Delete internal directory

This commit is contained in:
公明
2026-06-18 12:37:39 +08:00
committed by GitHub
parent 5e2b30c029
commit cf1f8515d9
269 changed files with 0 additions and 77207 deletions
File diff suppressed because it is too large Load Diff
@@ -1,81 +0,0 @@
package openai
import (
"encoding/json"
"strings"
)
// claudeReasoningRoundTripSep separates human-readable reasoning from a JSON payload of
// Anthropic thinking blocks (with signatures) for multi-turn extended thinking + tools.
// Not shown in UI (see DisplayReasoningContent).
const claudeReasoningRoundTripSep = "\n---CSAI_CLAUDE_THINKING_BLOCKS---\n"
// DisplayReasoningContent returns reasoning text suitable for the UI (strips internal
// Claude round-trip JSON suffix). Safe for DeepSeek/plain reasoning strings (no-op).
func DisplayReasoningContent(s string) string {
s = strings.TrimSpace(s)
if s == "" {
return ""
}
i := strings.LastIndex(s, claudeReasoningRoundTripSep)
if i < 0 {
return s
}
return strings.TrimSpace(s[:i])
}
func appendClaudeReasoningRoundTrip(display string, blocks []claudeContentBlock) string {
var payload []map[string]string
for _, b := range blocks {
if b.Type != "thinking" {
continue
}
payload = append(payload, map[string]string{
"type": b.Type,
"thinking": b.Thinking,
"signature": b.Signature,
})
}
if len(payload) == 0 {
return strings.TrimSpace(display)
}
js, err := json.Marshal(payload)
if err != nil {
return strings.TrimSpace(display)
}
d := strings.TrimSpace(display)
if d == "" {
return claudeReasoningRoundTripSep + string(js)
}
return d + claudeReasoningRoundTripSep + string(js)
}
// parseClaudeReasoningAssistantBlocks extracts Anthropic thinking blocks from an OpenAI-style
// reasoning_content string. When no suffix is present, blocks is nil (caller must not invent signatures).
func parseClaudeReasoningAssistantBlocks(reasoningContent string) (display string, blocks []claudeContentBlock) {
reasoningContent = strings.TrimSpace(reasoningContent)
if reasoningContent == "" {
return "", nil
}
idx := strings.LastIndex(reasoningContent, claudeReasoningRoundTripSep)
if idx < 0 {
return reasoningContent, nil
}
display = strings.TrimSpace(reasoningContent[:idx])
jsonPart := strings.TrimSpace(reasoningContent[idx+len(claudeReasoningRoundTripSep):])
var arr []struct {
Type string `json:"type"`
Thinking string `json:"thinking"`
Signature string `json:"signature"`
}
if err := json.Unmarshal([]byte(jsonPart), &arr); err != nil {
return reasoningContent, nil
}
for _, x := range arr {
if x.Type != "thinking" {
continue
}
blocks = append(blocks, claudeContentBlock{Type: "thinking", Thinking: x.Thinking, Signature: x.Signature})
}
return display, blocks
}
@@ -1,102 +0,0 @@
package openai
import (
"encoding/json"
"strings"
"testing"
)
func TestDisplayReasoningContent(t *testing.T) {
raw := "hello" + claudeReasoningRoundTripSep + `[{"type":"thinking","thinking":"x","signature":"sig"}]`
if d := DisplayReasoningContent(raw); d != "hello" {
t.Fatalf("got %q", d)
}
if DisplayReasoningContent("plain") != "plain" {
t.Fatal()
}
}
func TestAppendParseClaudeReasoningRoundTrip(t *testing.T) {
blocks := []claudeContentBlock{
{Type: "thinking", Thinking: "a", Signature: "s1"},
{Type: "thinking", Thinking: "b", Signature: "s2"},
}
s := appendClaudeReasoningRoundTrip("sum", blocks)
if !strings.Contains(s, claudeReasoningRoundTripSep) {
t.Fatal("missing sep")
}
display, back := parseClaudeReasoningAssistantBlocks(s)
if display != "sum" || len(back) != 2 {
t.Fatalf("display=%q len=%d", display, len(back))
}
if back[0].Signature != "s1" || back[1].Thinking != "b" {
t.Fatalf("%+v", back)
}
}
func TestConvertOpenAIToClaude_AssistantReasoningReplay(t *testing.T) {
rc := appendClaudeReasoningRoundTrip("vis", []claudeContentBlock{
{Type: "thinking", Thinking: "t1", Signature: "sig1"},
})
payload := map[string]interface{}{
"model": "claude-3-5-sonnet-latest",
"messages": []interface{}{
map[string]interface{}{
"role": "assistant",
"content": "out",
"reasoning_content": rc,
},
},
}
req, err := convertOpenAIToClaude(payload)
if err != nil {
t.Fatal(err)
}
if len(req.Messages) != 1 {
t.Fatalf("messages=%d", len(req.Messages))
}
blocks := req.Messages[0].Content.Blocks
if len(blocks) < 2 {
t.Fatalf("blocks=%d", len(blocks))
}
if blocks[0].Type != "thinking" || blocks[0].Signature != "sig1" {
t.Fatalf("first block %+v", blocks[0])
}
foundText := false
for _, b := range blocks {
if b.Type == "text" && b.Text == "out" {
foundText = true
}
}
if !foundText {
t.Fatalf("blocks=%+v", blocks)
}
}
func TestClaudeToOpenAIResponseJSON_Thinking(t *testing.T) {
claudeBody := []byte(`{
"id":"msg_1","type":"message","role":"assistant","model":"x","stop_reason":"end_turn",
"content":[
{"type":"thinking","thinking":"step","signature":"sigx"},
{"type":"text","text":"hi"}
]
}`)
oai, err := claudeToOpenAIResponseJSON(claudeBody)
if err != nil {
t.Fatal(err)
}
var wrap map[string]interface{}
if err := json.Unmarshal(oai, &wrap); err != nil {
t.Fatal(err)
}
choices := wrap["choices"].([]interface{})
ch0 := choices[0].(map[string]interface{})
msg := ch0["message"].(map[string]interface{})
rc, _ := msg["reasoning_content"].(string)
if !strings.Contains(rc, "step") || !strings.Contains(rc, claudeReasoningRoundTripSep) {
t.Fatalf("reasoning_content=%q", rc)
}
if msg["content"] != "hi" {
t.Fatal()
}
}
-149
View File
@@ -1,149 +0,0 @@
package openai
// eino_sse_sanitizer.go 解决 Eino 走 meguminnnnnnnnn/go-openai SDK 时,
// 中转站心跳/SSE 控制行累计 > 300 行触发 ErrTooManyEmptyStreamMessages
// (报错文案: "stream has sent too many empty messages")的问题。
//
// 触发链路:
// einoopenai.NewChatModel
// → eino-ext/libs/acl/openai → meguminnnnnnnnn/go-openai
// → streamReader.processLines() 对所有非 "data:" 行计数, > 300 即抛错。
//
// 中转站常见的非 data: 行(合法 SSE 但 SDK 不接受):
// ":" / ": keepalive" / ": ping" / "event: ping" / "retry: 3000"
// 以及思考型模型 prefill 期间穿插的大量心跳。
//
// 兜底策略: 在 HTTP transport 层把响应 Body 包一层 reader, 只放行 "data:"
// 开头的行, 把心跳/注释/事件类型行就地吞掉。下游 SDK 永远见不到非 data: 行,
// 计数器始终为 0, 该错误不可能再发生。
//
// 该层对调用方完全透明:
// - 仅当响应 Content-Type 是 text/event-stream 时介入;普通 JSON 响应原样透传
// - data: payload (含 [DONE] 与 {"error":...}) 一字节不改
// - 上游真断流 (EOF / connection reset / context cancel) 原样透传
import (
"bufio"
"bytes"
"io"
"net/http"
"strings"
)
const (
// einoSSEReaderBufSize 给 bufio 一个较大的初始缓冲, 避免单行大 JSON chunk
// (含工具调用 arguments / reasoning_content) 频繁触发缓冲区扩容。
einoSSEReaderBufSize = 64 * 1024
)
// einoSSESanitizingRoundTripper 包装下游 RoundTripper, 对 SSE 响应做行级清洗。
type einoSSESanitizingRoundTripper struct {
base http.RoundTripper
}
func (rt *einoSSESanitizingRoundTripper) RoundTrip(req *http.Request) (*http.Response, error) {
resp, err := rt.base.RoundTrip(req)
if err != nil || resp == nil {
return resp, err
}
if !isSSEResponse(resp) {
return resp, nil
}
resp.Body = newEinoSSESanitizingBody(resp.Body)
return resp, nil
}
// isSSEResponse 仅对 200 + text/event-stream 的响应做清洗;
// 错误响应 (4xx/5xx 通常是 application/json) 不动, 由 SDK 走原错误路径。
func isSSEResponse(resp *http.Response) bool {
if resp.StatusCode != http.StatusOK {
return false
}
ct := resp.Header.Get("Content-Type")
if ct == "" {
return false
}
ct = strings.ToLower(strings.TrimSpace(ct))
// 兼容 "text/event-stream", "text/event-stream; charset=utf-8" 等。
return strings.HasPrefix(ct, "text/event-stream")
}
// einoSSESanitizingBody 是包装后的响应体: 只放行 data: 行, 其它行吞掉。
type einoSSESanitizingBody struct {
upstream io.ReadCloser
reader *bufio.Reader
pending []byte // 已清洗、待返回给下游的字节 (永远以 \n 结尾的完整 data: 行)
err error // upstream 终态错误 (io.EOF 或网络错误)
}
func newEinoSSESanitizingBody(body io.ReadCloser) *einoSSESanitizingBody {
return &einoSSESanitizingBody{
upstream: body,
reader: bufio.NewReaderSize(body, einoSSEReaderBufSize),
}
}
func (b *einoSSESanitizingBody) Read(p []byte) (int, error) {
if len(p) == 0 {
return 0, nil
}
if len(b.pending) > 0 {
n := copy(p, b.pending)
b.pending = b.pending[n:]
return n, nil
}
// 从上游读, 直到攒出一行 data: 或拿到终态。
// 单次循环可能丢弃任意多行心跳, 但只放行至多一行 data: 后退出,
// 避免一次 Read 阻塞过久 / pending 缓冲过大。
for b.err == nil {
line, err := b.reader.ReadBytes('\n')
if len(line) > 0 {
if isPassThroughSSELine(line) {
if line[len(line)-1] != '\n' {
line = append(line, '\n')
}
b.pending = line
if err != nil {
b.err = err
}
break
}
// 非 data: 行 (空行 / ":" 注释 / event: / retry: / id: / 任何裸文本)
// 全部吞掉, 不向下游透出, 继续循环读下一行。
}
if err != nil {
b.err = err
break
}
}
if len(b.pending) > 0 {
n := copy(p, b.pending)
b.pending = b.pending[n:]
return n, nil
}
return 0, b.err
}
func (b *einoSSESanitizingBody) Close() error {
return b.upstream.Close()
}
// isPassThroughSSELine 判定该行是否需要原样放行给下游 SDK。
// 仅 "data:" (大小写不敏感, 可有任意前导空白) 开头的行需要保留。
// 注意: 不能用 TrimSpace 去尾部换行后再判, 否则 " data: x" 会被误判;
// 我们只 trim 前导空白, 与 SDK 内部 TrimSpace 后再正则 ^data:\s* 的语义一致。
func isPassThroughSSELine(line []byte) bool {
trimmed := bytes.TrimLeft(line, " \t")
if len(trimmed) < 5 {
return false
}
// 大小写不敏感比较前 5 字节是否为 "data:"。SSE 规范要求字段名小写,
// 但宽松匹配可以兼容个别中转站的非规范实现。
return (trimmed[0] == 'd' || trimmed[0] == 'D') &&
(trimmed[1] == 'a' || trimmed[1] == 'A') &&
(trimmed[2] == 't' || trimmed[2] == 'T') &&
(trimmed[3] == 'a' || trimmed[3] == 'A') &&
trimmed[4] == ':'
}
-303
View File
@@ -1,303 +0,0 @@
package openai
import (
"bufio"
"bytes"
"errors"
"io"
"net/http"
"net/http/httptest"
"regexp"
"strings"
"testing"
)
// 复现 meguminnnnnnnnn/go-openai 的 SSE 行计数算法 (默认 limit=300):
// - 逐行读
// - 非 "data:" 行 (空行 / ":" 注释 / event: / retry:) 累计 emptyMessagesCount
// - > 300 抛 ErrTooManyEmptyStreamMessages
// - 遇到 data: 行 reset, 返回 payload
//
// 这一算法与上游 SDK 的 stream_reader.go processLines() 严格一致 (验证依据见
// /Users/temp/go/pkg/mod/github.com/meguminnnnnnnnn/go-openai@v0.1.2/stream_reader.go)。
// 测试中只复刻 "限制触发" 这一行为, 用来回归验证 sanitizer 的根因修复。
var errTooManyEmptyStreamMessages = errors.New("stream has sent too many empty messages")
func sdkLikeRecvAll(body io.Reader, limit uint) ([]string, error) {
headerData := regexp.MustCompile(`^data:\s*`)
r := bufio.NewReader(body)
var payloads []string
for {
var emptyMessagesCount uint
var payload []byte
for {
line, err := r.ReadBytes('\n')
if err != nil {
if err == io.EOF {
return payloads, nil
}
return payloads, err
}
noSpace := bytes.TrimSpace(line)
if !headerData.Match(noSpace) {
emptyMessagesCount++
if emptyMessagesCount > limit {
return payloads, errTooManyEmptyStreamMessages
}
continue
}
payload = headerData.ReplaceAll(noSpace, nil)
break
}
if string(payload) == "[DONE]" {
return payloads, nil
}
payloads = append(payloads, string(payload))
}
}
func newSSEServer(t *testing.T, body string, contentType string, status int) *httptest.Server {
t.Helper()
return httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
if contentType != "" {
w.Header().Set("Content-Type", contentType)
}
w.WriteHeader(status)
_, _ = io.WriteString(w, body)
}))
}
func sanitizingClient(base *http.Client) *http.Client {
if base == nil {
base = &http.Client{}
}
cloned := *base
transport := base.Transport
if transport == nil {
transport = http.DefaultTransport
}
cloned.Transport = &einoSSESanitizingRoundTripper{base: transport}
return &cloned
}
func readAll(t *testing.T, body io.ReadCloser) string {
t.Helper()
defer body.Close()
out, err := io.ReadAll(body)
if err != nil {
t.Fatalf("read body: %v", err)
}
return string(out)
}
// 1) 仅 data: 行 → 一字节不改地透传。
func TestSSESanitizer_PassesDataLinesUnchanged(t *testing.T) {
body := "data: {\"a\":1}\ndata: {\"b\":2}\ndata: [DONE]\n"
srv := newSSEServer(t, body, "text/event-stream", 200)
defer srv.Close()
resp, err := sanitizingClient(nil).Get(srv.URL)
if err != nil {
t.Fatalf("get: %v", err)
}
got := readAll(t, resp.Body)
if got != body {
t.Fatalf("body mismatch:\nwant %q\ngot %q", body, got)
}
}
// 2) 心跳/注释/事件类型行被吞掉, 仅保留 data: 行。
func TestSSESanitizer_DropsHeartbeatsAndControlLines(t *testing.T) {
body := strings.Join([]string{
": keepalive",
"",
"event: ping",
"retry: 3000",
"id: 42",
"data: {\"x\":1}",
": ping",
"",
"data: {\"x\":2}",
"data: [DONE]",
"",
}, "\n")
srv := newSSEServer(t, body, "text/event-stream", 200)
defer srv.Close()
resp, err := sanitizingClient(nil).Get(srv.URL)
if err != nil {
t.Fatalf("get: %v", err)
}
got := readAll(t, resp.Body)
want := "data: {\"x\":1}\ndata: {\"x\":2}\ndata: [DONE]\n"
if got != want {
t.Fatalf("sanitized body mismatch:\nwant %q\ngot %q", want, got)
}
}
// 3) 根因回归: 上游堆 500 行心跳后才发 data:, 原始 SDK 算法会抛
// ErrTooManyEmptyStreamMessages, sanitize 之后必须能正常拿到所有 data:。
func TestSSESanitizer_ProtectsAgainstTooManyEmptyMessages(t *testing.T) {
const heartbeats = 500
var buf bytes.Buffer
for i := 0; i < heartbeats; i++ {
buf.WriteString(": keepalive\n")
}
buf.WriteString("data: {\"chunk\":1}\n")
buf.WriteString("data: {\"chunk\":2}\n")
buf.WriteString("data: [DONE]\n")
t.Run("baseline_without_sanitizer_must_fail", func(t *testing.T) {
_, err := sdkLikeRecvAll(bytes.NewReader(buf.Bytes()), 300)
if !errors.Is(err, errTooManyEmptyStreamMessages) {
t.Fatalf("expected ErrTooManyEmptyStreamMessages, got %v", err)
}
})
t.Run("with_sanitizer_must_succeed", func(t *testing.T) {
srv := newSSEServer(t, buf.String(), "text/event-stream", 200)
defer srv.Close()
resp, err := sanitizingClient(nil).Get(srv.URL)
if err != nil {
t.Fatalf("get: %v", err)
}
defer resp.Body.Close()
payloads, err := sdkLikeRecvAll(resp.Body, 300)
if err != nil {
t.Fatalf("sdk-like recv after sanitize: %v", err)
}
want := []string{`{"chunk":1}`, `{"chunk":2}`}
if len(payloads) != len(want) {
t.Fatalf("payload count mismatch: want %d got %d (%v)", len(want), len(payloads), payloads)
}
for i, w := range want {
if payloads[i] != w {
t.Fatalf("payload[%d] mismatch: want %q got %q", i, w, payloads[i])
}
}
})
}
// 4) 心跳穿插在 data: 之间也能正确清洗 (思考型模型 prefill 期间常见)。
func TestSSESanitizer_HeartbeatsInterleavedWithData(t *testing.T) {
var buf bytes.Buffer
buf.WriteString("data: {\"chunk\":1}\n")
for i := 0; i < 400; i++ {
buf.WriteString(": keepalive\n")
}
buf.WriteString("data: {\"chunk\":2}\n")
buf.WriteString("data: [DONE]\n")
srv := newSSEServer(t, buf.String(), "text/event-stream", 200)
defer srv.Close()
resp, err := sanitizingClient(nil).Get(srv.URL)
if err != nil {
t.Fatalf("get: %v", err)
}
defer resp.Body.Close()
payloads, err := sdkLikeRecvAll(resp.Body, 300)
if err != nil {
t.Fatalf("sdk-like recv: %v", err)
}
if got, want := len(payloads), 2; got != want {
t.Fatalf("payload count: want %d got %d", want, got)
}
}
// 5) 非 SSE 响应 (例如非流式 JSON) 不应被 sanitizer 介入。
func TestSSESanitizer_PassesNonSSEResponseUntouched(t *testing.T) {
body := `{"id":"x","object":"chat.completion","choices":[]}`
srv := newSSEServer(t, body, "application/json", 200)
defer srv.Close()
resp, err := sanitizingClient(nil).Get(srv.URL)
if err != nil {
t.Fatalf("get: %v", err)
}
got := readAll(t, resp.Body)
if got != body {
t.Fatalf("non-SSE body must be untouched:\nwant %q\ngot %q", body, got)
}
}
// 6) 错误响应 (4xx/5xx) 不应被 sanitize, 即使 Content-Type 是 SSE 也不动,
// 避免吞掉类似 "data: " 之外的错误正文。
func TestSSESanitizer_PassesNon200Untouched(t *testing.T) {
body := `{"error":{"message":"rate limit"}}`
srv := newSSEServer(t, body, "text/event-stream", 429)
defer srv.Close()
resp, err := sanitizingClient(nil).Get(srv.URL)
if err != nil {
t.Fatalf("get: %v", err)
}
got := readAll(t, resp.Body)
if got != body {
t.Fatalf("error body must be untouched:\nwant %q\ngot %q", body, got)
}
}
// 7) data: 行末尾若缺 \n (异常上游) sanitizer 也补齐, 保证下游按行解析。
func TestSSESanitizer_AppendsTrailingNewlineIfMissing(t *testing.T) {
body := "data: {\"a\":1}"
srv := newSSEServer(t, body, "text/event-stream", 200)
defer srv.Close()
resp, err := sanitizingClient(nil).Get(srv.URL)
if err != nil {
t.Fatalf("get: %v", err)
}
got := readAll(t, resp.Body)
want := "data: {\"a\":1}\n"
if got != want {
t.Fatalf("trailing newline:\nwant %q\ngot %q", want, got)
}
}
// 8) 大 chunk (一行数十 KB) 也能完整透传, 不被切断。
func TestSSESanitizer_LargeDataLinePassesIntact(t *testing.T) {
huge := strings.Repeat("x", 80*1024)
body := "data: {\"big\":\"" + huge + "\"}\ndata: [DONE]\n"
srv := newSSEServer(t, body, "text/event-stream", 200)
defer srv.Close()
resp, err := sanitizingClient(nil).Get(srv.URL)
if err != nil {
t.Fatalf("get: %v", err)
}
got := readAll(t, resp.Body)
if got != body {
t.Fatalf("large body length mismatch: want %d got %d", len(body), len(got))
}
}
// 9) isPassThroughSSELine 单元覆盖。
func TestIsPassThroughSSELine(t *testing.T) {
cases := []struct {
line string
want bool
}{
{"data: {\"a\":1}\n", true},
{"DATA: x\n", true},
{" data: x\n", true},
{"data:\n", true},
{"\n", false},
{"\r\n", false},
{": keepalive\n", false},
{":\n", false},
{"event: ping\n", false},
{"retry: 3000\n", false},
{"id: 42\n", false},
{"datax: y\n", false},
{"da", false},
}
for _, c := range cases {
if got := isPassThroughSSELine([]byte(c.line)); got != c.want {
t.Errorf("isPassThroughSSELine(%q) = %v, want %v", c.line, got, c.want)
}
}
}
@@ -1,56 +0,0 @@
package openai
import "testing"
func TestNormalizeStreamingDelta_RepeatedCharBoundary(t *testing.T) {
// 流式在重复数字边界分片:不得把 "43" 的首字符与 "194" 尾字符误合并。
cur, d := normalizeStreamingDelta("https://x:194", "43")
if want := "https://x:19443"; cur != want {
t.Fatalf("next: want %q got %q", want, cur)
}
if d != "43" {
t.Fatalf("delta: want %q got %q", "43", d)
}
}
func TestNormalizeStreamingDelta_CumulativePrefix(t *testing.T) {
cur, d := normalizeStreamingDelta("今天", "今天天气")
if cur != "今天天气" || d != "天气" {
t.Fatalf("got cur=%q d=%q", cur, d)
}
}
func TestNormalizeStreamingDelta_FullRetransmit(t *testing.T) {
cur, d := normalizeStreamingDelta("今天", "今天")
if d != "" || cur != "今天" {
t.Fatalf("got cur=%q d=%q", cur, d)
}
}
func TestNormalizeStreamingDelta_SingleRuneRepeated(t *testing.T) {
cur, d := normalizeStreamingDelta("呀", "呀")
if want := "呀呀"; cur != want {
t.Fatalf("next: want %q got %q", want, cur)
}
if d != "呀" {
t.Fatalf("delta: want %q got %q", "呀", d)
}
cur, d = normalizeStreamingDelta("4", "4")
if want := "44"; cur != want {
t.Fatalf("next: want %q got %q", want, cur)
}
if d != "4" {
t.Fatalf("delta: want %q got %q", "4", d)
}
}
func TestNormalizeStreamingDelta_CumulativeExtendsNumber(t *testing.T) {
// 已缓冲 "194" 后收到累计串 "19443"(注意 "1943" 并非 "19443" 的前缀,不能靠误写的中间态测 HasPrefix)。
cur, d := normalizeStreamingDelta("194", "19443")
if want := "19443"; cur != want {
t.Fatalf("next: want %q got %q", want, cur)
}
if d != "43" {
t.Fatalf("delta: want %q got %q", "43", d)
}
}
-537
View File
@@ -1,537 +0,0 @@
package openai
import (
"bufio"
"bytes"
"context"
"encoding/json"
"fmt"
"io"
"net/http"
"strings"
"time"
"unicode/utf8"
"cyberstrike-ai/internal/config"
"go.uber.org/zap"
)
// Client 统一封装与OpenAI兼容模型交互的HTTP客户端。
type Client struct {
httpClient *http.Client
config *config.OpenAIConfig
logger *zap.Logger
}
// APIError 表示OpenAI接口返回的非200错误。
type APIError struct {
StatusCode int
Body string
}
func (e *APIError) Error() string {
return fmt.Sprintf("openai api error: status=%d body=%s", e.StatusCode, e.Body)
}
// normalizeStreamingDelta 将可能是“累计片段/重发片段”的内容归一化为“纯增量”。
// 部分兼容网关会返回累计 content;若直接 append 会出现重复文本。
//
// 注意:
// - 不做「任意后缀与前缀重叠」合并;流式可能在重复字符边界分片("194"+"43"→"19443")。
// - HasPrefix 仅在 incoming 严格长于 current 时视为累计全文,否则会把分片产生的第二个相同
// 单字/单码点(叠字、44、22 等)误判为「整段重复」而吞字。
// - incoming==current 仅当 current 长度 >1 个码点时才视为整包重发;单码点重复必须走拼接。
// - 不再使用「current 以 incoming 结尾则丢弃」:否则 "1943"+"43" 会误吞增量(19443 显示成 1943)。
// 若网关重复发送尾部片段,应重复送完整累计串,由 HasPrefix 分支去重。
func normalizeStreamingDelta(current, incoming string) (next, delta string) {
if incoming == "" {
return current, ""
}
if current == "" {
return incoming, incoming
}
if strings.HasPrefix(incoming, current) && len(incoming) > len(current) {
return incoming, incoming[len(current):]
}
if incoming == current && utf8.RuneCountInString(current) > 1 {
return current, ""
}
return current + incoming, incoming
}
// NewClient 创建一个新的OpenAI客户端。
func NewClient(cfg *config.OpenAIConfig, httpClient *http.Client, logger *zap.Logger) *Client {
if httpClient == nil {
httpClient = http.DefaultClient
}
if logger == nil {
logger = zap.NewNop()
}
return &Client{
httpClient: httpClient,
config: cfg,
logger: logger,
}
}
// UpdateConfig 动态更新OpenAI配置。
func (c *Client) UpdateConfig(cfg *config.OpenAIConfig) {
c.config = cfg
}
// ChatCompletion 调用 /chat/completions 接口。
func (c *Client) ChatCompletion(ctx context.Context, payload interface{}, out interface{}) error {
if c == nil {
return fmt.Errorf("openai client is not initialized")
}
if c.config == nil {
return fmt.Errorf("openai config is nil")
}
if strings.TrimSpace(c.config.APIKey) == "" {
return fmt.Errorf("openai api key is empty")
}
if c.isClaude() {
return c.claudeChatCompletion(ctx, payload, out)
}
baseURL := strings.TrimSuffix(c.config.BaseURL, "/")
if baseURL == "" {
baseURL = "https://api.openai.com/v1"
}
body, err := json.Marshal(payload)
if err != nil {
return fmt.Errorf("marshal openai payload: %w", err)
}
c.logger.Debug("sending OpenAI chat completion request",
zap.Int("payloadSizeKB", len(body)/1024))
req, err := http.NewRequestWithContext(ctx, http.MethodPost, baseURL+"/chat/completions", bytes.NewReader(body))
if err != nil {
return fmt.Errorf("build openai request: %w", err)
}
req.Header.Set("Content-Type", "application/json")
req.Header.Set("Authorization", "Bearer "+c.config.APIKey)
requestStart := time.Now()
resp, err := c.httpClient.Do(req)
if err != nil {
return fmt.Errorf("call openai api: %w", err)
}
defer resp.Body.Close()
bodyChan := make(chan []byte, 1)
errChan := make(chan error, 1)
go func() {
responseBody, err := io.ReadAll(resp.Body)
if err != nil {
errChan <- err
return
}
bodyChan <- responseBody
}()
var respBody []byte
select {
case respBody = <-bodyChan:
case err := <-errChan:
return fmt.Errorf("read openai response: %w", err)
case <-ctx.Done():
return fmt.Errorf("read openai response timeout: %w", ctx.Err())
case <-time.After(25 * time.Minute):
return fmt.Errorf("read openai response timeout (25m)")
}
c.logger.Debug("received OpenAI response",
zap.Int("status", resp.StatusCode),
zap.Duration("duration", time.Since(requestStart)),
zap.Int("responseSizeKB", len(respBody)/1024),
)
if resp.StatusCode != http.StatusOK {
c.logger.Warn("OpenAI chat completion returned non-200",
zap.Int("status", resp.StatusCode),
zap.String("body", string(respBody)),
)
return &APIError{
StatusCode: resp.StatusCode,
Body: string(respBody),
}
}
if out != nil {
if err := json.Unmarshal(respBody, out); err != nil {
c.logger.Error("failed to unmarshal OpenAI response",
zap.Error(err),
zap.String("body", string(respBody)),
)
return fmt.Errorf("unmarshal openai response: %w", err)
}
}
return nil
}
// ChatCompletionStream 调用 /chat/completions 的流式模式(stream=true),并在每个 delta 到达时回调 onDelta。
// 返回最终拼接的 content(只拼 content delta;工具调用 delta 未做处理)。
func (c *Client) ChatCompletionStream(ctx context.Context, payload interface{}, onDelta func(delta string) error) (string, error) {
if c == nil {
return "", fmt.Errorf("openai client is not initialized")
}
if c.config == nil {
return "", fmt.Errorf("openai config is nil")
}
if strings.TrimSpace(c.config.APIKey) == "" {
return "", fmt.Errorf("openai api key is empty")
}
if c.isClaude() {
return c.claudeChatCompletionStream(ctx, payload, onDelta)
}
baseURL := strings.TrimSuffix(c.config.BaseURL, "/")
if baseURL == "" {
baseURL = "https://api.openai.com/v1"
}
body, err := json.Marshal(payload)
if err != nil {
return "", fmt.Errorf("marshal openai payload: %w", err)
}
req, err := http.NewRequestWithContext(ctx, http.MethodPost, baseURL+"/chat/completions", bytes.NewReader(body))
if err != nil {
return "", fmt.Errorf("build openai request: %w", err)
}
req.Header.Set("Content-Type", "application/json")
req.Header.Set("Authorization", "Bearer "+c.config.APIKey)
requestStart := time.Now()
resp, err := c.httpClient.Do(req)
if err != nil {
return "", fmt.Errorf("call openai api: %w", err)
}
defer resp.Body.Close()
// 非200:读完 body 返回
if resp.StatusCode != http.StatusOK {
respBody, readErr := io.ReadAll(resp.Body)
if readErr != nil {
c.logger.Warn("failed to read OpenAI error response body", zap.Error(readErr))
}
return "", &APIError{
StatusCode: resp.StatusCode,
Body: string(respBody),
}
}
type streamDelta struct {
// OpenAI 兼容流式通常使用 content;但部分兼容实现可能用 text。
Content string `json:"content,omitempty"`
Text string `json:"text,omitempty"`
}
type streamChoice struct {
Delta streamDelta `json:"delta"`
FinishReason *string `json:"finish_reason,omitempty"`
}
type streamResponse struct {
ID string `json:"id,omitempty"`
Choices []streamChoice `json:"choices"`
Error *struct {
Message string `json:"message"`
Type string `json:"type"`
} `json:"error,omitempty"`
}
reader := bufio.NewReader(resp.Body)
var full strings.Builder
fullText := ""
// 典型 SSE 结构:
// data: {...}\n\n
// data: [DONE]\n\n
for {
line, readErr := reader.ReadString('\n')
if readErr != nil {
if readErr == io.EOF {
break
}
return full.String(), fmt.Errorf("read openai stream: %w", readErr)
}
trimmed := strings.TrimSpace(line)
if trimmed == "" {
continue
}
if !strings.HasPrefix(trimmed, "data:") {
continue
}
dataStr := strings.TrimSpace(strings.TrimPrefix(trimmed, "data:"))
if dataStr == "[DONE]" {
break
}
var chunk streamResponse
if err := json.Unmarshal([]byte(dataStr), &chunk); err != nil {
// 解析失败跳过(兼容各种兼容层的差异)
continue
}
if chunk.Error != nil && strings.TrimSpace(chunk.Error.Message) != "" {
return full.String(), fmt.Errorf("openai stream error: %s", chunk.Error.Message)
}
if len(chunk.Choices) == 0 {
continue
}
delta := chunk.Choices[0].Delta.Content
if delta == "" {
delta = chunk.Choices[0].Delta.Text
}
if delta == "" {
continue
}
var deltaOut string
fullText, deltaOut = normalizeStreamingDelta(fullText, delta)
if deltaOut == "" {
continue
}
full.WriteString(deltaOut)
if onDelta != nil {
if err := onDelta(deltaOut); err != nil {
return full.String(), err
}
}
}
c.logger.Debug("received OpenAI stream completion",
zap.Duration("duration", time.Since(requestStart)),
zap.Int("contentLen", full.Len()),
)
return full.String(), nil
}
// StreamToolCall 流式工具调用的累积结果(arguments 以字符串形式拼接,留给上层再解析为 JSON)。
type StreamToolCall struct {
Index int
ID string
Type string
FunctionName string
FunctionArgsStr string
}
// ChatCompletionStreamWithToolCalls 流式模式:同时把 content delta 实时回调,并在结束后返回 tool_calls 和 finish_reason。
func (c *Client) ChatCompletionStreamWithToolCalls(
ctx context.Context,
payload interface{},
onContentDelta func(delta string) error,
) (string, []StreamToolCall, string, error) {
if c == nil {
return "", nil, "", fmt.Errorf("openai client is not initialized")
}
if c.config == nil {
return "", nil, "", fmt.Errorf("openai config is nil")
}
if strings.TrimSpace(c.config.APIKey) == "" {
return "", nil, "", fmt.Errorf("openai api key is empty")
}
if c.isClaude() {
return c.claudeChatCompletionStreamWithToolCalls(ctx, payload, onContentDelta)
}
baseURL := strings.TrimSuffix(c.config.BaseURL, "/")
if baseURL == "" {
baseURL = "https://api.openai.com/v1"
}
body, err := json.Marshal(payload)
if err != nil {
return "", nil, "", fmt.Errorf("marshal openai payload: %w", err)
}
req, err := http.NewRequestWithContext(ctx, http.MethodPost, baseURL+"/chat/completions", bytes.NewReader(body))
if err != nil {
return "", nil, "", fmt.Errorf("build openai request: %w", err)
}
req.Header.Set("Content-Type", "application/json")
req.Header.Set("Authorization", "Bearer "+c.config.APIKey)
requestStart := time.Now()
resp, err := c.httpClient.Do(req)
if err != nil {
return "", nil, "", fmt.Errorf("call openai api: %w", err)
}
defer resp.Body.Close()
if resp.StatusCode != http.StatusOK {
respBody, readErr := io.ReadAll(resp.Body)
if readErr != nil {
c.logger.Warn("failed to read OpenAI error response body", zap.Error(readErr))
}
return "", nil, "", &APIError{
StatusCode: resp.StatusCode,
Body: string(respBody),
}
}
// delta tool_calls 的增量结构
type toolCallFunctionDelta struct {
Name string `json:"name,omitempty"`
Arguments string `json:"arguments,omitempty"`
}
type toolCallDelta struct {
Index int `json:"index,omitempty"`
ID string `json:"id,omitempty"`
Type string `json:"type,omitempty"`
Function toolCallFunctionDelta `json:"function,omitempty"`
}
type streamDelta2 struct {
Content string `json:"content,omitempty"`
Text string `json:"text,omitempty"`
ToolCalls []toolCallDelta `json:"tool_calls,omitempty"`
}
type streamChoice2 struct {
Delta streamDelta2 `json:"delta"`
FinishReason *string `json:"finish_reason,omitempty"`
}
type streamResponse2 struct {
Choices []streamChoice2 `json:"choices"`
Error *struct {
Message string `json:"message"`
Type string `json:"type"`
} `json:"error,omitempty"`
}
type toolCallAccum struct {
id string
typ string
name string
args strings.Builder
}
toolCallAccums := make(map[int]*toolCallAccum)
reader := bufio.NewReader(resp.Body)
var full strings.Builder
fullText := ""
finishReason := ""
for {
line, readErr := reader.ReadString('\n')
if readErr != nil {
if readErr == io.EOF {
break
}
return full.String(), nil, finishReason, fmt.Errorf("read openai stream: %w", readErr)
}
trimmed := strings.TrimSpace(line)
if trimmed == "" {
continue
}
if !strings.HasPrefix(trimmed, "data:") {
continue
}
dataStr := strings.TrimSpace(strings.TrimPrefix(trimmed, "data:"))
if dataStr == "[DONE]" {
break
}
var chunk streamResponse2
if err := json.Unmarshal([]byte(dataStr), &chunk); err != nil {
// 兼容:解析失败跳过
continue
}
if chunk.Error != nil && strings.TrimSpace(chunk.Error.Message) != "" {
return full.String(), nil, finishReason, fmt.Errorf("openai stream error: %s", chunk.Error.Message)
}
if len(chunk.Choices) == 0 {
continue
}
choice := chunk.Choices[0]
if choice.FinishReason != nil && strings.TrimSpace(*choice.FinishReason) != "" {
finishReason = strings.TrimSpace(*choice.FinishReason)
}
delta := choice.Delta
content := delta.Content
if content == "" {
content = delta.Text
}
if content != "" {
var contentOut string
fullText, contentOut = normalizeStreamingDelta(fullText, content)
if contentOut != "" {
full.WriteString(contentOut)
if onContentDelta != nil {
if err := onContentDelta(contentOut); err != nil {
return full.String(), nil, finishReason, err
}
}
}
}
if len(delta.ToolCalls) > 0 {
for _, tc := range delta.ToolCalls {
acc, ok := toolCallAccums[tc.Index]
if !ok {
acc = &toolCallAccum{}
toolCallAccums[tc.Index] = acc
}
if tc.ID != "" {
acc.id = tc.ID
}
if tc.Type != "" {
acc.typ = tc.Type
}
if tc.Function.Name != "" {
acc.name = tc.Function.Name
}
if tc.Function.Arguments != "" {
acc.args.WriteString(tc.Function.Arguments)
}
}
}
}
// 组装 tool calls
indices := make([]int, 0, len(toolCallAccums))
for idx := range toolCallAccums {
indices = append(indices, idx)
}
// 手写简单排序(避免额外 import)
for i := 0; i < len(indices); i++ {
for j := i + 1; j < len(indices); j++ {
if indices[j] < indices[i] {
indices[i], indices[j] = indices[j], indices[i]
}
}
}
toolCalls := make([]StreamToolCall, 0, len(indices))
for _, idx := range indices {
acc := toolCallAccums[idx]
tc := StreamToolCall{
Index: idx,
ID: acc.id,
Type: acc.typ,
FunctionName: acc.name,
FunctionArgsStr: acc.args.String(),
}
toolCalls = append(toolCalls, tc)
}
c.logger.Debug("received OpenAI stream completion (tool_calls)",
zap.Duration("duration", time.Since(requestStart)),
zap.Int("contentLen", full.Len()),
zap.Int("toolCalls", len(toolCalls)),
zap.String("finishReason", finishReason),
)
if strings.TrimSpace(finishReason) == "" {
finishReason = "stop"
}
return full.String(), toolCalls, finishReason, nil
}
-20
View File
@@ -1,20 +0,0 @@
package openai
// SSEAccumulatedKey 为 SSE progress 事件 data 中的服务端权威流式全文快照字段。
// 前端应优先用该字段更新 buffer,避免对 delta 二次 normalize 导致叠字。
const SSEAccumulatedKey = "accumulated"
// WithSSEAccumulated 在 progress data 中附带当前流式累计全文(权威快照)。
func WithSSEAccumulated(data map[string]interface{}, accumulated string) map[string]interface{} {
if data == nil {
data = make(map[string]interface{}, 1)
}
data[SSEAccumulatedKey] = accumulated
return data
}
// NormalizeStreamingDelta 将可能是“累计片段/重发片段”的内容归一化为“纯增量”。
// 与 unexported normalizeStreamingDelta 相同,供 agent / multiagent 等包在发 SSE 前累计正文。
func NormalizeStreamingDelta(current, incoming string) (next, delta string) {
return normalizeStreamingDelta(current, incoming)
}
-88
View File
@@ -1,88 +0,0 @@
package openai
import (
"bytes"
"io"
"net/http"
"strings"
"github.com/bytedance/sonic"
"go.uber.org/zap"
)
// SummarizationRequestHeader marks chat/completion requests issued by Eino summarization
// middleware (via model.WithExtraHeader). The diagnostic transport logs empty-choices bodies
// only for these requests so main-agent traffic stays quiet.
const SummarizationRequestHeader = "X-CyberStrike-Summarization"
const summarizationDiagBodyMaxBytes = 8192
// AttachSummarizationDiagTransport wraps client.Transport to log raw API bodies when
// summarization receives HTTP 200 with an empty choices array.
func AttachSummarizationDiagTransport(client *http.Client, logger *zap.Logger) {
if client == nil || logger == nil {
return
}
base := client.Transport
if base == nil {
base = http.DefaultTransport
}
client.Transport = &summarizationDiagRoundTripper{base: base, logger: logger}
}
type summarizationDiagRoundTripper struct {
base http.RoundTripper
logger *zap.Logger
}
func (rt *summarizationDiagRoundTripper) RoundTrip(req *http.Request) (*http.Response, error) {
resp, err := rt.base.RoundTrip(req)
if err != nil || resp == nil || resp.Body == nil {
return resp, err
}
if !isSummarizationRequest(req) || !strings.Contains(strings.ToLower(resp.Header.Get("Content-Type")), "json") {
return resp, err
}
body, readErr := io.ReadAll(resp.Body)
_ = resp.Body.Close()
if readErr != nil {
resp.Body = io.NopCloser(bytes.NewReader(nil))
return resp, err
}
resp.Body = io.NopCloser(bytes.NewReader(body))
resp.ContentLength = int64(len(body))
if rt.logger != nil && summarizationResponseEmptyChoices(body) {
rt.logger.Warn("eino summarization: API returned empty choices",
zap.Int("status", resp.StatusCode),
zap.Int("response_bytes", len(body)),
zap.String("raw_body", truncateForLog(string(body), summarizationDiagBodyMaxBytes)),
)
}
return resp, err
}
func isSummarizationRequest(req *http.Request) bool {
if req == nil {
return false
}
return strings.TrimSpace(req.Header.Get(SummarizationRequestHeader)) == "1"
}
func summarizationResponseEmptyChoices(body []byte) bool {
var parsed struct {
Choices []any `json:"choices"`
}
if err := sonic.Unmarshal(body, &parsed); err != nil {
return false
}
return len(parsed.Choices) == 0
}
func truncateForLog(s string, maxBytes int) string {
if maxBytes <= 0 || len(s) <= maxBytes {
return s
}
return s[:maxBytes] + "…(truncated)"
}
@@ -1,47 +0,0 @@
package openai
import (
"io"
"net/http"
"strings"
"testing"
"go.uber.org/zap"
)
type staticRoundTripper struct {
status int
body string
}
func (s *staticRoundTripper) RoundTrip(req *http.Request) (*http.Response, error) {
return &http.Response{
StatusCode: s.status,
Header: http.Header{"Content-Type": []string{"application/json"}},
Body: io.NopCloser(strings.NewReader(s.body)),
}, nil
}
func TestSummarizationResponseEmptyChoices(t *testing.T) {
if !summarizationResponseEmptyChoices([]byte(`{"choices":[]}`)) {
t.Fatal("expected empty choices")
}
if summarizationResponseEmptyChoices([]byte(`{"choices":[{"index":0}]}`)) {
t.Fatal("expected non-empty choices")
}
}
func TestSummarizationDiagRoundTripper_SkipsWithoutHeader(t *testing.T) {
client := &http.Client{
Transport: &summarizationDiagRoundTripper{
base: &staticRoundTripper{status: 200, body: `{"choices":[]}`},
logger: zap.NewNop(),
},
}
req, _ := http.NewRequest(http.MethodPost, "https://example.com/v1/chat/completions", nil)
resp, err := client.Do(req)
if err != nil {
t.Fatal(err)
}
_ = resp.Body.Close()
}