smashimo/oh-my-pi
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge remote-tracking branch 'origin/farm/8b3cc0d0/increase-openai-stream-timeout'

Browse Source
This commit is contained in:
can1357
2026-07-01 02:57:28 +02:00
parent 6e9ae7adc7 20fe5029ac
commit f2f62f9b8a
3 changed files with 57 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files
packages/ai/test/openai-completions-progress-chunk.test.ts
+36
View File
@@ -11,6 +11,7 @@ const openAICompletionsModel = {
...(getBundledModel("openai", "gpt-4o-mini") as Model<"openai-completions">),
api: "openai-completions",
} satisfies Model<"openai-completions">;
const openAIResponsesModel = getBundledModel("openai", "gpt-5-mini") as Model<"openai-responses">;
function baseContext(): Context {
return {
@@ -190,6 +191,41 @@ describe("resolveOpenAICompat stream idle timeout", () => {
expect(openAICompletionsModel.compat.streamIdleTimeoutMs).toBeUndefined();
});
it("widens local OpenAI-compatible stream watchdogs", () => {
const completions = buildModel({
...openAICompletionsModel,
id: "qwen3-local",
name: "Qwen3 Local",
provider: "llama.cpp",
baseUrl: "http://localhost:8080/v1",
compat: openAICompletionsModel.compatConfig,
} as ModelSpec<"openai-completions">);
const responses = buildModel({
...openAIResponsesModel,
id: "qwen3-local",
name: "Qwen3 Local",
provider: "llama.cpp",
baseUrl: "http://localhost:8080/v1",
compat: openAIResponsesModel.compatConfig,
} as ModelSpec<"openai-responses">);
expect(completions.compat.streamIdleTimeoutMs).toBe(300_000);
expect(responses.compat.streamIdleTimeoutMs).toBe(300_000);
});
it("widens custom loopback OpenAI-compatible responses stream watchdogs", () => {
const model = buildModel({
...openAIResponsesModel,
id: "local-model",
name: "Local Model",
provider: "custom-local",
baseUrl: "http://127.0.0.1:8080/v1",
compat: openAIResponsesModel.compatConfig,
} as ModelSpec<"openai-responses">);
expect(model.compat.streamIdleTimeoutMs).toBe(300_000);
});
it("widens Xiaomi MiMo Pro stream watchdog (issue #1770)", () => {
const model = buildModel({
...openAICompletionsModel,
packages/catalog/CHANGELOG.md
+1
View File
@@ -5,6 +5,7 @@
### Fixed
- Made discovery request timeouts cancellable in the OpenAI-compatible resolvers. `fetchOpenAICompatibleModels` and `fetchLiteLLMRichModels` (used by the vLLM, LiteLLM, and LM Studio catalogs) previously received or armed an uncancellable `AbortSignal.timeout(...)`, leaving a pending timer on the heap after the request settled. They now accept a `timeoutMs` option and arm a cancellable `AbortController` whose backing timer is cleared the instant the fetch + body read settle, covering the whole operation and never leaving a signal armed on the heap.
- Widened stream watchdogs for local OpenAI-compatible backends (llama.cpp, LM Studio, vLLM, Ollama, and loopback custom providers) so cold model loads do not trip first-event timeouts. ([#3940](https://github.com/can1357/oh-my-pi/issues/3940))
## [16.2.10] - 2026-06-30
packages/catalog/src/compat/openai.ts
+20 -11
View File
@@ -57,6 +57,8 @@ function matchesKimiK27CodeFamily(spec: ModelSpec<"openai-completions">): boolea
const XIAOMI_MIMO_STREAM_IDLE_TIMEOUT_MS = 300_000;
/** Alibaba Coding Plan (coding-intl.dashscope) qwen models idle before the first event (issue #1770). */
const ALIBABA_CODING_PLAN_STREAM_IDLE_TIMEOUT_MS = 600_000;
/** Local OpenAI-compatible backends can spend minutes cold-loading a model before the first SSE event. */
const LOCAL_OPENAI_COMPAT_STREAM_IDLE_TIMEOUT_MS = 300_000;
const MINIMAX_PROVIDER_OR_ID_PATTERN = /minimax/i;
const DSML_HEALING_PROVIDERS = new Set([
"ollama",
@@ -294,6 +296,9 @@ export function buildOpenAICompat(spec: ModelSpec<"openai-completions">): Resolv
isMoonshotNative ||
isOpenCodeHost;
const isOpenCodeProvider = provider === "opencode-go" || provider === "opencode-zen";
const isLocalOpenAICompatBackend =
!PROXY_OPENAI_COMPAT_PROVIDERS.has(provider) &&
(LOCAL_OPENAI_COMPAT_PROVIDERS.has(provider) || hasLocalLoopbackBaseUrl(baseUrl));
const useMaxTokens =
isMistral ||
@@ -348,9 +353,10 @@ export function buildOpenAICompat(spec: ModelSpec<"openai-completions">): Resolv
isCopilotHost ||
isZenmuxHost);
// Stream-watchdog floor: GLM coding-plan SKUs, Kimi K2.6, and direct
// DeepSeek reasoning models can idle for minutes while reasoning; widen the
// idle timeout so warm-ups stop aborting and retrying.
// Stream-watchdog floor: GLM coding-plan SKUs, Kimi K2.6, direct
// DeepSeek reasoning models, and local OpenAI-compatible backends can idle
// for minutes while reasoning or cold-loading weights; widen the idle
// timeout so warm-ups stop aborting and retrying.
const streamIdleTimeoutMs =
GLM_CODING_PLAN_MODEL_PATTERN.test(spec.id) && (isZai || isZhipu)
? GLM_CODING_PLAN_STREAM_IDLE_TIMEOUT_MS
@@ -362,7 +368,9 @@ export function buildOpenAICompat(spec: ModelSpec<"openai-completions">): Resolv
? KIMI_K26_REASONING_STREAM_IDLE_TIMEOUT_MS
: spec.reasoning && isDirectDeepseekApi
? DEEPSEEK_REASONING_STREAM_IDLE_TIMEOUT_MS
: undefined;
: isLocalOpenAICompatBackend
? LOCAL_OPENAI_COMPAT_STREAM_IDLE_TIMEOUT_MS
: undefined;
// Fireworks "Fast" variants (`<id>-fast`) are served from the router
// namespace (`accounts/fireworks/routers/<id>-fast`), like Fire Pass, rather
@@ -482,9 +490,7 @@ export function buildOpenAICompat(spec: ModelSpec<"openai-completions">): Resolv
// when a thinking block actually exists on the turn
// (`nonEmptyThinkingBlocks.length > 0`), so the flag is a no-op on
// pure-text histories.
replayReasoningContent:
!PROXY_OPENAI_COMPAT_PROVIDERS.has(provider) &&
(LOCAL_OPENAI_COMPAT_PROVIDERS.has(provider) || hasLocalLoopbackBaseUrl(baseUrl)),
replayReasoningContent: isLocalOpenAICompatBackend,
// `preserve_thinking: true` makes the Qwen3.6+ chat template render
// `<think>...</think>` for older assistant turns too, instead of
// stripping it the moment a new user message moves them past
@@ -496,9 +502,7 @@ export function buildOpenAICompat(spec: ModelSpec<"openai-completions">): Resolv
// with `replayReasoningContent` above). Non-Qwen templates ignore the
// parameter, so the flag stays a no-op outside the Qwen path.
qwenPreserveThinking:
(thinkingFormat === "qwen" || thinkingFormat === "qwen-chat-template") &&
!PROXY_OPENAI_COMPAT_PROVIDERS.has(provider) &&
(LOCAL_OPENAI_COMPAT_PROVIDERS.has(provider) || hasLocalLoopbackBaseUrl(baseUrl)),
(thinkingFormat === "qwen" || thinkingFormat === "qwen-chat-template") && isLocalOpenAICompatBackend,
requiresAssistantContentForToolCalls: isKimiModel || isDirectDeepseekReasoning,
cacheControlFormat: isOpenRouter && spec.id.startsWith("anthropic/") ? "anthropic" : undefined,
openRouterRouting: undefined,
@@ -584,6 +588,9 @@ export function buildOpenAIResponsesCompat(spec: OpenAIResponsesSpecLike): Resol
const isAnthropicModel = id ? isClaudeModelId(id) || isAnthropicNamespacedModelId(id) : false;
const isDeepseekFamily = id ? isDeepseekModelIdOrName(id) || isDeepseekModelIdOrName(spec.name) : false;
const reasoningCapable = Boolean(spec.reasoning);
const isLocalOpenAICompatBackend =
!PROXY_OPENAI_COMPAT_PROVIDERS.has(spec.provider) &&
(LOCAL_OPENAI_COMPAT_PROVIDERS.has(spec.provider) || hasLocalLoopbackBaseUrl(baseUrl));
const compat: ResolvedOpenAIResponsesCompat = {
supportsDeveloperRole: isAzure || isOpenAIUrl || hostMatchesUrl(baseUrl, "githubCopilot"),
@@ -645,7 +652,9 @@ export function buildOpenAIResponsesCompat(spec: OpenAIResponsesSpecLike): Resol
emptyLengthFinishIsContextError: spec.provider === "ollama",
usesOpenAIToolCallIdLimit: spec.provider === "openai",
promptCacheSessionHeader: spec.provider === "xai-oauth" ? "x-grok-conv-id" : undefined,
streamIdleTimeoutMs: spec.compat?.streamIdleTimeoutMs,
streamIdleTimeoutMs: isLocalOpenAICompatBackend
? LOCAL_OPENAI_COMPAT_STREAM_IDLE_TIMEOUT_MS
: spec.compat?.streamIdleTimeoutMs,
};
applyCompatOverrides(compat, spec.compat);
if (spec.compat?.reasoningDisableMode === undefined) {