Add projects

backend/internal/thinking/provider/antigravity/apply.go

@@ -0,0 +1,220 @@
+// Package antigravity implements thinking configuration for Antigravity API format.
+//
+// Antigravity uses request.generationConfig.thinkingConfig.* path.
+// but requires additional normalization for Claude models:
+//   - Ensure thinking budget < max_tokens
+//   - Remove thinkingConfig if budget < minimum allowed
+package antigravity
+
+import (
+	"strings"
+
+	"github.com/router-for-me/CLIProxyAPI/v7/internal/registry"
+	"github.com/router-for-me/CLIProxyAPI/v7/internal/thinking"
+	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
+)
+
+// Applier applies thinking configuration for Antigravity API format.
+type Applier struct{}
+
+var _ thinking.ProviderApplier = (*Applier)(nil)
+
+// NewApplier creates a new Antigravity thinking applier.
+func NewApplier() *Applier {
+	return &Applier{}
+}
+
+func init() {
+	thinking.RegisterProvider("antigravity", NewApplier())
+}
+
+// Apply applies thinking configuration to Antigravity request body.
+//
+// For Claude models, additional constraints are applied:
+//   - Ensure thinking budget < max_tokens
+//   - Remove thinkingConfig if budget < minimum allowed
+func (a *Applier) Apply(body []byte, config thinking.ThinkingConfig, modelInfo *registry.ModelInfo) ([]byte, error) {
+	if thinking.IsUserDefinedModel(modelInfo) {
+		return a.applyCompatible(body, config, modelInfo)
+	}
+	if modelInfo.Thinking == nil {
+		return body, nil
+	}
+
+	if config.Mode != thinking.ModeBudget && config.Mode != thinking.ModeLevel && config.Mode != thinking.ModeNone && config.Mode != thinking.ModeAuto {
+		return body, nil
+	}
+
+	if len(body) == 0 || !gjson.ValidBytes(body) {
+		body = []byte(`{}`)
+	}
+
+	isClaude := strings.Contains(strings.ToLower(modelInfo.ID), "claude")
+
+	// ModeAuto: Always use Budget format with thinkingBudget=-1
+	if config.Mode == thinking.ModeAuto {
+		return a.applyBudgetFormat(body, config, modelInfo, isClaude)
+	}
+	if config.Mode == thinking.ModeBudget {
+		return a.applyBudgetFormat(body, config, modelInfo, isClaude)
+	}
+
+	// For non-auto modes, choose format based on model capabilities
+	support := modelInfo.Thinking
+	if len(support.Levels) > 0 {
+		return a.applyLevelFormat(body, config)
+	}
+	return a.applyBudgetFormat(body, config, modelInfo, isClaude)
+}
+
+func (a *Applier) applyCompatible(body []byte, config thinking.ThinkingConfig, modelInfo *registry.ModelInfo) ([]byte, error) {
+	if config.Mode != thinking.ModeBudget && config.Mode != thinking.ModeLevel && config.Mode != thinking.ModeNone && config.Mode != thinking.ModeAuto {
+		return body, nil
+	}
+
+	if len(body) == 0 || !gjson.ValidBytes(body) {
+		body = []byte(`{}`)
+	}
+
+	isClaude := false
+	if modelInfo != nil {
+		isClaude = strings.Contains(strings.ToLower(modelInfo.ID), "claude")
+	}
+
+	if config.Mode == thinking.ModeAuto {
+		return a.applyBudgetFormat(body, config, modelInfo, isClaude)
+	}
+
+	if config.Mode == thinking.ModeLevel || (config.Mode == thinking.ModeNone && config.Level != "") {
+		return a.applyLevelFormat(body, config)
+	}
+
+	return a.applyBudgetFormat(body, config, modelInfo, isClaude)
+}
+
+func (a *Applier) applyLevelFormat(body []byte, config thinking.ThinkingConfig) ([]byte, error) {
+	// Remove conflicting fields to avoid both thinkingLevel and thinkingBudget in output
+	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 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 == "" {
+			// 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 result, nil
+		}
+		if config.Level != "" {
+			result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.thinkingLevel", string(config.Level))
+		}
+		return applyAntigravityIncludeThoughts(result, body), nil
+	}
+
+	// Only handle ModeLevel - budget conversion should be done by upper layer
+	if config.Mode != thinking.ModeLevel {
+		return body, nil
+	}
+
+	level := string(config.Level)
+	result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.thinkingLevel", level)
+	return applyAntigravityIncludeThoughts(result, body), nil
+}
+
+func (a *Applier) applyBudgetFormat(body []byte, config thinking.ThinkingConfig, modelInfo *registry.ModelInfo, isClaude bool) ([]byte, error) {
+	// Remove conflicting fields to avoid both thinkingLevel and thinkingBudget in output
+	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 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
+
+	// Apply Claude-specific constraints first to get the final budget value
+	if isClaude && modelInfo != nil {
+		budget, result = a.normalizeClaudeBudget(budget, result, modelInfo)
+		// Check if the thinking amount was removed entirely. Summary visibility is
+		// independent, so retain an explicit includeThoughts control if present.
+		if budget == -2 {
+			return applyAntigravityIncludeThoughts(result, body), nil
+		}
+	}
+
+	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
+		}
+	}
+	return result
+}
+
+// normalizeClaudeBudget applies Claude-specific constraints to thinking budget.
+//
+// It handles:
+//   - Ensuring thinking budget < max_tokens
+//   - Removing thinkingConfig if budget < minimum allowed
+//
+// Returns the normalized budget and updated payload.
+// Returns budget=-2 as a sentinel indicating thinkingConfig was removed entirely.
+func (a *Applier) normalizeClaudeBudget(budget int, payload []byte, modelInfo *registry.ModelInfo) (int, []byte) {
+	if modelInfo == nil {
+		return budget, payload
+	}
+
+	// Get effective max tokens
+	effectiveMax, setDefaultMax := a.effectiveMaxTokens(payload, modelInfo)
+	if effectiveMax > 0 && budget >= effectiveMax {
+		budget = effectiveMax - 1
+	}
+
+	// Check minimum budget
+	minBudget := 0
+	if modelInfo.Thinking != nil {
+		minBudget = modelInfo.Thinking.Min
+	}
+	if minBudget > 0 && budget >= 0 && budget < minBudget {
+		// Budget is below minimum, remove thinking config entirely
+		payload, _ = sjson.DeleteBytes(payload, "request.generationConfig.thinkingConfig")
+		return -2, payload
+	}
+
+	// Set default max tokens if needed
+	if setDefaultMax && effectiveMax > 0 {
+		payload, _ = sjson.SetBytes(payload, "request.generationConfig.maxOutputTokens", effectiveMax)
+	}
+
+	return budget, payload
+}
+
+// effectiveMaxTokens returns the max tokens to cap thinking:
+// prefer request-provided maxOutputTokens; otherwise fall back to model default.
+// The boolean indicates whether the value came from the model default (and thus should be written back).
+func (a *Applier) effectiveMaxTokens(payload []byte, modelInfo *registry.ModelInfo) (max int, fromModel bool) {
+	if maxTok := gjson.GetBytes(payload, "request.generationConfig.maxOutputTokens"); maxTok.Exists() && maxTok.Int() > 0 {
+		return int(maxTok.Int()), false
+	}
+	if modelInfo != nil && modelInfo.MaxCompletionTokens > 0 {
+		return modelInfo.MaxCompletionTokens, true
+	}
+	return 0, false
+}

backend/internal/thinking/provider/claude/apply.go

@@ -0,0 +1,270 @@
+// Package claude implements thinking configuration scaffolding for Claude models.
+//
+// Claude models support two thinking control styles:
+//   - Manual thinking: thinking.type="enabled" with thinking.budget_tokens (token budget)
+//   - Adaptive thinking (Claude 4.6): thinking.type="adaptive" with output_config.effort (low/medium/high/max)
+//
+// Some Claude models support ZeroAllowed (sonnet-4-5, opus-4-5), while older models do not.
+// See: _bmad-output/planning-artifacts/architecture.md#Epic-6
+package claude
+
+import (
+	"github.com/router-for-me/CLIProxyAPI/v7/internal/registry"
+	"github.com/router-for-me/CLIProxyAPI/v7/internal/thinking"
+	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
+)
+
+// Applier implements thinking.ProviderApplier for Claude models.
+// This applier is stateless and holds no configuration.
+type Applier struct{}
+
+// NewApplier creates a new Claude thinking applier.
+func NewApplier() *Applier {
+	return &Applier{}
+}
+
+func init() {
+	thinking.RegisterProvider("claude", NewApplier())
+}
+
+// Apply applies thinking configuration to Claude request body.
+//
+// IMPORTANT: This method expects config to be pre-validated by thinking.ValidateConfig.
+// ValidateConfig handles:
+//   - Mode conversion (Level→Budget, Auto→Budget)
+//   - Budget clamping to model range
+//   - ZeroAllowed constraint enforcement
+//
+// Apply processes:
+//   - ModeBudget: manual thinking budget_tokens
+//   - ModeLevel: adaptive thinking effort (Claude 4.6)
+//   - ModeAuto: provider default adaptive/manual behavior
+//   - ModeNone: disabled
+//
+// Expected output format when enabled:
+//
+//	{
+//	  "thinking": {
+//	    "type": "enabled",
+//	    "budget_tokens": 16384
+//	  }
+//	}
+//
+// Expected output format for adaptive:
+//
+//	{
+//	  "thinking": {
+//	    "type": "adaptive"
+//	  },
+//	  "output_config": {
+//	    "effort": "high"
+//	  }
+//	}
+//
+// Expected output format when disabled:
+//
+//	{
+//	  "thinking": {
+//	    "type": "disabled"
+//	  }
+//	}
+func (a *Applier) Apply(body []byte, config thinking.ThinkingConfig, modelInfo *registry.ModelInfo) ([]byte, error) {
+	if thinking.IsUserDefinedModel(modelInfo) {
+		return applyCompatibleClaude(body, config)
+	}
+	if modelInfo.Thinking == nil {
+		return body, nil
+	}
+
+	if len(body) == 0 || !gjson.ValidBytes(body) {
+		body = []byte(`{}`)
+	}
+
+	supportsAdaptive := modelInfo != nil && modelInfo.Thinking != nil && len(modelInfo.Thinking.Levels) > 0
+
+	switch config.Mode {
+	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")
+		}
+		return result, nil
+
+	case thinking.ModeLevel:
+		// Adaptive thinking effort is only valid when the model advertises discrete levels.
+		// (Claude 4.6 uses output_config.effort.)
+		if supportsAdaptive && config.Level != "" {
+			result, _ := sjson.SetBytes(body, "thinking.type", "adaptive")
+			result, _ = sjson.DeleteBytes(result, "thinking.budget_tokens")
+			result, _ = sjson.SetBytes(result, "output_config.effort", string(config.Level))
+			return result, nil
+		}
+
+		// Fallback for non-adaptive Claude models: convert level to budget_tokens.
+		if budget, ok := thinking.ConvertLevelToBudget(string(config.Level)); ok {
+			config.Mode = thinking.ModeBudget
+			config.Budget = budget
+			config.Level = ""
+		} else {
+			return body, nil
+		}
+		fallthrough
+
+	case thinking.ModeBudget:
+		// Budget is expected to be pre-validated by ValidateConfig (clamped, ZeroAllowed enforced).
+		// Decide enabled/disabled based on budget value.
+		if config.Budget == 0 {
+			result, _ := sjson.SetBytes(body, "thinking.type", "disabled")
+			result, _ = sjson.DeleteBytes(result, "thinking.budget_tokens")
+			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")
+			}
+			return result, nil
+		}
+
+		result, _ := sjson.SetBytes(body, "thinking.type", "enabled")
+		result, _ = sjson.SetBytes(result, "thinking.budget_tokens", config.Budget)
+		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")
+		}
+
+		// Ensure max_tokens > thinking.budget_tokens (Anthropic API constraint).
+		result = a.normalizeClaudeBudget(result, config.Budget, modelInfo)
+		return result, nil
+
+	case thinking.ModeAuto:
+		// For Claude 4.6 models, auto maps to adaptive thinking with upstream defaults.
+		if supportsAdaptive {
+			result, _ := sjson.SetBytes(body, "thinking.type", "adaptive")
+			result, _ = sjson.DeleteBytes(result, "thinking.budget_tokens")
+			// Explicit effort is optional for adaptive thinking; omit it to allow upstream default.
+			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")
+			}
+			return result, nil
+		}
+
+		// Legacy fallback: enable thinking without specifying budget_tokens.
+		result, _ := sjson.SetBytes(body, "thinking.type", "enabled")
+		result, _ = sjson.DeleteBytes(result, "thinking.budget_tokens")
+		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")
+		}
+		return result, nil
+
+	default:
+		return body, nil
+	}
+}
+
+// normalizeClaudeBudget applies Claude-specific constraints to ensure max_tokens > budget_tokens.
+// Anthropic API requires this constraint; violating it returns a 400 error.
+func (a *Applier) normalizeClaudeBudget(body []byte, budgetTokens int, modelInfo *registry.ModelInfo) []byte {
+	if budgetTokens <= 0 {
+		return body
+	}
+
+	// Ensure the request satisfies Claude constraints:
+	//  1) Determine effective max_tokens (request overrides model default)
+	//  2) If budget_tokens >= max_tokens, reduce budget_tokens to max_tokens-1
+	//  3) If the adjusted budget falls below the model minimum, leave the request unchanged
+	//  4) If max_tokens came from model default, write it back into the request
+
+	effectiveMax, setDefaultMax := a.effectiveMaxTokens(body, modelInfo)
+	if setDefaultMax && effectiveMax > 0 {
+		body, _ = sjson.SetBytes(body, "max_tokens", effectiveMax)
+	}
+
+	// Compute the budget we would apply after enforcing budget_tokens < max_tokens.
+	adjustedBudget := budgetTokens
+	if effectiveMax > 0 && adjustedBudget >= effectiveMax {
+		adjustedBudget = effectiveMax - 1
+	}
+
+	minBudget := 0
+	if modelInfo != nil && modelInfo.Thinking != nil {
+		minBudget = modelInfo.Thinking.Min
+	}
+	if minBudget > 0 && adjustedBudget > 0 && adjustedBudget < minBudget {
+		// If enforcing the max_tokens constraint would push the budget below the model minimum,
+		// leave the request unchanged.
+		return body
+	}
+
+	if adjustedBudget != budgetTokens {
+		body, _ = sjson.SetBytes(body, "thinking.budget_tokens", adjustedBudget)
+	}
+
+	return body
+}
+
+// effectiveMaxTokens returns the max tokens to cap thinking:
+// prefer request-provided max_tokens; otherwise fall back to model default.
+// The boolean indicates whether the value came from the model default (and thus should be written back).
+func (a *Applier) effectiveMaxTokens(body []byte, modelInfo *registry.ModelInfo) (max int, fromModel bool) {
+	if maxTok := gjson.GetBytes(body, "max_tokens"); maxTok.Exists() && maxTok.Int() > 0 {
+		return int(maxTok.Int()), false
+	}
+	if modelInfo != nil && modelInfo.MaxCompletionTokens > 0 {
+		return modelInfo.MaxCompletionTokens, true
+	}
+	return 0, false
+}
+
+func applyCompatibleClaude(body []byte, config thinking.ThinkingConfig) ([]byte, error) {
+	if config.Mode != thinking.ModeBudget && config.Mode != thinking.ModeNone && config.Mode != thinking.ModeAuto && config.Mode != thinking.ModeLevel {
+		return body, nil
+	}
+
+	if len(body) == 0 || !gjson.ValidBytes(body) {
+		body = []byte(`{}`)
+	}
+
+	switch config.Mode {
+	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")
+		}
+		return result, nil
+	case thinking.ModeAuto:
+		result, _ := sjson.SetBytes(body, "thinking.type", "enabled")
+		result, _ = sjson.DeleteBytes(result, "thinking.budget_tokens")
+		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")
+		}
+		return result, nil
+	case thinking.ModeLevel:
+		// For user-defined models, interpret ModeLevel as Claude adaptive thinking effort.
+		// Upstream is responsible for validating whether the target model supports it.
+		if config.Level == "" {
+			return body, nil
+		}
+		result, _ := sjson.SetBytes(body, "thinking.type", "adaptive")
+		result, _ = sjson.DeleteBytes(result, "thinking.budget_tokens")
+		result, _ = sjson.SetBytes(result, "output_config.effort", string(config.Level))
+		return result, nil
+	default:
+		result, _ := sjson.SetBytes(body, "thinking.type", "enabled")
+		result, _ = sjson.SetBytes(result, "thinking.budget_tokens", config.Budget)
+		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")
+		}
+		return result, nil
+	}
+}

backend/internal/thinking/provider/codex/apply.go

@@ -0,0 +1,120 @@
+// Package codex implements thinking configuration for Codex (OpenAI Responses API) models.
+//
+// Codex models use the reasoning.effort format with discrete levels
+// (low/medium/high). This is similar to OpenAI but uses nested field
+// "reasoning.effort" instead of "reasoning_effort".
+// See: _bmad-output/planning-artifacts/architecture.md#Epic-8
+package codex
+
+import (
+	"github.com/router-for-me/CLIProxyAPI/v7/internal/registry"
+	"github.com/router-for-me/CLIProxyAPI/v7/internal/thinking"
+	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
+)
+
+// Applier implements thinking.ProviderApplier for Codex models.
+//
+// Codex-specific behavior:
+//   - Output format: reasoning.effort (string: low/medium/high/xhigh)
+//   - Level-only mode: no numeric budget support
+//   - Some models support ZeroAllowed (gpt-5.1, gpt-5.2)
+type Applier struct{}
+
+var _ thinking.ProviderApplier = (*Applier)(nil)
+
+// NewApplier creates a new Codex thinking applier.
+func NewApplier() *Applier {
+	return &Applier{}
+}
+
+func init() {
+	thinking.RegisterProvider("codex", NewApplier())
+}
+
+// Apply applies thinking configuration to Codex request body.
+//
+// Expected output format:
+//
+//	{
+//	  "reasoning": {
+//	    "effort": "high"
+//	  }
+//	}
+func (a *Applier) Apply(body []byte, config thinking.ThinkingConfig, modelInfo *registry.ModelInfo) ([]byte, error) {
+	if thinking.IsUserDefinedModel(modelInfo) {
+		return applyCompatibleCodex(body, config)
+	}
+	if modelInfo.Thinking == nil {
+		return body, nil
+	}
+
+	// Only handle ModeLevel and ModeNone; other modes pass through unchanged.
+	if config.Mode != thinking.ModeLevel && config.Mode != thinking.ModeNone {
+		return body, nil
+	}
+
+	if len(body) == 0 || !gjson.ValidBytes(body) {
+		body = []byte(`{}`)
+	}
+
+	if config.Mode == thinking.ModeLevel {
+		result, _ := sjson.SetBytes(body, "reasoning.effort", string(config.Level))
+		return result, nil
+	}
+
+	effort := ""
+	support := modelInfo.Thinking
+	if config.Budget == 0 {
+		if support.ZeroAllowed || thinking.HasLevel(support.Levels, string(thinking.LevelNone)) {
+			effort = string(thinking.LevelNone)
+		}
+	}
+	if effort == "" && config.Level != "" {
+		effort = string(config.Level)
+	}
+	if effort == "" && len(support.Levels) > 0 {
+		effort = support.Levels[0]
+	}
+	if effort == "" {
+		return body, nil
+	}
+
+	result, _ := sjson.SetBytes(body, "reasoning.effort", effort)
+	return result, nil
+}
+
+func applyCompatibleCodex(body []byte, config thinking.ThinkingConfig) ([]byte, error) {
+	if len(body) == 0 || !gjson.ValidBytes(body) {
+		body = []byte(`{}`)
+	}
+
+	var effort string
+	switch config.Mode {
+	case thinking.ModeLevel:
+		if config.Level == "" {
+			return body, nil
+		}
+		effort = string(config.Level)
+	case thinking.ModeNone:
+		effort = string(thinking.LevelNone)
+		if config.Level != "" {
+			effort = string(config.Level)
+		}
+	case thinking.ModeAuto:
+		// Auto mode for user-defined models: pass through as "auto"
+		effort = string(thinking.LevelAuto)
+	case thinking.ModeBudget:
+		// Budget mode: convert budget to level using threshold mapping
+		level, ok := thinking.ConvertBudgetToLevel(config.Budget)
+		if !ok {
+			return body, nil
+		}
+		effort = level
+	default:
+		return body, nil
+	}
+
+	result, _ := sjson.SetBytes(body, "reasoning.effort", effort)
+	return result, nil
+}

backend/internal/thinking/provider/gemini/apply.go

@@ -0,0 +1,182 @@
+// Package gemini implements thinking configuration for Gemini models.
+//
+// Gemini models have two formats:
+//   - Gemini 2.5: Uses thinkingBudget (numeric)
+//   - Gemini 3.x: Uses thinkingLevel (string: minimal/low/medium/high)
+//     or thinkingBudget=-1 for auto/dynamic mode
+//
+// Output format is determined by ThinkingConfig.Mode and ThinkingSupport.Levels:
+//   - ModeAuto: Always uses thinkingBudget=-1 (both Gemini 2.5 and 3.x)
+//   - len(Levels) > 0: Uses thinkingLevel (Gemini 3.x discrete levels)
+//   - len(Levels) == 0: Uses thinkingBudget (Gemini 2.5)
+package gemini
+
+import (
+	"github.com/router-for-me/CLIProxyAPI/v7/internal/registry"
+	"github.com/router-for-me/CLIProxyAPI/v7/internal/thinking"
+	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
+)
+
+// Applier applies thinking configuration for Gemini models.
+//
+// Gemini-specific behavior:
+//   - Gemini 2.5: thinkingBudget format, flash series supports ZeroAllowed
+//   - Gemini 3.x: thinkingLevel format, disable by removing thinkingConfig when zero is allowed
+//   - Use ThinkingSupport.Levels to decide output format
+type Applier struct{}
+
+// NewApplier creates a new Gemini thinking applier.
+func NewApplier() *Applier {
+	return &Applier{}
+}
+
+func init() {
+	thinking.RegisterProvider("gemini", NewApplier())
+}
+
+// Apply applies thinking configuration to Gemini request body.
+//
+// Expected output format (Gemini 2.5):
+//
+//	{
+//	  "generationConfig": {
+//	    "thinkingConfig": {
+//	      "thinkingBudget": 8192,
+//	      "includeThoughts": true
+//	    }
+//	  }
+//	}
+//
+// Expected output format (Gemini 3.x):
+//
+//	{
+//	  "generationConfig": {
+//	    "thinkingConfig": {
+//	      "thinkingLevel": "high",
+//	      "includeThoughts": true
+//	    }
+//	  }
+//	}
+func (a *Applier) Apply(body []byte, config thinking.ThinkingConfig, modelInfo *registry.ModelInfo) ([]byte, error) {
+	if thinking.IsUserDefinedModel(modelInfo) {
+		return a.applyCompatible(body, config)
+	}
+	if modelInfo.Thinking == nil {
+		return body, nil
+	}
+
+	if config.Mode != thinking.ModeBudget && config.Mode != thinking.ModeLevel && config.Mode != thinking.ModeNone && config.Mode != thinking.ModeAuto {
+		return body, nil
+	}
+
+	if len(body) == 0 || !gjson.ValidBytes(body) {
+		body = []byte(`{}`)
+	}
+
+	// Choose format based on config.Mode and model capabilities:
+	// - ModeLevel: use Level format (validation will reject unsupported levels)
+	// - ModeNone: use Level format if model has Levels, else Budget format
+	// - ModeBudget/ModeAuto: use Budget format
+	switch config.Mode {
+	case thinking.ModeLevel:
+		return a.applyLevelFormat(body, config)
+	case thinking.ModeNone:
+		// ModeNone: route based on model capability (has Levels or not)
+		if len(modelInfo.Thinking.Levels) > 0 {
+			return a.applyLevelFormat(body, config)
+		}
+		return a.applyBudgetFormat(body, config)
+	default:
+		return a.applyBudgetFormat(body, config)
+	}
+}
+
+func (a *Applier) applyCompatible(body []byte, config thinking.ThinkingConfig) ([]byte, error) {
+	if config.Mode != thinking.ModeBudget && config.Mode != thinking.ModeLevel && config.Mode != thinking.ModeNone && config.Mode != thinking.ModeAuto {
+		return body, nil
+	}
+
+	if len(body) == 0 || !gjson.ValidBytes(body) {
+		body = []byte(`{}`)
+	}
+
+	if config.Mode == thinking.ModeAuto {
+		return a.applyBudgetFormat(body, config)
+	}
+
+	if config.Mode == thinking.ModeLevel || (config.Mode == thinking.ModeNone && config.Level != "") {
+		return a.applyLevelFormat(body, config)
+	}
+
+	return a.applyBudgetFormat(body, config)
+}
+
+func (a *Applier) applyLevelFormat(body []byte, config thinking.ThinkingConfig) ([]byte, error) {
+	// ModeNone semantics:
+	//   - 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 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 == "" {
+			// 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 result, nil
+		}
+		if config.Level != "" {
+			result, _ = sjson.SetBytes(result, "generationConfig.thinkingConfig.thinkingLevel", string(config.Level))
+		}
+		return applyGeminiIncludeThoughts(result, body), nil
+	}
+
+	// Only handle ModeLevel - budget conversion should be done by upper layer
+	if config.Mode != thinking.ModeLevel {
+		return body, nil
+	}
+
+	level := string(config.Level)
+	result, _ = sjson.SetBytes(result, "generationConfig.thinkingConfig.thinkingLevel", level)
+	return applyGeminiIncludeThoughts(result, body), nil
+}
+
+func (a *Applier) applyBudgetFormat(body []byte, config thinking.ThinkingConfig) ([]byte, error) {
+	// Remove conflicting fields to avoid both thinkingLevel and thinkingBudget in output
+	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 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
+}
+
+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
+		}
+	}
+	return result
+}

backend/internal/thinking/provider/interactions/apply.go

@@ -0,0 +1,178 @@
+// Package interactions applies native Interactions thinking configuration.
+package interactions
+
+import (
+	"strings"
+
+	"github.com/router-for-me/CLIProxyAPI/v7/internal/registry"
+	"github.com/router-for-me/CLIProxyAPI/v7/internal/thinking"
+	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
+)
+
+// Applier implements thinking.ProviderApplier for the native Interactions API.
+type Applier struct{}
+
+// NewApplier creates a new Interactions thinking applier.
+func NewApplier() *Applier {
+	return &Applier{}
+}
+
+func init() {
+	thinking.RegisterProvider("interactions", NewApplier())
+}
+
+// Apply writes thinking configuration using native Interactions generation_config fields.
+func (a *Applier) Apply(body []byte, config thinking.ThinkingConfig, modelInfo *registry.ModelInfo) ([]byte, error) {
+	if config.Mode != thinking.ModeBudget && config.Mode != thinking.ModeLevel && config.Mode != thinking.ModeNone && config.Mode != thinking.ModeAuto {
+		return body, nil
+	}
+	if len(body) == 0 || !gjson.ValidBytes(body) {
+		body = []byte(`{}`)
+	}
+
+	result := stripInteractionsThinkingFields(body)
+	switch config.Mode {
+	case thinking.ModeLevel:
+		return applyInteractionsLevel(result, body, string(config.Level), modelInfo), nil
+	case thinking.ModeBudget:
+		return applyInteractionsBudget(result, body, config.Budget, modelInfo), nil
+	case thinking.ModeAuto:
+		return setInteractionsThinkingSummaries(result, body), nil
+	case thinking.ModeNone:
+		return applyInteractionsNone(result, body, config, modelInfo), nil
+	default:
+		return body, nil
+	}
+}
+
+func applyInteractionsBudget(result, original []byte, budget int, modelInfo *registry.ModelInfo) []byte {
+	level, ok := thinking.ConvertBudgetToLevel(budget)
+	if !ok {
+		return setInteractionsThinkingSummaries(result, original)
+	}
+	switch level {
+	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)
+	}
+}
+
+func applyInteractionsLevel(result, original []byte, level string, modelInfo *registry.ModelInfo) []byte {
+	level = normalizeInteractionsLevel(level, modelInfo)
+	if level != "" {
+		result, _ = sjson.SetBytes(result, "generation_config.thinking_level", level)
+	}
+	return setInteractionsThinkingSummaries(result, original)
+}
+
+func applyInteractionsNone(result, original []byte, config thinking.ThinkingConfig, modelInfo *registry.ModelInfo) []byte {
+	if config.Level != "" {
+		return applyInteractionsLevel(result, original, string(config.Level), modelInfo)
+	}
+	if config.Budget > 0 {
+		return applyInteractionsBudget(result, original, config.Budget, modelInfo)
+	}
+	// 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 {
+	result := body
+	for _, path := range []string{
+		"generation_config.thinking_level",
+		"generation_config.thinkingLevel",
+		"generation_config.thinking_budget",
+		"generation_config.thinkingBudget",
+		"generation_config.thinking_summaries",
+		"generation_config.thinkingSummaries",
+		"generation_config.thinking_config",
+		"generation_config.thinkingConfig",
+		"generationConfig.thinkingLevel",
+		"generationConfig.thinking_level",
+		"generationConfig.thinkingBudget",
+		"generationConfig.thinking_budget",
+		"generationConfig.thinkingSummaries",
+		"generationConfig.thinking_summaries",
+		"generationConfig.thinkingConfig",
+	} {
+		result, _ = sjson.DeleteBytes(result, path)
+	}
+	return result
+}
+
+func setInteractionsThinkingSummaries(result, original []byte) []byte {
+	if value, okValue := originalInteractionsThinkingSummaries(original); okValue {
+		result, _ = sjson.SetBytes(result, "generation_config.thinking_summaries", value)
+		return result
+	}
+	if includeThoughts, okValue := originalInteractionsIncludeThoughts(original); okValue {
+		value := "none"
+		if includeThoughts {
+			value = "auto"
+		}
+		result, _ = sjson.SetBytes(result, "generation_config.thinking_summaries", value)
+	}
+	return result
+}
+
+func originalInteractionsThinkingSummaries(body []byte) (string, bool) {
+	for _, path := range []string{
+		"generation_config.thinking_summaries",
+		"generation_config.thinkingSummaries",
+	} {
+		value := gjson.GetBytes(body, path)
+		if value.Type != gjson.String {
+			continue
+		}
+		switch normalized := strings.ToLower(strings.TrimSpace(value.String())); normalized {
+		case "auto", "none":
+			return normalized, true
+		}
+	}
+	return "", false
+}
+
+func originalInteractionsIncludeThoughts(body []byte) (bool, bool) {
+	for _, path := range []string{
+		"generation_config.thinking_config.include_thoughts",
+		"generation_config.thinking_config.includeThoughts",
+		"generation_config.thinkingConfig.include_thoughts",
+		"generation_config.thinkingConfig.includeThoughts",
+	} {
+		switch value := gjson.GetBytes(body, path); value.Type {
+		case gjson.True:
+			return true, true
+		case gjson.False:
+			return false, true
+		}
+	}
+	return false, false
+}
+
+func normalizeInteractionsLevel(level string, modelInfo *registry.ModelInfo) string {
+	level = strings.ToLower(strings.TrimSpace(level))
+	if level == "" || level == string(thinking.LevelNone) || level == string(thinking.LevelAuto) {
+		return ""
+	}
+	if modelInfo != nil && modelInfo.Thinking != nil && len(modelInfo.Thinking.Levels) > 0 {
+		for _, candidate := range modelInfo.Thinking.Levels {
+			if strings.EqualFold(candidate, level) {
+				return strings.ToLower(candidate)
+			}
+		}
+		return strings.ToLower(modelInfo.Thinking.Levels[len(modelInfo.Thinking.Levels)-1])
+	}
+	switch level {
+	case string(thinking.LevelMax), string(thinking.LevelXHigh):
+		return string(thinking.LevelHigh)
+	default:
+		return level
+	}
+}

backend/internal/thinking/provider/kimi/apply.go

@@ -0,0 +1,168 @@
+// Package kimi implements thinking configuration for Kimi (Moonshot AI) models.
+//
+// Kimi models use a native thinking object for both enabled and disabled thinking.
+// The top-level reasoning_effort field is accepted only as a legacy input by the
+// unified extraction layer and is removed from the final Kimi payload.
+package kimi
+
+import (
+	"fmt"
+
+	"github.com/router-for-me/CLIProxyAPI/v7/internal/registry"
+	"github.com/router-for-me/CLIProxyAPI/v7/internal/thinking"
+	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
+)
+
+// Applier implements thinking.ProviderApplier for Kimi models.
+//
+// Kimi-specific behavior:
+//   - Enabled thinking: thinking.type="enabled" + thinking.effort=<level>
+//   - Disabled thinking: thinking.type="disabled"
+//   - Supports budget-to-level conversion
+//   - Preserves existing thinking.keep when enabling or changing effort
+type Applier struct{}
+
+var _ thinking.ProviderApplier = (*Applier)(nil)
+
+// NewApplier creates a new Kimi thinking applier.
+func NewApplier() *Applier {
+	return &Applier{}
+}
+
+func init() {
+	thinking.RegisterProvider("kimi", NewApplier())
+}
+
+// Apply applies thinking configuration to Kimi request body.
+//
+// Expected output format (enabled):
+//
+//	{
+//	  "thinking": {
+//	    "type": "enabled",
+//	    "effort": "high"
+//	  }
+//	}
+//
+// Expected output format (disabled):
+//
+//	{
+//	  "thinking": {
+//	    "type": "disabled"
+//	  }
+//	}
+func (a *Applier) Apply(body []byte, config thinking.ThinkingConfig, modelInfo *registry.ModelInfo) ([]byte, error) {
+	if thinking.IsUserDefinedModel(modelInfo) {
+		return applyCompatibleKimi(body, config)
+	}
+	if modelInfo.Thinking == nil {
+		return body, nil
+	}
+
+	if len(body) == 0 || !gjson.ValidBytes(body) {
+		body = []byte(`{}`)
+	}
+
+	var effort string
+	switch config.Mode {
+	case thinking.ModeLevel:
+		if config.Level == "" {
+			return body, nil
+		}
+		effort = string(config.Level)
+	case thinking.ModeNone:
+		// Respect clamped fallback level for models that cannot disable thinking.
+		if config.Level != "" && config.Level != thinking.LevelNone {
+			effort = string(config.Level)
+			break
+		}
+		// Kimi requires explicit disabled thinking object.
+		return applyDisabledThinking(body)
+	case thinking.ModeBudget:
+		// Convert budget to level using threshold mapping
+		level, ok := thinking.ConvertBudgetToLevel(config.Budget)
+		if !ok {
+			return body, nil
+		}
+		effort = level
+	case thinking.ModeAuto:
+		// Auto mode maps to "auto" effort
+		effort = string(thinking.LevelAuto)
+	default:
+		return body, nil
+	}
+
+	if effort == "" {
+		return body, nil
+	}
+	return applyEnabledThinking(body, effort)
+}
+
+// applyCompatibleKimi applies thinking config for user-defined Kimi models.
+func applyCompatibleKimi(body []byte, config thinking.ThinkingConfig) ([]byte, error) {
+	if len(body) == 0 || !gjson.ValidBytes(body) {
+		body = []byte(`{}`)
+	}
+
+	var effort string
+	switch config.Mode {
+	case thinking.ModeLevel:
+		if config.Level == "" {
+			return body, nil
+		}
+		effort = string(config.Level)
+	case thinking.ModeNone:
+		if config.Level == "" || config.Level == thinking.LevelNone {
+			return applyDisabledThinking(body)
+		}
+		if config.Level != "" {
+			effort = string(config.Level)
+		}
+	case thinking.ModeAuto:
+		effort = string(thinking.LevelAuto)
+	case thinking.ModeBudget:
+		// Convert budget to level
+		level, ok := thinking.ConvertBudgetToLevel(config.Budget)
+		if !ok {
+			return body, nil
+		}
+		effort = level
+	default:
+		return body, nil
+	}
+
+	return applyEnabledThinking(body, effort)
+}
+
+func applyEnabledThinking(body []byte, effort string) ([]byte, error) {
+	result, errDeleteLegacyEffort := sjson.DeleteBytes(body, "reasoning_effort")
+	if errDeleteLegacyEffort != nil {
+		return body, fmt.Errorf("kimi thinking: failed to clear reasoning_effort: %w", errDeleteLegacyEffort)
+	}
+	result, errSetType := sjson.SetBytes(result, "thinking.type", "enabled")
+	if errSetType != nil {
+		return body, fmt.Errorf("kimi thinking: failed to set thinking.type: %w", errSetType)
+	}
+	result, errSetEffort := sjson.SetBytes(result, "thinking.effort", effort)
+	if errSetEffort != nil {
+		return body, fmt.Errorf("kimi thinking: failed to set thinking.effort: %w", errSetEffort)
+	}
+	return result, nil
+}
+
+func applyDisabledThinking(body []byte) ([]byte, error) {
+	result, errDeleteThinking := sjson.DeleteBytes(body, "thinking")
+	if errDeleteThinking != nil {
+		return body, fmt.Errorf("kimi thinking: failed to clear thinking object: %w", errDeleteThinking)
+	}
+	result, errDeleteEffort := sjson.DeleteBytes(result, "reasoning_effort")
+	if errDeleteEffort != nil {
+		return body, fmt.Errorf("kimi thinking: failed to clear reasoning_effort: %w", errDeleteEffort)
+	}
+	result, errSetType := sjson.SetBytes(result, "thinking.type", "disabled")
+	if errSetType != nil {
+		return body, fmt.Errorf("kimi thinking: failed to set thinking.type: %w", errSetType)
+	}
+	return result, nil
+}

backend/internal/thinking/provider/openai/apply.go

@@ -0,0 +1,117 @@
+// Package openai implements thinking configuration for OpenAI/Codex models.
+//
+// OpenAI models use the reasoning_effort format with discrete levels
+// (low/medium/high). Some models support xhigh and none levels.
+// See: _bmad-output/planning-artifacts/architecture.md#Epic-8
+package openai
+
+import (
+	"github.com/router-for-me/CLIProxyAPI/v7/internal/registry"
+	"github.com/router-for-me/CLIProxyAPI/v7/internal/thinking"
+	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
+)
+
+// Applier implements thinking.ProviderApplier for OpenAI models.
+//
+// OpenAI-specific behavior:
+//   - Output format: reasoning_effort (string: low/medium/high/xhigh)
+//   - Level-only mode: no numeric budget support
+//   - Some models support ZeroAllowed (gpt-5.1, gpt-5.2)
+type Applier struct{}
+
+var _ thinking.ProviderApplier = (*Applier)(nil)
+
+// NewApplier creates a new OpenAI thinking applier.
+func NewApplier() *Applier {
+	return &Applier{}
+}
+
+func init() {
+	thinking.RegisterProvider("openai", NewApplier())
+}
+
+// Apply applies thinking configuration to OpenAI request body.
+//
+// Expected output format:
+//
+//	{
+//	  "reasoning_effort": "high"
+//	}
+func (a *Applier) Apply(body []byte, config thinking.ThinkingConfig, modelInfo *registry.ModelInfo) ([]byte, error) {
+	if thinking.IsUserDefinedModel(modelInfo) {
+		return applyCompatibleOpenAI(body, config)
+	}
+	if modelInfo.Thinking == nil {
+		return body, nil
+	}
+
+	// Only handle ModeLevel and ModeNone; other modes pass through unchanged.
+	if config.Mode != thinking.ModeLevel && config.Mode != thinking.ModeNone {
+		return body, nil
+	}
+
+	if len(body) == 0 || !gjson.ValidBytes(body) {
+		body = []byte(`{}`)
+	}
+
+	if config.Mode == thinking.ModeLevel {
+		result, _ := sjson.SetBytes(body, "reasoning_effort", string(config.Level))
+		return result, nil
+	}
+
+	effort := ""
+	support := modelInfo.Thinking
+	if config.Budget == 0 {
+		if support.ZeroAllowed || thinking.HasLevel(support.Levels, string(thinking.LevelNone)) {
+			effort = string(thinking.LevelNone)
+		}
+	}
+	if effort == "" && config.Level != "" {
+		effort = string(config.Level)
+	}
+	if effort == "" && len(support.Levels) > 0 {
+		effort = support.Levels[0]
+	}
+	if effort == "" {
+		return body, nil
+	}
+
+	result, _ := sjson.SetBytes(body, "reasoning_effort", effort)
+	return result, nil
+}
+
+func applyCompatibleOpenAI(body []byte, config thinking.ThinkingConfig) ([]byte, error) {
+	if len(body) == 0 || !gjson.ValidBytes(body) {
+		body = []byte(`{}`)
+	}
+
+	var effort string
+	switch config.Mode {
+	case thinking.ModeLevel:
+		if config.Level == "" {
+			return body, nil
+		}
+		effort = string(config.Level)
+	case thinking.ModeNone:
+		effort = string(thinking.LevelNone)
+		if config.Level != "" {
+			effort = string(config.Level)
+		}
+	case thinking.ModeAuto:
+		// Auto mode for user-defined models: pass through as "auto"
+		effort = string(thinking.LevelAuto)
+	case thinking.ModeBudget:
+		// Budget mode: convert budget to level using threshold mapping
+		level, ok := thinking.ConvertBudgetToLevel(config.Budget)
+		if !ok {
+			return body, nil
+		}
+		effort = level
+	default:
+		return body, nil
+	}
+
+	result, _ := sjson.SetBytes(body, "reasoning_effort", effort)
+	return result, nil
+}

backend/internal/thinking/provider/xai/apply.go

@@ -0,0 +1,26 @@
+// Package xai implements thinking configuration for xAI Grok Responses API models.
+//
+// xAI models use the OpenAI Responses API compatible reasoning.effort format
+// with discrete levels.
+package xai
+
+import (
+	"github.com/router-for-me/CLIProxyAPI/v7/internal/thinking"
+	"github.com/router-for-me/CLIProxyAPI/v7/internal/thinking/provider/codex"
+)
+
+// Applier implements thinking.ProviderApplier for xAI models.
+type Applier struct {
+	codex.Applier
+}
+
+var _ thinking.ProviderApplier = (*Applier)(nil)
+
+// NewApplier creates a new xAI thinking applier.
+func NewApplier() *Applier {
+	return &Applier{}
+}
+
+func init() {
+	thinking.RegisterProvider("xai", NewApplier())
+}