feat: strengthen tool emulation prompting

Improve proxy-side tool instructions so models more reliably emit structured tool actions, and add focused tests covering prompt guidance and default action limits.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

tests/test_tool_call_bridge.py

@@ -65,6 +65,7 @@ from starlette.requests import Request
 from starlette.responses import JSONResponse, Response, StreamingResponse
 
 from app.anthropic_schema import AnthropicMessagesRequest
+from app.http.tool_emulation import EmulatedToolChoice, EmulatedToolDef, inject_tooling, parse_action_blocks
 from app.openai_schema import ChatCompletionsRequest, ResponsesRequest
 import app.main as main
 
@@ -2789,6 +2790,54 @@ class AdminIntrospectionEndpointTests(unittest.IsolatedAsyncioTestCase):
         self.assertEqual(ctx.exception.status_code, 401)
 
 
+class ToolEmulationPromptTests(unittest.TestCase):
+    def test_inject_tooling_adds_routing_hints_and_examples(self) -> None:
+        tools = [
+            EmulatedToolDef(
+                name="read_file",
+                description="Read a file",
+                input_schema={"type": "object", "properties": {"path": {"type": "string"}}},
+            ),
+            EmulatedToolDef(
+                name="bash",
+                description="Run shell commands",
+                input_schema={"type": "object", "properties": {"command": {"type": "string"}}},
+            ),
+        ]
+
+        injected = inject_tooling("system prompt", tools, EmulatedToolChoice(mode="auto"))
+
+        self.assertIn("Tool routing guide:", injected)
+        self.assertIn("use read_file", injected)
+        self.assertIn("use bash", injected)
+        self.assertIn("Core tool syntax examples.", injected)
+        self.assertIn("Example valid action block", injected)
+        self.assertIn("tool_choice=auto means you must decide whether the user request needs a tool", injected)
+
+    def test_inject_tooling_does_not_modify_plain_system_when_no_tools(self) -> None:
+        injected = inject_tooling("system prompt", [], EmulatedToolChoice(mode="auto"))
+        self.assertEqual(injected, "system prompt")
+
+    def test_parse_action_blocks_limits_default_to_five_calls(self) -> None:
+        tools = [
+            EmulatedToolDef(
+                name="lookup",
+                description="Lookup data",
+                input_schema={"type": "object", "properties": {"q": {"type": "string"}}},
+            )
+        ]
+        text = "\n".join(
+            f"```json action\n{{\"tool\":\"lookup\",\"parameters\":{{\"q\":\"item-{i}\"}}}}\n```"
+            for i in range(6)
+        )
+
+        calls, remaining = parse_action_blocks(text, tools)
+
+        self.assertEqual(len(calls), 5)
+        self.assertEqual([call.arguments["q"] for call in calls], [f"item-{i}" for i in range(5)])
+        self.assertIn('"q":"item-5"', remaining)
+
+
 class SessionCacheToolFingerprintTests(unittest.TestCase):
     def test_build_key_changes_with_tool_config(self) -> None:
         from app.session_cache import SessionCache