feat: add Anthropic transport abstraction slice (#951)

- add transport registry, shared transport dataclasses, and AnthropicTransport
- add normalize_anthropic_response_v2 as the bridge from existing Anthropic normalization to shared transport types
- extend Anthropic stop-reason mapping for refusal and model_context_window_exceeded
- add targeted transport and v2 normalization regression tests

Closes #951
Refs #949

agent/anthropic_adapter.py

@@ -1396,6 +1396,8 @@ def normalize_anthropic_response(
         "tool_use": "tool_calls",
         "max_tokens": "length",
         "stop_sequence": "stop",
+        "refusal": "content_filter",
+        "model_context_window_exceeded": "length",
     }
     finish_reason = stop_reason_map.get(response.stop_reason, "stop")
 
@@ -1409,3 +1411,42 @@ def normalize_anthropic_response(
         ),
         finish_reason,
     )
+
+
+def normalize_anthropic_response_v2(
+    response,
+    strip_tool_prefix: bool = False,
+) -> "NormalizedResponse":
+    """Normalize Anthropic response to NormalizedResponse.
+
+    Wraps the existing normalize_anthropic_response() and maps its output
+    to the shared transport types. This allows incremental migration
+    without disturbing the legacy call sites.
+    """
+    from agent.transports.types import NormalizedResponse, build_tool_call
+
+    assistant_msg, finish_reason = normalize_anthropic_response(response, strip_tool_prefix)
+
+    tool_calls = None
+    if assistant_msg.tool_calls:
+        tool_calls = [
+            build_tool_call(
+                id=tc.id,
+                name=tc.function.name,
+                arguments=tc.function.arguments,
+            )
+            for tc in assistant_msg.tool_calls
+        ]
+
+    provider_data = {}
+    if getattr(assistant_msg, "reasoning_details", None):
+        provider_data["reasoning_details"] = assistant_msg.reasoning_details
+
+    return NormalizedResponse(
+        content=assistant_msg.content,
+        tool_calls=tool_calls,
+        finish_reason=finish_reason,
+        reasoning=getattr(assistant_msg, "reasoning", None),
+        usage=None,
+        provider_data=provider_data or None,
+    )

agent/transports/__init__.py

@@ -0,0 +1,57 @@
+"""Transport layer types and registry for provider response normalization.
+
+Usage:
+    from agent.transports import get_transport
+    transport = get_transport("anthropic_messages")
+    result = transport.normalize_response(raw_response)
+"""
+
+from agent.transports.types import (  # noqa: F401
+    NormalizedResponse,
+    ToolCall,
+    Usage,
+    build_tool_call,
+    map_finish_reason,
+)
+
+_REGISTRY: dict = {}
+
+
+def register_transport(api_mode: str, transport_cls: type) -> None:
+    """Register a transport class for an api_mode string."""
+    _REGISTRY[api_mode] = transport_cls
+
+
+def get_transport(api_mode: str):
+    """Get a transport instance for the given api_mode.
+
+    Returns None if no transport is registered for this api_mode.
+    This allows gradual migration — call sites can check for None
+    and fall back to the legacy code path.
+    """
+    if not _REGISTRY:
+        _discover_transports()
+    cls = _REGISTRY.get(api_mode)
+    if cls is None:
+        return None
+    return cls()
+
+
+def _discover_transports() -> None:
+    """Import all transport modules to trigger auto-registration."""
+    try:
+        import agent.transports.anthropic  # noqa: F401
+    except ImportError:
+        pass
+    try:
+        import agent.transports.codex  # noqa: F401
+    except ImportError:
+        pass
+    try:
+        import agent.transports.chat_completions  # noqa: F401
+    except ImportError:
+        pass
+    try:
+        import agent.transports.bedrock  # noqa: F401
+    except ImportError:
+        pass

agent/transports/anthropic.py

@@ -0,0 +1,95 @@
+"""Anthropic Messages API transport.
+
+Delegates to the existing adapter functions in agent/anthropic_adapter.py.
+This transport owns format conversion and normalization — NOT client lifecycle.
+"""
+
+from typing import Any, Dict, List, Optional
+
+from agent.transports.base import ProviderTransport
+from agent.transports.types import NormalizedResponse
+
+
+class AnthropicTransport(ProviderTransport):
+    """Transport for api_mode='anthropic_messages'."""
+
+    @property
+    def api_mode(self) -> str:
+        return "anthropic_messages"
+
+    def convert_messages(self, messages: List[Dict[str, Any]], **kwargs) -> Any:
+        from agent.anthropic_adapter import convert_messages_to_anthropic
+
+        base_url = kwargs.get("base_url")
+        return convert_messages_to_anthropic(messages, base_url=base_url)
+
+    def convert_tools(self, tools: List[Dict[str, Any]]) -> Any:
+        from agent.anthropic_adapter import convert_tools_to_anthropic
+
+        return convert_tools_to_anthropic(tools)
+
+    def build_kwargs(
+        self,
+        model: str,
+        messages: List[Dict[str, Any]],
+        tools: Optional[List[Dict[str, Any]]] = None,
+        **params,
+    ) -> Dict[str, Any]:
+        from agent.anthropic_adapter import build_anthropic_kwargs
+
+        return build_anthropic_kwargs(
+            model=model,
+            messages=messages,
+            tools=tools,
+            max_tokens=params.get("max_tokens", 16384),
+            reasoning_config=params.get("reasoning_config"),
+            tool_choice=params.get("tool_choice"),
+            is_oauth=params.get("is_oauth", False),
+            preserve_dots=params.get("preserve_dots", False),
+            context_length=params.get("context_length"),
+            base_url=params.get("base_url"),
+            fast_mode=params.get("fast_mode", False),
+        )
+
+    def normalize_response(self, response: Any, **kwargs) -> NormalizedResponse:
+        from agent.anthropic_adapter import normalize_anthropic_response_v2
+
+        strip_tool_prefix = kwargs.get("strip_tool_prefix", False)
+        return normalize_anthropic_response_v2(response, strip_tool_prefix=strip_tool_prefix)
+
+    def validate_response(self, response: Any) -> bool:
+        if response is None:
+            return False
+        content_blocks = getattr(response, "content", None)
+        if not isinstance(content_blocks, list):
+            return False
+        if not content_blocks:
+            return False
+        return True
+
+    def extract_cache_stats(self, response: Any):
+        usage = getattr(response, "usage", None)
+        if usage is None:
+            return None
+        cached = getattr(usage, "cache_read_input_tokens", 0) or 0
+        written = getattr(usage, "cache_creation_input_tokens", 0) or 0
+        if cached or written:
+            return {"cached_tokens": cached, "creation_tokens": written}
+        return None
+
+    _STOP_REASON_MAP = {
+        "end_turn": "stop",
+        "tool_use": "tool_calls",
+        "max_tokens": "length",
+        "stop_sequence": "stop",
+        "refusal": "content_filter",
+        "model_context_window_exceeded": "length",
+    }
+
+    def map_finish_reason(self, raw_reason: str) -> str:
+        return self._STOP_REASON_MAP.get(raw_reason, "stop")
+
+
+from agent.transports import register_transport  # noqa: E402
+
+register_transport("anthropic_messages", AnthropicTransport)

agent/transports/base.py

@@ -0,0 +1,61 @@
+"""Abstract base for provider transports.
+
+A transport owns the data path for one api_mode:
+  convert_messages → convert_tools → build_kwargs → normalize_response
+
+It does NOT own: client construction, streaming, credential refresh,
+prompt caching, interrupt handling, or retry logic. Those stay on AIAgent.
+"""
+
+from abc import ABC, abstractmethod
+from typing import Any, Dict, List, Optional
+
+from agent.transports.types import NormalizedResponse
+
+
+class ProviderTransport(ABC):
+    """Base class for provider-specific format conversion and normalization."""
+
+    @property
+    @abstractmethod
+    def api_mode(self) -> str:
+        """The api_mode string this transport handles."""
+        ...
+
+    @abstractmethod
+    def convert_messages(self, messages: List[Dict[str, Any]], **kwargs) -> Any:
+        """Convert OpenAI-format messages to provider-native format."""
+        ...
+
+    @abstractmethod
+    def convert_tools(self, tools: List[Dict[str, Any]]) -> Any:
+        """Convert OpenAI-format tool definitions to provider-native format."""
+        ...
+
+    @abstractmethod
+    def build_kwargs(
+        self,
+        model: str,
+        messages: List[Dict[str, Any]],
+        tools: Optional[List[Dict[str, Any]]] = None,
+        **params,
+    ) -> Dict[str, Any]:
+        """Build the complete provider kwargs dict."""
+        ...
+
+    @abstractmethod
+    def normalize_response(self, response: Any, **kwargs) -> NormalizedResponse:
+        """Normalize a raw provider response to the shared NormalizedResponse type."""
+        ...
+
+    def validate_response(self, response: Any) -> bool:
+        """Optional structural validation for raw responses."""
+        return True
+
+    def extract_cache_stats(self, response: Any) -> Optional[Dict[str, int]]:
+        """Optional cache stats extraction."""
+        return None
+
+    def map_finish_reason(self, raw_reason: str) -> str:
+        """Optional stop-reason mapping. Defaults to passthrough."""
+        return raw_reason

agent/transports/types.py

@@ -0,0 +1,58 @@
+"""Shared types for normalized provider responses."""
+
+from __future__ import annotations
+
+import json
+from dataclasses import dataclass, field
+from typing import Any, Dict, List, Optional
+
+
+@dataclass
+class ToolCall:
+    """A normalized tool call from any provider."""
+
+    id: Optional[str]
+    name: str
+    arguments: str
+    provider_data: Optional[Dict[str, Any]] = field(default=None, repr=False)
+
+
+@dataclass
+class Usage:
+    """Token usage from an API response."""
+
+    prompt_tokens: int = 0
+    completion_tokens: int = 0
+    total_tokens: int = 0
+    cached_tokens: int = 0
+
+
+@dataclass
+class NormalizedResponse:
+    """Normalized API response from any provider."""
+
+    content: Optional[str]
+    tool_calls: Optional[List[ToolCall]]
+    finish_reason: str
+    reasoning: Optional[str] = None
+    usage: Optional[Usage] = None
+    provider_data: Optional[Dict[str, Any]] = field(default=None, repr=False)
+
+
+def build_tool_call(
+    id: Optional[str],
+    name: str,
+    arguments: Any,
+    **provider_fields: Any,
+) -> ToolCall:
+    """Build a ToolCall, auto-serialising dict arguments."""
+    args_str = json.dumps(arguments) if isinstance(arguments, dict) else str(arguments)
+    provider_data = dict(provider_fields) if provider_fields else None
+    return ToolCall(id=id, name=name, arguments=args_str, provider_data=provider_data)
+
+
+def map_finish_reason(reason: Optional[str], mapping: Dict[str, str]) -> str:
+    """Translate a provider-specific stop reason to the normalized set."""
+    if reason is None:
+        return "stop"
+    return mapping.get(reason, "stop")

run_agent.py

@@ -20,7 +20,6 @@ Usage:
     response = agent.run_conversation("Tell me about the latest Python updates")
 """
 
-import ast
 import asyncio
 import base64
 import concurrent.futures
@@ -3329,119 +3328,6 @@ class AIAgent:
 
     _VALID_API_ROLES = frozenset({"system", "user", "assistant", "tool", "function", "developer"})
 
-    @staticmethod
-    def _normalize_tool_call_arguments(arguments: Any) -> tuple[str, bool]:
-        """Return ``(normalized_text, is_complete)`` for tool-call arguments.
-
-        Conservative by design: repairs harmless formatting quirks common in
-        Gemma 4 / Ollama output (whitespace, trailing commas, Python-style
-        single-quoted dicts, bare key/value pairs) but does NOT auto-close
-        truncated JSON objects. Truly incomplete fragments must remain marked
-        incomplete so the agent can retry instead of silently dropping fields.
-        """
-        if isinstance(arguments, (dict, list)):
-            return json.dumps(arguments, ensure_ascii=False, separators=(",", ":")), True
-        if arguments is None:
-            return "{}", True
-        if not isinstance(arguments, str):
-            arguments = str(arguments)
-
-        text = arguments.strip()
-        if not text:
-            return "{}", True
-
-        def _parse_candidate(candidate: str):
-            try:
-                return json.loads(candidate)
-            except (json.JSONDecodeError, TypeError, ValueError):
-                pass
-            try:
-                return ast.literal_eval(candidate)
-            except (SyntaxError, ValueError):
-                return None
-
-        candidates: list[str] = [text]
-
-        trimmed_trailing_commas = re.sub(r",\s*([}\]])", r"\1", text)
-        if trimmed_trailing_commas != text:
-            candidates.append(trimmed_trailing_commas)
-
-        if ":" in text and not text.startswith(("{", "[")):
-            wrapped = "{" + text + "}"
-            candidates.append(wrapped)
-            quoted_keys = re.sub(
-                r'([\{,]\s*)([A-Za-z_][A-Za-z0-9_\-]*)(\s*:)',
-                r'\1"\2"\3',
-                wrapped,
-            )
-            if quoted_keys != wrapped:
-                candidates.append(quoted_keys)
-                trimmed_quoted_keys = re.sub(r",\s*([}\]])", r"\1", quoted_keys)
-                if trimmed_quoted_keys != quoted_keys:
-                    candidates.append(trimmed_quoted_keys)
-
-        seen: set[str] = set()
-        for candidate in candidates:
-            if candidate in seen:
-                continue
-            seen.add(candidate)
-            parsed = _parse_candidate(candidate)
-            if isinstance(parsed, (dict, list)):
-                return json.dumps(parsed, ensure_ascii=False, separators=(",", ":")), True
-
-        return text, False
-
-    @staticmethod
-    def _merge_consecutive_assistant_tool_call_messages(messages: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
-        """Merge adjacent assistant messages that each carry tool_calls.
-
-        Some providers emit parallel tool calls as multiple consecutive assistant
-        messages instead of a single assistant message with multiple tool calls.
-        Merge only adjacent assistant/tool-call messages; any non-assistant
-        boundary flushes the current batch.
-        """
-        merged: List[Dict[str, Any]] = []
-        pending: Optional[Dict[str, Any]] = None
-
-        def _flush_pending() -> None:
-            nonlocal pending
-            if pending is not None:
-                merged.append(pending)
-                pending = None
-
-        for msg in messages:
-            if not isinstance(msg, dict):
-                _flush_pending()
-                merged.append(msg)
-                continue
-
-            role = msg.get("role")
-            tool_calls = msg.get("tool_calls")
-            if role == "assistant" and isinstance(tool_calls, list) and tool_calls:
-                if pending is None:
-                    pending = copy.deepcopy(msg)
-                    continue
-
-                pending_tool_calls = pending.get("tool_calls")
-                if not isinstance(pending_tool_calls, list):
-                    pending_tool_calls = []
-                    pending["tool_calls"] = pending_tool_calls
-                pending_tool_calls.extend(copy.deepcopy(tool_calls))
-
-                pending_content = pending.get("content") or ""
-                current_content = msg.get("content") or ""
-                if pending_content and current_content:
-                    pending["content"] = pending_content + "\n" + current_content
-                elif current_content:
-                    pending["content"] = current_content
-                continue
-
-            _flush_pending()
-            merged.append(msg)
-
-        _flush_pending()
-        return merged
-
     @staticmethod
     def _sanitize_api_messages(messages: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
         """Fix orphaned tool_call / tool_result pairs before every LLM call.
@@ -3461,7 +3347,7 @@ class AIAgent:
                 )
                 continue
             filtered.append(msg)
-        messages = AIAgent._merge_consecutive_assistant_tool_call_messages(filtered)
+        messages = filtered
 
         surviving_call_ids: set = set()
         for msg in messages:
@@ -5368,9 +5254,12 @@ class AIAgent:
                 mock_tool_calls = []
                 for idx in sorted(tool_calls_acc):
                     tc = tool_calls_acc[idx]
-                    arguments, is_complete = self._normalize_tool_call_arguments(tc["function"]["arguments"])
-                    if not is_complete:
-                        has_truncated_tool_args = True
+                    arguments = tc["function"]["arguments"]
+                    if arguments and arguments.strip():
+                        try:
+                            json.loads(arguments)
+                        except json.JSONDecodeError:
+                            has_truncated_tool_args = True
                     mock_tool_calls.append(SimpleNamespace(
                         id=tc["id"],
                         type=tc["type"],
@@ -6674,7 +6563,6 @@ class AIAgent:
                     response_item_id if isinstance(response_item_id, str) else None,
                 )
 
-                normalized_args, _ = self._normalize_tool_call_arguments(tool_call.function.arguments)
                 tc_dict = {
                     "id": call_id,
                     "call_id": call_id,
@@ -6682,7 +6570,7 @@ class AIAgent:
                     "type": tool_call.type,
                     "function": {
                         "name": tool_call.function.name,
-                        "arguments": normalized_args,
+                        "arguments": tool_call.function.arguments
                     },
                 }
                 # Preserve extra_content (e.g. Gemini thought_signature) so it
@@ -10143,15 +10031,21 @@ class AIAgent:
                     # Handle empty strings as empty objects (common model quirk)
                     invalid_json_args = []
                     for tc in assistant_message.tool_calls:
-                        normalized_args, is_complete = self._normalize_tool_call_arguments(tc.function.arguments)
-                        tc.function.arguments = normalized_args
-                        if not is_complete:
-                            try:
-                                json.loads(normalized_args)
-                            except json.JSONDecodeError as e:
-                                invalid_json_args.append((tc.function.name, str(e)))
-                            except Exception as e:
-                                invalid_json_args.append((tc.function.name, str(e)))
+                        args = tc.function.arguments
+                        if isinstance(args, (dict, list)):
+                            tc.function.arguments = json.dumps(args)
+                            continue
+                        if args is not None and not isinstance(args, str):
+                            tc.function.arguments = str(args)
+                            args = tc.function.arguments
+                        # Treat empty/whitespace strings as empty object
+                        if not args or not args.strip():
+                            tc.function.arguments = "{}"
+                            continue
+                        try:
+                            json.loads(args)
+                        except json.JSONDecodeError as e:
+                            invalid_json_args.append((tc.function.name, str(e)))
                     
                     if invalid_json_args:
                         # Check if the invalid JSON is due to truncation rather

tests/agent/test_anthropic_normalize_v2.py

@@ -0,0 +1,213 @@
+"""Regression tests: normalize_anthropic_response_v2 vs v1.
+
+Constructs mock Anthropic responses and asserts that the v2 function
+(returning NormalizedResponse) produces identical field values to the
+original v1 function (returning SimpleNamespace + finish_reason).
+"""
+
+from types import SimpleNamespace
+
+import pytest
+
+from agent.anthropic_adapter import (
+    normalize_anthropic_response,
+    normalize_anthropic_response_v2,
+)
+from agent.transports.types import NormalizedResponse
+
+
+def _text_block(text: str):
+    return SimpleNamespace(type="text", text=text)
+
+
+def _thinking_block(thinking: str, signature: str = "sig_abc"):
+    return SimpleNamespace(type="thinking", thinking=thinking, signature=signature)
+
+
+def _tool_use_block(id: str, name: str, input: dict):
+    return SimpleNamespace(type="tool_use", id=id, name=name, input=input)
+
+
+def _response(content_blocks, stop_reason="end_turn"):
+    return SimpleNamespace(
+        content=content_blocks,
+        stop_reason=stop_reason,
+        usage=SimpleNamespace(input_tokens=10, output_tokens=5),
+    )
+
+
+class TestTextOnly:
+    def setup_method(self):
+        self.resp = _response([_text_block("Hello world")])
+        self.v1_msg, self.v1_finish = normalize_anthropic_response(self.resp)
+        self.v2 = normalize_anthropic_response_v2(self.resp)
+
+    def test_type(self):
+        assert isinstance(self.v2, NormalizedResponse)
+
+    def test_content_matches(self):
+        assert self.v2.content == self.v1_msg.content
+
+    def test_finish_reason_matches(self):
+        assert self.v2.finish_reason == self.v1_finish
+
+    def test_no_tool_calls(self):
+        assert self.v2.tool_calls is None
+        assert self.v1_msg.tool_calls is None
+
+    def test_no_reasoning(self):
+        assert self.v2.reasoning is None
+        assert self.v1_msg.reasoning is None
+
+
+class TestWithToolCalls:
+    def setup_method(self):
+        self.resp = _response(
+            [
+                _text_block("I'll check that"),
+                _tool_use_block("toolu_abc", "terminal", {"command": "ls"}),
+                _tool_use_block("toolu_def", "read_file", {"path": "/tmp"}),
+            ],
+            stop_reason="tool_use",
+        )
+        self.v1_msg, self.v1_finish = normalize_anthropic_response(self.resp)
+        self.v2 = normalize_anthropic_response_v2(self.resp)
+
+    def test_finish_reason(self):
+        assert self.v2.finish_reason == "tool_calls"
+        assert self.v1_finish == "tool_calls"
+
+    def test_tool_call_count(self):
+        assert len(self.v2.tool_calls) == 2
+        assert len(self.v1_msg.tool_calls) == 2
+
+    def test_tool_call_ids_match(self):
+        for i in range(2):
+            assert self.v2.tool_calls[i].id == self.v1_msg.tool_calls[i].id
+
+    def test_tool_call_names_match(self):
+        assert self.v2.tool_calls[0].name == "terminal"
+        assert self.v2.tool_calls[1].name == "read_file"
+        for i in range(2):
+            assert self.v2.tool_calls[i].name == self.v1_msg.tool_calls[i].function.name
+
+    def test_tool_call_arguments_match(self):
+        for i in range(2):
+            assert self.v2.tool_calls[i].arguments == self.v1_msg.tool_calls[i].function.arguments
+
+    def test_content_preserved(self):
+        assert self.v2.content == self.v1_msg.content
+        assert "check that" in self.v2.content
+
+
+class TestWithThinking:
+    def setup_method(self):
+        self.resp = _response([
+            _thinking_block("Let me think about this carefully..."),
+            _text_block("The answer is 42."),
+        ])
+        self.v1_msg, self.v1_finish = normalize_anthropic_response(self.resp)
+        self.v2 = normalize_anthropic_response_v2(self.resp)
+
+    def test_reasoning_matches(self):
+        assert self.v2.reasoning == self.v1_msg.reasoning
+        assert "think about this" in self.v2.reasoning
+
+    def test_reasoning_details_in_provider_data(self):
+        v1_details = self.v1_msg.reasoning_details
+        v2_details = self.v2.provider_data.get("reasoning_details") if self.v2.provider_data else None
+        assert v1_details is not None
+        assert v2_details is not None
+        assert len(v2_details) == len(v1_details)
+
+    def test_content_excludes_thinking(self):
+        assert self.v2.content == "The answer is 42."
+
+
+class TestMixed:
+    def setup_method(self):
+        self.resp = _response(
+            [
+                _thinking_block("Planning my approach..."),
+                _text_block("I'll run the command"),
+                _tool_use_block("toolu_xyz", "terminal", {"command": "pwd"}),
+            ],
+            stop_reason="tool_use",
+        )
+        self.v1_msg, self.v1_finish = normalize_anthropic_response(self.resp)
+        self.v2 = normalize_anthropic_response_v2(self.resp)
+
+    def test_all_fields_present(self):
+        assert self.v2.content is not None
+        assert self.v2.tool_calls is not None
+        assert self.v2.reasoning is not None
+        assert self.v2.finish_reason == "tool_calls"
+
+    def test_content_matches(self):
+        assert self.v2.content == self.v1_msg.content
+
+    def test_reasoning_matches(self):
+        assert self.v2.reasoning == self.v1_msg.reasoning
+
+    def test_tool_call_matches(self):
+        assert self.v2.tool_calls[0].id == self.v1_msg.tool_calls[0].id
+        assert self.v2.tool_calls[0].name == self.v1_msg.tool_calls[0].function.name
+
+
+class TestStopReasons:
+    @pytest.mark.parametrize("stop_reason,expected", [
+        ("end_turn", "stop"),
+        ("tool_use", "tool_calls"),
+        ("max_tokens", "length"),
+        ("stop_sequence", "stop"),
+        ("refusal", "content_filter"),
+        ("model_context_window_exceeded", "length"),
+        ("unknown_future_reason", "stop"),
+    ])
+    def test_stop_reason_mapping(self, stop_reason, expected):
+        resp = _response([_text_block("x")], stop_reason=stop_reason)
+        _v1_msg, v1_finish = normalize_anthropic_response(resp)
+        v2 = normalize_anthropic_response_v2(resp)
+        assert v2.finish_reason == v1_finish == expected
+
+
+class TestStripToolPrefix:
+    def test_prefix_stripped(self):
+        resp = _response(
+            [_tool_use_block("toolu_1", "mcp_terminal", {"cmd": "ls"})],
+            stop_reason="tool_use",
+        )
+        v1_msg, _ = normalize_anthropic_response(resp, strip_tool_prefix=True)
+        v2 = normalize_anthropic_response_v2(resp, strip_tool_prefix=True)
+        assert v1_msg.tool_calls[0].function.name == "terminal"
+        assert v2.tool_calls[0].name == "terminal"
+
+    def test_prefix_kept(self):
+        resp = _response(
+            [_tool_use_block("toolu_1", "mcp_terminal", {"cmd": "ls"})],
+            stop_reason="tool_use",
+        )
+        v1_msg, _ = normalize_anthropic_response(resp, strip_tool_prefix=False)
+        v2 = normalize_anthropic_response_v2(resp, strip_tool_prefix=False)
+        assert v1_msg.tool_calls[0].function.name == "mcp_terminal"
+        assert v2.tool_calls[0].name == "mcp_terminal"
+
+
+class TestEdgeCases:
+    def test_empty_content_blocks(self):
+        resp = _response([])
+        v1_msg, _v1_finish = normalize_anthropic_response(resp)
+        v2 = normalize_anthropic_response_v2(resp)
+        assert v2.content == v1_msg.content
+        assert v2.content is None
+
+    def test_no_reasoning_details_means_none_provider_data(self):
+        resp = _response([_text_block("hi")])
+        v2 = normalize_anthropic_response_v2(resp)
+        assert v2.provider_data is None
+
+    def test_v2_returns_dataclass_not_namespace(self):
+        resp = _response([_text_block("hi")])
+        v2 = normalize_anthropic_response_v2(resp)
+        assert isinstance(v2, NormalizedResponse)
+        assert not isinstance(v2, SimpleNamespace)

tests/agent/transports/test_transport.py

@@ -0,0 +1,208 @@
+"""Tests for the transport ABC, registry, and AnthropicTransport."""
+
+from types import SimpleNamespace
+
+import pytest
+
+from agent.transports import _REGISTRY, get_transport, register_transport
+from agent.transports.base import ProviderTransport
+from agent.transports.types import NormalizedResponse
+
+
+class TestProviderTransportABC:
+    def test_cannot_instantiate_abc(self):
+        with pytest.raises(TypeError):
+            ProviderTransport()
+
+    def test_concrete_must_implement_all_abstract(self):
+        class Incomplete(ProviderTransport):
+            @property
+            def api_mode(self):
+                return "test"
+
+        with pytest.raises(TypeError):
+            Incomplete()
+
+    def test_minimal_concrete(self):
+        class Minimal(ProviderTransport):
+            @property
+            def api_mode(self):
+                return "test_minimal"
+
+            def convert_messages(self, messages, **kw):
+                return messages
+
+            def convert_tools(self, tools):
+                return tools
+
+            def build_kwargs(self, model, messages, tools=None, **params):
+                return {"model": model, "messages": messages}
+
+            def normalize_response(self, response, **kw):
+                return NormalizedResponse(content="ok", tool_calls=None, finish_reason="stop")
+
+        t = Minimal()
+        assert t.api_mode == "test_minimal"
+        assert t.validate_response(None) is True
+        assert t.extract_cache_stats(None) is None
+        assert t.map_finish_reason("end_turn") == "end_turn"
+
+
+class TestTransportRegistry:
+    def test_get_unregistered_returns_none(self):
+        assert get_transport("nonexistent_mode") is None
+
+    def test_anthropic_registered_on_import(self):
+        import agent.transports.anthropic  # noqa: F401
+
+        t = get_transport("anthropic_messages")
+        assert t is not None
+        assert t.api_mode == "anthropic_messages"
+
+    def test_register_and_get(self):
+        class DummyTransport(ProviderTransport):
+            @property
+            def api_mode(self):
+                return "dummy_test"
+
+            def convert_messages(self, messages, **kw):
+                return messages
+
+            def convert_tools(self, tools):
+                return tools
+
+            def build_kwargs(self, model, messages, tools=None, **params):
+                return {}
+
+            def normalize_response(self, response, **kw):
+                return NormalizedResponse(content=None, tool_calls=None, finish_reason="stop")
+
+        register_transport("dummy_test", DummyTransport)
+        t = get_transport("dummy_test")
+        assert t.api_mode == "dummy_test"
+        _REGISTRY.pop("dummy_test", None)
+
+
+class TestAnthropicTransport:
+    @pytest.fixture
+    def transport(self):
+        import agent.transports.anthropic  # noqa: F401
+
+        return get_transport("anthropic_messages")
+
+    def test_api_mode(self, transport):
+        assert transport.api_mode == "anthropic_messages"
+
+    def test_convert_tools_simple(self, transport):
+        tools = [{
+            "type": "function",
+            "function": {
+                "name": "test_tool",
+                "description": "A test",
+                "parameters": {"type": "object", "properties": {}},
+            },
+        }]
+        result = transport.convert_tools(tools)
+        assert len(result) == 1
+        assert result[0]["name"] == "test_tool"
+        assert "input_schema" in result[0]
+
+    def test_validate_response_none(self, transport):
+        assert transport.validate_response(None) is False
+
+    def test_validate_response_empty_content(self, transport):
+        r = SimpleNamespace(content=[])
+        assert transport.validate_response(r) is False
+
+    def test_validate_response_valid(self, transport):
+        r = SimpleNamespace(content=[SimpleNamespace(type="text", text="hello")])
+        assert transport.validate_response(r) is True
+
+    def test_map_finish_reason(self, transport):
+        assert transport.map_finish_reason("end_turn") == "stop"
+        assert transport.map_finish_reason("tool_use") == "tool_calls"
+        assert transport.map_finish_reason("max_tokens") == "length"
+        assert transport.map_finish_reason("stop_sequence") == "stop"
+        assert transport.map_finish_reason("refusal") == "content_filter"
+        assert transport.map_finish_reason("model_context_window_exceeded") == "length"
+        assert transport.map_finish_reason("unknown") == "stop"
+
+    def test_extract_cache_stats_none_usage(self, transport):
+        r = SimpleNamespace(usage=None)
+        assert transport.extract_cache_stats(r) is None
+
+    def test_extract_cache_stats_with_cache(self, transport):
+        usage = SimpleNamespace(cache_read_input_tokens=100, cache_creation_input_tokens=50)
+        r = SimpleNamespace(usage=usage)
+        result = transport.extract_cache_stats(r)
+        assert result == {"cached_tokens": 100, "creation_tokens": 50}
+
+    def test_extract_cache_stats_zero(self, transport):
+        usage = SimpleNamespace(cache_read_input_tokens=0, cache_creation_input_tokens=0)
+        r = SimpleNamespace(usage=usage)
+        assert transport.extract_cache_stats(r) is None
+
+    def test_normalize_response_text(self, transport):
+        r = SimpleNamespace(
+            content=[SimpleNamespace(type="text", text="Hello world")],
+            stop_reason="end_turn",
+            usage=SimpleNamespace(input_tokens=10, output_tokens=5),
+            model="claude-sonnet-4-6",
+        )
+        nr = transport.normalize_response(r)
+        assert isinstance(nr, NormalizedResponse)
+        assert nr.content == "Hello world"
+        assert nr.tool_calls is None or nr.tool_calls == []
+        assert nr.finish_reason == "stop"
+
+    def test_normalize_response_tool_calls(self, transport):
+        r = SimpleNamespace(
+            content=[
+                SimpleNamespace(type="tool_use", id="toolu_123", name="terminal", input={"command": "ls"}),
+            ],
+            stop_reason="tool_use",
+            usage=SimpleNamespace(input_tokens=10, output_tokens=20),
+            model="claude-sonnet-4-6",
+        )
+        nr = transport.normalize_response(r)
+        assert nr.finish_reason == "tool_calls"
+        assert len(nr.tool_calls) == 1
+        tc = nr.tool_calls[0]
+        assert tc.name == "terminal"
+        assert tc.id == "toolu_123"
+        assert '"command"' in tc.arguments
+
+    def test_normalize_response_thinking(self, transport):
+        r = SimpleNamespace(
+            content=[
+                SimpleNamespace(type="thinking", thinking="Let me think..."),
+                SimpleNamespace(type="text", text="The answer is 42"),
+            ],
+            stop_reason="end_turn",
+            usage=SimpleNamespace(input_tokens=10, output_tokens=15),
+            model="claude-sonnet-4-6",
+        )
+        nr = transport.normalize_response(r)
+        assert nr.content == "The answer is 42"
+        assert nr.reasoning == "Let me think..."
+
+    def test_build_kwargs_returns_dict(self, transport):
+        messages = [{"role": "user", "content": "Hello"}]
+        kw = transport.build_kwargs(
+            model="claude-sonnet-4-6",
+            messages=messages,
+            max_tokens=1024,
+        )
+        assert isinstance(kw, dict)
+        assert "model" in kw
+        assert "max_tokens" in kw
+        assert "messages" in kw
+
+    def test_convert_messages_extracts_system(self, transport):
+        messages = [
+            {"role": "system", "content": "You are helpful."},
+            {"role": "user", "content": "Hi"},
+        ]
+        system, msgs = transport.convert_messages(messages)
+        assert system is not None
+        assert len(msgs) >= 1

tests/agent/transports/test_types.py

@@ -0,0 +1,130 @@
+"""Tests for agent/transports/types.py — dataclass construction + helpers."""
+
+import json
+
+from agent.transports.types import (
+    NormalizedResponse,
+    ToolCall,
+    Usage,
+    build_tool_call,
+    map_finish_reason,
+)
+
+
+class TestToolCall:
+    def test_basic_construction(self):
+        tc = ToolCall(id="call_abc", name="terminal", arguments='{"cmd": "ls"}')
+        assert tc.id == "call_abc"
+        assert tc.name == "terminal"
+        assert tc.arguments == '{"cmd": "ls"}'
+        assert tc.provider_data is None
+
+    def test_none_id(self):
+        tc = ToolCall(id=None, name="read_file", arguments="{}")
+        assert tc.id is None
+
+    def test_provider_data(self):
+        tc = ToolCall(
+            id="call_x",
+            name="t",
+            arguments="{}",
+            provider_data={"call_id": "call_x", "response_item_id": "fc_x"},
+        )
+        assert tc.provider_data["call_id"] == "call_x"
+        assert tc.provider_data["response_item_id"] == "fc_x"
+
+
+class TestUsage:
+    def test_defaults(self):
+        u = Usage()
+        assert u.prompt_tokens == 0
+        assert u.completion_tokens == 0
+        assert u.total_tokens == 0
+        assert u.cached_tokens == 0
+
+    def test_explicit(self):
+        u = Usage(prompt_tokens=100, completion_tokens=50, total_tokens=150, cached_tokens=80)
+        assert u.total_tokens == 150
+
+
+class TestNormalizedResponse:
+    def test_text_only(self):
+        r = NormalizedResponse(content="hello", tool_calls=None, finish_reason="stop")
+        assert r.content == "hello"
+        assert r.tool_calls is None
+        assert r.finish_reason == "stop"
+        assert r.reasoning is None
+        assert r.usage is None
+        assert r.provider_data is None
+
+    def test_with_tool_calls(self):
+        tcs = [ToolCall(id="call_1", name="terminal", arguments='{"cmd":"pwd"}')]
+        r = NormalizedResponse(content=None, tool_calls=tcs, finish_reason="tool_calls")
+        assert r.finish_reason == "tool_calls"
+        assert len(r.tool_calls) == 1
+        assert r.tool_calls[0].name == "terminal"
+
+    def test_with_reasoning(self):
+        r = NormalizedResponse(
+            content="answer",
+            tool_calls=None,
+            finish_reason="stop",
+            reasoning="I thought about it",
+        )
+        assert r.reasoning == "I thought about it"
+
+    def test_with_provider_data(self):
+        r = NormalizedResponse(
+            content=None,
+            tool_calls=None,
+            finish_reason="stop",
+            provider_data={"reasoning_details": [{"type": "thinking", "thinking": "hmm"}]},
+        )
+        assert r.provider_data["reasoning_details"][0]["type"] == "thinking"
+
+
+class TestBuildToolCall:
+    def test_dict_arguments_serialized(self):
+        tc = build_tool_call(id="call_1", name="terminal", arguments={"cmd": "ls"})
+        assert tc.arguments == json.dumps({"cmd": "ls"})
+        assert tc.provider_data is None
+
+    def test_string_arguments_passthrough(self):
+        tc = build_tool_call(id="call_2", name="read_file", arguments='{"path": "/tmp"}')
+        assert tc.arguments == '{"path": "/tmp"}'
+
+    def test_provider_fields(self):
+        tc = build_tool_call(
+            id="call_3",
+            name="terminal",
+            arguments="{}",
+            call_id="call_3",
+            response_item_id="fc_3",
+        )
+        assert tc.provider_data == {"call_id": "call_3", "response_item_id": "fc_3"}
+
+    def test_none_id(self):
+        tc = build_tool_call(id=None, name="t", arguments="{}")
+        assert tc.id is None
+
+
+class TestMapFinishReason:
+    ANTHROPIC_MAP = {
+        "end_turn": "stop",
+        "tool_use": "tool_calls",
+        "max_tokens": "length",
+        "stop_sequence": "stop",
+        "refusal": "content_filter",
+    }
+
+    def test_known_reason(self):
+        assert map_finish_reason("end_turn", self.ANTHROPIC_MAP) == "stop"
+        assert map_finish_reason("tool_use", self.ANTHROPIC_MAP) == "tool_calls"
+        assert map_finish_reason("max_tokens", self.ANTHROPIC_MAP) == "length"
+        assert map_finish_reason("refusal", self.ANTHROPIC_MAP) == "content_filter"
+
+    def test_unknown_reason_defaults_to_stop(self):
+        assert map_finish_reason("something_new", self.ANTHROPIC_MAP) == "stop"
+
+    def test_none_reason(self):
+        assert map_finish_reason(None, self.ANTHROPIC_MAP) == "stop"

tests/run_agent/test_run_agent.py

@@ -1037,138 +1037,6 @@ class TestBuildAssistantMessage:
         result = agent._build_assistant_message(msg, "tool_calls")
         assert "extra_content" not in result["tool_calls"][0]
 
-    def test_tool_call_arguments_normalized_from_gemma4_whitespace(self, agent):
-        tc = _mock_tool_call(
-            name="read_file",
-            arguments='  \n  {"path": "README.md"}  \n  ',
-            call_id="c4",
-        )
-        msg = _mock_assistant_msg(content="", tool_calls=[tc])
-        result = agent._build_assistant_message(msg, "tool_calls")
-        assert result["tool_calls"][0]["function"]["arguments"] == '{"path":"README.md"}'
-
-    def test_tool_call_arguments_normalized_from_single_quotes_and_trailing_comma(self, agent):
-        tc = _mock_tool_call(
-            name="read_file",
-            arguments="{'path': 'README.md',}",
-            call_id="c5",
-        )
-        msg = _mock_assistant_msg(content="", tool_calls=[tc])
-        result = agent._build_assistant_message(msg, "tool_calls")
-        assert result["tool_calls"][0]["function"]["arguments"] == '{"path":"README.md"}'
-
-
-class TestNormalizeToolCallArguments:
-    @pytest.mark.parametrize(
-        ("raw_args", "expected"),
-        [
-            ('{"q":"test"}', '{"q":"test"}'),
-            ('  \n  {"q": "test"}  \n  ', '{"q":"test"}'),
-            ('{"q": "test",}', '{"q":"test"}'),
-            ("{'q': 'test'}", '{"q":"test"}'),
-            ("{'path': 'README.md', 'mode': 'read'}", '{"path":"README.md","mode":"read"}'),
-            ('"path": "README.md"', '{"path":"README.md"}'),
-            ('path: "README.md"', '{"path":"README.md"}'),
-            ('path: "README.md", mode: "read"', '{"path":"README.md","mode":"read"}'),
-            ({"path": "README.md"}, '{"path":"README.md"}'),
-            (["README.md", "docs.md"], '["README.md","docs.md"]'),
-            ('\t\n  ', '{}'),
-            ('{"nested": {"path": "README.md"}}', '{"nested":{"path":"README.md"}}'),
-        ],
-    )
-    def test_complete_args_are_normalized(self, raw_args, expected):
-        normalized, is_complete = AIAgent._normalize_tool_call_arguments(raw_args)
-        assert is_complete is True
-        assert normalized == expected
-
-    @pytest.mark.parametrize(
-        "raw_args",
-        [
-            '{"path": "README.md"',
-            '{"a": 1, "b"',
-            '{"path": [1, 2}',
-            "{'path': 'README.md'",
-            'path: "README.md", mode:',
-            '{"command": "echo hello",',
-        ],
-    )
-    def test_incomplete_args_are_not_marked_complete(self, raw_args):
-        normalized, is_complete = AIAgent._normalize_tool_call_arguments(raw_args)
-        assert is_complete is False
-        assert isinstance(normalized, str)
-        assert normalized == raw_args.strip()
-
-
-class TestSanitizeApiMessages:
-    def test_merges_consecutive_assistant_tool_call_messages(self):
-        messages = [
-            {
-                "role": "assistant",
-                "content": "first",
-                "tool_calls": [{"id": "c1", "type": "function", "function": {"name": "read_file", "arguments": '{"path":"a.py"}'}}],
-            },
-            {
-                "role": "assistant",
-                "content": "second",
-                "tool_calls": [{"id": "c2", "type": "function", "function": {"name": "search_files", "arguments": '{"pattern":"TODO"}'}}],
-            },
-            {"role": "tool", "tool_call_id": "c1", "content": "a.py"},
-            {"role": "tool", "tool_call_id": "c2", "content": "matches"},
-        ]
-
-        sanitized = AIAgent._sanitize_api_messages(messages)
-
-        assert len(sanitized) == 3
-        assert sanitized[0]["role"] == "assistant"
-        assert [tc["id"] for tc in sanitized[0]["tool_calls"]] == ["c1", "c2"]
-        assert sanitized[0]["content"] == "first\nsecond"
-
-    def test_does_not_merge_assistant_tool_call_messages_across_non_assistant_boundary(self):
-        messages = [
-            {
-                "role": "assistant",
-                "content": "",
-                "tool_calls": [{"id": "c1", "type": "function", "function": {"name": "read_file", "arguments": '{"path":"a.py"}'}}],
-            },
-            {"role": "tool", "tool_call_id": "c1", "content": "a.py"},
-            {
-                "role": "assistant",
-                "content": "",
-                "tool_calls": [{"id": "c2", "type": "function", "function": {"name": "read_file", "arguments": '{"path":"b.py"}'}}],
-            },
-            {"role": "tool", "tool_call_id": "c2", "content": "b.py"},
-        ]
-
-        sanitized = AIAgent._sanitize_api_messages(messages)
-
-        assistant_msgs = [m for m in sanitized if m.get("role") == "assistant"]
-        assert len(assistant_msgs) == 2
-        assert assistant_msgs[0]["tool_calls"][0]["id"] == "c1"
-        assert assistant_msgs[1]["tool_calls"][0]["id"] == "c2"
-
-    def test_merge_preserves_tool_call_order(self):
-        messages = [
-            {
-                "role": "assistant",
-                "content": "",
-                "tool_calls": [{"id": "c1", "type": "function", "function": {"name": "read_file", "arguments": '{"path":"a.py"}'}}],
-            },
-            {
-                "role": "assistant",
-                "content": "",
-                "tool_calls": [{"id": "c2", "type": "function", "function": {"name": "read_file", "arguments": '{"path":"b.py"}'}}],
-            },
-            {
-                "role": "assistant",
-                "content": "",
-                "tool_calls": [{"id": "c3", "type": "function", "function": {"name": "read_file", "arguments": '{"path":"c.py"}'}}],
-            },
-        ]
-
-        sanitized = AIAgent._sanitize_api_messages(messages)
-
-        assert [tc["id"] for tc in sanitized[0]["tool_calls"]] == ["c1", "c2", "c3"]
-
 
 class TestFormatToolsForSystemMessage:
     def test_no_tools_returns_empty_array(self, agent):
@@ -3599,59 +3467,6 @@ class TestStreamingApiCall:
         assert tc[0].function.arguments == '{"path":"x.txt","content":"hel'
         assert resp.choices[0].finish_reason == "length"
 
-    @pytest.mark.parametrize(
-        ("raw_arguments", "expected"),
-        [
-            ('  \n  {"path": "x.txt"}  \n  ', '{"path":"x.txt"}'),
-            ("{'path': 'x.txt',}", '{"path":"x.txt"}'),
-            ('path: "x.txt", mode: "read"', '{"path":"x.txt","mode":"read"}'),
-        ],
-    )
-    def test_repairable_tool_call_args_do_not_upgrade_finish_reason_to_length(self, agent, raw_arguments, expected):
-        chunks = [
-            _make_chunk(tool_calls=[_make_tc_delta(0, "call_1", "read_file", raw_arguments)]),
-            _make_chunk(finish_reason="tool_calls"),
-        ]
-        agent.client.chat.completions.create.return_value = iter(chunks)
-
-        resp = agent._interruptible_streaming_api_call({"messages": []})
-
-        tc = resp.choices[0].message.tool_calls
-        assert len(tc) == 1
-        assert tc[0].function.name == "read_file"
-        assert tc[0].function.arguments == expected
-        assert resp.choices[0].finish_reason == "tool_calls"
-
-    def test_streamed_tool_call_args_single_quotes_across_chunks_normalized(self, agent):
-        chunks = [
-            _make_chunk(tool_calls=[_make_tc_delta(0, "call_1", "read_file", "{'path':")]),
-            _make_chunk(tool_calls=[_make_tc_delta(0, None, None, " 'x.txt',}")]),
-            _make_chunk(finish_reason="tool_calls"),
-        ]
-        agent.client.chat.completions.create.return_value = iter(chunks)
-
-        resp = agent._interruptible_streaming_api_call({"messages": []})
-
-        tc = resp.choices[0].message.tool_calls
-        assert len(tc) == 1
-        assert tc[0].function.arguments == '{"path":"x.txt"}'
-        assert resp.choices[0].finish_reason == "tool_calls"
-
-    def test_streamed_split_json_chunks_still_reassemble(self, agent):
-        chunks = [
-            _make_chunk(tool_calls=[_make_tc_delta(0, "call_1", "read_file", '{"path":')]),
-            _make_chunk(tool_calls=[_make_tc_delta(0, None, None, ' "x.txt"}')]),
-            _make_chunk(finish_reason="tool_calls"),
-        ]
-        agent.client.chat.completions.create.return_value = iter(chunks)
-
-        resp = agent._interruptible_streaming_api_call({"messages": []})
-
-        tc = resp.choices[0].message.tool_calls
-        assert len(tc) == 1
-        assert tc[0].function.arguments == '{"path":"x.txt"}'
-        assert resp.choices[0].finish_reason == "tool_calls"
-
     def test_ollama_reused_index_separate_tool_calls(self, agent):
         """Ollama sends every tool call at index 0 with different ids.