From b3046d29b9859c22c797b86cc99c52ac6ca1a2fc Mon Sep 17 00:00:00 2001
From: sususu <suchangshan@foxmail.com>
Date: Thu, 30 Jul 2026 21:53:30 +0800
Subject: [PATCH 1/9] feat(thinking): preserve cross-protocol summary
 visibility

---
 .../runtime/executor/aistudio_executor.go     |   2 +-
 .../executor/antigravity_executor_execute.go  |   4 +-
 .../executor/antigravity_executor_stream.go   |   2 +-
 .../executor/antigravity_executor_tokens.go   |   2 +-
 .../executor/claude_executor_execute.go       |   3 -
 .../executor/claude_executor_request.go       |  21 -
 .../executor/claude_executor_stream.go        |   3 -
 .../runtime/executor/claude_executor_test.go  |  39 --
 .../runtime/executor/codex_openai_images.go   |   2 +-
 .../runtime/executor/gemini_executor_test.go  |   4 +-
 .../executor/helps/model_capabilities.go      |  11 +-
 internal/runtime/executor/helps/thinking.go   |  12 +
 internal/runtime/executor/kimi_executor.go    |   4 +-
 internal/thinking/apply.go                    |  51 +-
 .../thinking/apply_configured_api_key_test.go |  33 ++
 .../thinking/provider/antigravity/apply.go    |  77 ++-
 internal/thinking/provider/claude/apply.go    |   4 +
 internal/thinking/provider/gemini/apply.go    |  77 +--
 .../thinking/provider/interactions/apply.go   |  67 ++-
 internal/thinking/strip.go                    |   4 +-
 internal/thinking/summary.go                  | 456 ++++++++++++++++++
 internal/thinking/summary_test.go             | 215 +++++++++
 .../claude/antigravity_claude_request.go      |   2 -
 .../claude/antigravity_claude_request_test.go |   4 +-
 .../interactions_antigravity_request.go       |  19 +-
 .../antigravity_openai_request.go             |  64 ++-
 .../antigravity_openai_request_test.go        |  61 ++-
 .../codex/claude/codex_claude_request.go      |   4 +-
 .../codex/gemini/codex_gemini_request.go      |   4 +-
 .../interactions_codex_request.go             |  24 +-
 .../chat-completions/codex_openai_request.go  |   4 +-
 .../gemini/claude/gemini_claude_request.go    |   2 -
 .../interactions_gemini_common.go             |  19 +-
 .../chat-completions/gemini_openai_request.go |   2 -
 .../gemini_openai-responses_request.go        |   2 -
 sdk/translator/registry.go                    |   5 +-
 sdk/translator/registry_summary_test.go       | 127 +++++
 test/summary_intent_translation_test.go       | 226 +++++++++
 test/thinking_conversion_test.go              | 373 ++++++++++++--
 39 files changed, 1678 insertions(+), 357 deletions(-)
 create mode 100644 internal/runtime/executor/helps/thinking.go
 create mode 100644 internal/thinking/summary.go
 create mode 100644 internal/thinking/summary_test.go
 create mode 100644 sdk/translator/registry_summary_test.go
 create mode 100644 test/summary_intent_translation_test.go

diff --git a/internal/runtime/executor/aistudio_executor.go b/internal/runtime/executor/aistudio_executor.go
index ba8b006a..d2e78eca 100644
--- a/internal/runtime/executor/aistudio_executor.go
+++ b/internal/runtime/executor/aistudio_executor.go
@@ -461,7 +461,7 @@ func (e *AIStudioExecutor) translateRequest(ctx context.Context, req cliproxyexe
 	originalPayload := originalPayloadSource
 	originalTranslated := helps.TranslateRequestWithCodexMultiAgentV2(ctx, opts.Headers, e.cfg, from, to, baseModel, originalPayload, stream)
 	payload := helps.TranslateRequestWithCodexMultiAgentV2(ctx, opts.Headers, e.cfg, from, to, baseModel, req.Payload, stream)
-	payload, err := thinking.ApplyThinking(payload, req.Model, from.String(), to.String(), e.Identifier())
+	payload, err := helps.ApplyThinkingWithSourcePayload(payload, req.Payload, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
 		return nil, translatedPayload{}, err
 	}
diff --git a/internal/runtime/executor/antigravity_executor_execute.go b/internal/runtime/executor/antigravity_executor_execute.go
index 471c9dc3..77bce648 100644
--- a/internal/runtime/executor/antigravity_executor_execute.go
+++ b/internal/runtime/executor/antigravity_executor_execute.go
@@ -68,7 +68,7 @@ func (e *AntigravityExecutor) Execute(ctx context.Context, auth *cliproxyauth.Au
 	originalTranslated := helps.TranslateRequestWithCodexMultiAgentV2(ctx, opts.Headers, e.cfg, from, to, baseModel, originalPayload, false)
 	translated := helps.TranslateRequestWithCodexMultiAgentV2(ctx, opts.Headers, e.cfg, from, to, baseModel, req.Payload, false)
 
-	translated, err = thinking.ApplyThinking(translated, req.Model, from.String(), to.String(), e.Identifier())
+	translated, err = helps.ApplyThinkingWithSourcePayload(translated, req.Payload, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
 		return resp, err
 	}
@@ -290,7 +290,7 @@ func (e *AntigravityExecutor) executeClaudeNonStream(ctx context.Context, auth *
 	originalTranslated := helps.TranslateRequestWithCodexMultiAgentV2(ctx, opts.Headers, e.cfg, from, to, baseModel, originalPayload, true)
 	translated := helps.TranslateRequestWithCodexMultiAgentV2(ctx, opts.Headers, e.cfg, from, to, baseModel, req.Payload, true)
 
-	translated, err = thinking.ApplyThinking(translated, req.Model, from.String(), to.String(), e.Identifier())
+	translated, err = helps.ApplyThinkingWithSourcePayload(translated, req.Payload, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
 		return resp, err
 	}
diff --git a/internal/runtime/executor/antigravity_executor_stream.go b/internal/runtime/executor/antigravity_executor_stream.go
index b90fc84f..d0aa0725 100644
--- a/internal/runtime/executor/antigravity_executor_stream.go
+++ b/internal/runtime/executor/antigravity_executor_stream.go
@@ -63,7 +63,7 @@ func (e *AntigravityExecutor) ExecuteStream(ctx context.Context, auth *cliproxya
 	originalTranslated := helps.TranslateRequestWithCodexMultiAgentV2(ctx, opts.Headers, e.cfg, from, to, baseModel, originalPayload, true)
 	translated := helps.TranslateRequestWithCodexMultiAgentV2(ctx, opts.Headers, e.cfg, from, to, baseModel, req.Payload, true)
 
-	translated, err = thinking.ApplyThinking(translated, req.Model, from.String(), to.String(), e.Identifier())
+	translated, err = helps.ApplyThinkingWithSourcePayload(translated, req.Payload, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
 		return nil, err
 	}
diff --git a/internal/runtime/executor/antigravity_executor_tokens.go b/internal/runtime/executor/antigravity_executor_tokens.go
index 98d1d561..45867ee8 100644
--- a/internal/runtime/executor/antigravity_executor_tokens.go
+++ b/internal/runtime/executor/antigravity_executor_tokens.go
@@ -50,7 +50,7 @@ func (e *AntigravityExecutor) CountTokens(ctx context.Context, auth *cliproxyaut
 	// Prepare payload once (doesn't depend on baseURL)
 	payload := helps.TranslateRequestWithCodexMultiAgentV2(ctx, opts.Headers, e.cfg, from, to, baseModel, req.Payload, false)
 
-	payload, err := thinking.ApplyThinking(payload, req.Model, from.String(), to.String(), e.Identifier())
+	payload, err := helps.ApplyThinkingWithSourcePayload(payload, req.Payload, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
 		return cliproxyexecutor.Response{}, err
 	}
diff --git a/internal/runtime/executor/claude_executor_execute.go b/internal/runtime/executor/claude_executor_execute.go
index 8f84ec6e..0a57fad1 100644
--- a/internal/runtime/executor/claude_executor_execute.go
+++ b/internal/runtime/executor/claude_executor_execute.go
@@ -67,9 +67,6 @@ func (e *ClaudeExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 	// Disable thinking if tool_choice forces tool use (Anthropic API constraint)
 	body = disableThinkingIfToolChoiceForced(body)
 	body = normalizeClaudeSamplingForUpstream(body)
-	// Claude OAuth (and this executor's redact-thinking beta) returns signature-only
-	// thinking blocks unless display is set to "summarized".
-	body = ensureClaudeThinkingDisplay(body)
 
 	// Auto-inject cache_control if missing (optimization for ClawdBot/clients without caching support)
 	if countCacheControls(body) == 0 {
diff --git a/internal/runtime/executor/claude_executor_request.go b/internal/runtime/executor/claude_executor_request.go
index 86874440..fdded306 100644
--- a/internal/runtime/executor/claude_executor_request.go
+++ b/internal/runtime/executor/claude_executor_request.go
@@ -78,27 +78,6 @@ func normalizeClaudeSamplingForUpstream(body []byte) []byte {
 	return body
 }
 
-// ensureClaudeThinkingDisplay defaults thinking.display to "summarized" when thinking
-// is active and the client did not set display. Without this, Claude backends that
-// enable redact-thinking return signature-only thinking blocks (empty thinking text).
-// Explicit client values such as "omitted" are preserved.
-func ensureClaudeThinkingDisplay(body []byte) []byte {
-	thinkingType := strings.ToLower(strings.TrimSpace(gjson.GetBytes(body, "thinking.type").String()))
-	switch thinkingType {
-	case "enabled", "adaptive", "auto":
-	default:
-		return body
-	}
-	if display := strings.TrimSpace(gjson.GetBytes(body, "thinking.display").String()); display != "" {
-		return body
-	}
-	out, err := sjson.SetBytes(body, "thinking.display", "summarized")
-	if err != nil {
-		return body
-	}
-	return out
-}
-
 type compositeReadCloser struct {
 	io.Reader
 	closers []func() error
diff --git a/internal/runtime/executor/claude_executor_stream.go b/internal/runtime/executor/claude_executor_stream.go
index 9167e056..83dc7cfb 100644
--- a/internal/runtime/executor/claude_executor_stream.go
+++ b/internal/runtime/executor/claude_executor_stream.go
@@ -67,9 +67,6 @@ func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 	// Disable thinking if tool_choice forces tool use (Anthropic API constraint)
 	body = disableThinkingIfToolChoiceForced(body)
 	body = normalizeClaudeSamplingForUpstream(body)
-	// Claude OAuth (and this executor's redact-thinking beta) returns signature-only
-	// thinking blocks unless display is set to "summarized".
-	body = ensureClaudeThinkingDisplay(body)
 
 	// Auto-inject cache_control if missing (optimization for ClawdBot/clients without caching support)
 	if countCacheControls(body) == 0 {
diff --git a/internal/runtime/executor/claude_executor_test.go b/internal/runtime/executor/claude_executor_test.go
index 3e2946f7..45831b18 100644
--- a/internal/runtime/executor/claude_executor_test.go
+++ b/internal/runtime/executor/claude_executor_test.go
@@ -3110,45 +3110,6 @@ func TestClaudeExecutor_ExecuteOpenAINonStreamRestoresOAuthToolNames(t *testing.
 	}
 }
 
-func TestEnsureClaudeThinkingDisplay_SetsSummarizedWhenMissing(t *testing.T) {
-	payload := []byte(`{"thinking":{"type":"adaptive"},"output_config":{"effort":"high"}}`)
-	out := ensureClaudeThinkingDisplay(payload)
-
-	if got := gjson.GetBytes(out, "thinking.display").String(); got != "summarized" {
-		t.Fatalf("thinking.display = %q, want summarized", got)
-	}
-	if got := gjson.GetBytes(out, "thinking.type").String(); got != "adaptive" {
-		t.Fatalf("thinking.type = %q, want adaptive", got)
-	}
-}
-
-func TestEnsureClaudeThinkingDisplay_PreservesExplicitValue(t *testing.T) {
-	payload := []byte(`{"thinking":{"type":"enabled","budget_tokens":2048,"display":"omitted"}}`)
-	out := ensureClaudeThinkingDisplay(payload)
-
-	if got := gjson.GetBytes(out, "thinking.display").String(); got != "omitted" {
-		t.Fatalf("thinking.display = %q, want omitted", got)
-	}
-}
-
-func TestEnsureClaudeThinkingDisplay_SkipsWhenThinkingDisabled(t *testing.T) {
-	payload := []byte(`{"thinking":{"type":"disabled"}}`)
-	out := ensureClaudeThinkingDisplay(payload)
-
-	if gjson.GetBytes(out, "thinking.display").Exists() {
-		t.Fatalf("thinking.display should not be set when thinking is disabled: %s", out)
-	}
-}
-
-func TestEnsureClaudeThinkingDisplay_SkipsWhenThinkingMissing(t *testing.T) {
-	payload := []byte(`{"messages":[{"role":"user","content":"hi"}]}`)
-	out := ensureClaudeThinkingDisplay(payload)
-
-	if gjson.GetBytes(out, "thinking").Exists() {
-		t.Fatalf("thinking should remain absent: %s", out)
-	}
-}
-
 func TestPrependToFirstUserMessage_KeepsToolResultBlocksFirst(t *testing.T) {
 	// A conversation that opens on an assistant tool_use makes the first user
 	// message a tool_result carrier. Anthropic requires those blocks to stay at
diff --git a/internal/runtime/executor/codex_openai_images.go b/internal/runtime/executor/codex_openai_images.go
index 6a514a6a..3251489e 100644
--- a/internal/runtime/executor/codex_openai_images.go
+++ b/internal/runtime/executor/codex_openai_images.go
@@ -674,7 +674,7 @@ func (e *CodexExecutor) prepareCodexOpenAIImageBody(body []byte, req cliproxyexe
 		mainModel = codexOpenAIImagesMainModel
 	}
 	var errThinking error
-	out, errThinking = thinking.ApplyThinking(out, mainModel, codexOpenAIImageSourceFormat, "codex", e.Identifier())
+	out, errThinking = helps.ApplyThinkingWithSourcePayload(out, body, mainModel, codexOpenAIImageSourceFormat, "codex", e.Identifier())
 	if errThinking != nil {
 		return nil, errThinking
 	}
diff --git a/internal/runtime/executor/gemini_executor_test.go b/internal/runtime/executor/gemini_executor_test.go
index 6a22e4e7..4b2a720f 100644
--- a/internal/runtime/executor/gemini_executor_test.go
+++ b/internal/runtime/executor/gemini_executor_test.go
@@ -671,8 +671,8 @@ func TestGeminiExecutorNativeInteractionsAppliesThinkingSuffix(t *testing.T) {
 	if got := gjson.GetBytes(upstreamBody, "generation_config.thinking_level").String(); got != "high" {
 		t.Fatalf("thinking_level = %q, want high. Body: %s", got, string(upstreamBody))
 	}
-	if got := gjson.GetBytes(upstreamBody, "generation_config.thinking_summaries").String(); got != "auto" {
-		t.Fatalf("thinking_summaries = %q, want auto. Body: %s", got, string(upstreamBody))
+	if gjson.GetBytes(upstreamBody, "generation_config.thinking_summaries").Exists() {
+		t.Fatalf("thinking_summaries should be absent without explicit summary intent. Body: %s", string(upstreamBody))
 	}
 }
 
diff --git a/internal/runtime/executor/helps/model_capabilities.go b/internal/runtime/executor/helps/model_capabilities.go
index 8021561c..8bf6723d 100644
--- a/internal/runtime/executor/helps/model_capabilities.go
+++ b/internal/runtime/executor/helps/model_capabilities.go
@@ -9,12 +9,13 @@ import (
 // ApplyRequestThinking preserves the registry lookup path unless the auth
 // manager bound an exact configured API-key model definition to this attempt.
 func ApplyRequestThinking(body []byte, req cliproxyexecutor.Request, opts cliproxyexecutor.Options, fromFormat, toFormat, provider string) ([]byte, error) {
+	sourceBody := opts.OriginalRequest
+	if len(sourceBody) == 0 {
+		sourceBody = req.Payload
+	}
 	if modelInfo, ok := cliproxyauth.ResolvedAPIKeyModelInfo(req); ok {
-		sourceBody := opts.OriginalRequest
-		if len(sourceBody) == 0 {
-			sourceBody = req.Payload
-		}
 		return thinking.ApplyThinkingWithModelInfo(body, sourceBody, req.Model, fromFormat, toFormat, provider, modelInfo)
 	}
-	return thinking.ApplyThinking(body, req.Model, fromFormat, toFormat, provider)
+	summaryConfig := thinking.ExtractSummaryConfig(sourceBody, fromFormat)
+	return thinking.ApplyThinkingWithSummary(body, req.Model, fromFormat, toFormat, provider, summaryConfig)
 }
diff --git a/internal/runtime/executor/helps/thinking.go b/internal/runtime/executor/helps/thinking.go
new file mode 100644
index 00000000..49f3155c
--- /dev/null
+++ b/internal/runtime/executor/helps/thinking.go
@@ -0,0 +1,12 @@
+package helps
+
+import "github.com/router-for-me/CLIProxyAPI/v7/internal/thinking"
+
+// ApplyThinkingWithSourcePayload preserves summary visibility from the original
+// client payload while applying thinking configuration to its translated target
+// payload. A target representation alone can lose an explicit disabled summary
+// before a model suffix changes Claude thinking from disabled to adaptive.
+func ApplyThinkingWithSourcePayload(body, sourcePayload []byte, model, fromFormat, toFormat, providerKey string) ([]byte, error) {
+	summary := thinking.ExtractSummaryConfig(sourcePayload, fromFormat)
+	return thinking.ApplyThinkingWithSummary(body, model, fromFormat, toFormat, providerKey, summary)
+}
diff --git a/internal/runtime/executor/kimi_executor.go b/internal/runtime/executor/kimi_executor.go
index ec270705..d3c88145 100644
--- a/internal/runtime/executor/kimi_executor.go
+++ b/internal/runtime/executor/kimi_executor.go
@@ -113,7 +113,7 @@ func (e *KimiExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req
 		return resp, fmt.Errorf("kimi executor: failed to set model in payload: %w", err)
 	}
 
-	body, err = thinking.ApplyThinking(body, req.Model, from.String(), "kimi", e.Identifier())
+	body, err = helps.ApplyThinkingWithSourcePayload(body, req.Payload, req.Model, from.String(), "kimi", e.Identifier())
 	if err != nil {
 		return resp, err
 	}
@@ -222,7 +222,7 @@ func (e *KimiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 		return nil, fmt.Errorf("kimi executor: failed to set model in payload: %w", err)
 	}
 
-	body, err = thinking.ApplyThinking(body, req.Model, from.String(), "kimi", e.Identifier())
+	body, err = helps.ApplyThinkingWithSourcePayload(body, req.Payload, req.Model, from.String(), "kimi", e.Identifier())
 	if err != nil {
 		return nil, err
 	}
diff --git a/internal/thinking/apply.go b/internal/thinking/apply.go
index 1d25de7e..c19369c7 100644
--- a/internal/thinking/apply.go
+++ b/internal/thinking/apply.go
@@ -162,16 +162,31 @@ func IsUserDefinedModel(modelInfo *registry.ModelInfo) bool {
 //	// Without suffix - uses body config
 //	result, err := thinking.ApplyThinking(body, "gemini-2.5-pro", "gemini", "gemini", "gemini")
 func ApplyThinking(body []byte, model string, fromFormat string, toFormat string, providerKey string) ([]byte, error) {
-	return applyThinking(body, nil, model, fromFormat, toFormat, providerKey, nil, false)
+	summaryConfig := ExtractSummaryConfig(body, toFormat)
+	return applyThinking(body, nil, model, fromFormat, toFormat, providerKey, nil, false, summaryConfig)
+}
+
+// ApplyThinkingWithSummary applies canonical thinking effort while preserving
+// summary visibility extracted from the original source request. Callers that
+// translate before applying thinking must pass the source config explicitly:
+// a target Claude body can temporarily lack display while disabled thinking is
+// being rewritten by a model suffix.
+func ApplyThinkingWithSummary(body []byte, model string, fromFormat string, toFormat string, providerKey string, summaryConfig SummaryConfig) ([]byte, error) {
+	return applyThinking(body, nil, model, fromFormat, toFormat, providerKey, nil, false, summaryConfig)
 }
 
 // ApplyThinkingWithModelInfo applies thinking with the exact configured model
-// definition selected for an API-key execution attempt.
+// definition selected for an API-key execution attempt while preserving summary
+// visibility from the original source body.
 func ApplyThinkingWithModelInfo(body, sourceBody []byte, model string, fromFormat string, toFormat string, providerKey string, modelInfo *registry.ModelInfo) ([]byte, error) {
-	return applyThinking(body, sourceBody, model, fromFormat, toFormat, providerKey, modelInfo, true)
+	summaryConfig := ExtractSummaryConfig(sourceBody, fromFormat)
+	if len(sourceBody) == 0 {
+		summaryConfig = ExtractSummaryConfig(body, toFormat)
+	}
+	return applyThinking(body, sourceBody, model, fromFormat, toFormat, providerKey, modelInfo, true, summaryConfig)
 }
 
-func applyThinking(body, sourceBody []byte, model string, fromFormat string, toFormat string, providerKey string, resolvedModelInfo *registry.ModelInfo, modelInfoResolved bool) ([]byte, error) {
+func applyThinking(body, sourceBody []byte, model string, fromFormat string, toFormat string, providerKey string, resolvedModelInfo *registry.ModelInfo, modelInfoResolved bool, summaryConfig SummaryConfig) ([]byte, error) {
 	providerFormat := strings.ToLower(strings.TrimSpace(toFormat))
 	if modelInfoResolved && providerFormat == "openai-response" {
 		providerFormat = "codex"
@@ -184,6 +199,9 @@ func applyThinking(body, sourceBody []byte, model string, fromFormat string, toF
 	if fromFormat == "" {
 		fromFormat = providerFormat
 	}
+	// Summary visibility is orthogonal to thinking effort. Keep the original
+	// source intent before a suffix-specific applier rewrites provider fields,
+	// then restore it after the canonical effort has been applied.
 	// 1. Route check: Get provider applier
 	applier := GetProviderApplier(providerFormat)
 	if applier == nil {
@@ -207,11 +225,11 @@ func applyThinking(body, sourceBody []byte, model string, fromFormat string, toF
 	// Unknown models are treated as user-defined so thinking config can still be applied.
 	// The upstream service is responsible for validating the configuration.
 	if IsUserDefinedModel(modelInfo) {
-		return applyUserDefinedModel(body, modelInfo, fromFormat, providerFormat, suffixResult)
+		return applyUserDefinedModel(body, modelInfo, fromFormat, providerFormat, suffixResult, summaryConfig)
 	}
 	if modelInfo.Thinking == nil {
 		config := extractThinkingConfig(body, providerFormat)
-		if hasThinkingConfig(config) {
+		if hasThinkingConfig(config) || summaryConfig.Mode != SummaryUnspecified {
 			log.WithFields(log.Fields{
 				"model":    baseModel,
 				"provider": providerFormat,
@@ -259,7 +277,7 @@ func applyThinking(body, sourceBody []byte, model string, fromFormat string, toF
 			"provider": providerFormat,
 			"model":    modelInfo.ID,
 		}).Debug("thinking: no config found, passthrough |")
-		return body, nil
+		return applySummaryConfigForModel(body, providerFormat, baseModel, modelInfo, summaryConfig), nil
 	}
 	if modelInfoResolved && config.Mode == ModeLevel && modelInfo != nil && modelInfo.Thinking != nil && shouldMapConfiguredHighIntent(fromFormat, providerFormat, modelInfo) {
 		config.Level = mapConfiguredHighIntent(config.Level, modelInfo)
@@ -296,8 +314,13 @@ func applyThinking(body, sourceBody []byte, model string, fromFormat string, toF
 		"level":    validated.Level,
 	}).Debug("thinking: processed config to apply |")
 
-	// 6. Apply configuration using provider-specific applier
-	return applier.Apply(body, *validated, modelInfo)
+	// 6. Apply configuration using provider-specific applier, then restore the
+	// target summary intent that was explicit before suffix processing.
+	applied, err := applier.Apply(body, *validated, modelInfo)
+	if err != nil {
+		return applied, err
+	}
+	return applySummaryConfigForModel(applied, providerFormat, baseModel, modelInfo, summaryConfig), nil
 }
 
 func shouldMapConfiguredHighIntent(fromFormat, toFormat string, modelInfo *registry.ModelInfo) bool {
@@ -386,7 +409,7 @@ func parseSuffixToConfig(rawSuffix, provider, model string) ThinkingConfig {
 
 // applyUserDefinedModel applies thinking configuration for user-defined models
 // without ThinkingSupport validation.
-func applyUserDefinedModel(body []byte, modelInfo *registry.ModelInfo, fromFormat, toFormat string, suffixResult SuffixResult) ([]byte, error) {
+func applyUserDefinedModel(body []byte, modelInfo *registry.ModelInfo, fromFormat, toFormat string, suffixResult SuffixResult, summaryConfig SummaryConfig) ([]byte, error) {
 	// Get model ID for logging
 	modelID := ""
 	if modelInfo != nil {
@@ -427,7 +450,7 @@ func applyUserDefinedModel(body []byte, modelInfo *registry.ModelInfo, fromForma
 			"model":    modelID,
 			"provider": toFormat,
 		}).Debug("thinking: user-defined model, passthrough (no config) |")
-		return body, nil
+		return applySummaryConfigForModel(body, toFormat, modelID, modelInfo, summaryConfig), nil
 	}
 
 	applier := GetProviderApplier(toFormat)
@@ -447,7 +470,11 @@ func applyUserDefinedModel(body []byte, modelInfo *registry.ModelInfo, fromForma
 		"budget":   config.Budget,
 		"level":    config.Level,
 	}).Debug("thinking: processed config to apply |")
-	return applier.Apply(body, config, modelInfo)
+	applied, err := applier.Apply(body, config, modelInfo)
+	if err != nil {
+		return applied, err
+	}
+	return applySummaryConfigForModel(applied, toFormat, modelID, modelInfo, summaryConfig), nil
 }
 
 func normalizeUserDefinedConfig(config ThinkingConfig, fromFormat, toFormat string) ThinkingConfig {
diff --git a/internal/thinking/apply_configured_api_key_test.go b/internal/thinking/apply_configured_api_key_test.go
index 81e908fb..5aa3ce9d 100644
--- a/internal/thinking/apply_configured_api_key_test.go
+++ b/internal/thinking/apply_configured_api_key_test.go
@@ -92,6 +92,39 @@ func TestApplyThinkingWithModelInfoKeepsSameFamilyValidationStrict(t *testing.T)
 	}
 }
 
+func TestApplyThinkingWithModelInfoAppliesSummaryOnlyClaudeVisibility(t *testing.T) {
+	modelInfo := &registry.ModelInfo{
+		ID:       "private-claude",
+		Type:     "claude",
+		Thinking: &registry.ThinkingSupport{Levels: []string{"high"}},
+	}
+	for _, test := range []struct {
+		name    string
+		source  string
+		display string
+	}{
+		{name: "enabled", source: `{"reasoning":{"summary":"auto"}}`, display: "summarized"},
+		{name: "disabled", source: `{"reasoning":{"summary":null}}`, display: "omitted"},
+	} {
+		t.Run(test.name, func(t *testing.T) {
+			out, err := thinking.ApplyThinkingWithModelInfo(
+				[]byte(`{"model":"private-claude","max_tokens":32000}`),
+				[]byte(test.source),
+				"private-claude", "openai-response", "claude", "claude", modelInfo,
+			)
+			if err != nil {
+				t.Fatalf("ApplyThinkingWithModelInfo() error = %v", err)
+			}
+			if got := gjson.GetBytes(out, "thinking.type").String(); got != "adaptive" {
+				t.Fatalf("thinking.type = %q, want adaptive; body=%s", got, out)
+			}
+			if got := gjson.GetBytes(out, "thinking.display").String(); got != test.display {
+				t.Fatalf("thinking.display = %q, want %q; body=%s", got, test.display, out)
+			}
+		})
+	}
+}
+
 func TestApplyThinkingWithModelInfoUsesOriginalResponsesEffort(t *testing.T) {
 	modelInfo := &registry.ModelInfo{
 		ID:       "claude-upstream",
diff --git a/internal/thinking/provider/antigravity/apply.go b/internal/thinking/provider/antigravity/apply.go
index cb0659f1..968ee09d 100644
--- a/internal/thinking/provider/antigravity/apply.go
+++ b/internal/thinking/provider/antigravity/apply.go
@@ -98,19 +98,19 @@ func (a *Applier) applyLevelFormat(body []byte, config thinking.ThinkingConfig)
 	result, _ := sjson.DeleteBytes(body, "request.generationConfig.thinkingConfig.thinkingBudget")
 	result, _ = sjson.DeleteBytes(result, "request.generationConfig.thinkingConfig.thinking_budget")
 	result, _ = sjson.DeleteBytes(result, "request.generationConfig.thinkingConfig.thinking_level")
-	// Normalize includeThoughts field name to avoid oneof conflicts in upstream JSON parsing.
+	// Normalize includeThoughts field name and retain only documented booleans.
+	result, _ = sjson.DeleteBytes(result, "request.generationConfig.thinkingConfig.includeThoughts")
 	result, _ = sjson.DeleteBytes(result, "request.generationConfig.thinkingConfig.include_thoughts")
 
 	if config.Mode == thinking.ModeNone {
 		if config.Budget == 0 && config.Level == "" {
 			result, _ = sjson.DeleteBytes(result, "request.generationConfig.thinkingConfig")
-			return result, nil
+			return applyAntigravityIncludeThoughts(result, body), nil
 		}
-		result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.includeThoughts", false)
 		if config.Level != "" {
 			result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.thinkingLevel", string(config.Level))
 		}
-		return result, nil
+		return applyAntigravityIncludeThoughts(result, body), nil
 	}
 
 	// Only handle ModeLevel - budget conversion should be done by upper layer
@@ -120,17 +120,7 @@ func (a *Applier) applyLevelFormat(body []byte, config thinking.ThinkingConfig)
 
 	level := string(config.Level)
 	result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.thinkingLevel", level)
-
-	// Respect user's explicit includeThoughts setting from original body; default to true if not set
-	// Support both camelCase and snake_case variants
-	includeThoughts := true
-	if inc := gjson.GetBytes(body, "request.generationConfig.thinkingConfig.includeThoughts"); inc.Exists() {
-		includeThoughts = inc.Bool()
-	} else if inc := gjson.GetBytes(body, "request.generationConfig.thinkingConfig.include_thoughts"); inc.Exists() {
-		includeThoughts = inc.Bool()
-	}
-	result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.includeThoughts", includeThoughts)
-	return result, nil
+	return applyAntigravityIncludeThoughts(result, body), nil
 }
 
 func (a *Applier) applyBudgetFormat(body []byte, config thinking.ThinkingConfig, modelInfo *registry.ModelInfo, isClaude bool) ([]byte, error) {
@@ -138,7 +128,8 @@ func (a *Applier) applyBudgetFormat(body []byte, config thinking.ThinkingConfig,
 	result, _ := sjson.DeleteBytes(body, "request.generationConfig.thinkingConfig.thinkingLevel")
 	result, _ = sjson.DeleteBytes(result, "request.generationConfig.thinkingConfig.thinking_level")
 	result, _ = sjson.DeleteBytes(result, "request.generationConfig.thinkingConfig.thinking_budget")
-	// Normalize includeThoughts field name to avoid oneof conflicts in upstream JSON parsing.
+	// Normalize includeThoughts field name and retain only documented booleans.
+	result, _ = sjson.DeleteBytes(result, "request.generationConfig.thinkingConfig.includeThoughts")
 	result, _ = sjson.DeleteBytes(result, "request.generationConfig.thinkingConfig.include_thoughts")
 
 	budget := config.Budget
@@ -146,46 +137,32 @@ func (a *Applier) applyBudgetFormat(body []byte, config thinking.ThinkingConfig,
 	// Apply Claude-specific constraints first to get the final budget value
 	if isClaude && modelInfo != nil {
 		budget, result = a.normalizeClaudeBudget(budget, result, modelInfo)
-		// Check if budget was removed entirely
+		// Check if the thinking amount was removed entirely. Summary visibility is
+		// independent, so retain an explicit includeThoughts control if present.
 		if budget == -2 {
-			return result, nil
+			return applyAntigravityIncludeThoughts(result, body), nil
 		}
 	}
 
-	// For ModeNone, always set includeThoughts to false regardless of user setting.
-	// This ensures that when user requests budget=0 (disable thinking output),
-	// the includeThoughts is correctly set to false even if budget is clamped to min.
-	if config.Mode == thinking.ModeNone {
-		result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.thinkingBudget", budget)
-		result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.includeThoughts", false)
-		return result, nil
-	}
-
-	// Determine includeThoughts: respect user's explicit setting from original body if provided
-	// Support both camelCase and snake_case variants
-	var includeThoughts bool
-	var userSetIncludeThoughts bool
-	if inc := gjson.GetBytes(body, "request.generationConfig.thinkingConfig.includeThoughts"); inc.Exists() {
-		includeThoughts = inc.Bool()
-		userSetIncludeThoughts = true
-	} else if inc := gjson.GetBytes(body, "request.generationConfig.thinkingConfig.include_thoughts"); inc.Exists() {
-		includeThoughts = inc.Bool()
-		userSetIncludeThoughts = true
-	}
-
-	if !userSetIncludeThoughts {
-		// No explicit setting, use default logic based on mode
-		switch config.Mode {
-		case thinking.ModeAuto:
-			includeThoughts = true
-		default:
-			includeThoughts = budget > 0
+	result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.thinkingBudget", budget)
+	return applyAntigravityIncludeThoughts(result, body), nil
+}
+
+func applyAntigravityIncludeThoughts(result, original []byte) []byte {
+	for _, path := range []string{
+		"request.generationConfig.thinkingConfig.includeThoughts",
+		"request.generationConfig.thinkingConfig.include_thoughts",
+	} {
+		switch value := gjson.GetBytes(original, path); value.Type {
+		case gjson.True:
+			result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.includeThoughts", true)
+			return result
+		case gjson.False:
+			result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.includeThoughts", false)
+			return result
 		}
 	}
-
-	result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.thinkingBudget", budget)
-	result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.includeThoughts", includeThoughts)
-	return result, nil
+	return result
 }
 
 // normalizeClaudeBudget applies Claude-specific constraints to thinking budget.
diff --git a/internal/thinking/provider/claude/apply.go b/internal/thinking/provider/claude/apply.go
index 140a8135..97f02849 100644
--- a/internal/thinking/provider/claude/apply.go
+++ b/internal/thinking/provider/claude/apply.go
@@ -87,6 +87,8 @@ func (a *Applier) Apply(body []byte, config thinking.ThinkingConfig, modelInfo *
 	case thinking.ModeNone:
 		result, _ := sjson.SetBytes(body, "thinking.type", "disabled")
 		result, _ = sjson.DeleteBytes(result, "thinking.budget_tokens")
+		// Summary display only applies to an active thinking block.
+		result, _ = sjson.DeleteBytes(result, "thinking.display")
 		result, _ = sjson.DeleteBytes(result, "output_config.effort")
 		if oc := gjson.GetBytes(result, "output_config"); oc.Exists() && oc.IsObject() && len(oc.Map()) == 0 {
 			result, _ = sjson.DeleteBytes(result, "output_config")
@@ -231,6 +233,8 @@ func applyCompatibleClaude(body []byte, config thinking.ThinkingConfig) ([]byte,
 	case thinking.ModeNone:
 		result, _ := sjson.SetBytes(body, "thinking.type", "disabled")
 		result, _ = sjson.DeleteBytes(result, "thinking.budget_tokens")
+		// Summary display only applies to an active thinking block.
+		result, _ = sjson.DeleteBytes(result, "thinking.display")
 		result, _ = sjson.DeleteBytes(result, "output_config.effort")
 		if oc := gjson.GetBytes(result, "output_config"); oc.Exists() && oc.IsObject() && len(oc.Map()) == 0 {
 			result, _ = sjson.DeleteBytes(result, "output_config")
diff --git a/internal/thinking/provider/gemini/apply.go b/internal/thinking/provider/gemini/apply.go
index 92a8d7ec..c332e9ef 100644
--- a/internal/thinking/provider/gemini/apply.go
+++ b/internal/thinking/provider/gemini/apply.go
@@ -114,27 +114,27 @@ func (a *Applier) applyCompatible(body []byte, config thinking.ThinkingConfig) (
 
 func (a *Applier) applyLevelFormat(body []byte, config thinking.ThinkingConfig) ([]byte, error) {
 	// ModeNone semantics:
-	//   - ModeNone + Budget=0: remove thinkingConfig to disable thinking
-	//   - ModeNone + Budget>0: forced to think but hide output (includeThoughts=false)
-	// ValidateConfig sets config.Level to the lowest level when ModeNone + Budget > 0.
+	//   - ModeNone + Budget=0: remove the thinking amount configuration.
+	//   - ModeNone + Budget>0: clamp to the model's lowest supported amount.
+	// Summary visibility remains independent and is restored only when explicitly set.
 
 	// Remove conflicting fields to avoid both thinkingLevel and thinkingBudget in output
 	result, _ := sjson.DeleteBytes(body, "generationConfig.thinkingConfig.thinkingBudget")
 	result, _ = sjson.DeleteBytes(result, "generationConfig.thinkingConfig.thinking_budget")
 	result, _ = sjson.DeleteBytes(result, "generationConfig.thinkingConfig.thinking_level")
-	// Normalize includeThoughts field name to avoid oneof conflicts in upstream JSON parsing.
+	// Normalize includeThoughts field name and retain only documented booleans.
+	result, _ = sjson.DeleteBytes(result, "generationConfig.thinkingConfig.includeThoughts")
 	result, _ = sjson.DeleteBytes(result, "generationConfig.thinkingConfig.include_thoughts")
 
 	if config.Mode == thinking.ModeNone {
 		if config.Budget == 0 && config.Level == "" {
 			result, _ = sjson.DeleteBytes(result, "generationConfig.thinkingConfig")
-			return result, nil
+			return applyGeminiIncludeThoughts(result, body), nil
 		}
-		result, _ = sjson.SetBytes(result, "generationConfig.thinkingConfig.includeThoughts", false)
 		if config.Level != "" {
 			result, _ = sjson.SetBytes(result, "generationConfig.thinkingConfig.thinkingLevel", string(config.Level))
 		}
-		return result, nil
+		return applyGeminiIncludeThoughts(result, body), nil
 	}
 
 	// Only handle ModeLevel - budget conversion should be done by upper layer
@@ -144,17 +144,7 @@ func (a *Applier) applyLevelFormat(body []byte, config thinking.ThinkingConfig)
 
 	level := string(config.Level)
 	result, _ = sjson.SetBytes(result, "generationConfig.thinkingConfig.thinkingLevel", level)
-
-	// Respect user's explicit includeThoughts setting from original body; default to true if not set
-	// Support both camelCase and snake_case variants
-	includeThoughts := true
-	if inc := gjson.GetBytes(body, "generationConfig.thinkingConfig.includeThoughts"); inc.Exists() {
-		includeThoughts = inc.Bool()
-	} else if inc := gjson.GetBytes(body, "generationConfig.thinkingConfig.include_thoughts"); inc.Exists() {
-		includeThoughts = inc.Bool()
-	}
-	result, _ = sjson.SetBytes(result, "generationConfig.thinkingConfig.includeThoughts", includeThoughts)
-	return result, nil
+	return applyGeminiIncludeThoughts(result, body), nil
 }
 
 func (a *Applier) applyBudgetFormat(body []byte, config thinking.ThinkingConfig) ([]byte, error) {
@@ -162,43 +152,28 @@ func (a *Applier) applyBudgetFormat(body []byte, config thinking.ThinkingConfig)
 	result, _ := sjson.DeleteBytes(body, "generationConfig.thinkingConfig.thinkingLevel")
 	result, _ = sjson.DeleteBytes(result, "generationConfig.thinkingConfig.thinking_level")
 	result, _ = sjson.DeleteBytes(result, "generationConfig.thinkingConfig.thinking_budget")
-	// Normalize includeThoughts field name to avoid oneof conflicts in upstream JSON parsing.
+	// Normalize includeThoughts field name and retain only documented booleans.
+	result, _ = sjson.DeleteBytes(result, "generationConfig.thinkingConfig.includeThoughts")
 	result, _ = sjson.DeleteBytes(result, "generationConfig.thinkingConfig.include_thoughts")
 
 	budget := config.Budget
+	result, _ = sjson.SetBytes(result, "generationConfig.thinkingConfig.thinkingBudget", budget)
+	return applyGeminiIncludeThoughts(result, body), nil
+}
 
-	// For ModeNone, always set includeThoughts to false regardless of user setting.
-	// This ensures that when user requests budget=0 (disable thinking output),
-	// the includeThoughts is correctly set to false even if budget is clamped to min.
-	if config.Mode == thinking.ModeNone {
-		result, _ = sjson.SetBytes(result, "generationConfig.thinkingConfig.thinkingBudget", budget)
-		result, _ = sjson.SetBytes(result, "generationConfig.thinkingConfig.includeThoughts", false)
-		return result, nil
-	}
-
-	// Determine includeThoughts: respect user's explicit setting from original body if provided
-	// Support both camelCase and snake_case variants
-	var includeThoughts bool
-	var userSetIncludeThoughts bool
-	if inc := gjson.GetBytes(body, "generationConfig.thinkingConfig.includeThoughts"); inc.Exists() {
-		includeThoughts = inc.Bool()
-		userSetIncludeThoughts = true
-	} else if inc := gjson.GetBytes(body, "generationConfig.thinkingConfig.include_thoughts"); inc.Exists() {
-		includeThoughts = inc.Bool()
-		userSetIncludeThoughts = true
-	}
-
-	if !userSetIncludeThoughts {
-		// No explicit setting, use default logic based on mode
-		switch config.Mode {
-		case thinking.ModeAuto:
-			includeThoughts = true
-		default:
-			includeThoughts = budget > 0
+func applyGeminiIncludeThoughts(result, original []byte) []byte {
+	for _, path := range []string{
+		"generationConfig.thinkingConfig.includeThoughts",
+		"generationConfig.thinkingConfig.include_thoughts",
+	} {
+		switch value := gjson.GetBytes(original, path); value.Type {
+		case gjson.True:
+			result, _ = sjson.SetBytes(result, "generationConfig.thinkingConfig.includeThoughts", true)
+			return result
+		case gjson.False:
+			result, _ = sjson.SetBytes(result, "generationConfig.thinkingConfig.includeThoughts", false)
+			return result
 		}
 	}
-
-	result, _ = sjson.SetBytes(result, "generationConfig.thinkingConfig.thinkingBudget", budget)
-	result, _ = sjson.SetBytes(result, "generationConfig.thinkingConfig.includeThoughts", includeThoughts)
-	return result, nil
+	return result
 }
diff --git a/internal/thinking/provider/interactions/apply.go b/internal/thinking/provider/interactions/apply.go
index 2951b511..c644f5ad 100644
--- a/internal/thinking/provider/interactions/apply.go
+++ b/internal/thinking/provider/interactions/apply.go
@@ -34,11 +34,11 @@ func (a *Applier) Apply(body []byte, config thinking.ThinkingConfig, modelInfo *
 	result := stripInteractionsThinkingFields(body)
 	switch config.Mode {
 	case thinking.ModeLevel:
-		return applyInteractionsLevel(result, body, string(config.Level), modelInfo, "auto"), nil
+		return applyInteractionsLevel(result, body, string(config.Level), modelInfo), nil
 	case thinking.ModeBudget:
-		return applyInteractionsBudget(result, body, config.Budget, modelInfo, "auto"), nil
+		return applyInteractionsBudget(result, body, config.Budget, modelInfo), nil
 	case thinking.ModeAuto:
-		return setInteractionsThinkingSummaries(result, body, "auto"), nil
+		return setInteractionsThinkingSummaries(result, body), nil
 	case thinking.ModeNone:
 		return applyInteractionsNone(result, body, config, modelInfo), nil
 	default:
@@ -46,38 +46,38 @@ func (a *Applier) Apply(body []byte, config thinking.ThinkingConfig, modelInfo *
 	}
 }
 
-func applyInteractionsBudget(result, original []byte, budget int, modelInfo *registry.ModelInfo, summariesFallback string) []byte {
+func applyInteractionsBudget(result, original []byte, budget int, modelInfo *registry.ModelInfo) []byte {
 	level, ok := thinking.ConvertBudgetToLevel(budget)
 	if !ok {
-		return result
+		return setInteractionsThinkingSummaries(result, original)
 	}
 	switch level {
-	case string(thinking.LevelNone):
-		return setInteractionsThinkingSummaries(result, original, "none")
-	case string(thinking.LevelAuto):
-		return setInteractionsThinkingSummaries(result, original, "auto")
+	case string(thinking.LevelNone), string(thinking.LevelAuto):
+		// Thinking amount and summary visibility are independent. Interactions has
+		// no wire-level "none" thinking level, so preserve only explicit summary
+		// intent and otherwise let the target model use its documented default.
+		return setInteractionsThinkingSummaries(result, original)
 	default:
-		return applyInteractionsLevel(result, original, level, modelInfo, summariesFallback)
+		return applyInteractionsLevel(result, original, level, modelInfo)
 	}
 }
 
-func applyInteractionsLevel(result, original []byte, level string, modelInfo *registry.ModelInfo, summariesFallback string) []byte {
+func applyInteractionsLevel(result, original []byte, level string, modelInfo *registry.ModelInfo) []byte {
 	level = normalizeInteractionsLevel(level, modelInfo)
-	if level == "" {
-		return result
+	if level != "" {
+		result, _ = sjson.SetBytes(result, "generation_config.thinking_level", level)
 	}
-	result, _ = sjson.SetBytes(result, "generation_config.thinking_level", level)
-	return setInteractionsThinkingSummaries(result, original, summariesFallback)
+	return setInteractionsThinkingSummaries(result, original)
 }
 
 func applyInteractionsNone(result, original []byte, config thinking.ThinkingConfig, modelInfo *registry.ModelInfo) []byte {
 	if config.Level != "" {
-		result = applyInteractionsLevel(result, original, string(config.Level), modelInfo, "none")
-	} else if config.Budget > 0 {
-		result = applyInteractionsBudget(result, original, config.Budget, modelInfo, "none")
+		return applyInteractionsLevel(result, original, string(config.Level), modelInfo)
 	}
-	result, _ = sjson.SetBytes(result, "generation_config.thinking_summaries", "none")
-	return result
+	if config.Budget > 0 {
+		return applyInteractionsBudget(result, original, config.Budget, modelInfo)
+	}
+	return setInteractionsThinkingSummaries(result, original)
 }
 
 func stripInteractionsThinkingFields(body []byte) []byte {
@@ -104,7 +104,7 @@ func stripInteractionsThinkingFields(body []byte) []byte {
 	return result
 }
 
-func setInteractionsThinkingSummaries(result, original []byte, fallback string) []byte {
+func setInteractionsThinkingSummaries(result, original []byte) []byte {
 	if value, okValue := originalInteractionsThinkingSummaries(original); okValue {
 		result, _ = sjson.SetBytes(result, "generation_config.thinking_summaries", value)
 		return result
@@ -112,16 +112,9 @@ func setInteractionsThinkingSummaries(result, original []byte, fallback string)
 	if includeThoughts, okValue := originalInteractionsIncludeThoughts(original); okValue {
 		value := "none"
 		if includeThoughts {
-			value = fallback
-			if value == "" {
-				value = "auto"
-			}
+			value = "auto"
 		}
 		result, _ = sjson.SetBytes(result, "generation_config.thinking_summaries", value)
-		return result
-	}
-	if fallback != "" {
-		result, _ = sjson.SetBytes(result, "generation_config.thinking_summaries", fallback)
 	}
 	return result
 }
@@ -132,8 +125,12 @@ func originalInteractionsThinkingSummaries(body []byte) (string, bool) {
 		"generation_config.thinkingSummaries",
 	} {
 		value := gjson.GetBytes(body, path)
-		if value.Exists() && value.Type == gjson.String {
-			return strings.ToLower(strings.TrimSpace(value.String())), true
+		if value.Type != gjson.String {
+			continue
+		}
+		switch normalized := strings.ToLower(strings.TrimSpace(value.String())); normalized {
+		case "auto", "none":
+			return normalized, true
 		}
 	}
 	return "", false
@@ -146,9 +143,11 @@ func originalInteractionsIncludeThoughts(body []byte) (bool, bool) {
 		"generation_config.thinkingConfig.include_thoughts",
 		"generation_config.thinkingConfig.includeThoughts",
 	} {
-		value := gjson.GetBytes(body, path)
-		if value.Exists() {
-			return value.Bool(), true
+		switch value := gjson.GetBytes(body, path); value.Type {
+		case gjson.True:
+			return true, true
+		case gjson.False:
+			return false, true
 		}
 	}
 	return false, false
diff --git a/internal/thinking/strip.go b/internal/thinking/strip.go
index f514a7bd..f60b7ff2 100644
--- a/internal/thinking/strip.go
+++ b/internal/thinking/strip.go
@@ -47,14 +47,14 @@ func StripThinkingConfig(body []byte, provider string) []byte {
 			"generation_config.thinkingConfig",
 		}
 	case "openai":
-		paths = []string{"reasoning_effort"}
+		paths = []string{"reasoning_effort", "reasoning"}
 	case "kimi":
 		paths = []string{
 			"reasoning_effort",
 			"thinking",
 		}
 	case "codex", "xai":
-		paths = []string{"reasoning.effort"}
+		paths = []string{"reasoning"}
 	default:
 		return body
 	}
diff --git a/internal/thinking/summary.go b/internal/thinking/summary.go
new file mode 100644
index 00000000..4a19dae1
--- /dev/null
+++ b/internal/thinking/summary.go
@@ -0,0 +1,456 @@
+package thinking
+
+import (
+	"strings"
+
+	"github.com/router-for-me/CLIProxyAPI/v7/internal/registry"
+	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
+)
+
+// SummaryMode represents whether the client explicitly requested reasoning summaries.
+type SummaryMode int
+
+const (
+	SummaryUnspecified SummaryMode = iota
+	SummaryDisabled
+	SummaryEnabled
+)
+
+// SummaryConfig is the provider-neutral reasoning-summary visibility intent.
+// Detail preserves protocols that distinguish auto, concise, and detailed summaries.
+type SummaryConfig struct {
+	Mode   SummaryMode
+	Detail string
+}
+
+// ExtractSummaryConfig reads protocol-specific summary visibility intent.
+//
+// OpenAI Chat is the one protocol where effort implies summaries: chat
+// completions has no summary field of its own, and clients that send
+// reasoning_effort have always received reasoning summaries here, so treating a
+// non-none effort as an explicit request preserves that contract. Every other
+// protocol carries a dedicated summary field, so effort alone means nothing.
+func ExtractSummaryConfig(body []byte, format string) SummaryConfig {
+	normalized := strings.ToLower(strings.TrimSpace(format))
+	// Check the format first so unsupported targets skip whole-body validation.
+	if !summaryFormatSupported(normalized) || len(body) == 0 || !gjson.ValidBytes(body) {
+		return SummaryConfig{}
+	}
+
+	switch normalized {
+	case "openai":
+		if config, ok := extractOpenAIExplicitSummaryConfig(body); ok {
+			return config
+		}
+		if effort := gjson.GetBytes(body, "reasoning_effort"); effort.Type == gjson.String {
+			value := strings.ToLower(strings.TrimSpace(effort.String()))
+			if value == "" {
+				return SummaryConfig{}
+			}
+			if value == "none" {
+				return SummaryConfig{Mode: SummaryDisabled}
+			}
+			return SummaryConfig{Mode: SummaryEnabled, Detail: "auto"}
+		}
+	case "openai-response", "codex":
+		if config, ok := responsesSummaryConfig(body, "reasoning.summary"); ok {
+			return config
+		}
+		if config, ok := responsesSummaryConfig(body, "reasoning.generate_summary"); ok {
+			return config
+		}
+	case "claude":
+		// Anthropic only accepts display alongside active adaptive/manual thinking.
+		if !claudeThinkingAcceptsDisplay(body) {
+			return SummaryConfig{}
+		}
+		if config, ok := claudeSummaryConfig(body, "thinking.display"); ok {
+			return config
+		}
+	case "gemini":
+		if config, ok := firstSummaryBoolConfig(body, []string{
+			"generationConfig.thinkingConfig.includeThoughts",
+			"generationConfig.thinkingConfig.include_thoughts",
+			"generation_config.thinking_config.include_thoughts",
+			"generation_config.thinking_config.includeThoughts",
+		}); ok {
+			return config
+		}
+	case "antigravity":
+		if config, ok := firstSummaryBoolConfig(body, []string{
+			"request.generationConfig.thinkingConfig.includeThoughts",
+			"request.generationConfig.thinkingConfig.include_thoughts",
+			"request.generationConfig.thinking_config.includeThoughts",
+			"request.generationConfig.thinking_config.include_thoughts",
+		}); ok {
+			return config
+		}
+	case "interactions":
+		for _, path := range []string{
+			"generation_config.thinking_summaries",
+			"generation_config.thinkingSummaries",
+		} {
+			if config, ok := interactionsSummaryConfig(body, path); ok {
+				return config
+			}
+		}
+	}
+
+	return SummaryConfig{}
+}
+
+// ApplySummaryConfig writes canonical summary intent in the target protocol.
+func ApplySummaryConfig(body []byte, format string, config SummaryConfig) []byte {
+	return ApplySummaryConfigForModel(body, format, "", config)
+}
+
+// ApplySummaryConfigForModel writes canonical summary intent in the target
+// protocol and uses target model capabilities when a valid target request must
+// activate thinking before it can request summaries.
+func ApplySummaryConfigForModel(body []byte, format, model string, config SummaryConfig) []byte {
+	return applySummaryConfigForModel(body, format, model, nil, config)
+}
+
+// applySummaryConfigForModel uses the resolved model definition when execution
+// selected a configured API-key model whose capability is not globally visible.
+func applySummaryConfigForModel(body []byte, format, model string, modelInfo *registry.ModelInfo, config SummaryConfig) []byte {
+	normalized := strings.ToLower(strings.TrimSpace(format))
+	if config.Mode == SummaryUnspecified || !summaryFormatSupported(normalized) || len(body) == 0 || !gjson.ValidBytes(body) {
+		return body
+	}
+
+	enabled := config.Mode == SummaryEnabled
+	switch normalized {
+	case "openai":
+		body = applyOpenAIChatSummaryConfig(body, model, enabled)
+	case "claude":
+		// Anthropic documents display as invalid with thinking.type=disabled and
+		// requires it alongside adaptive or enabled thinking. An explicit source
+		// visibility request is independent of thinking effort, so activate the
+		// target model's documented thinking mode before writing either
+		// summarized or omitted. Unspecified intent returns above and leaves the
+		// target's default untouched.
+		if !gjson.GetBytes(body, "thinking.type").Exists() {
+			body = enableClaudeThinkingForSummary(body, model, modelInfo)
+		}
+		if !claudeThinkingAcceptsDisplay(body) {
+			return body
+		}
+		value := "omitted"
+		if enabled {
+			value = "summarized"
+		}
+		body, _ = sjson.SetBytes(body, "thinking.display", value)
+	case "gemini":
+		body, _ = sjson.SetBytes(body, "generationConfig.thinkingConfig.includeThoughts", enabled)
+		for _, path := range []string{
+			"generationConfig.thinkingConfig.include_thoughts",
+			"generation_config.thinking_config.include_thoughts",
+			"generation_config.thinking_config.includeThoughts",
+		} {
+			body, _ = sjson.DeleteBytes(body, path)
+		}
+	case "antigravity":
+		body, _ = sjson.SetBytes(body, "request.generationConfig.thinkingConfig.includeThoughts", enabled)
+		for _, path := range []string{
+			"request.generationConfig.thinkingConfig.include_thoughts",
+			"request.generationConfig.thinking_config.include_thoughts",
+			"request.generationConfig.thinking_config.includeThoughts",
+		} {
+			body, _ = sjson.DeleteBytes(body, path)
+		}
+	case "interactions":
+		// Google Interactions only accepts auto or none. OpenAI's concise and
+		// detailed selectors therefore collapse to the supported enabled value.
+		value := "none"
+		if enabled {
+			value = "auto"
+		}
+		body, _ = sjson.SetBytes(body, "generation_config.thinking_summaries", value)
+		body, _ = sjson.DeleteBytes(body, "generation_config.thinkingSummaries")
+	case "openai-response", "codex":
+		if enabled {
+			body, _ = sjson.SetBytes(body, "reasoning.summary", normalizedSummaryDetail(config.Detail))
+			body, _ = sjson.DeleteBytes(body, "reasoning.generate_summary")
+			break
+		}
+		// Omitting the field is the documented way to disable summaries; an
+		// explicit null is not accepted by every Responses-compatible backend.
+		body, _ = sjson.DeleteBytes(body, "reasoning.summary")
+		body, _ = sjson.DeleteBytes(body, "reasoning.generate_summary")
+		if reasoning := gjson.GetBytes(body, "reasoning"); reasoning.IsObject() && len(reasoning.Map()) == 0 {
+			body, _ = sjson.DeleteBytes(body, "reasoning")
+		}
+	}
+	return body
+}
+
+// summaryFormatSupported reports whether a protocol carries summary visibility
+// intent that this package can read or write.
+func summaryFormatSupported(format string) bool {
+	switch format {
+	case "openai", "openai-response", "codex", "claude", "gemini", "antigravity", "interactions":
+		return true
+	default:
+		return false
+	}
+}
+
+// claudeThinkingAcceptsDisplay reports whether the body carries an active
+// thinking block that can hold a display field.
+func claudeThinkingAcceptsDisplay(body []byte) bool {
+	switch strings.ToLower(strings.TrimSpace(gjson.GetBytes(body, "thinking.type").String())) {
+	case "adaptive":
+		return true
+	case "enabled":
+		// This runs before ApplyThinking normalizes the request, so a missing
+		// budget_tokens is an unfinished body rather than inactive thinking.
+		// Only an explicit non-positive budget means thinking is off.
+		budget := gjson.GetBytes(body, "thinking.budget_tokens")
+		return budget.Type != gjson.Number || budget.Int() > 0
+	default:
+		return false
+	}
+}
+
+// applyOpenAIChatSummaryConfig writes summary visibility intent for the Chat
+// Completions protocol.
+//
+// Four dialects share this protocol and only OpenAI's is authoritative. OpenAI
+// documents no reasoning-visibility field at all (Chat Completions never returns
+// reasoning text) and rejects unknown body parameters, so reasoning_effort is the
+// only field that is always safe to write here. OpenRouter's documented
+// "reason but hide" bits (reasoning.exclude and its legacy include_reasoning
+// alias) are updated only when the body already carries them, which is exactly
+// when the upstream is known to understand them.
+func applyOpenAIChatSummaryConfig(body []byte, model string, enabled bool) []byte {
+	if gjson.GetBytes(body, "reasoning").IsObject() {
+		body, _ = sjson.SetBytes(body, "reasoning.exclude", !enabled)
+	}
+	if gjson.GetBytes(body, "include_reasoning").IsBool() {
+		body, _ = sjson.SetBytes(body, "include_reasoning", enabled)
+	}
+	if !enabled {
+		// Chat has no portable way to keep reasoning while hiding its summary.
+		// reasoning_effort:"none" would disable reasoning instead of hiding it,
+		// and Google documents that it is not even honored on Gemini 2.5 Pro or
+		// 3 models, so leave the effort the client asked for untouched.
+		return body
+	}
+	effort := gjson.GetBytes(body, "reasoning_effort")
+	if effort.Type != gjson.String || strings.TrimSpace(effort.String()) == "" || strings.EqualFold(strings.TrimSpace(effort.String()), "none") {
+		body, _ = sjson.SetBytes(body, "reasoning_effort", openAIChatSummaryEffort(body, model))
+	}
+	return body
+}
+
+// openAIChatSummaryEffort picks an active reasoning effort that the target model
+// documents. Chat exposes reasoning only while an effort is active, so a summary
+// request has to select one when the client left it unset.
+func openAIChatSummaryEffort(body []byte, model string) string {
+	baseModel := ParseSuffix(model).ModelName
+	if baseModel == "" {
+		baseModel = ParseSuffix(gjson.GetBytes(body, "model").String()).ModelName
+	}
+	modelInfo := registry.LookupModelInfo(baseModel, "openai")
+	if modelInfo == nil || modelInfo.Thinking == nil || len(modelInfo.Thinking.Levels) == 0 {
+		return "medium"
+	}
+
+	levels := make([]string, 0, len(modelInfo.Thinking.Levels))
+	for _, level := range modelInfo.Thinking.Levels {
+		normalized := strings.ToLower(strings.TrimSpace(level))
+		if normalized == "" || normalized == "none" {
+			continue
+		}
+		if normalized == "medium" {
+			return "medium"
+		}
+		levels = append(levels, normalized)
+	}
+	if len(levels) == 0 {
+		return "medium"
+	}
+	return levels[len(levels)/2]
+}
+
+func extractOpenAIExplicitSummaryConfig(body []byte) (SummaryConfig, bool) {
+	// Google's documented Chat Completions extension is the authoritative
+	// explicit visibility control when present, ahead of CPA compatibility
+	// aliases and Chat's reasoning_effort fallback.
+	for _, path := range []string{
+		"extra_body.google.thinking_config.include_thoughts",
+		"extra_body.google.thinking_config.includeThoughts",
+		"extra_body.google.thinkingConfig.include_thoughts",
+		"extra_body.google.thinkingConfig.includeThoughts",
+		"extra_body.extra_body.google.thinking_config.include_thoughts",
+		"extra_body.extra_body.google.thinking_config.includeThoughts",
+		"google.thinking_config.include_thoughts",
+		"google.thinking_config.includeThoughts",
+		"thinking.includeThoughts",
+		"thinking.include_thoughts",
+		"reasoning.includeThoughts",
+		"reasoning.include_thoughts",
+		"generationConfig.thinkingConfig.includeThoughts",
+		"generationConfig.thinkingConfig.include_thoughts",
+		"generation_config.thinking_config.include_thoughts",
+		"generation_config.thinking_config.includeThoughts",
+	} {
+		if config, ok := summaryBoolConfig(body, path); ok {
+			return config, true
+		}
+	}
+
+	for _, path := range []string{
+		"reasoning.summary",
+		"reasoning.generate_summary",
+	} {
+		if config, ok := responsesSummaryConfig(body, path); ok {
+			return config, true
+		}
+	}
+
+	// reasoning.exclude is OpenRouter's documented "reason but hide" bit, not an
+	// OpenAI wire field; include_reasoning is its documented legacy alias
+	// (include_reasoning: false is equivalent to reasoning: {exclude: true}).
+	// Only accept actual JSON booleans.
+	if exclude := gjson.GetBytes(body, "reasoning.exclude"); exclude.IsBool() {
+		if exclude.Bool() {
+			return SummaryConfig{Mode: SummaryDisabled}, true
+		}
+		return SummaryConfig{Mode: SummaryEnabled, Detail: "auto"}, true
+	}
+	if include := gjson.GetBytes(body, "include_reasoning"); include.IsBool() {
+		if include.Bool() {
+			return SummaryConfig{Mode: SummaryEnabled, Detail: "auto"}, true
+		}
+		return SummaryConfig{Mode: SummaryDisabled}, true
+	}
+	// OpenRouter's reasoning.enabled turns reasoning on "with no exclusions", so
+	// it also decides visibility when no dedicated bit was sent.
+	if enabled := gjson.GetBytes(body, "reasoning.enabled"); enabled.IsBool() {
+		if enabled.Bool() {
+			return SummaryConfig{Mode: SummaryEnabled, Detail: "auto"}, true
+		}
+		return SummaryConfig{Mode: SummaryDisabled}, true
+	}
+	return SummaryConfig{}, false
+}
+
+func firstSummaryBoolConfig(body []byte, paths []string) (SummaryConfig, bool) {
+	for _, path := range paths {
+		if config, ok := summaryBoolConfig(body, path); ok {
+			return config, true
+		}
+	}
+	return SummaryConfig{}, false
+}
+
+func summaryBoolConfig(body []byte, path string) (SummaryConfig, bool) {
+	switch value := gjson.GetBytes(body, path); value.Type {
+	case gjson.True:
+		return SummaryConfig{Mode: SummaryEnabled, Detail: "auto"}, true
+	case gjson.False:
+		return SummaryConfig{Mode: SummaryDisabled}, true
+	default:
+		return SummaryConfig{}, false
+	}
+}
+
+func responsesSummaryConfig(body []byte, path string) (SummaryConfig, bool) {
+	value := gjson.GetBytes(body, path)
+	if value.Raw == "" {
+		return SummaryConfig{}, false
+	}
+	if value.Type == gjson.Null {
+		return SummaryConfig{Mode: SummaryDisabled}, true
+	}
+	if value.Type != gjson.String {
+		return SummaryConfig{}, false
+	}
+
+	raw := strings.ToLower(strings.TrimSpace(value.String()))
+	switch raw {
+	case "auto", "concise", "detailed":
+		return SummaryConfig{Mode: SummaryEnabled, Detail: raw}, true
+	case "none":
+		// Compatibility with clients that expose a none enum; the OpenAI wire
+		// representation disables summaries by omitting the field.
+		return SummaryConfig{Mode: SummaryDisabled}, true
+	default:
+		return SummaryConfig{}, false
+	}
+}
+
+func claudeSummaryConfig(body []byte, path string) (SummaryConfig, bool) {
+	value := gjson.GetBytes(body, path)
+	if value.Type != gjson.String {
+		return SummaryConfig{}, false
+	}
+	switch strings.ToLower(strings.TrimSpace(value.String())) {
+	case "summarized":
+		return SummaryConfig{Mode: SummaryEnabled, Detail: "auto"}, true
+	case "omitted":
+		return SummaryConfig{Mode: SummaryDisabled}, true
+	default:
+		return SummaryConfig{}, false
+	}
+}
+
+func interactionsSummaryConfig(body []byte, path string) (SummaryConfig, bool) {
+	value := gjson.GetBytes(body, path)
+	if value.Type != gjson.String {
+		return SummaryConfig{}, false
+	}
+	switch strings.ToLower(strings.TrimSpace(value.String())) {
+	case "auto":
+		return SummaryConfig{Mode: SummaryEnabled, Detail: "auto"}, true
+	case "none":
+		return SummaryConfig{Mode: SummaryDisabled}, true
+	default:
+		return SummaryConfig{}, false
+	}
+}
+
+func enableClaudeThinkingForSummary(body []byte, model string, resolvedModelInfo *registry.ModelInfo) []byte {
+	modelInfo := resolvedModelInfo
+	if modelInfo == nil {
+		baseModel := ParseSuffix(model).ModelName
+		if baseModel == "" {
+			baseModel = ParseSuffix(gjson.GetBytes(body, "model").String()).ModelName
+		}
+		modelInfo = registry.LookupModelInfo(baseModel, "claude")
+	}
+	if modelInfo == nil || modelInfo.Thinking == nil {
+		return body
+	}
+
+	if len(modelInfo.Thinking.Levels) > 0 {
+		body, _ = sjson.SetBytes(body, "thinking.type", "adaptive")
+		body, _ = sjson.DeleteBytes(body, "thinking.budget_tokens")
+		return body
+	}
+
+	budget := modelInfo.Thinking.Min
+	if budget <= 0 {
+		return body
+	}
+	if maxTokens := gjson.GetBytes(body, "max_tokens"); maxTokens.Exists() && maxTokens.Int() <= int64(budget) {
+		return body
+	}
+	body, _ = sjson.SetBytes(body, "thinking.type", "enabled")
+	body, _ = sjson.SetBytes(body, "thinking.budget_tokens", budget)
+	return body
+}
+
+func normalizedSummaryDetail(detail string) string {
+	switch strings.ToLower(strings.TrimSpace(detail)) {
+	case "concise":
+		return "concise"
+	case "detailed":
+		return "detailed"
+	default:
+		return "auto"
+	}
+}
diff --git a/internal/thinking/summary_test.go b/internal/thinking/summary_test.go
new file mode 100644
index 00000000..6c9011f1
--- /dev/null
+++ b/internal/thinking/summary_test.go
@@ -0,0 +1,215 @@
+package thinking
+
+import (
+	"bytes"
+	"testing"
+
+	"github.com/tidwall/gjson"
+)
+
+func TestExtractSummaryConfig(t *testing.T) {
+	tests := []struct {
+		name       string
+		format     string
+		body       string
+		wantMode   SummaryMode
+		wantDetail string
+	}{
+		{name: "chat effort enables", format: "openai", body: `{"reasoning_effort":"high"}`, wantMode: SummaryEnabled, wantDetail: "auto"},
+		{name: "chat none disables", format: "openai", body: `{"reasoning_effort":"none"}`, wantMode: SummaryDisabled},
+		{name: "chat missing unspecified", format: "openai", body: `{}`, wantMode: SummaryUnspecified},
+		{name: "chat null effort unspecified", format: "openai", body: `{"reasoning_effort":null}`, wantMode: SummaryUnspecified},
+		{name: "chat non-string effort unspecified", format: "openai", body: `{"reasoning_effort":17}`, wantMode: SummaryUnspecified},
+		{name: "chat google extension false overrides effort", format: "openai", body: `{"reasoning_effort":"high","extra_body":{"google":{"thinking_config":{"include_thoughts":false}}}}`, wantMode: SummaryDisabled},
+		{name: "chat google extension true", format: "openai", body: `{"extra_body":{"google":{"thinking_config":{"include_thoughts":true}}}}`, wantMode: SummaryEnabled, wantDetail: "auto"},
+		{name: "chat exclude disables", format: "openai", body: `{"reasoning_effort":"high","reasoning":{"exclude":true}}`, wantMode: SummaryDisabled},
+		{name: "chat exclude false enables", format: "openai", body: `{"reasoning":{"effort":"high","exclude":false}}`, wantMode: SummaryEnabled, wantDetail: "auto"},
+		{name: "chat legacy include_reasoning false disables", format: "openai", body: `{"reasoning_effort":"high","include_reasoning":false}`, wantMode: SummaryDisabled},
+		{name: "chat legacy include_reasoning true enables", format: "openai", body: `{"include_reasoning":true}`, wantMode: SummaryEnabled, wantDetail: "auto"},
+		{name: "chat reasoning enabled false disables", format: "openai", body: `{"reasoning":{"enabled":false}}`, wantMode: SummaryDisabled},
+		{name: "chat reasoning enabled true enables", format: "openai", body: `{"reasoning":{"enabled":true}}`, wantMode: SummaryEnabled, wantDetail: "auto"},
+		{name: "chat exclude wins over include_reasoning", format: "openai", body: `{"reasoning":{"exclude":true},"include_reasoning":true}`, wantMode: SummaryDisabled},
+		{name: "chat non-boolean include_reasoning unspecified", format: "openai", body: `{"include_reasoning":"false"}`, wantMode: SummaryUnspecified},
+		{name: "responses effort alone unspecified", format: "openai-response", body: `{"reasoning":{"effort":"high"}}`, wantMode: SummaryUnspecified},
+		{name: "responses summary auto", format: "openai-response", body: `{"reasoning":{"effort":"high","summary":"auto"}}`, wantMode: SummaryEnabled, wantDetail: "auto"},
+		{name: "responses summary concise", format: "openai-response", body: `{"reasoning":{"summary":"concise"}}`, wantMode: SummaryEnabled, wantDetail: "concise"},
+		{name: "responses summary null", format: "openai-response", body: `{"reasoning":{"summary":null}}`, wantMode: SummaryDisabled},
+		{name: "responses boolean summary invalid", format: "openai-response", body: `{"reasoning":{"summary":true}}`, wantMode: SummaryUnspecified},
+		{name: "responses deprecated generate summary", format: "openai-response", body: `{"reasoning":{"generate_summary":"detailed"}}`, wantMode: SummaryEnabled, wantDetail: "detailed"},
+		{name: "claude summarized", format: "claude", body: `{"thinking":{"type":"adaptive","display":"summarized"}}`, wantMode: SummaryEnabled, wantDetail: "auto"},
+		{name: "claude omitted", format: "claude", body: `{"thinking":{"type":"enabled","budget_tokens":2048,"display":"omitted"}}`, wantMode: SummaryDisabled},
+		{name: "claude display without type is invalid", format: "claude", body: `{"thinking":{"display":"summarized"}}`, wantMode: SummaryUnspecified},
+		{name: "claude display with auto type is invalid", format: "claude", body: `{"thinking":{"type":"auto","display":"summarized"}}`, wantMode: SummaryUnspecified},
+		// ApplySummaryConfig runs before ApplyThinking fills budget_tokens, so an
+		// absent budget must not be read as inactive thinking.
+		{name: "claude enabled display without budget is valid", format: "claude", body: `{"thinking":{"type":"enabled","display":"summarized"}}`, wantMode: SummaryEnabled, wantDetail: "auto"},
+		{name: "claude enabled display with zero budget is invalid", format: "claude", body: `{"thinking":{"type":"enabled","budget_tokens":0,"display":"summarized"}}`, wantMode: SummaryUnspecified},
+		{name: "gemini include true", format: "gemini", body: `{"generationConfig":{"thinkingConfig":{"includeThoughts":true}}}`, wantMode: SummaryEnabled, wantDetail: "auto"},
+		{name: "gemini include false", format: "gemini", body: `{"generationConfig":{"thinkingConfig":{"includeThoughts":false}}}`, wantMode: SummaryDisabled},
+		{name: "antigravity include true", format: "antigravity", body: `{"request":{"generationConfig":{"thinkingConfig":{"includeThoughts":true}}}}`, wantMode: SummaryEnabled, wantDetail: "auto"},
+		{name: "interactions auto", format: "interactions", body: `{"generation_config":{"thinking_summaries":"auto"}}`, wantMode: SummaryEnabled, wantDetail: "auto"},
+		{name: "interactions none", format: "interactions", body: `{"generation_config":{"thinking_summaries":"none"}}`, wantMode: SummaryDisabled},
+		{name: "interactions detailed is invalid", format: "interactions", body: `{"generation_config":{"thinking_summaries":"detailed"}}`, wantMode: SummaryUnspecified},
+		{name: "interactions boolean is invalid", format: "interactions", body: `{"generation_config":{"thinking_summaries":true}}`, wantMode: SummaryUnspecified},
+		{name: "gemini string bool is invalid", format: "gemini", body: `{"generationConfig":{"thinkingConfig":{"includeThoughts":"true"}}}`, wantMode: SummaryUnspecified},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			got := ExtractSummaryConfig([]byte(test.body), test.format)
+			if got.Mode != test.wantMode || got.Detail != test.wantDetail {
+				t.Fatalf("ExtractSummaryConfig() = %+v, want mode=%v detail=%q", got, test.wantMode, test.wantDetail)
+			}
+		})
+	}
+}
+
+func TestApplySummaryConfig(t *testing.T) {
+	tests := []struct {
+		name   string
+		format string
+		body   string
+		config SummaryConfig
+		path   string
+		want   string
+	}{
+		{name: "chat enabled creates compatibility effort", format: "openai", config: SummaryConfig{Mode: SummaryEnabled}, path: "reasoning_effort", want: "medium"},
+		{name: "chat enabled preserves active effort", format: "openai", body: `{"reasoning_effort":"high"}`, config: SummaryConfig{Mode: SummaryEnabled}, path: "reasoning_effort", want: "high"},
+		// Chat cannot express "reason but hide", so disabling must not fall back to
+		// reasoning_effort:"none", which would disable reasoning altogether.
+		{name: "chat disabled preserves requested effort", format: "openai", body: `{"reasoning_effort":"high"}`, config: SummaryConfig{Mode: SummaryDisabled}, path: "reasoning_effort", want: "high"},
+		{name: "chat disabled sets openrouter exclude when present", format: "openai", body: `{"reasoning":{"effort":"high","exclude":false}}`, config: SummaryConfig{Mode: SummaryDisabled}, path: "reasoning.exclude", want: "true"},
+		{name: "chat enabled clears openrouter exclude when present", format: "openai", body: `{"reasoning":{"effort":"high","exclude":true}}`, config: SummaryConfig{Mode: SummaryEnabled}, path: "reasoning.exclude", want: "false"},
+		{name: "chat disabled updates legacy include_reasoning when present", format: "openai", body: `{"reasoning_effort":"high","include_reasoning":true}`, config: SummaryConfig{Mode: SummaryDisabled}, path: "include_reasoning", want: "false"},
+		{name: "chat disabled invents no openrouter field", format: "openai", body: `{"reasoning_effort":"high"}`, config: SummaryConfig{Mode: SummaryDisabled}, path: "reasoning", want: ""},
+		{name: "claude enabled", format: "claude", body: `{"thinking":{"type":"adaptive"}}`, config: SummaryConfig{Mode: SummaryEnabled}, path: "thinking.display", want: "summarized"},
+		{name: "claude disabled", format: "claude", body: `{"thinking":{"type":"enabled","budget_tokens":2048}}`, config: SummaryConfig{Mode: SummaryDisabled}, path: "thinking.display", want: "omitted"},
+		{name: "gemini enabled", format: "gemini", config: SummaryConfig{Mode: SummaryEnabled}, path: "generationConfig.thinkingConfig.includeThoughts", want: "true"},
+		{name: "gemini disabled", format: "gemini", config: SummaryConfig{Mode: SummaryDisabled}, path: "generationConfig.thinkingConfig.includeThoughts", want: "false"},
+		{name: "antigravity enabled", format: "antigravity", config: SummaryConfig{Mode: SummaryEnabled}, path: "request.generationConfig.thinkingConfig.includeThoughts", want: "true"},
+		{name: "interactions detail collapses to auto", format: "interactions", config: SummaryConfig{Mode: SummaryEnabled, Detail: "detailed"}, path: "generation_config.thinking_summaries", want: "auto"},
+		{name: "interactions disabled", format: "interactions", config: SummaryConfig{Mode: SummaryDisabled}, path: "generation_config.thinking_summaries", want: "none"},
+		{name: "responses concise", format: "openai-response", config: SummaryConfig{Mode: SummaryEnabled, Detail: "concise"}, path: "reasoning.summary", want: "concise"},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			body := test.body
+			if body == "" {
+				body = `{}`
+			}
+			out := ApplySummaryConfig([]byte(body), test.format, test.config)
+			if got := gjson.GetBytes(out, test.path).String(); got != test.want {
+				t.Fatalf("%s = %q, want %q; body=%s", test.path, got, test.want, out)
+			}
+		})
+	}
+}
+
+func TestApplySummaryConfigNormalizesTargetAliases(t *testing.T) {
+	tests := []struct {
+		format    string
+		body      string
+		canonical string
+		alias     string
+	}{
+		{format: "gemini", body: `{"generationConfig":{"thinkingConfig":{"include_thoughts":true}}}`, canonical: "generationConfig.thinkingConfig.includeThoughts", alias: "generationConfig.thinkingConfig.include_thoughts"},
+		{format: "antigravity", body: `{"request":{"generationConfig":{"thinkingConfig":{"include_thoughts":true}}}}`, canonical: "request.generationConfig.thinkingConfig.includeThoughts", alias: "request.generationConfig.thinkingConfig.include_thoughts"},
+		{format: "interactions", body: `{"generation_config":{"thinkingSummaries":"auto"}}`, canonical: "generation_config.thinking_summaries", alias: "generation_config.thinkingSummaries"},
+	}
+	for _, test := range tests {
+		out := ApplySummaryConfig([]byte(test.body), test.format, SummaryConfig{Mode: SummaryEnabled})
+		if !gjson.GetBytes(out, test.canonical).Exists() {
+			t.Fatalf("%s missing canonical field: %s", test.format, out)
+		}
+		if gjson.GetBytes(out, test.alias).Exists() {
+			t.Fatalf("%s retained alias %s: %s", test.format, test.alias, out)
+		}
+	}
+}
+
+// Anthropic requires thinking.type, and rejects display on a disabled block, so
+// display must never be written unless thinking is already active.
+func TestApplySummaryConfig_ClaudeDisplayRequiresActiveThinking(t *testing.T) {
+	bodies := []string{
+		`{}`,
+		`{"messages":[{"role":"user","content":"hi"}]}`,
+		`{"thinking":{"type":"disabled"}}`,
+	}
+	for _, mode := range []SummaryMode{SummaryEnabled, SummaryDisabled} {
+		for _, body := range bodies {
+			out := ApplySummaryConfig([]byte(body), "claude", SummaryConfig{Mode: mode})
+			if gjson.GetBytes(out, "thinking.display").Exists() {
+				t.Fatalf("mode %v wrote display without active thinking: %s", mode, out)
+			}
+			if !bytes.Equal(out, []byte(body)) {
+				t.Fatalf("mode %v changed body: got %s, want %s", mode, out, body)
+			}
+		}
+	}
+}
+
+func TestApplySummaryConfigForModel_ClaudeExplicitVisibilityUsesValidThinkingMode(t *testing.T) {
+	tests := []struct {
+		name        string
+		model       string
+		body        string
+		mode        SummaryMode
+		wantType    string
+		wantDisplay string
+		wantBudget  int64
+	}{
+		{name: "adaptive model summarized", model: "claude-opus-5", body: `{"model":"claude-opus-5","max_tokens":32000}`, mode: SummaryEnabled, wantType: "adaptive", wantDisplay: "summarized"},
+		{name: "adaptive model omitted", model: "claude-opus-5", body: `{"model":"claude-opus-5","max_tokens":32000}`, mode: SummaryDisabled, wantType: "adaptive", wantDisplay: "omitted"},
+		{name: "manual model summarized", model: "claude-haiku-4-5-20251001", body: `{"model":"claude-haiku-4-5-20251001","max_tokens":32000}`, mode: SummaryEnabled, wantType: "enabled", wantDisplay: "summarized", wantBudget: 1024},
+		{name: "manual model omitted", model: "claude-haiku-4-5-20251001", body: `{"model":"claude-haiku-4-5-20251001","max_tokens":32000}`, mode: SummaryDisabled, wantType: "enabled", wantDisplay: "omitted", wantBudget: 1024},
+	}
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			out := ApplySummaryConfigForModel([]byte(test.body), "claude", test.model, SummaryConfig{Mode: test.mode})
+			if got := gjson.GetBytes(out, "thinking.type").String(); got != test.wantType {
+				t.Fatalf("thinking.type = %q, want %q; body=%s", got, test.wantType, out)
+			}
+			if got := gjson.GetBytes(out, "thinking.display").String(); got != test.wantDisplay {
+				t.Fatalf("thinking.display = %q, want %q; body=%s", got, test.wantDisplay, out)
+			}
+			if test.wantBudget > 0 && gjson.GetBytes(out, "thinking.budget_tokens").Int() != test.wantBudget {
+				t.Fatalf("thinking.budget_tokens = %d, want %d; body=%s", gjson.GetBytes(out, "thinking.budget_tokens").Int(), test.wantBudget, out)
+			}
+		})
+	}
+}
+
+func TestApplySummaryConfig_ResponsesNormalizesDeprecatedGenerateSummary(t *testing.T) {
+	out := ApplySummaryConfig([]byte(`{"reasoning":{"generate_summary":"detailed"}}`), "openai-response", SummaryConfig{Mode: SummaryEnabled, Detail: "detailed"})
+	if got := gjson.GetBytes(out, "reasoning.summary").String(); got != "detailed" {
+		t.Fatalf("reasoning.summary = %q, want detailed; body=%s", got, out)
+	}
+	if gjson.GetBytes(out, "reasoning.generate_summary").Exists() {
+		t.Fatalf("deprecated reasoning.generate_summary remained: %s", out)
+	}
+}
+
+func TestApplySummaryConfig_ResponsesDisabledOmitsSummary(t *testing.T) {
+	out := ApplySummaryConfig([]byte(`{"reasoning":{"effort":"high","summary":"auto"}}`), "openai-response", SummaryConfig{Mode: SummaryDisabled})
+	if result := gjson.GetBytes(out, "reasoning.summary"); result.Exists() {
+		t.Fatalf("reasoning.summary = %s, want absent; body=%s", result.Raw, out)
+	}
+	if got := gjson.GetBytes(out, "reasoning.effort").String(); got != "high" {
+		t.Fatalf("reasoning.effort = %q, want high; body=%s", got, out)
+	}
+}
+
+func TestApplySummaryConfig_ResponsesDisabledDropsEmptyReasoning(t *testing.T) {
+	out := ApplySummaryConfig([]byte(`{"model":"gpt-5.4","reasoning":{"summary":"auto"}}`), "openai-response", SummaryConfig{Mode: SummaryDisabled})
+	if gjson.GetBytes(out, "reasoning").Exists() {
+		t.Fatalf("empty reasoning object left behind: %s", out)
+	}
+}
+
+func TestApplySummaryConfig_UnspecifiedLeavesBodyUnchanged(t *testing.T) {
+	body := []byte(`{"thinking":{"type":"adaptive"}}`)
+	if got := ApplySummaryConfig(body, "claude", SummaryConfig{}); !bytes.Equal(got, body) {
+		t.Fatalf("unspecified summary changed body: got %s, want %s", got, body)
+	}
+}
diff --git a/internal/translator/antigravity/claude/antigravity_claude_request.go b/internal/translator/antigravity/claude/antigravity_claude_request.go
index 58a1068e..b575679a 100644
--- a/internal/translator/antigravity/claude/antigravity_claude_request.go
+++ b/internal/translator/antigravity/claude/antigravity_claude_request.go
@@ -885,7 +885,6 @@ func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _
 			if b := t.Get("budget_tokens"); b.Exists() && b.Type == gjson.Number {
 				budget := int(b.Int())
 				out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.thinkingBudget", budget)
-				out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.includeThoughts", true)
 			}
 		case "adaptive", "auto":
 			// For adaptive thinking:
@@ -901,7 +900,6 @@ func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _
 			} else {
 				out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.thinkingLevel", "high")
 			}
-			out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.includeThoughts", true)
 		}
 	}
 	if v := gjson.GetBytes(rawJSON, "temperature"); v.Exists() && v.Type == gjson.Number {
diff --git a/internal/translator/antigravity/claude/antigravity_claude_request_test.go b/internal/translator/antigravity/claude/antigravity_claude_request_test.go
index 52586f29..c41aa752 100644
--- a/internal/translator/antigravity/claude/antigravity_claude_request_test.go
+++ b/internal/translator/antigravity/claude/antigravity_claude_request_test.go
@@ -2238,8 +2238,8 @@ func TestConvertClaudeRequestToAntigravity_ThinkingConfig(t *testing.T) {
 		if thinkingConfig.Get("thinkingBudget").Int() != 8000 {
 			t.Errorf("Expected thinkingBudget 8000, got %d", thinkingConfig.Get("thinkingBudget").Int())
 		}
-		if !thinkingConfig.Get("includeThoughts").Bool() {
-			t.Error("includeThoughts should be true")
+		if thinkingConfig.Get("includeThoughts").Exists() {
+			t.Error("includeThoughts should be absent without explicit Claude display intent")
 		}
 	} else {
 		t.Log("thinkingConfig not present - model may not be registered in test registry")
diff --git a/internal/translator/antigravity/interactions/interactions_antigravity_request.go b/internal/translator/antigravity/interactions/interactions_antigravity_request.go
index 123b42dc..2d00a4f3 100644
--- a/internal/translator/antigravity/interactions/interactions_antigravity_request.go
+++ b/internal/translator/antigravity/interactions/interactions_antigravity_request.go
@@ -707,20 +707,17 @@ func antigravityInputAudioMimeType(format string) string {
 }
 
 func antigravityThinkingSummariesIncludeThoughts(summary gjson.Result) (bool, bool) {
-	switch summary.Type {
-	case gjson.True:
+	if summary.Type != gjson.String {
+		return false, false
+	}
+	switch strings.ToLower(strings.TrimSpace(summary.String())) {
+	case "auto":
 		return true, true
-	case gjson.False:
+	case "none":
 		return false, true
-	case gjson.String:
-		switch strings.ToLower(strings.TrimSpace(summary.String())) {
-		case "", "none", "off", "false", "disabled":
-			return false, true
-		default:
-			return true, true
-		}
+	default:
+		return false, false
 	}
-	return false, false
 }
 
 func convertSnakeCaseKeysToCamelCaseForAntigravity(raw []byte) []byte {
diff --git a/internal/translator/antigravity/openai/chat-completions/antigravity_openai_request.go b/internal/translator/antigravity/openai/chat-completions/antigravity_openai_request.go
index 6c99515f..ee936642 100644
--- a/internal/translator/antigravity/openai/chat-completions/antigravity_openai_request.go
+++ b/internal/translator/antigravity/openai/chat-completions/antigravity_openai_request.go
@@ -5,6 +5,7 @@ package chat_completions
 import (
 	"strings"
 
+	"github.com/router-for-me/CLIProxyAPI/v7/internal/thinking"
 	"github.com/router-for-me/CLIProxyAPI/v7/internal/translator/antigravity/gemini"
 	translatorcommon "github.com/router-for-me/CLIProxyAPI/v7/internal/translator/common"
 	"github.com/router-for-me/CLIProxyAPI/v7/internal/translator/gemini/common"
@@ -51,14 +52,12 @@ func ConvertOpenAIRequestToAntigravity(modelName string, inputRawJSON []byte, _
 			thinkingPath := "request.generationConfig.thinkingConfig"
 			if effort == "auto" {
 				out, _ = sjson.SetBytes(out, thinkingPath+".thinkingBudget", -1)
-				out, _ = sjson.SetBytes(out, thinkingPath+".includeThoughts", true)
 			} else {
 				out, _ = sjson.SetBytes(out, thinkingPath+".thinkingLevel", effort)
-				out, _ = sjson.SetBytes(out, thinkingPath+".includeThoughts", effort != "none")
 			}
 		}
 	}
-	out = applyOpenAIThinkingCompatibilityToAntigravity(out, rawJSON, modelName)
+	out = applyOpenAIThinkingCompatibilityToAntigravity(out, rawJSON)
 
 	// Temperature/top_p/top_k/max_tokens
 	if tr := gjson.GetBytes(rawJSON, "temperature"); tr.Exists() && tr.Type == gjson.Number {
@@ -513,29 +512,10 @@ func applyOpenAIToolChoiceToAntigravity(out, rawJSON []byte, functionNameMap map
 	return out
 }
 
-func applyOpenAIThinkingCompatibilityToAntigravity(out []byte, rawJSON []byte, modelName string) []byte {
+func applyOpenAIThinkingCompatibilityToAntigravity(out []byte, rawJSON []byte) []byte {
 	out = normalizeAntigravityOpenAIThinkingConfig(out)
-
-	for _, path := range []string{
-		"thinking.includeThoughts",
-		"thinking.include_thoughts",
-		"reasoning.includeThoughts",
-		"reasoning.include_thoughts",
-	} {
-		if value := gjson.GetBytes(rawJSON, path); value.Exists() {
-			out = setAntigravityOpenAIBoolIfDifferent(out, "request.generationConfig.thinkingConfig.includeThoughts", value.Bool())
-		}
-	}
-
-	if exclude := gjson.GetBytes(rawJSON, "reasoning.exclude"); exclude.Exists() {
-		out = setAntigravityOpenAIBoolIfDifferent(out, "request.generationConfig.thinkingConfig.includeThoughts", !exclude.Bool())
-	}
-
-	if !gjson.GetBytes(out, "request.generationConfig.thinkingConfig.includeThoughts").Exists() && antigravityOpenAIDefaultIncludeThoughts(modelName) {
-		out = setAntigravityOpenAIBoolIfDifferent(out, "request.generationConfig.thinkingConfig.includeThoughts", true)
-	}
-
-	return normalizeAntigravityOpenAIThinkingConfig(out)
+	config := thinking.ExtractSummaryConfig(rawJSON, "openai")
+	return thinking.ApplySummaryConfig(out, "antigravity", config)
 }
 
 func normalizeAntigravityOpenAIThinkingConfig(out []byte) []byte {
@@ -543,11 +523,19 @@ func normalizeAntigravityOpenAIThinkingConfig(out []byte) []byte {
 		"request.generationConfig.thinking_config",
 		"request.generationConfig.thinkingConfig",
 	} {
-		if includeThoughts := gjson.GetBytes(out, prefix+".includeThoughts"); includeThoughts.Exists() {
-			out = setAntigravityOpenAIBoolIfDifferent(out, "request.generationConfig.thinkingConfig.includeThoughts", includeThoughts.Bool())
+		if sourcePath := prefix + ".includeThoughts"; gjson.GetBytes(out, sourcePath).Exists() {
+			includeThoughts := gjson.GetBytes(out, sourcePath)
+			out = setAntigravityOpenAIBoolResultIfValid(out, "request.generationConfig.thinkingConfig.includeThoughts", includeThoughts)
+			if includeThoughts.Type != gjson.True && includeThoughts.Type != gjson.False {
+				out, _ = sjson.DeleteBytes(out, sourcePath)
+			}
 		}
-		if includeThoughts := gjson.GetBytes(out, prefix+".include_thoughts"); includeThoughts.Exists() {
-			out = setAntigravityOpenAIBoolIfDifferent(out, "request.generationConfig.thinkingConfig.includeThoughts", includeThoughts.Bool())
+		if sourcePath := prefix + ".include_thoughts"; gjson.GetBytes(out, sourcePath).Exists() {
+			includeThoughts := gjson.GetBytes(out, sourcePath)
+			out = setAntigravityOpenAIBoolResultIfValid(out, "request.generationConfig.thinkingConfig.includeThoughts", includeThoughts)
+			if includeThoughts.Type != gjson.True && includeThoughts.Type != gjson.False {
+				out, _ = sjson.DeleteBytes(out, sourcePath)
+			}
 		}
 		if thinkingLevel := gjson.GetBytes(out, prefix+".thinkingLevel"); thinkingLevel.Exists() {
 			out = setAntigravityOpenAIRawIfDifferent(out, "request.generationConfig.thinkingConfig.thinkingLevel", thinkingLevel)
@@ -568,7 +556,7 @@ func normalizeAntigravityOpenAIThinkingConfig(out []byte) []byte {
 		"request.generationConfig.include_thoughts",
 	} {
 		if includeThoughts := gjson.GetBytes(out, path); includeThoughts.Exists() {
-			out = setAntigravityOpenAIBoolIfDifferent(out, "request.generationConfig.thinkingConfig.includeThoughts", includeThoughts.Bool())
+			out = setAntigravityOpenAIBoolResultIfValid(out, "request.generationConfig.thinkingConfig.includeThoughts", includeThoughts)
 		}
 	}
 
@@ -588,6 +576,17 @@ func normalizeAntigravityOpenAIThinkingConfig(out []byte) []byte {
 	return out
 }
 
+func setAntigravityOpenAIBoolResultIfValid(out []byte, path string, value gjson.Result) []byte {
+	switch value.Type {
+	case gjson.True:
+		return setAntigravityOpenAIBoolIfDifferent(out, path, true)
+	case gjson.False:
+		return setAntigravityOpenAIBoolIfDifferent(out, path, false)
+	default:
+		return out
+	}
+}
+
 func setAntigravityOpenAIBoolIfDifferent(out []byte, path string, value bool) []byte {
 	current := gjson.GetBytes(out, path)
 	if value && current.Type == gjson.True || !value && current.Type == gjson.False {
@@ -611,8 +610,3 @@ func setAntigravityOpenAIRawIfDifferent(out []byte, path string, value gjson.Res
 	}
 	return updated
 }
-
-func antigravityOpenAIDefaultIncludeThoughts(modelName string) bool {
-	modelName = strings.ToLower(modelName)
-	return strings.Contains(modelName, "gemini-3")
-}
diff --git a/internal/translator/antigravity/openai/chat-completions/antigravity_openai_request_test.go b/internal/translator/antigravity/openai/chat-completions/antigravity_openai_request_test.go
index 81907cbb..33bd0e4c 100644
--- a/internal/translator/antigravity/openai/chat-completions/antigravity_openai_request_test.go
+++ b/internal/translator/antigravity/openai/chat-completions/antigravity_openai_request_test.go
@@ -191,17 +191,27 @@ func TestConvertOpenAIRequestToAntigravitySkipsEmptyAssistantMessages(t *testing
 
 func TestConvertOpenAIRequestToAntigravityThinkingAliases(t *testing.T) {
 	tests := []struct {
-		name string
-		body string
-		want bool
+		name       string
+		body       string
+		wantExists bool
+		want       bool
 	}{
 		{
-			name: "Default Gemini include thoughts",
+			name: "Missing summary intent leaves include thoughts absent",
 			body: `{
 				"model":"gemini-3.1-pro-low",
 				"messages":[{"role":"user","content":"hi"}]
 			}`,
-			want: true,
+		},
+		{
+			name: "Reasoning effort enables thoughts",
+			body: `{
+				"model":"gemini-3.1-pro-low",
+				"messages":[{"role":"user","content":"hi"}],
+				"reasoning_effort":"high"
+			}`,
+			wantExists: true,
+			want:       true,
 		},
 		{
 			name: "GenerationConfig snake include thoughts",
@@ -210,7 +220,16 @@ func TestConvertOpenAIRequestToAntigravityThinkingAliases(t *testing.T) {
 				"messages":[{"role":"user","content":"hi"}],
 				"generationConfig":{"thinkingConfig":{"include_thoughts":true}}
 			}`,
-			want: true,
+			wantExists: true,
+			want:       true,
+		},
+		{
+			name: "String include thoughts is ignored",
+			body: `{
+				"model":"gemini-3.1-pro-low",
+				"messages":[{"role":"user","content":"hi"}],
+				"generationConfig":{"thinkingConfig":{"includeThoughts":"true"}}
+			}`,
 		},
 		{
 			name: "Top-level thinking include thoughts",
@@ -219,7 +238,8 @@ func TestConvertOpenAIRequestToAntigravityThinkingAliases(t *testing.T) {
 				"messages":[{"role":"user","content":"hi"}],
 				"thinking":{"include_thoughts":true}
 			}`,
-			want: true,
+			wantExists: true,
+			want:       true,
 		},
 		{
 			name: "Reasoning exclude false includes thoughts",
@@ -228,7 +248,8 @@ func TestConvertOpenAIRequestToAntigravityThinkingAliases(t *testing.T) {
 				"messages":[{"role":"user","content":"hi"}],
 				"reasoning":{"exclude":false}
 			}`,
-			want: true,
+			wantExists: true,
+			want:       true,
 		},
 		{
 			name: "Reasoning exclude true hides thoughts",
@@ -237,7 +258,19 @@ func TestConvertOpenAIRequestToAntigravityThinkingAliases(t *testing.T) {
 				"messages":[{"role":"user","content":"hi"}],
 				"reasoning":{"exclude":true}
 			}`,
-			want: false,
+			wantExists: true,
+			want:       false,
+		},
+		{
+			name: "Google extension disables thoughts",
+			body: `{
+				"model":"gemini-3.1-pro-low",
+				"messages":[{"role":"user","content":"hi"}],
+				"reasoning_effort":"high",
+				"extra_body":{"google":{"thinking_config":{"include_thoughts":false}}}
+			}`,
+			wantExists: true,
+			want:       false,
 		},
 	}
 
@@ -245,11 +278,13 @@ func TestConvertOpenAIRequestToAntigravityThinkingAliases(t *testing.T) {
 		t.Run(tt.name, func(t *testing.T) {
 			result := ConvertOpenAIRequestToAntigravity("gemini-3.1-pro-low", []byte(tt.body), false)
 			includeThoughts := gjson.GetBytes(result, "request.generationConfig.thinkingConfig.includeThoughts")
-			if !includeThoughts.Exists() {
-				t.Fatalf("includeThoughts missing. Output: %s", result)
+			if includeThoughts.Exists() != tt.wantExists {
+				t.Fatalf("includeThoughts exists = %v, want %v. Output: %s", includeThoughts.Exists(), tt.wantExists, result)
 			}
-			if got := includeThoughts.Bool(); got != tt.want {
-				t.Fatalf("includeThoughts = %v, want %v. Output: %s", got, tt.want, result)
+			if tt.wantExists {
+				if got := includeThoughts.Bool(); got != tt.want {
+					t.Fatalf("includeThoughts = %v, want %v. Output: %s", got, tt.want, result)
+				}
 			}
 			if snake := gjson.GetBytes(result, "request.generationConfig.thinkingConfig.include_thoughts"); snake.Exists() {
 				t.Fatalf("include_thoughts should be normalized away. Output: %s", result)
diff --git a/internal/translator/codex/claude/codex_claude_request.go b/internal/translator/codex/claude/codex_claude_request.go
index 3a27e850..20241bf2 100644
--- a/internal/translator/codex/claude/codex_claude_request.go
+++ b/internal/translator/codex/claude/codex_claude_request.go
@@ -341,7 +341,9 @@ func ConvertClaudeRequestToCodex(modelName string, inputRawJSON []byte, _ bool)
 		}
 	}
 	template, _ = sjson.SetBytes(template, "reasoning.effort", reasoningEffort)
-	template, _ = sjson.SetBytes(template, "reasoning.summary", "auto")
+	// OpenAI documents reasoning summaries as explicit opt-in output. Leave
+	// reasoning.summary to the source request's canonical summary intent instead
+	// of coupling it to reasoning effort.
 	serviceTier := normalizeCodexServiceTier(rootResult.Get("service_tier"))
 	if speed := rootResult.Get("speed"); speed.Type == gjson.String && speed.String() == "fast" {
 		serviceTier = "priority"
diff --git a/internal/translator/codex/gemini/codex_gemini_request.go b/internal/translator/codex/gemini/codex_gemini_request.go
index e61dc975..f5a03bdf 100644
--- a/internal/translator/codex/gemini/codex_gemini_request.go
+++ b/internal/translator/codex/gemini/codex_gemini_request.go
@@ -322,7 +322,9 @@ func ConvertGeminiRequestToCodex(modelName string, inputRawJSON []byte, _ bool)
 		// No thinking config, set default effort
 		out, _ = sjson.SetBytes(out, "reasoning.effort", "medium")
 	}
-	out, _ = sjson.SetBytes(out, "reasoning.summary", "auto")
+	// OpenAI documents reasoning summaries as explicit opt-in output. Leave
+	// reasoning.summary to the source request's canonical summary intent instead
+	// of coupling it to reasoning effort.
 	out, _ = sjson.SetBytes(out, "stream", true)
 	out, _ = sjson.SetBytes(out, "store", false)
 	out, _ = sjson.SetBytes(out, "include", []string{"reasoning.encrypted_content"})
diff --git a/internal/translator/codex/interactions/interactions_codex_request.go b/internal/translator/codex/interactions/interactions_codex_request.go
index 89b083f2..25287e89 100644
--- a/internal/translator/codex/interactions/interactions_codex_request.go
+++ b/internal/translator/codex/interactions/interactions_codex_request.go
@@ -155,17 +155,11 @@ func interactionsCodexReasoningSummary(cfg gjson.Result) string {
 		"thinkingSummaries",
 		"reasoning.summary",
 	} {
-		if value := cfg.Get(path); value.Exists() {
-			switch value.Type {
-			case gjson.True:
-				return "auto"
-			case gjson.False:
-				return "none"
-			case gjson.String:
-				summary := strings.ToLower(strings.TrimSpace(value.String()))
-				if summary != "" {
-					return summary
-				}
+		if value := cfg.Get(path); value.Type == gjson.String {
+			summary := strings.ToLower(strings.TrimSpace(value.String()))
+			switch summary {
+			case "auto", "none":
+				return summary
 			}
 		}
 	}
@@ -177,10 +171,10 @@ func interactionsCodexReasoningSummary(cfg gjson.Result) string {
 		"thinkingConfig.include_thoughts",
 		"thinkingConfig.includeThoughts",
 	} {
-		if value := cfg.Get(path); value.Exists() {
-			if value.Bool() {
-				return "auto"
-			}
+		switch value := cfg.Get(path); value.Type {
+		case gjson.True:
+			return "auto"
+		case gjson.False:
 			return "none"
 		}
 	}
diff --git a/internal/translator/codex/openai/chat-completions/codex_openai_request.go b/internal/translator/codex/openai/chat-completions/codex_openai_request.go
index 051d26ef..307df55d 100644
--- a/internal/translator/codex/openai/chat-completions/codex_openai_request.go
+++ b/internal/translator/codex/openai/chat-completions/codex_openai_request.go
@@ -64,7 +64,9 @@ func ConvertOpenAIRequestToCodex(modelName string, inputRawJSON []byte, stream b
 		out, _ = sjson.SetBytes(out, "reasoning.effort", "medium")
 	}
 	out, _ = sjson.SetBytes(out, "parallel_tool_calls", true)
-	out, _ = sjson.SetBytes(out, "reasoning.summary", "auto")
+	// OpenAI documents reasoning summaries as explicit opt-in output. Leave
+	// reasoning.summary to the source request's canonical summary intent instead
+	// of coupling it to reasoning effort.
 	out, _ = sjson.SetBytes(out, "include", []string{"reasoning.encrypted_content"})
 
 	// Model
diff --git a/internal/translator/gemini/claude/gemini_claude_request.go b/internal/translator/gemini/claude/gemini_claude_request.go
index 2df2009e..8a2259b7 100644
--- a/internal/translator/gemini/claude/gemini_claude_request.go
+++ b/internal/translator/gemini/claude/gemini_claude_request.go
@@ -266,7 +266,6 @@ func ConvertClaudeRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 			if b := t.Get("budget_tokens"); b.Exists() && b.Type == gjson.Number {
 				budget := int(b.Int())
 				out, _ = sjson.SetBytes(out, "generationConfig.thinkingConfig.thinkingBudget", budget)
-				out, _ = sjson.SetBytes(out, "generationConfig.thinkingConfig.includeThoughts", true)
 			}
 		case "adaptive", "auto":
 			// For adaptive thinking:
@@ -290,7 +289,6 @@ func ConvertClaudeRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 					out, _ = sjson.SetBytes(out, "generationConfig.thinkingConfig.thinkingLevel", "high")
 				}
 			}
-			out, _ = sjson.SetBytes(out, "generationConfig.thinkingConfig.includeThoughts", true)
 		}
 	}
 	if v := gjson.GetBytes(rawJSON, "temperature"); v.Exists() && v.Type == gjson.Number {
diff --git a/internal/translator/gemini/interactions/interactions_gemini_common.go b/internal/translator/gemini/interactions/interactions_gemini_common.go
index e27e815e..59db6641 100644
--- a/internal/translator/gemini/interactions/interactions_gemini_common.go
+++ b/internal/translator/gemini/interactions/interactions_gemini_common.go
@@ -449,20 +449,17 @@ func normalizeInteractionsGenerationConfig(out []byte) []byte {
 }
 
 func interactionsThinkingSummariesIncludeThoughts(summary gjson.Result) (bool, bool) {
-	switch summary.Type {
-	case gjson.True:
+	if summary.Type != gjson.String {
+		return false, false
+	}
+	switch strings.ToLower(strings.TrimSpace(summary.String())) {
+	case "auto":
 		return true, true
-	case gjson.False:
+	case "none":
 		return false, true
-	case gjson.String:
-		switch strings.ToLower(strings.TrimSpace(summary.String())) {
-		case "", "none", "off", "false", "disabled":
-			return false, true
-		default:
-			return true, true
-		}
+	default:
+		return false, false
 	}
-	return false, false
 }
 
 func copyInteractionsResponseModalities(out []byte, root gjson.Result) []byte {
diff --git a/internal/translator/gemini/openai/chat-completions/gemini_openai_request.go b/internal/translator/gemini/openai/chat-completions/gemini_openai_request.go
index 0eea0925..64731dc4 100644
--- a/internal/translator/gemini/openai/chat-completions/gemini_openai_request.go
+++ b/internal/translator/gemini/openai/chat-completions/gemini_openai_request.go
@@ -48,10 +48,8 @@ func ConvertOpenAIRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 			thinkingPath := "generationConfig.thinkingConfig"
 			if effort == "auto" {
 				out, _ = sjson.SetBytes(out, thinkingPath+".thinkingBudget", -1)
-				out, _ = sjson.SetBytes(out, thinkingPath+".includeThoughts", true)
 			} else {
 				out, _ = sjson.SetBytes(out, thinkingPath+".thinkingLevel", effort)
-				out, _ = sjson.SetBytes(out, thinkingPath+".includeThoughts", effort != "none")
 			}
 		}
 	}
diff --git a/internal/translator/gemini/openai/responses/gemini_openai-responses_request.go b/internal/translator/gemini/openai/responses/gemini_openai-responses_request.go
index 8ee3186a..6ebb4336 100644
--- a/internal/translator/gemini/openai/responses/gemini_openai-responses_request.go
+++ b/internal/translator/gemini/openai/responses/gemini_openai-responses_request.go
@@ -379,10 +379,8 @@ func ConvertOpenAIResponsesRequestToGemini(modelName string, inputRawJSON []byte
 			thinkingPath := "generationConfig.thinkingConfig"
 			if effort == "auto" {
 				out, _ = sjson.SetBytes(out, thinkingPath+".thinkingBudget", -1)
-				out, _ = sjson.SetBytes(out, thinkingPath+".includeThoughts", true)
 			} else {
 				out, _ = sjson.SetBytes(out, thinkingPath+".thinkingLevel", effort)
-				out, _ = sjson.SetBytes(out, thinkingPath+".includeThoughts", effort != "none")
 			}
 		}
 	}
diff --git a/sdk/translator/registry.go b/sdk/translator/registry.go
index ad4d351d..e9ef609f 100644
--- a/sdk/translator/registry.go
+++ b/sdk/translator/registry.go
@@ -4,6 +4,7 @@ import (
 	"context"
 	"sync"
 
+	"github.com/router-for-me/CLIProxyAPI/v7/internal/thinking"
 	log "github.com/sirupsen/logrus"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
@@ -56,6 +57,8 @@ func (r *Registry) SetPluginHooks(hooks PluginHooks) {
 // "model" field is still updated to match the resolved model name so that
 // client-side prefixes (e.g. "copilot/gpt-5-mini") are not leaked upstream.
 func (r *Registry) TranslateRequest(from, to Format, model string, rawJSON []byte, stream bool) []byte {
+	summaryConfig := thinking.ExtractSummaryConfig(rawJSON, from.String())
+
 	r.mu.RLock()
 	var fn RequestTransform
 	if byTarget, ok := r.requests[from]; ok {
@@ -85,7 +88,7 @@ func (r *Registry) TranslateRequest(from, to Format, model string, rawJSON []byt
 			}
 		}
 	}
-	return body
+	return thinking.ApplySummaryConfigForModel(body, to.String(), model, summaryConfig)
 }
 
 // HasRequestTransformer indicates whether a request translator exists.
diff --git a/sdk/translator/registry_summary_test.go b/sdk/translator/registry_summary_test.go
new file mode 100644
index 00000000..79312dd8
--- /dev/null
+++ b/sdk/translator/registry_summary_test.go
@@ -0,0 +1,127 @@
+package translator
+
+import (
+	"testing"
+
+	"github.com/tidwall/gjson"
+)
+
+func TestRegistryTranslateRequestAppliesSummaryIntent(t *testing.T) {
+	tests := []struct {
+		name       string
+		from       Format
+		to         Format
+		input      string
+		translated string
+		path       string
+		want       string
+		wantExists bool
+	}{
+		{
+			name:       "chat effort enables Claude summary",
+			from:       FormatOpenAI,
+			to:         FormatClaude,
+			input:      `{"reasoning_effort":"high"}`,
+			translated: `{"thinking":{"type":"adaptive"}}`,
+			path:       "thinking.display",
+			want:       "summarized",
+			wantExists: true,
+		},
+		{
+			name:       "responses effort alone leaves Claude display absent",
+			from:       FormatOpenAIResponse,
+			to:         FormatClaude,
+			input:      `{"reasoning":{"effort":"high"}}`,
+			translated: `{"thinking":{"type":"adaptive"}}`,
+			path:       "thinking.display",
+		},
+		{
+			name:       "responses summary enables Claude summary",
+			from:       FormatOpenAIResponse,
+			to:         FormatClaude,
+			input:      `{"reasoning":{"effort":"high","summary":"auto"}}`,
+			translated: `{"thinking":{"type":"adaptive"}}`,
+			path:       "thinking.display",
+			want:       "summarized",
+			wantExists: true,
+		},
+		{
+			name:       "responses null summary disables Gemini summaries",
+			from:       FormatOpenAIResponse,
+			to:         FormatGemini,
+			input:      `{"reasoning":{"effort":"high","summary":null}}`,
+			translated: `{"generationConfig":{"thinkingConfig":{"thinkingLevel":"high"}}}`,
+			path:       "generationConfig.thinkingConfig.includeThoughts",
+			want:       "false",
+			wantExists: true,
+		},
+		{
+			name:       "Google Chat extension overrides effort",
+			from:       FormatOpenAI,
+			to:         FormatGemini,
+			input:      `{"reasoning_effort":"high","extra_body":{"google":{"thinking_config":{"include_thoughts":false}}}}`,
+			translated: `{"generationConfig":{"thinkingConfig":{"thinkingLevel":"high","includeThoughts":true}}}`,
+			path:       "generationConfig.thinkingConfig.includeThoughts",
+			want:       "false",
+			wantExists: true,
+		},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			registry := NewRegistry()
+			registry.Register(test.from, test.to, func(_ string, _ []byte, _ bool) []byte {
+				return []byte(test.translated)
+			}, ResponseTransform{})
+			out := registry.TranslateRequest(test.from, test.to, "model", []byte(test.input), false)
+			result := gjson.GetBytes(out, test.path)
+			if result.Exists() != test.wantExists {
+				t.Fatalf("%s exists = %v, want %v; body=%s", test.path, result.Exists(), test.wantExists, out)
+			}
+			if test.wantExists && result.String() != test.want {
+				t.Fatalf("%s = %q, want %q; body=%s", test.path, result.String(), test.want, out)
+			}
+		})
+	}
+}
+
+func TestRegistryTranslateRequestMakesExplicitClaudeVisibilityValid(t *testing.T) {
+	tests := []struct {
+		name        string
+		input       string
+		wantDisplay string
+	}{
+		{name: "summary auto is visible", input: `{"reasoning":{"summary":"auto"},"input":"hi"}`, wantDisplay: "summarized"},
+		{name: "summary null is hidden", input: `{"reasoning":{"summary":null},"input":"hi"}`, wantDisplay: "omitted"},
+	}
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			registry := NewRegistry()
+			registry.Register(FormatOpenAIResponse, FormatClaude, func(_ string, _ []byte, _ bool) []byte {
+				return []byte(`{"model":"claude-opus-5","max_tokens":32000}`)
+			}, ResponseTransform{})
+			out := registry.TranslateRequest(
+				FormatOpenAIResponse,
+				FormatClaude,
+				"claude-opus-5",
+				[]byte(test.input),
+				false,
+			)
+			if got := gjson.GetBytes(out, "thinking.type").String(); got != "adaptive" {
+				t.Fatalf("thinking.type = %q, want adaptive; body=%s", got, out)
+			}
+			if got := gjson.GetBytes(out, "thinking.display").String(); got != test.wantDisplay {
+				t.Fatalf("thinking.display = %q, want %q; body=%s", got, test.wantDisplay, out)
+			}
+		})
+	}
+}
+
+func TestRegistryTranslateRequestPreservesNativeClaudeMissingDisplay(t *testing.T) {
+	registry := NewRegistry()
+	body := []byte(`{"model":"claude-opus-5","thinking":{"type":"adaptive"}}`)
+	out := registry.TranslateRequest(FormatClaude, FormatClaude, "claude-opus-5", body, true)
+	if gjson.GetBytes(out, "thinking.display").Exists() {
+		t.Fatalf("native Claude request without display gained one: %s", out)
+	}
+}
diff --git a/test/summary_intent_translation_test.go b/test/summary_intent_translation_test.go
new file mode 100644
index 00000000..e2cfdd99
--- /dev/null
+++ b/test/summary_intent_translation_test.go
@@ -0,0 +1,226 @@
+package test
+
+import (
+	"fmt"
+	"testing"
+	"time"
+
+	"github.com/router-for-me/CLIProxyAPI/v7/internal/registry"
+	"github.com/router-for-me/CLIProxyAPI/v7/internal/thinking"
+	_ "github.com/router-for-me/CLIProxyAPI/v7/internal/thinking/provider/antigravity"
+	_ "github.com/router-for-me/CLIProxyAPI/v7/internal/translator"
+	sdktranslator "github.com/router-for-me/CLIProxyAPI/v7/sdk/translator"
+	"github.com/tidwall/gjson"
+)
+
+func TestSummaryIntentTranslation(t *testing.T) {
+	tests := []struct {
+		name       string
+		from       sdktranslator.Format
+		to         sdktranslator.Format
+		body       string
+		path       string
+		want       string
+		wantExists bool
+	}{
+		{name: "Chat effort enables Claude summary", from: sdktranslator.FormatOpenAI, to: sdktranslator.FormatClaude, body: `{"model":"claude-opus-5","reasoning_effort":"high","messages":[{"role":"user","content":"hi"}]}`, path: "thinking.display", want: "summarized", wantExists: true},
+		// Anthropic rejects display next to a disabled thinking block, so a "none"
+		// effort must leave the field off rather than write "omitted".
+		{name: "Chat none leaves disabled Claude thinking without display", from: sdktranslator.FormatOpenAI, to: sdktranslator.FormatClaude, body: `{"model":"claude-opus-5","reasoning_effort":"none","messages":[{"role":"user","content":"hi"}]}`, path: "thinking.display"},
+		// Anthropic requires thinking.type. For an unregistered target CPA cannot
+		// safely guess adaptive versus manual thinking, so it must not emit an
+		// invalid display-only object. Registered targets are covered below.
+		{name: "Unknown Claude target does not get display only thinking", from: sdktranslator.FormatOpenAI, to: sdktranslator.FormatClaude, body: `{"model":"unregistered-claude-model","reasoning":{"exclude":false},"messages":[{"role":"user","content":"hi"}]}`, path: "thinking"},
+		{name: "Unknown Claude target from Interactions stays valid", from: sdktranslator.FormatInteractions, to: sdktranslator.FormatClaude, body: `{"model":"unregistered-claude-model","generation_config":{"thinking_summaries":"auto"},"input":"hi"}`, path: "thinking"},
+		{name: "Chat none omits Codex summary", from: sdktranslator.FormatOpenAI, to: sdktranslator.FormatCodex, body: `{"model":"gpt-5.4","reasoning_effort":"none","messages":[{"role":"user","content":"hi"}]}`, path: "reasoning.summary"},
+		// The Responses API makes reasoning.summary an explicit opt-in, so an
+		// absent source intent must remain absent when translated to Codex.
+		{name: "Claude absent display leaves Codex summary absent", from: sdktranslator.FormatClaude, to: sdktranslator.FormatCodex, body: `{"model":"gpt-5.4","max_tokens":1024,"thinking":{"type":"adaptive"},"output_config":{"effort":"high"},"messages":[{"role":"user","content":"hi"}]}`, path: "reasoning.summary"},
+		{name: "Gemini absent includeThoughts leaves Codex summary absent", from: sdktranslator.FormatGemini, to: sdktranslator.FormatCodex, body: `{"model":"gpt-5.4","generationConfig":{"thinkingConfig":{"thinkingLevel":"high"}},"contents":[{"role":"user","parts":[{"text":"hi"}]}]}`, path: "reasoning.summary"},
+		{name: "Claude summarized enables Codex summary", from: sdktranslator.FormatClaude, to: sdktranslator.FormatCodex, body: `{"model":"gpt-5.4","max_tokens":1024,"thinking":{"type":"adaptive","display":"summarized"},"output_config":{"effort":"high"},"messages":[{"role":"user","content":"hi"}]}`, path: "reasoning.summary", want: "auto", wantExists: true},
+		{name: "Interactions none omits Codex summary", from: sdktranslator.FormatInteractions, to: sdktranslator.FormatCodex, body: `{"model":"gpt-5.4","generation_config":{"thinking_level":"high","thinking_summaries":"none"},"input":"hi"}`, path: "reasoning.summary"},
+		{name: "Chat effort enables Codex summary", from: sdktranslator.FormatOpenAI, to: sdktranslator.FormatCodex, body: `{"model":"gpt-5.4","reasoning_effort":"high","messages":[{"role":"user","content":"hi"}]}`, path: "reasoning.summary", want: "auto", wantExists: true},
+		{name: "Responses summary only enables Chat compatibility effort", from: sdktranslator.FormatOpenAIResponse, to: sdktranslator.FormatOpenAI, body: `{"model":"gpt-5.4","reasoning":{"summary":"auto"},"input":"hi"}`, path: "reasoning_effort", want: "medium", wantExists: true},
+		// Chat has no field for "reason but hide": OpenAI documents none and rejects
+		// unknown parameters, so a disabled summary must leave the requested effort
+		// alone instead of turning reasoning off upstream.
+		{name: "Responses disabled summary keeps Chat effort", from: sdktranslator.FormatOpenAIResponse, to: sdktranslator.FormatOpenAI, body: `{"model":"gpt-5.4","reasoning":{"effort":"high","summary":null},"input":"hi"}`, path: "reasoning_effort", want: "high", wantExists: true},
+		{name: "Gemini disabled summary keeps Chat effort", from: sdktranslator.FormatGemini, to: sdktranslator.FormatOpenAI, body: `{"model":"gpt-5.4","generationConfig":{"thinkingConfig":{"thinkingLevel":"high","includeThoughts":false}},"contents":[{"role":"user","parts":[{"text":"hi"}]}]}`, path: "reasoning_effort", want: "high", wantExists: true},
+		{name: "Claude omitted display keeps Chat effort", from: sdktranslator.FormatClaude, to: sdktranslator.FormatOpenAI, body: `{"model":"gpt-5.4","thinking":{"type":"adaptive","display":"omitted"},"output_config":{"effort":"high"},"messages":[{"role":"user","content":"hi"}]}`, path: "reasoning_effort", want: "high", wantExists: true},
+		{name: "Chat without effort leaves Claude display absent", from: sdktranslator.FormatOpenAI, to: sdktranslator.FormatClaude, body: `{"model":"claude-opus-5","messages":[{"role":"user","content":"hi"}]}`, path: "thinking.display"},
+		{name: "Responses effort alone leaves Claude display absent", from: sdktranslator.FormatOpenAIResponse, to: sdktranslator.FormatClaude, body: `{"model":"claude-opus-5","reasoning":{"effort":"high"},"input":"hi"}`, path: "thinking.display"},
+		{name: "Responses summary enables Claude summary", from: sdktranslator.FormatOpenAIResponse, to: sdktranslator.FormatClaude, body: `{"model":"claude-opus-5","reasoning":{"effort":"high","summary":"auto"},"input":"hi"}`, path: "thinking.display", want: "summarized", wantExists: true},
+		{name: "Responses null summary disables Claude summary", from: sdktranslator.FormatOpenAIResponse, to: sdktranslator.FormatClaude, body: `{"model":"claude-opus-5","reasoning":{"effort":"high","summary":null},"input":"hi"}`, path: "thinking.display", want: "omitted", wantExists: true},
+		{name: "Chat effort enables Gemini summary", from: sdktranslator.FormatOpenAI, to: sdktranslator.FormatGemini, body: `{"model":"gemini-3.6-flash","reasoning_effort":"high","messages":[{"role":"user","content":"hi"}]}`, path: "generationConfig.thinkingConfig.includeThoughts", want: "true", wantExists: true},
+		{name: "Chat none disables Gemini summary", from: sdktranslator.FormatOpenAI, to: sdktranslator.FormatGemini, body: `{"model":"gemini-3.6-flash","reasoning_effort":"none","messages":[{"role":"user","content":"hi"}]}`, path: "generationConfig.thinkingConfig.includeThoughts", want: "false", wantExists: true},
+		{name: "Chat effort enables Antigravity summary", from: sdktranslator.FormatOpenAI, to: sdktranslator.FormatAntigravity, body: `{"model":"gemini-3.6-flash","reasoning_effort":"high","messages":[{"role":"user","content":"hi"}]}`, path: "request.generationConfig.thinkingConfig.includeThoughts", want: "true", wantExists: true},
+		{name: "Google Chat extension overrides Gemini summary", from: sdktranslator.FormatOpenAI, to: sdktranslator.FormatGemini, body: `{"model":"gemini-3.6-flash","reasoning_effort":"high","extra_body":{"google":{"thinking_config":{"include_thoughts":false}}},"messages":[{"role":"user","content":"hi"}]}`, path: "generationConfig.thinkingConfig.includeThoughts", want: "false", wantExists: true},
+		{name: "Responses effort alone leaves Gemini summary absent", from: sdktranslator.FormatOpenAIResponse, to: sdktranslator.FormatGemini, body: `{"model":"gemini-3.6-flash","reasoning":{"effort":"high"},"input":"hi"}`, path: "generationConfig.thinkingConfig.includeThoughts"},
+		{name: "Responses detailed summary enables Gemini summary", from: sdktranslator.FormatOpenAIResponse, to: sdktranslator.FormatGemini, body: `{"model":"gemini-3.6-flash","reasoning":{"effort":"high","summary":"detailed"},"input":"hi"}`, path: "generationConfig.thinkingConfig.includeThoughts", want: "true", wantExists: true},
+		{name: "Responses effort alone leaves Antigravity summary absent", from: sdktranslator.FormatOpenAIResponse, to: sdktranslator.FormatAntigravity, body: `{"model":"gemini-3.6-flash","reasoning":{"effort":"high"},"input":"hi"}`, path: "request.generationConfig.thinkingConfig.includeThoughts"},
+		{name: "Responses summary enables Antigravity summary", from: sdktranslator.FormatOpenAIResponse, to: sdktranslator.FormatAntigravity, body: `{"model":"gemini-3.6-flash","reasoning":{"effort":"high","summary":"auto"},"input":"hi"}`, path: "request.generationConfig.thinkingConfig.includeThoughts", want: "true", wantExists: true},
+		{name: "Chat effort enables Interactions summary", from: sdktranslator.FormatOpenAI, to: sdktranslator.FormatInteractions, body: `{"model":"gemini-3.6-flash","reasoning_effort":"high","messages":[{"role":"user","content":"hi"}]}`, path: "generation_config.thinking_summaries", want: "auto", wantExists: true},
+		{name: "Responses concise summary maps to Interactions auto", from: sdktranslator.FormatOpenAIResponse, to: sdktranslator.FormatInteractions, body: `{"model":"gemini-3.6-flash","reasoning":{"effort":"high","summary":"concise"},"input":"hi"}`, path: "generation_config.thinking_summaries", want: "auto", wantExists: true},
+		{name: "Native Claude summarized enables Gemini summary", from: sdktranslator.FormatClaude, to: sdktranslator.FormatGemini, body: `{"model":"claude-opus-5","thinking":{"type":"adaptive","display":"summarized"},"messages":[{"role":"user","content":"hi"}]}`, path: "generationConfig.thinkingConfig.includeThoughts", want: "true", wantExists: true},
+		{name: "Native Gemini disabled omits Claude summary", from: sdktranslator.FormatGemini, to: sdktranslator.FormatClaude, body: `{"model":"gemini-3.6-flash","generationConfig":{"thinkingConfig":{"thinkingLevel":"high","includeThoughts":false}},"contents":[{"role":"user","parts":[{"text":"hi"}]}]}`, path: "thinking.display", want: "omitted", wantExists: true},
+		{name: "Native Gemini absent summary leaves Claude display absent", from: sdktranslator.FormatGemini, to: sdktranslator.FormatClaude, body: `{"model":"gemini-3.6-flash","generationConfig":{"thinkingConfig":{"thinkingLevel":"high"}},"contents":[{"role":"user","parts":[{"text":"hi"}]}]}`, path: "thinking.display"},
+		{name: "Native Interactions auto enables Gemini summary", from: sdktranslator.FormatInteractions, to: sdktranslator.FormatGemini, body: `{"model":"gemini-3.6-flash","generation_config":{"thinking_level":"high","thinking_summaries":"auto"},"input":"hi"}`, path: "generationConfig.thinkingConfig.includeThoughts", want: "true", wantExists: true},
+		{name: "Native Interactions none omits Claude summary", from: sdktranslator.FormatInteractions, to: sdktranslator.FormatClaude, body: `{"model":"claude-opus-5","generation_config":{"thinking_level":"high","thinking_summaries":"none"},"input":"hi"}`, path: "thinking.display", want: "omitted", wantExists: true},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			out := sdktranslator.TranslateRequest(test.from, test.to, "", []byte(test.body), true)
+			result := gjson.GetBytes(out, test.path)
+			if result.Exists() != test.wantExists {
+				t.Fatalf("%s exists = %v, want %v; body=%s", test.path, result.Exists(), test.wantExists, out)
+			}
+			if test.wantExists && result.String() != test.want {
+				t.Fatalf("%s = %q, want %q; body=%s", test.path, result.String(), test.want, out)
+			}
+		})
+	}
+}
+
+func TestInvalidInteractionsSummaryDoesNotWriteTargetControl(t *testing.T) {
+	body := []byte(`{"model":"model","generation_config":{"thinking_summaries":"banana"},"input":"hi"}`)
+	for _, test := range []struct {
+		name string
+		to   sdktranslator.Format
+		path string
+	}{
+		{name: "Gemini", to: sdktranslator.FormatGemini, path: "generationConfig.thinkingConfig.includeThoughts"},
+		{name: "Antigravity", to: sdktranslator.FormatAntigravity, path: "request.generationConfig.thinkingConfig.includeThoughts"},
+		{name: "Codex", to: sdktranslator.FormatCodex, path: "reasoning.summary"},
+	} {
+		t.Run(test.name, func(t *testing.T) {
+			out := sdktranslator.TranslateRequest(sdktranslator.FormatInteractions, test.to, "model", body, false)
+			if result := gjson.GetBytes(out, test.path); result.Exists() {
+				t.Fatalf("invalid Interactions summary wrote %s=%s; body=%s", test.path, result.Raw, out)
+			}
+		})
+	}
+}
+
+func TestSummaryIntentFinalPipeline(t *testing.T) {
+	reg := registry.GetGlobalRegistry()
+	uid := fmt.Sprintf("summary-final-pipeline-%d", time.Now().UnixNano())
+	reg.RegisterClient(uid, "test", getTestModels())
+	defer reg.UnregisterClient(uid)
+
+	tests := []struct {
+		name       string
+		from       sdktranslator.Format
+		to         sdktranslator.Format
+		model      string
+		body       string
+		path       string
+		want       string
+		wantExists bool
+	}{
+		{name: "Responses summary only activates visible Claude thinking", from: sdktranslator.FormatOpenAIResponse, to: sdktranslator.FormatClaude, model: "claude-sonnet-4-6-model", body: `{"model":"claude-sonnet-4-6-model","reasoning":{"summary":"auto"},"input":"hi"}`, path: "thinking.display", want: "summarized", wantExists: true},
+		{name: "Responses null summary only activates hidden Claude thinking", from: sdktranslator.FormatOpenAIResponse, to: sdktranslator.FormatClaude, model: "claude-sonnet-4-6-model", body: `{"model":"claude-sonnet-4-6-model","reasoning":{"summary":null},"input":"hi"}`, path: "thinking.display", want: "omitted", wantExists: true},
+		{name: "Responses default keeps Claude display default", from: sdktranslator.FormatOpenAIResponse, to: sdktranslator.FormatClaude, model: "claude-sonnet-4-6-model", body: `{"model":"claude-sonnet-4-6-model","input":"hi"}`, path: "thinking.display"},
+		{name: "Chat summary alias only activates valid Claude thinking", from: sdktranslator.FormatOpenAI, to: sdktranslator.FormatClaude, model: "claude-sonnet-4-6-model", body: `{"model":"claude-sonnet-4-6-model","reasoning":{"exclude":false},"messages":[{"role":"user","content":"hi"}]}`, path: "thinking.display", want: "summarized", wantExists: true},
+		{name: "Interactions summary only activates valid Claude thinking", from: sdktranslator.FormatInteractions, to: sdktranslator.FormatClaude, model: "claude-sonnet-4-6-model", body: `{"model":"claude-sonnet-4-6-model","generation_config":{"thinking_summaries":"auto"},"input":"hi"}`, path: "thinking.display", want: "summarized", wantExists: true},
+		{name: "Claude suffix none removes otherwise enabled display", from: sdktranslator.FormatOpenAIResponse, to: sdktranslator.FormatClaude, model: "claude-sonnet-4-6-model(none)", body: `{"model":"claude-sonnet-4-6-model(none)","reasoning":{"summary":"auto"},"input":"hi"}`, path: "thinking.display"},
+		{name: "Claude suffix preserves explicit disabled summary", from: sdktranslator.FormatOpenAIResponse, to: sdktranslator.FormatClaude, model: "claude-sonnet-4-6-model(high)", body: `{"model":"claude-sonnet-4-6-model(high)","reasoning":{"summary":null},"input":"hi"}`, path: "thinking.display", want: "omitted", wantExists: true},
+		{name: "Responses effort alone stays omitted on Antigravity", from: sdktranslator.FormatOpenAIResponse, to: sdktranslator.FormatAntigravity, model: "antigravity-budget-model", body: `{"model":"antigravity-budget-model","reasoning":{"effort":"medium"},"input":"hi"}`, path: "request.generationConfig.thinkingConfig.includeThoughts"},
+		{name: "Responses summary reaches Antigravity", from: sdktranslator.FormatOpenAIResponse, to: sdktranslator.FormatAntigravity, model: "antigravity-budget-model", body: `{"model":"antigravity-budget-model","reasoning":{"effort":"medium","summary":"auto"},"input":"hi"}`, path: "request.generationConfig.thinkingConfig.includeThoughts", want: "true", wantExists: true},
+		{name: "Google Chat extension false survives Gemini applier", from: sdktranslator.FormatOpenAI, to: sdktranslator.FormatGemini, model: "gemini-mixed-model", body: `{"model":"gemini-mixed-model","reasoning_effort":"high","extra_body":{"google":{"thinking_config":{"include_thoughts":false}}},"messages":[{"role":"user","content":"hi"}]}`, path: "generationConfig.thinkingConfig.includeThoughts", want: "false", wantExists: true},
+		// Captured from isolated Claude Code 2.1.220 with
+		// alwaysThinkingEnabled:true. Sonnet uses adaptive thinking, while Haiku
+		// uses manual enabled thinking with a budget; both explicitly omit text.
+		{name: "Claude Code Sonnet omitted thinking reaches Gemini", from: sdktranslator.FormatClaude, to: sdktranslator.FormatGemini, model: "gemini-mixed-model", body: `{"model":"claude-sonnet-4-6","thinking":{"type":"adaptive","display":"omitted"},"output_config":{"effort":"high"},"messages":[{"role":"user","content":"hi"}]}`, path: "generationConfig.thinkingConfig.includeThoughts", want: "false", wantExists: true},
+		{name: "Claude Code Sonnet omitted thinking reaches Antigravity", from: sdktranslator.FormatClaude, to: sdktranslator.FormatAntigravity, model: "antigravity-budget-model", body: `{"model":"claude-sonnet-4-6","thinking":{"type":"adaptive","display":"omitted"},"output_config":{"effort":"high"},"messages":[{"role":"user","content":"hi"}]}`, path: "request.generationConfig.thinkingConfig.includeThoughts", want: "false", wantExists: true},
+		{name: "Claude Code Haiku omitted thinking reaches Gemini", from: sdktranslator.FormatClaude, to: sdktranslator.FormatGemini, model: "gemini-mixed-model", body: `{"model":"claude-haiku-4-5-20251001","thinking":{"type":"enabled","budget_tokens":31999,"display":"omitted"},"messages":[{"role":"user","content":"hi"}]}`, path: "generationConfig.thinkingConfig.includeThoughts", want: "false", wantExists: true},
+		{name: "Claude Code Haiku omitted thinking reaches Antigravity", from: sdktranslator.FormatClaude, to: sdktranslator.FormatAntigravity, model: "antigravity-budget-model", body: `{"model":"claude-haiku-4-5-20251001","thinking":{"type":"enabled","budget_tokens":31999,"display":"omitted"},"messages":[{"role":"user","content":"hi"}]}`, path: "request.generationConfig.thinkingConfig.includeThoughts", want: "false", wantExists: true},
+		{name: "Summary-only control is stripped for non-thinking Gemini model", from: sdktranslator.FormatOpenAIResponse, to: sdktranslator.FormatGemini, model: "no-thinking-model", body: `{"model":"no-thinking-model","reasoning":{"summary":"auto"},"input":"hi"}`, path: "generationConfig.thinkingConfig"},
+		{name: "Interactions level alone keeps summaries omitted", from: sdktranslator.FormatInteractions, to: sdktranslator.FormatInteractions, model: "level-model", body: `{"model":"level-model","generation_config":{"thinking_level":"high"},"input":"hi"}`, path: "generation_config.thinking_summaries"},
+		{name: "Interactions auto survives its applier", from: sdktranslator.FormatInteractions, to: sdktranslator.FormatInteractions, model: "level-model", body: `{"model":"level-model","generation_config":{"thinking_level":"high","thinking_summaries":"auto"},"input":"hi"}`, path: "generation_config.thinking_summaries", want: "auto", wantExists: true},
+		{name: "Deprecated Responses detail reaches Codex", from: sdktranslator.FormatOpenAIResponse, to: sdktranslator.FormatCodex, model: "level-model", body: `{"model":"level-model","reasoning":{"effort":"high","generate_summary":"detailed"},"input":"hi"}`, path: "reasoning.summary", want: "detailed", wantExists: true},
+		{name: "Gemini missing includeThoughts stays omitted on Claude", from: sdktranslator.FormatGemini, to: sdktranslator.FormatClaude, model: "claude-sonnet-4-6-model", body: `{"model":"claude-sonnet-4-6-model","generationConfig":{"thinkingConfig":{"thinkingLevel":"high"}},"contents":[{"role":"user","parts":[{"text":"hi"}]}]}`, path: "thinking.display"},
+		{name: "Gemini true includeThoughts reaches Claude", from: sdktranslator.FormatGemini, to: sdktranslator.FormatClaude, model: "claude-sonnet-4-6-model", body: `{"model":"claude-sonnet-4-6-model","generationConfig":{"thinkingConfig":{"thinkingLevel":"high","includeThoughts":true}},"contents":[{"role":"user","parts":[{"text":"hi"}]}]}`, path: "thinking.display", want: "summarized", wantExists: true},
+		{name: "Native Antigravity budget keeps visibility omitted", from: sdktranslator.FormatAntigravity, to: sdktranslator.FormatAntigravity, model: "antigravity-budget-model", body: `{"model":"antigravity-budget-model","request":{"generationConfig":{"thinkingConfig":{"thinkingBudget":8192}},"contents":[{"role":"user","parts":[{"text":"hi"}]}]}}`, path: "request.generationConfig.thinkingConfig.includeThoughts"},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			baseModel := thinking.ParseSuffix(test.model).ModelName
+			out := sdktranslator.TranslateRequest(test.from, test.to, baseModel, []byte(test.body), true)
+			var err error
+			out, err = thinking.ApplyThinkingWithSummary(out, test.model, test.from.String(), test.to.String(), test.to.String(), thinking.ExtractSummaryConfig([]byte(test.body), test.from.String()))
+			if err != nil {
+				t.Fatalf("ApplyThinking() error = %v; body=%s", err, out)
+			}
+			result := gjson.GetBytes(out, test.path)
+			if result.Exists() != test.wantExists {
+				t.Fatalf("%s exists = %v, want %v; body=%s", test.path, result.Exists(), test.wantExists, out)
+			}
+			if test.wantExists && result.String() != test.want {
+				t.Fatalf("%s = %q, want %q; body=%s", test.path, result.String(), test.want, out)
+			}
+			if test.to == sdktranslator.FormatClaude && gjson.GetBytes(out, "thinking.type").String() == "disabled" && gjson.GetBytes(out, "thinking.display").Exists() {
+				t.Fatalf("disabled Claude thinking retained display: %s", out)
+			}
+		})
+	}
+}
+
+func TestNativeClaudeMissingDisplayPreservesSignatureOnlyHistory(t *testing.T) {
+	body := []byte(`{"model":"claude-opus-5","thinking":{"type":"adaptive"},"messages":[{"role":"assistant","content":[{"type":"thinking","thinking":"","signature":"opus-signature"}]},{"role":"user","content":"continue"}]}`)
+	out := sdktranslator.TranslateRequest(sdktranslator.FormatClaude, sdktranslator.FormatClaude, "claude-opus-5", body, true)
+	if gjson.GetBytes(out, "thinking.display").Exists() {
+		t.Fatalf("native Claude request without display gained one: %s", out)
+	}
+	if got := gjson.GetBytes(out, "messages").Raw; got != gjson.GetBytes(body, "messages").Raw {
+		t.Fatalf("signature-only history changed: got %s, want %s", got, gjson.GetBytes(body, "messages").Raw)
+	}
+}
+
+// Antigravity wraps Gemini generateContent, where includeThoughts is an
+// independent opt-in. Thinking level/budget changes must preserve explicit
+// booleans and leave an omitted visibility control omitted.
+func TestAntigravityIncludeThoughtsPreservesExplicitness(t *testing.T) {
+	reg := registry.GetGlobalRegistry()
+	uid := fmt.Sprintf("antigravity-summary-default-%d", time.Now().UnixNano())
+	reg.RegisterClient(uid, "test", getTestModels())
+	defer reg.UnregisterClient(uid)
+
+	const contents = `"contents":[{"role":"user","parts":[{"text":"hi"}]}]`
+	tests := []struct {
+		name       string
+		model      string
+		body       string
+		want       string
+		wantExists bool
+	}{
+		{name: "suffix thinking without intent stays omitted", model: "antigravity-budget-model(medium)", body: `{"request":{` + contents + `}}`},
+		{name: "native budget without intent stays omitted", model: "antigravity-budget-model", body: `{"request":{"generationConfig":{"thinkingConfig":{"thinkingBudget":8192}},` + contents + `}}`},
+		{name: "explicit true is preserved", model: "antigravity-budget-model(medium)", body: `{"request":{"generationConfig":{"thinkingConfig":{"includeThoughts":true}},` + contents + `}}`, want: "true", wantExists: true},
+		{name: "explicit false is preserved", model: "antigravity-budget-model(medium)", body: `{"request":{"generationConfig":{"thinkingConfig":{"includeThoughts":false}},` + contents + `}}`, want: "false", wantExists: true},
+		{name: "explicit snake case false is preserved", model: "antigravity-budget-model(medium)", body: `{"request":{"generationConfig":{"thinkingConfig":{"include_thoughts":false}},` + contents + `}}`, want: "false", wantExists: true},
+		{name: "disabled thinking without summary intent stays omitted", model: "antigravity-budget-model(none)", body: `{"request":{` + contents + `}}`},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			out, err := thinking.ApplyThinking([]byte(test.body), test.model, "antigravity", "antigravity", "antigravity")
+			if err != nil {
+				t.Fatalf("ApplyThinking() error = %v; body=%s", err, out)
+			}
+			result := gjson.GetBytes(out, "request.generationConfig.thinkingConfig.includeThoughts")
+			if result.Exists() != test.wantExists {
+				t.Fatalf("includeThoughts exists = %v, want %v; body=%s", result.Exists(), test.wantExists, out)
+			}
+			if test.wantExists {
+				if got := fmt.Sprintf("%v", result.Bool()); got != test.want {
+					t.Fatalf("includeThoughts = %s, want %s; body=%s", got, test.want, out)
+				}
+			}
+			if gjson.GetBytes(out, "request.generationConfig.thinkingConfig.include_thoughts").Exists() {
+				t.Fatalf("snake_case includeThoughts left in payload: %s", out)
+			}
+		})
+	}
+}
diff --git a/test/thinking_conversion_test.go b/test/thinking_conversion_test.go
index 2a95d107..07dfb039 100644
--- a/test/thinking_conversion_test.go
+++ b/test/thinking_conversion_test.go
@@ -1456,25 +1456,30 @@ func TestThinkingE2EMatrix_Body(t *testing.T) {
 			includeThoughts: "false",
 			expectErr:       false,
 		},
-		// Case 31A: reasoning_effort=none with zero allowed → delete thinkingConfig
+		// Case 31A: reasoning_effort=none with zero allowed removes the amount but
+		// preserves Chat's explicit disabled summary intent.
 		{
-			name:        "31A",
-			from:        "openai",
-			to:          "gemini",
-			model:       "gemini-toggle-mixed-model",
-			inputJSON:   `{"model":"gemini-toggle-mixed-model","messages":[{"role":"user","content":"hi"}],"reasoning_effort":"none"}`,
-			expectField: "",
-			expectErr:   false,
+			name:            "31A",
+			from:            "openai",
+			to:              "gemini",
+			model:           "gemini-toggle-mixed-model",
+			inputJSON:       `{"model":"gemini-toggle-mixed-model","messages":[{"role":"user","content":"hi"}],"reasoning_effort":"none"}`,
+			expectField:     "generationConfig.thinkingConfig.includeThoughts",
+			expectValue:     "false",
+			includeThoughts: "false",
+			expectErr:       false,
 		},
-		// Case 31B: reasoning_effort=none with zero allowed to Antigravity → delete thinkingConfig
+		// Case 31B: the same explicit disabled intent survives Antigravity.
 		{
-			name:        "31B",
-			from:        "openai",
-			to:          "antigravity",
-			model:       "gemini-toggle-mixed-model",
-			inputJSON:   `{"model":"gemini-toggle-mixed-model","messages":[{"role":"user","content":"hi"}],"reasoning_effort":"none"}`,
-			expectField: "",
-			expectErr:   false,
+			name:            "31B",
+			from:            "openai",
+			to:              "antigravity",
+			model:           "gemini-toggle-mixed-model",
+			inputJSON:       `{"model":"gemini-toggle-mixed-model","messages":[{"role":"user","content":"hi"}],"reasoning_effort":"none"}`,
+			expectField:     "request.generationConfig.thinkingConfig.includeThoughts",
+			expectValue:     "false",
+			includeThoughts: "false",
+			expectErr:       false,
 		},
 		// Case 31C: reasoning.effort=none with zero allowed → delete thinkingConfig
 		{
@@ -2448,7 +2453,7 @@ func TestThinkingE2EProviderTargets(t *testing.T) {
 			expectValue: "high",
 		},
 
-		// Interactions target: native API uses generation_config.thinking_level and thinking_summaries.
+		// Interactions target: native API uses generation_config.thinking_level and optional thinking_summaries.
 		{
 			name:         "I1",
 			from:         "interactions",
@@ -2461,16 +2466,310 @@ func TestThinkingE2EProviderTargets(t *testing.T) {
 			expectValue2: "auto",
 		},
 		{
-			name:         "I2",
+			name:        "I2",
+			from:        "interactions",
+			to:          "interactions",
+			model:       "level-model(8192)",
+			inputJSON:   `{"model":"level-model(8192)","input":"hi"}`,
+			expectField: "generation_config.thinking_level",
+			expectValue: "medium",
+		},
+		// Responses client against a chat-shaped provider. Because thinking is read
+		// back off the translated body, this pair only works if the request translator
+		// rewrites reasoning.effort as reasoning_effort; nothing else covered it.
+		{
+			name:        "R1",
+			from:        "openai-response",
+			to:          "openai",
+			model:       "level-model",
+			inputJSON:   `{"model":"level-model","input":"hi","reasoning":{"effort":"high"}}`,
+			expectField: "reasoning_effort",
+			expectValue: "high",
+		},
+		{
+			name:        "R2",
+			from:        "openai-response",
+			to:          "openai",
+			model:       "level-model",
+			inputJSON:   `{"model":"level-model","input":"hi","reasoning":{"effort":"none"}}`,
+			expectField: "reasoning_effort",
+			expectValue: "minimal",
+		},
+		{
+			name:         "R3",
+			from:         "openai-response",
+			to:           "kimi",
+			model:        "kimi-toggle-thinking-model",
+			inputJSON:    `{"model":"kimi-toggle-thinking-model","input":"hi","reasoning":{"effort":"high"}}`,
+			expectField:  "thinking.type",
+			expectValue:  "enabled",
+			expectField2: "thinking.effort",
+			expectValue2: "high",
+			expectAbsent: []string{"reasoning_effort"},
+		},
+		{
+			name:            "R4",
+			from:            "openai-response",
+			to:              "antigravity",
+			model:           "antigravity-budget-model",
+			inputJSON:       `{"model":"antigravity-budget-model","input":"hi","reasoning":{"effort":"medium"}}`,
+			expectField:     "request.generationConfig.thinkingConfig.thinkingBudget",
+			expectValue:     "8192",
+			includeThoughts: "true",
+		},
+	}
+
+	runThinkingTests(t, cases)
+}
+
+// TestThinkingE2EInteractionsMatrix covers the Interactions protocol in both
+// directions, which the suffix and body matrices above barely touch.
+//
+// Interactions expresses thinking through generation_config.thinking_level and the
+// independent auto/none generation_config.thinking_summaries control. Compatibility
+// thinking_budget and none/auto level inputs map onto a documented target level. The
+// IN cases drive Interactions
+// as the provider from every client protocol; the OUT cases drive an Interactions
+// client against every provider, so an explicit on/off request has to survive the
+// round trip in both roles.
+func TestThinkingE2EInteractionsMatrix(t *testing.T) {
+	reg := registry.GetGlobalRegistry()
+	uid := fmt.Sprintf("thinking-e2e-interactions-%d", time.Now().UnixNano())
+
+	reg.RegisterClient(uid, "test", getTestModels())
+	defer reg.UnregisterClient(uid)
+
+	cases := []thinkingTestCase{
+		// Interactions as provider: explicit on from every client protocol.
+		{
+			name:        "IN1",
+			from:        "claude",
+			to:          "interactions",
+			model:       "level-model",
+			inputJSON:   `{"model":"level-model","max_tokens":1024,"messages":[{"role":"user","content":"hi"}],"thinking":{"type":"enabled","budget_tokens":10000}}`,
+			expectField: "generation_config.thinking_level",
+			expectValue: "high",
+		},
+		{
+			name:        "IN2",
+			from:        "openai",
+			to:          "interactions",
+			model:       "level-model",
+			inputJSON:   `{"model":"level-model","messages":[{"role":"user","content":"hi"}],"reasoning_effort":"minimal"}`,
+			expectField: "generation_config.thinking_level",
+			expectValue: "minimal",
+		},
+		{
+			name:        "IN3",
+			from:        "openai-response",
+			to:          "interactions",
+			model:       "level-model",
+			inputJSON:   `{"model":"level-model","input":"hi","reasoning":{"effort":"low"}}`,
+			expectField: "generation_config.thinking_level",
+			expectValue: "low",
+		},
+		{
+			name:        "IN4",
+			from:        "gemini",
+			to:          "interactions",
+			model:       "level-model",
+			inputJSON:   `{"model":"level-model","contents":[{"role":"user","parts":[{"text":"hi"}]}],"generationConfig":{"thinkingConfig":{"includeThoughts":true,"thinkingBudget":20000}}}`,
+			expectField: "generation_config.thinking_level",
+			expectValue: "high",
+		},
+		// A level the model does not publish falls back to its highest level.
+		{
+			name:        "IN5",
+			from:        "openai",
+			to:          "interactions",
+			model:       "level-subset-model",
+			inputJSON:   `{"model":"level-subset-model","messages":[{"role":"user","content":"hi"}],"reasoning_effort":"xhigh"}`,
+			expectField: "generation_config.thinking_level",
+			expectValue: "high",
+		},
+		// Interactions cannot fully disable this model, so thinking clamps to the
+		// lowest documented level. Summary visibility remains omitted unless the
+		// source independently requested it.
+		{
+			name:        "IN6",
+			from:        "claude",
+			to:          "interactions",
+			model:       "level-model",
+			inputJSON:   `{"model":"level-model","max_tokens":1024,"messages":[{"role":"user","content":"hi"}],"thinking":{"type":"disabled"}}`,
+			expectField: "generation_config.thinking_level",
+			expectValue: "minimal",
+		},
+		{
+			name:        "IN7",
+			from:        "openai",
+			to:          "interactions",
+			model:       "level-model(none)",
+			inputJSON:   `{"model":"level-model(none)","messages":[{"role":"user","content":"hi"}]}`,
+			expectField: "generation_config.thinking_level",
+			expectValue: "minimal",
+		},
+		{
+			name:        "IN8",
+			from:        "interactions",
+			to:          "interactions",
+			model:       "level-model",
+			inputJSON:   `{"model":"level-model","generation_config":{"thinking_level":"none"},"input":"hi"}`,
+			expectField: "generation_config.thinking_level",
+			expectValue: "minimal",
+		},
+		// Interactions supports auto as its only enabled summary selector.
+		{
+			name:         "IN9",
 			from:         "interactions",
 			to:           "interactions",
-			model:        "level-model(8192)",
-			inputJSON:    `{"model":"level-model(8192)","input":"hi"}`,
+			model:        "level-model",
+			inputJSON:    `{"model":"level-model","generation_config":{"thinking_level":"low","thinking_summaries":"auto"},"input":"hi"}`,
 			expectField:  "generation_config.thinking_level",
-			expectValue:  "medium",
+			expectValue:  "low",
 			expectField2: "generation_config.thinking_summaries",
 			expectValue2: "auto",
 		},
+		// A legacy thinking_budget maps onto the level enum.
+		{
+			name:        "IN10",
+			from:        "interactions",
+			to:          "interactions",
+			model:       "level-model",
+			inputJSON:   `{"model":"level-model","generation_config":{"thinking_budget":400},"input":"hi"}`,
+			expectField: "generation_config.thinking_level",
+			expectValue: "minimal",
+		},
+		// Auto on a model without dynamic thinking resolves to the mid-range level,
+		// the same normalization every other target gets.
+		{
+			name:        "IN11",
+			from:        "interactions",
+			to:          "interactions",
+			model:       "level-model",
+			inputJSON:   `{"model":"level-model","generation_config":{"thinking_budget":-1},"input":"hi"}`,
+			expectField: "generation_config.thinking_level",
+			expectValue: "medium",
+		},
+
+		// Interactions as client: explicit on has to reach every provider's own knob.
+		{
+			name:        "OUT1",
+			from:        "interactions",
+			to:          "claude",
+			model:       "claude-budget-model",
+			inputJSON:   `{"model":"claude-budget-model","generation_config":{"thinking_level":"medium"},"input":"hi"}`,
+			expectField: "thinking.budget_tokens",
+			expectValue: "8192",
+		},
+		{
+			name:        "OUT2",
+			from:        "interactions",
+			to:          "openai",
+			model:       "level-model",
+			inputJSON:   `{"model":"level-model","generation_config":{"thinking_level":"high"},"input":"hi"}`,
+			expectField: "reasoning_effort",
+			expectValue: "high",
+		},
+		{
+			name:        "OUT3",
+			from:        "interactions",
+			to:          "codex",
+			model:       "level-model",
+			inputJSON:   `{"model":"level-model","generation_config":{"thinking_level":"low"},"input":"hi"}`,
+			expectField: "reasoning.effort",
+			expectValue: "low",
+		},
+		{
+			name:        "OUT4",
+			from:        "interactions",
+			to:          "gemini",
+			model:       "gemini-budget-model",
+			inputJSON:   `{"model":"gemini-budget-model","generation_config":{"thinking_level":"medium"},"input":"hi"}`,
+			expectField: "generationConfig.thinkingConfig.thinkingBudget",
+			expectValue: "8192",
+		},
+		{
+			name:            "OUT5",
+			from:            "interactions",
+			to:              "antigravity",
+			model:           "antigravity-budget-model",
+			inputJSON:       `{"model":"antigravity-budget-model","generation_config":{"thinking_level":"medium"},"input":"hi"}`,
+			expectField:     "request.generationConfig.thinkingConfig.thinkingBudget",
+			expectValue:     "8192",
+			includeThoughts: "true",
+		},
+		{
+			name:         "OUT6",
+			from:         "interactions",
+			to:           "kimi",
+			model:        "kimi-toggle-thinking-model",
+			inputJSON:    `{"model":"kimi-toggle-thinking-model","generation_config":{"thinking_level":"high"},"input":"hi"}`,
+			expectField:  "thinking.type",
+			expectValue:  "enabled",
+			expectField2: "thinking.effort",
+			expectValue2: "high",
+		},
+		{
+			name:        "OUT7",
+			from:        "interactions",
+			to:          "xai",
+			model:       "xai-level-model",
+			inputJSON:   `{"model":"xai-level-model","generation_config":{"thinking_level":"high"},"input":"hi"}`,
+			expectField: "reasoning.effort",
+			expectValue: "high",
+		},
+		// Interactions as client: explicit off has to reach every provider's own way
+		// of saying no thinking.
+		{
+			name:        "OUT8",
+			from:        "interactions",
+			to:          "claude",
+			model:       "claude-budget-model",
+			inputJSON:   `{"model":"claude-budget-model","generation_config":{"thinking_level":"none"},"input":"hi"}`,
+			expectField: "thinking.type",
+			expectValue: "disabled",
+		},
+		{
+			name:            "OUT9",
+			from:            "interactions",
+			to:              "antigravity",
+			model:           "antigravity-budget-model",
+			inputJSON:       `{"model":"antigravity-budget-model","generation_config":{"thinking_level":"none"},"input":"hi"}`,
+			expectField:     "request.generationConfig.thinkingConfig.thinkingBudget",
+			expectValue:     "0",
+			includeThoughts: "false",
+		},
+		{
+			name:         "OUT10",
+			from:         "interactions",
+			to:           "kimi",
+			model:        "kimi-toggle-thinking-model",
+			inputJSON:    `{"model":"kimi-toggle-thinking-model","generation_config":{"thinking_level":"none"},"input":"hi"}`,
+			expectField:  "thinking.type",
+			expectValue:  "disabled",
+			expectAbsent: []string{"thinking.effort", "reasoning_effort"},
+		},
+		// A level+budget model that allows zero expresses off by dropping
+		// thinkingConfig entirely, so an Interactions client reaches the same shape a
+		// chat or Responses client does.
+		{
+			name:         "OUT11",
+			from:         "interactions",
+			to:           "gemini",
+			model:        "gemini-toggle-mixed-model",
+			inputJSON:    `{"model":"gemini-toggle-mixed-model","generation_config":{"thinking_level":"none"},"input":"hi"}`,
+			expectAbsent: []string{"generationConfig.thinkingConfig"},
+		},
+		// Auto reaches a dynamic-capable provider as dynamic thinking.
+		{
+			name:        "OUT12",
+			from:        "interactions",
+			to:          "gemini",
+			model:       "gemini-budget-model",
+			inputJSON:   `{"model":"gemini-budget-model","generation_config":{"thinking_level":"auto"},"input":"hi"}`,
+			expectField: "generationConfig.thinkingConfig.thinkingBudget",
+			expectValue: "-1",
+		},
 	}
 
 	runThinkingTests(t, cases)
@@ -3204,18 +3503,36 @@ func runThinkingTests(t *testing.T, cases []thinkingTestCase) {
 				assertField(tc.expectField3, tc.expectValue3)
 			}
 
-			if tc.includeThoughts != "" && (tc.to == "gemini" || tc.to == "antigravity") {
+			if tc.to == "gemini" || tc.to == "antigravity" {
 				path := "generationConfig.thinkingConfig.includeThoughts"
 				if tc.to == "antigravity" {
 					path = "request.generationConfig.thinkingConfig.includeThoughts"
 				}
-				itVal := gjson.GetBytes(body, path)
-				if !itVal.Exists() {
-					t.Fatalf("expected includeThoughts field not found, body=%s", string(body))
+				wantIncludeThoughts := ""
+				summaryConfig := thinking.ExtractSummaryConfig([]byte(tc.inputJSON), tc.from)
+				switch summaryConfig.Mode {
+				case thinking.SummaryEnabled:
+					wantIncludeThoughts = "true"
+				case thinking.SummaryDisabled:
+					wantIncludeThoughts = "false"
+				default:
+					// Thinking amount does not imply summary visibility. Keep the
+					// provider field absent when the source omitted its summary control.
 				}
-				actual := fmt.Sprintf("%v", itVal.Bool())
-				if actual != tc.includeThoughts {
-					t.Fatalf("includeThoughts: expected %s, got %s, body=%s", tc.includeThoughts, actual, string(body))
+
+				itVal := gjson.GetBytes(body, path)
+				if wantIncludeThoughts == "" {
+					if itVal.Exists() {
+						t.Fatalf("includeThoughts should be absent without summary intent, body=%s", string(body))
+					}
+				} else {
+					if !itVal.Exists() {
+						t.Fatalf("expected includeThoughts field not found, body=%s", string(body))
+					}
+					actual := fmt.Sprintf("%v", itVal.Bool())
+					if actual != wantIncludeThoughts {
+						t.Fatalf("includeThoughts: expected %s, got %s, body=%s", wantIncludeThoughts, actual, string(body))
+					}
 				}
 			}
 		})
-- 
2.51.2


From b63a38d566ab04bf2237084771d7bc303475031c Mon Sep 17 00:00:00 2001
From: sususu <suchangshan@foxmail.com>
Date: Thu, 30 Jul 2026 22:38:07 +0800
Subject: [PATCH 2/9] fix(thinking): avoid enabling Claude for hidden summaries

---
 .../thinking/apply_configured_api_key_test.go | 58 ++++++++++--------
 internal/thinking/summary.go                  | 12 ++--
 internal/thinking/summary_test.go             | 38 +++++++-----
 sdk/translator/registry_summary_test.go       | 61 ++++++++++---------
 test/summary_intent_translation_test.go       |  4 +-
 5 files changed, 98 insertions(+), 75 deletions(-)

diff --git a/internal/thinking/apply_configured_api_key_test.go b/internal/thinking/apply_configured_api_key_test.go
index 5aa3ce9d..a443491e 100644
--- a/internal/thinking/apply_configured_api_key_test.go
+++ b/internal/thinking/apply_configured_api_key_test.go
@@ -92,36 +92,44 @@ func TestApplyThinkingWithModelInfoKeepsSameFamilyValidationStrict(t *testing.T)
 	}
 }
 
-func TestApplyThinkingWithModelInfoAppliesSummaryOnlyClaudeVisibility(t *testing.T) {
+func TestApplyThinkingWithModelInfoAppliesEnabledSummaryOnlyClaudeVisibility(t *testing.T) {
 	modelInfo := &registry.ModelInfo{
 		ID:       "private-claude",
 		Type:     "claude",
 		Thinking: &registry.ThinkingSupport{Levels: []string{"high"}},
 	}
-	for _, test := range []struct {
-		name    string
-		source  string
-		display string
-	}{
-		{name: "enabled", source: `{"reasoning":{"summary":"auto"}}`, display: "summarized"},
-		{name: "disabled", source: `{"reasoning":{"summary":null}}`, display: "omitted"},
-	} {
-		t.Run(test.name, func(t *testing.T) {
-			out, err := thinking.ApplyThinkingWithModelInfo(
-				[]byte(`{"model":"private-claude","max_tokens":32000}`),
-				[]byte(test.source),
-				"private-claude", "openai-response", "claude", "claude", modelInfo,
-			)
-			if err != nil {
-				t.Fatalf("ApplyThinkingWithModelInfo() error = %v", err)
-			}
-			if got := gjson.GetBytes(out, "thinking.type").String(); got != "adaptive" {
-				t.Fatalf("thinking.type = %q, want adaptive; body=%s", got, out)
-			}
-			if got := gjson.GetBytes(out, "thinking.display").String(); got != test.display {
-				t.Fatalf("thinking.display = %q, want %q; body=%s", got, test.display, out)
-			}
-		})
+	out, err := thinking.ApplyThinkingWithModelInfo(
+		[]byte(`{"model":"private-claude","max_tokens":32000}`),
+		[]byte(`{"reasoning":{"summary":"auto"}}`),
+		"private-claude", "openai-response", "claude", "claude", modelInfo,
+	)
+	if err != nil {
+		t.Fatalf("ApplyThinkingWithModelInfo() error = %v", err)
+	}
+	if got := gjson.GetBytes(out, "thinking.type").String(); got != "adaptive" {
+		t.Fatalf("thinking.type = %q, want adaptive; body=%s", got, out)
+	}
+	if got := gjson.GetBytes(out, "thinking.display").String(); got != "summarized" {
+		t.Fatalf("thinking.display = %q, want summarized; body=%s", got, out)
+	}
+}
+
+func TestApplyThinkingWithModelInfoDoesNotActivateClaudeForDisabledSummary(t *testing.T) {
+	modelInfo := &registry.ModelInfo{
+		ID:       "private-claude",
+		Type:     "claude",
+		Thinking: &registry.ThinkingSupport{Levels: []string{"high"}},
+	}
+	out, err := thinking.ApplyThinkingWithModelInfo(
+		[]byte(`{"model":"private-claude","max_tokens":32000}`),
+		[]byte(`{"reasoning":{"summary":null}}`),
+		"private-claude", "openai-response", "claude", "claude", modelInfo,
+	)
+	if err != nil {
+		t.Fatalf("ApplyThinkingWithModelInfo() error = %v", err)
+	}
+	if gjson.GetBytes(out, "thinking").Exists() {
+		t.Fatalf("disabled summary activated Claude thinking: %s", out)
 	}
 }
 
diff --git a/internal/thinking/summary.go b/internal/thinking/summary.go
index 4a19dae1..430a35d8 100644
--- a/internal/thinking/summary.go
+++ b/internal/thinking/summary.go
@@ -126,12 +126,12 @@ func applySummaryConfigForModel(body []byte, format, model string, modelInfo *re
 		body = applyOpenAIChatSummaryConfig(body, model, enabled)
 	case "claude":
 		// Anthropic documents display as invalid with thinking.type=disabled and
-		// requires it alongside adaptive or enabled thinking. An explicit source
-		// visibility request is independent of thinking effort, so activate the
-		// target model's documented thinking mode before writing either
-		// summarized or omitted. Unspecified intent returns above and leaves the
-		// target's default untouched.
-		if !gjson.GetBytes(body, "thinking.type").Exists() {
+		// requires it alongside adaptive or enabled thinking. An enabled source
+		// summary needs an active target thinking mode. A disabled summary only
+		// hides an already-active target thinking mode; it must not enable thinking
+		// merely to hide a summary that would not otherwise exist. Unspecified
+		// intent returns above and leaves the target's default untouched.
+		if enabled && !gjson.GetBytes(body, "thinking.type").Exists() {
 			body = enableClaudeThinkingForSummary(body, model, modelInfo)
 		}
 		if !claudeThinkingAcceptsDisplay(body) {
diff --git a/internal/thinking/summary_test.go b/internal/thinking/summary_test.go
index 6c9011f1..59bd2eb4 100644
--- a/internal/thinking/summary_test.go
+++ b/internal/thinking/summary_test.go
@@ -149,29 +149,25 @@ func TestApplySummaryConfig_ClaudeDisplayRequiresActiveThinking(t *testing.T) {
 	}
 }
 
-func TestApplySummaryConfigForModel_ClaudeExplicitVisibilityUsesValidThinkingMode(t *testing.T) {
+func TestApplySummaryConfigForModel_ClaudeEnabledSummaryUsesValidThinkingMode(t *testing.T) {
 	tests := []struct {
-		name        string
-		model       string
-		body        string
-		mode        SummaryMode
-		wantType    string
-		wantDisplay string
-		wantBudget  int64
+		name       string
+		model      string
+		body       string
+		wantType   string
+		wantBudget int64
 	}{
-		{name: "adaptive model summarized", model: "claude-opus-5", body: `{"model":"claude-opus-5","max_tokens":32000}`, mode: SummaryEnabled, wantType: "adaptive", wantDisplay: "summarized"},
-		{name: "adaptive model omitted", model: "claude-opus-5", body: `{"model":"claude-opus-5","max_tokens":32000}`, mode: SummaryDisabled, wantType: "adaptive", wantDisplay: "omitted"},
-		{name: "manual model summarized", model: "claude-haiku-4-5-20251001", body: `{"model":"claude-haiku-4-5-20251001","max_tokens":32000}`, mode: SummaryEnabled, wantType: "enabled", wantDisplay: "summarized", wantBudget: 1024},
-		{name: "manual model omitted", model: "claude-haiku-4-5-20251001", body: `{"model":"claude-haiku-4-5-20251001","max_tokens":32000}`, mode: SummaryDisabled, wantType: "enabled", wantDisplay: "omitted", wantBudget: 1024},
+		{name: "adaptive model", model: "claude-opus-5", body: `{"model":"claude-opus-5","max_tokens":32000}`, wantType: "adaptive"},
+		{name: "manual model", model: "claude-haiku-4-5-20251001", body: `{"model":"claude-haiku-4-5-20251001","max_tokens":32000}`, wantType: "enabled", wantBudget: 1024},
 	}
 	for _, test := range tests {
 		t.Run(test.name, func(t *testing.T) {
-			out := ApplySummaryConfigForModel([]byte(test.body), "claude", test.model, SummaryConfig{Mode: test.mode})
+			out := ApplySummaryConfigForModel([]byte(test.body), "claude", test.model, SummaryConfig{Mode: SummaryEnabled})
 			if got := gjson.GetBytes(out, "thinking.type").String(); got != test.wantType {
 				t.Fatalf("thinking.type = %q, want %q; body=%s", got, test.wantType, out)
 			}
-			if got := gjson.GetBytes(out, "thinking.display").String(); got != test.wantDisplay {
-				t.Fatalf("thinking.display = %q, want %q; body=%s", got, test.wantDisplay, out)
+			if got := gjson.GetBytes(out, "thinking.display").String(); got != "summarized" {
+				t.Fatalf("thinking.display = %q, want summarized; body=%s", got, out)
 			}
 			if test.wantBudget > 0 && gjson.GetBytes(out, "thinking.budget_tokens").Int() != test.wantBudget {
 				t.Fatalf("thinking.budget_tokens = %d, want %d; body=%s", gjson.GetBytes(out, "thinking.budget_tokens").Int(), test.wantBudget, out)
@@ -180,6 +176,18 @@ func TestApplySummaryConfigForModel_ClaudeExplicitVisibilityUsesValidThinkingMod
 	}
 }
 
+// Disabling summaries must not activate Claude thinking. Doing so would add
+// reasoning tokens, latency, and cost to a request that asked only to hide output.
+func TestApplySummaryConfigForModel_ClaudeDisabledSummaryDoesNotEnableThinking(t *testing.T) {
+	for _, model := range []string{"claude-opus-5", "claude-haiku-4-5-20251001"} {
+		body := []byte(`{"model":"` + model + `","max_tokens":32000}`)
+		out := ApplySummaryConfigForModel(body, "claude", model, SummaryConfig{Mode: SummaryDisabled})
+		if gjson.GetBytes(out, "thinking").Exists() {
+			t.Fatalf("model %s gained thinking for a disabled summary: %s", model, out)
+		}
+	}
+}
+
 func TestApplySummaryConfig_ResponsesNormalizesDeprecatedGenerateSummary(t *testing.T) {
 	out := ApplySummaryConfig([]byte(`{"reasoning":{"generate_summary":"detailed"}}`), "openai-response", SummaryConfig{Mode: SummaryEnabled, Detail: "detailed"})
 	if got := gjson.GetBytes(out, "reasoning.summary").String(); got != "detailed" {
diff --git a/sdk/translator/registry_summary_test.go b/sdk/translator/registry_summary_test.go
index 79312dd8..312ee57b 100644
--- a/sdk/translator/registry_summary_test.go
+++ b/sdk/translator/registry_summary_test.go
@@ -85,35 +85,40 @@ func TestRegistryTranslateRequestAppliesSummaryIntent(t *testing.T) {
 	}
 }
 
-func TestRegistryTranslateRequestMakesExplicitClaudeVisibilityValid(t *testing.T) {
-	tests := []struct {
-		name        string
-		input       string
-		wantDisplay string
-	}{
-		{name: "summary auto is visible", input: `{"reasoning":{"summary":"auto"},"input":"hi"}`, wantDisplay: "summarized"},
-		{name: "summary null is hidden", input: `{"reasoning":{"summary":null},"input":"hi"}`, wantDisplay: "omitted"},
+func TestRegistryTranslateRequestActivatesClaudeForEnabledSummary(t *testing.T) {
+	registry := NewRegistry()
+	registry.Register(FormatOpenAIResponse, FormatClaude, func(_ string, _ []byte, _ bool) []byte {
+		return []byte(`{"model":"claude-opus-5","max_tokens":32000}`)
+	}, ResponseTransform{})
+	out := registry.TranslateRequest(
+		FormatOpenAIResponse,
+		FormatClaude,
+		"claude-opus-5",
+		[]byte(`{"reasoning":{"summary":"auto"},"input":"hi"}`),
+		false,
+	)
+	if got := gjson.GetBytes(out, "thinking.type").String(); got != "adaptive" {
+		t.Fatalf("thinking.type = %q, want adaptive; body=%s", got, out)
 	}
-	for _, test := range tests {
-		t.Run(test.name, func(t *testing.T) {
-			registry := NewRegistry()
-			registry.Register(FormatOpenAIResponse, FormatClaude, func(_ string, _ []byte, _ bool) []byte {
-				return []byte(`{"model":"claude-opus-5","max_tokens":32000}`)
-			}, ResponseTransform{})
-			out := registry.TranslateRequest(
-				FormatOpenAIResponse,
-				FormatClaude,
-				"claude-opus-5",
-				[]byte(test.input),
-				false,
-			)
-			if got := gjson.GetBytes(out, "thinking.type").String(); got != "adaptive" {
-				t.Fatalf("thinking.type = %q, want adaptive; body=%s", got, out)
-			}
-			if got := gjson.GetBytes(out, "thinking.display").String(); got != test.wantDisplay {
-				t.Fatalf("thinking.display = %q, want %q; body=%s", got, test.wantDisplay, out)
-			}
-		})
+	if got := gjson.GetBytes(out, "thinking.display").String(); got != "summarized" {
+		t.Fatalf("thinking.display = %q, want summarized; body=%s", got, out)
+	}
+}
+
+func TestRegistryTranslateRequestDoesNotActivateClaudeForDisabledSummary(t *testing.T) {
+	registry := NewRegistry()
+	registry.Register(FormatOpenAIResponse, FormatClaude, func(_ string, _ []byte, _ bool) []byte {
+		return []byte(`{"model":"claude-opus-5","max_tokens":32000}`)
+	}, ResponseTransform{})
+	out := registry.TranslateRequest(
+		FormatOpenAIResponse,
+		FormatClaude,
+		"claude-opus-5",
+		[]byte(`{"reasoning":{"summary":null},"input":"hi"}`),
+		false,
+	)
+	if gjson.GetBytes(out, "thinking").Exists() {
+		t.Fatalf("disabled summary activated Claude thinking: %s", out)
 	}
 }
 
diff --git a/test/summary_intent_translation_test.go b/test/summary_intent_translation_test.go
index e2cfdd99..06edfdad 100644
--- a/test/summary_intent_translation_test.go
+++ b/test/summary_intent_translation_test.go
@@ -119,7 +119,9 @@ func TestSummaryIntentFinalPipeline(t *testing.T) {
 		wantExists bool
 	}{
 		{name: "Responses summary only activates visible Claude thinking", from: sdktranslator.FormatOpenAIResponse, to: sdktranslator.FormatClaude, model: "claude-sonnet-4-6-model", body: `{"model":"claude-sonnet-4-6-model","reasoning":{"summary":"auto"},"input":"hi"}`, path: "thinking.display", want: "summarized", wantExists: true},
-		{name: "Responses null summary only activates hidden Claude thinking", from: sdktranslator.FormatOpenAIResponse, to: sdktranslator.FormatClaude, model: "claude-sonnet-4-6-model", body: `{"model":"claude-sonnet-4-6-model","reasoning":{"summary":null},"input":"hi"}`, path: "thinking.display", want: "omitted", wantExists: true},
+		// Disabling summaries alone must not activate Claude thinking: doing so adds
+		// reasoning tokens, latency, and cost to a request with no thinking effort.
+		{name: "Responses null summary alone keeps Claude thinking disabled", from: sdktranslator.FormatOpenAIResponse, to: sdktranslator.FormatClaude, model: "claude-sonnet-4-6-model", body: `{"model":"claude-sonnet-4-6-model","reasoning":{"summary":null},"input":"hi"}`, path: "thinking"},
 		{name: "Responses default keeps Claude display default", from: sdktranslator.FormatOpenAIResponse, to: sdktranslator.FormatClaude, model: "claude-sonnet-4-6-model", body: `{"model":"claude-sonnet-4-6-model","input":"hi"}`, path: "thinking.display"},
 		{name: "Chat summary alias only activates valid Claude thinking", from: sdktranslator.FormatOpenAI, to: sdktranslator.FormatClaude, model: "claude-sonnet-4-6-model", body: `{"model":"claude-sonnet-4-6-model","reasoning":{"exclude":false},"messages":[{"role":"user","content":"hi"}]}`, path: "thinking.display", want: "summarized", wantExists: true},
 		{name: "Interactions summary only activates valid Claude thinking", from: sdktranslator.FormatInteractions, to: sdktranslator.FormatClaude, model: "claude-sonnet-4-6-model", body: `{"model":"claude-sonnet-4-6-model","generation_config":{"thinking_summaries":"auto"},"input":"hi"}`, path: "thinking.display", want: "summarized", wantExists: true},
-- 
2.51.2


From 76008b4720823747560705a8b0f2baa2a10bd1df Mon Sep 17 00:00:00 2001
From: sususu <suchangshan@foxmail.com>
Date: Thu, 30 Jul 2026 22:44:01 +0800
Subject: [PATCH 3/9] fix(thinking): decouple Interactions effort summaries

---
 .../interactions_antigravity_request.go       |  5 ++--
 .../interactions_antigravity_test.go          | 29 +++++++++++++++++++
 test/summary_intent_translation_test.go       |  3 ++
 3 files changed, 35 insertions(+), 2 deletions(-)

diff --git a/internal/translator/antigravity/interactions/interactions_antigravity_request.go b/internal/translator/antigravity/interactions/interactions_antigravity_request.go
index 2d00a4f3..53d9df0e 100644
--- a/internal/translator/antigravity/interactions/interactions_antigravity_request.go
+++ b/internal/translator/antigravity/interactions/interactions_antigravity_request.go
@@ -195,12 +195,13 @@ func copyInteractionsReasoningToAntigravity(out []byte, root gjson.Result) []byt
 		effort = strings.ToLower(strings.TrimSpace(reasoning.Get("thinking_level").String()))
 	}
 	if effort != "" {
+		// Thinking amount and summary visibility are independent. This OpenAI-style
+		// compatibility alias controls only the amount; includeThoughts is written
+		// below only for an explicit Interactions summary selector.
 		if effort == "auto" {
 			out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.thinkingBudget", -1)
-			out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.includeThoughts", true)
 		} else {
 			out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.thinkingLevel", effort)
-			out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.includeThoughts", effort != "none")
 		}
 	}
 	if summary := reasoning.Get("summary"); summary.Exists() {
diff --git a/internal/translator/antigravity/interactions/interactions_antigravity_test.go b/internal/translator/antigravity/interactions/interactions_antigravity_test.go
index d6baa800..d0052a7b 100644
--- a/internal/translator/antigravity/interactions/interactions_antigravity_test.go
+++ b/internal/translator/antigravity/interactions/interactions_antigravity_test.go
@@ -68,6 +68,35 @@ func TestConvertInteractionsRequestToAntigravityPreservesGenerationConfig(t *tes
 	}
 }
 
+func TestConvertInteractionsReasoningToAntigravityKeepsSummaryIndependent(t *testing.T) {
+	tests := []struct {
+		name       string
+		reasoning  string
+		want       bool
+		wantExists bool
+	}{
+		{name: "effort only leaves summaries unspecified", reasoning: `{"effort":"high"}`},
+		{name: "explicit auto enables summaries", reasoning: `{"effort":"high","summary":"auto"}`, want: true, wantExists: true},
+		{name: "explicit none disables summaries", reasoning: `{"effort":"high","summary":"none"}`, wantExists: true},
+	}
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			body := []byte(`{"model":"antigravity-test","input":"hi","reasoning":` + test.reasoning + `}`)
+			out := ConvertInteractionsRequestToAntigravity("antigravity-test", body, false)
+			if got := gjson.GetBytes(out, "request.generationConfig.thinkingConfig.thinkingLevel").String(); got != "high" {
+				t.Fatalf("thinkingLevel = %q, want high. Output: %s", got, out)
+			}
+			includeThoughts := gjson.GetBytes(out, "request.generationConfig.thinkingConfig.includeThoughts")
+			if includeThoughts.Exists() != test.wantExists {
+				t.Fatalf("includeThoughts exists = %v, want %v. Output: %s", includeThoughts.Exists(), test.wantExists, out)
+			}
+			if test.wantExists && includeThoughts.Bool() != test.want {
+				t.Fatalf("includeThoughts = %v, want %v. Output: %s", includeThoughts.Bool(), test.want, out)
+			}
+		})
+	}
+}
+
 func TestConvertAntigravityResponseToInteractionsNonStream(t *testing.T) {
 	raw := []byte(`{"response":{"responseId":"resp_1","candidates":[{"content":{"role":"model","parts":[{"text":"ok"},{"functionCall":{"name":"lookup","id":"call_1","args":{"q":"x"}}}]},"finishReason":"STOP"}],"usageMetadata":{"promptTokenCount":3,"candidatesTokenCount":2,"totalTokenCount":5}}}`)
 	out := ConvertAntigravityResponseToInteractionsNonStream(context.Background(), "antigravity-test", nil, nil, raw, nil)
diff --git a/test/summary_intent_translation_test.go b/test/summary_intent_translation_test.go
index 06edfdad..5cb01362 100644
--- a/test/summary_intent_translation_test.go
+++ b/test/summary_intent_translation_test.go
@@ -140,6 +140,9 @@ func TestSummaryIntentFinalPipeline(t *testing.T) {
 		{name: "Summary-only control is stripped for non-thinking Gemini model", from: sdktranslator.FormatOpenAIResponse, to: sdktranslator.FormatGemini, model: "no-thinking-model", body: `{"model":"no-thinking-model","reasoning":{"summary":"auto"},"input":"hi"}`, path: "generationConfig.thinkingConfig"},
 		{name: "Interactions level alone keeps summaries omitted", from: sdktranslator.FormatInteractions, to: sdktranslator.FormatInteractions, model: "level-model", body: `{"model":"level-model","generation_config":{"thinking_level":"high"},"input":"hi"}`, path: "generation_config.thinking_summaries"},
 		{name: "Interactions auto survives its applier", from: sdktranslator.FormatInteractions, to: sdktranslator.FormatInteractions, model: "level-model", body: `{"model":"level-model","generation_config":{"thinking_level":"high","thinking_summaries":"auto"},"input":"hi"}`, path: "generation_config.thinking_summaries", want: "auto", wantExists: true},
+		{name: "Interactions reasoning effort leaves Antigravity summaries unspecified", from: sdktranslator.FormatInteractions, to: sdktranslator.FormatAntigravity, model: "antigravity-budget-model", body: `{"model":"antigravity-budget-model","reasoning":{"effort":"high"},"input":"hi"}`, path: "request.generationConfig.thinkingConfig.includeThoughts"},
+		{name: "Interactions reasoning summary auto reaches Antigravity", from: sdktranslator.FormatInteractions, to: sdktranslator.FormatAntigravity, model: "antigravity-budget-model", body: `{"model":"antigravity-budget-model","reasoning":{"effort":"high","summary":"auto"},"input":"hi"}`, path: "request.generationConfig.thinkingConfig.includeThoughts", want: "true", wantExists: true},
+		{name: "Interactions reasoning summary none reaches Antigravity", from: sdktranslator.FormatInteractions, to: sdktranslator.FormatAntigravity, model: "antigravity-budget-model", body: `{"model":"antigravity-budget-model","reasoning":{"effort":"high","summary":"none"},"input":"hi"}`, path: "request.generationConfig.thinkingConfig.includeThoughts", want: "false", wantExists: true},
 		{name: "Deprecated Responses detail reaches Codex", from: sdktranslator.FormatOpenAIResponse, to: sdktranslator.FormatCodex, model: "level-model", body: `{"model":"level-model","reasoning":{"effort":"high","generate_summary":"detailed"},"input":"hi"}`, path: "reasoning.summary", want: "detailed", wantExists: true},
 		{name: "Gemini missing includeThoughts stays omitted on Claude", from: sdktranslator.FormatGemini, to: sdktranslator.FormatClaude, model: "claude-sonnet-4-6-model", body: `{"model":"claude-sonnet-4-6-model","generationConfig":{"thinkingConfig":{"thinkingLevel":"high"}},"contents":[{"role":"user","parts":[{"text":"hi"}]}]}`, path: "thinking.display"},
 		{name: "Gemini true includeThoughts reaches Claude", from: sdktranslator.FormatGemini, to: sdktranslator.FormatClaude, model: "claude-sonnet-4-6-model", body: `{"model":"claude-sonnet-4-6-model","generationConfig":{"thinkingConfig":{"thinkingLevel":"high","includeThoughts":true}},"contents":[{"role":"user","parts":[{"text":"hi"}]}]}`, path: "thinking.display", want: "summarized", wantExists: true},
-- 
2.51.2


From 92b6bc4a868a3e5ecf2804f24239c213c9d6f575 Mon Sep 17 00:00:00 2001
From: sususu <suchangshan@foxmail.com>
Date: Thu, 30 Jul 2026 23:05:14 +0800
Subject: [PATCH 4/9] docs(thinking): clarify Claude model defaults

---
 internal/thinking/summary.go            | 18 +++++++++++++-----
 internal/thinking/summary_test.go       |  5 +++--
 test/summary_intent_translation_test.go |  7 ++++---
 3 files changed, 20 insertions(+), 10 deletions(-)

diff --git a/internal/thinking/summary.go b/internal/thinking/summary.go
index 430a35d8..5cca0a9b 100644
--- a/internal/thinking/summary.go
+++ b/internal/thinking/summary.go
@@ -126,11 +126,19 @@ func applySummaryConfigForModel(body []byte, format, model string, modelInfo *re
 		body = applyOpenAIChatSummaryConfig(body, model, enabled)
 	case "claude":
 		// Anthropic documents display as invalid with thinking.type=disabled and
-		// requires it alongside adaptive or enabled thinking. An enabled source
-		// summary needs an active target thinking mode. A disabled summary only
-		// hides an already-active target thinking mode; it must not enable thinking
-		// merely to hide a summary that would not otherwise exist. Unspecified
-		// intent returns above and leaves the target's default untouched.
+		// requires it alongside adaptive or enabled thinking. Model defaults differ:
+		// Opus 5 and Sonnet 5 default to adaptive thinking; Fable/Mythos 5 are always
+		// on. Opus 4.8/4.7/4.6, Sonnet 4.6, and the 4.5 models default to thinking
+		// off. The newest models also default display to omitted. Keeping a missing
+		// thinking block absent therefore preserves both kinds of model default;
+		// absence does not mean every Claude model runs without thinking. Only an
+		// enabled summary may activate a valid target thinking mode so that summarized
+		// text can be returned. A disabled summary only adds omitted to an
+		// already-active target mode.
+		//
+		// Anthropic docs:
+		// https://platform.claude.com/docs/en/build-with-claude/thinking
+		// https://platform.claude.com/docs/en/build-with-claude/thinking-troubleshooting#supported-models
 		if enabled && !gjson.GetBytes(body, "thinking.type").Exists() {
 			body = enableClaudeThinkingForSummary(body, model, modelInfo)
 		}
diff --git a/internal/thinking/summary_test.go b/internal/thinking/summary_test.go
index 59bd2eb4..1fed4147 100644
--- a/internal/thinking/summary_test.go
+++ b/internal/thinking/summary_test.go
@@ -176,8 +176,9 @@ func TestApplySummaryConfigForModel_ClaudeEnabledSummaryUsesValidThinkingMode(t
 	}
 }
 
-// Disabling summaries must not activate Claude thinking. Doing so would add
-// reasoning tokens, latency, and cost to a request that asked only to hide output.
+// Disabling summaries must not make CPA add a Claude thinking block. Absence
+// preserves the per-model default: newer models may still think by default,
+// while older models remain off.
 func TestApplySummaryConfigForModel_ClaudeDisabledSummaryDoesNotEnableThinking(t *testing.T) {
 	for _, model := range []string{"claude-opus-5", "claude-haiku-4-5-20251001"} {
 		body := []byte(`{"model":"` + model + `","max_tokens":32000}`)
diff --git a/test/summary_intent_translation_test.go b/test/summary_intent_translation_test.go
index 5cb01362..b53c7ca3 100644
--- a/test/summary_intent_translation_test.go
+++ b/test/summary_intent_translation_test.go
@@ -119,9 +119,10 @@ func TestSummaryIntentFinalPipeline(t *testing.T) {
 		wantExists bool
 	}{
 		{name: "Responses summary only activates visible Claude thinking", from: sdktranslator.FormatOpenAIResponse, to: sdktranslator.FormatClaude, model: "claude-sonnet-4-6-model", body: `{"model":"claude-sonnet-4-6-model","reasoning":{"summary":"auto"},"input":"hi"}`, path: "thinking.display", want: "summarized", wantExists: true},
-		// Disabling summaries alone must not activate Claude thinking: doing so adds
-		// reasoning tokens, latency, and cost to a request with no thinking effort.
-		{name: "Responses null summary alone keeps Claude thinking disabled", from: sdktranslator.FormatOpenAIResponse, to: sdktranslator.FormatClaude, model: "claude-sonnet-4-6-model", body: `{"model":"claude-sonnet-4-6-model","reasoning":{"summary":null},"input":"hi"}`, path: "thinking"},
+		// Summary visibility must not override Claude's per-model thinking default.
+		// Sonnet 4.6 defaults off; newer default-on models remain default-on without
+		// CPA injecting an explicit thinking block.
+		{name: "Responses null summary alone preserves Claude thinking default", from: sdktranslator.FormatOpenAIResponse, to: sdktranslator.FormatClaude, model: "claude-sonnet-4-6-model", body: `{"model":"claude-sonnet-4-6-model","reasoning":{"summary":null},"input":"hi"}`, path: "thinking"},
 		{name: "Responses default keeps Claude display default", from: sdktranslator.FormatOpenAIResponse, to: sdktranslator.FormatClaude, model: "claude-sonnet-4-6-model", body: `{"model":"claude-sonnet-4-6-model","input":"hi"}`, path: "thinking.display"},
 		{name: "Chat summary alias only activates valid Claude thinking", from: sdktranslator.FormatOpenAI, to: sdktranslator.FormatClaude, model: "claude-sonnet-4-6-model", body: `{"model":"claude-sonnet-4-6-model","reasoning":{"exclude":false},"messages":[{"role":"user","content":"hi"}]}`, path: "thinking.display", want: "summarized", wantExists: true},
 		{name: "Interactions summary only activates valid Claude thinking", from: sdktranslator.FormatInteractions, to: sdktranslator.FormatClaude, model: "claude-sonnet-4-6-model", body: `{"model":"claude-sonnet-4-6-model","generation_config":{"thinking_summaries":"auto"},"input":"hi"}`, path: "thinking.display", want: "summarized", wantExists: true},
-- 
2.51.2


From 5d307c195dd23c533f9c5ca9f59216e192101eab Mon Sep 17 00:00:00 2001
From: sususu <suchangshan@foxmail.com>
Date: Thu, 30 Jul 2026 23:26:20 +0800
Subject: [PATCH 5/9] fix(thinking): close summary translation gaps

---
 .../runtime/executor/aistudio_executor.go     |  2 +-
 .../executor/aistudio_executor_test.go        | 20 +++++
 internal/thinking/summary.go                  | 18 ++++-
 internal/thinking/summary_test.go             |  8 ++
 .../claude/gemini/claude_gemini_request.go    |  4 -
 sdk/translator/registry.go                    | 38 ++++++----
 sdk/translator/registry_summary_test.go       | 73 +++++++++++++++++++
 test/summary_intent_translation_test.go       | 38 ++++++++++
 8 files changed, 179 insertions(+), 22 deletions(-)

diff --git a/internal/runtime/executor/aistudio_executor.go b/internal/runtime/executor/aistudio_executor.go
index d2e78eca..3cabe5da 100644
--- a/internal/runtime/executor/aistudio_executor.go
+++ b/internal/runtime/executor/aistudio_executor.go
@@ -461,7 +461,7 @@ func (e *AIStudioExecutor) translateRequest(ctx context.Context, req cliproxyexe
 	originalPayload := originalPayloadSource
 	originalTranslated := helps.TranslateRequestWithCodexMultiAgentV2(ctx, opts.Headers, e.cfg, from, to, baseModel, originalPayload, stream)
 	payload := helps.TranslateRequestWithCodexMultiAgentV2(ctx, opts.Headers, e.cfg, from, to, baseModel, req.Payload, stream)
-	payload, err := helps.ApplyThinkingWithSourcePayload(payload, req.Payload, req.Model, from.String(), to.String(), e.Identifier())
+	payload, err := helps.ApplyThinkingWithSourcePayload(payload, originalPayloadSource, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
 		return nil, translatedPayload{}, err
 	}
diff --git a/internal/runtime/executor/aistudio_executor_test.go b/internal/runtime/executor/aistudio_executor_test.go
index 52ce6147..ea5bd8df 100644
--- a/internal/runtime/executor/aistudio_executor_test.go
+++ b/internal/runtime/executor/aistudio_executor_test.go
@@ -17,8 +17,28 @@ import (
 	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v7/sdk/cliproxy/executor"
 	"github.com/router-for-me/CLIProxyAPI/v7/sdk/cliproxy/usage"
 	sdktranslator "github.com/router-for-me/CLIProxyAPI/v7/sdk/translator"
+	"github.com/tidwall/gjson"
 )
 
+func TestAIStudioTranslateRequestPreservesSummaryFromOriginalRequest(t *testing.T) {
+	executor := NewAIStudioExecutor(&config.Config{}, "aistudio", nil)
+	req := cliproxyexecutor.Request{
+		Model:   "gemini-3.6-flash",
+		Payload: []byte(`{"model":"gemini-3.6-flash","input":"hi"}`),
+	}
+	opts := cliproxyexecutor.Options{
+		SourceFormat:    sdktranslator.FormatOpenAIResponse,
+		OriginalRequest: []byte(`{"model":"gemini-3.6-flash","reasoning":{"summary":"auto"},"input":"hi"}`),
+	}
+	payload, _, err := executor.translateRequest(context.Background(), req, opts, false)
+	if err != nil {
+		t.Fatalf("translateRequest() error = %v", err)
+	}
+	if !gjson.GetBytes(payload, "generationConfig.thinkingConfig.includeThoughts").Bool() {
+		t.Fatalf("original request summary intent was lost: %s", payload)
+	}
+}
+
 func TestAIStudioExecutorExecuteStartsTTFTBeforeRelayWait(t *testing.T) {
 	const authID = "aistudio-ttft-auth"
 	delay := 40 * time.Millisecond
diff --git a/internal/thinking/summary.go b/internal/thinking/summary.go
index 5cca0a9b..5f7d0ead 100644
--- a/internal/thinking/summary.go
+++ b/internal/thinking/summary.go
@@ -95,6 +95,14 @@ func ExtractSummaryConfig(body []byte, format string) SummaryConfig {
 				return config
 			}
 		}
+		if config, ok := firstSummaryBoolConfig(body, []string{
+			"generation_config.thinking_config.include_thoughts",
+			"generation_config.thinking_config.includeThoughts",
+			"generation_config.thinkingConfig.include_thoughts",
+			"generation_config.thinkingConfig.includeThoughts",
+		}); ok {
+			return config
+		}
 	}
 
 	return SummaryConfig{}
@@ -213,10 +221,14 @@ func claudeThinkingAcceptsDisplay(body []byte) bool {
 		return true
 	case "enabled":
 		// This runs before ApplyThinking normalizes the request, so a missing
-		// budget_tokens is an unfinished body rather than inactive thinking.
-		// Only an explicit non-positive budget means thinking is off.
+		// budget_tokens is an unfinished body rather than inactive thinking. CPA
+		// also accepts -1 as its compatibility representation for auto thinking.
 		budget := gjson.GetBytes(body, "thinking.budget_tokens")
-		return budget.Type != gjson.Number || budget.Int() > 0
+		if budget.Type != gjson.Number {
+			return true
+		}
+		value := budget.Int()
+		return value == -1 || value > 0
 	default:
 		return false
 	}
diff --git a/internal/thinking/summary_test.go b/internal/thinking/summary_test.go
index 1fed4147..da7d7bb2 100644
--- a/internal/thinking/summary_test.go
+++ b/internal/thinking/summary_test.go
@@ -44,11 +44,19 @@ func TestExtractSummaryConfig(t *testing.T) {
 		// absent budget must not be read as inactive thinking.
 		{name: "claude enabled display without budget is valid", format: "claude", body: `{"thinking":{"type":"enabled","display":"summarized"}}`, wantMode: SummaryEnabled, wantDetail: "auto"},
 		{name: "claude enabled display with zero budget is invalid", format: "claude", body: `{"thinking":{"type":"enabled","budget_tokens":0,"display":"summarized"}}`, wantMode: SummaryUnspecified},
+		{name: "claude auto compatibility budget summarized", format: "claude", body: `{"thinking":{"type":"enabled","budget_tokens":-1,"display":"summarized"}}`, wantMode: SummaryEnabled, wantDetail: "auto"},
+		{name: "claude auto compatibility budget omitted", format: "claude", body: `{"thinking":{"type":"enabled","budget_tokens":-1,"display":"omitted"}}`, wantMode: SummaryDisabled},
 		{name: "gemini include true", format: "gemini", body: `{"generationConfig":{"thinkingConfig":{"includeThoughts":true}}}`, wantMode: SummaryEnabled, wantDetail: "auto"},
 		{name: "gemini include false", format: "gemini", body: `{"generationConfig":{"thinkingConfig":{"includeThoughts":false}}}`, wantMode: SummaryDisabled},
 		{name: "antigravity include true", format: "antigravity", body: `{"request":{"generationConfig":{"thinkingConfig":{"includeThoughts":true}}}}`, wantMode: SummaryEnabled, wantDetail: "auto"},
 		{name: "interactions auto", format: "interactions", body: `{"generation_config":{"thinking_summaries":"auto"}}`, wantMode: SummaryEnabled, wantDetail: "auto"},
 		{name: "interactions none", format: "interactions", body: `{"generation_config":{"thinking_summaries":"none"}}`, wantMode: SummaryDisabled},
+		{name: "interactions nested snake include false", format: "interactions", body: `{"generation_config":{"thinking_config":{"include_thoughts":false}}}`, wantMode: SummaryDisabled},
+		{name: "interactions nested camel include true", format: "interactions", body: `{"generation_config":{"thinking_config":{"includeThoughts":true}}}`, wantMode: SummaryEnabled, wantDetail: "auto"},
+		{name: "interactions camel config snake include true", format: "interactions", body: `{"generation_config":{"thinkingConfig":{"include_thoughts":true}}}`, wantMode: SummaryEnabled, wantDetail: "auto"},
+		{name: "interactions camel config camel include false", format: "interactions", body: `{"generation_config":{"thinkingConfig":{"includeThoughts":false}}}`, wantMode: SummaryDisabled},
+		{name: "interactions enum wins over include alias", format: "interactions", body: `{"generation_config":{"thinking_summaries":"none","thinking_config":{"include_thoughts":true}}}`, wantMode: SummaryDisabled},
+		{name: "interactions string include alias is invalid", format: "interactions", body: `{"generation_config":{"thinking_config":{"include_thoughts":"false"}}}`, wantMode: SummaryUnspecified},
 		{name: "interactions detailed is invalid", format: "interactions", body: `{"generation_config":{"thinking_summaries":"detailed"}}`, wantMode: SummaryUnspecified},
 		{name: "interactions boolean is invalid", format: "interactions", body: `{"generation_config":{"thinking_summaries":true}}`, wantMode: SummaryUnspecified},
 		{name: "gemini string bool is invalid", format: "gemini", body: `{"generationConfig":{"thinkingConfig":{"includeThoughts":"true"}}}`, wantMode: SummaryUnspecified},
diff --git a/internal/translator/claude/gemini/claude_gemini_request.go b/internal/translator/claude/gemini/claude_gemini_request.go
index ccbaa9d0..b7c7bfa8 100644
--- a/internal/translator/claude/gemini/claude_gemini_request.go
+++ b/internal/translator/claude/gemini/claude_gemini_request.go
@@ -217,10 +217,6 @@ func ConvertGeminiRequestToClaude(modelName string, inputRawJSON []byte, stream
 							out, _ = sjson.SetBytes(out, "thinking.budget_tokens", budget)
 						}
 					}
-				} else if includeThoughts := thinkingConfig.Get("includeThoughts"); includeThoughts.Exists() && includeThoughts.Type == gjson.True {
-					out, _ = sjson.SetBytes(out, "thinking.type", "enabled")
-				} else if includeThoughts := thinkingConfig.Get("include_thoughts"); includeThoughts.Exists() && includeThoughts.Type == gjson.True {
-					out, _ = sjson.SetBytes(out, "thinking.type", "enabled")
 				}
 			}
 		}
diff --git a/sdk/translator/registry.go b/sdk/translator/registry.go
index e9ef609f..830d0355 100644
--- a/sdk/translator/registry.go
+++ b/sdk/translator/registry.go
@@ -70,25 +70,35 @@ func (r *Registry) TranslateRequest(from, to Format, model string, rawJSON []byt
 	body := rawJSON
 	if fn != nil {
 		body = fn(model, body, stream)
-	} else {
-		if model != "" && gjson.GetBytes(body, "model").String() != model {
-			if updated, err := sjson.SetBytes(body, "model", model); err != nil {
-				log.Warnf("translator: failed to normalize model in request fallback: %v", err)
-			} else {
-				body = updated
-			}
+		body = thinking.ApplySummaryConfigForModel(body, to.String(), model, summaryConfig)
+		if hooks != nil {
+			// Request normalizers run after native translation and own the final
+			// provider payload, including any summary field they remove.
+			body = hooks.NormalizeRequest(context.Background(), from, to, model, body, stream)
 		}
+		return body
 	}
 
-	if hooks != nil {
-		body = hooks.NormalizeRequest(context.Background(), from, to, model, body, stream)
-		if fn == nil {
-			if translated, ok := hooks.TranslateRequest(context.Background(), from, to, model, body, stream); ok {
-				body = translated
-			}
+	if model != "" && gjson.GetBytes(body, "model").String() != model {
+		if updated, err := sjson.SetBytes(body, "model", model); err != nil {
+			log.Warnf("translator: failed to normalize model in request fallback: %v", err)
+		} else {
+			body = updated
 		}
 	}
-	return thinking.ApplySummaryConfigForModel(body, to.String(), model, summaryConfig)
+	if hooks == nil {
+		// No translation occurred. Preserve the documented fallback shape instead
+		// of mixing target-protocol summary fields into the source payload.
+		return body
+	}
+
+	// Plugin request normalizers canonicalize the source before a plugin request
+	// translator gets a chance to handle a missing native route.
+	body = hooks.NormalizeRequest(context.Background(), from, to, model, body, stream)
+	if translated, ok := hooks.TranslateRequest(context.Background(), from, to, model, body, stream); ok {
+		body = thinking.ApplySummaryConfigForModel(translated, to.String(), model, summaryConfig)
+	}
+	return body
 }
 
 // HasRequestTransformer indicates whether a request translator exists.
diff --git a/sdk/translator/registry_summary_test.go b/sdk/translator/registry_summary_test.go
index 312ee57b..1b77b951 100644
--- a/sdk/translator/registry_summary_test.go
+++ b/sdk/translator/registry_summary_test.go
@@ -1,9 +1,11 @@
 package translator
 
 import (
+	"bytes"
 	"testing"
 
 	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
 )
 
 func TestRegistryTranslateRequestAppliesSummaryIntent(t *testing.T) {
@@ -130,3 +132,74 @@ func TestRegistryTranslateRequestPreservesNativeClaudeMissingDisplay(t *testing.
 		t.Fatalf("native Claude request without display gained one: %s", out)
 	}
 }
+
+func TestRegistryTranslateRequestDoesNotMixSummaryIntoFallback(t *testing.T) {
+	registry := NewRegistry()
+	body := []byte(`{"model":"gemini-3.6-flash","reasoning":{"summary":"auto"},"input":"hi"}`)
+	out := registry.TranslateRequest(FormatOpenAIResponse, FormatGemini, "gemini-3.6-flash", body, false)
+	if !bytes.Equal(out, body) {
+		t.Fatalf("missing translator changed fallback body: got %s, want %s", out, body)
+	}
+	if gjson.GetBytes(out, "generationConfig").Exists() {
+		t.Fatalf("missing translator mixed Gemini fields into Responses body: %s", out)
+	}
+}
+
+func TestRegistryTranslateRequestPluginMissDoesNotMixSummary(t *testing.T) {
+	registry := NewRegistry()
+	hooks := &fakePluginHooks{requestTranslateOK: false}
+	registry.SetPluginHooks(hooks)
+	body := []byte(`{"model":"gemini-3.6-flash","reasoning":{"summary":"auto"},"input":"hi"}`)
+	out := registry.TranslateRequest(FormatOpenAIResponse, FormatGemini, "gemini-3.6-flash", body, false)
+	if !bytes.Equal(out, body) {
+		t.Fatalf("plugin translation miss changed fallback body: got %s, want %s", out, body)
+	}
+	if gjson.GetBytes(out, "generationConfig").Exists() {
+		t.Fatalf("plugin translation miss mixed Gemini fields into Responses body: %s", out)
+	}
+}
+
+func TestRegistryTranslateRequestAppliesSummaryAfterPluginTranslation(t *testing.T) {
+	registry := NewRegistry()
+	hooks := &fakePluginHooks{
+		requestTranslateBody: []byte(`{"generationConfig":{"thinkingConfig":{"thinkingLevel":"high"}}}`),
+		requestTranslateOK:   true,
+	}
+	registry.SetPluginHooks(hooks)
+	out := registry.TranslateRequest(
+		FormatOpenAIResponse,
+		FormatGemini,
+		"gemini-3.6-flash",
+		[]byte(`{"reasoning":{"summary":"auto"},"input":"hi"}`),
+		false,
+	)
+	if !gjson.GetBytes(out, "generationConfig.thinkingConfig.includeThoughts").Bool() {
+		t.Fatalf("plugin-translated request lost canonical summary: %s", out)
+	}
+}
+
+func TestRegistryTranslateRequestNormalizerOwnsFinalSummaryField(t *testing.T) {
+	registry := NewRegistry()
+	registry.Register(FormatOpenAIResponse, FormatGemini, func(_ string, _ []byte, _ bool) []byte {
+		return []byte(`{"generationConfig":{"thinkingConfig":{"thinkingLevel":"high"}}}`)
+	}, ResponseTransform{})
+	hooks := &fakePluginHooks{normalizeRequest: func(body []byte) []byte {
+		if !gjson.GetBytes(body, "generationConfig.thinkingConfig.includeThoughts").Bool() {
+			t.Fatalf("normalizer did not receive canonical enabled summary: %s", body)
+		}
+		out, _ := sjson.DeleteBytes(body, "generationConfig.thinkingConfig.includeThoughts")
+		return out
+	}}
+	registry.SetPluginHooks(hooks)
+
+	out := registry.TranslateRequest(
+		FormatOpenAIResponse,
+		FormatGemini,
+		"gemini-3.6-flash",
+		[]byte(`{"reasoning":{"effort":"high","summary":"auto"},"input":"hi"}`),
+		false,
+	)
+	if gjson.GetBytes(out, "generationConfig.thinkingConfig.includeThoughts").Exists() {
+		t.Fatalf("summary post-processing overrode request normalizer: %s", out)
+	}
+}
diff --git a/test/summary_intent_translation_test.go b/test/summary_intent_translation_test.go
index b53c7ca3..59ad4b50 100644
--- a/test/summary_intent_translation_test.go
+++ b/test/summary_intent_translation_test.go
@@ -62,6 +62,7 @@ func TestSummaryIntentTranslation(t *testing.T) {
 		{name: "Chat effort enables Interactions summary", from: sdktranslator.FormatOpenAI, to: sdktranslator.FormatInteractions, body: `{"model":"gemini-3.6-flash","reasoning_effort":"high","messages":[{"role":"user","content":"hi"}]}`, path: "generation_config.thinking_summaries", want: "auto", wantExists: true},
 		{name: "Responses concise summary maps to Interactions auto", from: sdktranslator.FormatOpenAIResponse, to: sdktranslator.FormatInteractions, body: `{"model":"gemini-3.6-flash","reasoning":{"effort":"high","summary":"concise"},"input":"hi"}`, path: "generation_config.thinking_summaries", want: "auto", wantExists: true},
 		{name: "Native Claude summarized enables Gemini summary", from: sdktranslator.FormatClaude, to: sdktranslator.FormatGemini, body: `{"model":"claude-opus-5","thinking":{"type":"adaptive","display":"summarized"},"messages":[{"role":"user","content":"hi"}]}`, path: "generationConfig.thinkingConfig.includeThoughts", want: "true", wantExists: true},
+		{name: "Claude auto compatibility budget keeps Gemini summary", from: sdktranslator.FormatClaude, to: sdktranslator.FormatGemini, body: `{"model":"gemini-3.6-flash","thinking":{"type":"enabled","budget_tokens":-1,"display":"summarized"},"messages":[{"role":"user","content":"hi"}]}`, path: "generationConfig.thinkingConfig.includeThoughts", want: "true", wantExists: true},
 		{name: "Native Gemini disabled omits Claude summary", from: sdktranslator.FormatGemini, to: sdktranslator.FormatClaude, body: `{"model":"gemini-3.6-flash","generationConfig":{"thinkingConfig":{"thinkingLevel":"high","includeThoughts":false}},"contents":[{"role":"user","parts":[{"text":"hi"}]}]}`, path: "thinking.display", want: "omitted", wantExists: true},
 		{name: "Native Gemini absent summary leaves Claude display absent", from: sdktranslator.FormatGemini, to: sdktranslator.FormatClaude, body: `{"model":"gemini-3.6-flash","generationConfig":{"thinkingConfig":{"thinkingLevel":"high"}},"contents":[{"role":"user","parts":[{"text":"hi"}]}]}`, path: "thinking.display"},
 		{name: "Native Interactions auto enables Gemini summary", from: sdktranslator.FormatInteractions, to: sdktranslator.FormatGemini, body: `{"model":"gemini-3.6-flash","generation_config":{"thinking_level":"high","thinking_summaries":"auto"},"input":"hi"}`, path: "generationConfig.thinkingConfig.includeThoughts", want: "true", wantExists: true},
@@ -173,6 +174,43 @@ func TestSummaryIntentFinalPipeline(t *testing.T) {
 	}
 }
 
+func TestGeminiSummaryOnlyProducesValidClaudeThinking(t *testing.T) {
+	reg := registry.GetGlobalRegistry()
+	uid := fmt.Sprintf("gemini-summary-only-claude-%d", time.Now().UnixNano())
+	reg.RegisterClient(uid, "test", getTestModels())
+	defer reg.UnregisterClient(uid)
+
+	tests := []struct {
+		name       string
+		model      string
+		wantType   string
+		wantBudget int64
+	}{
+		{name: "adaptive model", model: "claude-sonnet-4-6-model", wantType: "adaptive"},
+		{name: "manual model", model: "claude-budget-model", wantType: "enabled", wantBudget: 1024},
+	}
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			body := []byte(`{"model":"` + test.model + `","generationConfig":{"thinkingConfig":{"includeThoughts":true}},"contents":[{"role":"user","parts":[{"text":"hi"}]}]}`)
+			out := sdktranslator.TranslateRequest(sdktranslator.FormatGemini, sdktranslator.FormatClaude, test.model, body, false)
+			if got := gjson.GetBytes(out, "thinking.type").String(); got != test.wantType {
+				t.Fatalf("thinking.type = %q, want %q; body=%s", got, test.wantType, out)
+			}
+			if got := gjson.GetBytes(out, "thinking.display").String(); got != "summarized" {
+				t.Fatalf("thinking.display = %q, want summarized; body=%s", got, out)
+			}
+			budget := gjson.GetBytes(out, "thinking.budget_tokens")
+			if test.wantBudget > 0 {
+				if budget.Int() != test.wantBudget {
+					t.Fatalf("thinking.budget_tokens = %d, want %d; body=%s", budget.Int(), test.wantBudget, out)
+				}
+			} else if budget.Exists() {
+				t.Fatalf("adaptive model retained budget_tokens: %s", out)
+			}
+		})
+	}
+}
+
 func TestNativeClaudeMissingDisplayPreservesSignatureOnlyHistory(t *testing.T) {
 	body := []byte(`{"model":"claude-opus-5","thinking":{"type":"adaptive"},"messages":[{"role":"assistant","content":[{"type":"thinking","thinking":"","signature":"opus-signature"}]},{"role":"user","content":"continue"}]}`)
 	out := sdktranslator.TranslateRequest(sdktranslator.FormatClaude, sdktranslator.FormatClaude, "claude-opus-5", body, true)
-- 
2.51.2


From 87ceaf83bb702ab4053623ffad78205df7f37d08 Mon Sep 17 00:00:00 2001
From: sususu <suchangshan@foxmail.com>
Date: Fri, 31 Jul 2026 00:05:59 +0800
Subject: [PATCH 6/9] fix(thinking): honor provider visibility semantics

---
 internal/thinking/apply.go                    | 25 ++++--
 .../thinking/apply_configured_api_key_test.go | 55 ++++++++++++
 .../thinking/provider/antigravity/apply.go    |  5 +-
 internal/thinking/provider/gemini/apply.go    |  5 +-
 internal/thinking/summary.go                  | 88 +++++++++----------
 internal/thinking/summary_test.go             | 46 +++++++++-
 test/summary_intent_translation_test.go       |  4 +-
 test/thinking_conversion_test.go              | 38 ++++----
 8 files changed, 187 insertions(+), 79 deletions(-)

diff --git a/internal/thinking/apply.go b/internal/thinking/apply.go
index c19369c7..a349f269 100644
--- a/internal/thinking/apply.go
+++ b/internal/thinking/apply.go
@@ -225,7 +225,7 @@ func applyThinking(body, sourceBody []byte, model string, fromFormat string, toF
 	// Unknown models are treated as user-defined so thinking config can still be applied.
 	// The upstream service is responsible for validating the configuration.
 	if IsUserDefinedModel(modelInfo) {
-		return applyUserDefinedModel(body, modelInfo, fromFormat, providerFormat, suffixResult, summaryConfig)
+		return applyUserDefinedModel(body, modelInfo, fromFormat, providerFormat, providerKey, suffixResult, summaryConfig)
 	}
 	if modelInfo.Thinking == nil {
 		config := extractThinkingConfig(body, providerFormat)
@@ -277,7 +277,7 @@ func applyThinking(body, sourceBody []byte, model string, fromFormat string, toF
 			"provider": providerFormat,
 			"model":    modelInfo.ID,
 		}).Debug("thinking: no config found, passthrough |")
-		return applySummaryConfigForModel(body, providerFormat, baseModel, modelInfo, summaryConfig), nil
+		return applySummaryConfigForProvider(body, providerFormat, baseModel, providerKey, modelInfo, summaryConfig), nil
 	}
 	if modelInfoResolved && config.Mode == ModeLevel && modelInfo != nil && modelInfo.Thinking != nil && shouldMapConfiguredHighIntent(fromFormat, providerFormat, modelInfo) {
 		config.Level = mapConfiguredHighIntent(config.Level, modelInfo)
@@ -320,7 +320,17 @@ func applyThinking(body, sourceBody []byte, model string, fromFormat string, toF
 	if err != nil {
 		return applied, err
 	}
-	return applySummaryConfigForModel(applied, providerFormat, baseModel, modelInfo, summaryConfig), nil
+	// A fully disabled amount takes precedence over visibility. Re-applying a
+	// summary-only field can recreate an otherwise removed provider config and
+	// make a default-on model think again.
+	if thinkingIsFullyDisabled(*validated) {
+		return applied, nil
+	}
+	return applySummaryConfigForProvider(applied, providerFormat, baseModel, providerKey, modelInfo, summaryConfig), nil
+}
+
+func thinkingIsFullyDisabled(config ThinkingConfig) bool {
+	return config.Mode == ModeNone && config.Budget == 0 && config.Level == ""
 }
 
 func shouldMapConfiguredHighIntent(fromFormat, toFormat string, modelInfo *registry.ModelInfo) bool {
@@ -409,7 +419,7 @@ func parseSuffixToConfig(rawSuffix, provider, model string) ThinkingConfig {
 
 // applyUserDefinedModel applies thinking configuration for user-defined models
 // without ThinkingSupport validation.
-func applyUserDefinedModel(body []byte, modelInfo *registry.ModelInfo, fromFormat, toFormat string, suffixResult SuffixResult, summaryConfig SummaryConfig) ([]byte, error) {
+func applyUserDefinedModel(body []byte, modelInfo *registry.ModelInfo, fromFormat, toFormat, providerKey string, suffixResult SuffixResult, summaryConfig SummaryConfig) ([]byte, error) {
 	// Get model ID for logging
 	modelID := ""
 	if modelInfo != nil {
@@ -450,7 +460,7 @@ func applyUserDefinedModel(body []byte, modelInfo *registry.ModelInfo, fromForma
 			"model":    modelID,
 			"provider": toFormat,
 		}).Debug("thinking: user-defined model, passthrough (no config) |")
-		return applySummaryConfigForModel(body, toFormat, modelID, modelInfo, summaryConfig), nil
+		return applySummaryConfigForProvider(body, toFormat, modelID, providerKey, modelInfo, summaryConfig), nil
 	}
 
 	applier := GetProviderApplier(toFormat)
@@ -474,7 +484,10 @@ func applyUserDefinedModel(body []byte, modelInfo *registry.ModelInfo, fromForma
 	if err != nil {
 		return applied, err
 	}
-	return applySummaryConfigForModel(applied, toFormat, modelID, modelInfo, summaryConfig), nil
+	if thinkingIsFullyDisabled(config) {
+		return applied, nil
+	}
+	return applySummaryConfigForProvider(applied, toFormat, modelID, providerKey, modelInfo, summaryConfig), nil
 }
 
 func normalizeUserDefinedConfig(config ThinkingConfig, fromFormat, toFormat string) ThinkingConfig {
diff --git a/internal/thinking/apply_configured_api_key_test.go b/internal/thinking/apply_configured_api_key_test.go
index a443491e..b056139e 100644
--- a/internal/thinking/apply_configured_api_key_test.go
+++ b/internal/thinking/apply_configured_api_key_test.go
@@ -133,6 +133,61 @@ func TestApplyThinkingWithModelInfoDoesNotActivateClaudeForDisabledSummary(t *te
 	}
 }
 
+func TestApplyThinkingWithModelInfoSummaryOnlyDoesNotInventOpenAIEffort(t *testing.T) {
+	modelInfo := &registry.ModelInfo{
+		ID:       "private-openai",
+		Type:     "openai",
+		Thinking: &registry.ThinkingSupport{Levels: []string{"high", "max"}},
+	}
+	out, err := thinking.ApplyThinkingWithModelInfo(
+		[]byte(`{"model":"private-openai","messages":[{"role":"user","content":"hi"}]}`),
+		[]byte(`{"model":"private-openai","reasoning":{"summary":"auto"},"input":"hi"}`),
+		"private-openai", "openai-response", "openai", "openai", modelInfo,
+	)
+	if err != nil {
+		t.Fatalf("ApplyThinkingWithModelInfo() error = %v; body=%s", err, out)
+	}
+	if gjson.GetBytes(out, "reasoning_effort").Exists() {
+		t.Fatalf("summary-only request invented reasoning_effort: %s", out)
+	}
+}
+
+func TestApplyThinkingWithSummaryKeepsOpenAIChatSuffixNone(t *testing.T) {
+	out, err := thinking.ApplyThinkingWithSummary(
+		[]byte(`{"model":"private-openai","messages":[{"role":"user","content":"hi"}]}`),
+		"private-openai(none)", "openai-response", "openai", "openai",
+		thinking.SummaryConfig{Mode: thinking.SummaryEnabled, Detail: "auto"},
+	)
+	if err != nil {
+		t.Fatalf("ApplyThinkingWithSummary() error = %v; body=%s", err, out)
+	}
+	if got := gjson.GetBytes(out, "reasoning_effort").String(); got != "none" {
+		t.Fatalf("reasoning_effort = %q, want none; body=%s", got, out)
+	}
+}
+
+func TestApplyThinkingWithModelInfoUsesOpenRouterVisibility(t *testing.T) {
+	modelInfo := &registry.ModelInfo{
+		ID:       "openrouter-model",
+		Type:     "openai-compatibility",
+		Thinking: &registry.ThinkingSupport{Levels: []string{"high", "max"}},
+	}
+	out, err := thinking.ApplyThinkingWithModelInfo(
+		[]byte(`{"model":"openrouter-model","messages":[{"role":"user","content":"hi"}]}`),
+		[]byte(`{"model":"openrouter-model","reasoning":{"summary":"auto"},"input":"hi"}`),
+		"openrouter-model", "openai-response", "openai", "openrouter", modelInfo,
+	)
+	if err != nil {
+		t.Fatalf("ApplyThinkingWithModelInfo() error = %v; body=%s", err, out)
+	}
+	if exclude := gjson.GetBytes(out, "reasoning.exclude"); !exclude.Exists() || exclude.Bool() {
+		t.Fatalf("OpenRouter summary visibility not enabled: %s", out)
+	}
+	if gjson.GetBytes(out, "reasoning_effort").Exists() {
+		t.Fatalf("OpenRouter summary visibility invented reasoning_effort: %s", out)
+	}
+}
+
 func TestApplyThinkingWithModelInfoUsesOriginalResponsesEffort(t *testing.T) {
 	modelInfo := &registry.ModelInfo{
 		ID:       "claude-upstream",
diff --git a/internal/thinking/provider/antigravity/apply.go b/internal/thinking/provider/antigravity/apply.go
index 968ee09d..6d2edbfa 100644
--- a/internal/thinking/provider/antigravity/apply.go
+++ b/internal/thinking/provider/antigravity/apply.go
@@ -104,8 +104,11 @@ func (a *Applier) applyLevelFormat(body []byte, config thinking.ThinkingConfig)
 
 	if config.Mode == thinking.ModeNone {
 		if config.Budget == 0 && config.Level == "" {
+			// With the amount fully disabled, visibility is irrelevant. Restoring
+			// includeThoughts alone would recreate thinkingConfig and let a
+			// default-on model think again.
 			result, _ = sjson.DeleteBytes(result, "request.generationConfig.thinkingConfig")
-			return applyAntigravityIncludeThoughts(result, body), nil
+			return result, nil
 		}
 		if config.Level != "" {
 			result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.thinkingLevel", string(config.Level))
diff --git a/internal/thinking/provider/gemini/apply.go b/internal/thinking/provider/gemini/apply.go
index c332e9ef..cc4f071e 100644
--- a/internal/thinking/provider/gemini/apply.go
+++ b/internal/thinking/provider/gemini/apply.go
@@ -128,8 +128,11 @@ func (a *Applier) applyLevelFormat(body []byte, config thinking.ThinkingConfig)
 
 	if config.Mode == thinking.ModeNone {
 		if config.Budget == 0 && config.Level == "" {
+			// With the amount fully disabled, visibility is irrelevant. Restoring
+			// includeThoughts alone would recreate thinkingConfig and let a
+			// default-on model think again.
 			result, _ = sjson.DeleteBytes(result, "generationConfig.thinkingConfig")
-			return applyGeminiIncludeThoughts(result, body), nil
+			return result, nil
 		}
 		if config.Level != "" {
 			result, _ = sjson.SetBytes(result, "generationConfig.thinkingConfig.thinkingLevel", string(config.Level))
diff --git a/internal/thinking/summary.go b/internal/thinking/summary.go
index 5f7d0ead..17951990 100644
--- a/internal/thinking/summary.go
+++ b/internal/thinking/summary.go
@@ -95,6 +95,12 @@ func ExtractSummaryConfig(body []byte, format string) SummaryConfig {
 				return config
 			}
 		}
+		// Existing Interactions translators accept the OpenAI-style top-level
+		// compatibility object. Keep the official generation_config selector
+		// authoritative when both are present.
+		if config, ok := interactionsSummaryConfig(body, "reasoning.summary"); ok {
+			return config
+		}
 		if config, ok := firstSummaryBoolConfig(body, []string{
 			"generation_config.thinking_config.include_thoughts",
 			"generation_config.thinking_config.includeThoughts",
@@ -123,6 +129,12 @@ func ApplySummaryConfigForModel(body []byte, format, model string, config Summar
 // applySummaryConfigForModel uses the resolved model definition when execution
 // selected a configured API-key model whose capability is not globally visible.
 func applySummaryConfigForModel(body []byte, format, model string, modelInfo *registry.ModelInfo, config SummaryConfig) []byte {
+	return applySummaryConfigForProvider(body, format, model, "", modelInfo, config)
+}
+
+// applySummaryConfigForProvider uses the execution provider identity for Chat
+// dialects whose visibility controls are not part of the OpenAI wire format.
+func applySummaryConfigForProvider(body []byte, format, model, provider string, modelInfo *registry.ModelInfo, config SummaryConfig) []byte {
 	normalized := strings.ToLower(strings.TrimSpace(format))
 	if config.Mode == SummaryUnspecified || !summaryFormatSupported(normalized) || len(body) == 0 || !gjson.ValidBytes(body) {
 		return body
@@ -131,7 +143,7 @@ func applySummaryConfigForModel(body []byte, format, model string, modelInfo *re
 	enabled := config.Mode == SummaryEnabled
 	switch normalized {
 	case "openai":
-		body = applyOpenAIChatSummaryConfig(body, model, enabled)
+		body = applyOpenAIChatSummaryConfig(body, provider, enabled)
 	case "claude":
 		// Anthropic documents display as invalid with thinking.type=disabled and
 		// requires it alongside adaptive or enabled thinking. Model defaults differ:
@@ -234,65 +246,45 @@ func claudeThinkingAcceptsDisplay(body []byte) bool {
 	}
 }
 
-// applyOpenAIChatSummaryConfig writes summary visibility intent for the Chat
-// Completions protocol.
+// applyOpenAIChatSummaryConfig writes only documented Chat visibility controls.
 //
-// Four dialects share this protocol and only OpenAI's is authoritative. OpenAI
-// documents no reasoning-visibility field at all (Chat Completions never returns
-// reasoning text) and rejects unknown body parameters, so reasoning_effort is the
-// only field that is always safe to write here. OpenRouter's documented
-// "reason but hide" bits (reasoning.exclude and its legacy include_reasoning
-// alias) are updated only when the body already carries them, which is exactly
-// when the upstream is known to understand them.
-func applyOpenAIChatSummaryConfig(body []byte, model string, enabled bool) []byte {
-	if gjson.GetBytes(body, "reasoning").IsObject() {
+// OpenAI Chat Completions exposes reasoning_effort but no reasoning summary or
+// visibility parameter. DeepSeek and Kimi Chat return reasoning_content while
+// thinking is active, but likewise document no independent hide/show switch.
+// Summary intent must therefore never invent or overwrite thinking effort for
+// those dialects. OpenRouter is the exception: reasoning.exclude is its
+// documented "reason but hide" control, and include_reasoning is its deprecated
+// inverse alias. Unknown OpenAI-compatible providers are handled conservatively
+// by updating those fields only when the payload already carries them.
+//
+// Docs:
+// https://developers.openai.com/api/reference/resources/chat/subresources/completions/methods/create
+// https://openrouter.ai/docs/guides/best-practices/reasoning-tokens
+// https://api-docs.deepseek.com/guides/thinking_mode
+// https://platform.kimi.ai/docs/api/chat
+func applyOpenAIChatSummaryConfig(body []byte, provider string, enabled bool) []byte {
+	if isOpenRouterProvider(provider) || gjson.GetBytes(body, "reasoning.exclude").IsBool() {
 		body, _ = sjson.SetBytes(body, "reasoning.exclude", !enabled)
 	}
 	if gjson.GetBytes(body, "include_reasoning").IsBool() {
 		body, _ = sjson.SetBytes(body, "include_reasoning", enabled)
 	}
-	if !enabled {
-		// Chat has no portable way to keep reasoning while hiding its summary.
-		// reasoning_effort:"none" would disable reasoning instead of hiding it,
-		// and Google documents that it is not even honored on Gemini 2.5 Pro or
-		// 3 models, so leave the effort the client asked for untouched.
-		return body
-	}
-	effort := gjson.GetBytes(body, "reasoning_effort")
-	if effort.Type != gjson.String || strings.TrimSpace(effort.String()) == "" || strings.EqualFold(strings.TrimSpace(effort.String()), "none") {
-		body, _ = sjson.SetBytes(body, "reasoning_effort", openAIChatSummaryEffort(body, model))
-	}
 	return body
 }
 
-// openAIChatSummaryEffort picks an active reasoning effort that the target model
-// documents. Chat exposes reasoning only while an effort is active, so a summary
-// request has to select one when the client left it unset.
-func openAIChatSummaryEffort(body []byte, model string) string {
-	baseModel := ParseSuffix(model).ModelName
-	if baseModel == "" {
-		baseModel = ParseSuffix(gjson.GetBytes(body, "model").String()).ModelName
-	}
-	modelInfo := registry.LookupModelInfo(baseModel, "openai")
-	if modelInfo == nil || modelInfo.Thinking == nil || len(modelInfo.Thinking.Levels) == 0 {
-		return "medium"
+func isOpenRouterProvider(provider string) bool {
+	provider = strings.ToLower(strings.TrimSpace(provider))
+	if provider == "openrouter" {
+		return true
 	}
-
-	levels := make([]string, 0, len(modelInfo.Thinking.Levels))
-	for _, level := range modelInfo.Thinking.Levels {
-		normalized := strings.ToLower(strings.TrimSpace(level))
-		if normalized == "" || normalized == "none" {
-			continue
-		}
-		if normalized == "medium" {
-			return "medium"
+	for _, part := range strings.FieldsFunc(provider, func(r rune) bool {
+		return r == '-' || r == '_' || r == '/' || r == '.' || r == ':'
+	}) {
+		if part == "openrouter" {
+			return true
 		}
-		levels = append(levels, normalized)
-	}
-	if len(levels) == 0 {
-		return "medium"
 	}
-	return levels[len(levels)/2]
+	return false
 }
 
 func extractOpenAIExplicitSummaryConfig(body []byte) (SummaryConfig, bool) {
diff --git a/internal/thinking/summary_test.go b/internal/thinking/summary_test.go
index da7d7bb2..84c110c9 100644
--- a/internal/thinking/summary_test.go
+++ b/internal/thinking/summary_test.go
@@ -55,6 +55,9 @@ func TestExtractSummaryConfig(t *testing.T) {
 		{name: "interactions nested camel include true", format: "interactions", body: `{"generation_config":{"thinking_config":{"includeThoughts":true}}}`, wantMode: SummaryEnabled, wantDetail: "auto"},
 		{name: "interactions camel config snake include true", format: "interactions", body: `{"generation_config":{"thinkingConfig":{"include_thoughts":true}}}`, wantMode: SummaryEnabled, wantDetail: "auto"},
 		{name: "interactions camel config camel include false", format: "interactions", body: `{"generation_config":{"thinkingConfig":{"includeThoughts":false}}}`, wantMode: SummaryDisabled},
+		{name: "interactions enum wins over compatibility reasoning", format: "interactions", body: `{"generation_config":{"thinking_summaries":"none"},"reasoning":{"summary":"auto"}}`, wantMode: SummaryDisabled},
+		{name: "interactions compatibility reasoning auto", format: "interactions", body: `{"reasoning":{"summary":"auto"}}`, wantMode: SummaryEnabled, wantDetail: "auto"},
+		{name: "interactions compatibility reasoning none", format: "interactions", body: `{"reasoning":{"summary":"none"}}`, wantMode: SummaryDisabled},
 		{name: "interactions enum wins over include alias", format: "interactions", body: `{"generation_config":{"thinking_summaries":"none","thinking_config":{"include_thoughts":true}}}`, wantMode: SummaryDisabled},
 		{name: "interactions string include alias is invalid", format: "interactions", body: `{"generation_config":{"thinking_config":{"include_thoughts":"false"}}}`, wantMode: SummaryUnspecified},
 		{name: "interactions detailed is invalid", format: "interactions", body: `{"generation_config":{"thinking_summaries":"detailed"}}`, wantMode: SummaryUnspecified},
@@ -81,8 +84,9 @@ func TestApplySummaryConfig(t *testing.T) {
 		path   string
 		want   string
 	}{
-		{name: "chat enabled creates compatibility effort", format: "openai", config: SummaryConfig{Mode: SummaryEnabled}, path: "reasoning_effort", want: "medium"},
+		{name: "chat enabled invents no effort", format: "openai", config: SummaryConfig{Mode: SummaryEnabled}, path: "reasoning_effort", want: ""},
 		{name: "chat enabled preserves active effort", format: "openai", body: `{"reasoning_effort":"high"}`, config: SummaryConfig{Mode: SummaryEnabled}, path: "reasoning_effort", want: "high"},
+		{name: "chat enabled preserves disabled effort", format: "openai", body: `{"reasoning_effort":"none"}`, config: SummaryConfig{Mode: SummaryEnabled}, path: "reasoning_effort", want: "none"},
 		// Chat cannot express "reason but hide", so disabling must not fall back to
 		// reasoning_effort:"none", which would disable reasoning altogether.
 		{name: "chat disabled preserves requested effort", format: "openai", body: `{"reasoning_effort":"high"}`, config: SummaryConfig{Mode: SummaryDisabled}, path: "reasoning_effort", want: "high"},
@@ -114,6 +118,46 @@ func TestApplySummaryConfig(t *testing.T) {
 	}
 }
 
+func TestApplySummaryConfig_OpenAIChatProviderDialects(t *testing.T) {
+	tests := []struct {
+		name         string
+		provider     string
+		body         string
+		mode         SummaryMode
+		wantExclude  string
+		wantExisting bool
+		wantEffort   string
+	}{
+		{name: "OpenAI does not invent visibility", provider: "openai", body: `{}`, mode: SummaryEnabled},
+		{name: "OpenRouter enables visibility", provider: "openrouter", body: `{}`, mode: SummaryEnabled, wantExclude: "false", wantExisting: true},
+		{name: "OpenRouter disables visibility", provider: "prod-openrouter", body: `{}`, mode: SummaryDisabled, wantExclude: "true", wantExisting: true},
+		{name: "DeepSeek preserves documented effort", provider: "deepseek", body: `{"reasoning_effort":"high"}`, mode: SummaryDisabled, wantEffort: "high"},
+		{name: "Kimi preserves documented K3 effort", provider: "kimi", body: `{"reasoning_effort":"max"}`, mode: SummaryEnabled, wantEffort: "max"},
+		{name: "Moonshot does not invent visibility", provider: "moonshot", body: `{"thinking":{"type":"enabled"}}`, mode: SummaryEnabled},
+		{name: "generic provider updates existing OpenRouter field", provider: "openai-compatibility", body: `{"reasoning":{"exclude":false}}`, mode: SummaryDisabled, wantExclude: "true", wantExisting: true},
+	}
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			out := applySummaryConfigForProvider([]byte(test.body), "openai", "model", test.provider, nil, SummaryConfig{Mode: test.mode})
+			exclude := gjson.GetBytes(out, "reasoning.exclude")
+			if exclude.Exists() != test.wantExisting {
+				t.Fatalf("reasoning.exclude exists = %v, want %v; body=%s", exclude.Exists(), test.wantExisting, out)
+			}
+			if test.wantExisting && exclude.String() != test.wantExclude {
+				t.Fatalf("reasoning.exclude = %q, want %q; body=%s", exclude.String(), test.wantExclude, out)
+			}
+			effort := gjson.GetBytes(out, "reasoning_effort")
+			if test.wantEffort == "" {
+				if effort.Exists() {
+					t.Fatalf("summary visibility invented reasoning_effort: %s", out)
+				}
+			} else if effort.String() != test.wantEffort {
+				t.Fatalf("reasoning_effort = %q, want %q; body=%s", effort.String(), test.wantEffort, out)
+			}
+		})
+	}
+}
+
 func TestApplySummaryConfigNormalizesTargetAliases(t *testing.T) {
 	tests := []struct {
 		format    string
diff --git a/test/summary_intent_translation_test.go b/test/summary_intent_translation_test.go
index 59ad4b50..cc1724f5 100644
--- a/test/summary_intent_translation_test.go
+++ b/test/summary_intent_translation_test.go
@@ -40,7 +40,7 @@ func TestSummaryIntentTranslation(t *testing.T) {
 		{name: "Claude summarized enables Codex summary", from: sdktranslator.FormatClaude, to: sdktranslator.FormatCodex, body: `{"model":"gpt-5.4","max_tokens":1024,"thinking":{"type":"adaptive","display":"summarized"},"output_config":{"effort":"high"},"messages":[{"role":"user","content":"hi"}]}`, path: "reasoning.summary", want: "auto", wantExists: true},
 		{name: "Interactions none omits Codex summary", from: sdktranslator.FormatInteractions, to: sdktranslator.FormatCodex, body: `{"model":"gpt-5.4","generation_config":{"thinking_level":"high","thinking_summaries":"none"},"input":"hi"}`, path: "reasoning.summary"},
 		{name: "Chat effort enables Codex summary", from: sdktranslator.FormatOpenAI, to: sdktranslator.FormatCodex, body: `{"model":"gpt-5.4","reasoning_effort":"high","messages":[{"role":"user","content":"hi"}]}`, path: "reasoning.summary", want: "auto", wantExists: true},
-		{name: "Responses summary only enables Chat compatibility effort", from: sdktranslator.FormatOpenAIResponse, to: sdktranslator.FormatOpenAI, body: `{"model":"gpt-5.4","reasoning":{"summary":"auto"},"input":"hi"}`, path: "reasoning_effort", want: "medium", wantExists: true},
+		{name: "Responses summary only invents no Chat effort", from: sdktranslator.FormatOpenAIResponse, to: sdktranslator.FormatOpenAI, body: `{"model":"gpt-5.4","reasoning":{"summary":"auto"},"input":"hi"}`, path: "reasoning_effort"},
 		// Chat has no field for "reason but hide": OpenAI documents none and rejects
 		// unknown parameters, so a disabled summary must leave the requested effort
 		// alone instead of turning reasoning off upstream.
@@ -127,10 +127,12 @@ func TestSummaryIntentFinalPipeline(t *testing.T) {
 		{name: "Responses default keeps Claude display default", from: sdktranslator.FormatOpenAIResponse, to: sdktranslator.FormatClaude, model: "claude-sonnet-4-6-model", body: `{"model":"claude-sonnet-4-6-model","input":"hi"}`, path: "thinking.display"},
 		{name: "Chat summary alias only activates valid Claude thinking", from: sdktranslator.FormatOpenAI, to: sdktranslator.FormatClaude, model: "claude-sonnet-4-6-model", body: `{"model":"claude-sonnet-4-6-model","reasoning":{"exclude":false},"messages":[{"role":"user","content":"hi"}]}`, path: "thinking.display", want: "summarized", wantExists: true},
 		{name: "Interactions summary only activates valid Claude thinking", from: sdktranslator.FormatInteractions, to: sdktranslator.FormatClaude, model: "claude-sonnet-4-6-model", body: `{"model":"claude-sonnet-4-6-model","generation_config":{"thinking_summaries":"auto"},"input":"hi"}`, path: "thinking.display", want: "summarized", wantExists: true},
+		{name: "Interactions compatibility summary activates valid Claude thinking", from: sdktranslator.FormatInteractions, to: sdktranslator.FormatClaude, model: "claude-sonnet-4-6-model", body: `{"model":"claude-sonnet-4-6-model","reasoning":{"summary":"auto"},"input":"hi"}`, path: "thinking.display", want: "summarized", wantExists: true},
 		{name: "Claude suffix none removes otherwise enabled display", from: sdktranslator.FormatOpenAIResponse, to: sdktranslator.FormatClaude, model: "claude-sonnet-4-6-model(none)", body: `{"model":"claude-sonnet-4-6-model(none)","reasoning":{"summary":"auto"},"input":"hi"}`, path: "thinking.display"},
 		{name: "Claude suffix preserves explicit disabled summary", from: sdktranslator.FormatOpenAIResponse, to: sdktranslator.FormatClaude, model: "claude-sonnet-4-6-model(high)", body: `{"model":"claude-sonnet-4-6-model(high)","reasoning":{"summary":null},"input":"hi"}`, path: "thinking.display", want: "omitted", wantExists: true},
 		{name: "Responses effort alone stays omitted on Antigravity", from: sdktranslator.FormatOpenAIResponse, to: sdktranslator.FormatAntigravity, model: "antigravity-budget-model", body: `{"model":"antigravity-budget-model","reasoning":{"effort":"medium"},"input":"hi"}`, path: "request.generationConfig.thinkingConfig.includeThoughts"},
 		{name: "Responses summary reaches Antigravity", from: sdktranslator.FormatOpenAIResponse, to: sdktranslator.FormatAntigravity, model: "antigravity-budget-model", body: `{"model":"antigravity-budget-model","reasoning":{"effort":"medium","summary":"auto"},"input":"hi"}`, path: "request.generationConfig.thinkingConfig.includeThoughts", want: "true", wantExists: true},
+		{name: "Responses null summary alone hides default Gemini thoughts", from: sdktranslator.FormatOpenAIResponse, to: sdktranslator.FormatGemini, model: "gemini-mixed-model", body: `{"model":"gemini-mixed-model","reasoning":{"summary":null},"input":"hi"}`, path: "generationConfig.thinkingConfig.includeThoughts", want: "false", wantExists: true},
 		{name: "Google Chat extension false survives Gemini applier", from: sdktranslator.FormatOpenAI, to: sdktranslator.FormatGemini, model: "gemini-mixed-model", body: `{"model":"gemini-mixed-model","reasoning_effort":"high","extra_body":{"google":{"thinking_config":{"include_thoughts":false}}},"messages":[{"role":"user","content":"hi"}]}`, path: "generationConfig.thinkingConfig.includeThoughts", want: "false", wantExists: true},
 		// Captured from isolated Claude Code 2.1.220 with
 		// alwaysThinkingEnabled:true. Sonnet uses adaptive thinking, while Haiku
diff --git a/test/thinking_conversion_test.go b/test/thinking_conversion_test.go
index 07dfb039..d71d6e35 100644
--- a/test/thinking_conversion_test.go
+++ b/test/thinking_conversion_test.go
@@ -1456,30 +1456,26 @@ func TestThinkingE2EMatrix_Body(t *testing.T) {
 			includeThoughts: "false",
 			expectErr:       false,
 		},
-		// Case 31A: reasoning_effort=none with zero allowed removes the amount but
-		// preserves Chat's explicit disabled summary intent.
+		// Case 31A: reasoning_effort=none with zero allowed removes the entire
+		// thinking config. includeThoughts alone would restore the model default.
 		{
-			name:            "31A",
-			from:            "openai",
-			to:              "gemini",
-			model:           "gemini-toggle-mixed-model",
-			inputJSON:       `{"model":"gemini-toggle-mixed-model","messages":[{"role":"user","content":"hi"}],"reasoning_effort":"none"}`,
-			expectField:     "generationConfig.thinkingConfig.includeThoughts",
-			expectValue:     "false",
-			includeThoughts: "false",
-			expectErr:       false,
+			name:        "31A",
+			from:        "openai",
+			to:          "gemini",
+			model:       "gemini-toggle-mixed-model",
+			inputJSON:   `{"model":"gemini-toggle-mixed-model","messages":[{"role":"user","content":"hi"}],"reasoning_effort":"none"}`,
+			expectField: "",
+			expectErr:   false,
 		},
-		// Case 31B: the same explicit disabled intent survives Antigravity.
+		// Case 31B: Antigravity keeps the same fully disabled representation.
 		{
-			name:            "31B",
-			from:            "openai",
-			to:              "antigravity",
-			model:           "gemini-toggle-mixed-model",
-			inputJSON:       `{"model":"gemini-toggle-mixed-model","messages":[{"role":"user","content":"hi"}],"reasoning_effort":"none"}`,
-			expectField:     "request.generationConfig.thinkingConfig.includeThoughts",
-			expectValue:     "false",
-			includeThoughts: "false",
-			expectErr:       false,
+			name:        "31B",
+			from:        "openai",
+			to:          "antigravity",
+			model:       "gemini-toggle-mixed-model",
+			inputJSON:   `{"model":"gemini-toggle-mixed-model","messages":[{"role":"user","content":"hi"}],"reasoning_effort":"none"}`,
+			expectField: "",
+			expectErr:   false,
 		},
 		// Case 31C: reasoning.effort=none with zero allowed → delete thinkingConfig
 		{
-- 
2.51.2


From 0c2ec7da235d4f5b1bcd9ab1c03ec77bccfb93f3 Mon Sep 17 00:00:00 2001
From: sususu <suchangshan@foxmail.com>
Date: Fri, 31 Jul 2026 12:45:13 +0800
Subject: [PATCH 7/9] fix(thinking): honor normalized summary payloads

---
 .../runtime/executor/aistudio_executor.go     |   2 +-
 .../executor/antigravity_executor_execute.go  |   4 +-
 .../executor/antigravity_executor_stream.go   |   2 +-
 .../executor/antigravity_executor_tokens.go   |   2 +-
 .../runtime/executor/codex_openai_images.go   |   2 +-
 .../executor/helps/model_capabilities.go      |  10 +-
 internal/runtime/executor/helps/thinking.go   |  62 ++++++-
 .../runtime/executor/helps/thinking_test.go   | 100 ++++++++++++
 internal/runtime/executor/kimi_executor.go    |   4 +-
 internal/thinking/apply.go                    |   7 +
 internal/thinking/summary.go                  |  16 ++
 internal/thinking/summary_test.go             |  12 ++
 test/thinking_conversion_test.go              | 153 +++++++++---------
 13 files changed, 277 insertions(+), 99 deletions(-)
 create mode 100644 internal/runtime/executor/helps/thinking_test.go

diff --git a/internal/runtime/executor/aistudio_executor.go b/internal/runtime/executor/aistudio_executor.go
index 3cabe5da..3cc37a09 100644
--- a/internal/runtime/executor/aistudio_executor.go
+++ b/internal/runtime/executor/aistudio_executor.go
@@ -461,7 +461,7 @@ func (e *AIStudioExecutor) translateRequest(ctx context.Context, req cliproxyexe
 	originalPayload := originalPayloadSource
 	originalTranslated := helps.TranslateRequestWithCodexMultiAgentV2(ctx, opts.Headers, e.cfg, from, to, baseModel, originalPayload, stream)
 	payload := helps.TranslateRequestWithCodexMultiAgentV2(ctx, opts.Headers, e.cfg, from, to, baseModel, req.Payload, stream)
-	payload, err := helps.ApplyThinkingWithSourcePayload(payload, originalPayloadSource, req.Model, from.String(), to.String(), e.Identifier())
+	payload, err := helps.ApplyThinkingWithSourcePayload(payload, req.Payload, originalPayloadSource, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
 		return nil, translatedPayload{}, err
 	}
diff --git a/internal/runtime/executor/antigravity_executor_execute.go b/internal/runtime/executor/antigravity_executor_execute.go
index 77bce648..6721bb01 100644
--- a/internal/runtime/executor/antigravity_executor_execute.go
+++ b/internal/runtime/executor/antigravity_executor_execute.go
@@ -68,7 +68,7 @@ func (e *AntigravityExecutor) Execute(ctx context.Context, auth *cliproxyauth.Au
 	originalTranslated := helps.TranslateRequestWithCodexMultiAgentV2(ctx, opts.Headers, e.cfg, from, to, baseModel, originalPayload, false)
 	translated := helps.TranslateRequestWithCodexMultiAgentV2(ctx, opts.Headers, e.cfg, from, to, baseModel, req.Payload, false)
 
-	translated, err = helps.ApplyThinkingWithSourcePayload(translated, req.Payload, req.Model, from.String(), to.String(), e.Identifier())
+	translated, err = helps.ApplyThinkingWithSourcePayload(translated, req.Payload, originalPayloadSource, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
 		return resp, err
 	}
@@ -290,7 +290,7 @@ func (e *AntigravityExecutor) executeClaudeNonStream(ctx context.Context, auth *
 	originalTranslated := helps.TranslateRequestWithCodexMultiAgentV2(ctx, opts.Headers, e.cfg, from, to, baseModel, originalPayload, true)
 	translated := helps.TranslateRequestWithCodexMultiAgentV2(ctx, opts.Headers, e.cfg, from, to, baseModel, req.Payload, true)
 
-	translated, err = helps.ApplyThinkingWithSourcePayload(translated, req.Payload, req.Model, from.String(), to.String(), e.Identifier())
+	translated, err = helps.ApplyThinkingWithSourcePayload(translated, req.Payload, originalPayloadSource, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
 		return resp, err
 	}
diff --git a/internal/runtime/executor/antigravity_executor_stream.go b/internal/runtime/executor/antigravity_executor_stream.go
index d0aa0725..98c7177c 100644
--- a/internal/runtime/executor/antigravity_executor_stream.go
+++ b/internal/runtime/executor/antigravity_executor_stream.go
@@ -63,7 +63,7 @@ func (e *AntigravityExecutor) ExecuteStream(ctx context.Context, auth *cliproxya
 	originalTranslated := helps.TranslateRequestWithCodexMultiAgentV2(ctx, opts.Headers, e.cfg, from, to, baseModel, originalPayload, true)
 	translated := helps.TranslateRequestWithCodexMultiAgentV2(ctx, opts.Headers, e.cfg, from, to, baseModel, req.Payload, true)
 
-	translated, err = helps.ApplyThinkingWithSourcePayload(translated, req.Payload, req.Model, from.String(), to.String(), e.Identifier())
+	translated, err = helps.ApplyThinkingWithSourcePayload(translated, req.Payload, originalPayloadSource, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
 		return nil, err
 	}
diff --git a/internal/runtime/executor/antigravity_executor_tokens.go b/internal/runtime/executor/antigravity_executor_tokens.go
index 45867ee8..523d7d2c 100644
--- a/internal/runtime/executor/antigravity_executor_tokens.go
+++ b/internal/runtime/executor/antigravity_executor_tokens.go
@@ -50,7 +50,7 @@ func (e *AntigravityExecutor) CountTokens(ctx context.Context, auth *cliproxyaut
 	// Prepare payload once (doesn't depend on baseURL)
 	payload := helps.TranslateRequestWithCodexMultiAgentV2(ctx, opts.Headers, e.cfg, from, to, baseModel, req.Payload, false)
 
-	payload, err := helps.ApplyThinkingWithSourcePayload(payload, req.Payload, req.Model, from.String(), to.String(), e.Identifier())
+	payload, err := helps.ApplyThinkingWithSourcePayload(payload, req.Payload, originalPayloadSource, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
 		return cliproxyexecutor.Response{}, err
 	}
diff --git a/internal/runtime/executor/codex_openai_images.go b/internal/runtime/executor/codex_openai_images.go
index 3251489e..18ef4418 100644
--- a/internal/runtime/executor/codex_openai_images.go
+++ b/internal/runtime/executor/codex_openai_images.go
@@ -674,7 +674,7 @@ func (e *CodexExecutor) prepareCodexOpenAIImageBody(body []byte, req cliproxyexe
 		mainModel = codexOpenAIImagesMainModel
 	}
 	var errThinking error
-	out, errThinking = helps.ApplyThinkingWithSourcePayload(out, body, mainModel, codexOpenAIImageSourceFormat, "codex", e.Identifier())
+	out, errThinking = helps.ApplyThinkingWithSourcePayload(out, body, body, mainModel, codexOpenAIImageSourceFormat, "codex", e.Identifier())
 	if errThinking != nil {
 		return nil, errThinking
 	}
diff --git a/internal/runtime/executor/helps/model_capabilities.go b/internal/runtime/executor/helps/model_capabilities.go
index 8bf6723d..fea97c5d 100644
--- a/internal/runtime/executor/helps/model_capabilities.go
+++ b/internal/runtime/executor/helps/model_capabilities.go
@@ -9,13 +9,13 @@ import (
 // ApplyRequestThinking preserves the registry lookup path unless the auth
 // manager bound an exact configured API-key model definition to this attempt.
 func ApplyRequestThinking(body []byte, req cliproxyexecutor.Request, opts cliproxyexecutor.Options, fromFormat, toFormat, provider string) ([]byte, error) {
-	sourceBody := opts.OriginalRequest
-	if len(sourceBody) == 0 {
-		sourceBody = req.Payload
+	originalSource := opts.OriginalRequest
+	if len(originalSource) == 0 {
+		originalSource = req.Payload
 	}
+	summaryConfig := translatedRequestSummaryConfig(body, req.Payload, originalSource, req.Model, fromFormat, toFormat)
 	if modelInfo, ok := cliproxyauth.ResolvedAPIKeyModelInfo(req); ok {
-		return thinking.ApplyThinkingWithModelInfo(body, sourceBody, req.Model, fromFormat, toFormat, provider, modelInfo)
+		return thinking.ApplyThinkingWithModelInfoAndSummary(body, originalSource, req.Model, fromFormat, toFormat, provider, modelInfo, summaryConfig)
 	}
-	summaryConfig := thinking.ExtractSummaryConfig(sourceBody, fromFormat)
 	return thinking.ApplyThinkingWithSummary(body, req.Model, fromFormat, toFormat, provider, summaryConfig)
 }
diff --git a/internal/runtime/executor/helps/thinking.go b/internal/runtime/executor/helps/thinking.go
index 49f3155c..9ad7a2e6 100644
--- a/internal/runtime/executor/helps/thinking.go
+++ b/internal/runtime/executor/helps/thinking.go
@@ -1,12 +1,64 @@
 package helps
 
-import "github.com/router-for-me/CLIProxyAPI/v7/internal/thinking"
+import (
+	"strings"
+
+	"github.com/router-for-me/CLIProxyAPI/v7/internal/thinking"
+	sdktranslator "github.com/router-for-me/CLIProxyAPI/v7/sdk/translator"
+)
 
 // ApplyThinkingWithSourcePayload preserves summary visibility from the original
 // client payload while applying thinking configuration to its translated target
-// payload. A target representation alone can lose an explicit disabled summary
-// before a model suffix changes Claude thinking from disabled to adaptive.
-func ApplyThinkingWithSourcePayload(body, sourcePayload []byte, model, fromFormat, toFormat, providerKey string) ([]byte, error) {
-	summary := thinking.ExtractSummaryConfig(sourcePayload, fromFormat)
+// payload. currentSourcePayload is the payload that was translated, while
+// originalSourcePayload retains intent removed by an earlier interceptor.
+func ApplyThinkingWithSourcePayload(body, currentSourcePayload, originalSourcePayload []byte, model, fromFormat, toFormat, providerKey string) ([]byte, error) {
+	summary := translatedRequestSummaryConfig(body, currentSourcePayload, originalSourcePayload, model, fromFormat, toFormat)
 	return thinking.ApplyThinkingWithSummary(body, model, fromFormat, toFormat, providerKey, summary)
 }
+
+// translatedRequestSummaryConfig gives the translated target payload precedence
+// so a plugin request normalizer can remove or rewrite a canonical summary field.
+// The original source is consulted only when the payload that was translated no
+// longer carries the inbound intent, or when the target could not represent that
+// intent until model-aware thinking is applied later (notably Claude).
+func translatedRequestSummaryConfig(body, currentSourcePayload, originalSourcePayload []byte, model, fromFormat, toFormat string) thinking.SummaryConfig {
+	fromFormat = strings.ToLower(strings.TrimSpace(fromFormat))
+	toFormat = strings.ToLower(strings.TrimSpace(toFormat))
+
+	var targetSummary thinking.SummaryConfig
+	if fromFormat == toFormat {
+		targetSummary = thinking.ExtractSummaryConfig(body, toFormat)
+	} else {
+		targetSummary = thinking.ExtractExplicitSummaryConfig(body, toFormat)
+	}
+	if targetSummary.Mode != thinking.SummaryUnspecified {
+		return targetSummary
+	}
+
+	currentSummary := thinking.ExtractSummaryConfig(currentSourcePayload, fromFormat)
+	originalSummary := thinking.ExtractSummaryConfig(originalSourcePayload, fromFormat)
+	if currentSummary.Mode == thinking.SummaryUnspecified {
+		return originalSummary
+	}
+
+	from := sdktranslator.FromString(fromFormat)
+	to := sdktranslator.FromString(toFormat)
+	if !sdktranslator.HasRequestTransformer(from, to) {
+		// A missing translation must remain source-shaped. Same-format requests
+		// were handled by targetSummary above, including explicit native aliases.
+		return thinking.SummaryConfig{}
+	}
+
+	candidate := thinking.ApplySummaryConfigForModel(body, toFormat, model, currentSummary)
+	if thinking.ExtractExplicitSummaryConfig(candidate, toFormat).Mode != thinking.SummaryUnspecified {
+		// Registry translation applied this field before plugin normalization. If
+		// it is absent now but can be represented on the normalized body, the
+		// normalizer deliberately removed it and must remain authoritative.
+		return thinking.SummaryConfig{}
+	}
+
+	// Some intents cannot be represented until the final model-aware pass. For
+	// example, Claude display is invalid on disabled thinking, but a suffix can
+	// subsequently activate adaptive thinking. Preserve the source in that case.
+	return currentSummary
+}
diff --git a/internal/runtime/executor/helps/thinking_test.go b/internal/runtime/executor/helps/thinking_test.go
new file mode 100644
index 00000000..69b18fda
--- /dev/null
+++ b/internal/runtime/executor/helps/thinking_test.go
@@ -0,0 +1,100 @@
+package helps_test
+
+import (
+	"context"
+	"testing"
+
+	helps "github.com/router-for-me/CLIProxyAPI/v7/internal/runtime/executor/helps"
+	_ "github.com/router-for-me/CLIProxyAPI/v7/internal/thinking/provider/gemini"
+	_ "github.com/router-for-me/CLIProxyAPI/v7/internal/translator"
+	sdktranslator "github.com/router-for-me/CLIProxyAPI/v7/sdk/translator"
+	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
+)
+
+type summaryRemovingPluginHooks struct {
+	t *testing.T
+}
+
+func (h *summaryRemovingPluginHooks) NormalizeRequest(_ context.Context, _, _ sdktranslator.Format, _ string, body []byte, _ bool) []byte {
+	h.t.Helper()
+	const path = "generationConfig.thinkingConfig.includeThoughts"
+	if !gjson.GetBytes(body, path).Bool() {
+		h.t.Fatalf("request normalizer did not receive enabled summary: %s", body)
+	}
+	out, _ := sjson.DeleteBytes(body, path)
+	return out
+}
+
+func (*summaryRemovingPluginHooks) TranslateRequest(context.Context, sdktranslator.Format, sdktranslator.Format, string, []byte, bool) ([]byte, bool) {
+	return nil, false
+}
+
+func (*summaryRemovingPluginHooks) NormalizeResponseBefore(context.Context, sdktranslator.Format, sdktranslator.Format, string, []byte, []byte, []byte, bool) []byte {
+	return nil
+}
+
+func (*summaryRemovingPluginHooks) TranslateResponse(context.Context, sdktranslator.Format, sdktranslator.Format, string, []byte, []byte, []byte, bool) ([]byte, bool) {
+	return nil, false
+}
+
+func (*summaryRemovingPluginHooks) NormalizeResponseAfter(context.Context, sdktranslator.Format, sdktranslator.Format, string, []byte, []byte, []byte, bool) []byte {
+	return nil
+}
+
+func TestApplyThinkingWithSourcePayloadPreservesNormalizerSummaryRemoval(t *testing.T) {
+	hooks := &summaryRemovingPluginHooks{t: t}
+	sdktranslator.SetPluginHooks(hooks)
+	t.Cleanup(func() { sdktranslator.SetPluginHooks(nil) })
+
+	source := []byte(`{"model":"gemini-3.6-flash","reasoning":{"effort":"high","summary":"auto"},"input":"hi"}`)
+	translated := sdktranslator.TranslateRequest(
+		sdktranslator.FormatOpenAIResponse,
+		sdktranslator.FormatGemini,
+		"gemini-3.6-flash",
+		source,
+		false,
+	)
+	const summaryPath = "generationConfig.thinkingConfig.includeThoughts"
+	if gjson.GetBytes(translated, summaryPath).Exists() {
+		t.Fatalf("request normalizer did not remove summary: %s", translated)
+	}
+
+	out, err := helps.ApplyThinkingWithSourcePayload(
+		translated,
+		source,
+		source,
+		"gemini-3.6-flash",
+		sdktranslator.FormatOpenAIResponse.String(),
+		sdktranslator.FormatGemini.String(),
+		"gemini",
+	)
+	if err != nil {
+		t.Fatalf("ApplyThinkingWithSourcePayload() error = %v", err)
+	}
+	if gjson.GetBytes(out, summaryPath).Exists() {
+		t.Fatalf("executor restored summary removed by request normalizer: %s", out)
+	}
+}
+
+func TestApplyThinkingWithSourcePayloadPreservesOriginalOnlySummary(t *testing.T) {
+	currentSource := []byte(`{"model":"gemini-3.6-flash","input":"hi"}`)
+	originalSource := []byte(`{"model":"gemini-3.6-flash","reasoning":{"summary":null},"input":"hi"}`)
+	body := []byte(`{"generationConfig":{"thinkingConfig":{"thinkingLevel":"high"}}}`)
+
+	out, err := helps.ApplyThinkingWithSourcePayload(
+		body,
+		currentSource,
+		originalSource,
+		"gemini-3.6-flash",
+		sdktranslator.FormatOpenAIResponse.String(),
+		sdktranslator.FormatGemini.String(),
+		"gemini",
+	)
+	if err != nil {
+		t.Fatalf("ApplyThinkingWithSourcePayload() error = %v", err)
+	}
+	if include := gjson.GetBytes(out, "generationConfig.thinkingConfig.includeThoughts"); !include.Exists() || include.Bool() {
+		t.Fatalf("original disabled summary was not preserved: %s", out)
+	}
+}
diff --git a/internal/runtime/executor/kimi_executor.go b/internal/runtime/executor/kimi_executor.go
index d3c88145..b9a89425 100644
--- a/internal/runtime/executor/kimi_executor.go
+++ b/internal/runtime/executor/kimi_executor.go
@@ -113,7 +113,7 @@ func (e *KimiExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req
 		return resp, fmt.Errorf("kimi executor: failed to set model in payload: %w", err)
 	}
 
-	body, err = helps.ApplyThinkingWithSourcePayload(body, req.Payload, req.Model, from.String(), "kimi", e.Identifier())
+	body, err = helps.ApplyThinkingWithSourcePayload(body, req.Payload, originalPayloadSource, req.Model, from.String(), "kimi", e.Identifier())
 	if err != nil {
 		return resp, err
 	}
@@ -222,7 +222,7 @@ func (e *KimiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 		return nil, fmt.Errorf("kimi executor: failed to set model in payload: %w", err)
 	}
 
-	body, err = helps.ApplyThinkingWithSourcePayload(body, req.Payload, req.Model, from.String(), "kimi", e.Identifier())
+	body, err = helps.ApplyThinkingWithSourcePayload(body, req.Payload, originalPayloadSource, req.Model, from.String(), "kimi", e.Identifier())
 	if err != nil {
 		return nil, err
 	}
diff --git a/internal/thinking/apply.go b/internal/thinking/apply.go
index a349f269..e9e3d34d 100644
--- a/internal/thinking/apply.go
+++ b/internal/thinking/apply.go
@@ -183,6 +183,13 @@ func ApplyThinkingWithModelInfo(body, sourceBody []byte, model string, fromForma
 	if len(sourceBody) == 0 {
 		summaryConfig = ExtractSummaryConfig(body, toFormat)
 	}
+	return ApplyThinkingWithModelInfoAndSummary(body, sourceBody, model, fromFormat, toFormat, providerKey, modelInfo, summaryConfig)
+}
+
+// ApplyThinkingWithModelInfoAndSummary applies the exact configured model
+// definition with a summary intent already resolved across source translation
+// and plugin normalization.
+func ApplyThinkingWithModelInfoAndSummary(body, sourceBody []byte, model string, fromFormat string, toFormat string, providerKey string, modelInfo *registry.ModelInfo, summaryConfig SummaryConfig) ([]byte, error) {
 	return applyThinking(body, sourceBody, model, fromFormat, toFormat, providerKey, modelInfo, true, summaryConfig)
 }
 
diff --git a/internal/thinking/summary.go b/internal/thinking/summary.go
index 17951990..72977ab4 100644
--- a/internal/thinking/summary.go
+++ b/internal/thinking/summary.go
@@ -114,6 +114,22 @@ func ExtractSummaryConfig(body []byte, format string) SummaryConfig {
 	return SummaryConfig{}
 }
 
+// ExtractExplicitSummaryConfig reads only explicit visibility controls from a
+// provider payload. Unlike ExtractSummaryConfig, OpenAI Chat reasoning_effort
+// is not treated as a summary proxy. This lets executor post-processing tell
+// whether a request normalizer retained or removed the translated target field.
+func ExtractExplicitSummaryConfig(body []byte, format string) SummaryConfig {
+	normalized := strings.ToLower(strings.TrimSpace(format))
+	if normalized != "openai" {
+		return ExtractSummaryConfig(body, normalized)
+	}
+	if len(body) == 0 || !gjson.ValidBytes(body) {
+		return SummaryConfig{}
+	}
+	config, _ := extractOpenAIExplicitSummaryConfig(body)
+	return config
+}
+
 // ApplySummaryConfig writes canonical summary intent in the target protocol.
 func ApplySummaryConfig(body []byte, format string, config SummaryConfig) []byte {
 	return ApplySummaryConfigForModel(body, format, "", config)
diff --git a/internal/thinking/summary_test.go b/internal/thinking/summary_test.go
index 84c110c9..e038fc14 100644
--- a/internal/thinking/summary_test.go
+++ b/internal/thinking/summary_test.go
@@ -75,6 +75,18 @@ func TestExtractSummaryConfig(t *testing.T) {
 	}
 }
 
+func TestExtractExplicitSummaryConfigDoesNotUseChatEffort(t *testing.T) {
+	body := []byte(`{"reasoning_effort":"high"}`)
+	if got := ExtractExplicitSummaryConfig(body, "openai"); got.Mode != SummaryUnspecified {
+		t.Fatalf("ExtractExplicitSummaryConfig() = %+v, want unspecified", got)
+	}
+
+	body = []byte(`{"reasoning_effort":"high","reasoning":{"exclude":true}}`)
+	if got := ExtractExplicitSummaryConfig(body, "openai"); got.Mode != SummaryDisabled {
+		t.Fatalf("ExtractExplicitSummaryConfig() = %+v, want disabled", got)
+	}
+}
+
 func TestApplySummaryConfig(t *testing.T) {
 	tests := []struct {
 		name   string
diff --git a/test/thinking_conversion_test.go b/test/thinking_conversion_test.go
index d71d6e35..45d709e3 100644
--- a/test/thinking_conversion_test.go
+++ b/test/thinking_conversion_test.go
@@ -241,7 +241,7 @@ func TestThinkingE2EMatrix_Suffix(t *testing.T) {
 			inputJSON:       `{"model":"level-subset-model(1)","messages":[{"role":"user","content":"hi"}]}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingLevel",
 			expectValue:     "low",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 		// Case 17A: auto → medium → clamped to low when low/high are equally close
@@ -277,7 +277,7 @@ func TestThinkingE2EMatrix_Suffix(t *testing.T) {
 			inputJSON:       `{"model":"gemini-budget-model(medium)","messages":[{"role":"user","content":"hi"}]}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "8192",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 		// Case 20: Effort xhigh → clamped to 20000 (max)
@@ -289,10 +289,10 @@ func TestThinkingE2EMatrix_Suffix(t *testing.T) {
 			inputJSON:       `{"model":"gemini-budget-model(xhigh)","messages":[{"role":"user","content":"hi"}]}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "20000",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
-		// Case 21: Effort none → clamped to 128 (min) → includeThoughts=false
+		// Case 21: Effort none → clamped to 128 (min)
 		{
 			name:            "21",
 			from:            "openai",
@@ -301,7 +301,7 @@ func TestThinkingE2EMatrix_Suffix(t *testing.T) {
 			inputJSON:       `{"model":"gemini-budget-model(none)","messages":[{"role":"user","content":"hi"}]}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "128",
-			includeThoughts: "false",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 		// Case 22: Effort auto → DynamicAllowed=true → -1
@@ -313,7 +313,7 @@ func TestThinkingE2EMatrix_Suffix(t *testing.T) {
 			inputJSON:       `{"model":"gemini-budget-model(auto)","messages":[{"role":"user","content":"hi"}]}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "-1",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 		// Case 23: Claude source no suffix → passthrough
@@ -335,7 +335,7 @@ func TestThinkingE2EMatrix_Suffix(t *testing.T) {
 			inputJSON:       `{"model":"gemini-budget-model(8192)","messages":[{"role":"user","content":"hi"}]}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "8192",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 		// Case 25: Budget 64000 → clamped to 20000 (max)
@@ -347,10 +347,10 @@ func TestThinkingE2EMatrix_Suffix(t *testing.T) {
 			inputJSON:       `{"model":"gemini-budget-model(64000)","messages":[{"role":"user","content":"hi"}]}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "20000",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
-		// Case 26: Budget 0 → clamped to 128 (min) → includeThoughts=false
+		// Case 26: Budget 0 → clamped to 128 (min)
 		{
 			name:            "26",
 			from:            "claude",
@@ -359,7 +359,7 @@ func TestThinkingE2EMatrix_Suffix(t *testing.T) {
 			inputJSON:       `{"model":"gemini-budget-model(0)","messages":[{"role":"user","content":"hi"}]}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "128",
-			includeThoughts: "false",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 		// Case 27: Budget -1 → DynamicAllowed=true → -1
@@ -371,7 +371,7 @@ func TestThinkingE2EMatrix_Suffix(t *testing.T) {
 			inputJSON:       `{"model":"gemini-budget-model(-1)","messages":[{"role":"user","content":"hi"}]}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "-1",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 
@@ -396,7 +396,7 @@ func TestThinkingE2EMatrix_Suffix(t *testing.T) {
 			inputJSON:       `{"model":"gemini-mixed-model(high)","messages":[{"role":"user","content":"hi"}]}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingLevel",
 			expectValue:     "high",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 		// Case 30: Effort xhigh → clamped to high
@@ -408,10 +408,10 @@ func TestThinkingE2EMatrix_Suffix(t *testing.T) {
 			inputJSON:       `{"model":"gemini-mixed-model(xhigh)","messages":[{"role":"user","content":"hi"}]}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingLevel",
 			expectValue:     "high",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
-		// Case 31: Effort none → clamped to low (min supported) → includeThoughts=false
+		// Case 31: Effort none → clamped to low (min supported)
 		{
 			name:            "31",
 			from:            "openai",
@@ -420,7 +420,7 @@ func TestThinkingE2EMatrix_Suffix(t *testing.T) {
 			inputJSON:       `{"model":"gemini-mixed-model(none)","messages":[{"role":"user","content":"hi"}]}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingLevel",
 			expectValue:     "low",
-			includeThoughts: "false",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 		// Case 32: Effort auto → DynamicAllowed=true → -1 (budget)
@@ -432,7 +432,7 @@ func TestThinkingE2EMatrix_Suffix(t *testing.T) {
 			inputJSON:       `{"model":"gemini-mixed-model(auto)","messages":[{"role":"user","content":"hi"}]}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "-1",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 		// Case 33: Claude source no suffix → passthrough
@@ -454,7 +454,7 @@ func TestThinkingE2EMatrix_Suffix(t *testing.T) {
 			inputJSON:       `{"model":"gemini-mixed-model(8192)","messages":[{"role":"user","content":"hi"}]}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "8192",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 		// Case 35: Budget 64000 → clamped to 32768 (max)
@@ -466,10 +466,10 @@ func TestThinkingE2EMatrix_Suffix(t *testing.T) {
 			inputJSON:       `{"model":"gemini-mixed-model(64000)","messages":[{"role":"user","content":"hi"}]}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "32768",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
-		// Case 36: Budget 0 → minimal → clamped to low (min level) → includeThoughts=false
+		// Case 36: Budget 0 → minimal → clamped to low (min level)
 		{
 			name:            "36",
 			from:            "claude",
@@ -478,7 +478,7 @@ func TestThinkingE2EMatrix_Suffix(t *testing.T) {
 			inputJSON:       `{"model":"gemini-mixed-model(0)","messages":[{"role":"user","content":"hi"}]}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingLevel",
 			expectValue:     "low",
-			includeThoughts: "false",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 		// Case 37: Budget -1 → DynamicAllowed=true → -1 (budget)
@@ -490,7 +490,7 @@ func TestThinkingE2EMatrix_Suffix(t *testing.T) {
 			inputJSON:       `{"model":"gemini-mixed-model(-1)","messages":[{"role":"user","content":"hi"}]}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "-1",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 
@@ -626,7 +626,7 @@ func TestThinkingE2EMatrix_Suffix(t *testing.T) {
 			inputJSON:       `{"model":"antigravity-budget-model(medium)","contents":[{"role":"user","parts":[{"text":"hi"}]}]}`,
 			expectField:     "request.generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "8192",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 		// Case 50: Effort xhigh → clamped to 20000 (max)
@@ -638,10 +638,10 @@ func TestThinkingE2EMatrix_Suffix(t *testing.T) {
 			inputJSON:       `{"model":"antigravity-budget-model(xhigh)","contents":[{"role":"user","parts":[{"text":"hi"}]}]}`,
 			expectField:     "request.generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "20000",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
-		// Case 51: Effort none → ZeroAllowed=true → 0 → includeThoughts=false
+		// Case 51: Effort none → ZeroAllowed=true → 0
 		{
 			name:            "51",
 			from:            "gemini",
@@ -650,7 +650,7 @@ func TestThinkingE2EMatrix_Suffix(t *testing.T) {
 			inputJSON:       `{"model":"antigravity-budget-model(none)","contents":[{"role":"user","parts":[{"text":"hi"}]}]}`,
 			expectField:     "request.generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "0",
-			includeThoughts: "false",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 		// Case 52: Effort auto → DynamicAllowed=true → -1
@@ -662,7 +662,7 @@ func TestThinkingE2EMatrix_Suffix(t *testing.T) {
 			inputJSON:       `{"model":"antigravity-budget-model(auto)","contents":[{"role":"user","parts":[{"text":"hi"}]}]}`,
 			expectField:     "request.generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "-1",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 		// Case 53: Claude to Antigravity no suffix → passthrough
@@ -684,7 +684,7 @@ func TestThinkingE2EMatrix_Suffix(t *testing.T) {
 			inputJSON:       `{"model":"antigravity-budget-model(8192)","messages":[{"role":"user","content":"hi"}]}`,
 			expectField:     "request.generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "8192",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 		// Case 55: Budget 64000 → clamped to 20000 (max)
@@ -696,10 +696,10 @@ func TestThinkingE2EMatrix_Suffix(t *testing.T) {
 			inputJSON:       `{"model":"antigravity-budget-model(64000)","messages":[{"role":"user","content":"hi"}]}`,
 			expectField:     "request.generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "20000",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
-		// Case 56: Budget 0 → ZeroAllowed=true → 0 → includeThoughts=false
+		// Case 56: Budget 0 → ZeroAllowed=true → 0
 		{
 			name:            "56",
 			from:            "claude",
@@ -708,7 +708,7 @@ func TestThinkingE2EMatrix_Suffix(t *testing.T) {
 			inputJSON:       `{"model":"antigravity-budget-model(0)","messages":[{"role":"user","content":"hi"}]}`,
 			expectField:     "request.generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "0",
-			includeThoughts: "false",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 		// Case 57: Budget -1 → DynamicAllowed=true → -1
@@ -720,7 +720,7 @@ func TestThinkingE2EMatrix_Suffix(t *testing.T) {
 			inputJSON:       `{"model":"antigravity-budget-model(-1)","messages":[{"role":"user","content":"hi"}]}`,
 			expectField:     "request.generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "-1",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 
@@ -927,7 +927,7 @@ func TestThinkingE2EMatrix_Suffix(t *testing.T) {
 			inputJSON:       `{"model":"user-defined-model(8192)","messages":[{"role":"user","content":"hi"}]}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "8192",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 		// Case 77: OpenAI to Claude budget 8192 → passthrough → 8192
@@ -950,7 +950,7 @@ func TestThinkingE2EMatrix_Suffix(t *testing.T) {
 			inputJSON:       `{"model":"user-defined-model(8192)","input":[{"role":"user","content":"hi"}]}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "8192",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 		// Case 79: OpenAI-Response to Claude budget 8192 → passthrough → 8192
@@ -1018,7 +1018,7 @@ func TestThinkingE2EMatrix_Suffix(t *testing.T) {
 			inputJSON:       `{"model":"gemini-budget-model(8192)","contents":[{"role":"user","parts":[{"text":"hi"}]}]}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "8192",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 		// Case 85: Gemini to Gemini, budget 64000 → clamped to Max
@@ -1030,7 +1030,7 @@ func TestThinkingE2EMatrix_Suffix(t *testing.T) {
 			inputJSON:       `{"model":"gemini-budget-model(64000)","contents":[{"role":"user","parts":[{"text":"hi"}]}]}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "20000",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 		// Case 86: Claude to Claude, budget 8192 → passthrough thinking.budget_tokens
@@ -1067,7 +1067,7 @@ func TestThinkingE2EMatrix_Suffix(t *testing.T) {
 			inputJSON:       `{"model":"gemini-budget-model(64000)","contents":[{"role":"user","parts":[{"text":"hi"}]}]}`,
 			expectField:     "request.generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "20000",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 		// Case 89: Gemini to Antigravity, budget 8192 → passthrough (normal value)
@@ -1079,7 +1079,7 @@ func TestThinkingE2EMatrix_Suffix(t *testing.T) {
 			inputJSON:       `{"model":"gemini-budget-model(8192)","contents":[{"role":"user","parts":[{"text":"hi"}]}]}`,
 			expectField:     "request.generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "8192",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 	}
@@ -1285,7 +1285,7 @@ func TestThinkingE2EMatrix_Body(t *testing.T) {
 			inputJSON:       `{"model":"level-subset-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"enabled","budget_tokens":1}}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingLevel",
 			expectValue:     "low",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 
@@ -1368,7 +1368,7 @@ func TestThinkingE2EMatrix_Body(t *testing.T) {
 			inputJSON:       `{"model":"gemini-budget-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"enabled","budget_tokens":8192}}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "8192",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 		// Case 25: thinking.budget_tokens=64000 → clamped to 20000
@@ -1380,10 +1380,10 @@ func TestThinkingE2EMatrix_Body(t *testing.T) {
 			inputJSON:       `{"model":"gemini-budget-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"enabled","budget_tokens":64000}}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "20000",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
-		// Case 26: thinking.budget_tokens=0 → clamped to 128 → includeThoughts=false
+		// Case 26: thinking.budget_tokens=0 → clamped to 128
 		{
 			name:            "26",
 			from:            "claude",
@@ -1392,7 +1392,7 @@ func TestThinkingE2EMatrix_Body(t *testing.T) {
 			inputJSON:       `{"model":"gemini-budget-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"enabled","budget_tokens":0}}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "128",
-			includeThoughts: "false",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 		// Case 27: thinking.budget_tokens=-1 → -1 (DynamicAllowed=true)
@@ -1404,7 +1404,7 @@ func TestThinkingE2EMatrix_Body(t *testing.T) {
 			inputJSON:       `{"model":"gemini-budget-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"enabled","budget_tokens":-1}}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "-1",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 
@@ -1528,7 +1528,7 @@ func TestThinkingE2EMatrix_Body(t *testing.T) {
 			inputJSON:       `{"model":"gemini-mixed-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"enabled","budget_tokens":8192}}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "8192",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 		// Case 35: thinking.budget_tokens=64000 → clamped to 32768 (keeps budget)
@@ -1540,10 +1540,10 @@ func TestThinkingE2EMatrix_Body(t *testing.T) {
 			inputJSON:       `{"model":"gemini-mixed-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"enabled","budget_tokens":64000}}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "32768",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
-		// Case 36: thinking.budget_tokens=0 → clamped to low → includeThoughts=false
+		// Case 36: thinking.budget_tokens=0 → clamped to low
 		{
 			name:            "36",
 			from:            "claude",
@@ -1552,7 +1552,7 @@ func TestThinkingE2EMatrix_Body(t *testing.T) {
 			inputJSON:       `{"model":"gemini-mixed-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"enabled","budget_tokens":0}}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingLevel",
 			expectValue:     "low",
-			includeThoughts: "false",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 		// Case 37: thinking.budget_tokens=-1 → -1 (DynamicAllowed=true)
@@ -1564,7 +1564,7 @@ func TestThinkingE2EMatrix_Body(t *testing.T) {
 			inputJSON:       `{"model":"gemini-mixed-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"enabled","budget_tokens":-1}}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "-1",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 
@@ -1700,7 +1700,7 @@ func TestThinkingE2EMatrix_Body(t *testing.T) {
 			inputJSON:       `{"model":"antigravity-budget-model","contents":[{"role":"user","parts":[{"text":"hi"}]}],"generationConfig":{"thinkingConfig":{"thinkingLevel":"medium"}}}`,
 			expectField:     "request.generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "8192",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 		// Case 50: thinkingLevel=xhigh → clamped to 20000
@@ -1712,7 +1712,7 @@ func TestThinkingE2EMatrix_Body(t *testing.T) {
 			inputJSON:       `{"model":"antigravity-budget-model","contents":[{"role":"user","parts":[{"text":"hi"}]}],"generationConfig":{"thinkingConfig":{"thinkingLevel":"xhigh"}}}`,
 			expectField:     "request.generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "20000",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 		// Case 51: thinkingLevel=none → 0 (ZeroAllowed=true)
@@ -1724,7 +1724,7 @@ func TestThinkingE2EMatrix_Body(t *testing.T) {
 			inputJSON:       `{"model":"antigravity-budget-model","contents":[{"role":"user","parts":[{"text":"hi"}]}],"generationConfig":{"thinkingConfig":{"thinkingLevel":"none"}}}`,
 			expectField:     "request.generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "0",
-			includeThoughts: "false",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 		// Case 52: thinkingBudget=-1 → -1 (DynamicAllowed=true)
@@ -1736,7 +1736,7 @@ func TestThinkingE2EMatrix_Body(t *testing.T) {
 			inputJSON:       `{"model":"antigravity-budget-model","contents":[{"role":"user","parts":[{"text":"hi"}]}],"generationConfig":{"thinkingConfig":{"thinkingBudget":-1}}}`,
 			expectField:     "request.generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "-1",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 		// Case 53: Claude no param → passthrough
@@ -1758,7 +1758,7 @@ func TestThinkingE2EMatrix_Body(t *testing.T) {
 			inputJSON:       `{"model":"antigravity-budget-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"enabled","budget_tokens":8192}}`,
 			expectField:     "request.generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "8192",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 		// Case 55: thinking.budget_tokens=64000 → clamped to 20000
@@ -1770,7 +1770,7 @@ func TestThinkingE2EMatrix_Body(t *testing.T) {
 			inputJSON:       `{"model":"antigravity-budget-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"enabled","budget_tokens":64000}}`,
 			expectField:     "request.generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "20000",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 		// Case 56: thinking.budget_tokens=0 → 0 (ZeroAllowed=true)
@@ -1782,7 +1782,7 @@ func TestThinkingE2EMatrix_Body(t *testing.T) {
 			inputJSON:       `{"model":"antigravity-budget-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"enabled","budget_tokens":0}}`,
 			expectField:     "request.generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "0",
-			includeThoughts: "false",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 		// Case 57: thinking.budget_tokens=-1 → -1 (DynamicAllowed=true)
@@ -1794,7 +1794,7 @@ func TestThinkingE2EMatrix_Body(t *testing.T) {
 			inputJSON:       `{"model":"antigravity-budget-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"enabled","budget_tokens":-1}}`,
 			expectField:     "request.generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "-1",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 
@@ -2024,7 +2024,7 @@ func TestThinkingE2EMatrix_Body(t *testing.T) {
 			inputJSON:       `{"model":"user-defined-model","input":[{"role":"user","content":"hi"}],"reasoning":{"effort":"medium"}}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "8192",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 		// Case 79: OpenAI-Response reasoning.effort=medium to Claude → 8192
@@ -2092,7 +2092,7 @@ func TestThinkingE2EMatrix_Body(t *testing.T) {
 			inputJSON:       `{"model":"gemini-budget-model","contents":[{"role":"user","parts":[{"text":"hi"}]}],"generationConfig":{"thinkingConfig":{"thinkingBudget":8192}}}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "8192",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 		// Case 85: Gemini to Gemini, thinkingBudget=64000 → exceeds Max error
@@ -2148,7 +2148,7 @@ func TestThinkingE2EMatrix_Body(t *testing.T) {
 			inputJSON:       `{"model":"gemini-budget-model","contents":[{"role":"user","parts":[{"text":"hi"}]}],"generationConfig":{"thinkingConfig":{"thinkingBudget":8192}}}`,
 			expectField:     "request.generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "8192",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 	}
@@ -2511,7 +2511,7 @@ func TestThinkingE2EProviderTargets(t *testing.T) {
 			inputJSON:       `{"model":"antigravity-budget-model","input":"hi","reasoning":{"effort":"medium"}}`,
 			expectField:     "request.generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "8192",
-			includeThoughts: "true",
+			includeThoughts: "",
 		},
 	}
 
@@ -2692,7 +2692,7 @@ func TestThinkingE2EInteractionsMatrix(t *testing.T) {
 			inputJSON:       `{"model":"antigravity-budget-model","generation_config":{"thinking_level":"medium"},"input":"hi"}`,
 			expectField:     "request.generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "8192",
-			includeThoughts: "true",
+			includeThoughts: "",
 		},
 		{
 			name:         "OUT6",
@@ -2733,7 +2733,7 @@ func TestThinkingE2EInteractionsMatrix(t *testing.T) {
 			inputJSON:       `{"model":"antigravity-budget-model","generation_config":{"thinking_level":"none"},"input":"hi"}`,
 			expectField:     "request.generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "0",
-			includeThoughts: "false",
+			includeThoughts: "",
 		},
 		{
 			name:         "OUT10",
@@ -3084,7 +3084,7 @@ func TestThinkingE2EClaudeAdaptive_Body(t *testing.T) {
 			inputJSON:       `{"model":"level-subset-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"high"}}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingLevel",
 			expectValue:     "high",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 		{
@@ -3095,7 +3095,7 @@ func TestThinkingE2EClaudeAdaptive_Body(t *testing.T) {
 			inputJSON:       `{"model":"gemini-budget-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"low"}}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "1024",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 		{
@@ -3106,7 +3106,7 @@ func TestThinkingE2EClaudeAdaptive_Body(t *testing.T) {
 			inputJSON:       `{"model":"gemini-budget-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"medium"}}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "8192",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 		{
@@ -3117,7 +3117,7 @@ func TestThinkingE2EClaudeAdaptive_Body(t *testing.T) {
 			inputJSON:       `{"model":"gemini-budget-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"high"}}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "20000",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 		{
@@ -3128,7 +3128,7 @@ func TestThinkingE2EClaudeAdaptive_Body(t *testing.T) {
 			inputJSON:       `{"model":"gemini-budget-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"}}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "20000",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 		{
@@ -3139,7 +3139,7 @@ func TestThinkingE2EClaudeAdaptive_Body(t *testing.T) {
 			inputJSON:       `{"model":"gemini-mixed-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"high"}}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingLevel",
 			expectValue:     "high",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 
@@ -3201,7 +3201,7 @@ func TestThinkingE2EClaudeAdaptive_Body(t *testing.T) {
 			inputJSON:       `{"model":"antigravity-budget-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"}}`,
 			expectField:     "request.generationConfig.thinkingConfig.thinkingBudget",
 			expectValue:     "20000",
-			includeThoughts: "true",
+			includeThoughts: "",
 			expectErr:       false,
 		},
 
@@ -3504,18 +3504,9 @@ func runThinkingTests(t *testing.T, cases []thinkingTestCase) {
 				if tc.to == "antigravity" {
 					path = "request.generationConfig.thinkingConfig.includeThoughts"
 				}
-				wantIncludeThoughts := ""
-				summaryConfig := thinking.ExtractSummaryConfig([]byte(tc.inputJSON), tc.from)
-				switch summaryConfig.Mode {
-				case thinking.SummaryEnabled:
-					wantIncludeThoughts = "true"
-				case thinking.SummaryDisabled:
-					wantIncludeThoughts = "false"
-				default:
-					// Thinking amount does not imply summary visibility. Keep the
-					// provider field absent when the source omitted its summary control.
-				}
-
+				// Each case declares its expected visibility independently from the
+				// extractor under test. Empty means the provider field must be absent.
+				wantIncludeThoughts := tc.includeThoughts
 				itVal := gjson.GetBytes(body, path)
 				if wantIncludeThoughts == "" {
 					if itVal.Exists() {
-- 
2.51.2


From c4dcd8703ad964ab7ca1f0c98b74948d7de7d1ce Mon Sep 17 00:00:00 2001
From: sususu <suchangshan@foxmail.com>
Date: Fri, 31 Jul 2026 13:28:14 +0800
Subject: [PATCH 8/9] fix(thinking): respect final summary authority

---
 .../executor/helps/model_capabilities_test.go | 88 ++++++++++++++++++-
 internal/thinking/apply.go                    | 11 +++
 .../thinking/apply_configured_api_key_test.go | 20 +++++
 internal/thinking/summary.go                  | 28 ++++++
 sdk/translator/registry.go                    |  7 +-
 sdk/translator/registry_summary_test.go       | 53 +++++++++++
 6 files changed, 202 insertions(+), 5 deletions(-)

diff --git a/internal/runtime/executor/helps/model_capabilities_test.go b/internal/runtime/executor/helps/model_capabilities_test.go
index c1e0b371..826c82e9 100644
--- a/internal/runtime/executor/helps/model_capabilities_test.go
+++ b/internal/runtime/executor/helps/model_capabilities_test.go
@@ -9,6 +9,7 @@ import (
 	"github.com/router-for-me/CLIProxyAPI/v7/internal/registry"
 	helps "github.com/router-for-me/CLIProxyAPI/v7/internal/runtime/executor/helps"
 	_ "github.com/router-for-me/CLIProxyAPI/v7/internal/thinking/provider/claude"
+	_ "github.com/router-for-me/CLIProxyAPI/v7/internal/translator"
 	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v7/sdk/cliproxy/auth"
 	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v7/sdk/cliproxy/executor"
 	sdktranslator "github.com/router-for-me/CLIProxyAPI/v7/sdk/translator"
@@ -16,8 +17,10 @@ import (
 )
 
 type configuredThinkingExecutor struct {
-	seenModel string
-	resolved  bool
+	seenModel        string
+	resolved         bool
+	translateRequest bool
+	translatedBody   []byte
 }
 
 func (*configuredThinkingExecutor) Identifier() string { return "claude" }
@@ -27,6 +30,10 @@ func (e *configuredThinkingExecutor) Execute(_ context.Context, _ *cliproxyauth.
 	modelInfo, resolved := cliproxyauth.ResolvedAPIKeyModelInfo(req)
 	e.resolved = resolved && modelInfo != nil
 	body := []byte(`{"thinking":{"type":"adaptive"},"output_config":{"effort":"low"}}`)
+	if e.translateRequest {
+		body = sdktranslator.TranslateRequest(opts.SourceFormat, sdktranslator.FormatClaude, req.Model, req.Payload, opts.Stream)
+		e.translatedBody = append(e.translatedBody[:0], body...)
+	}
 	out, err := helps.ApplyRequestThinking(body, req, opts, opts.SourceFormat.String(), "claude", "claude")
 	return cliproxyexecutor.Response{Payload: out}, err
 }
@@ -54,6 +61,83 @@ func (*configuredThinkingExecutor) HttpRequest(context.Context, *cliproxyauth.Au
 	return nil, nil
 }
 
+func TestApplyRequestThinkingUsesExactClaudeModeForSummaryOnlyRequest(t *testing.T) {
+	manager := cliproxyauth.NewManager(nil, nil, nil)
+	manager.SetConfig(&internalconfig.Config{
+		SDKConfig: internalconfig.SDKConfig{ForceModelPrefix: true},
+		ClaudeKey: []internalconfig.ClaudeKey{{
+			APIKey: "summary-selected-key",
+			Prefix: "summary-tenant",
+			Models: []internalconfig.ClaudeModel{{
+				Name:  "summary-shared-upstream",
+				Alias: "summary-public-model",
+				Thinking: &registry.ThinkingSupport{
+					Min: 1024,
+					Max: 16000,
+				},
+			}},
+		}},
+	})
+	executor := &configuredThinkingExecutor{translateRequest: true}
+	manager.RegisterExecutor(executor)
+	auth := &cliproxyauth.Auth{
+		ID:       "summary-selected-auth",
+		Provider: "claude",
+		Prefix:   "summary-tenant",
+		Attributes: map[string]string{
+			cliproxyauth.AttributeAuthKind: cliproxyauth.AuthKindAPIKey,
+			cliproxyauth.AttributeAPIKey:   "summary-selected-key",
+			cliproxyauth.AttributeSource:   "config:claude[0]",
+		},
+	}
+
+	modelRegistry := registry.GetGlobalRegistry()
+	modelRegistry.RegisterClient(auth.ID, auth.Provider, []*registry.ModelInfo{{
+		ID: "summary-tenant/summary-public-model", Type: "claude",
+	}})
+	modelRegistry.RegisterClient("summary-unrelated-auth", auth.Provider, []*registry.ModelInfo{{
+		ID: "summary-shared-upstream", Type: "claude",
+		Thinking: &registry.ThinkingSupport{Levels: []string{"high"}},
+	}})
+	t.Cleanup(func() {
+		modelRegistry.UnregisterClient(auth.ID)
+		modelRegistry.UnregisterClient("summary-unrelated-auth")
+	})
+	if registered, errRegister := manager.Register(t.Context(), auth); errRegister != nil {
+		t.Fatalf("Register() error = %v", errRegister)
+	} else if registered == nil {
+		t.Fatal("Register() returned nil auth")
+	}
+
+	original := []byte(`{"model":"summary-tenant/summary-public-model","reasoning":{"summary":"auto"},"input":"hi"}`)
+	response, errExecute := manager.Execute(t.Context(), []string{"claude"}, cliproxyexecutor.Request{
+		Model:   "summary-tenant/summary-public-model",
+		Payload: original,
+		Format:  sdktranslator.FormatOpenAIResponse,
+	}, cliproxyexecutor.Options{
+		SourceFormat:    sdktranslator.FormatOpenAIResponse,
+		OriginalRequest: original,
+	})
+	if errExecute != nil {
+		t.Fatalf("Execute() error = %v", errExecute)
+	}
+	if got := gjson.GetBytes(executor.translatedBody, "thinking.type").String(); got != "adaptive" {
+		t.Fatalf("pre-executor thinking.type = %q, want global adaptive trigger; body=%s", got, executor.translatedBody)
+	}
+	if got := gjson.GetBytes(response.Payload, "thinking.type").String(); got != "enabled" {
+		t.Fatalf("thinking.type = %q, want exact manual mode; body=%s", got, response.Payload)
+	}
+	if got := gjson.GetBytes(response.Payload, "thinking.budget_tokens").Int(); got != 1024 {
+		t.Fatalf("thinking.budget_tokens = %d, want exact minimum 1024; body=%s", got, response.Payload)
+	}
+	if got := gjson.GetBytes(response.Payload, "thinking.display").String(); got != "summarized" {
+		t.Fatalf("thinking.display = %q, want summarized; body=%s", got, response.Payload)
+	}
+	if gjson.GetBytes(response.Payload, "output_config.effort").Exists() {
+		t.Fatalf("manual thinking retained adaptive effort: %s", response.Payload)
+	}
+}
+
 func TestApplyRequestThinkingUsesSelectedPrefixedAPIKeyModel(t *testing.T) {
 	manager := cliproxyauth.NewManager(nil, nil, nil)
 	manager.SetConfig(&internalconfig.Config{
diff --git a/internal/thinking/apply.go b/internal/thinking/apply.go
index e9e3d34d..92e6161c 100644
--- a/internal/thinking/apply.go
+++ b/internal/thinking/apply.go
@@ -284,6 +284,17 @@ func applyThinking(body, sourceBody []byte, model string, fromFormat string, toF
 			"provider": providerFormat,
 			"model":    modelInfo.ID,
 		}).Debug("thinking: no config found, passthrough |")
+		if modelInfoResolved && providerFormat == "claude" && fromFormat != providerFormat && ExtractSummaryConfig(sourceBody, fromFormat).Mode == SummaryEnabled {
+			// Registry translation can only see aggregate model capabilities. For a
+			// cross-protocol summary-only request it may have activated adaptive
+			// thinking solely to make display valid. The selected API-key model is
+			// authoritative at execution time, so discard that inferred activation
+			// when the exact model supports only manual extended thinking. Use the
+			// source intent here even if a target normalizer removed display; in that
+			// case the inferred amount must disappear with it. Explicit native Claude
+			// thinking never reaches this cross-protocol branch.
+			body = stripInferredClaudeSummaryActivation(body, modelInfo)
+		}
 		return applySummaryConfigForProvider(body, providerFormat, baseModel, providerKey, modelInfo, summaryConfig), nil
 	}
 	if modelInfoResolved && config.Mode == ModeLevel && modelInfo != nil && modelInfo.Thinking != nil && shouldMapConfiguredHighIntent(fromFormat, providerFormat, modelInfo) {
diff --git a/internal/thinking/apply_configured_api_key_test.go b/internal/thinking/apply_configured_api_key_test.go
index b056139e..9c48c36d 100644
--- a/internal/thinking/apply_configured_api_key_test.go
+++ b/internal/thinking/apply_configured_api_key_test.go
@@ -114,6 +114,26 @@ func TestApplyThinkingWithModelInfoAppliesEnabledSummaryOnlyClaudeVisibility(t *
 	}
 }
 
+func TestApplyThinkingWithModelInfoAndSummaryDropsInferredClaudeModeWhenSummaryRemoved(t *testing.T) {
+	modelInfo := &registry.ModelInfo{
+		ID:       "private-manual-claude",
+		Type:     "claude",
+		Thinking: &registry.ThinkingSupport{Min: 1024, Max: 16000},
+	}
+	out, err := thinking.ApplyThinkingWithModelInfoAndSummary(
+		[]byte(`{"model":"private-manual-claude","max_tokens":32000,"thinking":{"type":"adaptive"}}`),
+		[]byte(`{"reasoning":{"summary":"auto"}}`),
+		"private-manual-claude", "openai-response", "claude", "claude", modelInfo,
+		thinking.SummaryConfig{},
+	)
+	if err != nil {
+		t.Fatalf("ApplyThinkingWithModelInfoAndSummary() error = %v", err)
+	}
+	if gjson.GetBytes(out, "thinking").Exists() {
+		t.Fatalf("removed summary retained globally inferred adaptive thinking: %s", out)
+	}
+}
+
 func TestApplyThinkingWithModelInfoDoesNotActivateClaudeForDisabledSummary(t *testing.T) {
 	modelInfo := &registry.ModelInfo{
 		ID:       "private-claude",
diff --git a/internal/thinking/summary.go b/internal/thinking/summary.go
index 72977ab4..34ae9010 100644
--- a/internal/thinking/summary.go
+++ b/internal/thinking/summary.go
@@ -441,6 +441,34 @@ func interactionsSummaryConfig(body []byte, path string) (SummaryConfig, bool) {
 	}
 }
 
+// stripInferredClaudeSummaryActivation removes a globally inferred adaptive
+// mode when the selected API-key model supports only manual extended thinking.
+// The exact model-aware summary pass can then activate enabled thinking with a
+// valid budget, or leave thinking absent when max_tokens cannot accommodate it.
+func stripInferredClaudeSummaryActivation(body []byte, modelInfo *registry.ModelInfo) []byte {
+	if modelInfo == nil || modelInfo.Thinking == nil || len(modelInfo.Thinking.Levels) > 0 || modelInfo.Thinking.Min <= 0 {
+		return body
+	}
+	if !strings.EqualFold(strings.TrimSpace(gjson.GetBytes(body, "thinking.type").String()), "adaptive") {
+		return body
+	}
+
+	for _, path := range []string{
+		"thinking.type",
+		"thinking.budget_tokens",
+		"thinking.display",
+		"output_config.effort",
+	} {
+		body, _ = sjson.DeleteBytes(body, path)
+	}
+	for _, path := range []string{"thinking", "output_config"} {
+		if object := gjson.GetBytes(body, path); object.Exists() && object.IsObject() && len(object.Map()) == 0 {
+			body, _ = sjson.DeleteBytes(body, path)
+		}
+	}
+	return body
+}
+
 func enableClaudeThinkingForSummary(body []byte, model string, resolvedModelInfo *registry.ModelInfo) []byte {
 	modelInfo := resolvedModelInfo
 	if modelInfo == nil {
diff --git a/sdk/translator/registry.go b/sdk/translator/registry.go
index 830d0355..6e9f0eed 100644
--- a/sdk/translator/registry.go
+++ b/sdk/translator/registry.go
@@ -57,8 +57,6 @@ func (r *Registry) SetPluginHooks(hooks PluginHooks) {
 // "model" field is still updated to match the resolved model name so that
 // client-side prefixes (e.g. "copilot/gpt-5-mini") are not leaked upstream.
 func (r *Registry) TranslateRequest(from, to Format, model string, rawJSON []byte, stream bool) []byte {
-	summaryConfig := thinking.ExtractSummaryConfig(rawJSON, from.String())
-
 	r.mu.RLock()
 	var fn RequestTransform
 	if byTarget, ok := r.requests[from]; ok {
@@ -69,6 +67,7 @@ func (r *Registry) TranslateRequest(from, to Format, model string, rawJSON []byt
 
 	body := rawJSON
 	if fn != nil {
+		summaryConfig := thinking.ExtractSummaryConfig(rawJSON, from.String())
 		body = fn(model, body, stream)
 		body = thinking.ApplySummaryConfigForModel(body, to.String(), model, summaryConfig)
 		if hooks != nil {
@@ -93,8 +92,10 @@ func (r *Registry) TranslateRequest(from, to Format, model string, rawJSON []byt
 	}
 
 	// Plugin request normalizers canonicalize the source before a plugin request
-	// translator gets a chance to handle a missing native route.
+	// translator gets a chance to handle a missing native route. Extract summary
+	// intent from that normalized source so a normalizer can remove or rewrite it.
 	body = hooks.NormalizeRequest(context.Background(), from, to, model, body, stream)
+	summaryConfig := thinking.ExtractSummaryConfig(body, from.String())
 	if translated, ok := hooks.TranslateRequest(context.Background(), from, to, model, body, stream); ok {
 		body = thinking.ApplySummaryConfigForModel(translated, to.String(), model, summaryConfig)
 	}
diff --git a/sdk/translator/registry_summary_test.go b/sdk/translator/registry_summary_test.go
index 1b77b951..16b03216 100644
--- a/sdk/translator/registry_summary_test.go
+++ b/sdk/translator/registry_summary_test.go
@@ -178,6 +178,59 @@ func TestRegistryTranslateRequestAppliesSummaryAfterPluginTranslation(t *testing
 	}
 }
 
+func TestRegistryTranslateRequestPluginNormalizerOwnsSourceSummaryIntent(t *testing.T) {
+	tests := []struct {
+		name       string
+		normalize  func([]byte) []byte
+		wantExists bool
+		want       bool
+	}{
+		{
+			name: "removed summary remains absent",
+			normalize: func(body []byte) []byte {
+				out, _ := sjson.DeleteBytes(body, "reasoning.summary")
+				return out
+			},
+		},
+		{
+			name: "disabled summary replaces enabled intent",
+			normalize: func(body []byte) []byte {
+				out, _ := sjson.SetBytes(body, "reasoning.summary", nil)
+				return out
+			},
+			wantExists: true,
+			want:       false,
+		},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			registry := NewRegistry()
+			hooks := &fakePluginHooks{
+				normalizeRequest:     test.normalize,
+				requestTranslateBody: []byte(`{"generationConfig":{"thinkingConfig":{"thinkingLevel":"high"}}}`),
+				requestTranslateOK:   true,
+			}
+			registry.SetPluginHooks(hooks)
+
+			out := registry.TranslateRequest(
+				FormatOpenAIResponse,
+				FormatGemini,
+				"gemini-3.6-flash",
+				[]byte(`{"reasoning":{"summary":"auto"},"input":"hi"}`),
+				false,
+			)
+			result := gjson.GetBytes(out, "generationConfig.thinkingConfig.includeThoughts")
+			if result.Exists() != test.wantExists {
+				t.Fatalf("includeThoughts exists = %v, want %v; body=%s", result.Exists(), test.wantExists, out)
+			}
+			if test.wantExists && result.Bool() != test.want {
+				t.Fatalf("includeThoughts = %v, want %v; body=%s", result.Bool(), test.want, out)
+			}
+		})
+	}
+}
+
 func TestRegistryTranslateRequestNormalizerOwnsFinalSummaryField(t *testing.T) {
 	registry := NewRegistry()
 	registry.Register(FormatOpenAIResponse, FormatGemini, func(_ string, _ []byte, _ bool) []byte {
-- 
2.51.2


From 24323ee4b7c33ec7268abf15728fe712888f90c4 Mon Sep 17 00:00:00 2001
From: sususu <suchangshan@foxmail.com>
Date: Fri, 31 Jul 2026 13:40:41 +0800
Subject: [PATCH 9/9] fix(thinking): drop disabled Interactions summaries

---
 internal/thinking/provider/interactions/apply.go | 5 ++++-
 test/summary_intent_translation_test.go          | 1 +
 2 files changed, 5 insertions(+), 1 deletion(-)

diff --git a/internal/thinking/provider/interactions/apply.go b/internal/thinking/provider/interactions/apply.go
index c644f5ad..b23f0d74 100644
--- a/internal/thinking/provider/interactions/apply.go
+++ b/internal/thinking/provider/interactions/apply.go
@@ -77,7 +77,10 @@ func applyInteractionsNone(result, original []byte, config thinking.ThinkingConf
 	if config.Budget > 0 {
 		return applyInteractionsBudget(result, original, config.Budget, modelInfo)
 	}
-	return setInteractionsThinkingSummaries(result, original)
+	// With the amount fully disabled, visibility is irrelevant. Restoring
+	// thinking_summaries alone could make a default-on model reason and return a
+	// summary despite the explicit none override.
+	return result
 }
 
 func stripInteractionsThinkingFields(body []byte) []byte {
diff --git a/test/summary_intent_translation_test.go b/test/summary_intent_translation_test.go
index cc1724f5..b19f0129 100644
--- a/test/summary_intent_translation_test.go
+++ b/test/summary_intent_translation_test.go
@@ -144,6 +144,7 @@ func TestSummaryIntentFinalPipeline(t *testing.T) {
 		{name: "Summary-only control is stripped for non-thinking Gemini model", from: sdktranslator.FormatOpenAIResponse, to: sdktranslator.FormatGemini, model: "no-thinking-model", body: `{"model":"no-thinking-model","reasoning":{"summary":"auto"},"input":"hi"}`, path: "generationConfig.thinkingConfig"},
 		{name: "Interactions level alone keeps summaries omitted", from: sdktranslator.FormatInteractions, to: sdktranslator.FormatInteractions, model: "level-model", body: `{"model":"level-model","generation_config":{"thinking_level":"high"},"input":"hi"}`, path: "generation_config.thinking_summaries"},
 		{name: "Interactions auto survives its applier", from: sdktranslator.FormatInteractions, to: sdktranslator.FormatInteractions, model: "level-model", body: `{"model":"level-model","generation_config":{"thinking_level":"high","thinking_summaries":"auto"},"input":"hi"}`, path: "generation_config.thinking_summaries", want: "auto", wantExists: true},
+		{name: "Interactions suffix none removes summary visibility", from: sdktranslator.FormatInteractions, to: sdktranslator.FormatInteractions, model: "gemini-toggle-mixed-model(none)", body: `{"model":"gemini-toggle-mixed-model(none)","generation_config":{"thinking_summaries":"auto"},"input":"hi"}`, path: "generation_config.thinking_summaries"},
 		{name: "Interactions reasoning effort leaves Antigravity summaries unspecified", from: sdktranslator.FormatInteractions, to: sdktranslator.FormatAntigravity, model: "antigravity-budget-model", body: `{"model":"antigravity-budget-model","reasoning":{"effort":"high"},"input":"hi"}`, path: "request.generationConfig.thinkingConfig.includeThoughts"},
 		{name: "Interactions reasoning summary auto reaches Antigravity", from: sdktranslator.FormatInteractions, to: sdktranslator.FormatAntigravity, model: "antigravity-budget-model", body: `{"model":"antigravity-budget-model","reasoning":{"effort":"high","summary":"auto"},"input":"hi"}`, path: "request.generationConfig.thinkingConfig.includeThoughts", want: "true", wantExists: true},
 		{name: "Interactions reasoning summary none reaches Antigravity", from: sdktranslator.FormatInteractions, to: sdktranslator.FormatAntigravity, model: "antigravity-budget-model", body: `{"model":"antigravity-budget-model","reasoning":{"effort":"high","summary":"none"},"input":"hi"}`, path: "request.generationConfig.thinkingConfig.includeThoughts", want: "false", wantExists: true},
-- 
2.51.2