Merge branch 'router-for-me:main' into master

Ravens2121 committed 11 months ago

Commit f302be52 parents cd4e84a 4360ed813 files · +579 −647Browse files

Unified · Split

13 files

internal/api/modules/amp/fallback_handlers.go +2 −2
internal/registry/model_definitions.go +1 −341
internal/runtime/executor/claude_executor.go +106 −45
internal/runtime/executor/codex_executor.go +12 −84
internal/runtime/executor/gemini_executor.go +8 −2
internal/runtime/executor/gemini_vertex_executor.go +29 −11
internal/runtime/executor/iflow_executor.go +8 −0
internal/runtime/executor/openai_compat_executor.go +8 −0
internal/runtime/executor/payload_helpers.go +48 −5
internal/runtime/executor/qwen_executor.go +9 −0
internal/util/gemini_thinking.go +0 −150
internal/util/thinking_suffix.go +327 −0
sdk/api/handlers/handlers.go +21 −7

internal/api/modules/amp/fallback_handlers.go+2−2

@@ -133,8 +133,8 @@ func (fh *FallbackHandler) WrapHandler(handler gin.HandlerFunc) gin.HandlerFunc · 132 unmodified lines
133	133	return
134	134	}
135	135
136		− // Normalize model (handles Gemini thinking suffixes)
137		− normalizedModel, _ := util.NormalizeGeminiThinkingModel(modelName)
	136	+ // Normalize model (handles dynamic thinking suffixes)
	137	+ normalizedModel, _ := util.NormalizeThinkingModel(modelName)
138	138
139	139	// Track resolved model for logging (may change if mapping is applied)
140	140	resolvedModel := normalizedModel

internal/registry/model_definitions.go+1−341

@@ -26,60 +26,6 @@ func GetClaudeModels() []*ModelInfo { · 25 unmodified lines
26	26	DisplayName: "Claude 4.5 Sonnet",
27	27	ContextLength: 200000,
28	28	MaxCompletionTokens: 64000,
29		− },
30		− {
31		− ID: "claude-sonnet-4-5-thinking",
32		− Object: "model",
33		− Created: 1759104000, // 2025-09-29
34		− OwnedBy: "anthropic",
35		− Type: "claude",
36		− DisplayName: "Claude 4.5 Sonnet Thinking",
37		− ContextLength: 200000,
38		− MaxCompletionTokens: 64000,
39		− Thinking: &ThinkingSupport{Min: 1024, Max: 100000, ZeroAllowed: false, DynamicAllowed: true},
40		− },
41		− {
42		− ID: "claude-opus-4-5-thinking",
43		− Object: "model",
44		− Created: 1761955200, // 2025-11-01
45		− OwnedBy: "anthropic",
46		− Type: "claude",
47		− DisplayName: "Claude 4.5 Opus Thinking",
48		− ContextLength: 200000,
49		− MaxCompletionTokens: 64000,
50		− Thinking: &ThinkingSupport{Min: 1024, Max: 100000, ZeroAllowed: false, DynamicAllowed: true},
51		− },
52		− {
53		− ID: "claude-opus-4-5-thinking-low",
54		− Object: "model",
55		− Created: 1761955200, // 2025-11-01
56		− OwnedBy: "anthropic",
57		− Type: "claude",
58		− DisplayName: "Claude 4.5 Opus Thinking Low",
59		− ContextLength: 200000,
60		− MaxCompletionTokens: 64000,
61		− Thinking: &ThinkingSupport{Min: 1024, Max: 100000, ZeroAllowed: false, DynamicAllowed: true},
62		− },
63		− {
64		− ID: "claude-opus-4-5-thinking-medium",
65		− Object: "model",
66		− Created: 1761955200, // 2025-11-01
67		− OwnedBy: "anthropic",
68		− Type: "claude",
69		− DisplayName: "Claude 4.5 Opus Thinking Medium",
70		− ContextLength: 200000,
71		− MaxCompletionTokens: 64000,
72		− Thinking: &ThinkingSupport{Min: 1024, Max: 100000, ZeroAllowed: false, DynamicAllowed: true},
73		− },
74		− {
75		− ID: "claude-opus-4-5-thinking-high",
76		− Object: "model",
77		− Created: 1761955200, // 2025-11-01
78		− OwnedBy: "anthropic",
79		− Type: "claude",
80		− DisplayName: "Claude 4.5 Opus Thinking High",
81		− ContextLength: 200000,
82		− MaxCompletionTokens: 64000,
83	29	Thinking: &ThinkingSupport{Min: 1024, Max: 100000, ZeroAllowed: false, DynamicAllowed: true},
84	30	},
85	31	{
@@ -92,6 +38,7 @@ func GetClaudeModels() []*ModelInfo { · 6 unmodified lines
92	38	Description: "Premium model combining maximum intelligence with practical performance",
93	39	ContextLength: 200000,
94	40	MaxCompletionTokens: 64000,
	41	+ Thinking: &ThinkingSupport{Min: 1024, Max: 100000, ZeroAllowed: false, DynamicAllowed: true},
95	42	},
96	43	{
97	44	ID: "claude-opus-4-1-20250805",
@@ -530,58 +477,6 @@ func GetOpenAIModels() []*ModelInfo { · 432 unmodified lines
530	477	MaxCompletionTokens: 128000,
531	478	SupportedParameters: []string{"tools"},
532	479	},
533		− {
534		− ID: "gpt-5-minimal",
535		− Object: "model",
536		− Created: 1754524800,
537		− OwnedBy: "openai",
538		− Type: "openai",
539		− Version: "gpt-5-2025-08-07",
540		− DisplayName: "GPT 5 Minimal",
541		− Description: "Stable version of GPT 5, The best model for coding and agentic tasks across domains.",
542		− ContextLength: 400000,
543		− MaxCompletionTokens: 128000,
544		− SupportedParameters: []string{"tools"},
545		− },
546		− {
547		− ID: "gpt-5-low",
548		− Object: "model",
549		− Created: 1754524800,
550		− OwnedBy: "openai",
551		− Type: "openai",
552		− Version: "gpt-5-2025-08-07",
553		− DisplayName: "GPT 5 Low",
554		− Description: "Stable version of GPT 5, The best model for coding and agentic tasks across domains.",
555		− ContextLength: 400000,
556		− MaxCompletionTokens: 128000,
557		− SupportedParameters: []string{"tools"},
558		− },
559		− {
560		− ID: "gpt-5-medium",
561		− Object: "model",
562		− Created: 1754524800,
563		− OwnedBy: "openai",
564		− Type: "openai",
565		− Version: "gpt-5-2025-08-07",
566		− DisplayName: "GPT 5 Medium",
567		− Description: "Stable version of GPT 5, The best model for coding and agentic tasks across domains.",
568		− ContextLength: 400000,
569		− MaxCompletionTokens: 128000,
570		− SupportedParameters: []string{"tools"},
571		− },
572		− {
573		− ID: "gpt-5-high",
574		− Object: "model",
575		− Created: 1754524800,
576		− OwnedBy: "openai",
577		− Type: "openai",
578		− Version: "gpt-5-2025-08-07",
579		− DisplayName: "GPT 5 High",
580		− Description: "Stable version of GPT 5, The best model for coding and agentic tasks across domains.",
581		− ContextLength: 400000,
582		− MaxCompletionTokens: 128000,
583		− SupportedParameters: []string{"tools"},
584		− },
585	480	{
586	481	ID: "gpt-5-codex",
587	482	Object: "model",
@@ -595,45 +490,6 @@ func GetOpenAIModels() []*ModelInfo { · 7 unmodified lines
595	490	MaxCompletionTokens: 128000,
596	491	SupportedParameters: []string{"tools"},
597	492	},
598		− {
599		− ID: "gpt-5-codex-low",
600		− Object: "model",
601		− Created: 1757894400,
602		− OwnedBy: "openai",
603		− Type: "openai",
604		− Version: "gpt-5-2025-09-15",
605		− DisplayName: "GPT 5 Codex Low",
606		− Description: "Stable version of GPT 5 Codex, The best model for coding and agentic tasks across domains.",
607		− ContextLength: 400000,
608		− MaxCompletionTokens: 128000,
609		− SupportedParameters: []string{"tools"},
610		− },
611		− {
612		− ID: "gpt-5-codex-medium",
613		− Object: "model",
614		− Created: 1757894400,
615		− OwnedBy: "openai",
616		− Type: "openai",
617		− Version: "gpt-5-2025-09-15",
618		− DisplayName: "GPT 5 Codex Medium",
619		− Description: "Stable version of GPT 5 Codex, The best model for coding and agentic tasks across domains.",
620		− ContextLength: 400000,
621		− MaxCompletionTokens: 128000,
622		− SupportedParameters: []string{"tools"},
623		− },
624		− {
625		− ID: "gpt-5-codex-high",
626		− Object: "model",
627		− Created: 1757894400,
628		− OwnedBy: "openai",
629		− Type: "openai",
630		− Version: "gpt-5-2025-09-15",
631		− DisplayName: "GPT 5 Codex High",
632		− Description: "Stable version of GPT 5 Codex, The best model for coding and agentic tasks across domains.",
633		− ContextLength: 400000,
634		− MaxCompletionTokens: 128000,
635		− SupportedParameters: []string{"tools"},
636		− },
637	493	{
638	494	ID: "gpt-5-codex-mini",
639	495	Object: "model",
@@ -647,32 +503,6 @@ func GetOpenAIModels() []*ModelInfo { · 7 unmodified lines
647	503	MaxCompletionTokens: 128000,
648	504	SupportedParameters: []string{"tools"},
649	505	},
650		− {
651		− ID: "gpt-5-codex-mini-medium",
652		− Object: "model",
653		− Created: 1762473600,
654		− OwnedBy: "openai",
655		− Type: "openai",
656		− Version: "gpt-5-2025-11-07",
657		− DisplayName: "GPT 5 Codex Mini Medium",
658		− Description: "Stable version of GPT 5 Codex Mini: cheaper, faster, but less capable version of GPT 5 Codex.",
659		− ContextLength: 400000,
660		− MaxCompletionTokens: 128000,
661		− SupportedParameters: []string{"tools"},
662		− },
663		− {
664		− ID: "gpt-5-codex-mini-high",
665		− Object: "model",
666		− Created: 1762473600,
667		− OwnedBy: "openai",
668		− Type: "openai",
669		− Version: "gpt-5-2025-11-07",
670		− DisplayName: "GPT 5 Codex Mini High",
671		− Description: "Stable version of GPT 5 Codex Mini: cheaper, faster, but less capable version of GPT 5 Codex.",
672		− ContextLength: 400000,
673		− MaxCompletionTokens: 128000,
674		− SupportedParameters: []string{"tools"},
675		− },
676	506	{
677	507	ID: "gpt-5.1",
678	508	Object: "model",
@@ -686,58 +516,6 @@ func GetOpenAIModels() []*ModelInfo { · 7 unmodified lines
686	516	MaxCompletionTokens: 128000,
687	517	SupportedParameters: []string{"tools"},
688	518	},
689		− {
690		− ID: "gpt-5.1-none",
691		− Object: "model",
692		− Created: 1762905600,
693		− OwnedBy: "openai",
694		− Type: "openai",
695		− Version: "gpt-5.1-2025-11-12",
696		− DisplayName: "GPT 5.1 Nothink",
697		− Description: "Stable version of GPT 5.1, The best model for coding and agentic tasks across domains.",
698		− ContextLength: 400000,
699		− MaxCompletionTokens: 128000,
700		− SupportedParameters: []string{"tools"},
701		− },
702		− {
703		− ID: "gpt-5.1-low",
704		− Object: "model",
705		− Created: 1762905600,
706		− OwnedBy: "openai",
707		− Type: "openai",
708		− Version: "gpt-5.1-2025-11-12",
709		− DisplayName: "GPT 5 Low",
710		− Description: "Stable version of GPT 5, The best model for coding and agentic tasks across domains.",
711		− ContextLength: 400000,
712		− MaxCompletionTokens: 128000,
713		− SupportedParameters: []string{"tools"},
714		− },
715		− {
716		− ID: "gpt-5.1-medium",
717		− Object: "model",
718		− Created: 1762905600,
719		− OwnedBy: "openai",
720		− Type: "openai",
721		− Version: "gpt-5.1-2025-11-12",
722		− DisplayName: "GPT 5.1 Medium",
723		− Description: "Stable version of GPT 5.1, The best model for coding and agentic tasks across domains.",
724		− ContextLength: 400000,
725		− MaxCompletionTokens: 128000,
726		− SupportedParameters: []string{"tools"},
727		− },
728		− {
729		− ID: "gpt-5.1-high",
730		− Object: "model",
731		− Created: 1762905600,
732		− OwnedBy: "openai",
733		− Type: "openai",
734		− Version: "gpt-5.1-2025-11-12",
735		− DisplayName: "GPT 5.1 High",
736		− Description: "Stable version of GPT 5.1, The best model for coding and agentic tasks across domains.",
737		− ContextLength: 400000,
738		− MaxCompletionTokens: 128000,
739		− SupportedParameters: []string{"tools"},
740		− },
741	519	{
742	520	ID: "gpt-5.1-codex",
743	521	Object: "model",
@@ -751,45 +529,6 @@ func GetOpenAIModels() []*ModelInfo { · 7 unmodified lines
751	529	MaxCompletionTokens: 128000,
752	530	SupportedParameters: []string{"tools"},
753	531	},
754		− {
755		− ID: "gpt-5.1-codex-low",
756		− Object: "model",
757		− Created: 1762905600,
758		− OwnedBy: "openai",
759		− Type: "openai",
760		− Version: "gpt-5.1-2025-11-12",
761		− DisplayName: "GPT 5.1 Codex Low",
762		− Description: "Stable version of GPT 5.1 Codex, The best model for coding and agentic tasks across domains.",
763		− ContextLength: 400000,
764		− MaxCompletionTokens: 128000,
765		− SupportedParameters: []string{"tools"},
766		− },
767		− {
768		− ID: "gpt-5.1-codex-medium",
769		− Object: "model",
770		− Created: 1762905600,
771		− OwnedBy: "openai",
772		− Type: "openai",
773		− Version: "gpt-5.1-2025-11-12",
774		− DisplayName: "GPT 5.1 Codex Medium",
775		− Description: "Stable version of GPT 5.1 Codex, The best model for coding and agentic tasks across domains.",
776		− ContextLength: 400000,
777		− MaxCompletionTokens: 128000,
778		− SupportedParameters: []string{"tools"},
779		− },
780		− {
781		− ID: "gpt-5.1-codex-high",
782		− Object: "model",
783		− Created: 1762905600,
784		− OwnedBy: "openai",
785		− Type: "openai",
786		− Version: "gpt-5.1-2025-11-12",
787		− DisplayName: "GPT 5.1 Codex High",
788		− Description: "Stable version of GPT 5.1 Codex, The best model for coding and agentic tasks across domains.",
789		− ContextLength: 400000,
790		− MaxCompletionTokens: 128000,
791		− SupportedParameters: []string{"tools"},
792		− },
793	532	{
794	533	ID: "gpt-5.1-codex-mini",
795	534	Object: "model",
@@ -803,33 +542,6 @@ func GetOpenAIModels() []*ModelInfo { · 7 unmodified lines
803	542	MaxCompletionTokens: 128000,
804	543	SupportedParameters: []string{"tools"},
805	544	},
806		− {
807		− ID: "gpt-5.1-codex-mini-medium",
808		− Object: "model",
809		− Created: 1762905600,
810		− OwnedBy: "openai",
811		− Type: "openai",
812		− Version: "gpt-5.1-2025-11-12",
813		− DisplayName: "GPT 5.1 Codex Mini Medium",
814		− Description: "Stable version of GPT 5.1 Codex Mini: cheaper, faster, but less capable version of GPT 5.1 Codex.",
815		− ContextLength: 400000,
816		− MaxCompletionTokens: 128000,
817		− SupportedParameters: []string{"tools"},
818		− },
819		− {
820		− ID: "gpt-5.1-codex-mini-high",
821		− Object: "model",
822		− Created: 1762905600,
823		− OwnedBy: "openai",
824		− Type: "openai",
825		− Version: "gpt-5.1-2025-11-12",
826		− DisplayName: "GPT 5.1 Codex Mini High",
827		− Description: "Stable version of GPT 5.1 Codex Mini: cheaper, faster, but less capable version of GPT 5.1 Codex.",
828		− ContextLength: 400000,
829		− MaxCompletionTokens: 128000,
830		− SupportedParameters: []string{"tools"},
831		− },
832		−
833	545	{
834	546	ID: "gpt-5.1-codex-max",
835	547	Object: "model",
@@ -843,58 +555,6 @@ func GetOpenAIModels() []*ModelInfo { · 7 unmodified lines
843	555	MaxCompletionTokens: 128000,
844	556	SupportedParameters: []string{"tools"},
845	557	},
846		− {
847		− ID: "gpt-5.1-codex-max-low",
848		− Object: "model",
849		− Created: 1763424000,
850		− OwnedBy: "openai",
851		− Type: "openai",
852		− Version: "gpt-5.1-max",
853		− DisplayName: "GPT 5.1 Codex Max Low",
854		− Description: "Stable version of GPT 5.1 Codex Max Low",
855		− ContextLength: 400000,
856		− MaxCompletionTokens: 128000,
857		− SupportedParameters: []string{"tools"},
858		− },
859		− {
860		− ID: "gpt-5.1-codex-max-medium",
861		− Object: "model",
862		− Created: 1763424000,
863		− OwnedBy: "openai",
864		− Type: "openai",
865		− Version: "gpt-5.1-max",
866		− DisplayName: "GPT 5.1 Codex Max Medium",
867		− Description: "Stable version of GPT 5.1 Codex Max Medium",
868		− ContextLength: 400000,
869		− MaxCompletionTokens: 128000,
870		− SupportedParameters: []string{"tools"},
871		− },
872		− {
873		− ID: "gpt-5.1-codex-max-high",
874		− Object: "model",
875		− Created: 1763424000,
876		− OwnedBy: "openai",
877		− Type: "openai",
878		− Version: "gpt-5.1-max",
879		− DisplayName: "GPT 5.1 Codex Max High",
880		− Description: "Stable version of GPT 5.1 Codex Max High",
881		− ContextLength: 400000,
882		− MaxCompletionTokens: 128000,
883		− SupportedParameters: []string{"tools"},
884		− },
885		− {
886		− ID: "gpt-5.1-codex-max-xhigh",
887		− Object: "model",
888		− Created: 1763424000,
889		− OwnedBy: "openai",
890		− Type: "openai",
891		− Version: "gpt-5.1-max",
892		− DisplayName: "GPT 5.1 Codex Max XHigh",
893		− Description: "Stable version of GPT 5.1 Codex Max XHigh",
894		− ContextLength: 400000,
895		− MaxCompletionTokens: 128000,
896		− SupportedParameters: []string{"tools"},
897		− },
898	558	}
899	559	}
900	560

internal/runtime/executor/claude_executor.go+106−45

@@ -54,15 +54,22 @@ func (e ClaudeExecutor) Execute(ctx context.Context, auth cliproxyauth.Auth, r · 53 unmodified lines
54	54	// Use streaming translation to preserve function calling, except for claude.
55	55	stream := from != to
56	56	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), stream)
57		− modelForUpstream := req.Model
58		− if modelOverride := e.resolveUpstreamModel(req.Model, auth); modelOverride != "" {
59		− body, _ = sjson.SetBytes(body, "model", modelOverride)
60		− modelForUpstream = modelOverride
	57	+ upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
	58	+ if upstreamModel == "" {
	59	+ upstreamModel = req.Model
	60	+ }
	61	+ if modelOverride := e.resolveUpstreamModel(upstreamModel, auth); modelOverride != "" {
	62	+ upstreamModel = modelOverride
	63	+ } else if !strings.EqualFold(upstreamModel, req.Model) {
	64	+ if modelOverride := e.resolveUpstreamModel(req.Model, auth); modelOverride != "" {
	65	+ upstreamModel = modelOverride
	66	+ }
61	67	}
62		− // Inject thinking config based on model suffix for thinking variants
63		− body = e.injectThinkingConfig(req.Model, body)
	68	+ body, _ = sjson.SetBytes(body, "model", upstreamModel)
	69	+ // Inject thinking config based on model metadata for thinking variants
	70	+ body = e.injectThinkingConfig(req.Model, req.Metadata, body)
64	71
65		− if !strings.HasPrefix(modelForUpstream, "claude-3-5-haiku") {
	72	+ if !strings.HasPrefix(upstreamModel, "claude-3-5-haiku") {
66	73	body = checkSystemInstructions(body)
67	74	}
68	75	body = applyPayloadConfig(e.cfg, req.Model, body)
@@ -161,11 +168,20 @@ func (e ClaudeExecutor) ExecuteStream(ctx context.Context, auth cliproxyauth.A · 92 unmodified lines
161	168	from := opts.SourceFormat
162	169	to := sdktranslator.FromString("claude")
163	170	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
164		− if modelOverride := e.resolveUpstreamModel(req.Model, auth); modelOverride != "" {
165		− body, _ = sjson.SetBytes(body, "model", modelOverride)
	171	+ upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
	172	+ if upstreamModel == "" {
	173	+ upstreamModel = req.Model
	174	+ }
	175	+ if modelOverride := e.resolveUpstreamModel(upstreamModel, auth); modelOverride != "" {
	176	+ upstreamModel = modelOverride
	177	+ } else if !strings.EqualFold(upstreamModel, req.Model) {
	178	+ if modelOverride := e.resolveUpstreamModel(req.Model, auth); modelOverride != "" {
	179	+ upstreamModel = modelOverride
	180	+ }
166	181	}
167		− // Inject thinking config based on model suffix for thinking variants
168		− body = e.injectThinkingConfig(req.Model, body)
	182	+ body, _ = sjson.SetBytes(body, "model", upstreamModel)
	183	+ // Inject thinking config based on model metadata for thinking variants
	184	+ body = e.injectThinkingConfig(req.Model, req.Metadata, body)
169	185	body = checkSystemInstructions(body)
170	186	body = applyPayloadConfig(e.cfg, req.Model, body)
171	187
@@ -295,13 +311,20 @@ func (e ClaudeExecutor) CountTokens(ctx context.Context, auth cliproxyauth.Aut · 123 unmodified lines
295	311	// Use streaming translation to preserve function calling, except for claude.
296	312	stream := from != to
297	313	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), stream)
298		− modelForUpstream := req.Model
299		− if modelOverride := e.resolveUpstreamModel(req.Model, auth); modelOverride != "" {
300		− body, _ = sjson.SetBytes(body, "model", modelOverride)
301		− modelForUpstream = modelOverride
	314	+ upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
	315	+ if upstreamModel == "" {
	316	+ upstreamModel = req.Model
	317	+ }
	318	+ if modelOverride := e.resolveUpstreamModel(upstreamModel, auth); modelOverride != "" {
	319	+ upstreamModel = modelOverride
	320	+ } else if !strings.EqualFold(upstreamModel, req.Model) {
	321	+ if modelOverride := e.resolveUpstreamModel(req.Model, auth); modelOverride != "" {
	322	+ upstreamModel = modelOverride
	323	+ }
302	324	}
	325	+ body, _ = sjson.SetBytes(body, "model", upstreamModel)
303	326
304		− if !strings.HasPrefix(modelForUpstream, "claude-3-5-haiku") {
	327	+ if !strings.HasPrefix(upstreamModel, "claude-3-5-haiku") {
305	328	body = checkSystemInstructions(body)
306	329	}
307	330
@@ -427,25 +450,15 @@ func extractAndRemoveBetas(body []byte) ([]string, []byte) { · 119 unmodified lines
427	450	return betas, body
428	451	}
429	452
430		−// injectThinkingConfig adds thinking configuration based on model name suffix
431		−func (e *ClaudeExecutor) injectThinkingConfig(modelName string, body []byte) []byte {
	453	+// injectThinkingConfig adds thinking configuration based on metadata or legacy suffixes.
	454	+func (e *ClaudeExecutor) injectThinkingConfig(modelName string, metadata map[string]any, body []byte) []byte {
432	455	// Only inject if thinking config is not already present
433	456	if gjson.GetBytes(body, "thinking").Exists() {
434	457	return body
435	458	}
436	459
437		− var budgetTokens int
438		− switch {
439		− case strings.HasSuffix(modelName, "-thinking-low"):
440		− budgetTokens = 1024
441		− case strings.HasSuffix(modelName, "-thinking-medium"):
442		− budgetTokens = 8192
443		− case strings.HasSuffix(modelName, "-thinking-high"):
444		− budgetTokens = 24576
445		− case strings.HasSuffix(modelName, "-thinking"):
446		− // Default thinking without suffix uses medium budget
447		− budgetTokens = 8192
448		− default:
	460	+ budgetTokens, ok := resolveClaudeThinkingBudget(modelName, metadata)
	461	+ if !ok \|\| budgetTokens <= 0 {
449	462	return body
450	463	}
451	464
@@ -454,6 +467,44 @@ func (e *ClaudeExecutor) injectThinkingConfig(modelName string, body []byte) []b · 2 unmodified lines
454	467	return body
455	468	}
456	469
	470	+func resolveClaudeThinkingBudget(modelName string, metadata map[string]any) (int, bool) {
	471	+ budget, include, effort, matched := util.ThinkingFromMetadata(metadata)
	472	+ if matched {
	473	+ if include != nil && !*include {
	474	+ return 0, false
	475	+ }
	476	+ if budget != nil {
	477	+ normalized := util.NormalizeThinkingBudget(modelName, *budget)
	478	+ if normalized > 0 {
	479	+ return normalized, true
	480	+ }
	481	+ return 0, false
	482	+ }
	483	+ if effort != nil {
	484	+ if derived, ok := util.ThinkingEffortToBudget(modelName, *effort); ok && derived > 0 {
	485	+ return derived, true
	486	+ }
	487	+ }
	488	+ }
	489	+ return claudeBudgetFromSuffix(modelName)
	490	+}
	491	+
	492	+func claudeBudgetFromSuffix(modelName string) (int, bool) {
	493	+ lower := strings.ToLower(strings.TrimSpace(modelName))
	494	+ switch {
	495	+ case strings.HasSuffix(lower, "-thinking-low"):
	496	+ return 1024, true
	497	+ case strings.HasSuffix(lower, "-thinking-medium"):
	498	+ return 8192, true
	499	+ case strings.HasSuffix(lower, "-thinking-high"):
	500	+ return 24576, true
	501	+ case strings.HasSuffix(lower, "-thinking"):
	502	+ return 8192, true
	503	+ default:
	504	+ return 0, false
	505	+ }
	506	+}
	507	+
457	508	// ensureMaxTokensForThinking ensures max_tokens > thinking.budget_tokens when thinking is enabled.
458	509	// Anthropic API requires this constraint; violating it returns a 400 error.
459	510	// This function should be called after all thinking configuration is finalized.
@@ -491,35 +542,45 @@ func ensureMaxTokensForThinking(modelName string, body []byte) []byte { · 31 unmodified lines
491	542	}
492	543
493	544	func (e ClaudeExecutor) resolveUpstreamModel(alias string, auth cliproxyauth.Auth) string {
494		− if alias == "" {
	545	+ trimmed := strings.TrimSpace(alias)
	546	+ if trimmed == "" {
495	547	return ""
496	548	}
497		− // Hardcoded mappings for thinking models to actual Claude model names
498		− switch alias {
499		− case "claude-opus-4-5-thinking", "claude-opus-4-5-thinking-low", "claude-opus-4-5-thinking-medium", "claude-opus-4-5-thinking-high":
500		− return "claude-opus-4-5-20251101"
501		− case "claude-sonnet-4-5-thinking":
502		− return "claude-sonnet-4-5-20250929"
503		− }
	549	+
504	550	entry := e.resolveClaudeConfig(auth)
505	551	if entry == nil {
506	552	return ""
507	553	}
	554	+
	555	+ normalizedModel, metadata := util.NormalizeThinkingModel(trimmed)
	556	+
	557	+ // Candidate names to match against configured aliases/names.
	558	+ candidates := []string{strings.TrimSpace(normalizedModel)}
	559	+ if !strings.EqualFold(normalizedModel, trimmed) {
	560	+ candidates = append(candidates, trimmed)
	561	+ }
	562	+ if original := util.ResolveOriginalModel(normalizedModel, metadata); original != "" && !strings.EqualFold(original, normalizedModel) {
	563	+ candidates = append(candidates, original)
	564	+ }
	565	+
508	566	for i := range entry.Models {
509	567	model := entry.Models[i]
510	568	name := strings.TrimSpace(model.Name)
511	569	modelAlias := strings.TrimSpace(model.Alias)
512		− if modelAlias != "" {
513		− if strings.EqualFold(modelAlias, alias) {
	570	+
	571	+ for _, candidate := range candidates {
	572	+ if candidate == "" {
	573	+ continue
	574	+ }
	575	+ if modelAlias != "" && strings.EqualFold(modelAlias, candidate) {
514	576	if name != "" {
515	577	return name
516	578	}
517		− return alias
	579	+ return candidate
	580	+ }
	581	+ if name != "" && strings.EqualFold(name, candidate) {
	582	+ return name
518	583	}
519		− continue
520		− }
521		− if name != "" && strings.EqualFold(name, alias) {
522		− return name
523	584	}
524	585	}
525	586	return ""

internal/runtime/executor/codex_executor.go+12−84

@@ -49,14 +49,14 @@ func (e CodexExecutor) Execute(ctx context.Context, auth cliproxyauth.Auth, re · 48 unmodified lines
49	49	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
50	50	defer reporter.trackFailure(ctx, &err)
51	51
	52	+ upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
	53	+
52	54	from := opts.SourceFormat
53	55	to := sdktranslator.FromString("codex")
54	56	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
55		−
56		− body = e.setReasoningEffortByAlias(req.Model, body)
57		−
	57	+ body = applyReasoningEffortMetadata(body, req.Metadata, req.Model)
58	58	body = applyPayloadConfig(e.cfg, req.Model, body)
59		−
	59	+ body, _ = sjson.SetBytes(body, "model", upstreamModel)
60	60	body, _ = sjson.SetBytes(body, "stream", true)
61	61	body, _ = sjson.DeleteBytes(body, "previous_response_id")
62	62
@@ -142,13 +142,16 @@ func (e CodexExecutor) ExecuteStream(ctx context.Context, auth cliproxyauth.Au · 79 unmodified lines
142	142	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
143	143	defer reporter.trackFailure(ctx, &err)
144	144
	145	+ upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
	146	+
145	147	from := opts.SourceFormat
146	148	to := sdktranslator.FromString("codex")
147	149	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
148	150
149		− body = e.setReasoningEffortByAlias(req.Model, body)
	151	+ body = applyReasoningEffortMetadata(body, req.Metadata, req.Model)
150	152	body = applyPayloadConfig(e.cfg, req.Model, body)
151	153	body, _ = sjson.DeleteBytes(body, "previous_response_id")
	154	+ body, _ = sjson.SetBytes(body, "model", upstreamModel)
152	155
153	156	url := strings.TrimSuffix(baseURL, "/") + "/responses"
154	157	httpReq, err := e.cacheHelper(ctx, from, url, req, body)
@@ -235,14 +238,16 @@ func (e CodexExecutor) ExecuteStream(ctx context.Context, auth cliproxyauth.Au · 80 unmodified lines
235	238	}
236	239
237	240	func (e CodexExecutor) CountTokens(ctx context.Context, auth cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
	241	+ upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
	242	+
238	243	from := opts.SourceFormat
239	244	to := sdktranslator.FromString("codex")
240	245	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
241	246
242	247	modelForCounting := req.Model
243	248
244		− body = e.setReasoningEffortByAlias(req.Model, body)
245		−
	249	+ body = applyReasoningEffortMetadata(body, req.Metadata, req.Model)
	250	+ body, _ = sjson.SetBytes(body, "model", upstreamModel)
246	251	body, _ = sjson.DeleteBytes(body, "previous_response_id")
247	252	body, _ = sjson.SetBytes(body, "stream", false)
248	253
@@ -261,83 +266,6 @@ func (e CodexExecutor) CountTokens(ctx context.Context, auth cliproxyauth.Auth · 12 unmodified lines
261	266	return cliproxyexecutor.Response{Payload: []byte(translated)}, nil
262	267	}
263	268
264		−func (e *CodexExecutor) setReasoningEffortByAlias(modelName string, payload []byte) []byte {
265		− if util.InArray([]string{"gpt-5", "gpt-5-minimal", "gpt-5-low", "gpt-5-medium", "gpt-5-high"}, modelName) {
266		− payload, _ = sjson.SetBytes(payload, "model", "gpt-5")
267		− switch modelName {
268		− case "gpt-5-minimal":
269		− payload, _ = sjson.SetBytes(payload, "reasoning.effort", "minimal")
270		− case "gpt-5-low":
271		− payload, _ = sjson.SetBytes(payload, "reasoning.effort", "low")
272		− case "gpt-5-medium":
273		− payload, _ = sjson.SetBytes(payload, "reasoning.effort", "medium")
274		− case "gpt-5-high":
275		− payload, _ = sjson.SetBytes(payload, "reasoning.effort", "high")
276		− }
277		− } else if util.InArray([]string{"gpt-5-codex", "gpt-5-codex-low", "gpt-5-codex-medium", "gpt-5-codex-high"}, modelName) {
278		− payload, _ = sjson.SetBytes(payload, "model", "gpt-5-codex")
279		− switch modelName {
280		− case "gpt-5-codex-low":
281		− payload, _ = sjson.SetBytes(payload, "reasoning.effort", "low")
282		− case "gpt-5-codex-medium":
283		− payload, _ = sjson.SetBytes(payload, "reasoning.effort", "medium")
284		− case "gpt-5-codex-high":
285		− payload, _ = sjson.SetBytes(payload, "reasoning.effort", "high")
286		− }
287		− } else if util.InArray([]string{"gpt-5-codex-mini", "gpt-5-codex-mini-medium", "gpt-5-codex-mini-high"}, modelName) {
288		− payload, _ = sjson.SetBytes(payload, "model", "gpt-5-codex-mini")
289		− switch modelName {
290		− case "gpt-5-codex-mini-medium":
291		− payload, _ = sjson.SetBytes(payload, "reasoning.effort", "medium")
292		− case "gpt-5-codex-mini-high":
293		− payload, _ = sjson.SetBytes(payload, "reasoning.effort", "high")
294		− }
295		− } else if util.InArray([]string{"gpt-5.1", "gpt-5.1-none", "gpt-5.1-low", "gpt-5.1-medium", "gpt-5.1-high"}, modelName) {
296		− payload, _ = sjson.SetBytes(payload, "model", "gpt-5.1")
297		− switch modelName {
298		− case "gpt-5.1-none":
299		− payload, _ = sjson.SetBytes(payload, "reasoning.effort", "none")
300		− case "gpt-5.1-low":
301		− payload, _ = sjson.SetBytes(payload, "reasoning.effort", "low")
302		− case "gpt-5.1-medium":
303		− payload, _ = sjson.SetBytes(payload, "reasoning.effort", "medium")
304		− case "gpt-5.1-high":
305		− payload, _ = sjson.SetBytes(payload, "reasoning.effort", "high")
306		− }
307		− } else if util.InArray([]string{"gpt-5.1-codex", "gpt-5.1-codex-low", "gpt-5.1-codex-medium", "gpt-5.1-codex-high"}, modelName) {
308		− payload, _ = sjson.SetBytes(payload, "model", "gpt-5.1-codex")
309		− switch modelName {
310		− case "gpt-5.1-codex-low":
311		− payload, _ = sjson.SetBytes(payload, "reasoning.effort", "low")
312		− case "gpt-5.1-codex-medium":
313		− payload, _ = sjson.SetBytes(payload, "reasoning.effort", "medium")
314		− case "gpt-5.1-codex-high":
315		− payload, _ = sjson.SetBytes(payload, "reasoning.effort", "high")
316		− }
317		− } else if util.InArray([]string{"gpt-5.1-codex-mini", "gpt-5.1-codex-mini-medium", "gpt-5.1-codex-mini-high"}, modelName) {
318		− payload, _ = sjson.SetBytes(payload, "model", "gpt-5.1-codex-mini")
319		− switch modelName {
320		− case "gpt-5.1-codex-mini-medium":
321		− payload, _ = sjson.SetBytes(payload, "reasoning.effort", "medium")
322		− case "gpt-5.1-codex-mini-high":
323		− payload, _ = sjson.SetBytes(payload, "reasoning.effort", "high")
324		− }
325		− } else if util.InArray([]string{"gpt-5.1-codex-max", "gpt-5.1-codex-max-low", "gpt-5.1-codex-max-medium", "gpt-5.1-codex-max-high", "gpt-5.1-codex-max-xhigh"}, modelName) {
326		− payload, _ = sjson.SetBytes(payload, "model", "gpt-5.1-codex-max")
327		− switch modelName {
328		− case "gpt-5.1-codex-max-low":
329		− payload, _ = sjson.SetBytes(payload, "reasoning.effort", "low")
330		− case "gpt-5.1-codex-max-medium":
331		− payload, _ = sjson.SetBytes(payload, "reasoning.effort", "medium")
332		− case "gpt-5.1-codex-max-high":
333		− payload, _ = sjson.SetBytes(payload, "reasoning.effort", "high")
334		− case "gpt-5.1-codex-max-xhigh":
335		− payload, _ = sjson.SetBytes(payload, "reasoning.effort", "xhigh")
336		− }
337		− }
338		− return payload
339		−}
340		−
341	269	func tokenizerForCodexModel(model string) (tokenizer.Codec, error) {
342	270	sanitized := strings.ToLower(strings.TrimSpace(model))
343	271	switch {

internal/runtime/executor/gemini_executor.go+8−2

@@ -75,6 +75,8 @@ func (e GeminiExecutor) Execute(ctx context.Context, auth cliproxyauth.Auth, r · 74 unmodified lines
75	75	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
76	76	defer reporter.trackFailure(ctx, &err)
77	77
	78	+ upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
	79	+
78	80	// Official Gemini API via API key or OAuth bearer
79	81	from := opts.SourceFormat
80	82	to := sdktranslator.FromString("gemini")
@@ -85,6 +87,7 @@ func (e GeminiExecutor) Execute(ctx context.Context, auth cliproxyauth.Auth, r · 4 unmodified lines
85	87	body = util.StripThinkingConfigIfUnsupported(req.Model, body)
86	88	body = fixGeminiImageAspectRatio(req.Model, body)
87	89	body = applyPayloadConfig(e.cfg, req.Model, body)
	90	+ body, _ = sjson.SetBytes(body, "model", upstreamModel)
88	91
89	92	action := "generateContent"
90	93	if req.Metadata != nil {
@@ -93,7 +96,7 @@ func (e GeminiExecutor) Execute(ctx context.Context, auth cliproxyauth.Auth, r · 2 unmodified lines
93	96	}
94	97	}
95	98	baseURL := resolveGeminiBaseURL(auth)
96		− url := fmt.Sprintf("%s/%s/models/%s:%s", baseURL, glAPIVersion, req.Model, action)
	99	+ url := fmt.Sprintf("%s/%s/models/%s:%s", baseURL, glAPIVersion, upstreamModel, action)
97	100	if opts.Alt != "" && action != "countTokens" {
98	101	url = url + fmt.Sprintf("?$alt=%s", opts.Alt)
99	102	}
@@ -167,6 +170,8 @@ func (e GeminiExecutor) ExecuteStream(ctx context.Context, auth cliproxyauth.A · 67 unmodified lines
167	170	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
168	171	defer reporter.trackFailure(ctx, &err)
169	172
	173	+ upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
	174	+
170	175	from := opts.SourceFormat
171	176	to := sdktranslator.FromString("gemini")
172	177	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
@@ -176,9 +181,10 @@ func (e GeminiExecutor) ExecuteStream(ctx context.Context, auth cliproxyauth.A · 3 unmodified lines
176	181	body = util.StripThinkingConfigIfUnsupported(req.Model, body)
177	182	body = fixGeminiImageAspectRatio(req.Model, body)
178	183	body = applyPayloadConfig(e.cfg, req.Model, body)
	184	+ body, _ = sjson.SetBytes(body, "model", upstreamModel)
179	185
180	186	baseURL := resolveGeminiBaseURL(auth)
181		− url := fmt.Sprintf("%s/%s/models/%s:%s", baseURL, glAPIVersion, req.Model, "streamGenerateContent")
	187	+ url := fmt.Sprintf("%s/%s/models/%s:%s", baseURL, glAPIVersion, upstreamModel, "streamGenerateContent")
182	188	if opts.Alt == "" {
183	189	url = url + "?alt=sse"
184	190	} else {

internal/runtime/executor/gemini_vertex_executor.go+29−11

@@ -105,10 +105,12 @@ func (e GeminiVertexExecutor) CountTokens(ctx context.Context, auth cliproxyau · 104 unmodified lines
105	105
106	106	// countTokensWithServiceAccount handles token counting using service account credentials.
107	107	func (e GeminiVertexExecutor) countTokensWithServiceAccount(ctx context.Context, auth cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options, projectID, location string, saJSON []byte) (cliproxyexecutor.Response, error) {
	108	+ upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
	109	+
108	110	from := opts.SourceFormat
109	111	to := sdktranslator.FromString("gemini")
110	112	translatedReq := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
111		− if budgetOverride, includeOverride, ok := util.GeminiThinkingFromMetadata(req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
	113	+ if budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(req.Model, req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
112	114	if budgetOverride != nil {
113	115	norm := util.NormalizeThinkingBudget(req.Model, *budgetOverride)
114	116	budgetOverride = &norm
@@ -117,13 +119,14 @@ func (e *GeminiVertexExecutor) countTokensWithServiceAccount(ctx context.Context · 2 unmodified lines
117	119	}
118	120	translatedReq = util.StripThinkingConfigIfUnsupported(req.Model, translatedReq)
119	121	translatedReq = fixGeminiImageAspectRatio(req.Model, translatedReq)
	122	+ translatedReq, _ = sjson.SetBytes(translatedReq, "model", upstreamModel)
120	123	respCtx := context.WithValue(ctx, "alt", opts.Alt)
121	124	translatedReq, _ = sjson.DeleteBytes(translatedReq, "tools")
122	125	translatedReq, _ = sjson.DeleteBytes(translatedReq, "generationConfig")
123	126	translatedReq, _ = sjson.DeleteBytes(translatedReq, "safetySettings")
124	127
125	128	baseURL := vertexBaseURL(location)
126		− url := fmt.Sprintf("%s/%s/projects/%s/locations/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, projectID, location, req.Model, "countTokens")
	129	+ url := fmt.Sprintf("%s/%s/projects/%s/locations/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, projectID, location, upstreamModel, "countTokens")
127	130
128	131	httpReq, errNewReq := http.NewRequestWithContext(respCtx, http.MethodPost, url, bytes.NewReader(translatedReq))
129	132	if errNewReq != nil {
@@ -191,10 +194,12 @@ func (e *GeminiVertexExecutor) countTokensWithServiceAccount(ctx context.Context · 61 unmodified lines
191	194
192	195	// countTokensWithAPIKey handles token counting using API key credentials.
193	196	func (e GeminiVertexExecutor) countTokensWithAPIKey(ctx context.Context, auth cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options, apiKey, baseURL string) (cliproxyexecutor.Response, error) {
	197	+ upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
	198	+
194	199	from := opts.SourceFormat
195	200	to := sdktranslator.FromString("gemini")
196	201	translatedReq := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
197		− if budgetOverride, includeOverride, ok := util.GeminiThinkingFromMetadata(req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
	202	+ if budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(req.Model, req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
198	203	if budgetOverride != nil {
199	204	norm := util.NormalizeThinkingBudget(req.Model, *budgetOverride)
200	205	budgetOverride = &norm
@@ -203,6 +208,7 @@ func (e GeminiVertexExecutor) countTokensWithAPIKey(ctx context.Context, auth · 2 unmodified lines
203	208	}
204	209	translatedReq = util.StripThinkingConfigIfUnsupported(req.Model, translatedReq)
205	210	translatedReq = fixGeminiImageAspectRatio(req.Model, translatedReq)
	211	+ translatedReq, _ = sjson.SetBytes(translatedReq, "model", upstreamModel)
206	212	respCtx := context.WithValue(ctx, "alt", opts.Alt)
207	213	translatedReq, _ = sjson.DeleteBytes(translatedReq, "tools")
208	214	translatedReq, _ = sjson.DeleteBytes(translatedReq, "generationConfig")
@@ -286,10 +292,12 @@ func (e *GeminiVertexExecutor) executeWithServiceAccount(ctx context.Context, au · 77 unmodified lines
286	292	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
287	293	defer reporter.trackFailure(ctx, &err)
288	294
	295	+ upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
	296	+
289	297	from := opts.SourceFormat
290	298	to := sdktranslator.FromString("gemini")
291	299	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
292		− if budgetOverride, includeOverride, ok := util.GeminiThinkingFromMetadata(req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
	300	+ if budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(req.Model, req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
293	301	if budgetOverride != nil {
294	302	norm := util.NormalizeThinkingBudget(req.Model, *budgetOverride)
295	303	budgetOverride = &norm
@@ -301,6 +309,7 @@ func (e *GeminiVertexExecutor) executeWithServiceAccount(ctx context.Context, au · 5 unmodified lines
301	309	body = util.StripThinkingConfigIfUnsupported(req.Model, body)
302	310	body = fixGeminiImageAspectRatio(req.Model, body)
303	311	body = applyPayloadConfig(e.cfg, req.Model, body)
	312	+ body, _ = sjson.SetBytes(body, "model", upstreamModel)
304	313
305	314	action := "generateContent"
306	315	if req.Metadata != nil {
@@ -309,7 +318,7 @@ func (e *GeminiVertexExecutor) executeWithServiceAccount(ctx context.Context, au · 2 unmodified lines
309	318	}
310	319	}
311	320	baseURL := vertexBaseURL(location)
312		− url := fmt.Sprintf("%s/%s/projects/%s/locations/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, projectID, location, req.Model, action)
	321	+ url := fmt.Sprintf("%s/%s/projects/%s/locations/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, projectID, location, upstreamModel, action)
313	322	if opts.Alt != "" && action != "countTokens" {
314	323	url = url + fmt.Sprintf("?$alt=%s", opts.Alt)
315	324	}
@@ -383,10 +392,12 @@ func (e GeminiVertexExecutor) executeWithAPIKey(ctx context.Context, auth clip · 67 unmodified lines
383	392	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
384	393	defer reporter.trackFailure(ctx, &err)
385	394
	395	+ upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
	396	+
386	397	from := opts.SourceFormat
387	398	to := sdktranslator.FromString("gemini")
388	399	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
389		− if budgetOverride, includeOverride, ok := util.GeminiThinkingFromMetadata(req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
	400	+ if budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(req.Model, req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
390	401	if budgetOverride != nil {
391	402	norm := util.NormalizeThinkingBudget(req.Model, *budgetOverride)
392	403	budgetOverride = &norm
@@ -398,6 +409,7 @@ func (e GeminiVertexExecutor) executeWithAPIKey(ctx context.Context, auth clip · 5 unmodified lines
398	409	body = util.StripThinkingConfigIfUnsupported(req.Model, body)
399	410	body = fixGeminiImageAspectRatio(req.Model, body)
400	411	body = applyPayloadConfig(e.cfg, req.Model, body)
	412	+ body, _ = sjson.SetBytes(body, "model", upstreamModel)
401	413
402	414	action := "generateContent"
403	415	if req.Metadata != nil {
@@ -410,7 +422,7 @@ func (e GeminiVertexExecutor) executeWithAPIKey(ctx context.Context, auth clip · 6 unmodified lines
410	422	if baseURL == "" {
411	423	baseURL = "https://generativelanguage.googleapis.com"
412	424	}
413		− url := fmt.Sprintf("%s/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, req.Model, action)
	425	+ url := fmt.Sprintf("%s/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, upstreamModel, action)
414	426	if opts.Alt != "" && action != "countTokens" {
415	427	url = url + fmt.Sprintf("?$alt=%s", opts.Alt)
416	428	}
@@ -481,10 +493,12 @@ func (e *GeminiVertexExecutor) executeStreamWithServiceAccount(ctx context.Conte · 64 unmodified lines
481	493	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
482	494	defer reporter.trackFailure(ctx, &err)
483	495
	496	+ upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
	497	+
484	498	from := opts.SourceFormat
485	499	to := sdktranslator.FromString("gemini")
486	500	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
487		− if budgetOverride, includeOverride, ok := util.GeminiThinkingFromMetadata(req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
	501	+ if budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(req.Model, req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
488	502	if budgetOverride != nil {
489	503	norm := util.NormalizeThinkingBudget(req.Model, *budgetOverride)
490	504	budgetOverride = &norm
@@ -496,9 +510,10 @@ func (e *GeminiVertexExecutor) executeStreamWithServiceAccount(ctx context.Conte · 5 unmodified lines
496	510	body = util.StripThinkingConfigIfUnsupported(req.Model, body)
497	511	body = fixGeminiImageAspectRatio(req.Model, body)
498	512	body = applyPayloadConfig(e.cfg, req.Model, body)
	513	+ body, _ = sjson.SetBytes(body, "model", upstreamModel)
499	514
500	515	baseURL := vertexBaseURL(location)
501		− url := fmt.Sprintf("%s/%s/projects/%s/locations/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, projectID, location, req.Model, "streamGenerateContent")
	516	+ url := fmt.Sprintf("%s/%s/projects/%s/locations/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, projectID, location, upstreamModel, "streamGenerateContent")
502	517	if opts.Alt == "" {
503	518	url = url + "?alt=sse"
504	519	} else {
@@ -595,10 +610,12 @@ func (e *GeminiVertexExecutor) executeStreamWithAPIKey(ctx context.Context, auth · 90 unmodified lines
595	610	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
596	611	defer reporter.trackFailure(ctx, &err)
597	612
	613	+ upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
	614	+
598	615	from := opts.SourceFormat
599	616	to := sdktranslator.FromString("gemini")
600	617	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
601		− if budgetOverride, includeOverride, ok := util.GeminiThinkingFromMetadata(req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
	618	+ if budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(req.Model, req.Metadata); ok && util.ModelSupportsThinking(req.Model) {
602	619	if budgetOverride != nil {
603	620	norm := util.NormalizeThinkingBudget(req.Model, *budgetOverride)
604	621	budgetOverride = &norm
@@ -610,12 +627,13 @@ func (e *GeminiVertexExecutor) executeStreamWithAPIKey(ctx context.Context, auth · 5 unmodified lines
610	627	body = util.StripThinkingConfigIfUnsupported(req.Model, body)
611	628	body = fixGeminiImageAspectRatio(req.Model, body)
612	629	body = applyPayloadConfig(e.cfg, req.Model, body)
	630	+ body, _ = sjson.SetBytes(body, "model", upstreamModel)
613	631
614	632	// For API key auth, use simpler URL format without project/location
615	633	if baseURL == "" {
616	634	baseURL = "https://generativelanguage.googleapis.com"
617	635	}
618		− url := fmt.Sprintf("%s/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, req.Model, "streamGenerateContent")
	636	+ url := fmt.Sprintf("%s/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, upstreamModel, "streamGenerateContent")
619	637	if opts.Alt == "" {
620	638	url = url + "?alt=sse"
621	639	} else {

internal/runtime/executor/iflow_executor.go+8−0

@@ -57,6 +57,10 @@ func (e IFlowExecutor) Execute(ctx context.Context, auth cliproxyauth.Auth, re · 56 unmodified lines
57	57	from := opts.SourceFormat
58	58	to := sdktranslator.FromString("openai")
59	59	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
	60	+ body = applyReasoningEffortMetadataChatCompletions(body, req.Metadata, req.Model)
	61	+ if upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata); upstreamModel != "" {
	62	+ body, _ = sjson.SetBytes(body, "model", upstreamModel)
	63	+ }
60	64	body = applyPayloadConfig(e.cfg, req.Model, body)
61	65
62	66	endpoint := strings.TrimSuffix(baseURL, "/") + iflowDefaultEndpoint
@@ -139,6 +143,10 @@ func (e IFlowExecutor) ExecuteStream(ctx context.Context, auth cliproxyauth.Au · 76 unmodified lines
139	143	to := sdktranslator.FromString("openai")
140	144	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
141	145
	146	+ body = applyReasoningEffortMetadataChatCompletions(body, req.Metadata, req.Model)
	147	+ if upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata); upstreamModel != "" {
	148	+ body, _ = sjson.SetBytes(body, "model", upstreamModel)
	149	+ }
142	150	// Ensure tools array exists to avoid provider quirks similar to Qwen's behaviour.
143	151	toolsResult := gjson.GetBytes(body, "tools")
144	152	if toolsResult.Exists() && toolsResult.IsArray() && len(toolsResult.Array()) == 0 {

internal/runtime/executor/openai_compat_executor.go+8−0

@@ -58,6 +58,10 @@ func (e OpenAICompatExecutor) Execute(ctx context.Context, auth cliproxyauth.A · 57 unmodified lines
58	58	translated = e.overrideModel(translated, modelOverride)
59	59	}
60	60	translated = applyPayloadConfigWithRoot(e.cfg, req.Model, to.String(), "", translated)
	61	+ translated = applyReasoningEffortMetadataChatCompletions(translated, req.Metadata, req.Model)
	62	+ if upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata); upstreamModel != "" {
	63	+ translated, _ = sjson.SetBytes(translated, "model", upstreamModel)
	64	+ }
61	65
62	66	url := strings.TrimSuffix(baseURL, "/") + "/chat/completions"
63	67	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(translated))
@@ -143,6 +147,10 @@ func (e OpenAICompatExecutor) ExecuteStream(ctx context.Context, auth cliproxy · 79 unmodified lines
143	147	translated = e.overrideModel(translated, modelOverride)
144	148	}
145	149	translated = applyPayloadConfigWithRoot(e.cfg, req.Model, to.String(), "", translated)
	150	+ translated = applyReasoningEffortMetadataChatCompletions(translated, req.Metadata, req.Model)
	151	+ if upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata); upstreamModel != "" {
	152	+ translated, _ = sjson.SetBytes(translated, "model", upstreamModel)
	153	+ }
146	154
147	155	url := strings.TrimSuffix(baseURL, "/") + "/chat/completions"
148	156	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(translated))

internal/runtime/executor/payload_helpers.go+48−5

@@ -12,8 +12,8 @@ import ( · 11 unmodified lines
12	12	// applyThinkingMetadata applies thinking config from model suffix metadata (e.g., -reasoning, -thinking-N)
13	13	// for standard Gemini format payloads. It normalizes the budget when the model supports thinking.
14	14	func applyThinkingMetadata(payload []byte, metadata map[string]any, model string) []byte {
15		− budgetOverride, includeOverride, ok := util.GeminiThinkingFromMetadata(metadata)
16		− if !ok {
	15	+ budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(model, metadata)
	16	+ if !ok \|\| (budgetOverride == nil && includeOverride == nil) {
17	17	return payload
18	18	}
19	19	if !util.ModelSupportsThinking(model) {
@@ -29,17 +29,60 @@ func applyThinkingMetadata(payload []byte, metadata map[string]any, model string · 9 unmodified lines
29	29	// applyThinkingMetadataCLI applies thinking config from model suffix metadata (e.g., -reasoning, -thinking-N)
30	30	// for Gemini CLI format payloads (nested under "request"). It normalizes the budget when the model supports thinking.
31	31	func applyThinkingMetadataCLI(payload []byte, metadata map[string]any, model string) []byte {
32		− budgetOverride, includeOverride, ok := util.GeminiThinkingFromMetadata(metadata)
33		− if !ok {
	32	+ budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(model, metadata)
	33	+ if !ok \|\| (budgetOverride == nil && includeOverride == nil) {
34	34	return payload
35	35	}
36		− if budgetOverride != nil && util.ModelSupportsThinking(model) {
	36	+ if !util.ModelSupportsThinking(model) {
	37	+ return payload
	38	+ }
	39	+ if budgetOverride != nil {
37	40	norm := util.NormalizeThinkingBudget(model, *budgetOverride)
38	41	budgetOverride = &norm
39	42	}
40	43	return util.ApplyGeminiCLIThinkingConfig(payload, budgetOverride, includeOverride)
41	44	}
42	45
	46	+// applyReasoningEffortMetadata applies reasoning effort overrides (reasoning.effort) when present in metadata.
	47	+// It avoids overwriting an existing reasoning.effort field and only applies to models that support thinking.
	48	+func applyReasoningEffortMetadata(payload []byte, metadata map[string]any, model string) []byte {
	49	+ if len(metadata) == 0 {
	50	+ return payload
	51	+ }
	52	+ if !util.ModelSupportsThinking(model) {
	53	+ return payload
	54	+ }
	55	+ if gjson.GetBytes(payload, "reasoning.effort").Exists() {
	56	+ return payload
	57	+ }
	58	+ if effort, ok := util.ReasoningEffortFromMetadata(metadata); ok && effort != "" {
	59	+ if updated, err := sjson.SetBytes(payload, "reasoning.effort", effort); err == nil {
	60	+ return updated
	61	+ }
	62	+ }
	63	+ return payload
	64	+}
	65	+
	66	+// applyReasoningEffortMetadataChatCompletions applies reasoning_effort (OpenAI chat completions field)
	67	+// when present in metadata. It avoids overwriting an existing reasoning_effort field.
	68	+func applyReasoningEffortMetadataChatCompletions(payload []byte, metadata map[string]any, model string) []byte {
	69	+ if len(metadata) == 0 {
	70	+ return payload
	71	+ }
	72	+ if !util.ModelSupportsThinking(model) {
	73	+ return payload
	74	+ }
	75	+ if gjson.GetBytes(payload, "reasoning_effort").Exists() {
	76	+ return payload
	77	+ }
	78	+ if effort, ok := util.ReasoningEffortFromMetadata(metadata); ok && effort != "" {
	79	+ if updated, err := sjson.SetBytes(payload, "reasoning_effort", effort); err == nil {
	80	+ return updated
	81	+ }
	82	+ }
	83	+ return payload
	84	+}
	85	+
43	86	// applyPayloadConfig applies payload default and override rules from configuration
44	87	// to the given JSON payload for the specified model.
45	88	// Defaults only fill missing fields, while overrides always overwrite existing values.

internal/runtime/executor/qwen_executor.go+9−0

@@ -12,6 +12,7 @@ import ( · 11 unmodified lines
12	12
13	13	qwenauth "github.com/router-for-me/CLIProxyAPI/v6/internal/auth/qwen"
14	14	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
	15	+ "github.com/router-for-me/CLIProxyAPI/v6/internal/util"
15	16	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
16	17	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
17	18	sdktranslator "github.com/router-for-me/CLIProxyAPI/v6/sdk/translator"
@@ -50,6 +51,10 @@ func (e QwenExecutor) Execute(ctx context.Context, auth cliproxyauth.Auth, req · 32 unmodified lines
50	51	from := opts.SourceFormat
51	52	to := sdktranslator.FromString("openai")
52	53	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
	54	+ body = applyReasoningEffortMetadataChatCompletions(body, req.Metadata, req.Model)
	55	+ if upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata); upstreamModel != "" {
	56	+ body, _ = sjson.SetBytes(body, "model", upstreamModel)
	57	+ }
53	58	body = applyPayloadConfig(e.cfg, req.Model, body)
54	59
55	60	url := strings.TrimSuffix(baseURL, "/") + "/chat/completions"
@@ -121,6 +126,10 @@ func (e QwenExecutor) ExecuteStream(ctx context.Context, auth cliproxyauth.Aut · 65 unmodified lines
121	126	to := sdktranslator.FromString("openai")
122	127	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
123	128
	129	+ body = applyReasoningEffortMetadataChatCompletions(body, req.Metadata, req.Model)
	130	+ if upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata); upstreamModel != "" {
	131	+ body, _ = sjson.SetBytes(body, "model", upstreamModel)
	132	+ }
124	133	toolsResult := gjson.GetBytes(body, "tools")
125	134	// I'm addressing the Qwen3 "poisoning" issue, which is caused by the model needing a tool to be defined. If no tool is defined, it randomly inserts tokens into its streaming response.
126	135	// This will have no real consequences. It's just to scare Qwen3.

internal/util/gemini_thinking.go+0−150

@@ -1,8 +1,6 @@
1	1	package util
2	2
3	3	import (
4		− "encoding/json"
5		− "strconv"
6	4	"strings"
7	5
8	6	"github.com/tidwall/gjson"
@@ -15,80 +13,6 @@ const ( · 6 unmodified lines
15	13	GeminiOriginalModelMetadataKey = "gemini_original_model"
16	14	)
17	15
18		−func ParseGeminiThinkingSuffix(model string) (string, int, bool, bool) {
19		− if model == "" {
20		− return model, nil, nil, false
21		− }
22		− lower := strings.ToLower(model)
23		− if !strings.HasPrefix(lower, "gemini-") {
24		− return model, nil, nil, false
25		− }
26		−
27		− if strings.HasSuffix(lower, "-nothinking") {
28		− base := model[:len(model)-len("-nothinking")]
29		− budgetValue := 0
30		− if strings.HasPrefix(lower, "gemini-2.5-pro") {
31		− budgetValue = 128
32		− }
33		− include := false
34		− return base, &budgetValue, &include, true
35		− }
36		−
37		− // Handle "-reasoning" suffix: enables thinking with dynamic budget (-1)
38		− // Maps: gemini-2.5-flash-reasoning -> gemini-2.5-flash with thinkingBudget=-1
39		− if strings.HasSuffix(lower, "-reasoning") {
40		− base := model[:len(model)-len("-reasoning")]
41		− budgetValue := -1 // Dynamic budget
42		− include := true
43		− return base, &budgetValue, &include, true
44		− }
45		−
46		− idx := strings.LastIndex(lower, "-thinking-")
47		− if idx == -1 {
48		− return model, nil, nil, false
49		− }
50		−
51		− digits := model[idx+len("-thinking-"):]
52		− if digits == "" {
53		− return model, nil, nil, false
54		− }
55		− end := len(digits)
56		− for i := 0; i < len(digits); i++ {
57		− if digits[i] < '0' \|\| digits[i] > '9' {
58		− end = i
59		− break
60		− }
61		− }
62		− if end == 0 {
63		− return model, nil, nil, false
64		− }
65		− valueStr := digits[:end]
66		− value, err := strconv.Atoi(valueStr)
67		− if err != nil {
68		− return model, nil, nil, false
69		− }
70		− base := model[:idx]
71		− budgetValue := value
72		− return base, &budgetValue, nil, true
73		−}
74		−
75		−func NormalizeGeminiThinkingModel(modelName string) (string, map[string]any) {
76		− baseModel, budget, include, matched := ParseGeminiThinkingSuffix(modelName)
77		− if !matched {
78		− return baseModel, nil
79		− }
80		− metadata := map[string]any{
81		− GeminiOriginalModelMetadataKey: modelName,
82		− }
83		− if budget != nil {
84		− metadata[GeminiThinkingBudgetMetadataKey] = *budget
85		− }
86		− if include != nil {
87		− metadata[GeminiIncludeThoughtsMetadataKey] = *include
88		− }
89		− return baseModel, metadata
90		−}
91		−
92	16	func ApplyGeminiThinkingConfig(body []byte, budget int, includeThoughts bool) []byte {
93	17	if budget == nil && includeThoughts == nil {
94	18	return body
@@ -133,80 +57,6 @@ func ApplyGeminiCLIThinkingConfig(body []byte, budget int, includeThoughts boo · 38 unmodified lines
133	57	return updated
134	58	}
135	59
136		−func GeminiThinkingFromMetadata(metadata map[string]any) (int, bool, bool) {
137		− if len(metadata) == 0 {
138		− return nil, nil, false
139		− }
140		− var (
141		− budgetPtr *int
142		− includePtr *bool
143		− matched bool
144		− )
145		− if rawBudget, ok := metadata[GeminiThinkingBudgetMetadataKey]; ok {
146		− switch v := rawBudget.(type) {
147		− case int:
148		− budget := v
149		− budgetPtr = &budget
150		− matched = true
151		− case int32:
152		− budget := int(v)
153		− budgetPtr = &budget
154		− matched = true
155		− case int64:
156		− budget := int(v)
157		− budgetPtr = &budget
158		− matched = true
159		− case float64:
160		− budget := int(v)
161		− budgetPtr = &budget
162		− matched = true
163		− case json.Number:
164		− if val, err := v.Int64(); err == nil {
165		− budget := int(val)
166		− budgetPtr = &budget
167		− matched = true
168		− }
169		− }
170		− }
171		− if rawInclude, ok := metadata[GeminiIncludeThoughtsMetadataKey]; ok {
172		− switch v := rawInclude.(type) {
173		− case bool:
174		− include := v
175		− includePtr = &include
176		− matched = true
177		− case string:
178		− if parsed, err := strconv.ParseBool(v); err == nil {
179		− include := parsed
180		− includePtr = &include
181		− matched = true
182		− }
183		− case json.Number:
184		− if val, err := v.Int64(); err == nil {
185		− include := val != 0
186		− includePtr = &include
187		− matched = true
188		− }
189		− case int:
190		− include := v != 0
191		− includePtr = &include
192		− matched = true
193		− case int32:
194		− include := v != 0
195		− includePtr = &include
196		− matched = true
197		− case int64:
198		− include := v != 0
199		− includePtr = &include
200		− matched = true
201		− case float64:
202		− include := v != 0
203		− includePtr = &include
204		− matched = true
205		− }
206		− }
207		− return budgetPtr, includePtr, matched
208		−}
209		−
210	60	// modelsWithDefaultThinking lists models that should have thinking enabled by default
211	61	// when no explicit thinkingConfig is provided.
212	62	var modelsWithDefaultThinking = map[string]bool{

internal/util/thinking_suffix.go+327−0

@@ -0,0 +1,327 @@
	1	+package util
	2	+
	3	+import (
	4	+ "encoding/json"
	5	+ "strconv"
	6	+ "strings"
	7	+)
	8	+
	9	+const (
	10	+ ThinkingBudgetMetadataKey = "thinking_budget"
	11	+ ThinkingIncludeThoughtsMetadataKey = "thinking_include_thoughts"
	12	+ ReasoningEffortMetadataKey = "reasoning_effort"
	13	+ ThinkingOriginalModelMetadataKey = "thinking_original_model"
	14	+)
	15	+
	16	+// NormalizeThinkingModel parses dynamic thinking suffixes on model names and returns
	17	+// the normalized base model with extracted metadata. Supported patterns:
	18	+// - "-thinking-<number>" extracts a numeric budget
	19	+// - "-thinking-<level>" extracts a reasoning effort level (minimal/low/medium/high/xhigh/auto/none)
	20	+// - "-thinking" maps to a default reasoning effort of "medium"
	21	+// - "-reasoning" maps to dynamic budget (-1) and include_thoughts=true
	22	+// - "-nothinking" maps to budget=0 and include_thoughts=false
	23	+func NormalizeThinkingModel(modelName string) (string, map[string]any) {
	24	+ if modelName == "" {
	25	+ return modelName, nil
	26	+ }
	27	+
	28	+ lower := strings.ToLower(modelName)
	29	+ baseModel := modelName
	30	+
	31	+ var (
	32	+ budgetOverride *int
	33	+ includeThoughts *bool
	34	+ reasoningEffort *string
	35	+ matched bool
	36	+ )
	37	+
	38	+ switch {
	39	+ case strings.HasSuffix(lower, "-nothinking"):
	40	+ baseModel = modelName[:len(modelName)-len("-nothinking")]
	41	+ budget := 0
	42	+ include := false
	43	+ budgetOverride = &budget
	44	+ includeThoughts = &include
	45	+ matched = true
	46	+ case strings.HasSuffix(lower, "-reasoning"):
	47	+ baseModel = modelName[:len(modelName)-len("-reasoning")]
	48	+ budget := -1
	49	+ include := true
	50	+ budgetOverride = &budget
	51	+ includeThoughts = &include
	52	+ matched = true
	53	+ default:
	54	+ if idx := strings.LastIndex(lower, "-thinking-"); idx != -1 {
	55	+ value := modelName[idx+len("-thinking-"):]
	56	+ if value != "" {
	57	+ if parsed, ok := parseIntPrefix(value); ok {
	58	+ baseModel = modelName[:idx]
	59	+ budgetOverride = &parsed
	60	+ matched = true
	61	+ } else if effort, okEffort := normalizeReasoningEffort(value); okEffort {
	62	+ baseModel = modelName[:idx]
	63	+ reasoningEffort = &effort
	64	+ matched = true
	65	+ }
	66	+ }
	67	+ } else if strings.HasSuffix(lower, "-thinking") {
	68	+ baseModel = modelName[:len(modelName)-len("-thinking")]
	69	+ effort := "medium"
	70	+ reasoningEffort = &effort
	71	+ matched = true
	72	+ }
	73	+ }
	74	+
	75	+ if !matched {
	76	+ return baseModel, nil
	77	+ }
	78	+
	79	+ metadata := map[string]any{
	80	+ ThinkingOriginalModelMetadataKey: modelName,
	81	+ }
	82	+ if budgetOverride != nil {
	83	+ metadata[ThinkingBudgetMetadataKey] = *budgetOverride
	84	+ }
	85	+ if includeThoughts != nil {
	86	+ metadata[ThinkingIncludeThoughtsMetadataKey] = *includeThoughts
	87	+ }
	88	+ if reasoningEffort != nil {
	89	+ metadata[ReasoningEffortMetadataKey] = *reasoningEffort
	90	+ }
	91	+ return baseModel, metadata
	92	+}
	93	+
	94	+// ThinkingFromMetadata extracts thinking overrides from metadata produced by NormalizeThinkingModel.
	95	+// It accepts both the new generic keys and legacy Gemini-specific keys.
	96	+func ThinkingFromMetadata(metadata map[string]any) (int, bool, *string, bool) {
	97	+ if len(metadata) == 0 {
	98	+ return nil, nil, nil, false
	99	+ }
	100	+
	101	+ var (
	102	+ budgetPtr *int
	103	+ includePtr *bool
	104	+ effortPtr *string
	105	+ matched bool
	106	+ )
	107	+
	108	+ readBudget := func(key string) {
	109	+ if budgetPtr != nil {
	110	+ return
	111	+ }
	112	+ if raw, ok := metadata[key]; ok {
	113	+ if v, okNumber := parseNumberToInt(raw); okNumber {
	114	+ budget := v
	115	+ budgetPtr = &budget
	116	+ matched = true
	117	+ }
	118	+ }
	119	+ }
	120	+
	121	+ readInclude := func(key string) {
	122	+ if includePtr != nil {
	123	+ return
	124	+ }
	125	+ if raw, ok := metadata[key]; ok {
	126	+ switch v := raw.(type) {
	127	+ case bool:
	128	+ val := v
	129	+ includePtr = &val
	130	+ matched = true
	131	+ case *bool:
	132	+ if v != nil {
	133	+ val := *v
	134	+ includePtr = &val
	135	+ matched = true
	136	+ }
	137	+ }
	138	+ }
	139	+ }
	140	+
	141	+ readEffort := func(key string) {
	142	+ if effortPtr != nil {
	143	+ return
	144	+ }
	145	+ if raw, ok := metadata[key]; ok {
	146	+ if val, okStr := raw.(string); okStr && strings.TrimSpace(val) != "" {
	147	+ normalized := strings.ToLower(strings.TrimSpace(val))
	148	+ effortPtr = &normalized
	149	+ matched = true
	150	+ }
	151	+ }
	152	+ }
	153	+
	154	+ readBudget(ThinkingBudgetMetadataKey)
	155	+ readBudget(GeminiThinkingBudgetMetadataKey)
	156	+ readInclude(ThinkingIncludeThoughtsMetadataKey)
	157	+ readInclude(GeminiIncludeThoughtsMetadataKey)
	158	+ readEffort(ReasoningEffortMetadataKey)
	159	+ readEffort("reasoning.effort")
	160	+
	161	+ return budgetPtr, includePtr, effortPtr, matched
	162	+}
	163	+
	164	+// ResolveThinkingConfigFromMetadata derives thinking budget/include overrides,
	165	+// converting reasoning effort strings into budgets when possible.
	166	+func ResolveThinkingConfigFromMetadata(model string, metadata map[string]any) (int, bool, bool) {
	167	+ budget, include, effort, matched := ThinkingFromMetadata(metadata)
	168	+ if !matched {
	169	+ return nil, nil, false
	170	+ }
	171	+
	172	+ if budget == nil && effort != nil {
	173	+ if derived, ok := ThinkingEffortToBudget(model, *effort); ok {
	174	+ budget = &derived
	175	+ }
	176	+ }
	177	+ return budget, include, budget != nil \|\| include != nil \|\| effort != nil
	178	+}
	179	+
	180	+// ReasoningEffortFromMetadata resolves a reasoning effort string from metadata,
	181	+// inferring "auto" and "none" when budgets request dynamic or disabled thinking.
	182	+func ReasoningEffortFromMetadata(metadata map[string]any) (string, bool) {
	183	+ budget, include, effort, matched := ThinkingFromMetadata(metadata)
	184	+ if !matched {
	185	+ return "", false
	186	+ }
	187	+ if effort != nil && *effort != "" {
	188	+ return *effort, true
	189	+ }
	190	+ if budget != nil {
	191	+ switch *budget {
	192	+ case -1:
	193	+ return "auto", true
	194	+ case 0:
	195	+ return "none", true
	196	+ }
	197	+ }
	198	+ if include != nil && !*include {
	199	+ return "none", true
	200	+ }
	201	+ return "", true
	202	+}
	203	+
	204	+// ThinkingEffortToBudget maps reasoning effort levels to approximate budgets,
	205	+// clamping the result to the model's supported range.
	206	+func ThinkingEffortToBudget(model, effort string) (int, bool) {
	207	+ if effort == "" {
	208	+ return 0, false
	209	+ }
	210	+ switch strings.ToLower(effort) {
	211	+ case "none":
	212	+ return 0, true
	213	+ case "auto":
	214	+ return NormalizeThinkingBudget(model, -1), true
	215	+ case "minimal":
	216	+ return NormalizeThinkingBudget(model, 512), true
	217	+ case "low":
	218	+ return NormalizeThinkingBudget(model, 1024), true
	219	+ case "medium":
	220	+ return NormalizeThinkingBudget(model, 8192), true
	221	+ case "high":
	222	+ return NormalizeThinkingBudget(model, 24576), true
	223	+ case "xhigh":
	224	+ return NormalizeThinkingBudget(model, 32768), true
	225	+ default:
	226	+ return 0, false
	227	+ }
	228	+}
	229	+
	230	+// ResolveOriginalModel returns the original model name stored in metadata (if present),
	231	+// otherwise falls back to the provided model.
	232	+func ResolveOriginalModel(model string, metadata map[string]any) string {
	233	+ normalize := func(name string) string {
	234	+ if name == "" {
	235	+ return ""
	236	+ }
	237	+ if base, _ := NormalizeThinkingModel(name); base != "" {
	238	+ return base
	239	+ }
	240	+ return strings.TrimSpace(name)
	241	+ }
	242	+
	243	+ if metadata != nil {
	244	+ if v, ok := metadata[ThinkingOriginalModelMetadataKey]; ok {
	245	+ if s, okStr := v.(string); okStr && strings.TrimSpace(s) != "" {
	246	+ if base := normalize(s); base != "" {
	247	+ return base
	248	+ }
	249	+ }
	250	+ }
	251	+ if v, ok := metadata[GeminiOriginalModelMetadataKey]; ok {
	252	+ if s, okStr := v.(string); okStr && strings.TrimSpace(s) != "" {
	253	+ if base := normalize(s); base != "" {
	254	+ return base
	255	+ }
	256	+ }
	257	+ }
	258	+ }
	259	+ // Fallback: try to re-normalize the model name when metadata was dropped.
	260	+ if base := normalize(model); base != "" {
	261	+ return base
	262	+ }
	263	+ return model
	264	+}
	265	+
	266	+func parseIntPrefix(value string) (int, bool) {
	267	+ if value == "" {
	268	+ return 0, false
	269	+ }
	270	+ digits := strings.TrimLeft(value, "-")
	271	+ if digits == "" {
	272	+ return 0, false
	273	+ }
	274	+ end := len(digits)
	275	+ for i := 0; i < len(digits); i++ {
	276	+ if digits[i] < '0' \|\| digits[i] > '9' {
	277	+ end = i
	278	+ break
	279	+ }
	280	+ }
	281	+ if end == 0 {
	282	+ return 0, false
	283	+ }
	284	+ val, err := strconv.Atoi(digits[:end])
	285	+ if err != nil {
	286	+ return 0, false
	287	+ }
	288	+ return val, true
	289	+}
	290	+
	291	+func parseNumberToInt(raw any) (int, bool) {
	292	+ switch v := raw.(type) {
	293	+ case int:
	294	+ return v, true
	295	+ case int32:
	296	+ return int(v), true
	297	+ case int64:
	298	+ return int(v), true
	299	+ case float64:
	300	+ return int(v), true
	301	+ case json.Number:
	302	+ if val, err := v.Int64(); err == nil {
	303	+ return int(val), true
	304	+ }
	305	+ case string:
	306	+ if strings.TrimSpace(v) == "" {
	307	+ return 0, false
	308	+ }
	309	+ if parsed, err := strconv.Atoi(strings.TrimSpace(v)); err == nil {
	310	+ return parsed, true
	311	+ }
	312	+ }
	313	+ return 0, false
	314	+}
	315	+
	316	+func normalizeReasoningEffort(value string) (string, bool) {
	317	+ if value == "" {
	318	+ return "", false
	319	+ }
	320	+ effort := strings.ToLower(strings.TrimSpace(value))
	321	+ switch effort {
	322	+ case "minimal", "low", "medium", "high", "xhigh", "auto", "none":
	323	+ return effort, true
	324	+ default:
	325	+ return "", false
	326	+ }
	327	+}

sdk/api/handlers/handlers.go+21−7

@@ -343,18 +343,32 @@ func (h *BaseAPIHandler) getRequestDetails(modelName string) (providers []string · 342 unmodified lines
343	343
344	344	providerName, extractedModelName, isDynamic := h.parseDynamicModel(resolvedModelName)
345	345
346		− // First, normalize the model name to handle suffixes like "-thinking-128"
347		− // This needs to happen before determining the provider for non-dynamic models.
348		− normalizedModel, metadata = normalizeModelMetadata(resolvedModelName)
	346	+ targetModelName := resolvedModelName
	347	+ if isDynamic {
	348	+ targetModelName = extractedModelName
	349	+ }
	350	+
	351	+ // Normalize the model name to handle dynamic thinking suffixes before determining the provider.
	352	+ normalizedModel, metadata = normalizeModelMetadata(targetModelName)
349	353
350	354	if isDynamic {
351	355	providers = []string{providerName}
352		− // For dynamic models, the extractedModelName is already normalized by parseDynamicModel
353		− // so we use it as the final normalizedModel.
354		− normalizedModel = extractedModelName
355	356	} else {
356	357	// For non-dynamic models, use the normalizedModel to get the provider name.
357	358	providers = util.GetProviderName(normalizedModel)
	359	+ if len(providers) == 0 && metadata != nil {
	360	+ if originalRaw, ok := metadata[util.ThinkingOriginalModelMetadataKey]; ok {
	361	+ if originalModel, okStr := originalRaw.(string); okStr {
	362	+ originalModel = strings.TrimSpace(originalModel)
	363	+ if originalModel != "" && !strings.EqualFold(originalModel, normalizedModel) {
	364	+ if altProviders := util.GetProviderName(originalModel); len(altProviders) > 0 {
	365	+ providers = altProviders
	366	+ normalizedModel = originalModel
	367	+ }
	368	+ }
	369	+ }
	370	+ }
	371	+ }
358	372	}
359	373
360	374	if len(providers) == 0 {
@@ -402,7 +416,7 @@ func cloneBytes(src []byte) []byte { · 41 unmodified lines
402	416	}
403	417
404	418	func normalizeModelMetadata(modelName string) (string, map[string]any) {
405		− return util.NormalizeGeminiThinkingModel(modelName)
	419	+ return util.NormalizeThinkingModel(modelName)
406	420	}
407	421
408	422	func cloneMetadata(src map[string]any) map[string]any {