refactor(gateway): extract FIM routing

This commit is contained in:
Mark IJbema
2026-05-25 17:13:01 +02:00
parent d635a15809
commit 2bf49bddb6
3 changed files with 143 additions and 125 deletions
+140
View File
@@ -0,0 +1,140 @@
import { HEADER_FEATURE, KILO_API_BASE } from "../api/constants.js"
import { buildKiloHeaders } from "../headers.js"
type Auth = any
type FimProvider = "kilo" | "mistral" | "inception"
interface FimTarget {
provider: FimProvider
model: string
urls: string[]
}
const FIM_TIMEOUT_MS = 30_000
const KILO_FIM_URL = KILO_API_BASE + "/api/fim/completions"
const MISTRAL_FIM_URL = "https://api.mistral.ai/v1/fim/completions"
const CODESTRAL_FIM_URL = "https://codestral.mistral.ai/v1/fim/completions"
const INCEPTION_FIM_URL = "https://api.inceptionlabs.ai/v1/fim/completions"
export function resolveFimTarget(model?: string): FimTarget {
if (model === "mistral/codestral-2508") {
return { provider: "mistral", model: "codestral-2508", urls: [MISTRAL_FIM_URL, CODESTRAL_FIM_URL] }
}
if (model === "inception-direct/mercury-edit-2") {
return { provider: "inception", model: "mercury-edit-2", urls: [INCEPTION_FIM_URL] }
}
return { provider: "kilo", model: model ?? "mistralai/codestral-2501", urls: [KILO_FIM_URL] }
}
async function getProxyAuth(Auth: Auth) {
const auth = await Auth.get("kilo")
const token = auth?.type === "api" ? auth.key : auth?.type === "oauth" ? auth.access : undefined
return {
auth,
token,
organizationId: auth?.type === "oauth" ? auth.accountId : undefined,
}
}
async function getProviderKey(Auth: Auth, provider: FimProvider) {
const auth = await Auth.get(provider)
return auth?.type === "api" ? auth.key : undefined
}
async function fetchFim(
target: FimTarget,
url: string,
fallbacks: string[],
key: string,
input: {
prefix: string
suffix: string
maxTokens: number
temperature: number
signal: AbortSignal
organizationId?: string
},
): Promise<Response> {
console.info(`[FIM] request provider=${target.provider} model=${target.model} url=${url}`)
const response = await fetch(url, {
method: "POST",
headers: {
"Content-Type": "application/json",
Authorization: `Bearer ${key}`,
...(target.provider === "kilo" ? buildKiloHeaders(undefined, { kilocodeOrganizationId: input.organizationId }) : {}),
...(target.provider === "kilo" ? { [HEADER_FEATURE]: "autocomplete" } : {}),
},
signal: input.signal,
body: JSON.stringify({
model: target.model,
prompt: input.prefix,
suffix: input.suffix,
max_tokens: input.maxTokens,
temperature: input.temperature,
stream: true,
}),
})
const [next] = fallbacks
if (response.status === 401 && next) return fetchFim(target, next, fallbacks.slice(1), key, input)
return response
}
export function createFimHandler(Auth: Auth) {
return async (c: any) => {
const { prefix, suffix, model, maxTokens, temperature } = c.req.valid("json")
const target = resolveFimTarget(model)
const fimMaxTokens = maxTokens ?? 256
const fimTemperature = temperature ?? 0.2
const proxy = target.provider === "kilo" ? await getProxyAuth(Auth) : undefined
const token = target.provider === "kilo" ? proxy?.token : await getProviderKey(Auth, target.provider)
if (target.provider === "kilo" && !proxy?.auth) {
return c.json({ error: "Not authenticated with Kilo Gateway" }, 401)
}
if (target.provider === "kilo" && !token) {
return c.json({ error: "No valid token found" }, 401)
}
if (!token) {
return c.json({ error: `Missing ${target.provider} provider API key` }, 401)
}
const signal = AbortSignal.any([c.req.raw.signal, AbortSignal.timeout(FIM_TIMEOUT_MS)])
let response: Response
try {
const [url] = target.urls
if (!url) return c.json({ error: "No FIM endpoint configured" }, 500 as any)
response = await fetchFim(target, url, target.urls.slice(1), token, {
prefix,
suffix,
maxTokens: fimMaxTokens,
temperature: fimTemperature,
signal,
organizationId: proxy?.organizationId,
})
} catch (err) {
if (err instanceof DOMException && err.name === "TimeoutError") {
return c.json({ error: "FIM request timed out" }, 504 as any)
}
if (signal.aborted) return c.json({ error: "FIM request canceled" }, 499 as any)
throw err
}
if (!response.ok) {
const text = await response.text()
return c.json({ error: `FIM request failed: ${response.status} ${text}` }, response.status as any)
}
return new Response(response.body, {
headers: {
"Content-Type": "text/event-stream",
"Cache-Control": "no-cache",
Connection: "keep-alive",
},
})
}
}
+2 -124
View File
@@ -18,6 +18,7 @@ import {
import { buildKiloHeaders } from "../headers.js"
import type { ImportDeps, DrizzleDb } from "../cloud-sessions.js"
import { fetchCloudSession, fetchCloudSessionForImport, importSessionToDb } from "../cloud-sessions.js"
import { createFimHandler } from "./fim.js"
// Type definitions for OpenCode dependencies (injected at runtime)
type Hono = any
@@ -29,14 +30,6 @@ type Auth = any
type ModelCache = { clear: (providerID: string) => void }
type Z = any
type FimProvider = "kilo" | "mistral" | "inception"
interface FimTarget {
provider: FimProvider
model: string
urls: string[]
}
interface KiloRoutesDeps extends ImportDeps {
Hono: new () => Hono
describeRoute: DescribeRoute
@@ -49,22 +42,6 @@ interface KiloRoutesDeps extends ImportDeps {
InstanceStore: { disposeAllInstances(): Promise<void> }
}
const FIM_TIMEOUT_MS = 30_000
const KILO_FIM_URL = KILO_API_BASE + "/api/fim/completions"
const MISTRAL_FIM_URL = "https://api.mistral.ai/v1/fim/completions"
const CODESTRAL_FIM_URL = "https://codestral.mistral.ai/v1/fim/completions"
const INCEPTION_FIM_URL = "https://api.inceptionlabs.ai/v1/fim/completions"
export function resolveFimTarget(model?: string): FimTarget {
if (model === "mistral/codestral-2508") {
return { provider: "mistral", model: "codestral-2508", urls: [MISTRAL_FIM_URL, CODESTRAL_FIM_URL] }
}
if (model === "inception-direct/mercury-edit-2") {
return { provider: "inception", model: "mercury-edit-2", urls: [INCEPTION_FIM_URL] }
}
return { provider: "kilo", model: model ?? "mistralai/codestral-2501", urls: [KILO_FIM_URL] }
}
/**
* Create Kilo Gateway routes with OpenCode dependencies injected
*
@@ -169,52 +146,6 @@ export function createKiloRoutes(deps: KiloRoutesDeps) {
}
}
const getProviderKey = async (provider: FimProvider) => {
const auth = await Auth.get(provider)
return auth?.type === "api" ? auth.key : undefined
}
const fetchFim = async (
target: FimTarget,
url: string,
fallbacks: string[],
key: string,
input: {
prefix: string
suffix: string
maxTokens: number
temperature: number
signal: AbortSignal
organizationId?: string
},
): Promise<Response> => {
console.info(`[FIM] request provider=${target.provider} model=${target.model} url=${url}`)
const response = await fetch(url, {
method: "POST",
headers: {
"Content-Type": "application/json",
Authorization: `Bearer ${key}`,
...(target.provider === "kilo"
? buildKiloHeaders(undefined, { kilocodeOrganizationId: input.organizationId })
: {}),
...(target.provider === "kilo" ? { [HEADER_FEATURE]: "autocomplete" } : {}),
},
signal: input.signal,
body: JSON.stringify({
model: target.model,
prompt: input.prefix,
suffix: input.suffix,
max_tokens: input.maxTokens,
temperature: input.temperature,
stream: true,
}),
})
const [next] = fallbacks
if (response.status === 401 && next) return fetchFim(target, next, fallbacks.slice(1), key, input)
return response
}
return new Hono()
.get(
"/profile",
@@ -406,60 +337,7 @@ export function createKiloRoutes(deps: KiloRoutesDeps) {
temperature: z.number().optional(),
}),
),
async (c: any) => {
const { prefix, suffix, model, maxTokens, temperature } = c.req.valid("json")
const target = resolveFimTarget(model)
const fimMaxTokens = maxTokens ?? 256
const fimTemperature = temperature ?? 0.2
const proxy = target.provider === "kilo" ? await getProxyAuth() : undefined
const token = target.provider === "kilo" ? proxy?.token : await getProviderKey(target.provider)
if (target.provider === "kilo" && !proxy?.auth) {
return c.json({ error: "Not authenticated with Kilo Gateway" }, 401)
}
if (target.provider === "kilo" && !token) {
return c.json({ error: "No valid token found" }, 401)
}
if (!token) {
return c.json({ error: `Missing ${target.provider} provider API key` }, 401)
}
const signal = AbortSignal.any([c.req.raw.signal, AbortSignal.timeout(FIM_TIMEOUT_MS)])
let response: Response
try {
const [url] = target.urls
if (!url) return c.json({ error: "No FIM endpoint configured" }, 500 as any)
response = await fetchFim(target, url, target.urls.slice(1), token, {
prefix,
suffix,
maxTokens: fimMaxTokens,
temperature: fimTemperature,
signal,
organizationId: proxy?.organizationId,
})
} catch (err) {
if (err instanceof DOMException && err.name === "TimeoutError")
return c.json({ error: "FIM request timed out" }, 504 as any)
if (signal.aborted) return c.json({ error: "FIM request canceled" }, 499 as any)
throw err
}
if (!response.ok) {
const text = await response.text()
return c.json({ error: `FIM request failed: ${response.status} ${text}` }, response.status as any)
}
return new Response(response.body, {
headers: {
"Content-Type": "text/event-stream",
"Cache-Control": "no-cache",
Connection: "keep-alive",
},
})
},
createFimHandler(Auth),
)
.post(
"/audio/transcriptions",
+1 -1
View File
@@ -1,5 +1,5 @@
import { describe, expect, test } from "bun:test"
import { resolveFimTarget } from "../src/server/routes"
import { resolveFimTarget } from "../src/server/fim"
describe("FIM target resolution", () => {
test("keeps gateway autocomplete models on Kilo Gateway", () => {