Skip to content

armgit / ECC

Public Imported from GitHub
0

The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.

fix(llm): align Claude provider with current Anthropic API (#2133)

quadcent8973d0f6c53c
diff --git a/src/llm/cli/selector.py b/src/llm/cli/selector.py
index 4419b2bc6..87b513e21 100644
--- a/src/llm/cli/selector.py
+++ b/src/llm/cli/selector.py
@@ -105,9 +105,9 @@ def interactive_select(
if models_per_provider is None:
models_per_provider = {
"claude": [
- ("claude-opus-4-5", "Claude Opus 4.5 - Most capable"),
- ("claude-sonnet-4-7", "Claude Sonnet 4.7 - Balanced"),
- ("claude-haiku-4-7", "Claude Haiku 4.7 - Fast"),
+ ("claude-opus-4-8", "Claude Opus 4.8 - Most capable"),
+ ("claude-sonnet-4-6", "Claude Sonnet 4.6 - Balanced"),
+ ("claude-haiku-4-5", "Claude Haiku 4.5 - Fast"),
],
"openai": [
("gpt-4o", "GPT-4o - Most capable"),
diff --git a/src/llm/providers/claude.py b/src/llm/providers/claude.py
index 2383db4e2..55cce8951 100644
--- a/src/llm/providers/claude.py
+++ b/src/llm/providers/claude.py
@@ -13,7 +13,14 @@ from llm.core.interface import (
LLMProvider,
RateLimitError,
)
-from llm.core.types import LLMInput, LLMOutput, Message, ModelInfo, ProviderType, ToolCall
+from llm.core.types import LLMInput, LLMOutput, ModelInfo, ProviderType, Role, ToolCall
+
+_DEFAULT_MODEL = "claude-sonnet-4-6"
+_OPUS_ADAPTIVE_ONLY_PREFIXES = ("claude-opus-4-7", "claude-opus-4-8")
+
+
+def _uses_adaptive_thinking_only(model: str) -> bool:
+ return any(model.startswith(prefix) for prefix in _OPUS_ADAPTIVE_ONLY_PREFIXES)
class ClaudeProvider(LLMProvider):
@@ -23,77 +30,90 @@ class ClaudeProvider(LLMProvider):
self.client = Anthropic(api_key=api_key or os.environ.get("ANTHROPIC_API_KEY"), base_url=base_url)
self._models = [
ModelInfo(
- name="claude-opus-4-5",
+ name="claude-opus-4-8",
provider=ProviderType.CLAUDE,
supports_tools=True,
supports_vision=True,
- max_tokens=8192,
- context_window=200000,
+ max_tokens=64000,
+ context_window=1_000_000,
),
ModelInfo(
- name="claude-sonnet-4-7",
+ name="claude-sonnet-4-6",
provider=ProviderType.CLAUDE,
supports_tools=True,
supports_vision=True,
- max_tokens=8192,
- context_window=200000,
+ max_tokens=64000,
+ context_window=1_000_000,
),
ModelInfo(
- name="claude-haiku-4-7",
+ name="claude-haiku-4-5",
provider=ProviderType.CLAUDE,
supports_tools=True,
- supports_vision=False,
- max_tokens=4096,
- context_window=200000,
+ supports_vision=True,
+ max_tokens=16000,
+ context_window=200_000,
),
]
def generate(self, input: LLMInput) -> LLMOutput:
try:
+ model = input.model or _DEFAULT_MODEL
+ system_parts = [msg.content for msg in input.messages if msg.role == Role.SYSTEM]
+ api_messages = [
+ msg.to_dict() for msg in input.messages if msg.role not in (Role.SYSTEM,)
+ ]
+
params: dict[str, Any] = {
- "model": input.model or "claude-sonnet-4-7",
- "messages": [msg.to_dict() for msg in input.messages],
- "temperature": input.temperature,
+ "model": model,
+ "messages": api_messages,
+ "max_tokens": input.max_tokens if input.max_tokens else 16000,
+ "cache_control": {"type": "ephemeral"},
}
- if input.max_tokens:
- params["max_tokens"] = input.max_tokens
- else:
- params["max_tokens"] = 8192 # required by Anthropic API
- if input.tools:
- params["tools"] = [tool.to_anthropic_tool() for tool in input.tools]
+ if system_parts:
+ params["system"] = "\n\n".join(system_parts)
+ if input.tools:
+ params["tools"] = [tool.to_anthropic_tool() for tool in input.tools]
+ if not _uses_adaptive_thinking_only(model):
+ params["temperature"] = input.temperature
+ if _uses_adaptive_thinking_only(model):
+ params["thinking"] = {"type": "adaptive"}
response = self.client.messages.create(**params)
- text_parts: list[str] = []
- tool_calls: list[ToolCall] = []
- for block in response.content or []:
- block_type = getattr(block, "type", None)
- if block_type == "text":
- text = getattr(block, "text", "")
- if text:
- text_parts.append(text)
- elif block_type == "tool_use":
- raw_arguments = getattr(block, "input", {})
- arguments = (
- raw_arguments.copy()
- if isinstance(raw_arguments, dict)
- else getattr(raw_arguments, "__dict__", {}).copy()
- )
- tool_calls.append(
- ToolCall(
- id=getattr(block, "id", ""),
- name=getattr(block, "name", ""),
- arguments=arguments,
- )
- )
-
- return LLMOutput(
- content="".join(text_parts),
- tool_calls=tool_calls or None,
+ text_parts: list[str] = []
+ tool_calls: list[ToolCall] = []
+ for block in response.content or []:
+ block_type = getattr(block, "type", None)
+ if block_type == "text":
+ text = getattr(block, "text", "")
+ if text:
+ text_parts.append(text)
+ elif block_type == "tool_use":
+ raw_arguments = getattr(block, "input", {})
+ arguments = (
+ raw_arguments.copy()
+ if isinstance(raw_arguments, dict)
+ else getattr(raw_arguments, "__dict__", {}).copy()
+ )
+ tool_calls.append(
+ ToolCall(
+ id=getattr(block, "id", ""),
+ name=getattr(block, "name", ""),
+ arguments=arguments,
+ )
+ )
+
+ return LLMOutput(
+ content="".join(text_parts),
+ tool_calls=tool_calls or None,
model=response.model,
usage={
"input_tokens": response.usage.input_tokens,
"output_tokens": response.usage.output_tokens,
+ "cache_creation_input_tokens": getattr(
+ response.usage, "cache_creation_input_tokens", 0
+ ),
+ "cache_read_input_tokens": getattr(response.usage, "cache_read_input_tokens", 0),
},
stop_reason=response.stop_reason,
)
@@ -114,4 +134,4 @@ class ClaudeProvider(LLMProvider):
return bool(self.client.api_key)
def get_default_model(self) -> str:
- return "claude-sonnet-4-7"
+ return _DEFAULT_MODEL