diff --git a/config.example.yaml b/config.example.yaml
index 786f14559..af9bf6a06 100644
--- a/config.example.yaml
+++ b/config.example.yaml
@@ -908,3 +908,16 @@ nonstream-keepalive-interval: 0
 #       params: # JSON paths (gjson/sjson syntax) to remove from the payload
 #         - "generationConfig.thinkingConfig.thinkingBudget"
 #         - "generationConfig.responseJsonSchema"
+
+# Optional translator configuration
+# translator:
+#   # When true, prior assistant thinking/reasoning content is carried over into
+#   # a labeled "Prior assistant reasoning (unverified context)" system
+#   # instruction when translating to a target protocol (e.g. plain OpenAI chat
+#   # completions) that has no canonical thought field. Default is false.
+#   #
+#   # This is a fallback for non-canonical targets: the most recent 3 blocks are
+#   # kept, each capped at 4000 runes, and older or truncated content is marked.
+#   # Canonical compatibility targets (isCompat / reasoning_content) are not
+#   # affected.
+#   carry-over-thinking-in-system: false
diff --git a/internal/config/config.go b/internal/config/config.go
index 0d8fb234f..109e37d92 100644
--- a/internal/config/config.go
+++ b/internal/config/config.go
@@ -180,6 +180,9 @@ type Config struct {
 	// Payload defines default and override rules for provider payload parameters.
 	Payload PayloadConfig `yaml:"payload" json:"payload"`
 
+	// Translator controls cross-format request translation behavior.
+	Translator TranslatorConfig `yaml:"translator" json:"translator"`
+
 	// IncognitoBrowser opens OAuth URLs in an incognito/private browser window.
 	IncognitoBrowser bool `yaml:"incognito-browser" json:"incognito-browser"`
 
diff --git a/internal/config/config_types.go b/internal/config/config_types.go
index 39970e93a..e4b725872 100644
--- a/internal/config/config_types.go
+++ b/internal/config/config_types.go
@@ -366,6 +366,15 @@ type PayloadModelRule struct {
 	NotExist []string `yaml:"not-exist" json:"not-exist"`
 }
 
+// TranslatorConfig controls cross-format request translation behavior.
+type TranslatorConfig struct {
+	// CarryOverThinkingInSystem moves prior assistant reasoning/thinking into a
+	// labeled system instruction when the target protocol has no canonical thought
+	// field (e.g. plain OpenAI chat completions). Default false preserves strict
+	// protocol behavior.
+	CarryOverThinkingInSystem bool `yaml:"carry-over-thinking-in-system" json:"carry-over-thinking-in-system"`
+}
+
 // CloakConfig configures request cloaking for non-Claude-Code clients.
 // Cloaking disguises API requests to appear as originating from the official Claude Code CLI.
 type CloakConfig struct {
diff --git a/internal/runtime/executor/helps/carry_over.go b/internal/runtime/executor/helps/carry_over.go
new file mode 100644
index 000000000..2ed05dd4f
--- /dev/null
+++ b/internal/runtime/executor/helps/carry_over.go
@@ -0,0 +1,312 @@
+package helps
+
+import (
+	"fmt"
+	"strings"
+
+	sigcompat "github.com/router-for-me/CLIProxyAPI/v7/internal/signature"
+	"github.com/router-for-me/CLIProxyAPI/v7/internal/thinking"
+	translatorcommon "github.com/router-for-me/CLIProxyAPI/v7/internal/translator/common"
+	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
+)
+
+const (
+	carryOverLabel        = "Prior assistant reasoning (unverified context)"
+	carryOverMaxBlocks    = 3
+	carryOverMaxBlockSize = 4000
+)
+
+// CarryOverThinkingToSystem extracts reasoning_content from assistant messages
+// in an OpenAI Chat Completions payload and rewrites it as a labeled system
+// instruction. It drops assistant messages that are empty after reasoning is
+// removed, including assistant messages that were already empty. OpenAI rejects
+// empty assistant messages, so removing them is intentional. Existing first
+// system message is extended; otherwise a new one is inserted.
+//
+// The function does not add reasoning to response bodies; it is only for
+// request bodies being sent to a target without a canonical thought field.
+func CarryOverThinkingToSystem(payload []byte) []byte {
+	if len(payload) == 0 || !gjson.ValidBytes(payload) {
+		return payload
+	}
+
+	messages := gjson.GetBytes(payload, "messages")
+	if !messages.Exists() || !messages.IsArray() {
+		return payload
+	}
+
+	var reasoningBlocks []string
+	keptMessages := make([][]byte, 0, len(messages.Array()))
+
+	messages.ForEach(func(_, msg gjson.Result) bool {
+		role := msg.Get("role").String()
+
+		var reasoning string
+		if role == "assistant" {
+			if rc := msg.Get("reasoning_content"); rc.Exists() && rc.Type == gjson.String {
+				reasoning = rc.String()
+			}
+		}
+
+		if reasoning != "" {
+			reasoningBlocks = append(reasoningBlocks, reasoning)
+		}
+
+		updated := []byte(msg.Raw)
+		if msg.Get("reasoning_content").Exists() {
+			updated, _ = sjson.DeleteBytes(updated, "reasoning_content")
+		}
+
+		if role == "assistant" && !assistantMessageHasContent(updated) {
+			return true
+		}
+
+		keptMessages = append(keptMessages, updated)
+		return true
+	})
+
+	if len(reasoningBlocks) == 0 {
+		return payload
+	}
+
+	systemText := carryOverLabel + ":\n\n" + formatCarryOverText(reasoningBlocks)
+
+	if len(keptMessages) > 0 && gjson.GetBytes(keptMessages[0], "role").String() == "system" {
+		keptMessages[0] = mergeCarryOverIntoSystemMessage(keptMessages[0], systemText)
+	} else {
+		systemMsg := []byte(`{"role":"system","content":""}`)
+		systemMsg, _ = sjson.SetBytes(systemMsg, "content", systemText)
+		keptMessages = append([][]byte{systemMsg}, keptMessages...)
+	}
+
+	return translatorcommon.SetRawArrayItems(payload, "messages", keptMessages)
+}
+
+func assistantMessageHasContent(msg []byte) bool {
+	if gjson.GetBytes(msg, "tool_calls").IsArray() && len(gjson.GetBytes(msg, "tool_calls").Array()) > 0 {
+		return true
+	}
+
+	c := gjson.GetBytes(msg, "content")
+	if !c.Exists() || c.Type == gjson.Null {
+		return false
+	}
+
+	if c.Type == gjson.String {
+		return strings.TrimSpace(c.String()) != ""
+	}
+
+	if c.IsArray() && len(c.Array()) > 0 {
+		for _, part := range c.Array() {
+			if part.Get("type").String() == "text" {
+				if strings.TrimSpace(part.Get("text").String()) != "" {
+					return true
+				}
+			} else if part.Get("type").Exists() {
+				return true
+			}
+		}
+	}
+
+	return false
+}
+
+func formatCarryOverText(blocks []string) string {
+	omitted := 0
+	if len(blocks) > carryOverMaxBlocks {
+		omitted = len(blocks) - carryOverMaxBlocks
+		blocks = blocks[len(blocks)-carryOverMaxBlocks:]
+	}
+
+	var parts []string
+	if omitted > 0 {
+		parts = append(parts, fmt.Sprintf("[... %d older reasoning block(s) omitted; showing the most recent %d.]", omitted, carryOverMaxBlocks))
+	}
+
+	for i, block := range blocks {
+		if i > 0 || omitted > 0 {
+			parts = append(parts, "")
+		}
+
+		runes := []rune(block)
+		if len(runes) > carryOverMaxBlockSize {
+			block = string(runes[:carryOverMaxBlockSize]) + "\n\n... [reasoning truncated]"
+		}
+		parts = append(parts, block)
+	}
+
+	return strings.Join(parts, "\n")
+}
+
+func newCarryOverTextPart(text string) []byte {
+	part := []byte(`{"type":"text","text":""}`)
+	part, _ = sjson.SetBytes(part, "text", text)
+	return part
+}
+
+func mergeCarryOverIntoSystemMessage(msg []byte, carryOverText string) []byte {
+	c := gjson.GetBytes(msg, "content")
+
+	switch {
+	case !c.Exists() || c.Type == gjson.Null:
+		msg, _ = sjson.SetBytes(msg, "content", carryOverText)
+
+	case c.Type == gjson.String:
+		merged := carryOverText + "\n\n" + c.String()
+		msg, _ = sjson.SetBytes(msg, "content", merged)
+
+	case c.IsArray():
+		items := [][]byte{newCarryOverTextPart(carryOverText)}
+		c.ForEach(func(_, part gjson.Result) bool {
+			switch {
+			case part.IsObject():
+				items = append(items, []byte(part.Raw))
+			case part.Type == gjson.String:
+				items = append(items, newCarryOverTextPart(part.String()))
+			}
+			return true
+		})
+
+		msg, _ = sjson.SetRawBytes(msg, "content", translatorcommon.JoinRawArray(items))
+
+	default:
+		msg, _ = sjson.SetBytes(msg, "content", carryOverText)
+	}
+
+	return msg
+}
+
+// carryOverClaudeSource extracts unsigned assistant thinking blocks from a
+// Claude request and rewrites them as a top-level system instruction. Signed
+// thinking with a compatible signature is left in place so the normal registry
+// path can map it to reasoning_content. This runs before native translation so
+// unsigned thinking is not dropped; plugin NormalizeRequest hooks then run on
+// the translated provider payload and own the final OpenAI-shaped request.
+func carryOverClaudeSource(payload []byte) []byte {
+	if len(payload) == 0 || !gjson.ValidBytes(payload) {
+		return payload
+	}
+
+	messages := gjson.GetBytes(payload, "messages")
+	if !messages.Exists() || !messages.IsArray() {
+		return payload
+	}
+
+	var blocks []string
+	keptMessages := make([][]byte, 0, len(messages.Array()))
+
+	messages.ForEach(func(_, msg gjson.Result) bool {
+		role := msg.Get("role").String()
+		if role != "assistant" {
+			keptMessages = append(keptMessages, []byte(msg.Raw))
+			return true
+		}
+
+		content := msg.Get("content")
+		if !content.IsArray() {
+			keptMessages = append(keptMessages, []byte(msg.Raw))
+			return true
+		}
+
+		var keptParts [][]byte
+		hasToolUse := false
+		extractedFromThis := false
+
+		content.ForEach(func(_, part gjson.Result) bool {
+			partType := part.Get("type").String()
+			if partType == "tool_use" {
+				hasToolUse = true
+			}
+			if partType != "thinking" {
+				if part.IsObject() {
+					keptParts = append(keptParts, []byte(part.Raw))
+				}
+				return true
+			}
+
+			text := thinking.GetThinkingText(part)
+			if strings.TrimSpace(text) == "" {
+				return true
+			}
+
+			if isUnsignedClaudeThinking(part) {
+				extractedFromThis = true
+				blocks = append(blocks, text)
+				return true
+			}
+
+			keptParts = append(keptParts, []byte(part.Raw))
+			return true
+		})
+
+		if !extractedFromThis {
+			keptMessages = append(keptMessages, []byte(msg.Raw))
+			return true
+		}
+
+		if len(keptParts) == 0 && !hasToolUse {
+			// assistant turn was only unsigned thinking; drop it
+			return true
+		}
+
+		updated := []byte(msg.Raw)
+		updated, _ = sjson.SetRawBytes(updated, "content", translatorcommon.JoinRawArray(keptParts))
+		keptMessages = append(keptMessages, updated)
+		return true
+	})
+
+	if len(blocks) == 0 {
+		return payload
+	}
+
+	systemText := carryOverLabel + ":\n\n" + formatCarryOverText(blocks)
+	payload = injectClaudeCarryOverSystem(payload, systemText)
+	return translatorcommon.SetRawArrayItems(payload, "messages", keptMessages)
+}
+
+func isUnsignedClaudeThinking(part gjson.Result) bool {
+	sig := part.Get("signature").String()
+	if strings.TrimSpace(sig) == "" {
+		return true
+	}
+	_, ok := sigcompat.CompatibleSignatureForProvider(sigcompat.SignatureProviderGPT, sig)
+	return !ok
+}
+
+func injectClaudeCarryOverSystem(payload []byte, carryOverText string) []byte {
+	system := gjson.GetBytes(payload, "system")
+
+	switch {
+	case !system.Exists() || system.Type == gjson.Null:
+		payload, _ = sjson.SetBytes(payload, "system", carryOverText)
+
+	case system.Type == gjson.String:
+		var merged string
+		if strings.TrimSpace(system.String()) != "" {
+			merged = carryOverText + "\n\n" + system.String()
+		} else {
+			merged = carryOverText
+		}
+		payload, _ = sjson.SetBytes(payload, "system", merged)
+
+	case system.IsArray():
+		items := [][]byte{newCarryOverTextPart(carryOverText)}
+		system.ForEach(func(_, part gjson.Result) bool {
+			switch {
+			case part.IsObject():
+				items = append(items, []byte(part.Raw))
+			case part.Type == gjson.String:
+				items = append(items, newCarryOverTextPart(part.String()))
+			}
+			return true
+		})
+
+		payload, _ = sjson.SetRawBytes(payload, "system", translatorcommon.JoinRawArray(items))
+
+	default:
+		payload, _ = sjson.SetBytes(payload, "system", carryOverText)
+	}
+
+	return payload
+}
diff --git a/internal/runtime/executor/helps/carry_over_test.go b/internal/runtime/executor/helps/carry_over_test.go
new file mode 100644
index 000000000..8691fcfa4
--- /dev/null
+++ b/internal/runtime/executor/helps/carry_over_test.go
@@ -0,0 +1,380 @@
+package helps
+
+import (
+	"context"
+	"encoding/base64"
+	"strings"
+	"testing"
+
+	"github.com/router-for-me/CLIProxyAPI/v7/internal/config"
+	sdktranslator "github.com/router-for-me/CLIProxyAPI/v7/sdk/translator"
+	"github.com/tidwall/gjson"
+)
+
+func TestCarryOverThinkingToSystem_MovesReasoningToSystemMessage(t *testing.T) {
+	input := []byte(`{
+		"model": "test",
+		"messages": [
+			{"role": "user", "content": "hello"},
+			{"role": "assistant", "content": "answer", "reasoning_content": "I should be helpful."}
+		]
+	}`)
+
+	out := CarryOverThinkingToSystem(input)
+
+	if gjson.GetBytes(out, "messages.0.role").String() != "system" {
+		t.Fatalf("expected first message to be system, got %s", gjson.GetBytes(out, "messages.0.role").String())
+	}
+	content := gjson.GetBytes(out, "messages.0.content").String()
+	if !strings.Contains(content, carryOverLabel) {
+		t.Fatalf("expected system content to contain %q, got %q", carryOverLabel, content)
+	}
+	if !strings.Contains(content, "I should be helpful") {
+		t.Fatalf("expected system content to contain reasoning, got %q", content)
+	}
+
+	if gjson.GetBytes(out, "messages.2.reasoning_content").Exists() {
+		t.Fatalf("reasoning_content should be removed from assistant message")
+	}
+
+	assistant := gjson.GetBytes(out, "messages.2")
+	if assistant.Get("role").String() != "assistant" || assistant.Get("content").String() != "answer" {
+		t.Fatalf("assistant message should be preserved, got %s", assistant.Raw)
+	}
+}
+
+func TestCarryOverThinkingToSystem_DropsEmptyAssistantMessage(t *testing.T) {
+	input := []byte(`{
+		"model": "test",
+		"messages": [
+			{"role": "user", "content": "hello"},
+			{"role": "assistant", "content": "", "reasoning_content": "only reasoning"}
+		]
+	}`)
+
+	out := CarryOverThinkingToSystem(input)
+
+	if gjson.GetBytes(out, "messages.#").Int() != 2 {
+		t.Fatalf("expected 2 messages, got %d: %s", gjson.GetBytes(out, "messages.#").Int(), string(out))
+	}
+	if gjson.GetBytes(out, "messages.1.role").String() != "user" {
+		t.Fatalf("user message should remain second")
+	}
+}
+
+func TestCarryOverThinkingToSystem_KeepsAssistantWithToolCalls(t *testing.T) {
+	input := []byte(`{
+		"model": "test",
+		"messages": [
+			{"role": "assistant", "content": "", "tool_calls": [{"id":"1","type":"function"}], "reasoning_content": "tool planning"}
+		]
+	}`)
+
+	out := CarryOverThinkingToSystem(input)
+
+	if gjson.GetBytes(out, "messages.#").Int() != 2 {
+		t.Fatalf("expected 2 messages, got %s", string(out))
+	}
+	if !gjson.GetBytes(out, "messages.1.tool_calls").Exists() {
+		t.Fatalf("tool_calls should be preserved")
+	}
+	if gjson.GetBytes(out, "messages.1.reasoning_content").Exists() {
+		t.Fatalf("reasoning_content should be removed")
+	}
+}
+
+func TestCarryOverThinkingToSystem_MergesIntoExistingSystemMessage(t *testing.T) {
+	input := []byte(`{
+		"model": "test",
+		"messages": [
+			{"role": "system", "content": "You are a helpful assistant."},
+			{"role": "assistant", "reasoning_content": "thinking", "content": "hi"}
+		]
+	}`)
+
+	out := CarryOverThinkingToSystem(input)
+
+	content := gjson.GetBytes(out, "messages.0.content").String()
+	if !strings.HasPrefix(content, carryOverLabel) {
+		t.Fatalf("expected carry-over label at start of system content, got %q", content)
+	}
+	if !strings.Contains(content, "You are a helpful assistant.") {
+		t.Fatalf("expected original system content to be preserved, got %q", content)
+	}
+}
+
+func TestCarryOverThinkingToSystem_MergesIntoExistingSystemMessageArray(t *testing.T) {
+	input := []byte(`{
+		"model": "test",
+		"messages": [
+			{"role": "system", "content": [{"type":"text","text":"base"}]},
+			{"role": "assistant", "reasoning_content": "thinking", "content": "hi"}
+		]
+	}`)
+
+	out := CarryOverThinkingToSystem(input)
+
+	firstType := gjson.GetBytes(out, "messages.0.content.0.type").String()
+	if firstType != "text" {
+		t.Fatalf("expected first content part to be text, got %q", firstType)
+	}
+	if !strings.Contains(gjson.GetBytes(out, "messages.0.content.0.text").String(), carryOverLabel) {
+		t.Fatalf("expected carry-over text in first content part, got %q", gjson.GetBytes(out, "messages.0.content.0.text").String())
+	}
+	if gjson.GetBytes(out, "messages.0.content.1.text").String() != "base" {
+		t.Fatalf("expected original content part to be preserved, got %q", gjson.GetBytes(out, "messages.0.content.1.text").String())
+	}
+}
+
+func TestCarryOverThinkingToSystem_BoundsAndTruncates(t *testing.T) {
+	// Build 5 reasoning blocks
+	blocks := []string{"old1", "old2", "mid", "recent", "newest"}
+	var msgs []string
+	for _, b := range blocks {
+		msgs = append(msgs, `{"role":"assistant","content":"","reasoning_content":"`+b+`"}`)
+	}
+	input := []byte(`{"model":"test","messages":[` + strings.Join(msgs, ",") + `]}`)
+
+	out := CarryOverThinkingToSystem(input)
+
+	system := gjson.GetBytes(out, "messages.0.content").String()
+	if !strings.Contains(system, "older reasoning block(s) omitted") {
+		t.Fatalf("expected omission marker, got %q", system)
+	}
+	if strings.Contains(system, "old1") || strings.Contains(system, "old2") {
+		t.Fatalf("expected old1 and old2 to be omitted, got %q", system)
+	}
+	if !strings.Contains(system, "mid") || !strings.Contains(system, "recent") || !strings.Contains(system, "newest") {
+		t.Fatalf("expected mid, recent, newest to be present, got %q", system)
+	}
+}
+
+func TestCarryOverThinkingToSystem_TruncatesLongBlock(t *testing.T) {
+	long := strings.Repeat("x", carryOverMaxBlockSize+50)
+	input := []byte(`{"model":"test","messages":[{"role":"assistant","content":"","reasoning_content":"` + long + `"}]}`)
+
+	out := CarryOverThinkingToSystem(input)
+
+	system := gjson.GetBytes(out, "messages.0.content").String()
+	if !strings.Contains(system, "... [reasoning truncated]") {
+		t.Fatalf("expected truncation marker, got %q", system)
+	}
+}
+
+func TestCarryOverThinkingToSystem_PreservesStringArrayElements(t *testing.T) {
+	input := []byte(`{
+		"model": "test",
+		"messages": [
+			{"role": "system", "content": ["plain string", {"type":"text","text":"object part"}]},
+			{"role": "assistant", "reasoning_content": "thinking", "content": "hi"}
+		]
+	}`)
+
+	out := CarryOverThinkingToSystem(input)
+
+	if gjson.GetBytes(out, "messages.0.content.1.type").String() != "text" || gjson.GetBytes(out, "messages.0.content.1.text").String() != "plain string" {
+		t.Fatalf("expected plain string to be wrapped and preserved, got %s", string(out))
+	}
+	if gjson.GetBytes(out, "messages.0.content.2.text").String() != "object part" {
+		t.Fatalf("expected object part to be preserved, got %s", string(out))
+	}
+}
+
+func TestCarryOverThinkingToSystem_NoReasoningLeavesPayloadUnchanged(t *testing.T) {
+	input := []byte(`{"model":"test","messages":[{"role":"user","content":"hello"}]}`)
+	out := CarryOverThinkingToSystem(input)
+	if string(out) != string(input) {
+		t.Fatalf("expected payload to be unchanged, got %s", string(out))
+	}
+}
+
+func TestTranslateRequestWithAPIKeyModelCompatibility_CarryOverClaudeToOpenAI(t *testing.T) {
+	cfg := &config.Config{
+		Translator: config.TranslatorConfig{
+			CarryOverThinkingInSystem: true,
+		},
+	}
+
+	claudePayload := []byte(`{
+		"model": "claude-3-opus",
+		"messages": [
+			{"role": "user", "content": "hello"},
+			{"role": "assistant", "content": [{"type":"text","text":"hi"},{"type":"thinking","thinking":"internal reasoning"}]}
+		]
+	}`)
+
+	out := TranslateRequestWithAPIKeyModelCompatibility(context.Background(), nil, cfg, sdktranslator.FormatClaude, sdktranslator.FormatOpenAI, "test", claudePayload, false, false)
+
+	firstRole := gjson.GetBytes(out, "messages.0.role").String()
+	if firstRole != "system" {
+		t.Fatalf("expected first message role to be system, got %q", firstRole)
+	}
+	if !strings.Contains(gjson.GetBytes(out, "messages.0.content").String(), "internal reasoning") {
+		t.Fatalf("expected reasoning in system message, got %s", string(out))
+	}
+	if gjson.GetBytes(out, "messages.#").Int() != 3 {
+		t.Fatalf("expected 3 messages (system, user, assistant), got %d: %s", gjson.GetBytes(out, "messages.#").Int(), string(out))
+	}
+	for _, msg := range gjson.GetBytes(out, "messages").Array() {
+		if msg.Get("reasoning_content").Exists() {
+			t.Fatalf("no message should have reasoning_content, got %s", msg.Raw)
+		}
+	}
+}
+
+func TestTranslateRequestWithAPIKeyModelCompatibility_RespectsCompat(t *testing.T) {
+	cfg := &config.Config{
+		Translator: config.TranslatorConfig{
+			CarryOverThinkingInSystem: true,
+		},
+	}
+
+	claudePayload := []byte(`{
+		"model": "claude-3-opus",
+		"messages": [
+			{"role": "user", "content": "hello"},
+			{"role": "assistant", "content": [{"type":"text","text":"hi"},{"type":"thinking","thinking":"internal reasoning"}]}
+		]
+	}`)
+
+	out := TranslateRequestWithAPIKeyModelCompatibility(context.Background(), nil, cfg, sdktranslator.FormatClaude, sdktranslator.FormatOpenAI, "test", claudePayload, false, true)
+
+	if gjson.GetBytes(out, "messages.0.role").String() == "system" {
+		t.Fatalf("system carry-over should not happen when isCompat is true")
+	}
+	if !gjson.GetBytes(out, "messages.1.reasoning_content").Exists() {
+		t.Fatalf("expected canonical reasoning_content on compat path, got %s", string(out))
+	}
+}
+
+func TestTranslateRequestWithAPIKeyModelCompatibility_DisabledByDefault(t *testing.T) {
+	cfg := &config.Config{} // default false
+
+	claudePayload := []byte(`{
+		"model": "claude-3-opus",
+		"messages": [
+			{"role": "assistant", "content": [{"type":"thinking","thinking":"internal reasoning"},{"type":"text","text":"hi"}]}
+		]
+	}`)
+
+	out := TranslateRequestWithAPIKeyModelCompatibility(context.Background(), nil, cfg, sdktranslator.FormatClaude, sdktranslator.FormatOpenAI, "test", claudePayload, false, false)
+
+	if gjson.GetBytes(out, "messages.0.role").String() == "system" {
+		t.Fatalf("carry-over should not happen when disabled")
+	}
+	// Non-compat default drops unsigned thinking.
+	if gjson.GetBytes(out, "messages.0.reasoning_content").Exists() {
+		t.Fatalf("unsigned thinking should not become reasoning_content on default non-compat path, got %s", string(out))
+	}
+}
+
+func validGPTChatReasoningSignature() string {
+	raw := make([]byte, 1+8+16+16+32)
+	raw[0] = 0x80
+	raw[8] = 1
+	for i := 9; i < len(raw); i++ {
+		raw[i] = byte(i)
+	}
+	return base64.URLEncoding.EncodeToString(raw)
+}
+
+func TestTranslateRequestWithAPIKeyModelCompatibility_CarryOverKeepsSignedReasoningContent(t *testing.T) {
+	cfg := &config.Config{
+		Translator: config.TranslatorConfig{
+			CarryOverThinkingInSystem: true,
+		},
+	}
+
+	sig := validGPTChatReasoningSignature()
+	claudePayload := []byte(`{
+		"model": "claude-3-opus",
+		"messages": [
+			{"role": "assistant", "content": [
+				{"type":"thinking","thinking":"unsigned fallback reasoning"},
+				{"type":"thinking","thinking":"signed canonical reasoning","signature":"` + sig + `"},
+				{"type":"text","text":"hi"}
+			]}
+		]
+	}`)
+
+	out := TranslateRequestWithAPIKeyModelCompatibility(context.Background(), nil, cfg, sdktranslator.FormatClaude, sdktranslator.FormatOpenAI, "test", claudePayload, false, false)
+
+	system := gjson.GetBytes(out, "messages.0.content").String()
+	if !strings.Contains(system, "unsigned fallback reasoning") {
+		t.Fatalf("expected unsigned reasoning in system message, got %s", string(out))
+	}
+	if strings.Contains(system, "signed canonical reasoning") {
+		t.Fatalf("signed reasoning should stay as reasoning_content, got %s", string(out))
+	}
+
+	assistant := gjson.GetBytes(out, "messages.1")
+	if assistant.Get("role").String() != "assistant" {
+		t.Fatalf("expected assistant message, got %s", assistant.Raw)
+	}
+	if !strings.Contains(assistant.Get("reasoning_content").String(), "signed canonical reasoning") {
+		t.Fatalf("expected signed reasoning as reasoning_content, got %s", string(out))
+	}
+	if assistant.Get("content.0.text").String() != "hi" {
+		t.Fatalf("expected assistant content to be preserved, got %s", string(out))
+	}
+}
+
+func TestTranslateRequestWithAPIKeyModelCompatibility_CarryOverMergesWithSystem(t *testing.T) {
+	cfg := &config.Config{
+		Translator: config.TranslatorConfig{
+			CarryOverThinkingInSystem: true,
+		},
+	}
+
+	claudePayload := []byte(`{
+		"model": "claude-3-opus",
+		"system": "Base instructions",
+		"messages": [
+			{"role": "assistant", "content": [{"type":"thinking","thinking":"prior reasoning"}]}
+		]
+	}`)
+
+	out := TranslateRequestWithAPIKeyModelCompatibility(context.Background(), nil, cfg, sdktranslator.FormatClaude, sdktranslator.FormatOpenAI, "test", claudePayload, false, false)
+
+	system := gjson.GetBytes(out, "messages.0.content").String()
+	if !strings.Contains(system, carryOverLabel) {
+		t.Fatalf("expected carry-over label in system message, got %q", system)
+	}
+	if !strings.Contains(system, "Base instructions") {
+		t.Fatalf("expected original system instructions to be preserved, got %q", system)
+	}
+	if !strings.Contains(system, "prior reasoning") {
+		t.Fatalf("expected prior reasoning in system message, got %q", system)
+	}
+}
+
+func TestTranslateRequestWithAPIKeyModelCompatibility_CarryOverKeepsToolCalls(t *testing.T) {
+	cfg := &config.Config{
+		Translator: config.TranslatorConfig{
+			CarryOverThinkingInSystem: true,
+		},
+	}
+
+	claudePayload := []byte(`{
+		"model": "claude-3-opus",
+		"messages": [
+			{"role": "assistant", "content": [
+				{"type":"thinking","thinking":"tool planning"},
+				{"type":"tool_use","id":"tu_1","name":"do_work","input":{"x":1}}
+			]}
+		]
+	}`)
+
+	out := TranslateRequestWithAPIKeyModelCompatibility(context.Background(), nil, cfg, sdktranslator.FormatClaude, sdktranslator.FormatOpenAI, "test", claudePayload, false, false)
+
+	if gjson.GetBytes(out, "messages.#").Int() != 2 {
+		t.Fatalf("expected 2 messages, got %d: %s", gjson.GetBytes(out, "messages.#").Int(), string(out))
+	}
+	system := gjson.GetBytes(out, "messages.0.content").String()
+	if !strings.Contains(system, "tool planning") {
+		t.Fatalf("expected tool planning in system message, got %q", system)
+	}
+	if !gjson.GetBytes(out, "messages.1.tool_calls").Exists() {
+		t.Fatalf("expected assistant tool_calls to be preserved, got %s", string(out))
+	}
+}
diff --git a/internal/runtime/executor/helps/codex_multi_agent_v2.go b/internal/runtime/executor/helps/codex_multi_agent_v2.go
index 4e2209f86..a7665a748 100644
--- a/internal/runtime/executor/helps/codex_multi_agent_v2.go
+++ b/internal/runtime/executor/helps/codex_multi_agent_v2.go
@@ -69,6 +69,23 @@ func sameByteSlice(a, b []byte) bool {
 // request translators when a configured API-key model enables compatibility mode.
 func TranslateRequestWithAPIKeyModelCompatibility(ctx context.Context, headers http.Header, cfg *config.Config, from, to sdktranslator.Format, model string, payload []byte, stream, isCompat bool) []byte {
 	if !isCompat {
+		if cfg != nil && cfg.Translator.CarryOverThinkingInSystem && to == sdktranslator.FormatOpenAI {
+			working := payload
+			if from == sdktranslator.FormatClaude {
+				// Extract unsigned assistant thinking before native translation so
+				// it is not dropped. The registry then translates and runs plugin
+				// NormalizeRequest hooks on the final OpenAI-shaped payload.
+				// Signed thinking stays in place and maps to reasoning_content.
+				working = carryOverClaudeSource(working)
+			}
+			translated := TranslateRequestWithCodexMultiAgentV2(ctx, headers, cfg, from, to, model, working, stream)
+			if from != sdktranslator.FormatClaude {
+				// Other sources (e.g. openai-response) may already expose prior
+				// reasoning as reasoning_content in the translated payload.
+				translated = CarryOverThinkingToSystem(translated)
+			}
+			return translated
+		}
 		return TranslateRequestWithCodexMultiAgentV2(ctx, headers, cfg, from, to, model, payload, stream)
 	}
 	if from == sdktranslator.FormatOpenAIResponse && to != sdktranslator.FormatCodex && to != sdktranslator.FormatOpenAIResponse {