feat: keep short context inline

API.en.md

@@ -770,8 +770,8 @@ Reads runtime settings and status, including:
 - `runtime` (`account_max_inflight`, `account_max_queue`, `global_max_inflight`, `token_refresh_interval_hours`)
 - `responses` / `embeddings`
 - `auto_delete` (`mode`: `none` / `single` / `all`; legacy `sessions=true` is still treated as `all`)
-- `current_input_file` (`enabled` defaults to `true`, plus `min_chars`)
-- `thinking_injection` (`enabled` defaults to `true`, `prompt`, and `default_prompt`)
+- `current_input_file` (`enabled` defaults to `true`, plus `min_chars`; `inline_max_tokens` defaults to `30000`)
+- `thinking_injection` (`enabled` defaults to `false`, `prompt`, and `default_prompt`)
 - `model_aliases`
 - `env_backed`, `needs_vercel_sync`
 - `parser_v2` (`mode`: `off` / `shadow` / `enforce`; `env_override`: whether an env var is active; read-only, ignored on PUT)
@@ -786,7 +786,7 @@ Hot-updates runtime settings. Supported fields:
 - `responses.store_ttl_seconds`
 - `embeddings.provider`
 - `auto_delete.mode`
-- `current_input_file.enabled` / `current_input_file.min_chars`
+- `current_input_file.enabled` / `current_input_file.min_chars` / `current_input_file.inline_max_tokens`
 - `thinking_injection.enabled` / `thinking_injection.prompt`
 - `model_aliases`
 - `toolcall` policy is fixed and is no longer writable through settings

API.md

@@ -776,8 +776,8 @@ data: {"type":"message_stop"}
 - `runtime`（`account_max_inflight`、`account_max_queue`、`global_max_inflight`、`token_refresh_interval_hours`）
 - `responses` / `embeddings`
 - `auto_delete`（`mode`：`none` / `single` / `all`；旧配置 `sessions=true` 仍按 `all` 处理）
-- `current_input_file`（`enabled` 默认返回 `true`、`min_chars`）
-- `thinking_injection`（`enabled` 默认返回 `true`、`prompt`、`default_prompt`）
+- `current_input_file`（`enabled` 默认返回 `true`、`min_chars`、`inline_max_tokens` 默认返回 `30000`）
+- `thinking_injection`（`enabled` 默认返回 `false`、`prompt`、`default_prompt`）
 - `model_aliases`
 - `env_backed`、`needs_vercel_sync`
 - `parser_v2`（`mode`：`off` / `shadow` / `enforce`；`env_override`：是否被环境变量覆盖；只读，不接受 PUT 写入）
@@ -792,7 +792,7 @@ data: {"type":"message_stop"}
 - `responses.store_ttl_seconds`
 - `embeddings.provider`
 - `auto_delete.mode`
-- `current_input_file.enabled` / `current_input_file.min_chars`
+- `current_input_file.enabled` / `current_input_file.min_chars` / `current_input_file.inline_max_tokens`
 - `thinking_injection.enabled` / `thinking_injection.prompt`
 - `model_aliases`
 - `toolcall` 策略已固定，不再作为可写入字段

config.example.json

@@ -48,7 +48,8 @@
   },
   "current_input_file": {
     "enabled": true,
-    "min_chars": 0
+    "min_chars": 0,
+    "inline_max_tokens": 30000
   },
   "thinking_injection": {
     "enabled": false,

docs/prompt-compatibility.md

@@ -283,9 +283,9 @@ OpenAI 的文件上传现在不再是“只传文件本体”的通用路径，
 
 兼容层现在只保留 `current_input_file` 这一种拆分方式；旧的 `history_split` 配置字段已移除，读取旧配置时会忽略它且不会再写回。
 
-- `current_input_file` 默认开启；它在统一 completion runtime 入口全局生效，用于把“完整上下文”合并进 `DS2API_HISTORY.txt` 上下文文件。当最新 user turn 的纯文本长度达到 `current_input_file.min_chars`（默认 `0`）时，runtime 会上传一个文件名为 `DS2API_HISTORY.txt` 的上下文文件。文件内容会先经过各协议入口的标准化，再序列化成按轮次编号的 `DS2API_HISTORY.txt` 风格 transcript，带有 `# DS2API_HISTORY.txt` 标题和 `=== N. ROLE ===` 分段；如果当前请求带有可用 tools 且 tool choice 不是 `none`，runtime 还会上传 `DS2API_TOOLS.txt`，只承载本次请求的工具名称、描述和参数 schema。live prompt 中则会给出一个 continuation 语气的 user 消息，引导模型从 `DS2API_HISTORY.txt` 的最新状态继续推进，并直接回答最新请求；如果有 `DS2API_TOOLS.txt`，live prompt 会额外要求模型把该文件视为可调用工具和 schema 的权威来源，避免把任务拉回起点，也避免把大段 schema 再次内联到 prompt。
+- `current_input_file` 默认开启；它在统一 completion runtime 入口全局生效，并采用 inline-first 策略。runtime 会先按拆分前的完整 `PromptTokenText` 估算整体上下文 token；当整体上下文不超过 `current_input_file.inline_max_tokens`（默认 `30000`）时，请求保持完整 inline prompt，不上传生成的上下文或工具文件。只有最新 user turn 的纯文本长度达到 `current_input_file.min_chars`（默认 `0`）且整体上下文超过 inline 阈值时，runtime 才会把“完整上下文”合并进 `DS2API_HISTORY.txt` 上下文文件。文件内容会先经过各协议入口的标准化，再序列化成按轮次编号的 `DS2API_HISTORY.txt` 风格 transcript，带有 `# DS2API_HISTORY.txt` 标题和 `=== N. ROLE ===` 分段；如果当前请求带有可用 tools 且 tool choice 不是 `none`，runtime 还会上传 `DS2API_TOOLS.txt`，只承载本次请求的工具名称、描述和参数 schema。live prompt 中则会给出一个 continuation 语气的 user 消息，引导模型从 `DS2API_HISTORY.txt` 的最新状态继续推进，并直接回答最新请求；如果有 `DS2API_TOOLS.txt`，live prompt 会额外要求模型把该文件视为可调用工具和 schema 的权威来源，避免把任务拉回起点，也避免把大段 schema 再次内联到 prompt。
 - 如果 `current_input_file.enabled=false`，请求会直接透传，不上传任何拆分上下文文件。
-- 即使触发 `current_input_file` 后 live prompt 被缩短，对客户端回包里的上下文 token 统计，仍会沿用**拆分前的完整 prompt 语义**做计数，而不是按缩短后的占位 prompt 计算；否则会把真实上下文显著算小。
+- 对客户端回包里的上下文 token 统计，短上下文 inline 路径直接按完整 prompt 计数；触发 `current_input_file` 后 live prompt 被缩短时，仍会沿用**拆分前的完整 prompt 语义**做计数，而不是按缩短后的占位 prompt 计算；否则会把真实上下文显著算小。
 - 如果空输出重试最终需要切换托管账号，runtime 会在新账号下重新上传已生成的 `DS2API_HISTORY.txt` 和可选 `DS2API_TOOLS.txt`，再用新的 generated `file_id` 重建 completion payload；客户端原有文件引用会保留在 generated 文件之后。这样可以避免把旧账号不可见的 generated file_id 带到新账号请求里。
 - runtime 会为 `DS2API_HISTORY.txt`、`DS2API_TOOLS.txt` 和 live prompt 计算稳定 SHA-256 内容 hash，并写入 completion request 日志字段；`DS2API_TOOLS.txt` 还会使用账号隔离、`model_type` 隔离、内容 hash 隔离的 5 分钟内存缓存。同一账号在短时间内复用完全相同的工具 schema 时，会复用已上传的 tools `file_id`；`DS2API_HISTORY.txt` 只记录 hash，不做 file_id 缓存，避免把长对话历史大量留存在本地内存。
 

internal/config/codec.go

@@ -42,7 +42,7 @@ func (c Config) MarshalJSON() ([]byte, error) {
 		m["embeddings"] = c.Embeddings
 	}
 	m["auto_delete"] = c.AutoDelete
-	if c.CurrentInputFile.Enabled != nil || c.CurrentInputFile.MinChars != 0 {
+	if c.CurrentInputFile.Enabled != nil || c.CurrentInputFile.MinChars != 0 || c.CurrentInputFile.InlineMaxTokens != 0 {
 		m["current_input_file"] = c.CurrentInputFile
 	}
 	if c.ThinkingInjection.Enabled != nil || strings.TrimSpace(c.ThinkingInjection.Prompt) != "" {
@@ -204,8 +204,9 @@ func (c Config) Clone() Config {
 		Embeddings: c.Embeddings,
 		AutoDelete: c.AutoDelete,
 		CurrentInputFile: CurrentInputFileConfig{
-			Enabled:  cloneBoolPtr(c.CurrentInputFile.Enabled),
-			MinChars: c.CurrentInputFile.MinChars,
+			Enabled:         cloneBoolPtr(c.CurrentInputFile.Enabled),
+			MinChars:        c.CurrentInputFile.MinChars,
+			InlineMaxTokens: c.CurrentInputFile.InlineMaxTokens,
 		},
 		ThinkingInjection: ThinkingInjectionConfig{
 			Enabled: cloneBoolPtr(c.ThinkingInjection.Enabled),

internal/config/config.go

@@ -179,8 +179,9 @@ type AutoDeleteConfig struct {
 }
 
 type CurrentInputFileConfig struct {
-	Enabled  *bool `json:"enabled,omitempty"`
-	MinChars int   `json:"min_chars,omitempty"`
+	Enabled         *bool `json:"enabled,omitempty"`
+	MinChars        int   `json:"min_chars,omitempty"`
+	InlineMaxTokens int   `json:"inline_max_tokens,omitempty"`
 }
 
 type ThinkingInjectionConfig struct {

internal/config/store_accessors.go

@@ -160,6 +160,15 @@ func (s *Store) CurrentInputFileMinChars() int {
 	return s.cfg.CurrentInputFile.MinChars
 }
 
+func (s *Store) CurrentInputFileInlineMaxTokens() int {
+	s.mu.RLock()
+	defer s.mu.RUnlock()
+	if s.cfg.CurrentInputFile.InlineMaxTokens > 0 {
+		return s.cfg.CurrentInputFile.InlineMaxTokens
+	}
+	return 30000
+}
+
 func (s *Store) ThinkingInjectionEnabled() bool {
 	s.mu.RLock()
 	defer s.mu.RUnlock()

internal/config/store_accessors_test.go

@@ -10,9 +10,12 @@ func TestStoreCurrentInputFileAccessors(t *testing.T) {
 	if got := store.CurrentInputFileMinChars(); got != 0 {
 		t.Fatalf("default current input file min_chars=%d want=0", got)
 	}
+	if got := store.CurrentInputFileInlineMaxTokens(); got != 30000 {
+		t.Fatalf("default current input file inline_max_tokens=%d want=30000", got)
+	}
 
 	enabled := false
-	store.cfg.CurrentInputFile = CurrentInputFileConfig{Enabled: &enabled, MinChars: 12345}
+	store.cfg.CurrentInputFile = CurrentInputFileConfig{Enabled: &enabled, MinChars: 12345, InlineMaxTokens: 4096}
 	if store.CurrentInputFileEnabled() {
 		t.Fatal("expected current input file disabled")
 	}
@@ -25,6 +28,9 @@ func TestStoreCurrentInputFileAccessors(t *testing.T) {
 	if got := store.CurrentInputFileMinChars(); got != 12345 {
 		t.Fatalf("current input file min_chars=%d want=12345", got)
 	}
+	if got := store.CurrentInputFileInlineMaxTokens(); got != 4096 {
+		t.Fatalf("current input file inline_max_tokens=%d want=4096", got)
+	}
 }
 
 func TestStoreThinkingInjectionAccessors(t *testing.T) {

internal/config/validation.go

@@ -118,6 +118,9 @@ func ValidateCurrentInputFileConfig(currentInputFile CurrentInputFileConfig) err
 	if currentInputFile.MinChars != 0 {
 		return ValidateIntRange("current_input_file.min_chars", currentInputFile.MinChars, 1, 100000000, true)
 	}
+	if currentInputFile.InlineMaxTokens != 0 {
+		return ValidateIntRange("current_input_file.inline_max_tokens", currentInputFile.InlineMaxTokens, 1, 100000000, true)
+	}
 	return nil
 }
 

internal/config/validation_test.go

@@ -44,6 +44,11 @@ func TestValidateConfigRejectsInvalidValues(t *testing.T) {
 			cfg:  Config{CurrentInputFile: CurrentInputFileConfig{MinChars: -1}},
 			want: "current_input_file.min_chars",
 		},
+		{
+			name: "current input file inline max tokens",
+			cfg:  Config{CurrentInputFile: CurrentInputFileConfig{InlineMaxTokens: -1}},
+			want: "current_input_file.inline_max_tokens",
+		},
 	}
 
 	for _, tc := range tests {

internal/httpapi/admin/handler_settings_test.go

@@ -64,6 +64,9 @@ func TestGetSettingsIncludesCurrentInputFileDefaults(t *testing.T) {
 	if got := intFrom(currentInputFile["min_chars"]); got != 0 {
 		t.Fatalf("expected current_input_file.min_chars=0, got %d body=%v", got, body)
 	}
+	if got := intFrom(currentInputFile["inline_max_tokens"]); got != 30000 {
+		t.Fatalf("expected current_input_file.inline_max_tokens=30000, got %d body=%v", got, body)
+	}
 	thinkingInjection, _ := body["thinking_injection"].(map[string]any)
 	if got := boolFrom(thinkingInjection["enabled"]); got {
 		t.Fatalf("expected thinking_injection.enabled=false, body=%v", body)
@@ -187,8 +190,9 @@ func TestUpdateSettingsCurrentInputFile(t *testing.T) {
 	h := newAdminTestHandler(t, `{"keys":["k1"],"history_split":{"enabled":true,"trigger_after_turns":2}}`)
 	payload := map[string]any{
 		"current_input_file": map[string]any{
-			"enabled":   true,
-			"min_chars": 12345,
+			"enabled":           true,
+			"min_chars":         12345,
+			"inline_max_tokens": 4096,
 		},
 	}
 	b, _ := json.Marshal(payload)
@@ -205,13 +209,16 @@ func TestUpdateSettingsCurrentInputFile(t *testing.T) {
 	if snap.CurrentInputFile.MinChars != 12345 {
 		t.Fatalf("expected current_input_file.min_chars=12345, got %#v", snap.CurrentInputFile)
 	}
+	if snap.CurrentInputFile.InlineMaxTokens != 4096 {
+		t.Fatalf("expected current_input_file.inline_max_tokens=4096, got %#v", snap.CurrentInputFile)
+	}
 	if !h.Store.CurrentInputFileEnabled() {
 		t.Fatal("expected current input file accessor to stay enabled")
 	}
 }
 
 func TestUpdateSettingsCurrentInputFilePartialUpdatePreservesEnabled(t *testing.T) {
-	h := newAdminTestHandler(t, `{"keys":["k1"],"current_input_file":{"enabled":false,"min_chars":777}}`)
+	h := newAdminTestHandler(t, `{"keys":["k1"],"current_input_file":{"enabled":false,"min_chars":777,"inline_max_tokens":8192}}`)
 	payload := map[string]any{
 		"current_input_file": map[string]any{
 			"min_chars": 5000,
@@ -231,13 +238,17 @@ func TestUpdateSettingsCurrentInputFilePartialUpdatePreservesEnabled(t *testing.
 	if snap.CurrentInputFile.MinChars != 5000 {
 		t.Fatalf("expected current_input_file.min_chars=5000, got %#v", snap.CurrentInputFile)
 	}
+	if snap.CurrentInputFile.InlineMaxTokens != 8192 {
+		t.Fatalf("expected current_input_file.inline_max_tokens preserved, got %#v", snap.CurrentInputFile)
+	}
 }
 
 func TestUpdateSettingsCurrentInputFilePartialUpdatePreservesMinChars(t *testing.T) {
-	h := newAdminTestHandler(t, `{"keys":["k1"],"current_input_file":{"enabled":false,"min_chars":777}}`)
+	h := newAdminTestHandler(t, `{"keys":["k1"],"current_input_file":{"enabled":false,"min_chars":777,"inline_max_tokens":8192}}`)
 	payload := map[string]any{
 		"current_input_file": map[string]any{
-			"enabled": true,
+			"enabled":           true,
+			"inline_max_tokens": 16384,
 		},
 	}
 	b, _ := json.Marshal(payload)
@@ -254,6 +265,9 @@ func TestUpdateSettingsCurrentInputFilePartialUpdatePreservesMinChars(t *testing
 	if snap.CurrentInputFile.MinChars != 777 {
 		t.Fatalf("expected current_input_file.min_chars to remain 777, got %#v", snap.CurrentInputFile)
 	}
+	if snap.CurrentInputFile.InlineMaxTokens != 16384 {
+		t.Fatalf("expected current_input_file.inline_max_tokens=16384, got %#v", snap.CurrentInputFile)
+	}
 }
 
 func TestUpdateSettingsLogPartialUpdatePreservesOmittedFields(t *testing.T) {

internal/httpapi/admin/settings/handler_settings_parse.go

@@ -151,6 +151,13 @@ func parseSettingsUpdateRequest(req map[string]any) (*config.AdminConfig, *confi
 			}
 			cfg.MinChars = n
 		}
+		if v, exists := raw["inline_max_tokens"]; exists {
+			n := intFrom(v)
+			if err := config.ValidateIntRange("current_input_file.inline_max_tokens", n, 0, 100000000, true); err != nil {
+				return nil, nil, nil, nil, nil, nil, nil, nil, nil, err
+			}
+			cfg.InlineMaxTokens = n
+		}
 		if err := config.ValidateCurrentInputFileConfig(*cfg); err != nil {
 			return nil, nil, nil, nil, nil, nil, nil, nil, nil, err
 		}

internal/httpapi/admin/settings/handler_settings_read.go

@@ -32,8 +32,9 @@ func (h *Handler) getSettings(w http.ResponseWriter, _ *http.Request) {
 		"embeddings":  snap.Embeddings,
 		"auto_delete": snap.AutoDelete,
 		"current_input_file": map[string]any{
-			"enabled":   h.Store.CurrentInputFileEnabled(),
-			"min_chars": h.Store.CurrentInputFileMinChars(),
+			"enabled":           h.Store.CurrentInputFileEnabled(),
+			"min_chars":         h.Store.CurrentInputFileMinChars(),
+			"inline_max_tokens": h.Store.CurrentInputFileInlineMaxTokens(),
 		},
 		"thinking_injection": map[string]any{
 			"enabled":        h.Store.ThinkingInjectionEnabled(),

internal/httpapi/admin/settings/handler_settings_write.go

@@ -30,6 +30,7 @@ func (h *Handler) updateSettings(w http.ResponseWriter, r *http.Request) {
 	}
 	currentInputEnabledSet := hasNestedSettingsKey(req, "current_input_file", "enabled")
 	currentInputMinCharsSet := hasNestedSettingsKey(req, "current_input_file", "min_chars")
+	currentInputInlineMaxTokensSet := hasNestedSettingsKey(req, "current_input_file", "inline_max_tokens")
 	thinkingInjectionEnabledSet := hasNestedSettingsKey(req, "thinking_injection", "enabled")
 	thinkingInjectionPromptSet := hasNestedSettingsKey(req, "thinking_injection", "prompt")
 	logLevelSet := hasNestedSettingsKey(req, "log", "level")
@@ -76,6 +77,9 @@ func (h *Handler) updateSettings(w http.ResponseWriter, r *http.Request) {
 			if currentInputMinCharsSet {
 				c.CurrentInputFile.MinChars = currentInputCfg.MinChars
 			}
+			if currentInputInlineMaxTokensSet {
+				c.CurrentInputFile.InlineMaxTokens = currentInputCfg.InlineMaxTokens
+			}
 		}
 		if thinkingInjCfg != nil {
 			if thinkingInjectionEnabledSet {

internal/httpapi/admin/shared/deps.go

@@ -35,6 +35,7 @@ type ConfigStore interface {
 	AutoDeleteMode() string
 	CurrentInputFileEnabled() bool
 	CurrentInputFileMinChars() int
+	CurrentInputFileInlineMaxTokens() int
 	ThinkingInjectionEnabled() bool
 	ThinkingInjectionPrompt() string
 	AutoDeleteSessions() bool

internal/httpapi/openai/history/current_input_cache_test.go

@@ -17,6 +17,13 @@ func (currentInputCacheTestStore) CurrentInputFileEnabled() bool { return true }
 func (currentInputCacheTestStore) CurrentInputFileMinChars() int { return 0 }
 func (currentInputCacheTestStore) ContextEngineMode() string     { return "off" }
 
+type inlineFirstTestStore struct{}
+
+func (inlineFirstTestStore) CurrentInputFileEnabled() bool        { return true }
+func (inlineFirstTestStore) CurrentInputFileMinChars() int        { return 0 }
+func (inlineFirstTestStore) CurrentInputFileInlineMaxTokens() int { return 30000 }
+func (inlineFirstTestStore) ContextEngineMode() string            { return "off" }
+
 type currentInputCacheTestDS struct {
 	uploads []dsclient.UploadFileRequest
 }
@@ -26,6 +33,28 @@ func (d *currentInputCacheTestDS) UploadFile(_ context.Context, _ *auth.RequestA
 	return &dsclient.UploadFileResult{ID: fmt.Sprintf("file-%d", len(d.uploads))}, nil
 }
 
+func TestApplyCurrentInputFileKeepsShortContextInline(t *testing.T) {
+	ds := &currentInputCacheTestDS{}
+	svc := Service{Store: inlineFirstTestStore{}, DS: ds}
+	stdReq := cacheTestStandardRequest()
+	stdReq.FinalPrompt = "system rule\n\nsearch docs"
+	stdReq.PromptTokenText = stdReq.FinalPrompt
+
+	out, err := svc.ApplyCurrentInputFile(context.Background(), &auth.RequestAuth{DeepSeekToken: "token"}, stdReq)
+	if err != nil {
+		t.Fatalf("apply failed: %v", err)
+	}
+	if len(ds.uploads) != 0 {
+		t.Fatalf("expected short context to stay inline without uploads, got %d", len(ds.uploads))
+	}
+	if out.CurrentInputFileApplied {
+		t.Fatalf("expected current input file to remain unapplied for short context")
+	}
+	if out.FinalPrompt != stdReq.FinalPrompt || out.PromptTokenText != stdReq.PromptTokenText {
+		t.Fatalf("expected inline request to pass through unchanged, got %#v", out)
+	}
+}
+
 func TestApplyCurrentInputFileCachesToolsFileByStableHash(t *testing.T) {
 	ResetCurrentInputToolsFileCacheForTesting()
 	ds := &currentInputCacheTestDS{}