fix(cli): honor enabled tools when serving (#737)

* fix(cli): honor enabled tools when serving

* fix(server): preserve tools for streaming agents

frontend/src/components/Chat/InputArea.tsx

@@ -5,6 +5,7 @@ import { useAppStore, generateId } from '../../lib/store';
 import { streamChat, streamResearch } from '../../lib/sse';
 import { fetchSavings, getBase } from '../../lib/api';
 import { listConnectors, getSyncStatus } from '../../lib/connectors-api';
+import { serializeToolCallArguments } from '../../lib/tool-call';
 import { MicButton } from './MicButton';
 import { useSpeech } from '../../hooks/useSpeech';
 import type {
@@ -389,7 +390,7 @@ export function InputArea() {
             const tc: ToolCallInfo = {
               id: generateId(),
               tool: data.tool,
-              arguments: data.arguments || '',
+              arguments: serializeToolCallArguments(data.arguments),
               status: 'running',
             };
             toolCalls.push(tc);
@@ -400,7 +401,7 @@ export function InputArea() {
             updateLastAssistant(convId, accumulatedContent, [...toolCalls]);
             useAppStore.getState().addLogEntry({
               timestamp: Date.now(), level: 'info', category: 'tool',
-              message: `Calling ${data.tool}(${data.arguments || ''})`,
+              message: `Calling ${data.tool}(${serializeToolCallArguments(data.arguments)})`,
             });
           } catch {}
         } else if (eventName === 'tool_call_end') {

frontend/src/components/Chat/ToolCallCard.tsx

@@ -1,6 +1,7 @@
 import { useState } from 'react';
 import { ChevronDown, ChevronRight, Loader2, CheckCircle2, XCircle } from 'lucide-react';
 import type { ToolCallInfo } from '../../types';
+import { serializeToolCallArguments } from '../../lib/tool-call';
 
 interface Props {
   toolCall: ToolCallInfo;
@@ -35,7 +36,10 @@ export function ToolCallCard({ toolCall }: Props) {
   const [expanded, setExpanded] = useState(false);
   const config = statusConfig[toolCall.status];
   const StatusIcon = config.icon;
-  const preview = previewArgs(toolCall.arguments);
+  // Persisted conversations may contain the pre-fix object payload despite
+  // the TypeScript contract, so normalize again at the final render boundary.
+  const argumentsText = serializeToolCallArguments(toolCall.arguments);
+  const preview = previewArgs(argumentsText);
 
   return (
     <div
@@ -95,7 +99,7 @@ export function ToolCallCard({ toolCall }: Props) {
           className="px-2.5 pb-2 pt-0.5"
           style={{ borderTop: '1px solid var(--color-border-subtle, var(--color-border))' }}
         >
-          {toolCall.arguments && (
+          {argumentsText && (
             <div className="mt-1.5">
               <div
                 style={{
@@ -120,7 +124,7 @@ export function ToolCallCard({ toolCall }: Props) {
                   wordBreak: 'break-all',
                 }}
               >
-                {formatJson(toolCall.arguments)}
+                {formatJson(argumentsText)}
               </pre>
             </div>
           )}

frontend/src/lib/api.ts

@@ -1,5 +1,6 @@
 import type { ModelInfo, SavingsData, ServerInfo } from '../types';
 import { SUPABASE_ANON_KEY, SUPABASE_URL } from './supabase';
+import { serializeToolCallArguments } from './tool-call';
 
 // ---------------------------------------------------------------------------
 // Supabase config
@@ -741,7 +742,7 @@ export async function sendAgentMessage(
               const parsed = JSON.parse(data);
               callbacks?.onToolCallStart?.({
                 tool: parsed.tool,
-                arguments: parsed.arguments ?? '',
+                arguments: serializeToolCallArguments(parsed.arguments),
               });
             } catch {
               /* skip */

frontend/src/lib/store.tool-call-repair.test.ts

@@ -0,0 +1,122 @@
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
+
+const CONVERSATIONS_KEY = 'openjarvis-conversations';
+
+class MemoryStorage {
+  private store = new Map<string, string>();
+
+  getItem(key: string): string | null {
+    return this.store.get(key) ?? null;
+  }
+
+  setItem(key: string, value: string): void {
+    this.store.set(key, String(value));
+  }
+
+  removeItem(key: string): void {
+    this.store.delete(key);
+  }
+}
+
+beforeEach(() => {
+  vi.resetModules();
+  (globalThis as unknown as { localStorage: MemoryStorage }).localStorage =
+    new MemoryStorage();
+});
+
+afterEach(() => {
+  (globalThis as unknown as { localStorage?: MemoryStorage }).localStorage =
+    undefined;
+});
+
+describe('persisted tool calls', () => {
+  it('repairs parsed argument objects while loading conversations', async () => {
+    localStorage.setItem(
+      CONVERSATIONS_KEY,
+      JSON.stringify({
+        version: 1,
+        activeId: 'conversation-1',
+        conversations: {
+          'conversation-1': {
+            id: 'conversation-1',
+            title: 'Broken chat',
+            createdAt: 1,
+            updatedAt: 1,
+            model: 'test-model',
+            messages: [
+              {
+                id: 'assistant-1',
+                role: 'assistant',
+                content: '',
+                timestamp: 1,
+                toolCalls: [
+                  {
+                    id: 'call-1',
+                    tool: 'web_search',
+                    arguments: { query: 'python' },
+                    status: 'success',
+                  },
+                ],
+              },
+            ],
+          },
+        },
+      }),
+    );
+
+    const { useAppStore } = await import('./store');
+
+    expect(useAppStore.getState().messages[0].toolCalls?.[0].arguments).toBe(
+      '{"query":"python"}',
+    );
+    const repaired = JSON.parse(localStorage.getItem(CONVERSATIONS_KEY) ?? '{}');
+    expect(
+      repaired.conversations['conversation-1'].messages[0].toolCalls[0].arguments,
+    ).toBe('{"query":"python"}');
+  });
+
+  it('keeps repaired conversations in memory when writeback fails', async () => {
+    localStorage.setItem(
+      CONVERSATIONS_KEY,
+      JSON.stringify({
+        version: 1,
+        activeId: 'conversation-1',
+        conversations: {
+          'conversation-1': {
+            id: 'conversation-1',
+            title: 'Readable chat',
+            createdAt: 1,
+            updatedAt: 1,
+            model: 'test-model',
+            messages: [
+              {
+                id: 'assistant-1',
+                role: 'assistant',
+                content: '',
+                timestamp: 1,
+                toolCalls: [
+                  {
+                    id: 'call-1',
+                    tool: 'web_search',
+                    arguments: { query: 'python' },
+                    status: 'success',
+                  },
+                ],
+              },
+            ],
+          },
+        },
+      }),
+    );
+    vi.spyOn(localStorage, 'setItem').mockImplementation(() => {
+      throw new DOMException('Storage quota exceeded', 'QuotaExceededError');
+    });
+
+    const { useAppStore } = await import('./store');
+
+    expect(useAppStore.getState().messages).toHaveLength(1);
+    expect(useAppStore.getState().messages[0].toolCalls?.[0].arguments).toBe(
+      '{"query":"python"}',
+    );
+  });
+});

frontend/src/lib/store.ts

@@ -16,6 +16,7 @@ import type {
 } from '../types';
 import type { ManagedAgent } from './api';
 import { isEmbedOnlyModel } from './model-capabilities';
+import { serializeToolCallArguments } from './tool-call';
 
 export interface CachedConnector {
   connector_id: string;
@@ -55,7 +56,30 @@ function loadConversations(): ConversationStore {
     const raw = localStorage.getItem(CONVERSATIONS_KEY);
     if (!raw) return { version: 1, conversations: {}, activeId: null };
     const parsed = JSON.parse(raw);
-    if (parsed.version === 1) return parsed;
+    if (parsed.version === 1) {
+      let repaired = false;
+      for (const conversation of Object.values(parsed.conversations ?? {}) as Conversation[]) {
+        for (const message of conversation.messages ?? []) {
+          for (const toolCall of message.toolCalls ?? []) {
+            const argumentsText = serializeToolCallArguments(toolCall.arguments);
+            if (argumentsText !== toolCall.arguments) {
+              toolCall.arguments = argumentsText;
+              repaired = true;
+            }
+          }
+        }
+      }
+      if (repaired) {
+        try {
+          localStorage.setItem(CONVERSATIONS_KEY, JSON.stringify(parsed));
+        } catch {
+          // Keep the repaired conversations usable in memory when storage is
+          // read-only or full. A failed best-effort writeback must not make
+          // otherwise readable conversation history disappear from the UI.
+        }
+      }
+      return parsed;
+    }
     return { version: 1, conversations: {}, activeId: null };
   } catch {
     return { version: 1, conversations: {}, activeId: null };

frontend/src/lib/tool-call.test.ts

@@ -0,0 +1,22 @@
+import { describe, expect, it } from 'vitest';
+
+import { serializeToolCallArguments } from './tool-call';
+
+describe('serializeToolCallArguments', () => {
+  it('preserves JSON strings', () => {
+    expect(serializeToolCallArguments('{"query":"python"}')).toBe(
+      '{"query":"python"}',
+    );
+  });
+
+  it('serializes parsed argument objects', () => {
+    expect(serializeToolCallArguments({ query: 'python' })).toBe(
+      '{"query":"python"}',
+    );
+  });
+
+  it('uses an empty string for missing arguments', () => {
+    expect(serializeToolCallArguments(null)).toBe('');
+    expect(serializeToolCallArguments(undefined)).toBe('');
+  });
+});

frontend/src/lib/tool-call.ts

@@ -0,0 +1,11 @@
+/** Convert tool-call arguments from API or persisted data into display-safe text. */
+export function serializeToolCallArguments(value: unknown): string {
+  if (typeof value === 'string') return value;
+  if (value == null) return '';
+
+  try {
+    return JSON.stringify(value) ?? String(value);
+  } catch {
+    return String(value);
+  }
+}

src/openjarvis/cli/serve.py

@@ -25,6 +25,30 @@ from openjarvis.intelligence import (
 
 logger = logging.getLogger(__name__)
 
+_DEFAULT_TOOLS = frozenset({"think", "calculator", "web_search"})
+
+
+def _resolve_allowed_tools(config: object) -> tuple[set[str], bool]:
+    """Return configured tool names and whether the selection was explicit.
+
+    ``tools.enabled`` is the canonical setting used by ``SystemBuilder`` and
+    the interactive CLI.  ``agent.tools`` remains as a backward-compatible
+    fallback, followed by the server's default tool set when neither is set.
+    """
+    configured = config.tools.enabled or config.agent.tools
+    if not configured:
+        return set(_DEFAULT_TOOLS), False
+
+    if isinstance(configured, list):
+        allowed = {
+            tool.strip()
+            for tool in configured
+            if isinstance(tool, str) and tool.strip()
+        }
+    else:
+        allowed = {tool.strip() for tool in configured.split(",") if tool.strip()}
+    return allowed, True
+
 
 def _unique_model_ids(model_ids: list[str]) -> list[str]:
     """Return model ids in first-seen order without duplicates."""
@@ -96,7 +120,7 @@ def _resolve_server_model(
     "--agent",
     "agent_name",
     default=None,
-    help="Agent for non-streaming requests (simple, orchestrator, react, openhands).",
+    help="Agent for chat requests (simple, orchestrator, react, openhands).",
 )
 @click.pass_context
 def serve(
@@ -305,21 +329,7 @@ def serve(
                     from openjarvis.core.registry import ToolRegistry
                     from openjarvis.tools._stubs import BaseTool
 
-                    _DEFAULT_TOOLS = {"think", "calculator", "web_search"}
-                    configured = config.agent.tools
-                    if configured:
-                        if isinstance(configured, list):
-                            allowed = {
-                                t.strip()
-                                for t in configured
-                                if isinstance(t, str) and t.strip()
-                            }
-                        else:
-                            allowed = {
-                                t.strip() for t in configured.split(",") if t.strip()
-                            }
-                    else:
-                        allowed = _DEFAULT_TOOLS
+                    allowed, tools_configured = _resolve_allowed_tools(config)
 
                     tools = []
                     for name in ToolRegistry.keys():
@@ -336,7 +346,7 @@ def serve(
                     # MCP server tools from config.tools.mcp.servers
                     # (#461 — these were silently dropped).
                     mcp_tools = managed_mcp_tools
-                    if configured:
+                    if tools_configured:
                         mcp_tools = [
                             tool
                             for tool in managed_mcp_tools
@@ -406,23 +416,7 @@ def serve(
                         from openjarvis.core.registry import ToolRegistry
                         from openjarvis.tools._stubs import BaseTool
 
-                        _DEFAULT_TOOLS = {"think", "calculator", "web_search"}
-                        configured = config.agent.tools
-                        if configured:
-                            if isinstance(configured, list):
-                                _allowed = {
-                                    t.strip()
-                                    for t in configured
-                                    if isinstance(t, str) and t.strip()
-                                }
-                            else:
-                                _allowed = {
-                                    t.strip()
-                                    for t in configured.split(",")
-                                    if t.strip()
-                                }
-                        else:
-                            _allowed = _DEFAULT_TOOLS
+                        _allowed, _tools_configured = _resolve_allowed_tools(config)
 
                         for _tname in ToolRegistry.keys():
                             if _tname not in _allowed:
@@ -436,7 +430,7 @@ def serve(
                         # Reuse the process-owned MCP pool so channels do not
                         # open a second transport to every configured server.
                         _ch_mcp_tools = managed_mcp_tools
-                        if configured:
+                        if _tools_configured:
                             _ch_mcp_tools = [
                                 tool
                                 for tool in managed_mcp_tools

src/openjarvis/server/routes.py

@@ -200,12 +200,14 @@ async def chat_completions(request_body: ChatCompletionRequest, request: Request
         # When the client passes `tools`, stream the model's raw
         # OpenAI-compat function-calling decision directly from the engine
         # (bypassing the agent) — the streaming mirror of the non-streaming
-        # #454 fix.  Routing tools through the agent stream bridge ignored
-        # `request_body.tools`, ran the agent's own tool loop, and
-        # word-split generic filler content into fake token deltas, so the
-        # caller's tool_calls were dropped entirely (the streaming analog of
-        # #414).  For plain chat (no tools), stream token-by-token directly
-        # from the engine for true real-time output.
+        # #454 fix. Routing client-supplied tools through a server-side agent
+        # would execute the agent's different tool set and drop the raw tool
+        # call the caller expects (#414).
+        #
+        # Without client-supplied tools, keep streaming requests on the
+        # configured server agent so its server-side tool loop is available
+        # to the desktop UI and other stream:true clients (#735). Fall back to
+        # direct token streaming when no tool-bearing agent is configured.
         if request_body.tools:
             return await _handle_stream_tools(
                 engine,
@@ -216,6 +218,16 @@ async def chat_completions(request_body: ChatCompletionRequest, request: Request
                 bus=getattr(request.app.state, "bus", None),
                 memory_service=getattr(request.app.state, "memory_service", None),
             )
+        if agent is not None and getattr(agent, "_tools", None):
+            return await _handle_agent_stream(
+                agent,
+                model,
+                request_body,
+                complexity_info,
+                trace_store=getattr(request.app.state, "trace_store", None),
+                bus=getattr(request.app.state, "bus", None),
+                memory_service=getattr(request.app.state, "memory_service", None),
+            )
         return await _handle_stream(
             engine,
             model,
@@ -547,6 +559,114 @@ def _handle_agent(
     )
 
 
+async def _handle_agent_stream(
+    agent,
+    model: str,
+    req: ChatCompletionRequest,
+    complexity_info=None,
+    *,
+    trace_store=None,
+    bus=None,
+    memory_service=None,
+):
+    """Run the configured agent and return its result as an SSE response.
+
+    Agents own the tool-execution loop, which is synchronous today.  Run that
+    loop in a worker thread and stream its final answer once complete.  This
+    keeps ``stream:true`` clients (including the desktop UI) on the same agent
+    and configured toolkit as non-streaming requests instead of bypassing the
+    agent and silently dropping server-side tools.
+
+    Requests that explicitly supply OpenAI ``tools`` continue to use
+    ``_handle_stream_tools`` so their raw tool-call deltas are preserved.
+    """
+    chunk_id = f"chatcmpl-{uuid.uuid4().hex[:12]}"
+    query_text = ""
+    for message in reversed(req.messages):
+        if message.role == "user" and message.content:
+            query_text = message.content
+            break
+
+    async def generate():
+        first_chunk = ChatCompletionChunk(
+            id=chunk_id,
+            model=model,
+            choices=[StreamChoice(delta=DeltaMessage(role="assistant"))],
+        )
+        yield f"data: {first_chunk.model_dump_json()}\n\n"
+
+        try:
+            response = await asyncio.to_thread(
+                _handle_agent,
+                agent,
+                model,
+                req,
+                complexity_info,
+                trace_store=trace_store,
+                bus=bus,
+            )
+        except Exception as exc:
+            logging.getLogger("openjarvis.server").error(
+                "Agent stream error: %s",
+                exc,
+                exc_info=True,
+            )
+            error_chunk = ChatCompletionChunk(
+                id=chunk_id,
+                model=model,
+                choices=[
+                    StreamChoice(
+                        delta=DeltaMessage(
+                            content=f"Sorry, an error occurred: {exc}",
+                        ),
+                        finish_reason="stop",
+                    )
+                ],
+            )
+            yield f"data: {error_chunk.model_dump_json()}\n\n"
+            yield "data: [DONE]\n\n"
+            return
+
+        content = _response_content(response)
+        if content:
+            content_chunk = ChatCompletionChunk(
+                id=chunk_id,
+                model=model,
+                choices=[StreamChoice(delta=DeltaMessage(content=content))],
+            )
+            yield f"data: {content_chunk.model_dump_json()}\n\n"
+
+        import json as _json
+
+        finish_chunk = ChatCompletionChunk(
+            id=chunk_id,
+            model=model,
+            choices=[
+                StreamChoice(delta=DeltaMessage(), finish_reason="stop"),
+            ],
+        )
+        finish_data = _json.loads(finish_chunk.model_dump_json())
+        finish_data["usage"] = response.usage.model_dump()
+        if complexity_info is not None:
+            finish_data["complexity"] = complexity_info.model_dump()
+        yield f"data: {_json.dumps(finish_data)}\n\n"
+
+        _record_completed_exchange(
+            memory_service,
+            query_text,
+            content,
+            bus=bus,
+            source="server.chat.stream",
+        )
+        yield "data: [DONE]\n\n"
+
+    return StreamingResponse(
+        generate(),
+        media_type="text/event-stream",
+        headers={"Cache-Control": "no-cache", "Connection": "keep-alive"},
+    )
+
+
 async def _handle_stream_tools(
     engine,
     model: str,
@@ -690,11 +810,10 @@ async def _handle_stream(
 ):
     """Stream response using SSE format.
 
-    This path streams straight from the engine, bypassing the agent /
+    This no-agent fallback streams straight from the engine, bypassing the
     ``TraceCollector``. When *trace_store* is set we accumulate the streamed
     tokens and record a minimal ``Trace`` once the stream completes
-    successfully — otherwise streamed chats (the desktop GUI's main path)
-    would never populate ``traces.db``.
+    successfully.
     """
     import time
 

src/openjarvis/server/stream_bridge.py

@@ -110,6 +110,12 @@ class AgentStreamBridge:
 
     def _format_named_event(self, name: str, data: dict) -> str:
         """Format an SSE event with an explicit ``event:`` field."""
+        if name == "tool_call_start" and not isinstance(data.get("arguments"), str):
+            # The in-process event bus uses parsed arguments for trace/eval
+            # consumers, while the web SSE contract expects their JSON text.
+            # Copy before normalizing so other subscribers keep the object.
+            data = dict(data)
+            data["arguments"] = json.dumps(data.get("arguments"))
         return f"event: {name}\ndata: {json.dumps(data)}\n\n"
 
     def _run_agent(self) -> object:

tests/cli/test_serve_tools.py

@@ -0,0 +1,53 @@
+"""Regression tests for tool selection during ``jarvis serve`` startup."""
+
+from __future__ import annotations
+
+import pytest
+
+from openjarvis.cli.serve import _resolve_allowed_tools
+from openjarvis.core.config import JarvisConfig
+
+
+@pytest.mark.parametrize(
+    "configured",
+    [
+        "code_interpreter,file_read",
+        ["code_interpreter", "file_read"],
+    ],
+)
+def test_tools_enabled_is_used_by_serve(configured):
+    config = JarvisConfig()
+    config.tools.enabled = configured
+
+    allowed, explicit = _resolve_allowed_tools(config)
+
+    assert allowed == {"code_interpreter", "file_read"}
+    assert explicit is True
+
+
+def test_tools_enabled_takes_precedence_over_legacy_agent_tools():
+    config = JarvisConfig()
+    config.tools.enabled = "file_read"
+    config.agent.tools = "calculator"
+
+    allowed, explicit = _resolve_allowed_tools(config)
+
+    assert allowed == {"file_read"}
+    assert explicit is True
+
+
+def test_agent_tools_remains_a_backward_compatible_fallback():
+    config = JarvisConfig()
+    config.agent.tools = "file_read"
+
+    allowed, explicit = _resolve_allowed_tools(config)
+
+    assert allowed == {"file_read"}
+    assert explicit is True
+
+
+def test_serve_defaults_tools_when_no_selection_is_configured():
+    allowed, explicit = _resolve_allowed_tools(JarvisConfig())
+
+    assert allowed == {"think", "calculator", "web_search"}
+    assert explicit is False

tests/server/test_model_management.py

@@ -213,6 +213,7 @@ class TestStreamingResilience:
         engine = _make_engine()
         agent = MagicMock()
         agent.agent_id = "simple"
+        agent._tools = []
         agent.run.return_value = AgentResult(
             content="agent response",
             turns=1,

tests/server/test_routes.py

@@ -534,6 +534,94 @@ class TestChatCompletions:
                     content += delta_content
         assert content == "Hello world"
 
+    def test_streaming_without_client_tools_uses_configured_agent(self):
+        """Server-side tools remain available to streaming web clients (#735)."""
+        from openjarvis.agents.orchestrator import OrchestratorAgent
+        from openjarvis.core.types import ToolResult
+        from openjarvis.tools._stubs import BaseTool, ToolSpec
+
+        executions: list[str] = []
+
+        class _FileReadTool(BaseTool):
+            @property
+            def spec(self):
+                return ToolSpec(
+                    name="file_read",
+                    description="Read a file",
+                    parameters={
+                        "type": "object",
+                        "properties": {"path": {"type": "string"}},
+                    },
+                )
+
+            def execute(self, **params):
+                executions.append(params["path"])
+                return ToolResult(
+                    tool_name="file_read",
+                    content="README fixture contents",
+                    success=True,
+                )
+
+        engine = _make_engine(content="ENGINE BYPASS")
+        engine.generate.side_effect = [
+            {
+                "content": "",
+                "tool_calls": [
+                    {
+                        "id": "call_1",
+                        "name": "file_read",
+                        "arguments": '{"path": "README.md"}',
+                    }
+                ],
+                "usage": {},
+            },
+            {
+                "content": "README fixture contents",
+                "finish_reason": "stop",
+                "usage": {},
+            },
+        ]
+        agent = OrchestratorAgent(
+            engine,
+            "test-model",
+            tools=[_FileReadTool()],
+            bus=EventBus(),
+            max_turns=3,
+            temperature=0.7,
+            max_tokens=128,
+            system_prompt="Use the configured tools.",
+        )
+        app = create_app(
+            engine,
+            "test-model",
+            agent=agent,
+            bus=EventBus(),
+            config=_test_config(),
+        )
+        client = TestClient(app)
+
+        resp = client.post(
+            "/v1/chat/completions",
+            json={
+                "model": "test-model",
+                "messages": [{"role": "user", "content": "Read README.md"}],
+                "stream": True,
+            },
+        )
+
+        assert resp.status_code == 200
+        content = ""
+        for line in resp.text.strip().split("\n"):
+            if not line.startswith("data:") or "[DONE]" in line:
+                continue
+            data = json.loads(line[5:].strip())
+            delta = data.get("choices", [{}])[0].get("delta", {})
+            content += delta.get("content") or ""
+
+        assert content == "README fixture contents"
+        assert executions == ["README.md"]
+        assert engine.generate.call_count == 2
+
     def test_streaming_with_tools_emits_tool_calls_and_bypasses_agent(self):
         """Regression for the streaming analog of #414.
 

tests/server/test_stream_bridge.py

@@ -0,0 +1,30 @@
+import json
+
+from openjarvis.server.stream_bridge import AgentStreamBridge
+
+
+def test_tool_call_start_serializes_arguments_for_sse_without_mutating_event():
+    bridge = object.__new__(AgentStreamBridge)
+    event_data = {
+        "tool": "web_search",
+        "arguments": {"query": "python"},
+        "agent": "agent-1",
+    }
+
+    event = bridge._format_named_event("tool_call_start", event_data)
+    payload = json.loads(event.split("data: ", 1)[1])
+
+    assert payload["arguments"] == '{"query": "python"}'
+    assert event_data["arguments"] == {"query": "python"}
+
+
+def test_tool_call_start_preserves_already_serialized_arguments():
+    bridge = object.__new__(AgentStreamBridge)
+
+    event = bridge._format_named_event(
+        "tool_call_start",
+        {"tool": "web_search", "arguments": '{"query":"python"}'},
+    )
+    payload = json.loads(event.split("data: ", 1)[1])
+
+    assert payload["arguments"] == '{"query":"python"}'