Skip to content

fix: route muse-spark to /v1/responses (Zen requires Responses API) #2

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
ConTresillo wants to merge 1 commit into
base: main
Choose a base branch
from
Open

fix: route muse-spark to /v1/responses (Zen requires Responses API) #2

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
107 changes: 100 additions & 7 deletions zen-proxy.mjs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -162,6 +162,41 @@ function retryableUpstream(status, body) {
return /model is unavailable|not supported|only be used in opencode|no such model|does not exist|overloaded|temporarily.*limit|upstream request failed/i.test(msg)
}

function isResponsesModel(id) {
return /muse-spark/i.test(String(id))
}

function chatMessagesToInput(messages) {
if (!Array.isArray(messages)) return String(messages ?? "")
return messages
.map((m) => {
const role = m?.role ?? "user"
const c = m?.content
const text = typeof c === "string" ? c : Array.isArray(c) ? c.map((p) => p?.text ?? p?.content ?? "").join("\n") : String(c ?? "")
if (role === "system") return `[SYSTEM] ${text}`
if (role === "user") return text
if (role === "assistant") return text
return `${role}: ${text}`
})
.join("\n\n")
}

function responsesOutputToText(data) {
if (!data || typeof data !== "object") return ""
if (typeof data.output_text === "string" && data.output_text) return data.output_text
const out = data.output
if (Array.isArray(out)) {
const texts = []
for (const item of out) {
if (item?.type === "message" && Array.isArray(item.content)) {
for (const c of item.content) if (c?.type === "output_text" && c.text) texts.push(c.text)
} else if (item?.type === "text" && item.text) texts.push(item.text)
}
if (texts.length) return texts.join("\n\n")
}
return ""
}

function toBool(v, dflt) {
if (v === undefined || v === null) return dflt
if (v === false || v === 0 || v === "0" || v === "false") return false
Expand Down Expand Up @@ -320,10 +355,26 @@ async function handleChat(req, res) {
let used = requested
for (const model of candidates) {
used = model
const payload = { ...body, model }
const useResponses = isResponsesModel(model)
const endpoint = useResponses ? "/responses" : "/chat/completions"
let payload
if (useResponses) {
const input = chatMessagesToInput(body.messages)
payload = { model, input, stream: !!body.stream }
if (body.temperature != null) payload.temperature = body.temperature
if (body.top_p != null) payload.top_p = body.top_p
// Muse Spark is reasoning-heavy (~900 reasoning tokens for a simple prompt).
// Forwarding a tiny max_tokens/max_output_tokens (e.g. 5) truncates before any
// visible text is produced (upstream returns reasoning only + empty output_text).
// Only forward when the caller asks for a reasonably large budget.
const mt = body.max_completion_tokens ?? body.max_tokens
if (mt != null && Number(mt) >= 256) payload.max_output_tokens = Number(mt)
} else {
payload = { ...body, model }
}
let upstreamRes
try {
upstreamRes = await fetch(`${config.upstream}/chat/completions`, {
upstreamRes = await fetch(`${config.upstream}${endpoint}`, {
method: "POST",
headers: zenHeaders(req, auth),
body: JSON.stringify(payload),
Expand All @@ -338,12 +389,45 @@ async function handleChat(req, res) {
if (upstreamRes.ok) {
const contentType = upstreamRes.headers.get("content-type") ?? "application/json"
if (isStream) {
// For responses API streaming the SSE shape differs; for now treat as non-stream fallback
if (useResponses) {
try {
const data = await upstreamRes.json()
const text = responsesOutputToText(data)
const chatResp = {
id: data.id ?? `chatcmpl-${Date.now()}`,
object: "chat.completion",
created: Math.floor(Date.now() / 1000),
model: requested,
choices: [{ index: 0, message: { role: "assistant", content: text }, finish_reason: "stop" }],
usage: data.usage ?? { prompt_tokens: 0, completion_tokens: 0, total_tokens: 0 },
}
recordReq(req, `${requested}→${model}`, Date.now() - start, 200)
return json(res, 200, chatResp)
} catch {
recordReq(req, requested, Date.now() - start, 502)
return json(res, 502, { error: { type: "upstream_error", message: "bad upstream response" } })
}
}
relayStream(req, res, upstreamRes, requested)
res.on("finish", () => recordReq(req, `${requested}→${model}`, Date.now() - start, 200))
return
}
try {
const content = await upstreamRes.json()
if (useResponses) {
const text = responsesOutputToText(content)
const chatResp = {
id: content.id ?? `chatcmpl-${Date.now()}`,
object: "chat.completion",
created: Math.floor(Date.now() / 1000),
model: requested,
choices: [{ index: 0, message: { role: "assistant", content: text }, finish_reason: "stop" }],
usage: content.usage ?? { prompt_tokens: 0, completion_tokens: 0, total_tokens: 0 },
}
recordReq(req, `${requested}→${model}`, Date.now() - start, 200)
return json(res, 200, chatResp)
}
if (content && typeof content === "object") content.model = requested
recordReq(req, `${requested}→${model}`, Date.now() - start, 200)
return json(res, 200, content)
Expand Down Expand Up @@ -515,7 +599,12 @@ async function syncModels() {
while (idx < candidates.length) {
const id = candidates[idx++]
try {
const r = await fetch(`${config.upstream}/chat/completions`, {
const useResponses = isResponsesModel(id)
const endpoint = useResponses ? "/responses" : "/chat/completions"
const probeBody = useResponses
? { model: id, input: "ping" }
: { model: id, messages: [{ role: "user", content: "ping" }], max_tokens: 5 }
const r = await fetch(`${config.upstream}${endpoint}`, {
method: "POST",
headers: (() => {
const h = {
Expand All @@ -528,7 +617,7 @@ async function syncModels() {
if (sess) h["x-opencode-session"] = sess
return h
})(),
body: JSON.stringify({ model: id, messages: [{ role: "user", content: "ping" }], max_tokens: 5 }),
body: JSON.stringify(probeBody),
signal: AbortSignal.timeout(Math.min(config.timeoutMs, 30_000)),
})
let bodyErr = ""
Expand Down Expand Up @@ -756,7 +845,10 @@ async function handleTest(req, res) {
typeof body.zenKey === "string" && body.zenKey.trim()
? `Bearer ${body.zenKey.trim()}`
: (authForUpstream(req) ?? "Bearer public")
const upstreamRes = await fetch(`${config.upstream}/chat/completions`, {
const useResponses = isResponsesModel(model)
const endpoint = useResponses ? "/responses" : "/chat/completions"
const probeBody = useResponses ? { model, input: "ping" } : { model, messages: [{ role: "user", content: "ping" }], max_tokens: 5 }
const upstreamRes = await fetch(`${config.upstream}${endpoint}`, {
method: "POST",
headers: (() => {
const h = {
Expand All @@ -771,13 +863,14 @@ async function handleTest(req, res) {
if (!ipOmit(ip)) h["x-real-ip"] = ip
return h
})(),
body: JSON.stringify({ model, messages: [{ role: "user", content: "ping" }], max_tokens: 5 }),
body: JSON.stringify(probeBody),
signal: AbortSignal.timeout(config.timeoutMs),
})
let detail = ""
try {
const parsed = await upstreamRes.json()
detail = parsed.error?.message ?? parsed.choices?.[0]?.message?.content ?? ""
if (useResponses) detail = responsesOutputToText(parsed) || parsed.error?.message || ""
else detail = parsed.error?.message ?? parsed.choices?.[0]?.message?.content ?? ""
} catch {}
json(res, 200, { ok: upstreamRes.ok, model, status: upstreamRes.status, ms: Date.now() - start, detail })
} catch (err) {
Expand Down