Fix DeepSeek max reasoning effort

commit: c2267b3f6d6674c51b183ca9b59d700bd1d1b5a5
author: patriceckhart · 27d ago · 2026-09-25 06:04 UTC
parent: c5883a4
signature: unsigned
refs: tag: v0.3.97

4 files · +42 -1

M README.md +1 -1


@@ -701,7 +701,7 @@ By default this uses:
701	701		- model: `deepseek-v4-pro`
702	702		- base URL: `https://api.deepseek.com/v1`
703	703
704		-	Catalog ships with `deepseek-v4-pro` (reasoning) and `deepseek-v4-flash`. These are exactly the IDs returned by `GET https://api.deepseek.com/models` today. You can add additional model IDs to `models.json` under the `deepseek` provider.
	704	+	Catalog ships with `deepseek-v4-pro` (reasoning) and `deepseek-v4-flash`. These are exactly the IDs returned by `GET https://api.deepseek.com/models` today. You can add additional model IDs to `models.json` under the `deepseek` provider. For direct DeepSeek reasoning models, `/reasoning` offers off, low, high, and max. DeepSeek maps medium and xhigh to high, while max sends the distinct `reasoning_effort: "max"` value. Other OpenAI-compatible providers retain their own effort limits.
705	705
706	706		Credential lookup order for DeepSeek:
707	707


@@ -701,7 +701,7 @@ By default this uses:
701		- model: `deepseek-v4-pro`	701		- model: `deepseek-v4-pro`
702		- base URL: `https://api.deepseek.com/v1`	702		- base URL: `https://api.deepseek.com/v1`
703			703
704	-	Catalog ships with `deepseek-v4-pro` (reasoning) and `deepseek-v4-flash`. These are exactly the IDs returned by `GET https://api.deepseek.com/models` today. You can add additional model IDs to `models.json` under the `deepseek` provider.	704	+	Catalog ships with `deepseek-v4-pro` (reasoning) and `deepseek-v4-flash`. These are exactly the IDs returned by `GET https://api.deepseek.com/models` today. You can add additional model IDs to `models.json` under the `deepseek` provider. For direct DeepSeek reasoning models, `/reasoning` offers off, low, high, and max. DeepSeek maps medium and xhigh to high, while max sends the distinct `reasoning_effort: "max"` value. Other OpenAI-compatible providers retain their own effort limits.
705			705
706		Credential lookup order for DeepSeek:	706		Credential lookup order for DeepSeek:
707			707

M packages/provider/openai.go +3 -0


@@ -263,6 +263,9 @@ func (c openaiClient) buildRequest(req Request) (oaiRequest, error) {
263	263		out.MaxCompletionTok = &maxTok
264	264		}
265	265		effort := OpenAIReasoningEffort(reasoning)
	266	+	if c.Name() == "deepseek" && reasoning == "max" {
	267	+	effort = "max"
	268	+	}
266	269		if usesAdaptiveThinking(m) {
267	270		// Some gateways expose adaptive-thinking Anthropic models through
268	271		// the OpenAI-compatible chat-completions wire. They accept the


@@ -263,6 +263,9 @@ func (c openaiClient) buildRequest(req Request) (oaiRequest, error) {
263	out.MaxCompletionTok = &maxTok	263		out.MaxCompletionTok = &maxTok
264	}	264		}
265	effort := OpenAIReasoningEffort(reasoning)	265		effort := OpenAIReasoningEffort(reasoning)
		266	+	if c.Name() == "deepseek" && reasoning == "max" {
		267	+	effort = "max"
		268	+	}
266	if usesAdaptiveThinking(m) {	269		if usesAdaptiveThinking(m) {
267	// Some gateways expose adaptive-thinking Anthropic models through	270		// Some gateways expose adaptive-thinking Anthropic models through
268	// the OpenAI-compatible chat-completions wire. They accept the	271		// the OpenAI-compatible chat-completions wire. They accept the

M packages/provider/reasoning.go +7 -0


@@ -54,6 +54,10 @@ func defaultReasoningLevels(model Model) []string {
54	54		}
55	55		return []string{"", "minimum", "low", "medium", "high"}
56	56		}
	57	+	if model.Provider == "deepseek" {
	58	+	// DeepSeek maps medium and xhigh to high; max is a distinct effort.
	59	+	return []string{"", "low", "high", "max"}
	60	+	}
57	61		if model.AdaptiveThinkingCompat {
58	62		return []string{"", "high"}
59	63		}

@@ -112,6 +116,9 @@ func ClampReasoningForModel(model Model, level string) string {
112	116		return candidate
113	117		}
114	118		}
	119	+	if model.Provider == "deepseek" && (normalized == "medium" \|\| normalized == "xhigh") && containsReasoningLevel(available, "high") {
	120	+	return "high"
	121	+	}
115	122		return nearestReasoningLevel(available, normalized)
116	123		}
117	124


@@ -54,6 +54,10 @@ func defaultReasoningLevels(model Model) []string {
54	}	54		}
55	return []string{"", "minimum", "low", "medium", "high"}	55		return []string{"", "minimum", "low", "medium", "high"}
56	}	56		}
		57	+	if model.Provider == "deepseek" {
		58	+	// DeepSeek maps medium and xhigh to high; max is a distinct effort.
		59	+	return []string{"", "low", "high", "max"}
		60	+	}
57	if model.AdaptiveThinkingCompat {	61		if model.AdaptiveThinkingCompat {
58	return []string{"", "high"}	62		return []string{"", "high"}
59	}	63		}

@@ -112,6 +116,9 @@ func ClampReasoningForModel(model Model, level string) string {
112	return candidate	116		return candidate
113	}	117		}
114	}	118		}
		119	+	if model.Provider == "deepseek" && (normalized == "medium" \|\| normalized == "xhigh") && containsReasoningLevel(available, "high") {
		120	+	return "high"
		121	+	}
115	return nearestReasoningLevel(available, normalized)	122		return nearestReasoningLevel(available, normalized)
116	}	123		}
117		124

M packages/provider/reasoning_test.go +31 -0


@@ -13,6 +13,7 @@ func TestAvailableReasoningLevels(t *testing.T) {
13	13		}{
14	14		{name: "unsupported", model: Model{}, want: []string{""}},
15	15		{name: "generic openai", model: Model{Reasoning: true}, want: []string{"", "low", "medium", "high"}},
	16	+	{name: "direct deepseek", model: Model{Provider: "deepseek", Reasoning: true}, want: []string{"", "low", "high", "max"}},
16	17		{name: "responses", model: Model{Reasoning: true, API: APIResponses, ID: "gpt-5.5"}, want: []string{"", "low", "medium", "high", "xhigh"}},
17	18		{name: "responses native max", model: Model{Reasoning: true, API: APIResponses, ID: "gpt-5.6-sol"}, want: []string{"", "low", "medium", "high", "xhigh", "max"}},
18	19		{name: "astra native max", model: Model{Reasoning: true, API: APIResponses, ID: "gpt-6-astra"}, want: []string{"", "low", "medium", "high", "xhigh", "max"}},

@@ -63,6 +64,8 @@ func TestClampReasoningForModel(t *testing.T) {
63	64		}{
64	65		{name: "unsupported", model: Model{}, level: "high", want: ""},
65	66		{name: "generic max", model: Model{Reasoning: true}, level: "max", want: "high"},
	67	+	{name: "deepseek xhigh maps to high", model: Model{Provider: "deepseek", Reasoning: true}, level: "xhigh", want: "high"},
	68	+	{name: "deepseek max stays max", model: Model{Provider: "deepseek", Reasoning: true}, level: "max", want: "max"},
66	69		{name: "adaptive minimum", model: Model{Reasoning: true, AdaptiveThinking: true}, level: "minimum", want: "low"},
67	70		{name: "gemini pro medium", model: Model{Provider: "google", ID: "gemini-3-pro", Reasoning: true}, level: "medium", want: "high"},
68	71		{name: "native max", model: Model{Reasoning: true, API: APIResponses, ID: "gpt-5.6-sol"}, level: "max", want: "max"},

@@ -89,6 +92,34 @@ func TestClampReasoningForModel(t *testing.T) {
89	92		}
90	93		}
91	94
	95	+	func TestDeepSeekReasoningEffort(t *testing.T) {
	96	+	client := NewDeepSeek("test", "").(*openaiClient)
	97	+	for _, model := range []string{"deepseek-v4-pro", "deepseek-v4-flash"} {
	98	+	m, err := FindModel("deepseek", model)
	99	+	if err != nil {
	100	+	t.Fatal(err)
	101	+	}
	102	+	if got := AvailableReasoningLevels(m); !slices.Equal(got, []string{"", "low", "high", "max"}) {
	103	+	t.Fatalf("%s: levels = %q, want off/low/high/max", model, got)
	104	+	}
	105	+	for _, tc := range []struct{ requested, want string }{
	106	+	{"low", "low"},
	107	+	{"medium", "high"},
	108	+	{"high", "high"},
	109	+	{"xhigh", "high"},
	110	+	{"max", "max"},
	111	+	} {
	112	+	wire, err := client.buildRequest(Request{Model: model, Reasoning: tc.requested})
	113	+	if err != nil {
	114	+	t.Fatal(err)
	115	+	}
	116	+	if wire.ReasoningEffort != tc.want {
	117	+	t.Errorf("%s %s: reasoning_effort = %q, want %q", model, tc.requested, wire.ReasoningEffort, tc.want)
	118	+	}
	119	+	}
	120	+	}
	121	+	}
	122	+
92	123		func TestOpenAIRequestUsesReasoningLevelMap(t *testing.T) {
93	124		SetLiveModels([]Model{{
94	125		Provider: "custom",


@@ -13,6 +13,7 @@ func TestAvailableReasoningLevels(t *testing.T) {
13	}{	13		}{
14	{name: "unsupported", model: Model{}, want: []string{""}},	14		{name: "unsupported", model: Model{}, want: []string{""}},
15	{name: "generic openai", model: Model{Reasoning: true}, want: []string{"", "low", "medium", "high"}},	15		{name: "generic openai", model: Model{Reasoning: true}, want: []string{"", "low", "medium", "high"}},
		16	+	{name: "direct deepseek", model: Model{Provider: "deepseek", Reasoning: true}, want: []string{"", "low", "high", "max"}},
16	{name: "responses", model: Model{Reasoning: true, API: APIResponses, ID: "gpt-5.5"}, want: []string{"", "low", "medium", "high", "xhigh"}},	17		{name: "responses", model: Model{Reasoning: true, API: APIResponses, ID: "gpt-5.5"}, want: []string{"", "low", "medium", "high", "xhigh"}},
17	{name: "responses native max", model: Model{Reasoning: true, API: APIResponses, ID: "gpt-5.6-sol"}, want: []string{"", "low", "medium", "high", "xhigh", "max"}},	18		{name: "responses native max", model: Model{Reasoning: true, API: APIResponses, ID: "gpt-5.6-sol"}, want: []string{"", "low", "medium", "high", "xhigh", "max"}},
18	{name: "astra native max", model: Model{Reasoning: true, API: APIResponses, ID: "gpt-6-astra"}, want: []string{"", "low", "medium", "high", "xhigh", "max"}},	19		{name: "astra native max", model: Model{Reasoning: true, API: APIResponses, ID: "gpt-6-astra"}, want: []string{"", "low", "medium", "high", "xhigh", "max"}},

@@ -63,6 +64,8 @@ func TestClampReasoningForModel(t *testing.T) {
63	}{	64		}{
64	{name: "unsupported", model: Model{}, level: "high", want: ""},	65		{name: "unsupported", model: Model{}, level: "high", want: ""},
65	{name: "generic max", model: Model{Reasoning: true}, level: "max", want: "high"},	66		{name: "generic max", model: Model{Reasoning: true}, level: "max", want: "high"},
		67	+	{name: "deepseek xhigh maps to high", model: Model{Provider: "deepseek", Reasoning: true}, level: "xhigh", want: "high"},
		68	+	{name: "deepseek max stays max", model: Model{Provider: "deepseek", Reasoning: true}, level: "max", want: "max"},
66	{name: "adaptive minimum", model: Model{Reasoning: true, AdaptiveThinking: true}, level: "minimum", want: "low"},	69		{name: "adaptive minimum", model: Model{Reasoning: true, AdaptiveThinking: true}, level: "minimum", want: "low"},
67	{name: "gemini pro medium", model: Model{Provider: "google", ID: "gemini-3-pro", Reasoning: true}, level: "medium", want: "high"},	70		{name: "gemini pro medium", model: Model{Provider: "google", ID: "gemini-3-pro", Reasoning: true}, level: "medium", want: "high"},
68	{name: "native max", model: Model{Reasoning: true, API: APIResponses, ID: "gpt-5.6-sol"}, level: "max", want: "max"},	71		{name: "native max", model: Model{Reasoning: true, API: APIResponses, ID: "gpt-5.6-sol"}, level: "max", want: "max"},

@@ -89,6 +92,34 @@ func TestClampReasoningForModel(t *testing.T) {
89	}	92		}
90	}	93		}
91		94
		95	+	func TestDeepSeekReasoningEffort(t *testing.T) {
		96	+	client := NewDeepSeek("test", "").(*openaiClient)
		97	+	for _, model := range []string{"deepseek-v4-pro", "deepseek-v4-flash"} {
		98	+	m, err := FindModel("deepseek", model)
		99	+	if err != nil {
		100	+	t.Fatal(err)
		101	+	}
		102	+	if got := AvailableReasoningLevels(m); !slices.Equal(got, []string{"", "low", "high", "max"}) {
		103	+	t.Fatalf("%s: levels = %q, want off/low/high/max", model, got)
		104	+	}
		105	+	for _, tc := range []struct{ requested, want string }{
		106	+	{"low", "low"},
		107	+	{"medium", "high"},
		108	+	{"high", "high"},
		109	+	{"xhigh", "high"},
		110	+	{"max", "max"},
		111	+	} {
		112	+	wire, err := client.buildRequest(Request{Model: model, Reasoning: tc.requested})
		113	+	if err != nil {
		114	+	t.Fatal(err)
		115	+	}
		116	+	if wire.ReasoningEffort != tc.want {
		117	+	t.Errorf("%s %s: reasoning_effort = %q, want %q", model, tc.requested, wire.ReasoningEffort, tc.want)
		118	+	}
		119	+	}
		120	+	}
		121	+	}
		122	+
92	func TestOpenAIRequestUsesReasoningLevelMap(t *testing.T) {	123		func TestOpenAIRequestUsesReasoningLevelMap(t *testing.T) {
93	SetLiveModels([]Model{{	124		SetLiveModels([]Model{{
94	Provider: "custom",	125		Provider: "custom",