refactor: remove legacy compatibility configuration and UI components

2026-05-05 17:05:32 +08:00 · 2026-05-03 04:14:19 +08:00
parent 2f7cb473fc
commit 1286b02247
58 changed files with 262 additions and 517 deletions
--- a/API.en.md
+++ b/API.en.md
@@ -733,7 +733,6 @@ Reads runtime settings and status, including:
 - `success`
 - `admin` (`has_password_hash`, `jwt_expire_hours`, `jwt_valid_after_unix`, `default_password_warning`)
 - `runtime` (`account_max_inflight`, `account_max_queue`, `global_max_inflight`, `token_refresh_interval_hours`)
- `compat` (`wide_input_strict_output`, `strip_reference_markers`)
 - `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`)
@@ -747,7 +746,6 @@ Hot-updates runtime settings. Supported fields:

 - `admin.jwt_expire_hours`
 - `runtime.account_max_inflight` / `runtime.account_max_queue` / `runtime.global_max_inflight` / `runtime.token_refresh_interval_hours`
- `compat.wide_input_strict_output` / `compat.strip_reference_markers`
 - `responses.store_ttl_seconds`
 - `embeddings.provider`
 - `auto_delete.mode`
@@ -776,9 +774,9 @@ Imports full config with:

 The request can send config directly, or wrapped as `{"config": {...}, "mode":"merge"}`.
 Query params `?mode=merge` / `?mode=replace` are also supported.
-`replace` mode replaces the full config shape while preserving Vercel sync metadata. `merge` mode merges `keys`, `api_keys`, `accounts`, and `model_aliases`, and overwrites non-empty fields under `admin`, `runtime`, `responses`, and `embeddings`. Manage `compat`, `auto_delete`, and `current_input_file` via `/admin/settings` or the config file; legacy `toolcall` fields are ignored.
+`replace` mode replaces the full config shape while preserving Vercel sync metadata. `merge` mode merges `keys`, `api_keys`, `accounts`, and `model_aliases`, and overwrites non-empty fields under `admin`, `runtime`, `responses`, and `embeddings`. Manage `auto_delete` and `current_input_file` via `/admin/settings` or the config file; legacy `compat` and `toolcall` fields are ignored.

-> Note: `merge` mode does not update `compat`, `auto_delete`, or `current_input_file`.
+> Note: `merge` mode does not update `auto_delete` or `current_input_file`.

 ### `GET /admin/config/export`

--- a/API.md
+++ b/API.md
@@ -740,7 +740,6 @@ data: {"type":"message_stop"}
 - `success`
 - `admin`（`has_password_hash`、`jwt_expire_hours`、`jwt_valid_after_unix`、`default_password_warning`）
 - `runtime`（`account_max_inflight`、`account_max_queue`、`global_max_inflight`、`token_refresh_interval_hours`）
- `compat`（`wide_input_strict_output`、`strip_reference_markers`）
 - `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`）
@@ -754,7 +753,6 @@ data: {"type":"message_stop"}

 - `admin.jwt_expire_hours`
 - `runtime.account_max_inflight` / `runtime.account_max_queue` / `runtime.global_max_inflight` / `runtime.token_refresh_interval_hours`
- `compat.wide_input_strict_output` / `compat.strip_reference_markers`
 - `responses.store_ttl_seconds`
 - `embeddings.provider`
 - `auto_delete.mode`
@@ -783,9 +781,9 @@ data: {"type":"message_stop"}

 请求可直接传配置对象，或使用 `{"config": {...}, "mode":"merge"}` 包裹格式。
 也支持在查询参数里传 `?mode=merge` / `?mode=replace`。
-`replace` 模式会按完整配置结构替换（保留 Vercel 同步元信息）；`merge` 模式会合并 `keys`、`api_keys`、`accounts`、`model_aliases`，并覆盖 `admin`、`runtime`、`responses`、`embeddings` 中的非空字段。`compat`、`auto_delete`、`current_input_file` 建议通过 `/admin/settings` 或配置文件管理；`toolcall` 相关字段会被忽略。
+`replace` 模式会按完整配置结构替换（保留 Vercel 同步元信息）；`merge` 模式会合并 `keys`、`api_keys`、`accounts`、`model_aliases`，并覆盖 `admin`、`runtime`、`responses`、`embeddings` 中的非空字段。`auto_delete`、`current_input_file` 建议通过 `/admin/settings` 或配置文件管理；`compat` 与 `toolcall` 相关字段会被忽略。

-> 注意：`merge` 模式不会更新 `compat`、`auto_delete`、`current_input_file`。
+> 注意：`merge` 模式不会更新 `auto_delete`、`current_input_file`。

 ### `GET /admin/config/export`

--- a/config.example.json
+++ b/config.example.json
@@ -43,10 +43,6 @@
    "gpt-5.3-codex": "deepseek-v4-pro",
    "o3": "deepseek-v4-pro"
  },
-  "compat": {
-    "wide_input_strict_output": true,
-    "strip_reference_markers": true
-  },
  "responses": {
    "store_ttl_seconds": 900
  },
--- a/docs/ARCHITECTURE.en.md
+++ b/docs/ARCHITECTURE.en.md
@@ -65,7 +65,7 @@ ds2api/
 │   ├── textclean/                        # Text cleanup
 │   ├── toolcall/                         # Tool-call parsing and repair
 │   ├── toolstream/                       # Go streaming tool-call anti-leak and delta detection
-│   ├── translatorcliproxy/               # Cross-protocol translation bridge
+│   ├── translatorcliproxy/               # Vercel/fallback/test protocol translation bridge
 │   ├── util/                             # Shared utility helpers
 │   ├── version/                          # Version query/compare
 │   └── webui/                            # WebUI static hosting logic
@@ -187,12 +187,12 @@ flowchart LR

 - `internal/server`: router tree + middlewares (health, protocol routes, Admin/WebUI).
 - `internal/httpapi/openai/*`: OpenAI HTTP surface split into chat, responses, files, embeddings, history, and shared packages; chat/responses share the promptcompat, stream, and toolcall semantics.
- `internal/httpapi/{claude,gemini}`: protocol adapters that normalize into the same prompt compatibility semantics; direct paths share DeepSeek session/PoW/completion execution through `completionruntime`, while Vercel/proxy paths can still translate through `translatorcliproxy` into the OpenAI handler.
+- `internal/httpapi/{claude,gemini}`: protocol adapters that normalize into the same prompt compatibility semantics; normal direct paths must share DeepSeek session/PoW/completion execution through `completionruntime`, while `translatorcliproxy` is reserved for Vercel prepare/release, missing-backend fallback, and regression tests.
 - `internal/httpapi/requestbody`: shared HTTP body reading, JSON pre-validation, and UTF-8 error helpers across protocol adapters.
 - `internal/promptcompat`: compatibility core for turning OpenAI/Claude/Gemini requests into DeepSeek web-chat plain-text context.
 - `internal/assistantturn`: Go output-side canonical semantics, converting DeepSeek SSE collection results and stream finalization state into assistant turns and centralizing thinking, tool call, citation, usage, stop/error behavior.
 - `internal/completionruntime`: shared Go completion execution helpers for DeepSeek session/PoW/call startup, non-stream collection, and empty-output retry; streaming paths use it to start upstream requests, continue to use `internal/stream` for real-time consumption, and use `assistantturn` during finalization.
- `internal/translatorcliproxy`: structure translation between Claude/Gemini and OpenAI.
+- `internal/translatorcliproxy`: bridge compatibility layer for Claude/Gemini and OpenAI shape translation; it is not the main business protocol conversion center.
 - `internal/deepseek/{client,protocol,transport}`: upstream requests, sessions, PoW adaptation, protocol constants, and transport details.
 - `internal/js/chat-stream` + `api/chat-stream.js`: Vercel Node streaming bridge; Go prepare/release owns auth, account lease, and completion payload assembly, while Node relays real-time SSE with Go-aligned finalization and tool sieve semantics.
 - `internal/stream` + `internal/sse`: Go stream parsing and incremental assembly.
--- a/docs/ARCHITECTURE.md
+++ b/docs/ARCHITECTURE.md
@@ -65,7 +65,7 @@ ds2api/
 │   ├── textclean/                        # 文本清洗
 │   ├── toolcall/                         # 工具调用解析与修复
 │   ├── toolstream/                       # Go 流式 tool call 防泄漏与增量检测
-│   ├── translatorcliproxy/               # 多协议互转桥
+│   ├── translatorcliproxy/               # Vercel/fallback/测试用协议互转桥
 │   ├── util/                             # 通用工具函数
 │   ├── version/                          # 版本查询/比较
 │   └── webui/                            # WebUI 静态托管相关逻辑
@@ -187,12 +187,12 @@ flowchart LR

 - `internal/server`：路由树和中间件挂载（健康检查、协议入口、Admin/WebUI）。
 - `internal/httpapi/openai/*`：OpenAI HTTP surface，按 chat、responses、files、embeddings、history、shared 拆分；chat/responses 共享 promptcompat、stream、toolcall 等核心语义。
- `internal/httpapi/{claude,gemini}`：协议输入输出适配，归一到同一套 prompt compatibility 语义；直连路径通过 `completionruntime` 共享 DeepSeek session/PoW/completion 调用，Vercel/代理路径仍可经 `translatorcliproxy` 转到 OpenAI handler。
+- `internal/httpapi/{claude,gemini}`：协议输入输出适配，归一到同一套 prompt compatibility 语义；正常直连路径必须通过 `completionruntime` 共享 DeepSeek session/PoW/completion 调用，`translatorcliproxy` 仅保留给 Vercel prepare/release、后端缺失 fallback 和回归测试。
 - `internal/httpapi/requestbody`：跨协议复用的请求体读取、JSON 解码前置校验与 UTF-8 错误处理辅助。
 - `internal/promptcompat`：OpenAI/Claude/Gemini 请求到 DeepSeek 网页纯文本上下文的兼容内核。
 - `internal/assistantturn`：Go 输出侧统一语义层，把 DeepSeek SSE 收集结果和流式收尾状态归一成 assistant turn，集中处理 thinking、tool call、citation、usage、stop/error 语义。
 - `internal/completionruntime`：Go surface 共享的 completion 执行辅助，负责 DeepSeek session/PoW/call 启动、非流式 collect 和 empty-output retry；流式路径复用它启动上游请求，继续用 `internal/stream` 做实时消费，并在最终收尾阶段接入 `assistantturn`。
- `internal/translatorcliproxy`：Claude/Gemini 与 OpenAI 结构互转。
+- `internal/translatorcliproxy`：Claude/Gemini 与 OpenAI 结构互转的桥接兼容层，不作为主业务协议转换中心。
 - `internal/deepseek/{client,protocol,transport}`：上游请求、会话、PoW 适配、协议常量与传输层。
 - `internal/js/chat-stream` + `api/chat-stream.js`：Vercel Node 流式桥；Go prepare/release 管理鉴权、账号租约和 completion payload，Node 侧负责实时 SSE 转发并保持 Go 对齐的终结态和 tool sieve 语义。
 - `internal/stream` + `internal/sse`：Go 流式解析与增量处理。
--- a/docs/DEVELOPMENT.md
+++ b/docs/DEVELOPMENT.md
@@ -68,7 +68,7 @@ gofmt -w <changed-go-files>
 3. 请求归一化：`internal/promptcompat` 或协议转换包。
 4. 上游请求：`internal/deepseek/client`。
 5. 流式输出：`internal/stream`、`internal/sse`、`internal/toolstream`。
-6. 响应格式：`internal/format/*` 或 `internal/translatorcliproxy`。
+6. 响应格式：主路径看 `internal/assistantturn` 与 `internal/format/*`；`internal/translatorcliproxy` 只用于 Vercel/fallback/test 桥接。

 对话记录页面问题优先检查：

--- a/docs/prompt-compatibility.md
+++ b/docs/prompt-compatibility.md
@@ -108,7 +108,7 @@ DS2API 当前的核心思路，不是把客户端传来的 `messages`、`tools`
 - 对外返回给客户端的 `prompt_tokens` / `input_tokens` / `promptTokenCount` 不再按“最后一条消息”或字符粗估近似返回，而是基于**完整上下文 prompt**做 tokenizer 计数；为了避免上下文实际超限但客户端误以为还能塞下，请求侧上下文 token 会额外保守上浮一点，宁可略大也不低估。
 - 当前 `/v1/chat/completions` 业务路径仍是“每次请求新建一个远端 `chat_session_id`，并默认发送 `parent_message_id: null`”；因此 DS2API 对外默认表现为“新会话 + prompt 拼历史”，而不是复用 DeepSeek 原生会话树。
 - 但 DeepSeek 远端本身支持同一 `chat_session_id` 的跨轮次持续对话。2026-04-27 已用项目内现有 DeepSeek client 做过一次不改业务代码的双轮实测：同一 `chat_session_id` 下，第 1 轮返回 `request_message_id=1` / `response_message_id=2` / 文本 `SESSION_TEST_ONE`；第 2 轮重新获取一次 PoW，并发送 `parent_message_id=2` 后，成功返回 `request_message_id=3` / `response_message_id=4` / 文本 `SESSION_TEST_TWO`。这说明“同远端会话持续聊天”能力存在，且每轮需要携带正确的 parent/message 链接信息，同时重新获取对应轮次可用的 PoW。
- OpenAI Chat / Responses 原生走统一 OpenAI 标准化与 DeepSeek payload 组装；Claude / Gemini 会尽量复用 OpenAI prompt/tool 语义，其中 Gemini 直接复用 `promptcompat.BuildOpenAIPromptForAdapter`。Go 主服务新增 `completionruntime` 启动层，统一执行 DeepSeek session/PoW/call；输出侧新增 `assistantturn` 语义层：非流式 OpenAI Chat / Responses / Claude / Gemini 会把 DeepSeek SSE 收集结果先归一成同一份 assistant turn，再分别渲染成各协议原生外形；流式 OpenAI Chat / Responses / Claude / Gemini 继续保持各协议实时 SSE framing，但最终收尾的 tool fallback、schema 归一、usage、empty-output / content-filter 错误语义同样由 `assistantturn` 判定。Claude / Gemini 的常规 Go 主路径不再依赖内部 `httptest` 转发到 OpenAI handler；`translatorcliproxy` 仍保留用于 Vercel bridge、兼容工具和回归测试。
+- OpenAI Chat / Responses 原生走统一 OpenAI 标准化与 DeepSeek payload 组装；Claude / Gemini 会尽量复用 OpenAI prompt/tool 语义，其中 Gemini 直接复用 `promptcompat.BuildOpenAIPromptForAdapter`。Go 主服务新增 `completionruntime` 启动层，统一执行 DeepSeek session/PoW/call；输出侧新增 `assistantturn` 语义层：非流式 OpenAI Chat / Responses / Claude / Gemini 会把 DeepSeek SSE 收集结果先归一成同一份 assistant turn，再分别渲染成各协议原生外形；流式 OpenAI Chat / Responses / Claude / Gemini 继续保持各协议实时 SSE framing，但最终收尾的 tool fallback、schema 归一、usage、empty-output / content-filter 错误语义同样由 `assistantturn` 判定。Claude / Gemini 的常规 Go 主路径不再依赖内部 `httptest` 转发到 OpenAI handler；`translatorcliproxy` 仅保留用于 Vercel bridge、后端缺失 fallback 和回归测试，不作为主业务协议转换中心。
 - Vercel Node 流式路径本轮不迁移，仍使用现有 Node bridge / stream-tool-sieve 实现；后续若变更 Node 流式语义，需要按 `assistantturn` 的 Go canonical 输出语义同步对齐。
 - 客户端传入的 thinking / reasoning 开关会被归一到下游 `thinking_enabled`。Gemini `generationConfig.thinkingConfig.thinkingBudget` 会翻译成同一套 thinking 开关；关闭时即使上游返回 `response/thinking_content`，兼容层也不会把它当作可见正文输出。若最终解析出的模型名带 `-nothinking` 后缀，则会无条件强制关闭 thinking，优先级高于请求体中的 `thinking` / `reasoning` / `reasoning_effort`。未显式关闭时，各 surface 会按解析后的 DeepSeek 模型默认能力开启 thinking，并用各自协议的原生形态暴露：OpenAI Chat 为 `reasoning_content`，OpenAI Responses 为 `response.reasoning.delta` / `reasoning` content，Claude 为 `thinking` block / `thinking_delta`，Gemini 为 `thought: true` part。
 - 对 OpenAI Chat / Responses 的非流式收尾，如果最终可见正文为空，兼容层会优先尝试把思维链中的独立 DSML / XML 工具块当作真实工具调用解析出来。流式链路也会在收尾阶段做同样的 fallback 检测，但不会因为思维链内容去中途拦截或改写流式输出；真正的工具识别始终基于原始上游文本，而不是基于“已经做过可见输出清洗”的版本，因此即使最终可见层会剥离完整 leaked DSML / XML `tool_calls` wrapper、并抑制全空参数或无效 wrapper 块，也不会影响真实工具调用转成结构化 `tool_calls` / `function_call`。补发结果会作为本轮 assistant 的结构化 `tool_calls` / `function_call` 输出返回，而不是塞进 `content` 文本；如果客户端没有开启 thinking / reasoning，思维链只用于检测，不会作为 `reasoning_content` 或可见正文暴露。只有正文为空且思维链里也没有可执行工具调用时，才继续按空回复错误处理。
--- a/internal/assistantturn/turn.go
+++ b/internal/assistantturn/turn.go
@@ -51,6 +51,20 @@ type Turn struct {
 	Error             *OutputError
 }

+type FinalizeOptions struct {
+	AlreadyEmittedToolCalls bool
+}
+
+type FinalOutcome struct {
+	FinishReason     string
+	Error            *OutputError
+	Usage            Usage
+	HasToolCalls     bool
+	HasVisibleText   bool
+	HasVisibleOutput bool
+	ShouldFail       bool
+}
+
 type BuildOptions struct {
 	Model                 string
 	Prompt                string
@@ -215,6 +229,48 @@ func ShouldRetryEmptyOutput(turn Turn, attempts, maxAttempts int) bool {
 		strings.TrimSpace(turn.Thinking) == ""
 }

+func FinalizeTurn(turn Turn, opts FinalizeOptions) FinalOutcome {
+	hasToolCalls := len(turn.ToolCalls) > 0 || opts.AlreadyEmittedToolCalls
+	hasVisibleText := strings.TrimSpace(turn.Text) != ""
+	hasVisibleThinking := strings.TrimSpace(turn.Thinking) != ""
+	err := turn.Error
+	if hasToolCalls {
+		err = nil
+	}
+	finishReason := FinishReason(turn)
+	if hasToolCalls {
+		finishReason = "tool_calls"
+	}
+	return FinalOutcome{
+		FinishReason:     finishReason,
+		Error:            err,
+		Usage:            turn.Usage,
+		HasToolCalls:     hasToolCalls,
+		HasVisibleText:   hasVisibleText,
+		HasVisibleOutput: hasVisibleText || hasVisibleThinking || hasToolCalls,
+		ShouldFail:       err != nil,
+	}
+}
+
+func OpenAIChatUsage(turn Turn) map[string]any {
+	return map[string]any{
+		"prompt_tokens":     turn.Usage.InputTokens,
+		"completion_tokens": turn.Usage.OutputTokens,
+		"total_tokens":      turn.Usage.TotalTokens,
+		"completion_tokens_details": map[string]any{
+			"reasoning_tokens": turn.Usage.ReasoningTokens,
+		},
+	}
+}
+
+func OpenAIResponsesUsage(turn Turn) map[string]any {
+	return map[string]any{
+		"input_tokens":  turn.Usage.InputTokens,
+		"output_tokens": turn.Usage.OutputTokens,
+		"total_tokens":  turn.Usage.TotalTokens,
+	}
+}
+
 func FinishReason(turn Turn) string {
 	switch turn.StopReason {
 	case StopReasonToolCalls:
--- a/internal/assistantturn/turn_test.go
+++ b/internal/assistantturn/turn_test.go
@@ -98,3 +98,30 @@ func TestBuildTurnFromStreamSnapshotAlreadyEmittedToolAvoidsEmptyError(t *testin
 		t.Fatalf("stop reason mismatch: %q", turn.StopReason)
 	}
 }
+
+func TestFinalizeTurnStopOutcome(t *testing.T) {
+	turn := BuildTurnFromCollected(sse.CollectResult{Text: "hello"}, BuildOptions{})
+	outcome := FinalizeTurn(turn, FinalizeOptions{})
+	if outcome.ShouldFail {
+		t.Fatalf("unexpected failure: %#v", outcome.Error)
+	}
+	if outcome.FinishReason != "stop" || !outcome.HasVisibleText || !outcome.HasVisibleOutput {
+		t.Fatalf("unexpected outcome: %#v", outcome)
+	}
+}
+
+func TestFinalizeTurnToolCallsOutcome(t *testing.T) {
+	turn := BuildTurnFromStreamSnapshot(StreamSnapshot{AlreadyEmittedCalls: true}, BuildOptions{})
+	outcome := FinalizeTurn(turn, FinalizeOptions{AlreadyEmittedToolCalls: true})
+	if outcome.ShouldFail || outcome.FinishReason != "tool_calls" || !outcome.HasToolCalls {
+		t.Fatalf("unexpected tool outcome: %#v", outcome)
+	}
+}
+
+func TestFinalizeTurnContentFilterOutcome(t *testing.T) {
+	turn := BuildTurnFromCollected(sse.CollectResult{ContentFilter: true}, BuildOptions{})
+	outcome := FinalizeTurn(turn, FinalizeOptions{})
+	if !outcome.ShouldFail || outcome.Error == nil || outcome.Error.Code != "content_filter" {
+		t.Fatalf("expected content filter failure, got %#v", outcome)
+	}
+}
--- a/internal/config/codec.go
+++ b/internal/config/codec.go
@@ -35,9 +35,6 @@ func (c Config) MarshalJSON() ([]byte, error) {
 	if c.Runtime.AccountMaxInflight > 0 || c.Runtime.AccountMaxQueue > 0 || c.Runtime.GlobalMaxInflight > 0 || c.Runtime.TokenRefreshIntervalHours > 0 {
 		m["runtime"] = c.Runtime
 	}
-	if c.Compat.WideInputStrictOutput != nil || c.Compat.StripReferenceMarkers != nil {
-		m["compat"] = c.Compat
-	}
 	if c.Responses.StoreTTLSeconds > 0 {
 		m["responses"] = c.Responses
 	}
@@ -100,9 +97,7 @@ func (c *Config) UnmarshalJSON(b []byte) error {
 				return fmt.Errorf("invalid field %q: %w", k, err)
 			}
 		case "compat":
-			if err := json.Unmarshal(v, &c.Compat); err != nil {
-				return fmt.Errorf("invalid field %q: %w", k, err)
-			}
+			// Removed field ignored instead of persisted.
 		case "toolcall":
 			// Legacy field ignored. Toolcall policy is fixed and no longer configurable.
 		case "responses":
@@ -155,13 +150,9 @@ func (c Config) Clone() Config {
 		ModelAliases: cloneStringMap(c.ModelAliases),
 		Admin:        c.Admin,
 		Runtime:      c.Runtime,
-		Compat: CompatConfig{
-			WideInputStrictOutput: cloneBoolPtr(c.Compat.WideInputStrictOutput),
-			StripReferenceMarkers: cloneBoolPtr(c.Compat.StripReferenceMarkers),
-		},
-		Responses:  c.Responses,
-		Embeddings: c.Embeddings,
-		AutoDelete: c.AutoDelete,
+		Responses:    c.Responses,
+		Embeddings:   c.Embeddings,
+		AutoDelete:   c.AutoDelete,
 		CurrentInputFile: CurrentInputFileConfig{
 			Enabled:  cloneBoolPtr(c.CurrentInputFile.Enabled),
 			MinChars: c.CurrentInputFile.MinChars,
--- a/internal/config/config.go
+++ b/internal/config/config.go
@@ -15,7 +15,6 @@ type Config struct {
 	ModelAliases      map[string]string       `json:"model_aliases,omitempty"`
 	Admin             AdminConfig             `json:"admin,omitempty"`
 	Runtime           RuntimeConfig           `json:"runtime,omitempty"`
-	Compat            CompatConfig            `json:"compat,omitempty"`
 	Responses         ResponsesConfig         `json:"responses,omitempty"`
 	Embeddings        EmbeddingsConfig        `json:"embeddings,omitempty"`
 	AutoDelete        AutoDeleteConfig        `json:"auto_delete"`
@@ -141,11 +140,6 @@ func (c *Config) normalizeModelAliases() {
 	}
 }

-type CompatConfig struct {
-	WideInputStrictOutput *bool `json:"wide_input_strict_output,omitempty"`
-	StripReferenceMarkers *bool `json:"strip_reference_markers,omitempty"`
-}
-
 type AdminConfig struct {
 	PasswordHash      string `json:"password_hash,omitempty"`
 	JWTExpireHours    int    `json:"jwt_expire_hours,omitempty"`
--- a/internal/config/config_edge_test.go
+++ b/internal/config/config_edge_test.go
@@ -163,8 +163,6 @@ func TestLowerFunction(t *testing.T) {
 // ─── Config.MarshalJSON / UnmarshalJSON roundtrip ────────────────────

 func TestConfigJSONRoundtrip(t *testing.T) {
-	trueVal := true
-	falseVal := false
 	cfg := Config{
 		Keys:         []string{"key1", "key2"},
 		Accounts:     []Account{{Email: "user@example.com", Password: "pass", Token: "tok"}},
@@ -175,10 +173,6 @@ func TestConfigJSONRoundtrip(t *testing.T) {
 		Runtime: RuntimeConfig{
 			TokenRefreshIntervalHours: 12,
 		},
-		Compat: CompatConfig{
-			WideInputStrictOutput: &trueVal,
-			StripReferenceMarkers: &falseVal,
-		},
 		VercelSyncHash: "hash123",
 		VercelSyncTime: 1234567890,
 		AdditionalFields: map[string]any{
@@ -211,12 +205,6 @@ func TestConfigJSONRoundtrip(t *testing.T) {
 	if decoded.AutoDelete.Mode != "single" {
 		t.Fatalf("unexpected auto delete mode: %#v", decoded.AutoDelete.Mode)
 	}
-	if decoded.Compat.WideInputStrictOutput == nil || !*decoded.Compat.WideInputStrictOutput {
-		t.Fatalf("unexpected compat wide_input_strict_output: %#v", decoded.Compat.WideInputStrictOutput)
-	}
-	if decoded.Compat.StripReferenceMarkers == nil || *decoded.Compat.StripReferenceMarkers {
-		t.Fatalf("unexpected compat strip_reference_markers: %#v", decoded.Compat.StripReferenceMarkers)
-	}
 	if decoded.VercelSyncHash != "hash123" {
 		t.Fatalf("unexpected vercel sync hash: %q", decoded.VercelSyncHash)
 	}
@@ -301,14 +289,10 @@ func TestConfigUnmarshalJSONIgnoresRemovedHistorySplit(t *testing.T) {
 // ─── Config.Clone ────────────────────────────────────────────────────

 func TestConfigCloneIsDeepCopy(t *testing.T) {
-	falseVal := false
 	cfg := Config{
-		Keys:         []string{"key1"},
-		Accounts:     []Account{{Email: "user@test.com", Token: "token"}},
-		ModelAliases: map[string]string{"claude-sonnet-4-6": "deepseek-v4-flash"},
-		Compat: CompatConfig{
-			StripReferenceMarkers: &falseVal,
-		},
+		Keys:             []string{"key1"},
+		Accounts:         []Account{{Email: "user@test.com", Token: "token"}},
+		ModelAliases:     map[string]string{"claude-sonnet-4-6": "deepseek-v4-flash"},
 		AdditionalFields: map[string]any{"custom": "value"},
 	}

@@ -318,9 +302,6 @@ func TestConfigCloneIsDeepCopy(t *testing.T) {
 	cfg.Keys[0] = "modified"
 	cfg.Accounts[0].Email = "modified@test.com"
 	cfg.ModelAliases["claude-sonnet-4-6"] = "modified-model"
-	if cfg.Compat.StripReferenceMarkers != nil {
-		*cfg.Compat.StripReferenceMarkers = true
-	}

 	// Cloned should not be affected
 	if cloned.Keys[0] != "key1" {
@@ -332,9 +313,6 @@ func TestConfigCloneIsDeepCopy(t *testing.T) {
 	if cloned.ModelAliases["claude-sonnet-4-6"] != "deepseek-v4-flash" {
 		t.Fatalf("clone model aliases was affected: %#v", cloned.ModelAliases)
 	}
-	if cloned.Compat.StripReferenceMarkers == nil || *cloned.Compat.StripReferenceMarkers {
-		t.Fatalf("clone compat was affected: %#v", cloned.Compat.StripReferenceMarkers)
-	}
 }

 func TestConfigCloneNilMaps(t *testing.T) {
@@ -473,53 +451,9 @@ func TestStoreFindAccountNotFound(t *testing.T) {
 	}
 }

-func TestStoreCompatWideInputStrictOutputDefaultTrue(t *testing.T) {
-	t.Setenv("DS2API_CONFIG_JSON", `{"keys":["k1"],"accounts":[]}`)
-	store := LoadStore()
-	if !store.CompatWideInputStrictOutput() {
-		t.Fatal("expected default wide_input_strict_output=true when unset")
-	}
-}
-
-func TestStoreCompatWideInputStrictOutputCanDisable(t *testing.T) {
-	t.Setenv("DS2API_CONFIG_JSON", `{"keys":["k1"],"accounts":[],"compat":{"wide_input_strict_output":false}}`)
-	store := LoadStore()
-	if store.CompatWideInputStrictOutput() {
-		t.Fatal("expected wide_input_strict_output=false when explicitly configured")
-	}
-
-	snap := store.Snapshot()
-	data, err := snap.MarshalJSON()
-	if err != nil {
-		t.Fatalf("marshal failed: %v", err)
-	}
-	var out map[string]any
-	if err := json.Unmarshal(data, &out); err != nil {
-		t.Fatalf("decode failed: %v", err)
-	}
-	rawCompat, ok := out["compat"].(map[string]any)
-	if !ok {
-		t.Fatalf("expected compat in marshaled output, got %#v", out)
-	}
-	if rawCompat["wide_input_strict_output"] != false {
-		t.Fatalf("expected explicit false in compat, got %#v", rawCompat)
-	}
-}
-
-func TestStoreCompatStripReferenceMarkersDefaultTrue(t *testing.T) {
-	t.Setenv("DS2API_CONFIG_JSON", `{"keys":["k1"],"accounts":[]}`)
-	store := LoadStore()
-	if !store.CompatStripReferenceMarkers() {
-		t.Fatal("expected default strip_reference_markers=true when unset")
-	}
-}
-
-func TestStoreCompatStripReferenceMarkersCanDisable(t *testing.T) {
+func TestStoreIgnoresRemovedCompatConfig(t *testing.T) {
 	t.Setenv("DS2API_CONFIG_JSON", `{"keys":["k1"],"accounts":[],"compat":{"strip_reference_markers":false}}`)
 	store := LoadStore()
-	if store.CompatStripReferenceMarkers() {
-		t.Fatal("expected strip_reference_markers=false when explicitly configured")
-	}

 	snap := store.Snapshot()
 	data, err := snap.MarshalJSON()
@@ -530,12 +464,8 @@ func TestStoreCompatStripReferenceMarkersCanDisable(t *testing.T) {
 	if err := json.Unmarshal(data, &out); err != nil {
 		t.Fatalf("decode failed: %v", err)
 	}
-	rawCompat, ok := out["compat"].(map[string]any)
-	if !ok {
-		t.Fatalf("expected compat in marshaled output, got %#v", out)
-	}
-	if rawCompat["strip_reference_markers"] != false {
-		t.Fatalf("expected explicit false in compat, got %#v", rawCompat)
+	if _, ok := out["compat"]; ok {
+		t.Fatalf("expected removed compat field not to marshal, got %#v", out)
 	}
 }

--- a/internal/config/store_accessors.go
+++ b/internal/config/store_accessors.go
@@ -21,24 +21,6 @@ func (s *Store) ModelAliases() map[string]string {
 	return out
 }

-func (s *Store) CompatWideInputStrictOutput() bool {
-	s.mu.RLock()
-	defer s.mu.RUnlock()
-	if s.cfg.Compat.WideInputStrictOutput == nil {
-		return true
-	}
-	return *s.cfg.Compat.WideInputStrictOutput
-}
-
-func (s *Store) CompatStripReferenceMarkers() bool {
-	s.mu.RLock()
-	defer s.mu.RUnlock()
-	if s.cfg.Compat.StripReferenceMarkers == nil {
-		return true
-	}
-	return *s.cfg.Compat.StripReferenceMarkers
-}
-
 func (s *Store) ToolcallMode() string {
 	return "feature_match"
 }
--- a/internal/httpapi/admin/settings/handler_settings_parse.go
+++ b/internal/httpapi/admin/settings/handler_settings_parse.go
@@ -21,11 +21,10 @@ func boolFrom(v any) bool {
 	}
 }

-func parseSettingsUpdateRequest(req map[string]any) (*config.AdminConfig, *config.RuntimeConfig, *config.CompatConfig, *config.ResponsesConfig, *config.EmbeddingsConfig, *config.AutoDeleteConfig, *config.CurrentInputFileConfig, *config.ThinkingInjectionConfig, map[string]string, error) {
+func parseSettingsUpdateRequest(req map[string]any) (*config.AdminConfig, *config.RuntimeConfig, *config.ResponsesConfig, *config.EmbeddingsConfig, *config.AutoDeleteConfig, *config.CurrentInputFileConfig, *config.ThinkingInjectionConfig, map[string]string, error) {
 	var (
 		adminCfg        *config.AdminConfig
 		runtimeCfg      *config.RuntimeConfig
-		compatCfg       *config.CompatConfig
 		respCfg         *config.ResponsesConfig
 		embCfg          *config.EmbeddingsConfig
 		autoDeleteCfg   *config.AutoDeleteConfig
@@ -39,7 +38,7 @@ func parseSettingsUpdateRequest(req map[string]any) (*config.AdminConfig, *confi
 		if v, exists := raw["jwt_expire_hours"]; exists {
 			n := intFrom(v)
 			if err := config.ValidateIntRange("admin.jwt_expire_hours", n, 1, 720, true); err != nil {
-				return nil, nil, nil, nil, nil, nil, nil, nil, nil, err
+				return nil, nil, nil, nil, nil, nil, nil, nil, err
 			}
 			cfg.JWTExpireHours = n
 		}
@@ -51,56 +50,43 @@ func parseSettingsUpdateRequest(req map[string]any) (*config.AdminConfig, *confi
 		if v, exists := raw["account_max_inflight"]; exists {
 			n := intFrom(v)
 			if err := config.ValidateIntRange("runtime.account_max_inflight", n, 1, 256, true); err != nil {
-				return nil, nil, nil, nil, nil, nil, nil, nil, nil, err
+				return nil, nil, nil, nil, nil, nil, nil, nil, err
 			}
 			cfg.AccountMaxInflight = n
 		}
 		if v, exists := raw["account_max_queue"]; exists {
 			n := intFrom(v)
 			if err := config.ValidateIntRange("runtime.account_max_queue", n, 1, 200000, true); err != nil {
-				return nil, nil, nil, nil, nil, nil, nil, nil, nil, err
+				return nil, nil, nil, nil, nil, nil, nil, nil, err
 			}
 			cfg.AccountMaxQueue = n
 		}
 		if v, exists := raw["global_max_inflight"]; exists {
 			n := intFrom(v)
 			if err := config.ValidateIntRange("runtime.global_max_inflight", n, 1, 200000, true); err != nil {
-				return nil, nil, nil, nil, nil, nil, nil, nil, nil, err
+				return nil, nil, nil, nil, nil, nil, nil, nil, err
 			}
 			cfg.GlobalMaxInflight = n
 		}
 		if v, exists := raw["token_refresh_interval_hours"]; exists {
 			n := intFrom(v)
 			if err := config.ValidateIntRange("runtime.token_refresh_interval_hours", n, 1, 720, true); err != nil {
-				return nil, nil, nil, nil, nil, nil, nil, nil, nil, err
+				return nil, nil, nil, nil, nil, nil, nil, nil, err
 			}
 			cfg.TokenRefreshIntervalHours = n
 		}
 		if cfg.AccountMaxInflight > 0 && cfg.GlobalMaxInflight > 0 && cfg.GlobalMaxInflight < cfg.AccountMaxInflight {
-			return nil, nil, nil, nil, nil, nil, nil, nil, nil, fmt.Errorf("runtime.global_max_inflight must be >= runtime.account_max_inflight")
+			return nil, nil, nil, nil, nil, nil, nil, nil, fmt.Errorf("runtime.global_max_inflight must be >= runtime.account_max_inflight")
 		}
 		runtimeCfg = cfg
 	}

-	if raw, ok := req["compat"].(map[string]any); ok {
-		cfg := &config.CompatConfig{}
-		if v, exists := raw["wide_input_strict_output"]; exists {
-			b := boolFrom(v)
-			cfg.WideInputStrictOutput = &b
-		}
-		if v, exists := raw["strip_reference_markers"]; exists {
-			b := boolFrom(v)
-			cfg.StripReferenceMarkers = &b
-		}
-		compatCfg = cfg
-	}
-
 	if raw, ok := req["responses"].(map[string]any); ok {
 		cfg := &config.ResponsesConfig{}
 		if v, exists := raw["store_ttl_seconds"]; exists {
 			n := intFrom(v)
 			if err := config.ValidateIntRange("responses.store_ttl_seconds", n, 30, 86400, true); err != nil {
-				return nil, nil, nil, nil, nil, nil, nil, nil, nil, err
+				return nil, nil, nil, nil, nil, nil, nil, nil, err
 			}
 			cfg.StoreTTLSeconds = n
 		}
@@ -112,7 +98,7 @@ func parseSettingsUpdateRequest(req map[string]any) (*config.AdminConfig, *confi
 		if v, exists := raw["provider"]; exists {
 			p := strings.TrimSpace(fmt.Sprintf("%v", v))
 			if err := config.ValidateTrimmedString("embeddings.provider", p, false); err != nil {
-				return nil, nil, nil, nil, nil, nil, nil, nil, nil, err
+				return nil, nil, nil, nil, nil, nil, nil, nil, err
 			}
 			cfg.Provider = p
 		}
@@ -138,7 +124,7 @@ func parseSettingsUpdateRequest(req map[string]any) (*config.AdminConfig, *confi
 		if v, exists := raw["mode"]; exists {
 			mode := strings.ToLower(strings.TrimSpace(fmt.Sprintf("%v", v)))
 			if err := config.ValidateAutoDeleteMode(mode); err != nil {
-				return nil, nil, nil, nil, nil, nil, nil, nil, nil, err
+				return nil, nil, nil, nil, nil, nil, nil, nil, err
 			}
 			if mode == "" {
 				mode = "none"
@@ -160,12 +146,12 @@ func parseSettingsUpdateRequest(req map[string]any) (*config.AdminConfig, *confi
 		if v, exists := raw["min_chars"]; exists {
 			n := intFrom(v)
 			if err := config.ValidateIntRange("current_input_file.min_chars", n, 0, 100000000, true); err != nil {
-				return nil, nil, nil, nil, nil, nil, nil, nil, nil, err
+				return nil, nil, nil, nil, nil, nil, nil, nil, err
 			}
 			cfg.MinChars = n
 		}
 		if err := config.ValidateCurrentInputFileConfig(*cfg); err != nil {
-			return nil, nil, nil, nil, nil, nil, nil, nil, nil, err
+			return nil, nil, nil, nil, nil, nil, nil, nil, err
 		}
 		currentInputCfg = cfg
 	}
@@ -182,5 +168,5 @@ func parseSettingsUpdateRequest(req map[string]any) (*config.AdminConfig, *confi
 		thinkingInjCfg = cfg
 	}

-	return adminCfg, runtimeCfg, compatCfg, respCfg, embCfg, autoDeleteCfg, currentInputCfg, thinkingInjCfg, aliasMap, nil
+	return adminCfg, runtimeCfg, respCfg, embCfg, autoDeleteCfg, currentInputCfg, thinkingInjCfg, aliasMap, nil
 }
--- a/internal/httpapi/admin/settings/handler_settings_read.go
+++ b/internal/httpapi/admin/settings/handler_settings_read.go
@@ -27,7 +27,6 @@ func (h *Handler) getSettings(w http.ResponseWriter, _ *http.Request) {
 			"global_max_inflight":          h.Store.RuntimeGlobalMaxInflight(recommended),
 			"token_refresh_interval_hours": h.Store.RuntimeTokenRefreshIntervalHours(),
 		},
-		"compat":      snap.Compat,
 		"responses":   snap.Responses,
 		"embeddings":  snap.Embeddings,
 		"auto_delete": snap.AutoDelete,
--- a/internal/httpapi/admin/settings/handler_settings_write.go
+++ b/internal/httpapi/admin/settings/handler_settings_write.go
@@ -17,7 +17,7 @@ func (h *Handler) updateSettings(w http.ResponseWriter, r *http.Request) {
 		return
 	}

-	adminCfg, runtimeCfg, compatCfg, responsesCfg, embeddingsCfg, autoDeleteCfg, currentInputCfg, thinkingInjCfg, aliasMap, err := parseSettingsUpdateRequest(req)
+	adminCfg, runtimeCfg, responsesCfg, embeddingsCfg, autoDeleteCfg, currentInputCfg, thinkingInjCfg, aliasMap, err := parseSettingsUpdateRequest(req)
 	if err != nil {
 		writeJSON(w, http.StatusBadRequest, map[string]any{"detail": err.Error()})
 		return
@@ -53,14 +53,6 @@ func (h *Handler) updateSettings(w http.ResponseWriter, r *http.Request) {
 				c.Runtime.TokenRefreshIntervalHours = runtimeCfg.TokenRefreshIntervalHours
 			}
 		}
-		if compatCfg != nil {
-			if compatCfg.WideInputStrictOutput != nil {
-				c.Compat.WideInputStrictOutput = compatCfg.WideInputStrictOutput
-			}
-			if compatCfg.StripReferenceMarkers != nil {
-				c.Compat.StripReferenceMarkers = compatCfg.StripReferenceMarkers
-			}
-		}
 		if responsesCfg != nil && responsesCfg.StoreTTLSeconds > 0 {
 			c.Responses.StoreTTLSeconds = responsesCfg.StoreTTLSeconds
 		}
--- a/internal/httpapi/admin/shared/deps.go
+++ b/internal/httpapi/admin/shared/deps.go
@@ -37,7 +37,6 @@ type ConfigStore interface {
 	CurrentInputFileMinChars() int
 	ThinkingInjectionEnabled() bool
 	ThinkingInjectionPrompt() string
-	CompatStripReferenceMarkers() bool
 	AutoDeleteSessions() bool
 }

--- a/internal/httpapi/claude/deps.go
+++ b/internal/httpapi/claude/deps.go
@@ -23,7 +23,6 @@ type DeepSeekCaller interface {

 type ConfigReader interface {
 	ModelAliases() map[string]string
-	CompatStripReferenceMarkers() bool
 	CurrentInputFileEnabled() bool
 	CurrentInputFileMinChars() int
 }
--- a/internal/httpapi/claude/deps_injection_test.go
+++ b/internal/httpapi/claude/deps_injection_test.go
@@ -7,7 +7,6 @@ type mockClaudeConfig struct {
 }

 func (m mockClaudeConfig) ModelAliases() map[string]string { return m.aliases }
-func (mockClaudeConfig) CompatStripReferenceMarkers() bool { return true }
 func (mockClaudeConfig) CurrentInputFileEnabled() bool     { return true }
 func (mockClaudeConfig) CurrentInputFileMinChars() int     { return 0 }

--- a/internal/httpapi/claude/handler_messages.go
+++ b/internal/httpapi/claude/handler_messages.go
@@ -84,7 +84,7 @@ func (h *Handler) handleClaudeDirect(w http.ResponseWriter, r *http.Request) boo
 		return true
 	}
 	result, outErr := completionruntime.ExecuteNonStreamWithRetry(r.Context(), h.DS, a, norm.Standard, completionruntime.Options{
-		StripReferenceMarkers: h.compatStripReferenceMarkers(),
+		StripReferenceMarkers: stripReferenceMarkersEnabled(),
 		RetryEnabled:          true,
 		CurrentInputFile:      h.Store,
 	})
@@ -290,7 +290,7 @@ func (h *Handler) handleClaudeStreamRealtime(w http.ResponseWriter, r *http.Requ
 		messages,
 		thinkingEnabled,
 		searchEnabled,
-		h.compatStripReferenceMarkers(),
+		stripReferenceMarkersEnabled(),
 		toolNames,
 		toolsRaw,
 		buildClaudePromptTokenText(messages, thinkingEnabled),
--- a/internal/httpapi/claude/handler_routes.go
+++ b/internal/httpapi/claude/handler_routes.go
@@ -21,11 +21,8 @@ type Handler struct {
 	OpenAI OpenAIChatRunner
 }

-func (h *Handler) compatStripReferenceMarkers() bool {
-	if h == nil || h.Store == nil {
-		return true
-	}
-	return h.Store.CompatStripReferenceMarkers()
+func stripReferenceMarkersEnabled() bool {
+	return true
 }

 var (
--- a/internal/httpapi/claude/proxy_vercel_test.go
+++ b/internal/httpapi/claude/proxy_vercel_test.go
@@ -14,9 +14,8 @@ type claudeProxyStoreStub struct {

 func (s claudeProxyStoreStub) ModelAliases() map[string]string { return s.aliases }

-func (claudeProxyStoreStub) CompatStripReferenceMarkers() bool { return true }
-func (claudeProxyStoreStub) CurrentInputFileEnabled() bool     { return true }
-func (claudeProxyStoreStub) CurrentInputFileMinChars() int     { return 0 }
+func (claudeProxyStoreStub) CurrentInputFileEnabled() bool { return true }
+func (claudeProxyStoreStub) CurrentInputFileMinChars() int { return 0 }

 type openAIProxyStub struct {
 	status int
--- a/internal/httpapi/claude/stream_runtime_finalize.go
+++ b/internal/httpapi/claude/stream_runtime_finalize.go
@@ -133,6 +133,9 @@ func (s *claudeStreamRuntime) finalize(stopReason string) {
 		ToolsRaw:              s.toolsRaw,
 	})
 	finalText := turn.Text
+	outcome := assistantturn.FinalizeTurn(turn, assistantturn.FinalizeOptions{
+		AlreadyEmittedToolCalls: s.toolCallsDetected,
+	})

 	if s.bufferToolContent && !s.toolCallsDetected {
 		if len(turn.ToolCalls) > 0 {
@@ -169,7 +172,7 @@ func (s *claudeStreamRuntime) finalize(stopReason string) {
 		}
 	}

-	if s.toolCallsDetected {
+	if outcome.HasToolCalls {
 		stopReason = "tool_use"
 	}

@@ -180,7 +183,7 @@ func (s *claudeStreamRuntime) finalize(stopReason string) {
 			"stop_sequence": nil,
 		},
 		"usage": map[string]any{
-			"output_tokens": turn.Usage.OutputTokens,
+			"output_tokens": outcome.Usage.OutputTokens,
 		},
 	})
 	s.send("message_stop", map[string]any{"type": "message_stop"})
--- a/internal/httpapi/claude/stream_status_test.go
+++ b/internal/httpapi/claude/stream_status_test.go
@@ -23,9 +23,8 @@ type streamStatusClaudeStoreStub struct{}

 func (streamStatusClaudeStoreStub) ModelAliases() map[string]string { return nil }

-func (streamStatusClaudeStoreStub) CompatStripReferenceMarkers() bool { return true }
-func (streamStatusClaudeStoreStub) CurrentInputFileEnabled() bool     { return true }
-func (streamStatusClaudeStoreStub) CurrentInputFileMinChars() int     { return 0 }
+func (streamStatusClaudeStoreStub) CurrentInputFileEnabled() bool { return true }
+func (streamStatusClaudeStoreStub) CurrentInputFileMinChars() int { return 0 }

 func captureClaudeStatusMiddleware(statuses *[]int) func(http.Handler) http.Handler {
 	return func(next http.Handler) http.Handler {
--- a/internal/httpapi/gemini/deps.go
+++ b/internal/httpapi/gemini/deps.go
@@ -23,7 +23,6 @@ type DeepSeekCaller interface {

 type ConfigReader interface {
 	ModelAliases() map[string]string
-	CompatStripReferenceMarkers() bool
 	CurrentInputFileEnabled() bool
 	CurrentInputFileMinChars() int
 }
--- a/internal/httpapi/gemini/handler_generate.go
+++ b/internal/httpapi/gemini/handler_generate.go
@@ -81,7 +81,7 @@ func (h *Handler) handleGeminiDirect(w http.ResponseWriter, r *http.Request, str
 		return true
 	}
 	result, outErr := completionruntime.ExecuteNonStreamWithRetry(r.Context(), h.DS, a, stdReq, completionruntime.Options{
-		StripReferenceMarkers: h.compatStripReferenceMarkers(),
+		StripReferenceMarkers: stripReferenceMarkersEnabled(),
 		RetryEnabled:          true,
 		CurrentInputFile:      h.Store,
 	})
@@ -294,7 +294,7 @@ func (h *Handler) handleNonStreamGenerateContent(w http.ResponseWriter, resp *ht
 	}

 	result := sse.CollectStream(resp, thinkingEnabled, true)
-	stripReferenceMarkers := h.compatStripReferenceMarkers()
+	stripReferenceMarkers := stripReferenceMarkersEnabled()
 	writeJSON(w, http.StatusOK, buildGeminiGenerateContentResponse(
 		model,
 		finalPrompt,
--- a/internal/httpapi/gemini/handler_routes.go
+++ b/internal/httpapi/gemini/handler_routes.go
@@ -18,11 +18,8 @@ type Handler struct {
 }

 //nolint:unused // used by native Gemini stream/non-stream runtime helpers.
-func (h *Handler) compatStripReferenceMarkers() bool {
-	if h == nil || h.Store == nil {
-		return true
-	}
-	return h.Store.CompatStripReferenceMarkers()
+func stripReferenceMarkersEnabled() bool {
+	return true
 }

 func RegisterRoutes(r chi.Router, h *Handler) {
--- a/internal/httpapi/gemini/handler_stream_runtime.go
+++ b/internal/httpapi/gemini/handler_stream_runtime.go
@@ -29,7 +29,7 @@ func (h *Handler) handleStreamGenerateContent(w http.ResponseWriter, r *http.Req

 	rc := http.NewResponseController(w)
 	_, canFlush := w.(http.Flusher)
-	runtime := newGeminiStreamRuntime(w, rc, canFlush, model, finalPrompt, thinkingEnabled, searchEnabled, h.compatStripReferenceMarkers(), toolNames, toolsRaw)
+	runtime := newGeminiStreamRuntime(w, rc, canFlush, model, finalPrompt, thinkingEnabled, searchEnabled, stripReferenceMarkersEnabled(), toolNames, toolsRaw)

 	initialType := "text"
 	if thinkingEnabled {
@@ -192,6 +192,7 @@ func (s *geminiStreamRuntime) finalize() {
 		ToolNames:             s.toolNames,
 		ToolsRaw:              s.toolsRaw,
 	})
+	outcome := assistantturn.FinalizeTurn(turn, assistantturn.FinalizeOptions{})

 	if s.bufferContent {
 		parts := buildGeminiPartsFromTurn(turn)
@@ -224,9 +225,9 @@ func (s *geminiStreamRuntime) finalize() {
 		},
 		"modelVersion": s.model,
 		"usageMetadata": map[string]any{
-			"promptTokenCount":     turn.Usage.InputTokens,
-			"candidatesTokenCount": turn.Usage.OutputTokens,
-			"totalTokenCount":      turn.Usage.TotalTokens,
+			"promptTokenCount":     outcome.Usage.InputTokens,
+			"candidatesTokenCount": outcome.Usage.OutputTokens,
+			"totalTokenCount":      outcome.Usage.TotalTokens,
 		},
 	})
 }
--- a/internal/httpapi/gemini/handler_test.go
+++ b/internal/httpapi/gemini/handler_test.go
@@ -18,10 +18,9 @@ import (

 type testGeminiConfig struct{}

-func (testGeminiConfig) ModelAliases() map[string]string   { return nil }
-func (testGeminiConfig) CompatStripReferenceMarkers() bool { return true }
-func (testGeminiConfig) CurrentInputFileEnabled() bool     { return true }
-func (testGeminiConfig) CurrentInputFileMinChars() int     { return 0 }
+func (testGeminiConfig) ModelAliases() map[string]string { return nil }
+func (testGeminiConfig) CurrentInputFileEnabled() bool   { return true }
+func (testGeminiConfig) CurrentInputFileMinChars() int   { return 0 }

 type testGeminiAuth struct {
 	a   *auth.RequestAuth
--- a/internal/httpapi/openai/chat/chat_history_test.go
+++ b/internal/httpapi/openai/chat/chat_history_test.go
@@ -57,7 +57,7 @@ func blockChatHistoryDetailDir(t *testing.T, detailDir string) func() {
 func TestChatCompletionsNonStreamPersistsHistory(t *testing.T) {
 	historyStore := newTestChatHistoryStore(t)
 	h := &Handler{
-		Store:       mockOpenAIConfig{wideInput: true},
+		Store:       mockOpenAIConfig{},
 		Auth:        streamStatusAuthStub{},
 		DS:          streamStatusDSStub{resp: makeOpenAISSEHTTPResponse(`data: {"p":"response/content","v":"hello world"}`, `data: [DONE]`)},
 		ChatHistory: historyStore,
@@ -216,7 +216,7 @@ func TestHandleStreamContextCancelledMarksHistoryStopped(t *testing.T) {
 func TestChatCompletionsSkipsAdminWebUISource(t *testing.T) {
 	historyStore := newTestChatHistoryStore(t)
 	h := &Handler{
-		Store:       mockOpenAIConfig{wideInput: true},
+		Store:       mockOpenAIConfig{},
 		Auth:        streamStatusAuthStub{},
 		DS:          streamStatusDSStub{resp: makeOpenAISSEHTTPResponse(`data: {"p":"response/content","v":"hello world"}`, `data: [DONE]`)},
 		ChatHistory: historyStore,
@@ -248,7 +248,7 @@ func TestChatCompletionsSkipsHistoryWhenDisabled(t *testing.T) {
 		t.Fatalf("disable history store failed: %v", err)
 	}
 	h := &Handler{
-		Store:       mockOpenAIConfig{wideInput: true},
+		Store:       mockOpenAIConfig{},
 		Auth:        streamStatusAuthStub{},
 		DS:          streamStatusDSStub{resp: makeOpenAISSEHTTPResponse(`data: {"p":"response/content","v":"hello world"}`, `data: [DONE]`)},
 		ChatHistory: historyStore,
@@ -278,7 +278,6 @@ func TestChatCompletionsCurrentInputFilePersistsNeutralPrompt(t *testing.T) {
 	ds := &inlineUploadDSStub{}
 	h := &Handler{
 		Store: mockOpenAIConfig{
-			wideInput:           true,
 			currentInputEnabled: true,
 		},
 		Auth:        streamStatusAuthStub{},
--- a/internal/httpapi/openai/chat/chat_stream_runtime.go
+++ b/internal/httpapi/openai/chat/chat_stream_runtime.go
@@ -230,7 +230,6 @@ func (s *chatStreamRuntime) finalize(finishReason string, deferEmptyOutput bool)
 	s.finalThinking = turn.Thinking
 	s.finalText = turn.Text
 	if len(turn.ToolCalls) > 0 && !s.toolCallsDoneEmitted {
-		finishReason = "tool_calls"
 		s.sendDelta(map[string]any{
 			"tool_calls": formatFinalStreamToolCallsWithStableIDs(turn.ToolCalls, s.streamToolCallIDs, s.toolsRaw),
 		})
@@ -241,7 +240,6 @@ func (s *chatStreamRuntime) finalize(finishReason string, deferEmptyOutput bool)
 		for _, evt := range toolstream.Flush(&s.toolSieve, s.toolNames) {
 			if len(evt.ToolCalls) > 0 {
 				batch.flush()
-				finishReason = "tool_calls"
 				s.toolCallsEmitted = true
 				s.toolCallsDoneEmitted = true
 				s.sendDelta(map[string]any{
@@ -261,14 +259,11 @@ func (s *chatStreamRuntime) finalize(finishReason string, deferEmptyOutput bool)
 		batch.flush()
 	}

-	if len(turn.ToolCalls) > 0 || s.toolCallsEmitted {
-		finishReason = "tool_calls"
-	}
-	if len(turn.ToolCalls) == 0 && !s.toolCallsEmitted && strings.TrimSpace(turn.Text) == "" {
-		status, message, code := upstreamEmptyOutputDetail(finishReason == "content_filter", turn.Text, turn.Thinking)
-		if turn.Error != nil {
-			status, message, code = turn.Error.Status, turn.Error.Message, turn.Error.Code
-		}
+	outcome := assistantturn.FinalizeTurn(turn, assistantturn.FinalizeOptions{
+		AlreadyEmittedToolCalls: s.toolCallsEmitted || s.toolCallsDoneEmitted,
+	})
+	if outcome.ShouldFail {
+		status, message, code := outcome.Error.Status, outcome.Error.Message, outcome.Error.Code
 		if deferEmptyOutput {
 			s.finalErrorStatus = status
 			s.finalErrorMessage = message
@@ -278,31 +273,20 @@ func (s *chatStreamRuntime) finalize(finishReason string, deferEmptyOutput bool)
 		s.sendFailedChunk(status, message, code)
 		return true
 	}
-	usage := chatUsageFromTurn(turn)
-	s.finalFinishReason = finishReason
+	usage := assistantturn.OpenAIChatUsage(turn)
+	s.finalFinishReason = outcome.FinishReason
 	s.finalUsage = usage
 	s.sendChunk(openaifmt.BuildChatStreamChunk(
 		s.completionID,
 		s.created,
 		s.model,
-		[]map[string]any{openaifmt.BuildChatStreamFinishChoice(0, finishReason)},
+		[]map[string]any{openaifmt.BuildChatStreamFinishChoice(0, outcome.FinishReason)},
 		usage,
 	))
 	s.sendDone()
 	return true
 }

-func chatUsageFromTurn(turn assistantturn.Turn) map[string]any {
-	return map[string]any{
-		"prompt_tokens":     turn.Usage.InputTokens,
-		"completion_tokens": turn.Usage.OutputTokens,
-		"total_tokens":      turn.Usage.TotalTokens,
-		"completion_tokens_details": map[string]any{
-			"reasoning_tokens": turn.Usage.ReasoningTokens,
-		},
-	}
-}
-
 func (s *chatStreamRuntime) onParsed(parsed sse.LineResult) streamengine.ParsedDecision {
 	if !parsed.Parsed {
 		return streamengine.ParsedDecision{}
--- a/internal/httpapi/openai/chat/empty_retry_runtime.go
+++ b/internal/httpapi/openai/chat/empty_retry_runtime.go
@@ -93,7 +93,7 @@ func (h *Handler) collectChatNonStreamAttempt(w http.ResponseWriter, resp *http.
 		Model:                 model,
 		Prompt:                usagePrompt,
 		SearchEnabled:         searchEnabled,
-		StripReferenceMarkers: h.compatStripReferenceMarkers(),
+		StripReferenceMarkers: stripReferenceMarkersEnabled(),
 		ToolNames:             toolNames,
 		ToolsRaw:              toolsRaw,
 	})
@@ -223,7 +223,7 @@ func (h *Handler) prepareChatStreamRuntime(w http.ResponseWriter, resp *http.Res
 	}
 	streamRuntime := newChatStreamRuntime(
 		w, rc, canFlush, completionID, time.Now().Unix(), model, finalPrompt,
-		thinkingEnabled, searchEnabled, h.compatStripReferenceMarkers(), toolNames, toolsRaw,
+		thinkingEnabled, searchEnabled, stripReferenceMarkersEnabled(), toolNames, toolsRaw,
 		toolChoice,
 		len(toolNames) > 0, h.toolcallFeatureMatchEnabled() && h.toolcallEarlyEmitHighConfidence(),
 	)
--- a/internal/httpapi/openai/chat/handler.go
+++ b/internal/httpapi/openai/chat/handler.go
@@ -35,11 +35,8 @@ type streamLease struct {
 	ExpiresAt time.Time
 }

-func (h *Handler) compatStripReferenceMarkers() bool {
-	if h == nil {
-		return true
-	}
-	return shared.CompatStripReferenceMarkers(h.Store)
+func stripReferenceMarkersEnabled() bool {
+	return true
 }

 func (h *Handler) applyCurrentInputFile(ctx context.Context, a *auth.RequestAuth, stdReq promptcompat.StandardRequest) (promptcompat.StandardRequest, error) {
@@ -108,22 +105,10 @@ func cleanVisibleOutput(text string, stripReferenceMarkers bool) string {
 	return shared.CleanVisibleOutput(text, stripReferenceMarkers)
 }

-func replaceCitationMarkersWithLinks(text string, links map[int]string) string {
-	return shared.ReplaceCitationMarkersWithLinks(text, links)
-}
-
-func shouldWriteUpstreamEmptyOutputError(text, thinking string) bool {
-	return shared.ShouldWriteUpstreamEmptyOutputError(text, thinking)
-}
-
 func upstreamEmptyOutputDetail(contentFilter bool, text, thinking string) (int, string, string) {
 	return shared.UpstreamEmptyOutputDetail(contentFilter, text, thinking)
 }

-func writeUpstreamEmptyOutputError(w http.ResponseWriter, text, thinking string, contentFilter bool) bool {
-	return shared.WriteUpstreamEmptyOutputError(w, text, thinking, contentFilter)
-}
-
 func emptyOutputRetryEnabled() bool {
 	return shared.EmptyOutputRetryEnabled()
 }
--- a/internal/httpapi/openai/chat/handler_chat.go
+++ b/internal/httpapi/openai/chat/handler_chat.go
@@ -8,6 +8,7 @@ import (
 	"strings"
 	"time"

+	"ds2api/internal/assistantturn"
 	"ds2api/internal/auth"
 	"ds2api/internal/completionruntime"
 	"ds2api/internal/config"
@@ -79,7 +80,7 @@ func (h *Handler) ChatCompletions(w http.ResponseWriter, r *http.Request) {

 	if !stdReq.Stream {
 		result, outErr := completionruntime.ExecuteNonStreamWithRetry(r.Context(), h.DS, a, stdReq, completionruntime.Options{
-			StripReferenceMarkers: h.compatStripReferenceMarkers(),
+			StripReferenceMarkers: stripReferenceMarkersEnabled(),
 			RetryEnabled:          true,
 			CurrentInputFile:      h.Store,
 		})
@@ -92,10 +93,10 @@ func (h *Handler) ChatCompletions(w http.ResponseWriter, r *http.Request) {
 			return
 		}
 		respBody := openaifmt.BuildChatCompletionWithToolCalls(result.SessionID, stdReq.ResponseModel, result.Turn.Prompt, result.Turn.Thinking, result.Turn.Text, result.Turn.ToolCalls, stdReq.ToolsRaw)
-		respBody["usage"] = chatUsageFromTurn(result.Turn)
-		finishReason := chatFinishReason(respBody)
+		respBody["usage"] = assistantturn.OpenAIChatUsage(result.Turn)
+		finishReason := assistantturn.FinalizeTurn(result.Turn, assistantturn.FinalizeOptions{}).FinishReason
 		if historySession != nil {
-			historySession.success(http.StatusOK, result.Turn.Thinking, result.Turn.Text, finishReason, chatUsageFromTurn(result.Turn))
+			historySession.success(http.StatusOK, result.Turn.Thinking, result.Turn.Text, finishReason, assistantturn.OpenAIChatUsage(result.Turn))
 		}
 		writeJSON(w, http.StatusOK, respBody)
 		return
@@ -162,33 +163,29 @@ func (h *Handler) handleNonStream(w http.ResponseWriter, resp *http.Response, co
 	}
 	result := sse.CollectStream(resp, thinkingEnabled, true)

-	stripReferenceMarkers := h.compatStripReferenceMarkers()
-	finalThinking := cleanVisibleOutput(result.Thinking, stripReferenceMarkers)
-	finalText := cleanVisibleOutput(result.Text, stripReferenceMarkers)
-	if searchEnabled {
-		finalText = replaceCitationMarkersWithLinks(finalText, result.CitationLinks)
-	}
-	detected := detectAssistantToolCalls(result.Text, finalText, result.Thinking, result.ToolDetectionThinking, toolNames)
-	if shouldWriteUpstreamEmptyOutputError(finalText, finalThinking) && len(detected.Calls) == 0 {
-		status, message, code := upstreamEmptyOutputDetail(result.ContentFilter, finalText, finalThinking)
+	turn := assistantturn.BuildTurnFromCollected(result, assistantturn.BuildOptions{
+		Model:                 model,
+		Prompt:                finalPrompt,
+		RefFileTokens:         refFileTokens,
+		SearchEnabled:         searchEnabled,
+		StripReferenceMarkers: stripReferenceMarkersEnabled(),
+		ToolNames:             toolNames,
+		ToolsRaw:              toolsRaw,
+		ToolChoice:            promptcompat.DefaultToolChoicePolicy(),
+	})
+	outcome := assistantturn.FinalizeTurn(turn, assistantturn.FinalizeOptions{})
+	if outcome.ShouldFail {
+		status, message, code := outcome.Error.Status, outcome.Error.Message, outcome.Error.Code
 		if historySession != nil {
-			historySession.error(status, message, code, finalThinking, finalText)
+			historySession.error(status, message, code, turn.Thinking, turn.Text)
 		}
-		writeUpstreamEmptyOutputError(w, finalText, finalThinking, result.ContentFilter)
+		writeOpenAIErrorWithCode(w, status, message, code)
 		return
 	}
-	respBody := openaifmt.BuildChatCompletionWithToolCalls(completionID, model, finalPrompt, finalThinking, finalText, detected.Calls, toolsRaw)
-	if refFileTokens > 0 {
-		addRefFileTokensToUsage(respBody, refFileTokens)
-	}
-	finishReason := "stop"
-	if choices, ok := respBody["choices"].([]map[string]any); ok && len(choices) > 0 {
-		if fr, _ := choices[0]["finish_reason"].(string); strings.TrimSpace(fr) != "" {
-			finishReason = fr
-		}
-	}
+	respBody := openaifmt.BuildChatCompletionWithToolCalls(completionID, model, finalPrompt, turn.Thinking, turn.Text, turn.ToolCalls, toolsRaw)
+	respBody["usage"] = assistantturn.OpenAIChatUsage(turn)
 	if historySession != nil {
-		historySession.success(http.StatusOK, finalThinking, finalText, finishReason, openaifmt.BuildChatUsageForModel(model, finalPrompt, finalThinking, finalText, refFileTokens))
+		historySession.success(http.StatusOK, turn.Thinking, turn.Text, outcome.FinishReason, assistantturn.OpenAIChatUsage(turn))
 	}
 	writeJSON(w, http.StatusOK, respBody)
 }
@@ -216,7 +213,7 @@ func (h *Handler) handleStream(w http.ResponseWriter, r *http.Request, resp *htt
 	created := time.Now().Unix()
 	bufferToolContent := len(toolNames) > 0
 	emitEarlyToolDeltas := h.toolcallFeatureMatchEnabled() && h.toolcallEarlyEmitHighConfidence()
-	stripReferenceMarkers := h.compatStripReferenceMarkers()
+	stripReferenceMarkers := stripReferenceMarkersEnabled()
 	initialType := "text"
 	if thinkingEnabled {
 		initialType = "thinking"
--- a/internal/httpapi/openai/chat/handler_chat_auto_delete_test.go
+++ b/internal/httpapi/openai/chat/handler_chat_auto_delete_test.go
@@ -75,7 +75,6 @@ func TestChatCompletionsAutoDeleteModes(t *testing.T) {
 			}
 			h := &Handler{
 				Store: mockOpenAIConfig{
-					wideInput:      true,
 					autoDeleteMode: tc.mode,
 				},
 				Auth: streamStatusAuthStub{},
@@ -123,7 +122,6 @@ func TestAutoDeleteRemoteSessionIgnoresCanceledParentContext(t *testing.T) {
 	ds := &autoDeleteCtxDSStub{}
 	h := &Handler{
 		Store: mockOpenAIConfig{
-			wideInput:      true,
 			autoDeleteMode: "single",
 		},
 		DS: ds,
--- a/internal/httpapi/openai/chat/handler_toolcall_test.go
+++ b/internal/httpapi/openai/chat/handler_toolcall_test.go
@@ -133,18 +133,13 @@ func TestHandleNonStreamReturns429WhenUpstreamHasOnlyThinking(t *testing.T) {
 	rec := httptest.NewRecorder()

 	h.handleNonStream(rec, resp, "cid-thinking-only", "deepseek-v4-pro", "prompt", 0, true, false, nil, nil, nil)
-	if rec.Code != http.StatusOK {
-		t.Fatalf("expected status 200 for thinking-only upstream output, got %d body=%s", rec.Code, rec.Body.String())
+	if rec.Code != http.StatusTooManyRequests {
+		t.Fatalf("expected status 429 for thinking-only upstream output, got %d body=%s", rec.Code, rec.Body.String())
 	}
 	out := decodeJSONBody(t, rec.Body.String())
-	choices, _ := out["choices"].([]any)
-	if len(choices) == 0 {
-		t.Fatal("expected at least one choice")
-	}
-	first, _ := choices[0].(map[string]any)
-	msg, _ := first["message"].(map[string]any)
-	if asString(msg["reasoning_content"]) != "Only thinking" {
-		t.Fatalf("expected reasoning_content='Only thinking', got %#v", msg)
+	errObj, _ := out["error"].(map[string]any)
+	if asString(errObj["code"]) != "upstream_empty_output" {
+		t.Fatalf("expected code=upstream_empty_output, got %#v", out)
 	}
 }

--- a/internal/httpapi/openai/chat/test_helpers_test.go
+++ b/internal/httpapi/openai/chat/test_helpers_test.go
@@ -12,7 +12,6 @@ import (

 type mockOpenAIConfig struct {
 	aliases             map[string]string
-	wideInput           bool
 	autoDeleteMode      string
 	toolMode            string
 	earlyEmit           string
@@ -24,11 +23,7 @@ type mockOpenAIConfig struct {
 	thinkingPrompt      string
 }

-func (m mockOpenAIConfig) ModelAliases() map[string]string { return m.aliases }
-func (m mockOpenAIConfig) CompatWideInputStrictOutput() bool {
-	return m.wideInput
-}
-func (m mockOpenAIConfig) CompatStripReferenceMarkers() bool   { return true }
+func (m mockOpenAIConfig) ModelAliases() map[string]string     { return m.aliases }
 func (m mockOpenAIConfig) ToolcallMode() string                { return m.toolMode }
 func (m mockOpenAIConfig) ToolcallEarlyEmitConfidence() string { return m.earlyEmit }
 func (m mockOpenAIConfig) ResponsesStoreTTLSeconds() int       { return m.responsesTTL }
--- a/internal/httpapi/openai/chat/vercel_prepare_test.go
+++ b/internal/httpapi/openai/chat/vercel_prepare_test.go
@@ -94,7 +94,6 @@ func TestHandleVercelStreamPrepareAppliesCurrentInputFile(t *testing.T) {
 	ds := &inlineUploadDSStub{}
 	h := &Handler{
 		Store: mockOpenAIConfig{
-			wideInput:           true,
 			currentInputEnabled: true,
 		},
 		Auth: streamStatusAuthStub{},
@@ -151,7 +150,6 @@ func TestHandleVercelStreamPrepareMapsCurrentInputFileManagedAuthFailureTo401(t
 	}
 	h := &Handler{
 		Store: mockOpenAIConfig{
-			wideInput:           true,
 			currentInputEnabled: true,
 		},
 		Auth: streamStatusManagedAuthStub{},
--- a/internal/httpapi/openai/chat/vercel_stream.go
+++ b/internal/httpapi/openai/chat/vercel_stream.go
@@ -109,13 +109,10 @@ func (h *Handler) handleVercelStreamPrepare(w http.ResponseWriter, r *http.Reque
 		"final_prompt":     stdReq.FinalPrompt,
 		"thinking_enabled": stdReq.Thinking,
 		"search_enabled":   stdReq.Search,
-		"compat": map[string]any{
-			"strip_reference_markers": h.compatStripReferenceMarkers(),
-		},
-		"tool_names":     stdReq.ToolNames,
-		"deepseek_token": a.DeepSeekToken,
-		"pow_header":     powHeader,
-		"payload":        payload,
+		"tool_names":       stdReq.ToolNames,
+		"deepseek_token":   a.DeepSeekToken,
+		"pow_header":       powHeader,
+		"payload":          payload,
 	})
 }

--- a/internal/httpapi/openai/deps_injection_test.go
+++ b/internal/httpapi/openai/deps_injection_test.go
@@ -1,6 +1,7 @@
 package openai

 import (
+	"strings"
 	"testing"

 	"ds2api/internal/promptcompat"
@@ -8,7 +9,6 @@ import (

 type mockOpenAIConfig struct {
 	aliases             map[string]string
-	wideInput           bool
 	autoDeleteMode      string
 	toolMode            string
 	earlyEmit           string
@@ -20,11 +20,7 @@ type mockOpenAIConfig struct {
 	thinkingPrompt      string
 }

-func (m mockOpenAIConfig) ModelAliases() map[string]string { return m.aliases }
-func (m mockOpenAIConfig) CompatWideInputStrictOutput() bool {
-	return m.wideInput
-}
-func (m mockOpenAIConfig) CompatStripReferenceMarkers() bool   { return true }
+func (m mockOpenAIConfig) ModelAliases() map[string]string     { return m.aliases }
 func (m mockOpenAIConfig) ToolcallMode() string                { return m.toolMode }
 func (m mockOpenAIConfig) ToolcallEarlyEmitConfidence() string { return m.earlyEmit }
 func (m mockOpenAIConfig) ResponsesStoreTTLSeconds() int       { return m.responsesTTL }
@@ -53,7 +49,6 @@ func TestNormalizeOpenAIChatRequestWithConfigInterface(t *testing.T) {
 		aliases: map[string]string{
 			"my-model": "deepseek-v4-flash-search",
 		},
-		wideInput: true,
 	}
 	req := map[string]any{
 		"model":    "my-model",
@@ -72,7 +67,7 @@ func TestNormalizeOpenAIChatRequestWithConfigInterface(t *testing.T) {
 }

 func TestNormalizeOpenAIChatRequestDisablesThinkingForNoThinkingModel(t *testing.T) {
-	cfg := mockOpenAIConfig{wideInput: true}
+	cfg := mockOpenAIConfig{}
 	req := map[string]any{
 		"model":            "deepseek-v4-pro-nothinking",
 		"messages":         []any{map[string]any{"role": "user", "content": "hello"}},
@@ -93,28 +88,22 @@ func TestNormalizeOpenAIChatRequestDisablesThinkingForNoThinkingModel(t *testing
 	}
 }

-func TestNormalizeOpenAIResponsesRequestWideInputPolicyFromInterface(t *testing.T) {
+func TestNormalizeOpenAIResponsesRequestAlwaysAcceptsWideInput(t *testing.T) {
 	req := map[string]any{
 		"model": "deepseek-v4-flash",
 		"input": "hi",
 	}

-	_, err := promptcompat.NormalizeOpenAIResponsesRequest(mockOpenAIConfig{
-		aliases:   map[string]string{},
-		wideInput: false,
-	}, req, "")
-	if err == nil {
-		t.Fatal("expected error when wide input is disabled and only input is provided")
-	}
-
 	out, err := promptcompat.NormalizeOpenAIResponsesRequest(mockOpenAIConfig{
-		aliases:   map[string]string{},
-		wideInput: true,
+		aliases: map[string]string{},
 	}, req, "")
 	if err != nil {
-		t.Fatalf("unexpected error when wide input is enabled: %v", err)
+		t.Fatalf("unexpected error for wide input request: %v", err)
 	}
 	if out.Surface != "openai_responses" {
 		t.Fatalf("unexpected surface: %q", out.Surface)
 	}
+	if !strings.Contains(out.FinalPrompt, "<｜User｜>hi") {
+		t.Fatalf("unexpected final prompt: %q", out.FinalPrompt)
+	}
 }
--- a/internal/httpapi/openai/file_inline_upload_test.go
+++ b/internal/httpapi/openai/file_inline_upload_test.go
@@ -151,7 +151,7 @@ func TestPreprocessInlineFileInputsDeduplicatesIdenticalPayloads(t *testing.T) {

 func TestChatCompletionsUploadsInlineFilesBeforeCompletion(t *testing.T) {
 	ds := &inlineUploadDSStub{}
-	h := &openAITestSurface{Store: mockOpenAIConfig{wideInput: true}, Auth: streamStatusAuthStub{}, DS: ds}
+	h := &openAITestSurface{Store: mockOpenAIConfig{}, Auth: streamStatusAuthStub{}, DS: ds}
 	reqBody := `{"model":"deepseek-v4-vision","messages":[{"role":"user","content":[{"type":"input_text","text":"hi"},{"type":"image_url","image_url":{"url":"data:image/png;base64,QUJDRA=="}}]}],"stream":false}`
 	req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", strings.NewReader(reqBody))
 	req.Header.Set("Authorization", "Bearer direct-token")
@@ -180,7 +180,7 @@ func TestChatCompletionsUploadsInlineFilesBeforeCompletion(t *testing.T) {

 func TestResponsesUploadsInlineFilesBeforeCompletion(t *testing.T) {
 	ds := &inlineUploadDSStub{}
-	h := &openAITestSurface{Store: mockOpenAIConfig{wideInput: true}, Auth: streamStatusAuthStub{}, DS: ds}
+	h := &openAITestSurface{Store: mockOpenAIConfig{}, Auth: streamStatusAuthStub{}, DS: ds}
 	r := chi.NewRouter()
 	registerOpenAITestRoutes(r, h)
 	reqBody := `{"model":"deepseek-v4-pro","input":[{"role":"user","content":[{"type":"input_text","text":"hi"},{"type":"input_image","image_url":{"url":"data:image/png;base64,QUJDRA=="}}]}],"stream":false}`
@@ -208,7 +208,7 @@ func TestResponsesUploadsInlineFilesBeforeCompletion(t *testing.T) {

 func TestChatCompletionsInlineUploadFailureReturnsBadRequest(t *testing.T) {
 	ds := &inlineUploadDSStub{}
-	h := &openAITestSurface{Store: mockOpenAIConfig{wideInput: true}, Auth: streamStatusAuthStub{}, DS: ds}
+	h := &openAITestSurface{Store: mockOpenAIConfig{}, Auth: streamStatusAuthStub{}, DS: ds}
 	reqBody := `{"model":"deepseek-v4-flash","messages":[{"role":"user","content":[{"type":"image_url","image_url":{"url":"data:image/png;base64,%%%"}}]}],"stream":false}`
 	req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", strings.NewReader(reqBody))
 	req.Header.Set("Authorization", "Bearer direct-token")
@@ -227,7 +227,7 @@ func TestChatCompletionsInlineUploadFailureReturnsBadRequest(t *testing.T) {

 func TestChatCompletionsInlineUploadLimitReturnsBadRequest(t *testing.T) {
 	ds := &inlineUploadDSStub{}
-	h := &openAITestSurface{Store: mockOpenAIConfig{wideInput: true}, Auth: streamStatusAuthStub{}, DS: ds}
+	h := &openAITestSurface{Store: mockOpenAIConfig{}, Auth: streamStatusAuthStub{}, DS: ds}
 	content := []any{map[string]any{"type": "input_text", "text": "hi"}}
 	for i := 0; i < 51; i++ {
 		content = append(content, map[string]any{
@@ -266,7 +266,7 @@ func TestChatCompletionsInlineUploadLimitReturnsBadRequest(t *testing.T) {

 func TestResponsesInlineUploadFailureReturnsInternalServerError(t *testing.T) {
 	ds := &inlineUploadDSStub{uploadErr: errors.New("boom")}
-	h := &openAITestSurface{Store: mockOpenAIConfig{wideInput: true}, Auth: streamStatusAuthStub{}, DS: ds}
+	h := &openAITestSurface{Store: mockOpenAIConfig{}, Auth: streamStatusAuthStub{}, DS: ds}
 	r := chi.NewRouter()
 	registerOpenAITestRoutes(r, h)
 	reqBody := `{"model":"deepseek-v4-flash","input":[{"role":"user","content":[{"type":"image_url","image_url":{"url":"data:image/png;base64,QUJDRA=="}}]}],"stream":false}`
@@ -289,7 +289,7 @@ func TestVercelPrepareUploadsInlineFilesBeforeLeasePayload(t *testing.T) {
 	t.Setenv("VERCEL", "1")
 	t.Setenv("DS2API_VERCEL_INTERNAL_SECRET", "stream-secret")
 	ds := &inlineUploadDSStub{}
-	h := &openAITestSurface{Store: mockOpenAIConfig{wideInput: true}, Auth: streamStatusAuthStub{}, DS: ds}
+	h := &openAITestSurface{Store: mockOpenAIConfig{}, Auth: streamStatusAuthStub{}, DS: ds}
 	r := chi.NewRouter()
 	registerOpenAITestRoutes(r, h)
 	reqBody := `{"model":"deepseek-v4-flash","messages":[{"role":"user","content":[{"type":"input_text","text":"hi"},{"type":"image_url","image_url":{"url":"data:image/png;base64,QUJDRA=="}}]}],"stream":true}`
--- a/internal/httpapi/openai/files_route_test.go
+++ b/internal/httpapi/openai/files_route_test.go
@@ -120,7 +120,7 @@ func newMultipartUploadRequest(t *testing.T, purpose string, filename string, da

 func TestFilesRouteUploadSuccess(t *testing.T) {
 	ds := &filesRouteDSStub{}
-	h := &openAITestSurface{Store: mockOpenAIConfig{wideInput: true}, Auth: streamStatusAuthStub{}, DS: ds}
+	h := &openAITestSurface{Store: mockOpenAIConfig{}, Auth: streamStatusAuthStub{}, DS: ds}
 	r := chi.NewRouter()
 	registerOpenAITestRoutes(r, h)

@@ -160,7 +160,7 @@ func TestFilesRouteUploadSuccess(t *testing.T) {

 func TestFilesRouteUploadIncludesAccountIDForManagedAccount(t *testing.T) {
 	ds := &filesRouteDSStub{}
-	h := &openAITestSurface{Store: mockOpenAIConfig{wideInput: true}, Auth: managedFilesAuthStub{}, DS: ds}
+	h := &openAITestSurface{Store: mockOpenAIConfig{}, Auth: managedFilesAuthStub{}, DS: ds}
 	r := chi.NewRouter()
 	registerOpenAITestRoutes(r, h)

@@ -188,7 +188,7 @@ func TestFilesRouteRetrieveSuccess(t *testing.T) {
 		Purpose:  "assistants",
 		Status:   "processed",
 	}}
-	h := &openAITestSurface{Store: mockOpenAIConfig{wideInput: true}, Auth: managedFilesAuthStub{}, DS: ds}
+	h := &openAITestSurface{Store: mockOpenAIConfig{}, Auth: managedFilesAuthStub{}, DS: ds}
 	r := chi.NewRouter()
 	registerOpenAITestRoutes(r, h)

@@ -214,7 +214,7 @@ func TestFilesRouteRetrieveSuccess(t *testing.T) {

 func TestFilesRouteRetrieveNotFound(t *testing.T) {
 	ds := &filesRouteDSStub{err: dsclient.ErrUploadFileNotFound}
-	h := &openAITestSurface{Store: mockOpenAIConfig{wideInput: true}, Auth: streamStatusAuthStub{}, DS: ds}
+	h := &openAITestSurface{Store: mockOpenAIConfig{}, Auth: streamStatusAuthStub{}, DS: ds}
 	r := chi.NewRouter()
 	registerOpenAITestRoutes(r, h)

@@ -229,7 +229,7 @@ func TestFilesRouteRetrieveNotFound(t *testing.T) {
 }

 func TestFilesRouteRejectsNonMultipart(t *testing.T) {
-	h := &openAITestSurface{Store: mockOpenAIConfig{wideInput: true}, Auth: streamStatusAuthStub{}, DS: &filesRouteDSStub{}}
+	h := &openAITestSurface{Store: mockOpenAIConfig{}, Auth: streamStatusAuthStub{}, DS: &filesRouteDSStub{}}
 	r := chi.NewRouter()
 	registerOpenAITestRoutes(r, h)

@@ -245,7 +245,7 @@ func TestFilesRouteRejectsNonMultipart(t *testing.T) {
 }

 func TestFilesRouteRequiresFileField(t *testing.T) {
-	h := &openAITestSurface{Store: mockOpenAIConfig{wideInput: true}, Auth: streamStatusAuthStub{}, DS: &filesRouteDSStub{}}
+	h := &openAITestSurface{Store: mockOpenAIConfig{}, Auth: streamStatusAuthStub{}, DS: &filesRouteDSStub{}}
 	r := chi.NewRouter()
 	registerOpenAITestRoutes(r, h)

--- a/internal/httpapi/openai/history_split_test.go
+++ b/internal/httpapi/openai/history_split_test.go
@@ -96,7 +96,6 @@ func TestApplyCurrentInputFileSkipsShortInputWhenThresholdNotReached(t *testing.
 	ds := &inlineUploadDSStub{}
 	h := &openAITestSurface{
 		Store: mockOpenAIConfig{
-			wideInput:           true,
 			currentInputEnabled: true,
 			currentInputMin:     10,
 		},
@@ -129,7 +128,6 @@ func TestApplyThinkingInjectionAppendsLatestUserPrompt(t *testing.T) {
 	ds := &inlineUploadDSStub{}
 	h := &openAITestSurface{
 		Store: mockOpenAIConfig{
-			wideInput:         true,
 			thinkingInjection: boolPtr(true),
 		},
 		DS: ds,
@@ -161,7 +159,6 @@ func TestApplyThinkingInjectionUsesCustomPrompt(t *testing.T) {
 	ds := &inlineUploadDSStub{}
 	h := &openAITestSurface{
 		Store: mockOpenAIConfig{
-			wideInput:         true,
 			thinkingInjection: boolPtr(true),
 			thinkingPrompt:    "custom thinking format",
 		},
@@ -191,7 +188,6 @@ func TestApplyCurrentInputFileDisabledPassThrough(t *testing.T) {
 	ds := &inlineUploadDSStub{}
 	h := &openAITestSurface{
 		Store: mockOpenAIConfig{
-			wideInput:           true,
 			currentInputEnabled: false,
 		},
 		DS: ds,
@@ -224,7 +220,6 @@ func TestApplyCurrentInputFileUploadsFirstTurnWithNumberedHistoryTranscript(t *t
 	ds := &inlineUploadDSStub{}
 	h := &openAITestSurface{
 		Store: mockOpenAIConfig{
-			wideInput:           true,
 			currentInputEnabled: true,
 			currentInputMin:     10,
 			thinkingInjection:   boolPtr(true),
@@ -294,7 +289,6 @@ func TestApplyCurrentInputFilePreservesFullContextPromptForTokenCounting(t *test
 	ds := &inlineUploadDSStub{}
 	h := &openAITestSurface{
 		Store: mockOpenAIConfig{
-			wideInput:           true,
 			currentInputEnabled: true,
 			currentInputMin:     0,
 			thinkingInjection:   boolPtr(true),
@@ -340,7 +334,6 @@ func TestApplyCurrentInputFileUploadsFullContextFile(t *testing.T) {
 	ds := &inlineUploadDSStub{}
 	h := &openAITestSurface{
 		Store: mockOpenAIConfig{
-			wideInput:           true,
 			currentInputEnabled: true,
 			currentInputMin:     0,
 			thinkingInjection:   boolPtr(true),
@@ -391,7 +384,6 @@ func TestApplyCurrentInputFileCarriesHistoryText(t *testing.T) {
 	ds := &inlineUploadDSStub{}
 	h := &openAITestSurface{
 		Store: mockOpenAIConfig{
-			wideInput:           true,
 			currentInputEnabled: true,
 		},
 		DS: ds,
@@ -424,7 +416,6 @@ func TestChatCompletionsCurrentInputFileUploadsContextAndKeepsNeutralPrompt(t *t
 	ds := &inlineUploadDSStub{}
 	h := &openAITestSurface{
 		Store: mockOpenAIConfig{
-			wideInput:           true,
 			currentInputEnabled: true,
 		},
 		Auth: streamStatusAuthStub{},
@@ -495,7 +486,6 @@ func TestResponsesCurrentInputFileUploadsContextAndKeepsNeutralPrompt(t *testing
 	ds := &inlineUploadDSStub{}
 	h := &openAITestSurface{
 		Store: mockOpenAIConfig{
-			wideInput:           true,
 			currentInputEnabled: true,
 		},
 		Auth: streamStatusAuthStub{},
@@ -553,7 +543,6 @@ func TestChatCompletionsCurrentInputFileMapsManagedAuthFailureTo401(t *testing.T
 	}
 	h := &openAITestSurface{
 		Store: mockOpenAIConfig{
-			wideInput:           true,
 			currentInputEnabled: true,
 		},
 		Auth: streamStatusManagedAuthStub{},
@@ -585,7 +574,6 @@ func TestResponsesCurrentInputFileMapsDirectAuthFailureTo401(t *testing.T) {
 	}
 	h := &openAITestSurface{
 		Store: mockOpenAIConfig{
-			wideInput:           true,
 			currentInputEnabled: true,
 		},
 		Auth: streamStatusAuthStub{},
@@ -617,7 +605,6 @@ func TestChatCompletionsCurrentInputFileUploadFailureReturnsInternalServerError(
 	ds := &inlineUploadDSStub{uploadErr: errors.New("boom")}
 	h := &openAITestSurface{
 		Store: mockOpenAIConfig{
-			wideInput:           true,
 			currentInputEnabled: true,
 		},
 		Auth: streamStatusAuthStub{},
@@ -646,7 +633,6 @@ func TestCurrentInputFileWorksAcrossAutoDeleteModes(t *testing.T) {
 			ds := &inlineUploadDSStub{}
 			h := &openAITestSurface{
 				Store: mockOpenAIConfig{
-					wideInput:           true,
 					autoDeleteMode:      mode,
 					currentInputEnabled: true,
 				},
--- a/internal/httpapi/openai/responses/empty_retry_runtime.go
+++ b/internal/httpapi/openai/responses/empty_retry_runtime.go
@@ -74,7 +74,7 @@ func (h *Handler) prepareResponsesStreamRuntime(w http.ResponseWriter, resp *htt
 	}
 	streamRuntime := newResponsesStreamRuntime(
 		w, rc, canFlush, responseID, model, finalPrompt, thinkingEnabled, searchEnabled,
-		h.compatStripReferenceMarkers(), toolNames, toolsRaw, len(toolNames) > 0,
+		stripReferenceMarkersEnabled(), toolNames, toolsRaw, len(toolNames) > 0,
 		h.toolcallFeatureMatchEnabled() && h.toolcallEarlyEmitHighConfidence(),
 		toolChoice, traceID, func(obj map[string]any) {
 			h.getResponseStore().put(owner, responseID, obj)
--- a/internal/httpapi/openai/responses/handler.go
+++ b/internal/httpapi/openai/responses/handler.go
@@ -11,7 +11,6 @@ import (
 	"ds2api/internal/httpapi/openai/history"
 	"ds2api/internal/httpapi/openai/shared"
 	"ds2api/internal/promptcompat"
-	"ds2api/internal/toolcall"
 	"ds2api/internal/toolstream"
 )

@@ -29,11 +28,8 @@ type Handler struct {
 	responses   *responseStore
 }

-func (h *Handler) compatStripReferenceMarkers() bool {
-	if h == nil {
-		return true
-	}
-	return shared.CompatStripReferenceMarkers(h.Store)
+func stripReferenceMarkersEnabled() bool {
+	return true
 }

 func (h *Handler) applyCurrentInputFile(ctx context.Context, a *auth.RequestAuth, stdReq promptcompat.StandardRequest) (promptcompat.StandardRequest, error) {
@@ -98,18 +94,6 @@ func cleanVisibleOutput(text string, stripReferenceMarkers bool) string {
 	return shared.CleanVisibleOutput(text, stripReferenceMarkers)
 }

-func replaceCitationMarkersWithLinks(text string, links map[int]string) string {
-	return shared.ReplaceCitationMarkersWithLinks(text, links)
-}
-
-func upstreamEmptyOutputDetail(contentFilter bool, text, thinking string) (int, string, string) {
-	return shared.UpstreamEmptyOutputDetail(contentFilter, text, thinking)
-}
-
-func writeUpstreamEmptyOutputError(w http.ResponseWriter, text, thinking string, contentFilter bool) bool {
-	return shared.WriteUpstreamEmptyOutputError(w, text, thinking, contentFilter)
-}
-
 func emptyOutputRetryEnabled() bool {
 	return shared.EmptyOutputRetryEnabled()
 }
@@ -129,7 +113,3 @@ func usagePromptWithEmptyOutputRetry(originalPrompt string, retryAttempts int) s
 func filterIncrementalToolCallDeltasByAllowed(deltas []toolstream.ToolCallDelta, seenNames map[int]string) []toolstream.ToolCallDelta {
 	return shared.FilterIncrementalToolCallDeltasByAllowed(deltas, seenNames)
 }
-
-func detectAssistantToolCalls(rawText, visibleText, exposedThinking, detectionThinking string, toolNames []string) toolcall.ToolCallParseResult {
-	return shared.DetectAssistantToolCalls(rawText, visibleText, exposedThinking, detectionThinking, toolNames)
-}
--- a/internal/httpapi/openai/responses/responses_handler.go
+++ b/internal/httpapi/openai/responses/responses_handler.go
@@ -11,6 +11,7 @@ import (
 	"github.com/go-chi/chi/v5"
 	"github.com/google/uuid"

+	"ds2api/internal/assistantturn"
 	"ds2api/internal/auth"
 	"ds2api/internal/completionruntime"
 	"ds2api/internal/config"
@@ -96,7 +97,7 @@ func (h *Handler) Responses(w http.ResponseWriter, r *http.Request) {
 	responseID := "resp_" + strings.ReplaceAll(uuid.NewString(), "-", "")
 	if !stdReq.Stream {
 		result, outErr := completionruntime.ExecuteNonStreamWithRetry(r.Context(), h.DS, a, stdReq, completionruntime.Options{
-			StripReferenceMarkers: h.compatStripReferenceMarkers(),
+			StripReferenceMarkers: stripReferenceMarkersEnabled(),
 			RetryEnabled:          true,
 			CurrentInputFile:      h.Store,
 		})
@@ -105,7 +106,7 @@ func (h *Handler) Responses(w http.ResponseWriter, r *http.Request) {
 			return
 		}
 		responseObj := openaifmt.BuildResponseObjectWithToolCalls(responseID, stdReq.ResponseModel, result.Turn.Prompt, result.Turn.Thinking, result.Turn.Text, result.Turn.ToolCalls, stdReq.ToolsRaw)
-		responseObj["usage"] = responsesUsageFromTurn(result.Turn)
+		responseObj["usage"] = assistantturn.OpenAIResponsesUsage(result.Turn)
 		h.getResponseStore().put(owner, responseID, responseObj)
 		writeJSON(w, http.StatusOK, responseObj)
 		return
@@ -132,28 +133,26 @@ func (h *Handler) handleResponsesNonStream(w http.ResponseWriter, resp *http.Res
 		return
 	}
 	result := sse.CollectStream(resp, thinkingEnabled, true)
-	stripReferenceMarkers := h.compatStripReferenceMarkers()
-	sanitizedThinking := cleanVisibleOutput(result.Thinking, stripReferenceMarkers)
-	sanitizedText := cleanVisibleOutput(result.Text, stripReferenceMarkers)
-	if searchEnabled {
-		sanitizedText = replaceCitationMarkersWithLinks(sanitizedText, result.CitationLinks)
-	}
-	textParsed := detectAssistantToolCalls(result.Text, sanitizedText, result.Thinking, result.ToolDetectionThinking, toolNames)
-	if len(textParsed.Calls) == 0 && writeUpstreamEmptyOutputError(w, sanitizedText, sanitizedThinking, result.ContentFilter) {
-		return
-	}
-	logResponsesToolPolicyRejection(traceID, toolChoice, textParsed, "text")

-	callCount := len(textParsed.Calls)
-	if toolChoice.IsRequired() && callCount == 0 {
-		writeOpenAIErrorWithCode(w, http.StatusUnprocessableEntity, "tool_choice requires at least one valid tool call.", "tool_choice_violation")
+	turn := assistantturn.BuildTurnFromCollected(result, assistantturn.BuildOptions{
+		Model:                 model,
+		Prompt:                finalPrompt,
+		RefFileTokens:         refFileTokens,
+		SearchEnabled:         searchEnabled,
+		StripReferenceMarkers: stripReferenceMarkersEnabled(),
+		ToolNames:             toolNames,
+		ToolsRaw:              toolsRaw,
+		ToolChoice:            toolChoice,
+	})
+	logResponsesToolPolicyRejection(traceID, toolChoice, turn.ParsedToolCalls, "text")
+	outcome := assistantturn.FinalizeTurn(turn, assistantturn.FinalizeOptions{})
+	if outcome.ShouldFail {
+		writeOpenAIErrorWithCode(w, outcome.Error.Status, outcome.Error.Message, outcome.Error.Code)
 		return
 	}

-	responseObj := openaifmt.BuildResponseObjectWithToolCalls(responseID, model, finalPrompt, sanitizedThinking, sanitizedText, textParsed.Calls, toolsRaw)
-	if refFileTokens > 0 {
-		addRefFileTokensToUsage(responseObj, refFileTokens)
-	}
+	responseObj := openaifmt.BuildResponseObjectWithToolCalls(responseID, model, finalPrompt, turn.Thinking, turn.Text, turn.ToolCalls, toolsRaw)
+	responseObj["usage"] = assistantturn.OpenAIResponsesUsage(turn)
 	h.getResponseStore().put(owner, responseID, responseObj)
 	writeJSON(w, http.StatusOK, responseObj)
 }
@@ -178,7 +177,7 @@ func (h *Handler) handleResponsesStream(w http.ResponseWriter, r *http.Request,
 	}
 	bufferToolContent := len(toolNames) > 0
 	emitEarlyToolDeltas := h.toolcallFeatureMatchEnabled() && h.toolcallEarlyEmitHighConfidence()
-	stripReferenceMarkers := h.compatStripReferenceMarkers()
+	stripReferenceMarkers := stripReferenceMarkersEnabled()

 	streamRuntime := newResponsesStreamRuntime(
 		w,
--- a/internal/httpapi/openai/responses/responses_stream_runtime_core.go
+++ b/internal/httpapi/openai/responses/responses_stream_runtime_core.go
@@ -194,15 +194,11 @@ func (s *responsesStreamRuntime) finalize(finishReason string, deferEmptyOutput

 	s.closeMessageItem()

-	if turn.Error != nil && turn.Error.Code == "tool_choice_violation" {
-		s.failResponse(turn.Error.Status, turn.Error.Message, turn.Error.Code)
-		return true
-	}
-	if len(detected) == 0 && strings.TrimSpace(turn.Text) == "" {
-		status, message, code := upstreamEmptyOutputDetail(finishReason == "content_filter", turn.Text, turn.Thinking)
-		if turn.Error != nil {
-			status, message, code = turn.Error.Status, turn.Error.Message, turn.Error.Code
-		}
+	outcome := assistantturn.FinalizeTurn(turn, assistantturn.FinalizeOptions{
+		AlreadyEmittedToolCalls: s.toolCallsEmitted || s.toolCallsDoneEmitted,
+	})
+	if outcome.ShouldFail {
+		status, message, code := outcome.Error.Status, outcome.Error.Message, outcome.Error.Code
 		if deferEmptyOutput {
 			s.finalErrorStatus = status
 			s.finalErrorMessage = message
@@ -223,14 +219,6 @@ func (s *responsesStreamRuntime) finalize(finishReason string, deferEmptyOutput
 	return true
 }

-func responsesUsageFromTurn(turn assistantturn.Turn) map[string]any {
-	return map[string]any{
-		"input_tokens":  turn.Usage.InputTokens,
-		"output_tokens": turn.Usage.OutputTokens,
-		"total_tokens":  turn.Usage.TotalTokens,
-	}
-}
-
 func (s *responsesStreamRuntime) logToolPolicyRejections(textParsed toolcall.ToolCallParseResult) {
 	logRejected := func(parsed toolcall.ToolCallParseResult, channel string) {
 		rejected := filteredRejectedToolNamesForLog(parsed.RejectedToolNames)
--- a/internal/httpapi/openai/responses/responses_stream_test.go
+++ b/internal/httpapi/openai/responses/responses_stream_test.go
@@ -453,25 +453,13 @@ func TestHandleResponsesNonStreamReturns429WhenUpstreamHasOnlyThinking(t *testin
 	}

 	h.handleResponsesNonStream(rec, resp, "owner-a", "resp_test", "deepseek-v4-pro", "prompt", 0, true, false, nil, nil, promptcompat.DefaultToolChoicePolicy(), "")
-	if rec.Code != http.StatusOK {
-		t.Fatalf("expected 200 for thinking-only upstream output, got %d body=%s", rec.Code, rec.Body.String())
+	if rec.Code != http.StatusTooManyRequests {
+		t.Fatalf("expected 429 for thinking-only upstream output, got %d body=%s", rec.Code, rec.Body.String())
 	}
 	out := decodeJSONBody(t, rec.Body.String())
-	output, _ := out["output"].([]any)
-	if len(output) == 0 {
-		t.Fatal("expected at least one output item")
-	}
-	first, _ := output[0].(map[string]any)
-	content, _ := first["content"].([]any)
-	if len(content) == 0 {
-		t.Fatal("expected at least one content item")
-	}
-	firstContent, _ := content[0].(map[string]any)
-	if asString(firstContent["type"]) != "reasoning" {
-		t.Fatalf("expected reasoning type, got %v", firstContent["type"])
-	}
-	if asString(firstContent["text"]) != "Only thinking" {
-		t.Fatalf("expected text='Only thinking', got %v", firstContent["text"])
+	errObj, _ := out["error"].(map[string]any)
+	if asString(errObj["code"]) != "upstream_empty_output" {
+		t.Fatalf("expected code=upstream_empty_output, got %#v", out)
 	}
 }

--- a/internal/httpapi/openai/shared/deps.go
+++ b/internal/httpapi/openai/shared/deps.go
@@ -35,8 +35,6 @@ type DeepSeekCaller interface {

 type ConfigReader interface {
 	ModelAliases() map[string]string
-	CompatWideInputStrictOutput() bool
-	CompatStripReferenceMarkers() bool
 	ToolcallMode() string
 	ToolcallEarlyEmitConfidence() string
 	ResponsesStoreTTLSeconds() int
@@ -56,13 +54,6 @@ type Deps struct {
 	ChatHistory *chathistory.Store
 }

-func CompatStripReferenceMarkers(store ConfigReader) bool {
-	if store == nil {
-		return true
-	}
-	return store.CompatStripReferenceMarkers()
-}
-
 var WriteJSON = util.WriteJSON

 var _ AuthResolver = (*auth.Resolver)(nil)
--- a/internal/httpapi/openai/stream_status_test.go
+++ b/internal/httpapi/openai/stream_status_test.go
@@ -135,7 +135,7 @@ func captureStatusMiddleware(statuses *[]int) func(http.Handler) http.Handler {
 func TestChatCompletionsStreamStatusCapturedAs200(t *testing.T) {
 	statuses := make([]int, 0, 1)
 	h := &openAITestSurface{
-		Store: mockOpenAIConfig{wideInput: true},
+		Store: mockOpenAIConfig{},
 		Auth:  streamStatusAuthStub{},
 		DS:    streamStatusDSStub{resp: makeOpenAISSEHTTPResponse(`data: {"p":"response/content","v":"hello"}`, "data: [DONE]")},
 	}
@@ -164,7 +164,7 @@ func TestChatCompletionsStreamStatusCapturedAs200(t *testing.T) {
 func TestResponsesStreamStatusCapturedAs200(t *testing.T) {
 	statuses := make([]int, 0, 1)
 	h := &openAITestSurface{
-		Store: mockOpenAIConfig{wideInput: true},
+		Store: mockOpenAIConfig{},
 		Auth:  streamStatusAuthStub{},
 		DS:    streamStatusDSStub{resp: makeOpenAISSEHTTPResponse(`data: {"p":"response/content","v":"hello"}`, "data: [DONE]")},
 	}
@@ -193,7 +193,7 @@ func TestResponsesStreamStatusCapturedAs200(t *testing.T) {
 func TestChatCompletionsStreamContentFilterStopsNormallyWithoutLeak(t *testing.T) {
 	statuses := make([]int, 0, 1)
 	h := &openAITestSurface{
-		Store: mockOpenAIConfig{wideInput: true},
+		Store: mockOpenAIConfig{},
 		Auth:  streamStatusAuthStub{},
 		DS: streamStatusDSStub{resp: makeOpenAISSEHTTPResponse(
 			`data: {"p":"response/content","v":"合法前缀"}`,
@@ -243,7 +243,7 @@ func TestChatCompletionsStreamContentFilterStopsNormallyWithoutLeak(t *testing.T
 func TestChatCompletionsStreamEmitsFailureFrameWhenUpstreamOutputEmpty(t *testing.T) {
 	statuses := make([]int, 0, 1)
 	h := &openAITestSurface{
-		Store: mockOpenAIConfig{wideInput: true},
+		Store: mockOpenAIConfig{},
 		Auth:  streamStatusAuthStub{},
 		DS:    streamStatusDSStub{resp: makeOpenAISSEHTTPResponse("data: [DONE]")},
 	}
@@ -289,7 +289,7 @@ func TestChatCompletionsStreamRetriesEmptyOutputOnSameSession(t *testing.T) {
 		makeOpenAISSEHTTPResponse(`data: {"p":"response/content","v":"visible"}`, "data: [DONE]"),
 	}}
 	h := &openAITestSurface{
-		Store: mockOpenAIConfig{wideInput: true},
+		Store: mockOpenAIConfig{},
 		Auth:  streamStatusAuthStub{},
 		DS:    ds,
 	}
@@ -349,7 +349,7 @@ func TestChatCompletionsNonStreamRetriesThinkingOnlyOutput(t *testing.T) {
 		makeOpenAISSEHTTPResponse(`data: {"p":"response/content","v":"visible"}`, "data: [DONE]"),
 	}}
 	h := &openAITestSurface{
-		Store: mockOpenAIConfig{wideInput: true},
+		Store: mockOpenAIConfig{},
 		Auth:  streamStatusAuthStub{},
 		DS:    ds,
 	}
@@ -388,7 +388,7 @@ func TestChatCompletionsContentFilterDoesNotRetry(t *testing.T) {
 		makeOpenAISSEHTTPResponse(`data: {"p":"response/content","v":"visible"}`, "data: [DONE]"),
 	}}
 	h := &openAITestSurface{
-		Store: mockOpenAIConfig{wideInput: true},
+		Store: mockOpenAIConfig{},
 		Auth:  streamStatusAuthStub{},
 		DS:    ds,
 	}
@@ -410,7 +410,7 @@ func TestChatCompletionsContentFilterDoesNotRetry(t *testing.T) {
 func TestResponsesStreamUsageIgnoresBatchAccumulatedTokenUsage(t *testing.T) {
 	statuses := make([]int, 0, 1)
 	h := &openAITestSurface{
-		Store: mockOpenAIConfig{wideInput: true},
+		Store: mockOpenAIConfig{},
 		Auth:  streamStatusAuthStub{},
 		DS: streamStatusDSStub{resp: makeOpenAISSEHTTPResponse(
 			`data: {"p":"response/content","v":"hello"}`,
@@ -461,7 +461,7 @@ func TestResponsesStreamRetriesThinkingOnlyOutput(t *testing.T) {
 		makeOpenAISSEHTTPResponse(`data: {"p":"response/content","v":"visible"}`, "data: [DONE]"),
 	}}
 	h := &openAITestSurface{
-		Store: mockOpenAIConfig{wideInput: true},
+		Store: mockOpenAIConfig{},
 		Auth:  streamStatusAuthStub{},
 		DS:    ds,
 	}
@@ -500,7 +500,7 @@ func TestResponsesNonStreamRetriesThinkingOnlyOutput(t *testing.T) {
 		makeOpenAISSEHTTPResponse(`data: {"p":"response/content","v":"visible"}`, "data: [DONE]"),
 	}}
 	h := &openAITestSurface{
-		Store: mockOpenAIConfig{wideInput: true},
+		Store: mockOpenAIConfig{},
 		Auth:  streamStatusAuthStub{},
 		DS:    ds,
 	}
@@ -546,7 +546,7 @@ func TestResponsesNonStreamRetriesThinkingOnlyOutput(t *testing.T) {
 func TestResponsesNonStreamUsageIgnoresPromptAndOutputTokenUsage(t *testing.T) {
 	statuses := make([]int, 0, 1)
 	h := &openAITestSurface{
-		Store: mockOpenAIConfig{wideInput: true},
+		Store: mockOpenAIConfig{},
 		Auth:  streamStatusAuthStub{},
 		DS: streamStatusDSStub{resp: makeOpenAISSEHTTPResponse(
 			`data: {"p":"response/content","v":"ok"}`,
--- a/internal/js/chat-stream/vercel_stream_impl.js
+++ b/internal/js/chat-stream/vercel_stream_impl.js
@@ -17,7 +17,6 @@ const {
  resolveToolcallPolicy,
  formatIncrementalToolCallDeltas,
  filterIncrementalToolCallDeltasByAllowed,
-  boolDefaultTrue,
  resetStreamToolCallState,
 } = require('./toolcall_policy');
 const { createChatCompletionEmitter, createDeltaCoalescer } = require('./stream_emitter');
@@ -58,7 +57,7 @@ async function handleVercelStream(req, res, rawBody, payload) {
  const toolPolicy = resolveToolcallPolicy(prep.body, payload.tools);
  const toolNames = toolPolicy.toolNames;
  const emitEarlyToolDeltas = toolPolicy.emitEarlyToolDeltas;
-  const stripReferenceMarkers = boolDefaultTrue(prep.body.compat && prep.body.compat.strip_reference_markers);
+  const stripReferenceMarkers = true;

  if (!model || !leaseID || !deepseekToken || !initialPowHeader || !completionPayload) {
    writeOpenAIError(res, 500, 'invalid vercel prepare response');
--- a/internal/promptcompat/request_normalize.go
+++ b/internal/promptcompat/request_normalize.go
@@ -10,7 +10,6 @@ import (

 type ConfigReader interface {
 	ModelAliases() map[string]string
-	CompatWideInputStrictOutput() bool
 }

 func NormalizeOpenAIChatRequest(store ConfigReader, req map[string]any, traceID string) (StandardRequest, error) {
@@ -74,17 +73,7 @@ func NormalizeOpenAIResponsesRequest(store ConfigReader, req map[string]any, tra
 		thinkingEnabled = false
 	}

-	// Keep width-control as an explicit policy hook even if current default is true.
-	allowWideInput := true
-	if store != nil {
-		allowWideInput = store.CompatWideInputStrictOutput()
-	}
-	var messagesRaw []any
-	if allowWideInput {
-		messagesRaw = ResponsesMessagesFromRequest(req)
-	} else if msgs, ok := req["messages"].([]any); ok && len(msgs) > 0 {
-		messagesRaw = msgs
-	}
+	messagesRaw := ResponsesMessagesFromRequest(req)
 	if len(messagesRaw) == 0 {
 		return StandardRequest{}, fmt.Errorf("request must include 'input' or 'messages'")
 	}
--- a/tests/node/chat-stream.test.js
+++ b/tests/node/chat-stream.test.js
@@ -137,7 +137,6 @@ async function runMockVercelStreamSequence(upstreamSequences, prepareOverrides =
    final_prompt: 'hello',
    thinking_enabled: false,
    search_enabled: false,
-    compat: { strip_reference_markers: true },
    tool_names: [],
    deepseek_token: 'deepseek-token',
    pow_header: 'pow-header',
@@ -310,7 +309,6 @@ test('vercel stream reuses prior PoW when refresh fails', async () => {
        final_prompt: 'hello',
        thinking_enabled: false,
        search_enabled: false,
-        compat: { strip_reference_markers: true },
        tool_names: [],
        deepseek_token: 'deepseek-token',
        pow_header: 'pow-header-initial',
--- a/webui/src/features/settings/CompatibilitySection.jsx
+++ b/webui/src/features/settings/CompatibilitySection.jsx
@@ -1,34 +0,0 @@
-import { ShieldAlert } from 'lucide-react'
-
-export default function CompatibilitySection({ t, form, setForm }) {
-    return (
-        <div className="bg-card border border-border rounded-xl p-5 space-y-4">
-            <div className="flex items-center gap-2">
-                <ShieldAlert className="w-4 h-4 text-muted-foreground" />
-                <h3 className="font-semibold">{t('settings.compatibilityTitle')}</h3>
-            </div>
-            <p className="text-sm text-muted-foreground">{t('settings.compatibilityDesc')}</p>
-            <div className="flex items-center justify-between gap-4">
-                <label className="text-sm font-medium">{t('settings.stripReferenceMarkers')}</label>
-                <button
-                    type="button"
-                    role="switch"
-                    aria-checked={form.compat?.strip_reference_markers ?? true}
-                    onClick={() => setForm((prev) => ({
-                        ...prev,
-                        compat: { ...prev.compat, strip_reference_markers: !Boolean(prev.compat?.strip_reference_markers ?? true) },
-                    }))}
-                    className={`relative inline-flex h-6 w-11 items-center rounded-full transition-colors ${
-                        form.compat?.strip_reference_markers ?? true ? 'bg-primary' : 'bg-muted'
-                    }`}
-                >
-                    <span
-                        className={`inline-block h-4 w-4 transform rounded-full bg-white transition-transform ${
-                            form.compat?.strip_reference_markers ?? true ? 'translate-x-6' : 'translate-x-1'
-                        }`}
-                    />
-                </button>
-            </div>
-        </div>
-    )
-}
--- a/webui/src/features/settings/SettingsContainer.jsx
+++ b/webui/src/features/settings/SettingsContainer.jsx
@@ -6,7 +6,6 @@ import SecuritySection from './SecuritySection'
 import RuntimeSection from './RuntimeSection'
 import BehaviorSection from './BehaviorSection'
 import CurrentInputFileSection from './CurrentInputFileSection'
-import CompatibilitySection from './CompatibilitySection'
 import AutoDeleteSection from './AutoDeleteSection'
 import ModelSection from './ModelSection'
 import BackupSection from './BackupSection'
@@ -98,8 +97,6 @@ export default function SettingsContainer({ onRefresh, onMessage, authFetch, onF

            <CurrentInputFileSection t={t} form={form} setForm={setForm} />

-            <CompatibilitySection t={t} form={form} setForm={setForm} />
-
            <AutoDeleteSection t={t} form={form} setForm={setForm} />

            <ModelSection t={t} form={form} setForm={setForm} />
--- a/webui/src/features/settings/useSettingsForm.js
+++ b/webui/src/features/settings/useSettingsForm.js
@@ -13,7 +13,6 @@ const MAX_AUTO_FETCH_FAILURES = 3
 const DEFAULT_FORM = {
    admin: { jwt_expire_hours: 24 },
    runtime: { account_max_inflight: 2, account_max_queue: 10, global_max_inflight: 10, token_refresh_interval_hours: 6 },
-    compat: { strip_reference_markers: true },
    responses: { store_ttl_seconds: 900 },
    embeddings: { provider: '' },
    auto_delete: { mode: 'none' },
@@ -60,9 +59,6 @@ function fromServerForm(data) {
            global_max_inflight: Number(data.runtime?.global_max_inflight || 10),
            token_refresh_interval_hours: Number(data.runtime?.token_refresh_interval_hours || 6),
        },
-        compat: {
-            strip_reference_markers: data.compat?.strip_reference_markers ?? true,
-        },
        responses: {
            store_ttl_seconds: Number(data.responses?.store_ttl_seconds || 900),
        },
@@ -95,9 +91,6 @@ function toServerPayload(form) {
            global_max_inflight: Number(form.runtime.global_max_inflight),
            token_refresh_interval_hours: Number(form.runtime.token_refresh_interval_hours),
        },
-        compat: {
-            strip_reference_markers: Boolean(form.compat?.strip_reference_markers ?? true),
-        },
        responses: { store_ttl_seconds: Number(form.responses.store_ttl_seconds) },
        embeddings: { provider: String(form.embeddings.provider || '').trim() },
        auto_delete: { mode: normalizeAutoDeleteMode(form.auto_delete) },
--- a/webui/src/locales/en.json
+++ b/webui/src/locales/en.json
@@ -397,9 +397,6 @@
        "currentInputFileDesc": "Enabled by default. Once the character threshold is reached, upload the full context as a DS2API_HISTORY.txt context file.",
        "currentInputFileMinChars": "Current input threshold (characters)",
        "currentInputFileHelp": "Default is 0, which uses independent split for any non-empty input.",
-        "compatibilityTitle": "Compatibility",
-        "compatibilityDesc": "Compatibility controls that keep stream output closer to the wire format or safer for the web UI.",
-        "stripReferenceMarkers": "Strip [reference:N] markers",
        "modelTitle": "Model mapping",
        "modelAliases": "Global model aliases (JSON)",
        "autoDeleteTitle": "Session Cleanup Policy",
@@ -485,4 +482,4 @@
            "four": "Trigger a redeploy to apply the updated environment variables."
        }
    }
-}
+}
--- a/webui/src/locales/zh.json
+++ b/webui/src/locales/zh.json
@@ -397,9 +397,6 @@
        "currentInputFileDesc": "默认开启。达到字符阈值后，将完整上下文上传为 DS2API_HISTORY.txt 上下文文件。",
        "currentInputFileMinChars": "当前输入阈值（字符数）",
        "currentInputFileHelp": "默认 0，表示只要有输入就会使用独立拆分。",
-        "compatibilityTitle": "兼容性设置",
-        "compatibilityDesc": "用于控制输出格式兼容性，避免把模型原始流里的标记直接暴露到前端。",
-        "stripReferenceMarkers": "移除 [reference:N] 标记",
        "modelTitle": "模型映射",
        "modelAliases": "全局模型映射（JSON）",
        "autoDeleteTitle": "会话删除策略",
@@ -485,4 +482,4 @@
            "four": "触发重新部署以应用新的环境变量。"
        }
    }
-}
+}