diff --git a/internal/adapter/openai/stream_status_test.go b/internal/adapter/openai/stream_status_test.go index 1601a7c..34de14f 100644 --- a/internal/adapter/openai/stream_status_test.go +++ b/internal/adapter/openai/stream_status_test.go @@ -239,7 +239,7 @@ func TestChatCompletionsStreamContentFilterStopsNormallyWithoutLeak(t *testing.T } } -func TestResponsesStreamUsageOverridesFromBatchAccumulatedTokenUsage(t *testing.T) { +func TestResponsesStreamUsageIgnoresBatchAccumulatedTokenUsage(t *testing.T) { statuses := make([]int, 0, 1) h := &Handler{ Store: mockOpenAIConfig{wideInput: true}, @@ -282,12 +282,12 @@ func TestResponsesStreamUsageOverridesFromBatchAccumulatedTokenUsage(t *testing. if usage == nil { t.Fatalf("expected usage in response payload, got %#v", resp) } - if got, _ := usage["output_tokens"].(float64); int(got) != 190 { - t.Fatalf("expected output_tokens=190, got %#v", usage["output_tokens"]) + if got, _ := usage["output_tokens"].(float64); int(got) == 190 { + t.Fatalf("expected upstream accumulated token usage to be ignored, got %#v", usage["output_tokens"]) } } -func TestResponsesNonStreamUsageOverridesPromptAndOutputTokenUsage(t *testing.T) { +func TestResponsesNonStreamUsageIgnoresPromptAndOutputTokenUsage(t *testing.T) { statuses := make([]int, 0, 1) h := &Handler{ Store: mockOpenAIConfig{wideInput: true}, @@ -322,13 +322,13 @@ func TestResponsesNonStreamUsageOverridesPromptAndOutputTokenUsage(t *testing.T) if usage == nil { t.Fatalf("expected usage object, got %#v", out) } - if got, _ := usage["input_tokens"].(float64); int(got) != 11 { - t.Fatalf("expected input_tokens=11, got %#v", usage["input_tokens"]) + input, _ := usage["input_tokens"].(float64) + output, _ := usage["output_tokens"].(float64) + total, _ := usage["total_tokens"].(float64) + if int(output) == 29 { + t.Fatalf("expected upstream completion token usage to be ignored, got %#v", usage["output_tokens"]) } - if got, _ := usage["output_tokens"].(float64); int(got) != 29 { - t.Fatalf("expected output_tokens=29, got %#v", usage["output_tokens"]) - } - if got, _ := usage["total_tokens"].(float64); int(got) != 40 { - t.Fatalf("expected total_tokens=40, got %#v", usage["total_tokens"]) + if int(total) != int(input)+int(output) { + t.Fatalf("expected total_tokens=input_tokens+output_tokens, usage=%#v", usage) } } diff --git a/internal/js/chat-stream/sse_parse_impl.js b/internal/js/chat-stream/sse_parse_impl.js index 10b85f0..0ce2896 100644 --- a/internal/js/chat-stream/sse_parse_impl.js +++ b/internal/js/chat-stream/sse_parse_impl.js @@ -443,70 +443,10 @@ function hasContentFilterStatusValue(v) { } function extractAccumulatedTokenUsage(chunk) { - const usage = findAccumulatedTokenUsage(chunk); - return usage || { prompt: 0, output: 0 }; -} - -function findAccumulatedTokenUsage(v) { - if (Array.isArray(v)) { - for (const item of v) { - const u = findAccumulatedTokenUsage(item); - if (u) return u; - } - return null; - } - if (!v || typeof v !== 'object') { - return null; - } - const pathValue = asString(v.p); - if (pathValue && pathValue.toLowerCase().includes('accumulated_token_usage')) { - const n = toInt(v.v); - if (n > 0) { - return { prompt: 0, output: n }; - } - } - if (pathValue && pathValue.toLowerCase().includes('token_usage')) { - const u = v.v; - if (u && typeof u === 'object') { - const p = toInt(u.prompt_tokens); - const c = toInt(u.completion_tokens); - if (p > 0 || c > 0) { - return { prompt: p, output: c }; - } - } - } - const direct = toInt(v.accumulated_token_usage); - if (direct > 0) { - return { prompt: 0, output: direct }; - } - if (v.token_usage && typeof v.token_usage === 'object') { - const p = toInt(v.token_usage.prompt_tokens); - const c = toInt(v.token_usage.completion_tokens); - if (p > 0 || c > 0) { - return { prompt: p, output: c }; - } - } - for (const value of Object.values(v)) { - const u = findAccumulatedTokenUsage(value); - if (u) return u; - } - return null; -} - -function toInt(v) { - if (typeof v === 'number' && Number.isFinite(v)) { - return Math.trunc(v); - } - if (typeof v === 'string' && v.trim() !== '') { - const n = Number(v); - if (Number.isFinite(n)) { - return Math.trunc(n); - } - } - if (typeof v !== 'number') { - return 0; - } - return Number.isFinite(v) ? Math.trunc(v) : 0; + // 临时策略：忽略上游 usage 字段（accumulated_token_usage / token_usage）， + // 统一使用内部估算计数，避免上下文累计口径误差。 + void chunk; + return { prompt: 0, output: 0 }; } function formatErrorMessage(v) { diff --git a/internal/sse/line_test.go b/internal/sse/line_test.go index ae6e9ac..c6c5368 100644 --- a/internal/sse/line_test.go +++ b/internal/sse/line_test.go @@ -26,7 +26,7 @@ func TestParseDeepSeekContentLineContentFilter(t *testing.T) { } } -func TestParseDeepSeekContentLineContentFilterCodeIncludesOutputTokens(t *testing.T) { +func TestParseDeepSeekContentLineContentFilterCodeIgnoresUpstreamOutputTokens(t *testing.T) { res := ParseDeepSeekContentLine( []byte(`data: {"code":"content_filter","accumulated_token_usage":99}`), false, "text", @@ -34,8 +34,8 @@ func TestParseDeepSeekContentLineContentFilterCodeIncludesOutputTokens(t *testin if !res.Parsed || !res.Stop || !res.ContentFilter { t.Fatalf("expected content-filter stop result: %#v", res) } - if res.OutputTokens != 99 { - t.Fatalf("expected output token usage 99, got %d", res.OutputTokens) + if res.OutputTokens != 0 { + t.Fatalf("expected upstream output token usage to be ignored, got %d", res.OutputTokens) } } @@ -46,27 +46,27 @@ func TestParseDeepSeekContentLineContentFilterStatus(t *testing.T) { } } -func TestParseDeepSeekContentLineCapturesAccumulatedTokenUsage(t *testing.T) { +func TestParseDeepSeekContentLineIgnoresAccumulatedTokenUsage(t *testing.T) { res := ParseDeepSeekContentLine([]byte(`data: {"p":"response","o":"BATCH","v":[{"p":"accumulated_token_usage","v":1383},{"p":"quasi_status","v":"FINISHED"}]}`), false, "text") - if res.OutputTokens != 1383 { - t.Fatalf("expected output token usage 1383, got %d", res.OutputTokens) + if res.OutputTokens != 0 { + t.Fatalf("expected accumulated token usage ignored, got %d", res.OutputTokens) } } -func TestParseDeepSeekContentLineCapturesAccumulatedTokenUsageString(t *testing.T) { +func TestParseDeepSeekContentLineIgnoresAccumulatedTokenUsageString(t *testing.T) { res := ParseDeepSeekContentLine([]byte(`data: {"p":"response","o":"BATCH","v":[{"p":"accumulated_token_usage","v":"190"},{"p":"quasi_status","v":"FINISHED"}]}`), false, "text") - if res.OutputTokens != 190 { - t.Fatalf("expected output token usage 190, got %d", res.OutputTokens) + if res.OutputTokens != 0 { + t.Fatalf("expected accumulated token usage string ignored, got %d", res.OutputTokens) } } -func TestParseDeepSeekContentLineErrorIncludesOutputTokens(t *testing.T) { +func TestParseDeepSeekContentLineErrorIgnoresUpstreamOutputTokens(t *testing.T) { res := ParseDeepSeekContentLine([]byte(`data: {"error":"boom","accumulated_token_usage":123}`), false, "text") if !res.Parsed || !res.Stop { t.Fatalf("expected stop on error: %#v", res) } - if res.OutputTokens != 123 { - t.Fatalf("expected output token usage 123 on error, got %d", res.OutputTokens) + if res.OutputTokens != 0 { + t.Fatalf("expected output token usage ignored on error, got %d", res.OutputTokens) } } diff --git a/internal/sse/parser.go b/internal/sse/parser.go index 051619e..1c4c943 100644 --- a/internal/sse/parser.go +++ b/internal/sse/parser.go @@ -3,8 +3,6 @@ package sse import ( "bytes" "encoding/json" - "math" - "strconv" "strings" "ds2api/internal/deepseek" @@ -365,84 +363,8 @@ func hasContentFilterStatusValue(v any) bool { } func extractAccumulatedTokenUsage(chunk map[string]any) (int, int) { - return findAccumulatedTokenUsage(chunk) -} - -func findAccumulatedTokenUsage(v any) (int, int) { - switch x := v.(type) { - case map[string]any: - if p, _ := x["p"].(string); strings.Contains(strings.ToLower(p), "accumulated_token_usage") { - if n, ok := toInt(x["v"]); ok && n > 0 { - return 0, n - } - } - if p, _ := x["p"].(string); strings.Contains(strings.ToLower(p), "token_usage") { - if m, ok := x["v"].(map[string]any); ok { - p, _ := toInt(m["prompt_tokens"]) - c, _ := toInt(m["completion_tokens"]) - if p > 0 || c > 0 { - return p, c - } - } - } - if n, ok := toInt(x["accumulated_token_usage"]); ok && n > 0 { - return 0, n - } - if usage, ok := x["token_usage"].(map[string]any); ok { - p, _ := toInt(usage["prompt_tokens"]) - c, _ := toInt(usage["completion_tokens"]) - if p > 0 || c > 0 { - return p, c - } - } - for _, vv := range x { - if p, c := findAccumulatedTokenUsage(vv); p > 0 || c > 0 { - return p, c - } - } - case []any: - for _, item := range x { - if p, c := findAccumulatedTokenUsage(item); p > 0 || c > 0 { - return p, c - } - } - } + // 临时策略：忽略上游 usage 字段（accumulated_token_usage / token_usage）， + // 由下游统一使用内部估算 token 计数，避免上下文累计口径导致单次输出偏差过大。 + _ = chunk return 0, 0 } - -func toInt(v any) (int, bool) { - switch x := v.(type) { - case int: - return x, true - case int32: - return int(x), true - case int64: - return int(x), true - case float64: - if math.IsNaN(x) || math.IsInf(x, 0) { - return 0, false - } - return int(x), true - case json.Number: - i, err := x.Int64() - if err != nil { - return 0, false - } - return int(i), true - case string: - s := strings.TrimSpace(x) - if s == "" { - return 0, false - } - if i, err := strconv.Atoi(s); err == nil { - return i, true - } - f, err := strconv.ParseFloat(s, 64) - if err != nil || math.IsNaN(f) || math.IsInf(f, 0) { - return 0, false - } - return int(f), true - default: - return 0, false - } -} diff --git a/internal/sse/parser_test.go b/internal/sse/parser_test.go index 89c5356..2f7b22c 100644 --- a/internal/sse/parser_test.go +++ b/internal/sse/parser_test.go @@ -28,8 +28,8 @@ func TestExtractTokenUsage(t *testing.T) { }, } p, c := extractAccumulatedTokenUsage(chunk) - if p != 123 || c != 456 { - t.Fatalf("expected 123/456, got %d/%d", p, c) + if p != 0 || c != 0 { + t.Fatalf("expected upstream usage ignored as 0/0, got %d/%d", p, c) } } diff --git a/internal/sse/raw_stream_token_replay_test.go b/internal/sse/raw_stream_token_replay_test.go index ddf0d7c..6dbe5fc 100644 --- a/internal/sse/raw_stream_token_replay_test.go +++ b/internal/sse/raw_stream_token_replay_test.go @@ -40,8 +40,8 @@ func TestRawStreamSamplesTokenReplay(t *testing.T) { if expectedTokens <= 0 { t.Fatalf("expected positive token usage from raw stream, got %d", expectedTokens) } - if parsedTokens != expectedTokens { - t.Fatalf("token mismatch parsed=%d expected=%d", parsedTokens, expectedTokens) + if parsedTokens != 0 { + t.Fatalf("expected parser to ignore upstream token usage, got parsed=%d expectedRaw=%d", parsedTokens, expectedTokens) } }) } diff --git a/tests/compat/expected/sse_content_filter_status.json b/tests/compat/expected/sse_content_filter_status.json index bf70dfb..9119fdc 100644 --- a/tests/compat/expected/sse_content_filter_status.json +++ b/tests/compat/expected/sse_content_filter_status.json @@ -3,6 +3,6 @@ "finished": true, "new_type": "text", "content_filter": true, - "output_tokens": 77, + "output_tokens": 0, "error_message": "" } diff --git a/tests/node/chat-stream.test.js b/tests/node/chat-stream.test.js index d1cc859..9315b15 100644 --- a/tests/node/chat-stream.test.js +++ b/tests/node/chat-stream.test.js @@ -248,7 +248,7 @@ test('parseChunkForContent strips reference markers from fragment content', () = assert.deepEqual(parsed.parts, [{ text: '广州天气 多云', type: 'text' }]); }); -test('parseChunkForContent detects content_filter status and carries output tokens', () => { +test('parseChunkForContent detects content_filter status and ignores upstream output tokens', () => { const chunk = { p: 'response', v: [ @@ -260,7 +260,7 @@ test('parseChunkForContent detects content_filter status and carries output toke assert.equal(parsed.parsed, true); assert.equal(parsed.finished, true); assert.equal(parsed.contentFilter, true); - assert.equal(parsed.outputTokens, 77); + assert.equal(parsed.outputTokens, 0); assert.deepEqual(parsed.parts, []); }); @@ -275,11 +275,11 @@ test('parseChunkForContent keeps error branches distinct from content_filter sta assert.equal(parsed.finished, true); assert.equal(parsed.contentFilter, false); assert.equal(parsed.errorMessage.length > 0, true); - assert.equal(parsed.outputTokens, 88); + assert.equal(parsed.outputTokens, 0); assert.deepEqual(parsed.parts, []); }); -test('parseChunkForContent preserves output tokens on FINISHED lines', () => { +test('parseChunkForContent ignores output tokens on FINISHED lines', () => { const parsed = parseChunkForContent( { p: 'response/status', v: 'FINISHED', accumulated_token_usage: 190 }, false, @@ -288,11 +288,11 @@ test('parseChunkForContent preserves output tokens on FINISHED lines', () => { assert.equal(parsed.parsed, true); assert.equal(parsed.finished, true); assert.equal(parsed.contentFilter, false); - assert.equal(parsed.outputTokens, 190); + assert.equal(parsed.outputTokens, 0); assert.deepEqual(parsed.parts, []); }); -test('parseChunkForContent captures output tokens from response BATCH status snapshots', () => { +test('parseChunkForContent ignores output tokens from response BATCH status snapshots', () => { const parsed = parseChunkForContent( { p: 'response', @@ -308,7 +308,7 @@ test('parseChunkForContent captures output tokens from response BATCH status sna assert.equal(parsed.parsed, true); assert.equal(parsed.finished, false); assert.equal(parsed.contentFilter, false); - assert.equal(parsed.outputTokens, 190); + assert.equal(parsed.outputTokens, 0); assert.deepEqual(parsed.parts, []); }); @@ -321,7 +321,7 @@ test('parseChunkForContent matches FINISHED case-insensitively on status paths', assert.equal(parsed.parsed, true); assert.equal(parsed.finished, true); assert.equal(parsed.contentFilter, false); - assert.equal(parsed.outputTokens, 190); + assert.equal(parsed.outputTokens, 0); assert.deepEqual(parsed.parts, []); }); @@ -334,7 +334,7 @@ test('parseChunkForContent filters INCOMPLETE status text without stopping strea assert.equal(parsed.parsed, true); assert.equal(parsed.finished, false); assert.equal(parsed.contentFilter, false); - assert.equal(parsed.outputTokens, 190); + assert.equal(parsed.outputTokens, 0); assert.deepEqual(parsed.parts, []); });