gracefully fall back when model chat template doesn't support tools

Models like gemma3 use Jinja chat templates that only support alternating
user/assistant roles and cannot handle the tool/assistant(tool_calls) message
roles introduced by the agentic web-search loop.

When a 500 response containing "Jinja" is detected, the client now resets the
message history to the original user prompt, clears the tool schemas, and
retries — allowing the model to answer from training data rather than erroring
out. This covers two failure modes:

  1. The model rejects the request immediately because the tools parameter
     causes its template to inject an incompatible system message.
  2. The model successfully emits a tool_call but then fails on the follow-up
     request because it can't process the resulting "tool" role message.

Adds three unit tests covering the fallback paths and the non-Jinja error case.

Signed-off-by: Eric Curtin <eric.curtin@docker.com>

Author: ericcurtin

cmd/cli/commands/run.go

@@ -16,6 +16,7 @@ import (
 	"github.com/docker/model-runner/cmd/cli/commands/completion"
 	"github.com/docker/model-runner/cmd/cli/desktop"
 	"github.com/docker/model-runner/cmd/cli/readline"
+	"github.com/docker/model-runner/cmd/cli/tools"
 	"github.com/docker/model-runner/pkg/inference"
 	"github.com/docker/model-runner/pkg/inference/scheduling"
 	"github.com/fatih/color"
@@ -24,6 +25,15 @@ import (
 	"golang.org/x/term"
 )
 
+// defaultTools returns the tools enabled by default for interactive sessions.
+// Web search can be disabled by setting DOCKER_MODEL_NO_WEBSEARCH=1.
+func defaultTools() []desktop.ClientTool {
+	if os.Getenv("DOCKER_MODEL_NO_WEBSEARCH") != "" {
+		return nil
+	}
+	return []desktop.ClientTool{&tools.WebSearchTool{}}
+}
+
 // readMultilineInput reads input from stdin, supporting both single-line and multiline input.
 // For multiline input, it detects triple-quoted strings and shows continuation prompts.
 func readMultilineInput(cmd *cobra.Command, scanner *bufio.Scanner) (string, error) {
@@ -632,11 +642,13 @@ func chatWithMarkdownContext(ctx context.Context, cmd *cobra.Command, client *de
 	// This reflects exactly what the model receives.
 	processedUserMessage = buildUserMessage(prompt, imageURLs)
 
+	activeTools := defaultTools()
+
 	if !useMarkdown {
 		// Simple case: just stream as plain text
 		assistantResponse, err = client.ChatWithMessagesContext(ctx, model, conversationHistory, prompt, imageURLs, func(content string) {
 			cmd.Print(content)
-		}, false)
+		}, false, activeTools...)
 		return assistantResponse, processedUserMessage, err
 	}
 
@@ -655,7 +667,7 @@ func chatWithMarkdownContext(ctx context.Context, cmd *cobra.Command, client *de
 		} else if rendered != "" {
 			cmd.Print(rendered)
 		}
-	}, true)
+	}, true, activeTools...)
 	if err != nil {
 		return assistantResponse, processedUserMessage, err
 	}

cmd/cli/desktop/api.go

@@ -1,8 +1,37 @@
 package desktop
 
+// Tool represents an OpenAI function tool definition.
+type Tool struct {
+	Type     string       `json:"type"`
+	Function ToolFunction `json:"function"`
+}
+
+// ToolFunction holds the schema for a tool.
+type ToolFunction struct {
+	Name        string `json:"name"`
+	Description string `json:"description,omitempty"`
+	Parameters  any    `json:"parameters,omitempty"`
+}
+
+// ToolCall represents a tool call in a message or streaming delta.
+type ToolCall struct {
+	ID       string           `json:"id,omitempty"`
+	Type     string           `json:"type,omitempty"`
+	Index    int              `json:"index"`
+	Function ToolCallFunction `json:"function"`
+}
+
+// ToolCallFunction holds the name and accumulated arguments for a tool call.
+type ToolCallFunction struct {
+	Name      string `json:"name,omitempty"`
+	Arguments string `json:"arguments,omitempty"`
+}
+
 type OpenAIChatMessage struct {
-	Role    string      `json:"role"`
-	Content interface{} `json:"content"` // Can be string or []ContentPart for multimodal
+	Role       string     `json:"role"`
+	Content    any        `json:"content,omitempty"` // Can be string or []ContentPart for multimodal
+	ToolCalls  []ToolCall `json:"tool_calls,omitempty"`
+	ToolCallID string     `json:"tool_call_id,omitempty"`
 }
 
 // ContentPart represents a part of multimodal content (text or image)
@@ -21,6 +50,7 @@ type OpenAIChatRequest struct {
 	Model    string              `json:"model"`
 	Messages []OpenAIChatMessage `json:"messages"`
 	Stream   bool                `json:"stream"`
+	Tools    []Tool              `json:"tools,omitempty"`
 }
 
 type OpenAIChatResponse struct {
@@ -30,13 +60,15 @@ type OpenAIChatResponse struct {
 	Model   string `json:"model"`
 	Choices []struct {
 		Delta struct {
-			Content          string `json:"content"`
-			Role             string `json:"role,omitempty"`
-			ReasoningContent string `json:"reasoning_content,omitempty"`
+			Content          string     `json:"content"`
+			Role             string     `json:"role,omitempty"`
+			ReasoningContent string     `json:"reasoning_content,omitempty"`
+			ToolCalls        []ToolCall `json:"tool_calls,omitempty"`
 		} `json:"delta"`
 		Message struct {
-			Content string `json:"content"`
-			Role    string `json:"role,omitempty"`
+			Content   string     `json:"content"`
+			Role      string     `json:"role,omitempty"`
+			ToolCalls []ToolCall `json:"tool_calls,omitempty"`
 		} `json:"message"`
 		Index        int    `json:"index"`
 		FinishReason string `json:"finish_reason"`