model: align gemma4 chat template handling (#17182)

Incorporate the upstream Gemma4 chat template refinements for tool-calling stability, turn closure, and multi-turn reasoning. This updates the native renderer and checked-in HF template fixtures to keep adjacent assistant/tool continuations in the same model turn, add the post-tool thought-channel cue when thinking is enabled, and match Google's default of not replaying historical thinking before a later user turn.

Also preserve null tool arguments through Gemma4 rendering/parsing and extend the Jinja2 parity coverage for these upstream behaviors.

model/parsers/gemma4.go

@@ -74,6 +74,12 @@ func (p *Gemma4Parser) Init(tools []api.Tool, lastMessage *api.Message, thinkVal
 		return tools
 	}
 
+	if lastMessage != nil && lastMessage.Role == "tool" {
+		p.state = Gemma4CollectingThinking
+		p.needsChannelNameStrip = false
+		return tools
+	}
+
 	if prefill && lastMessage.Content != "" {
 		p.state = Gemma4CollectingContent
 		return tools

model/parsers/gemma4_test.go

@@ -102,6 +102,23 @@ func TestGemma4Parser(t *testing.T) {
 				},
 			},
 		},
+		{
+			name:  "tool_call_with_null_arg",
+			input: `<|tool_call>call:set_optional{value:null,nested:{value:null}}<tool_call|>`,
+			expectedToolCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name: "set_optional",
+						Arguments: testArgs(map[string]any{
+							"value": nil,
+							"nested": map[string]any{
+								"value": nil,
+							},
+						}),
+					},
+				},
+			},
+		},
 		{
 			name:  "tool_call_with_nested_object",
 			input: `<|tool_call>call:process{config:{enabled:true,name:<|"|>test<|"|>}}<tool_call|>`,
@@ -508,6 +525,26 @@ func TestGemma4Parser_Streaming(t *testing.T) {
 	}
 }
 
+func TestGemma4Parser_ToolResponseContinuationStartsInThinking(t *testing.T) {
+	parser := &Gemma4Parser{hasThinkingSupport: true}
+	parser.Init(nil, &api.Message{Role: "tool", Content: "go.mod\ngo.sum\n"}, &api.ThinkValue{Value: true})
+
+	content, thinking, toolCalls, err := parser.Add("The user asked for files.<channel|>go.mod and go.sum", true)
+	if err != nil {
+		t.Fatalf("Add() error = %v", err)
+	}
+
+	if diff := cmp.Diff("go.mod and go.sum", content); diff != "" {
+		t.Errorf("content mismatch (-want +got):\n%s", diff)
+	}
+	if diff := cmp.Diff("The user asked for files.", thinking); diff != "" {
+		t.Errorf("thinking mismatch (-want +got):\n%s", diff)
+	}
+	if len(toolCalls) != 0 {
+		t.Fatalf("expected no tool calls, got %+v", toolCalls)
+	}
+}
+
 func TestGemma4Parser_StreamingToolCall(t *testing.T) {
 	parser := &Gemma4Parser{hasThinkingSupport: false}
 	parser.Init(nil, nil, nil)

model/renderers/gemma4.go

@@ -28,7 +28,7 @@ func (r *Gemma4Renderer) Render(messages []api.Message, tools []api.Tool, thinkV
 	var sb strings.Builder
 	imageOffset := 0
 
-	// BOS token — Gemma 4 models have add_bos_token=false in their tokenizer
+	// BOS token: Gemma 4 models have add_bos_token=false in their tokenizer
 	// config, so the tokenizer does not auto-prepend BOS. We must emit it
 	// explicitly in the rendered prompt, matching the HF chat template.
 	sb.WriteString("<bos>")
@@ -67,6 +67,7 @@ func (r *Gemma4Renderer) Render(messages []api.Message, tools []api.Tool, thinkV
 	}
 
 	var prevMessageType string
+	var prevNonToolRole string
 
 	// Consecutive tool messages are folded into the preceding assistant turn,
 	// and adjacent assistant messages continue in the same model turn.
@@ -81,12 +82,12 @@ func (r *Gemma4Renderer) Render(messages []api.Message, tools []api.Tool, thinkV
 			role = "model"
 		}
 
-		continueSameModelTurn := role == "model" && r.previousNonToolRole(loopMessages, i) == "assistant"
+		continueSameModelTurn := role == "model" && prevNonToolRole == "assistant"
 		if !continueSameModelTurn {
 			sb.WriteString("<|turn>" + role + "\n")
 		}
 
-		if message.Role == "assistant" && message.Thinking != "" && i > lastUserIdx && len(message.ToolCalls) > 0 {
+		if message.Role == "assistant" && message.Thinking != "" && i > lastUserIdx {
 			sb.WriteString("<|channel>thought\n")
 			sb.WriteString(message.Thinking)
 			sb.WriteString("\n<channel|>")
@@ -123,11 +124,15 @@ func (r *Gemma4Renderer) Render(messages []api.Message, tools []api.Tool, thinkV
 			messageHadContent = r.messageHasContent(message)
 		}
 
+		nextNonToolRole := r.nextNonToolRole(loopMessages, i)
+		continuesIntoNext := role == "model" && nextNonToolRole == "assistant" && (len(message.ToolCalls) == 0 || toolResponsesEmitted)
 		if prevMessageType == "tool_call" && !toolResponsesEmitted {
 			sb.WriteString("<|tool_response>")
-		} else if !(toolResponsesEmitted && !messageHadContent) {
+		} else if !continuesIntoNext && !(toolResponsesEmitted && !messageHadContent && nextNonToolRole == "") {
 			sb.WriteString("<turn|>\n")
 		}
+
+		prevNonToolRole = message.Role
 	}
 
 	// Generation prompt.
@@ -136,6 +141,8 @@ func (r *Gemma4Renderer) Render(messages []api.Message, tools []api.Tool, thinkV
 		if r.emptyBlockOnNothink && !hasThink {
 			sb.WriteString("<|channel>thought\n<channel|>")
 		}
+	} else if prevMessageType == "tool_response" && hasThink {
+		sb.WriteString("<|channel>thought\n")
 	}
 
 	return sb.String(), nil
@@ -181,8 +188,8 @@ func (r *Gemma4Renderer) renderToolResponseContent(msg api.Message, imageOffset
 	return sb.String()
 }
 
-func (r *Gemma4Renderer) previousNonToolRole(messages []api.Message, idx int) string {
+func (r *Gemma4Renderer) nextNonToolRole(messages []api.Message, idx int) string {
-	for i := idx - 1; i >= 0; i-- {
+	for i := idx + 1; i < len(messages); i++ {
 		if messages[i].Role != "tool" {
 			return messages[i].Role
 		}
@@ -699,6 +706,8 @@ func (r *Gemma4Renderer) formatToolResponseBlock(toolName, response string) stri
 
 func (r *Gemma4Renderer) formatArgValue(value any) string {
 	switch v := value.(type) {
+	case nil:
+		return "null"
 	case string:
 		return g4Q + v + g4Q
 	case bool:
@@ -747,6 +756,8 @@ func (r *Gemma4Renderer) formatMapValue(m map[string]any) string {
 
 func (r *Gemma4Renderer) formatSchemaValue(value any) string {
 	switch v := value.(type) {
+	case nil:
+		return "null"
 	case string:
 		return g4Q + v + g4Q
 	case bool:

model/renderers/gemma4_reference_test.go

@@ -906,7 +906,7 @@ func TestGemma4RendererMatchesReference(t *testing.T) {
 		},
 		// === Additional edge cases ported from original tests ===
 		{
-			// Assistant content with tool call — template emits tool_calls before content
+			// Assistant content with tool call: template emits tool_calls before content
 			name: "assistant_content_with_tool_call",
 			messages: []api.Message{
 				{Role: "user", Content: "Weather?"},
@@ -988,9 +988,23 @@ func TestGemma4RendererMatchesReference(t *testing.T) {
 				"<|turn>user\nWeather?<turn|>\n" +
 				"<|turn>model\n<|tool_call>call:get_weather{city:" + q + "Tokyo" + q + "}<tool_call|>" +
 				"<|tool_response>response:get_weather{value:" + q + `{"temperature": 15, "weather": "sunny"}` + q + "}<tool_response|>" +
+				"<turn|>\n" +
 				"<|turn>user\nThanks!<turn|>\n" +
 				"<|turn>model\n",
 		},
+		{
+			name: "adjacent_assistants_continue_same_model_turn",
+			messages: []api.Message{
+				{Role: "user", Content: "Start"},
+				{Role: "assistant", Content: "One."},
+				{Role: "assistant", Content: "Two."},
+				{Role: "user", Content: "More"},
+			},
+			expected: "<bos><|turn>user\nStart<turn|>\n" +
+				"<|turn>model\nOne.Two.<turn|>\n" +
+				"<|turn>user\nMore<turn|>\n" +
+				"<|turn>model\n",
+		},
 		// === Ordering and whitespace edge cases ===
 		{
 			// Tool call arguments are sorted alphabetically
@@ -1127,7 +1141,7 @@ Hi<turn|>
 				"<|turn>model\n",
 		},
 		{
-			// Property with no description — just type
+			// Property with no description: just type
 			name:     "property_no_description",
 			messages: []api.Message{{Role: "user", Content: "Count"}},
 			tools:    countTool(),
@@ -1407,7 +1421,7 @@ Hi<turn|>
 
 		// === Round 5: coding agent patterns ===
 		{
-			// Chained tool calls — assistant calls tool, gets result, calls another
+			// Chained tool calls: assistant calls tool, gets result, calls another
 			// tool, gets result, then the model responds. No user messages in between.
 			name: "chained_tool_calls",
 			messages: []api.Message{
@@ -1486,6 +1500,58 @@ Hi<turn|>
 				"<|turn>model\n<|tool_call>call:bash{command:" + q + q + "}<tool_call|>" +
 				"<|tool_response>response:bash{value:" + q + "error" + q + "}<tool_response|>",
 		},
+		{
+			name: "null_tool_arguments",
+			messages: []api.Message{
+				{Role: "user", Content: "Set optional"},
+				{Role: "assistant", ToolCalls: []api.ToolCall{{
+					Function: api.ToolCallFunction{Name: "set_optional", Arguments: testArgs(map[string]any{
+						"nested": map[string]any{"value": nil},
+						"value":  nil,
+					})},
+				}}},
+			},
+			expected: "<bos><|turn>user\nSet optional<turn|>\n" +
+				"<|turn>model\n<|tool_call>call:set_optional{nested:{value:null},value:null}<tool_call|><|tool_response>",
+		},
+		{
+			name: "historical_tool_call_thinking_omitted_before_later_user",
+			messages: []api.Message{
+				{Role: "user", Content: "List files"},
+				{Role: "assistant", Thinking: "Need the file list before answering.", ToolCalls: []api.ToolCall{{
+					Function: api.ToolCallFunction{Name: "bash", Arguments: testArgs(map[string]any{"command": "ls"})},
+				}}},
+				{Role: "tool", ToolName: "bash", Content: "file1.txt"},
+				{Role: "user", Content: "Now answer"},
+			},
+			tools: bashSmallTool(),
+			think: thinkTrue(),
+			expected: "<bos><|turn>system\n<|think|>\n" + bashSmallDeclRef + "<turn|>\n" +
+				"<|turn>user\nList files<turn|>\n" +
+				"<|turn>model\n" +
+				"<|tool_call>call:bash{command:" + q + "ls" + q + "}<tool_call|>" +
+				"<|tool_response>response:bash{value:" + q + "file1.txt" + q + "}<tool_response|>" +
+				"<turn|>\n" +
+				"<|turn>user\nNow answer<turn|>\n" +
+				"<|turn>model\n",
+		},
+		{
+			name: "terminal_tool_response_adds_thinking_cue_when_thinking_enabled",
+			messages: []api.Message{
+				{Role: "user", Content: "List files"},
+				{Role: "assistant", ToolCalls: []api.ToolCall{{
+					Function: api.ToolCallFunction{Name: "bash", Arguments: testArgs(map[string]any{"command": "ls"})},
+				}}},
+				{Role: "tool", ToolName: "bash", Content: "file1.txt"},
+			},
+			tools: bashSmallTool(),
+			think: thinkTrue(),
+			expected: "<bos><|turn>system\n<|think|>\n" + bashSmallDeclRef + "<turn|>\n" +
+				"<|turn>user\nList files<turn|>\n" +
+				"<|turn>model\n<|tool_call>call:bash{command:" + q + "ls" + q + "}<tool_call|>" +
+				"<|tool_response>response:bash{value:" + q + "file1.txt" + q + "}<tool_response|>" +
+				"<|channel>thought\n",
+		},
 	}
 
 	verifyJinja2 := os.Getenv("VERIFY_JINJA2") != ""
@@ -1561,6 +1627,50 @@ func TestGemma4LargeRendererOmitsEmptyThoughtBlockWhenThinkingEnabled(t *testing
 	assert.NotContains(t, got, "<|channel>thought\n<channel|>")
 }
 
+func TestGemma4RendererCanonicalToolContinuations(t *testing.T) {
+	q := `<|"|>`
+	messages := []api.Message{
+		{Role: "user", Content: "Use the tool"},
+		{Role: "assistant", Content: "First."},
+		{Role: "assistant", Content: "Checking.", ToolCalls: []api.ToolCall{{
+			Function: api.ToolCallFunction{
+				Name:      "bash",
+				Arguments: testArgs(map[string]any{"command": nil}),
+			},
+		}}},
+		{Role: "tool", ToolName: "bash", Content: "done"},
+		{Role: "assistant", Content: "Done."},
+		{Role: "user", Content: "Next"},
+	}
+
+	base := "<bos><|turn>system\n" + bashSmallDeclRef + "<turn|>\n" +
+		"<|turn>user\nUse the tool<turn|>\n" +
+		"<|turn>model\nFirst." +
+		"<|tool_call>call:bash{command:null}<tool_call|>" +
+		"<|tool_response>response:bash{value:" + q + "done" + q + "}<tool_response|>" +
+		"Checking.Done.<turn|>\n" +
+		"<|turn>user\nNext<turn|>\n<|turn>model\n"
+
+	tests := []struct {
+		name     string
+		renderer *Gemma4Renderer
+		want     string
+	}{
+		{name: "small", renderer: &Gemma4Renderer{}, want: base},
+		{name: "31b", renderer: &Gemma4Renderer{emptyBlockOnNothink: true}, want: base + "<|channel>thought\n<channel|>"},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got, err := tt.renderer.Render(messages, bashSmallTool(), nil)
+			assert.NoError(t, err)
+			assert.Equal(t, tt.want, got)
+			assert.NotContains(t, got, "First.<turn|>\n<|tool_call>")
+			assert.NotContains(t, got, "Checking.<turn|>\nDone.")
+		})
+	}
+}
+
 func TestGemma4RendererMatchesJinja2ExpandedParity(t *testing.T) {
 	if os.Getenv("VERIFY_JINJA2") == "" {
 		t.Skip("set VERIFY_JINJA2=1 to run expanded Jinja2 parity checks")

model/renderers/testdata/gemma4_31b_chat_template.jinja2

@@ -1,3 +1,9 @@
+{#
+  Template: Google Gemma 4 Canonical Chat Template
+  Author: Google Gemma Engineering Team
+  Published: 2026-07-09
+  Context: Fixed tool-calling loops, turn closures, and thinking content-ordering.
+#}
 {%- macro format_parameters(properties, required, filter_keys=false) -%}
     {%- set standard_keys = ['description', 'type', 'properties', 'required', 'nullable'] -%}
     {%- set ns = namespace(found_first=false) -%}
@@ -116,7 +122,9 @@
     }
 {%- endmacro -%}
 {%- macro format_argument(argument, escape_keys=True) -%}
-    {%- if argument is string -%}
+    {%- if argument is none -%}
+        {{- 'null' -}}
+    {%- elif argument is string -%}
         {{- '<|"|>' + argument + '<|"|>' -}}
     {%- elif argument is boolean -%}
         {{- 'true' if argument else 'false' -}}
@@ -172,18 +180,21 @@
     {{- '<tool_response|>' -}}
 {%- endmacro -%}
 
-{%- set ns = namespace(prev_message_type=None) -%}
+{#- ===== SETUP ===== -#}
+{%- set ns = namespace(prev_message_type=None, prev_non_tool_role=None) -%}
 {%- set loop_messages = messages -%}
+{%- set enable_thinking = enable_thinking | default(false) -%}
+{%- set preserve_thinking = preserve_thinking | default(false) -%}
 {{- bos_token -}}
 {#- Handle System/Tool Definitions Block -#}
-{%- if (enable_thinking is defined and enable_thinking) or tools or messages[0]['role'] in ['system', 'developer'] -%}
+{%- if enable_thinking or tools or (messages and messages[0]['role'] in ['system', 'developer']) -%}
     {{- '<|turn>system\n' -}}
     {#- Inject Thinking token at the very top of the FIRST system turn -#}
-    {%- if enable_thinking is defined and enable_thinking -%}
+    {%- if enable_thinking -%}
         {{- '<|think|>\n' -}}
         {%- set ns.prev_message_type = 'think' -%}
     {%- endif -%}
-    {%- if messages[0]['role'] in ['system', 'developer'] -%}
+    {%- if messages and messages[0]['role'] in ['system', 'developer'] -%}
         {%- if messages[0]['content'] is string -%}
             {{- messages[0]['content'] | trim -}}
         {%- elif messages[0]['content'] is sequence -%}
@@ -217,31 +228,22 @@
     {%- if message['role'] != 'tool' -%}
     {%- set ns.prev_message_type = None -%}
     {%- set role = 'model' if message['role'] == 'assistant' else message['role'] -%}
-    {#- Detect continuation: suppress duplicate <|turn>model when previous non-tool message was also assistant -#}
+    {#- Detect continuation using tracked state - O(1) instead of O(n) backward scan -#}
-    {%- set prev_nt = namespace(role=None, found=false) -%}
+    {%- set continue_same_model_turn = (role == 'model' and ns.prev_non_tool_role == 'assistant') -%}
-    {%- if loop.index0 > 0 -%}
-        {%- for j in range(loop.index0 - 1, -1, -1) -%}
-            {%- if not prev_nt.found -%}
-                {%- if loop_messages[j]['role'] != 'tool' -%}
-                    {%- set prev_nt.role = loop_messages[j]['role'] -%}
-                    {%- set prev_nt.found = true -%}
-                {%- endif -%}
-            {%- endif -%}
-        {%- endfor -%}
-    {%- endif -%}
-    {%- set continue_same_model_turn = (role == 'model' and prev_nt.role == 'assistant') -%}
     {%- if not continue_same_model_turn -%}
         {{- '<|turn>' + role + '\n' }}
+
     {%- endif -%}
 
     {#- Render reasoning/reasoning_content as thinking channel -#}
     {%- set thinking_text = message.get('reasoning') or message.get('reasoning_content') -%}
-    {%- if thinking_text and loop.index0 > ns_turn.last_user_idx and message.get('tool_calls') -%}
+    {%- set thinking_gate = (loop.index0 > ns_turn.last_user_idx) or (preserve_thinking and message.get('tool_calls')) -%}
+    {%- if thinking_text and thinking_gate -%}
         {{- '<|channel>thought\n' + thinking_text + '\n<channel|>' -}}
     {%- endif -%}
 
-            {%- if message['tool_calls'] -%}
+            {%- if message.get('tool_calls') -%}
-                {%- for tool_call in message['tool_calls'] -%}
+                {%- for tool_call in message.get('tool_calls') -%}
                     {%- set function = tool_call['function'] -%}
                     {{- '<|tool_call>call:' + function['name'] + '{' -}}
                     {%- if function['arguments'] is mapping -%}
@@ -251,8 +253,13 @@
                             {%- set ns_args.found_first = true -%}
                             {{- key -}}:{{- format_argument(value, escape_keys=False) -}}
                         {%- endfor -%}
-                    {%- elif function['arguments'] is string -%}
+                    {%- elif function['arguments'] is none -%}
-                        {{- function['arguments'] -}}
+                    {%- else -%}
+                        {{- raise_exception(
+                            "chat_template: tool_calls[].function.arguments must be a "
+                            "JSON object (mapping), not a string. Deserialize arguments "
+                            "before passing to the template."
+                        ) -}}
                     {%- endif -%}
                     {{- '}<tool_call|>' -}}
                 {%- endfor -%}
@@ -262,8 +269,8 @@
             {%- set ns_tr_out = namespace(flag=false) -%}
             {%- if message.get('tool_responses') -%}
                 {#- Legacy: tool_responses embedded on the assistant message (Google/Gemma native) -#}
-                {%- for tool_response in message['tool_responses'] -%}
+                {%- for tool_response in message.get('tool_responses') -%}
-                    {{- format_tool_response_block(tool_response['name'] | default('unknown'), tool_response['response']) -}}
+                    {{- format_tool_response_block(tool_response['name'] | default('unknown', true), tool_response['response']) -}}
                     {%- set ns_tr_out.flag = true -%}
                     {%- set ns.prev_message_type = 'tool_response' -%}
                 {%- endfor -%}
@@ -277,8 +284,8 @@
                     {%- else -%}
                         {%- set follow = loop_messages[k] -%}
                         {#- Resolve tool_call_id to function name -#}
-                        {%- set ns_tname = namespace(name=follow.get('name') | default('unknown')) -%}
+                        {%- set ns_tname = namespace(name=follow.get('name') or 'unknown') -%}
-                        {%- for tc in message['tool_calls'] -%}
+                        {%- for tc in message.get('tool_calls') -%}
                             {%- if tc.get('id') == follow.get('tool_call_id') -%}
                                 {%- set ns_tname.name = tc['function']['name'] -%}
                             {%- endif -%}
@@ -295,6 +302,15 @@
                                 {%- endif -%}
                             {%- endfor -%}
                             {{- format_tool_response_block(ns_tname.name, ns_txt.s) -}}
+                            {%- for part in tool_body -%}
+                                {%- if part.get('type') in ['image', 'image_url'] -%}
+                                    {{- '<|image|>' -}}
+                                {%- elif part.get('type') in ['audio', 'input_audio'] -%}
+                                    {{- '<|audio|>' -}}
+                                {%- elif part.get('type') == 'video' -%}
+                                    {{- '<|video|>' -}}
+                                {%- endif -%}
+                            {%- endfor -%}
                         {%- else -%}
                             {{- format_tool_response_block(ns_tname.name, tool_body) -}}
                         {%- endif -%}
@@ -305,29 +321,26 @@
             {%- endif -%}
 
             {%- set captured_content -%}
-            {%- if message['content'] is string -%}
+            {%- if message.get('content') is string -%}
                 {%- if role == 'model' -%}
                     {{- strip_thinking(message['content']) -}}
                 {%- else -%}
                     {{- message['content'] | trim -}}
                 {%- endif -%}
-            {%- elif message['content'] is sequence -%}
+            {%- elif message.get('content') is sequence -%}
                 {%- for item in message['content'] -%}
-                    {%- if item['type'] == 'text' -%}
+                    {%- if item.get('type') == 'text' -%}
                         {%- if role == 'model' -%}
                             {{- strip_thinking(item['text']) -}}
                         {%- else -%}
                             {{- item['text'] | trim -}}
                         {%- endif -%}
-                    {%- elif item['type'] == 'image' -%}
+                    {%- elif item.get('type') in ['image', 'image_url'] -%}
                         {{- '<|image|>' -}}
-                        {%- set ns.prev_message_type = 'image' -%}
+                    {%- elif item.get('type') in ['audio', 'input_audio'] -%}
-                    {%- elif item['type'] == 'audio' -%}
                         {{- '<|audio|>' -}}
-                        {%- set ns.prev_message_type = 'audio' -%}
+                    {%- elif item.get('type') == 'video' -%}
-                    {%- elif item['type'] == 'video' -%}
                         {{- '<|video|>' -}}
-                        {%- set ns.prev_message_type = 'video' -%}
                     {%- endif -%}
                 {%- endfor -%}
             {%- endif -%}
@@ -336,19 +349,42 @@
             {{- captured_content -}}
             {%- set has_content = captured_content | trim | length > 0 -%}
 
+        {#- Forward-scan: find next non-tool message role for continuation detection -#}
+        {%- set next_nt = namespace(role=None, found=false) -%}
+        {%- for j in range(loop.index0 + 1, loop_messages | length) -%}
+            {%- if not next_nt.found -%}
+                {%- if loop_messages[j]['role'] != 'tool' -%}
+                    {%- set next_nt.role = loop_messages[j]['role'] -%}
+                    {%- set next_nt.found = true -%}
+                {%- endif -%}
+            {%- endif -%}
+        {%- endfor -%}
+
+        {%- set continues_into_next = (
+            role == 'model'
+            and next_nt.role == 'assistant'
+            and (not message.get('tool_calls') or ns_tr_out.flag)
+        ) -%}
+
         {%- if ns.prev_message_type == 'tool_call' and not ns_tr_out.flag -%}
             {{- '<|tool_response>' -}}
-        {%- elif not (ns_tr_out.flag and not has_content) -%}
+        {%- elif continues_into_next -%}
+        {%- elif not (ns_tr_out.flag and not has_content and not next_nt.found) -%}
             {{- '<turn|>\n' -}}
         {%- endif -%}
+
+    {#- Track previous non-tool role for next iteration (avoids O(n) backward scan) -#}
+    {%- set ns.prev_non_tool_role = message['role'] -%}
     {%- endif -%}
 {%- endfor -%}
 
 {%- if add_generation_prompt -%}
     {%- if ns.prev_message_type != 'tool_response' and ns.prev_message_type != 'tool_call' -%}
         {{- '<|turn>model\n' -}}
-        {%- if not enable_thinking | default(false) -%}
+        {%- if not enable_thinking -%}
             {{- '<|channel>thought\n<channel|>' -}}
         {%- endif -%}
+    {%- elif ns.prev_message_type == 'tool_response' and enable_thinking -%}
+        {{- '<|channel>thought\n' -}}
     {%- endif -%}
-{%- endif -%}
+{%- endif -%}

model/renderers/testdata/gemma4_e2b_chat_template.jinja2

@@ -1,3 +1,9 @@
+{#
+  Template: Google Gemma 4 Canonical Chat Template
+  Author: Google Gemma Engineering Team
+  Published: 2026-07-09
+  Context: Fixed tool-calling loops, turn closures, and thinking content-ordering.
+#}
 {%- macro format_parameters(properties, required, filter_keys=false) -%}
     {%- set standard_keys = ['description', 'type', 'properties', 'required', 'nullable'] -%}
     {%- set ns = namespace(found_first=false) -%}
@@ -116,7 +122,9 @@
     }
 {%- endmacro -%}
 {%- macro format_argument(argument, escape_keys=True) -%}
-    {%- if argument is string -%}
+    {%- if argument is none -%}
+        {{- 'null' -}}
+    {%- elif argument is string -%}
         {{- '<|"|>' + argument + '<|"|>' -}}
     {%- elif argument is boolean -%}
         {{- 'true' if argument else 'false' -}}
@@ -172,18 +180,21 @@
     {{- '<tool_response|>' -}}
 {%- endmacro -%}
 
-{%- set ns = namespace(prev_message_type=None) -%}
+{#- ===== SETUP ===== -#}
+{%- set ns = namespace(prev_message_type=None, prev_non_tool_role=None) -%}
 {%- set loop_messages = messages -%}
+{%- set enable_thinking = enable_thinking | default(false) -%}
+{%- set preserve_thinking = preserve_thinking | default(false) -%}
 {{- bos_token -}}
 {#- Handle System/Tool Definitions Block -#}
-{%- if (enable_thinking is defined and enable_thinking) or tools or messages[0]['role'] in ['system', 'developer'] -%}
+{%- if enable_thinking or tools or (messages and messages[0]['role'] in ['system', 'developer']) -%}
     {{- '<|turn>system\n' -}}
     {#- Inject Thinking token at the very top of the FIRST system turn -#}
-    {%- if enable_thinking is defined and enable_thinking -%}
+    {%- if enable_thinking -%}
         {{- '<|think|>\n' -}}
         {%- set ns.prev_message_type = 'think' -%}
     {%- endif -%}
-    {%- if messages[0]['role'] in ['system', 'developer'] -%}
+    {%- if messages and messages[0]['role'] in ['system', 'developer'] -%}
         {%- if messages[0]['content'] is string -%}
             {{- messages[0]['content'] | trim -}}
         {%- elif messages[0]['content'] is sequence -%}
@@ -217,31 +228,22 @@
     {%- if message['role'] != 'tool' -%}
     {%- set ns.prev_message_type = None -%}
     {%- set role = 'model' if message['role'] == 'assistant' else message['role'] -%}
-    {#- Detect continuation: suppress duplicate <|turn>model when previous non-tool message was also assistant -#}
+    {#- Detect continuation using tracked state - O(1) instead of O(n) backward scan -#}
-    {%- set prev_nt = namespace(role=None, found=false) -%}
+    {%- set continue_same_model_turn = (role == 'model' and ns.prev_non_tool_role == 'assistant') -%}
-    {%- if loop.index0 > 0 -%}
-        {%- for j in range(loop.index0 - 1, -1, -1) -%}
-            {%- if not prev_nt.found -%}
-                {%- if loop_messages[j]['role'] != 'tool' -%}
-                    {%- set prev_nt.role = loop_messages[j]['role'] -%}
-                    {%- set prev_nt.found = true -%}
-                {%- endif -%}
-            {%- endif -%}
-        {%- endfor -%}
-    {%- endif -%}
-    {%- set continue_same_model_turn = (role == 'model' and prev_nt.role == 'assistant') -%}
     {%- if not continue_same_model_turn -%}
         {{- '<|turn>' + role + '\n' }}
+
     {%- endif -%}
 
     {#- Render reasoning/reasoning_content as thinking channel -#}
     {%- set thinking_text = message.get('reasoning') or message.get('reasoning_content') -%}
-    {%- if thinking_text and loop.index0 > ns_turn.last_user_idx and message.get('tool_calls') -%}
+    {%- set thinking_gate = (loop.index0 > ns_turn.last_user_idx) or (preserve_thinking and message.get('tool_calls')) -%}
+    {%- if thinking_text and thinking_gate -%}
         {{- '<|channel>thought\n' + thinking_text + '\n<channel|>' -}}
     {%- endif -%}
 
-            {%- if message['tool_calls'] -%}
+            {%- if message.get('tool_calls') -%}
-                {%- for tool_call in message['tool_calls'] -%}
+                {%- for tool_call in message.get('tool_calls') -%}
                     {%- set function = tool_call['function'] -%}
                     {{- '<|tool_call>call:' + function['name'] + '{' -}}
                     {%- if function['arguments'] is mapping -%}
@@ -251,8 +253,13 @@
                             {%- set ns_args.found_first = true -%}
                             {{- key -}}:{{- format_argument(value, escape_keys=False) -}}
                         {%- endfor -%}
-                    {%- elif function['arguments'] is string -%}
+                    {%- elif function['arguments'] is none -%}
-                        {{- function['arguments'] -}}
+                    {%- else -%}
+                        {{- raise_exception(
+                            "chat_template: tool_calls[].function.arguments must be a "
+                            "JSON object (mapping), not a string. Deserialize arguments "
+                            "before passing to the template."
+                        ) -}}
                     {%- endif -%}
                     {{- '}<tool_call|>' -}}
                 {%- endfor -%}
@@ -262,8 +269,8 @@
             {%- set ns_tr_out = namespace(flag=false) -%}
             {%- if message.get('tool_responses') -%}
                 {#- Legacy: tool_responses embedded on the assistant message (Google/Gemma native) -#}
-                {%- for tool_response in message['tool_responses'] -%}
+                {%- for tool_response in message.get('tool_responses') -%}
-                    {{- format_tool_response_block(tool_response['name'] | default('unknown'), tool_response['response']) -}}
+                    {{- format_tool_response_block(tool_response['name'] | default('unknown', true), tool_response['response']) -}}
                     {%- set ns_tr_out.flag = true -%}
                     {%- set ns.prev_message_type = 'tool_response' -%}
                 {%- endfor -%}
@@ -277,8 +284,8 @@
                     {%- else -%}
                         {%- set follow = loop_messages[k] -%}
                         {#- Resolve tool_call_id to function name -#}
-                        {%- set ns_tname = namespace(name=follow.get('name') | default('unknown')) -%}
+                        {%- set ns_tname = namespace(name=follow.get('name') or 'unknown') -%}
-                        {%- for tc in message['tool_calls'] -%}
+                        {%- for tc in message.get('tool_calls') -%}
                             {%- if tc.get('id') == follow.get('tool_call_id') -%}
                                 {%- set ns_tname.name = tc['function']['name'] -%}
                             {%- endif -%}
@@ -295,6 +302,15 @@
                                 {%- endif -%}
                             {%- endfor -%}
                             {{- format_tool_response_block(ns_tname.name, ns_txt.s) -}}
+                            {%- for part in tool_body -%}
+                                {%- if part.get('type') in ['image', 'image_url'] -%}
+                                    {{- '<|image|>' -}}
+                                {%- elif part.get('type') in ['audio', 'input_audio'] -%}
+                                    {{- '<|audio|>' -}}
+                                {%- elif part.get('type') == 'video' -%}
+                                    {{- '<|video|>' -}}
+                                {%- endif -%}
+                            {%- endfor -%}
                         {%- else -%}
                             {{- format_tool_response_block(ns_tname.name, tool_body) -}}
                         {%- endif -%}
@@ -305,29 +321,26 @@
             {%- endif -%}
 
             {%- set captured_content -%}
-            {%- if message['content'] is string -%}
+            {%- if message.get('content') is string -%}
                 {%- if role == 'model' -%}
                     {{- strip_thinking(message['content']) -}}
                 {%- else -%}
                     {{- message['content'] | trim -}}
                 {%- endif -%}
-            {%- elif message['content'] is sequence -%}
+            {%- elif message.get('content') is sequence -%}
                 {%- for item in message['content'] -%}
-                    {%- if item['type'] == 'text' -%}
+                    {%- if item.get('type') == 'text' -%}
                         {%- if role == 'model' -%}
                             {{- strip_thinking(item['text']) -}}
                         {%- else -%}
                             {{- item['text'] | trim -}}
                         {%- endif -%}
-                    {%- elif item['type'] == 'image' -%}
+                    {%- elif item.get('type') in ['image', 'image_url'] -%}
                         {{- '<|image|>' -}}
-                        {%- set ns.prev_message_type = 'image' -%}
+                    {%- elif item.get('type') in ['audio', 'input_audio'] -%}
-                    {%- elif item['type'] == 'audio' -%}
                         {{- '<|audio|>' -}}
-                        {%- set ns.prev_message_type = 'audio' -%}
+                    {%- elif item.get('type') == 'video' -%}
-                    {%- elif item['type'] == 'video' -%}
                         {{- '<|video|>' -}}
-                        {%- set ns.prev_message_type = 'video' -%}
                     {%- endif -%}
                 {%- endfor -%}
             {%- endif -%}
@@ -336,16 +349,39 @@
             {{- captured_content -}}
             {%- set has_content = captured_content | trim | length > 0 -%}
 
+        {#- Forward-scan: find next non-tool message role for continuation detection -#}
+        {%- set next_nt = namespace(role=None, found=false) -%}
+        {%- for j in range(loop.index0 + 1, loop_messages | length) -%}
+            {%- if not next_nt.found -%}
+                {%- if loop_messages[j]['role'] != 'tool' -%}
+                    {%- set next_nt.role = loop_messages[j]['role'] -%}
+                    {%- set next_nt.found = true -%}
+                {%- endif -%}
+            {%- endif -%}
+        {%- endfor -%}
+
+        {%- set continues_into_next = (
+            role == 'model'
+            and next_nt.role == 'assistant'
+            and (not message.get('tool_calls') or ns_tr_out.flag)
+        ) -%}
+
         {%- if ns.prev_message_type == 'tool_call' and not ns_tr_out.flag -%}
             {{- '<|tool_response>' -}}
-        {%- elif not (ns_tr_out.flag and not has_content) -%}
+        {%- elif continues_into_next -%}
+        {%- elif not (ns_tr_out.flag and not has_content and not next_nt.found) -%}
             {{- '<turn|>\n' -}}
         {%- endif -%}
+
+    {#- Track previous non-tool role for next iteration (avoids O(n) backward scan) -#}
+    {%- set ns.prev_non_tool_role = message['role'] -%}
     {%- endif -%}
 {%- endfor -%}
 
 {%- if add_generation_prompt -%}
     {%- if ns.prev_message_type != 'tool_response' and ns.prev_message_type != 'tool_call' -%}
         {{- '<|turn>model\n' -}}
+    {%- elif ns.prev_message_type == 'tool_response' and enable_thinking -%}
+        {{- '<|channel>thought\n' -}}
     {%- endif -%}
-{%- endif -%}
+{%- endif -%}