mirror of
https://gitcode.com/JianFeeeee/ModelRouter.git
synced 2026-10-03 23:54:06 +00:00
生产故障:pi 客户端报 `model "AUTO" returned a completed response with no content`, 重试延迟 8 秒。实测 AUTO 20 次有 2 次返回空 content,全部是 claude-opus-4-8。 根因(直连上游抓包确认):思考型模型先吐 reasoning_content,max_tokens 小到 思考阶段就把预算用完时,上游返回 200 / finish_reason=length,28 个 chunk 全是 reasoning_content、content 一片空白。runTier 只看 err == nil 就当成功返回, 客户端拿到一个空响应。 修复: - resultIsEmpty:非流式路径把「无 content、无 tool_calls、无 image」的响应当作 slot 失败继续降级。注意 ReasoningContent 不算内容——客户端要的是文本,为 另一个模型的思考阶段扣住请求比降级更糟。 - peekStream:流式路径在出现首个真实内容前缓冲 reasoning 前导,流结束仍无内容 则回报空结果,让 chainDrive 换 slot。缓冲只覆盖思考前导,拿到内容后立即 转发。tool_calls delta 算内容,agent 回合不会被误判。 - 3 条判据 + 3 个变异(恒 false / 恒 true / reasoning 算内容)全部被捕获。 同时修三个 WebUI 布局缺陷(都靠截图而非 DOM 断言发现): - 插件侧栏项只渲染图标没有标题:btn.innerHTML 只塞 pluginIconHTML(pg.icon), 与原生页的「图标 + <span>标题</span>」不一致,侧栏是一排无名图标。 - 计费维度表 8 列挤在 465px 卡片里:table-layout:fixed 把每列压到 62px, 23/80 个单元格溢出、数字互相重叠。改为 6 列(token 细分合并为 「输入(新鲜+缓存)」,细分进 title)+ table-layout:auto,实测 0/60 溢出。 - 数字列 word-break:break-all 让每个字符独占一行(USD 0.56 竖排成 U/S/D), 改 nowrap + 容器横向滚动。
128 lines
4.4 KiB
Go
128 lines
4.4 KiB
Go
package scheduler
|
|
|
|
import (
|
|
"context"
|
|
"encoding/json"
|
|
"strings"
|
|
"testing"
|
|
"time"
|
|
|
|
"llmsproxy/internal/types"
|
|
)
|
|
|
|
// TestResultIsEmptyMatchesTheReasoningOnlyResponse guards the production bug
|
|
// behind "model AUTO returned a completed response with no content".
|
|
//
|
|
// Production measurement: 2 of 20 AUTO requests returned zero content, both
|
|
// served by claude-opus-4-8. Direct upstream capture showed 28 chunks of
|
|
// `reasoning_content` and NO `content`, finish_reason=length — the token budget
|
|
// was consumed by thinking before any text was emitted. runTier accepted it
|
|
// because err == nil.
|
|
func TestResultIsEmptyMatchesTheReasoningOnlyResponse(t *testing.T) {
|
|
// Exactly what the upstream returned, as a decoded response.
|
|
onlyReasoning := &types.UnifiedResponse{
|
|
Model: "claude-opus-4-8",
|
|
ReasoningContent: "Alright, the user just said \"hi\". Simple greeting…",
|
|
FinishReason: "length",
|
|
TokenUsage: types.TokenUsage{Prompt: 14, Completion: 30, Total: 44},
|
|
}
|
|
if !resultIsEmpty(onlyReasoning) {
|
|
t.Error("a reasoning-only response with usage must count as empty; " +
|
|
"it is what the client sees as \"completed with no content\"")
|
|
}
|
|
|
|
// The same response WITH text is a normal answer.
|
|
if resultIsEmpty(&types.UnifiedResponse{Content: "Hi! 👋", ReasoningContent: "hmm"}) {
|
|
t.Error("a response with content is not empty")
|
|
}
|
|
// A tool-calling agent turn has no text and is perfectly valid.
|
|
if resultIsEmpty(&types.UnifiedResponse{
|
|
ToolCalls: []types.ToolCall{{Name: "read_file", Arguments: map[string]interface{}{"path": "x"}}},
|
|
}) {
|
|
t.Error("a tool-call-only response must NOT count as empty — agents " +
|
|
"legitimately produce tool calls with no text")
|
|
}
|
|
// An image slot returns no text by design.
|
|
if resultIsEmpty(&types.UnifiedResponse{
|
|
ImageData: []types.ImageData{{URL: "http://x/y.png"}},
|
|
}) {
|
|
t.Error("an image response must NOT count as empty")
|
|
}
|
|
// Whitespace-only content is as useless to a client as none at all.
|
|
if !resultIsEmpty(&types.UnifiedResponse{Content: " \n\t"}) {
|
|
t.Error("whitespace-only content must count as empty")
|
|
}
|
|
if resultIsEmpty(nil) {
|
|
t.Error("nil is not an empty response; it is an absent one")
|
|
}
|
|
}
|
|
|
|
// TestPeekStreamHoldsReasoningAndReportsEmpty is the streaming half: a channel
|
|
// of reasoning-only chunks must be reported as empty so chainDrive degrades.
|
|
func TestPeekStreamHoldsReasoningAndReportsEmpty(t *testing.T) {
|
|
in := make(chan types.UnifiedChunk, 8)
|
|
for i := 0; i < 5; i++ {
|
|
in <- types.UnifiedChunk{ReasoningContent: "thinking…"}
|
|
}
|
|
in <- types.UnifiedChunk{Done: true, FinishReason: "length"}
|
|
close(in)
|
|
|
|
_, peek := peekStream(in)
|
|
if peek() {
|
|
t.Error("a reasoning-only stream must be reported empty")
|
|
}
|
|
}
|
|
|
|
// TestPeekStreamForwardsContentAfterReasoning: the common case still works,
|
|
// and the reasoning preamble is not forwarded ahead of the content (holding it
|
|
// back is what keeps the degrade path available).
|
|
func TestPeekStreamForwardsContentAfterReasoning(t *testing.T) {
|
|
in := make(chan types.UnifiedChunk, 8)
|
|
in <- types.UnifiedChunk{ReasoningContent: "let me think"}
|
|
in <- types.UnifiedChunk{Content: "Hi"}
|
|
in <- types.UnifiedChunk{Content: "!"}
|
|
in <- types.UnifiedChunk{Done: true, FinishReason: "stop"}
|
|
close(in)
|
|
|
|
out, peek := peekStream(in)
|
|
if !peek() {
|
|
t.Fatal("a stream with content must be reported as having content")
|
|
}
|
|
var got []string
|
|
for ck := range out {
|
|
if ck.Content != "" {
|
|
got = append(got, ck.Content)
|
|
}
|
|
if strings.TrimSpace(ck.ReasoningContent) != "" {
|
|
t.Error("reasoning preamble must not be forwarded before content; " +
|
|
"that is what pins the client to a stream it cannot escape")
|
|
}
|
|
}
|
|
if strings.Join(got, "") != "Hi!" {
|
|
t.Errorf("forwarded content = %q, want %q", got, "Hi!")
|
|
}
|
|
}
|
|
|
|
// TestPeekStreamDoesNotDeadlockOnToolCalls: a tool-call delta is content for
|
|
// this purpose and must unblock peek immediately.
|
|
func TestPeekStreamDoesNotDeadlockOnToolCalls(t *testing.T) {
|
|
raw, _ := json.Marshal([]types.ToolCall{{Name: "ls"}})
|
|
in := make(chan types.UnifiedChunk, 4)
|
|
in <- types.UnifiedChunk{ToolCalls: raw}
|
|
in <- types.UnifiedChunk{Done: true, FinishReason: "tool_calls"}
|
|
close(in)
|
|
|
|
_, peek := peekStream(in)
|
|
done := make(chan bool, 1)
|
|
go func() { done <- peek() }()
|
|
select {
|
|
case ok := <-done:
|
|
if !ok {
|
|
t.Error("a tool-call delta must count as content")
|
|
}
|
|
case <-time.After(2 * time.Second):
|
|
t.Fatal("peek blocked on a tool-call-only stream")
|
|
}
|
|
_ = context.Background()
|
|
}
|