvePrimaryInferenceModel } from "@/app/lib/inferenceModel"
const REQUEST_TIMEOUT_MS = Number.parseInt(process.env.MCP_INSTALL_ASSIST_TIMEOUT_MS || "45000", 10)
const MODEL = process.env.MCP_PREFLIGHT_LLM_MODEL || process.env.OPENAI_MODEL || "gpt-4o-mini"
const BASE_URL = (process.env.MCP_PREFLIGHT_LLM_BASE_URL || process.env.OPENAI_BASE_URL || process.env.VLLM_BASE_URL || "http://localhost:8000/v1").replace(/\/+$/, "")
function extractJsonObject(value: string) {
const trimmed = value.trim()
@@ -34,8 +33,10 @@ export async function POST(request: Request) {
const body = await request.json().catch(() => ({}))
const prompt = typeof body?.prompt === "string" ? body.prompt.trim() : ""
if (!prompt) throw new Error("install prompt is required")
const model = await resolvePrimaryInferenceModel()
if (!model) throw new Error("No primary inference model is configured")
const response = await fetch(`${BASE_URL}/chat/completions`, {
const response = await fetch(`${resolveOpenAiCompatibleBaseUrl()}/chat/completions`, {
method: "POST",
signal: controller.signal,
headers: {
@@ -45,7 +46,7 @@ export async function POST(request: Request) {
: {}),
},
body: JSON.stringify({
model: MODEL,
model,
temperature: 0.1,
response_format: { type: "json_object" },
messages: [
@@ -88,7 +89,7 @@ export async function POST(request: Request) {
uploadEntrypoint: typeof suggestion?.uploadEntrypoint === "string" ? suggestion.uploadEntrypoint : "",
notes: typeof suggestion?.notes === "string" ? suggestion.notes : "",
},
model: MODEL,
model,
})
} catch (error) {
const message = error instanceof Error ? error.message : "Failed to assist MCP install"
+6 -5
View File
@@ -1,8 +1,7 @@
import { NextResponse } from "next/server"
import { resolveOpenAiCompatibleBaseUrl, resolvePrimaryInferenceModel } from "@/app/lib/inferenceModel"
const REQUEST_TIMEOUT_MS = Number.parseInt(process.env.MCP_REGISTRY_ASSIST_TIMEOUT_MS || "45000", 10)
const MODEL = process.env.MCP_PREFLIGHT_LLM_MODEL || process.env.OPENAI_MODEL || "gpt-4o-mini"
const BASE_URL = (process.env.MCP_PREFLIGHT_LLM_BASE_URL || process.env.OPENAI_BASE_URL || process.env.VLLM_BASE_URL || "http://localhost:8000/v1").replace(/\/+$/, "")
function extractJsonObject(value: string) {
const trimmed = value.trim()
@@ -19,8 +18,10 @@ export async function POST(request: Request) {
const body = await request.json().catch(() => ({}))
const prompt = typeof body?.prompt === "string" ? body.prompt.trim() : ""
if (!prompt) throw new Error("registry prompt is required")
const model = await resolvePrimaryInferenceModel()
if (!model) throw new Error("No primary inference model is configured")
const response = await fetch(`${BASE_URL}/chat/completions`, {
const response = await fetch(`${resolveOpenAiCompatibleBaseUrl()}/chat/completions`, {
method: "POST",
signal: controller.signal,
headers: {
@@ -30,7 +31,7 @@ export async function POST(request: Request) {
: {}),
},
body: JSON.stringify({
model: MODEL,
model,
temperature: 0.1,
response_format: { type: "json_object" },
messages: [
@@ -56,7 +57,7 @@ export async function POST(request: Request) {
baseUrl: typeof suggestion?.baseUrl === "string" ? suggestion.baseUrl : "",
description: typeof suggestion?.description === "string" ? suggestion.description : "",
},
model: MODEL,
model,
})
} catch (error) {
const message = error instanceof Error ? error.message : "Failed to assist registry setup"
+1 -1
View File
@@ -373,7 +373,7 @@ export async function POST(request: Request) {
attempted: true,
ok: false,
provider: "openai-compatible",
model: process.env.MCP_PREFLIGHT_LLM_MODEL || process.env.OPENAI_MODEL || "",
model: "",
baseUrl: process.env.MCP_PREFLIGHT_LLM_BASE_URL || process.env.OPENAI_BASE_URL || process.env.VLLM_BASE_URL || "",
summary: "LLM-assisted repair could not complete.",
changes: [],
+50
View File
@@ -0,0 +1,50 @@
import { execFile } from "node:child_process"
import { promisify } from "node:util"
import { OPENSHELL_BIN, hostCommandEnv } from "./hostCommands"
const execFileAsync = promisify(execFile)
function stripAnsi(value: string) {
return value.replace(/\x1B\[[0-?]*[ -/]*[@-~]/g, "")
}
function readField(output: string, label: string) {
const match = output.match(new RegExp(`^\\s*${label}:\\s*(.+?)\\s*$`, "im"))
return match ? match[1].trim() : ""
}
function parseRouteModel(output: string, label: "Gateway inference" | "System inference") {
const otherLabel = label === "Gateway inference" ? "System inference" : "Gateway inference"
const pattern = new RegExp(`${label}:\\s*([\\s\\S]*?)(?:\\n\\s*${otherLabel}:|$)`, "i")
const section = output.match(pattern)?.[1] ?? ""
if (!section || /Not configured/i.test(section)) return ""
return readField(section, "Model")
}
async function readOpenShellInferenceModel() {
const { stdout } = await execFileAsync(OPENSHELL_BIN, ["inference", "get"], {
env: hostCommandEnv({
OPENSHELL_GATEWAY: process.env.OPENSHELL_GATEWAY || "nemoclaw",
}),
timeout: 30000,
maxBuffer: 5 * 1024 * 1024,
})
const output = stripAnsi(String(stdout).trim())
return parseRouteModel(output, "Gateway inference") || parseRouteModel(output, "System inference")
}
export async function resolvePrimaryInferenceModel(explicitModel?: string | null) {
if (explicitModel?.trim()) return explicitModel.trim()
if (process.env.MCP_PREFLIGHT_LLM_MODEL?.trim()) return process.env.MCP_PREFLIGHT_LLM_MODEL.trim()
try {
const configured = await readOpenShellInferenceModel()
if (configured) return configured
} catch {
// Callers surface endpoint failures; model resolution falls through to explicit env compatibility.
}
return process.env.OPENAI_MODEL?.trim() || process.env.VLLM_MODEL?.trim() || ""
}
export function resolveOpenAiCompatibleBaseUrl() {
return (process.env.MCP_PREFLIGHT_LLM_BASE_URL || process.env.OPENAI_BASE_URL || process.env.VLLM_BASE_URL || "http://localhost:8000/v1").replace(/\/+$/, "")
}
+6 -14
View File
@@ -1,5 +1,6 @@
import { readdir, readFile, stat, writeFile } from "node:fs/promises"
import path from "node:path"
import { resolveOpenAiCompatibleBaseUrl, resolvePrimaryInferenceModel } from "./inferenceModel"
import { getSandboxInferenceConfig } from "./sandboxInferenceStore"
import type { McpPreflightResult } from "./mcpPreflight"
import type { McpServerInstall } from "./mcpServerStore"
@@ -46,12 +47,6 @@ const MAX_CONTEXT_FILE_BYTES = Number.parseInt(process.env.MCP_PREFLIGHT_REPAIR_
const MAX_CONTEXT_FILES = Number.parseInt(process.env.MCP_PREFLIGHT_REPAIR_MAX_FILES || "18", 10)
const MAX_REPAIR_FILE_BYTES = Number.parseInt(process.env.MCP_PREFLIGHT_REPAIR_MAX_OUTPUT_FILE_BYTES || String(512 * 1024), 10)
const REQUEST_TIMEOUT_MS = Number.parseInt(process.env.MCP_PREFLIGHT_REPAIR_TIMEOUT_MS || "60000", 10)
const DEFAULT_MODEL = process.env.MCP_PREFLIGHT_LLM_MODEL || process.env.OPENAI_MODEL || "gpt-4o-mini"
const DEFAULT_BASE_URL = process.env.MCP_PREFLIGHT_LLM_BASE_URL || process.env.OPENAI_BASE_URL || process.env.VLLM_BASE_URL || "http://localhost:8000/v1"
function cleanBaseUrl(value: string) {
return value.replace(/\/+$/, "")
}
function relativeInside(root: string, candidate: string) {
const fullPath = path.resolve(root, candidate)
@@ -110,20 +105,16 @@ async function readContextFiles(root: string) {
}
async function resolveRepairRoute(sandboxId?: string | null) {
let explicitModel = ""
if (sandboxId) {
const config = await getSandboxInferenceConfig(sandboxId)
const enabled = config.routes.filter((route) => route.enabled)
const primary = enabled.find((route) => route.id === config.primaryRouteId) || enabled[0]
if (primary?.model) {
return {
model: process.env.MCP_PREFLIGHT_LLM_MODEL || primary.model,
baseUrl: cleanBaseUrl(DEFAULT_BASE_URL),
}
}
explicitModel = primary?.model || ""
}
return {
model: DEFAULT_MODEL,
baseUrl: cleanBaseUrl(DEFAULT_BASE_URL),
model: await resolvePrimaryInferenceModel(explicitModel),
baseUrl: resolveOpenAiCompatibleBaseUrl(),
}
}
@@ -137,6 +128,7 @@ function extractJsonObject(value: string) {
async function requestRepairPlan(context: RepairContext, files: Awaited<ReturnType<typeof readContextFiles>>) {
const route = await resolveRepairRoute(context.sandboxId)
if (!route.model) throw new Error("No primary inference model is configured")
const controller = new AbortController()
const timeout = setTimeout(() => controller.abort(), REQUEST_TIMEOUT_MS)
try {