Normalize inference endpoint URLs

This commit is contained in:
mmckeen-nv
2026-04-29 13:01:59 -07:00
parent 10f5a0b6aa
commit e55dc7b9e0
5 changed files with 63 additions and 7 deletions
+7 -2
View File
@@ -1,6 +1,7 @@
import { NextResponse } from "next/server"
import { execFile } from "node:child_process"
import { promisify } from "node:util"
import { normalizeInferenceBaseUrlForGateway } from "@/app/lib/inferenceEndpointUrl"
import { OPENSHELL_BIN, hostCommandEnv } from "@/app/lib/hostCommands"
const execFileAsync = promisify(execFile)
@@ -144,6 +145,10 @@ export async function POST(request: Request) {
const model = optionalString(body?.model)
const baseUrl = optionalString(body?.baseUrl)
const type = normalizeProviderType(requestedType, name, baseUrl)
const normalizedBaseUrl = normalizeInferenceBaseUrlForGateway(
baseUrl,
type === "anthropic" ? "anthropic" : "openai",
)
const apiKey = optionalString(body?.apiKey)
const credentialKey = optionalString(body?.credentialKey) || "OPENAI_API_KEY"
const makeActive = body?.makeActive !== false
@@ -153,7 +158,7 @@ export async function POST(request: Request) {
if (makeActive && !model) throw new Error("model is required when making the endpoint active")
if (!/^[A-Z_][A-Z0-9_]*$/.test(credentialKey)) throw new Error("credential key must look like an environment variable name")
if (baseUrl) new URL(baseUrl)
if (normalizedBaseUrl) new URL(normalizedBaseUrl)
const existingNames = (await readProviders()).map((provider) => provider.name).filter(Boolean)
const providerArgs = existingNames.includes(name)
@@ -163,7 +168,7 @@ export async function POST(request: Request) {
const providerEnv = apiKey ? { [credentialKey]: apiKey } : {}
if (!existingNames.includes(name) && !apiKey) providerArgs.push("--from-existing")
if (apiKey) providerArgs.push("--credential", credentialKey)
if (baseUrl) providerArgs.push("--config", `OPENAI_BASE_URL=${baseUrl}`)
if (normalizedBaseUrl) providerArgs.push("--config", `OPENAI_BASE_URL=${normalizedBaseUrl}`)
const providerResult = await runOpenShell(providerArgs, providerEnv)
let inferenceResult = null
+3 -1
View File
@@ -4,6 +4,7 @@ import { existsSync } from "node:fs"
import { readFile, writeFile } from "node:fs/promises"
import path from "node:path"
import { promisify } from "node:util"
import { normalizeInferenceBaseUrlForGateway } from "@/app/lib/inferenceEndpointUrl"
import { inspectSandbox } from "@/app/lib/openshellHost"
import {
commandExists,
@@ -127,7 +128,8 @@ function applyConfiguredInferenceToOnboardEnv(env: NodeJS.ProcessEnv, route: Con
}
if (provider === "custom") {
const endpointUrl = process.env.NEMOCLAW_ENDPOINT_URL || process.env.OPENAI_BASE_URL || process.env.VLLM_BASE_URL || ""
if (endpointUrl) env.NEMOCLAW_ENDPOINT_URL = endpointUrl
const normalizedEndpointUrl = normalizeInferenceBaseUrlForGateway(endpointUrl, "openai")
if (normalizedEndpointUrl) env.NEMOCLAW_ENDPOINT_URL = normalizedEndpointUrl
const providerKey = process.env.NEMOCLAW_PROVIDER_KEY || process.env.OPENAI_API_KEY || process.env.VLLM_API_KEY || ""
if (providerKey) env.NEMOCLAW_PROVIDER_KEY = providerKey
}
+42
View File
@@ -0,0 +1,42 @@
export type InferenceEndpointFlavor = "anthropic" | "openai"
const HOST_REACHABLE_FROM_OPENSHELL = "host.docker.internal"
const LOCAL_GATEWAY_HOSTS = new Set(["localhost", "127.0.0.1", "0.0.0.0", "::1"])
function stripEndpointSuffix(pathname = "", suffixes: string[] = []) {
for (const suffix of suffixes) {
if (pathname === suffix) return ""
if (pathname.endsWith(suffix)) return pathname.slice(0, -suffix.length)
}
return pathname
}
function endpointSuffixes(flavor: InferenceEndpointFlavor) {
return flavor === "anthropic"
? ["/v1/messages", "/v1/models", "/v1", "/messages", "/models"]
: ["/responses", "/chat/completions", "/completions", "/models"]
}
function ensureScheme(value: string) {
return /^[a-z][a-z0-9+.-]*:\/\//i.test(value) ? value : `http://${value}`
}
export function normalizeInferenceBaseUrlForGateway(value: string, flavor: InferenceEndpointFlavor = "openai") {
const raw = String(value || "").trim()
if (!raw) return ""
try {
const url = new URL(ensureScheme(raw))
url.search = ""
url.hash = ""
if (LOCAL_GATEWAY_HOSTS.has(url.hostname.toLowerCase())) {
url.hostname = HOST_REACHABLE_FROM_OPENSHELL
}
let pathname = stripEndpointSuffix(url.pathname.replace(/\/+$/, ""), endpointSuffixes(flavor))
pathname = pathname.replace(/\/+$/, "")
url.pathname = pathname || "/"
return url.pathname === "/" ? url.origin : `${url.origin}${url.pathname}`
} catch {
return raw.replace(/[?#].*$/, "").replace(/\/+$/, "")
}
}
+3 -3
View File
@@ -7,11 +7,11 @@ const execFileAsync = promisify(execFile)
const DOCKER_BIN = process.env.DOCKER_BIN || "docker"
const OPENSHELL_CLUSTER_CONTAINER = process.env.OPENSHELL_CLUSTER_CONTAINER || "openshell-cluster-nemoclaw"
function modelEntry(route: SandboxInferenceRoute, openClawModelRef: string, compat: Record<string, unknown> | null) {
function modelEntry(route: SandboxInferenceRoute, compat: Record<string, unknown> | null) {
return {
...(compat ? { compat } : {}),
id: route.model,
name: openClawModelRef,
name: route.model,
reasoning: false,
input: ["text"],
cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
@@ -78,7 +78,7 @@ function buildOpenClawConfig(current: any, routes: SandboxInferenceRoute[], prim
api: resolved.api,
models: [],
}
providers[resolved.providerKey].models.push(modelEntry(route, resolved.modelRef, resolved.compat))
providers[resolved.providerKey].models.push(modelEntry(route, resolved.compat))
if (route.id === primary.id) primaryModelRef = resolved.modelRef
}
+8 -1
View File
@@ -8,6 +8,7 @@ const deleteRoutePath = path.join(root, 'app/api/sandbox/delete/route.ts')
const hostCommandsPath = path.join(root, 'app/lib/hostCommands.ts')
const telemetryRoutePath = path.join(root, 'app/api/telemetry/real/route.ts')
const inferenceRoutePath = path.join(root, 'app/api/inference/route.ts')
const inferenceEndpointUrlPath = path.join(root, 'app/lib/inferenceEndpointUrl.ts')
const ollamaModelsRoutePath = path.join(root, 'app/api/ollama/models/route.ts')
const sandboxInferenceRoutePath = path.join(root, 'app/api/sandbox/[sandboxId]/inference/route.ts')
const sandboxInferenceApplyRoutePath = path.join(root, 'app/api/sandbox/[sandboxId]/inference/apply/route.ts')
@@ -28,6 +29,7 @@ const deleteRouteSource = await readFile(deleteRoutePath, 'utf8')
const hostCommandsSource = await readFile(hostCommandsPath, 'utf8')
const telemetryRouteSource = await readFile(telemetryRoutePath, 'utf8')
const inferenceRouteSource = await readFile(inferenceRoutePath, 'utf8')
const inferenceEndpointUrlSource = await readFile(inferenceEndpointUrlPath, 'utf8')
const ollamaModelsRouteSource = await readFile(ollamaModelsRoutePath, 'utf8')
const sandboxInferenceRouteSource = await readFile(sandboxInferenceRoutePath, 'utf8')
const sandboxInferenceApplyRouteSource = await readFile(sandboxInferenceApplyRoutePath, 'utf8')
@@ -91,7 +93,11 @@ assert.match(inferenceRouteSource, /\["provider", "delete", name\]/, 'inference
assert.match(inferenceRouteSource, /\["inference", "set", "--provider", name, "--model", model\]/, 'inference route must set the active OpenShell inference route')
assert.match(inferenceRouteSource, /isVllmEndpoint\(name, baseUrl\)/, 'inference route must skip verification for vLLM-compatible endpoints')
assert.match(inferenceRouteSource, /normalizeProviderType\(requestedType, name, baseUrl\)/, 'inference route must force vLLM endpoints through the OpenAI-compatible provider type')
assert.match(inferenceRouteSource, /OPENAI_BASE_URL=\$\{baseUrl\}/, 'inference route must pass endpoint URL as provider config')
assert.match(inferenceRouteSource, /normalizeInferenceBaseUrlForGateway\(\s*baseUrl,/, 'inference route must normalize endpoint URLs for the OpenShell gateway pod')
assert.match(inferenceRouteSource, /OPENAI_BASE_URL=\$\{normalizedBaseUrl\}/, 'inference route must pass the normalized endpoint URL as provider config')
assert.match(createRouteSource, /normalizeInferenceBaseUrlForGateway\(endpointUrl, "openai"\)/, 'sandbox create must normalize env-derived OpenAI-compatible endpoints before NemoClaw onboarding')
assert.match(inferenceEndpointUrlSource, /host\.docker\.internal/, 'endpoint normalization must rewrite host-local URLs to a gateway-reachable host alias')
assert.match(inferenceEndpointUrlSource, /\/chat\/completions/, 'endpoint normalization must strip concrete OpenAI endpoint suffixes to the base URL')
assert.match(inferencePanelSource, /INFERENCE ENDPOINTS/, 'settings UI must expose inference endpoint configuration')
assert.doesNotMatch(inferencePanelSource, /NVIDIA Preset|useNvidiaInferencePreset|aws\/anthropic\/claude-opus-4-5/, 'settings UI must not expose NVIDIA-corp-specific inference presets')
assert.match(inferencePanelSource, /useVllmPreset/, 'settings UI must expose a vLLM endpoint preset')
@@ -119,6 +125,7 @@ assert.match(sandboxInferencePanelSource, /Add Route/, 'selected sandbox UI must
assert.match(sandboxInferenceApplyLibSource, /\/sandbox\/\.openclaw\/openclaw\.json/, 'live apply must patch OpenClaw config inside the running sandbox')
assert.match(sandboxInferenceApplyLibSource, /sha256sum \/sandbox\/\.openclaw\/openclaw\.json > \/sandbox\/\.openclaw\/\.config-hash/, 'live apply must refresh the immutable OpenClaw config hash')
assert.match(sandboxInferenceApplyLibSource, /\["inference", "set", "--no-verify", "--provider", primary\.provider, "--model", primary\.model\]/, 'live apply must point OpenShell inference at the primary route')
assert.match(sandboxInferenceApplyLibSource, /name: route\.model/, 'live apply must keep model display names raw while primary refs remain provider-qualified')
assert.doesNotMatch(sandboxRestartRouteSource, /delete", "pod"|kubectl delete/, 'restart route must not delete the backing sandbox pod')
assert.match(sandboxRestartRouteSource, /restartMode: "openclaw-runtime"/, 'restart route must restart only the in-sandbox OpenClaw runtime')
assert.match(sandboxRestartRouteSource, /waitForSandboxReady\(sandboxName, 15000, 1000\)/, 'restart route must verify OpenShell reports the sandbox ready before runtime restart')