klaude-code 2.6.0__py3-none-any.whl → 2.8.0__py3-none-any.whl

klaude_code/config/assets/builtin_config.yaml

@@ -6,12 +6,14 @@ provider_list:
   protocol: anthropic
   api_key: ${ANTHROPIC_API_KEY}
   model_list:
+
   - model_name: sonnet
     model_id: claude-sonnet-4-5-20250929
     context_limit: 200000
     provider_routing:
       sort: throughput
     cost: {input: 3, output: 15, cache_read: 0.3, cache_write: 3.75}
+
   - model_name: opus
     model_id: claude-opus-4-5-20251101
     context_limit: 200000
@@ -20,10 +22,13 @@ provider_list:
       type: enabled
       budget_tokens: 2048
     cost: {input: 5, output: 25, cache_read: 0.5, cache_write: 6.25}
+
+
 - provider_name: openai
   protocol: responses
   api_key: ${OPENAI_API_KEY}
   model_list:
+
   - model_name: gpt-5.2-high
     model_id: gpt-5.2
     max_tokens: 128000
@@ -33,6 +38,7 @@ provider_list:
       reasoning_effort: high
       reasoning_summary: detailed
     cost: {input: 1.75, output: 14, cache_read: 0.17}
+
   - model_name: gpt-5.2-medium
     model_id: gpt-5.2
     context_limit: 400000
@@ -41,6 +47,7 @@ provider_list:
       reasoning_effort: medium
       reasoning_summary: concise
     cost: {input: 1.75, output: 14, cache_read: 0.17}
+
   - model_name: gpt-5.2-low
     model_id: gpt-5.2
     context_limit: 400000
@@ -49,6 +56,7 @@ provider_list:
       reasoning_effort: low
       reasoning_summary: concise
     cost: {input: 1.75, output: 14, cache_read: 0.17}
+
   - model_name: gpt-5.2-fast
     model_id: gpt-5.2
     context_limit: 400000
@@ -56,6 +64,7 @@ provider_list:
     thinking:
       reasoning_effort: none
     cost: {input: 1.75, output: 14, cache_read: 0.17}
+
   - model_name: gpt-5.1-codex-max
     model_id: gpt-5.1-codex-max
     max_tokens: 128000
@@ -64,10 +73,13 @@ provider_list:
       reasoning_effort: medium
       reasoning_summary: detailed
     cost: {input: 1.25, output: 10, cache_read: 0.13}
+
+
 - provider_name: openrouter
   protocol: openrouter
   api_key: ${OPENROUTER_API_KEY}
   model_list:
+
   - model_name: gpt-5.2-high
     model_id: openai/gpt-5.2
     max_tokens: 128000
@@ -77,6 +89,7 @@ provider_list:
       reasoning_effort: high
       reasoning_summary: detailed
     cost: {input: 1.75, output: 14, cache_read: 0.17}
+
   - model_name: gpt-5.2-medium
     model_id: openai/gpt-5.2
     max_tokens: 128000
@@ -86,6 +99,7 @@ provider_list:
       reasoning_effort: medium
       reasoning_summary: concise
     cost: {input: 1.75, output: 14, cache_read: 0.17}
+
   - model_name: kimi
     model_id: moonshotai/kimi-k2-thinking
     context_limit: 262144
@@ -93,16 +107,19 @@ provider_list:
       only:
       - moonshotai/turbo
     cost: {input: 0.6, output: 2.5, cache_read: 0.15}
+
   - model_name: haiku
     model_id: anthropic/claude-haiku-4.5
     context_limit: 200000
     cost: {input: 1, output: 5, cache_read: 0.1, cache_write: 1.25}
+
   - model_name: sonnet
     model_id: anthropic/claude-4.5-sonnet
     context_limit: 200000
     provider_routing:
       sort: throughput
     cost: {input: 3, output: 15, cache_read: 0.3, cache_write: 3.75}
+
   - model_name: opus
     model_id: anthropic/claude-4.5-opus
     context_limit: 200000
@@ -110,18 +127,21 @@ provider_list:
       type: enabled
       budget_tokens: 2048
     cost: {input: 5, output: 25, cache_read: 0.5, cache_write: 6.25}
+
   - model_name: gemini-pro
     model_id: google/gemini-3-pro-preview
     context_limit: 1048576
     thinking:
       reasoning_effort: high
     cost: {input: 2, output: 12, cache_read: 0.2}
+
   - model_name: gemini-flash
     model_id: google/gemini-3-flash-preview
     context_limit: 1048576
     thinking:
       reasoning_effort: medium
     cost: {input: 0.5, output: 3, cache_read: 0.05}
+
   - model_name: nano-banana-pro
     model_id: google/gemini-3-pro-image-preview
     context_limit: 66000
@@ -129,6 +149,7 @@ provider_list:
     - image
     - text
     cost: {input: 2, output: 12, cache_read: 0.2, image: 120}
+
   - model_name: nano-banana
     model_id: google/gemini-2.5-flash-image
     context_limit: 33000
@@ -136,6 +157,7 @@ provider_list:
     - image
     - text
     cost: {input: 0.3, output: 2.5, cache_read: 0.03, image: 30}
+
   - model_name: grok
     model_id: x-ai/grok-4.1-fast
     context_limit: 2000000
@@ -143,10 +165,12 @@ provider_list:
       type: enabled
       budget_tokens: 2048
     cost: {input: 0.2, output: 0.5, cache_read: 0.05}
+
   - model_name: minimax
     model_id: minimax/minimax-m2.1
     context_limit: 204800
     cost: {input: 0.3, output: 1.2, cache_read: 0.03}
+
   - model_name: glm
     model_id: z-ai/glm-4.7
     context_limit: 200000
@@ -154,6 +178,7 @@ provider_list:
       only:
       - z-ai
     cost: {input: 0.44, output: 1.74, cache_read: 0.04}
+
   - model_name: seedream
     model_id: bytedance-seed/seedream-4.5
     context_limit: 4000
@@ -161,6 +186,7 @@ provider_list:
     modalities:
     - image
     - text
+
   - model_name: flux
     model_id: black-forest-labs/flux.2-max
     context_limit: 47000
@@ -168,22 +194,27 @@ provider_list:
     modalities:
     - image
     - text
+
+
 - provider_name: google
   protocol: google
   api_key: ${GOOGLE_API_KEY}
   model_list:
+
   - model_name: gemini-pro
     model_id: gemini-3-pro-preview
     context_limit: 1048576
     thinking:
       reasoning_effort: high
     cost: {input: 2, output: 12, cache_read: 0.2}
+
   - model_name: gemini-flash
     model_id: gemini-3-flash-preview
     context_limit: 1048576
     thinking:
       reasoning_effort: medium
     cost: {input: 0.5, output: 3, cache_read: 0.05}
+
   - model_name: nano-banana-pro
     model_id: gemini-3-pro-image-preview
     context_limit: 66000
@@ -191,6 +222,7 @@ provider_list:
     - image
     - text
     cost: {input: 2, output: 12, cache_read: 0.2, image: 120}
+
   - model_name: nano-banana
     model_id: gemini-2.5-flash-image
     context_limit: 33000
@@ -198,21 +230,27 @@ provider_list:
     - image
     - text
     cost: {input: 0.3, output: 2.5, cache_read: 0.03, image: 30}
+
+
 - provider_name: bedrock
   protocol: bedrock
   aws_access_key: ${AWS_ACCESS_KEY_ID}
   aws_secret_key: ${AWS_SECRET_ACCESS_KEY}
   aws_region: ${AWS_REGION}
   model_list:
+
   - model_name: sonnet
     model_id: us.anthropic.claude-sonnet-4-5-20250929-v1:0
     context_limit: 200000
     cost: {input: 3, output: 15, cache_read: 0.3, cache_write: 3.75}
+
+
 - provider_name: deepseek
   protocol: anthropic
   api_key: ${DEEPSEEK_API_KEY}
   base_url: https://api.deepseek.com/anthropic
   model_list:
+
   - model_name: deepseek
     model_id: deepseek-reasoner
     context_limit: 128000
@@ -220,11 +258,14 @@ provider_list:
       type: enabled
       budget_tokens: 2048
     cost: {input: 2, output: 3, cache_read: 0.2, currency: CNY}
+
+
 - provider_name: moonshot
   protocol: anthropic
   api_key: ${MOONSHOT_API_KEY}
   base_url: https://api.moonshot.cn/anthropic
   model_list:
+
   - model_name: kimi
     model_id: kimi-k2-thinking
     context_limit: 262144
@@ -232,13 +273,17 @@ provider_list:
       type: enabled
       budget_tokens: 8192
     cost: {input: 4, output: 16, cache_read: 1, currency: CNY}
+
+
 - provider_name: claude-max
   protocol: claude_oauth
   model_list:
+
   - model_name: sonnet
     model_id: claude-sonnet-4-5-20250929
     context_limit: 200000
     cost: {input: 3, output: 15, cache_read: 0.3, cache_write: 3.75}
+
   - model_name: opus
     model_id: claude-opus-4-5-20251101
     context_limit: 200000
@@ -246,17 +291,80 @@ provider_list:
       type: enabled
       budget_tokens: 2048
     cost: {input: 5, output: 25, cache_read: 0.5, cache_write: 6.25}
+
   - model_name: haiku
     model_id: claude-haiku-4-5-20251001
     context_limit: 200000
     cost: {input: 1, output: 5, cache_read: 0.1, cache_write: 1.25}
+
+
 - provider_name: codex
   protocol: codex_oauth
   model_list:
+
   - model_name: gpt-5.2-codex
     model_id: gpt-5.2-codex
     thinking:
       reasoning_effort: medium
+      reasoning_summary: auto
     context_limit: 400000
     max_tokens: 128000
     cost: {input: 1.75, output: 14, cache_read: 0.17}
+
+  - model_name: gpt-5.2-high
+    model_id: gpt-5.2
+    max_tokens: 128000
+    context_limit: 400000
+    verbosity: high
+    thinking:
+      reasoning_effort: high
+      reasoning_summary: detailed
+    cost: {input: 1.75, output: 14, cache_read: 0.17}
+
+
+- provider_name: antigravity
+  protocol: antigravity
+  model_list:
+  - model_name: opus
+    model_id: claude-opus-4-5-thinking
+    context_limit: 200000
+    max_tokens: 64000
+    thinking:
+      type: enabled
+      budget_tokens: 10240
+  - model_name: sonnet
+    model_id: claude-sonnet-4-5
+    context_limit: 200000
+    max_tokens: 64000
+  - model_name: sonnet-thinking
+    model_id: claude-sonnet-4-5-thinking
+    context_limit: 200000
+    max_tokens: 64000
+    thinking:
+      type: enabled
+      budget_tokens: 10240
+  - model_name: gemini-pro-high
+    model_id: gemini-3-pro-high
+    context_limit: 1048576
+    max_tokens: 65535
+    thinking:
+      reasoning_effort: high
+  - model_name: gemini-pro-low
+    model_id: gemini-3-pro-low
+    context_limit: 1048576
+    max_tokens: 65535
+    thinking:
+      reasoning_effort: low
+  - model_name: gemini-flash
+    model_id: gemini-3-flash
+    context_limit: 1048576
+    max_tokens: 65535
+    thinking:
+      reasoning_effort: medium
+  - model_name: gpt-oss
+    model_id: gpt-oss-120b-medium
+    context_limit: 131072
+    max_tokens: 32768
+
+
+compact_model: gemini-flash

klaude_code/config/builtin_config.py

@@ -13,7 +13,7 @@ from typing import TYPE_CHECKING, Any
 import yaml
 
 if TYPE_CHECKING:
-    from klaude_code.config.config import ProviderConfig
+    from klaude_code.config.config import Config
 
 
 @dataclass(frozen=True)
@@ -48,16 +48,10 @@ def _load_builtin_yaml() -> dict[str, Any]:
     return data
 
 
-def get_builtin_provider_configs() -> list["ProviderConfig"]:
-    """Load built-in provider configurations from YAML asset."""
+def get_builtin_config() -> "Config":
+    """Load built-in configuration from YAML asset."""
     # Import here to avoid circular import
-    from klaude_code.config.config import ProviderConfig
+    from klaude_code.config.config import Config
 
     data = _load_builtin_yaml()
-    return [ProviderConfig.model_validate(p) for p in data.get("provider_list", [])]
-
-
-def get_builtin_sub_agent_models() -> dict[str, str]:
-    """Load built-in sub agent model mappings from YAML asset."""
-    data = _load_builtin_yaml()
-    return data.get("sub_agent_models", {})
+    return Config.model_validate(data)

klaude_code/config/config.py

@@ -11,8 +11,7 @@ from pydantic import BaseModel, Field, ValidationError, model_validator
 from klaude_code.auth.env import get_auth_env
 from klaude_code.config.builtin_config import (
     SUPPORTED_API_KEYS,
-    get_builtin_provider_configs,
-    get_builtin_sub_agent_models,
+    get_builtin_config,
 )
 from klaude_code.log import log
 from klaude_code.protocol import llm_param
@@ -108,6 +107,15 @@ class ProviderConfig(llm_param.LLMConfigProviderParameter):
             # Consider available if logged in. Token refresh happens on-demand.
             return state is None
 
+        if self.protocol == LLMClientProtocol.ANTIGRAVITY:
+            # Antigravity uses OAuth authentication, not API key
+            from klaude_code.auth.antigravity.token_manager import AntigravityTokenManager
+
+            token_manager = AntigravityTokenManager()
+            state = token_manager.get_state()
+            # Consider available if logged in. Token refresh happens on-demand.
+            return state is None
+
         if self.protocol == LLMClientProtocol.BEDROCK:
             # Bedrock uses AWS credentials, not API key. Region is always required.
             _, resolved_profile = parse_env_var_syntax(self.aws_profile)
@@ -161,6 +169,7 @@ class UserConfig(BaseModel):
     """User configuration (what gets saved to disk)."""
 
     main_model: str | None = None
+    compact_model: str | None = None
     sub_agent_models: dict[str, str] = Field(default_factory=dict)
     theme: str | None = None
     provider_list: list[UserProviderConfig] = Field(default_factory=lambda: [])
@@ -183,6 +192,7 @@ class Config(BaseModel):
     """Merged configuration (builtin + user) for runtime use."""
 
     main_model: str | None = None
+    compact_model: str | None = None
     sub_agent_models: dict[str, str] = Field(default_factory=dict)
     theme: str | None = None
     provider_list: list[ProviderConfig] = Field(default_factory=lambda: [])
@@ -286,6 +296,7 @@ class Config(BaseModel):
                 not in {
                     llm_param.LLMClientProtocol.CODEX_OAUTH,
                     llm_param.LLMClientProtocol.CLAUDE_OAUTH,
+                    llm_param.LLMClientProtocol.ANTIGRAVITY,
                     llm_param.LLMClientProtocol.BEDROCK,
                 }
                 and not api_key
@@ -313,6 +324,7 @@ class Config(BaseModel):
                 not in {
                     llm_param.LLMClientProtocol.CODEX_OAUTH,
                     llm_param.LLMClientProtocol.CLAUDE_OAUTH,
+                    llm_param.LLMClientProtocol.ANTIGRAVITY,
                     llm_param.LLMClientProtocol.BEDROCK,
                 }
                 and not api_key
@@ -389,6 +401,7 @@ class Config(BaseModel):
 
         # Sync user-modifiable fields from merged config to user config
         user_config.main_model = self.main_model
+        user_config.compact_model = self.compact_model
         user_config.sub_agent_models = self.sub_agent_models
         user_config.theme = self.theme
         # Note: provider_list is NOT synced - user providers are already in user_config
@@ -407,6 +420,7 @@ def get_example_config() -> UserConfig:
     """Generate example config for user reference (will be commented out)."""
     return UserConfig(
         main_model="opus",
+        compact_model="gemini-flash",
         sub_agent_models={"explore": "haiku", "webagent": "sonnet", "task": "sonnet"},
         provider_list=[
             UserProviderConfig(
@@ -434,11 +448,7 @@ def get_example_config() -> UserConfig:
 
 def _get_builtin_config() -> Config:
     """Load built-in provider configurations."""
-    # Re-validate to ensure compatibility with current ProviderConfig class
-    # (needed for tests that may monkeypatch the class)
-    providers = [ProviderConfig.model_validate(p.model_dump()) for p in get_builtin_provider_configs()]
-    sub_agent_models = get_builtin_sub_agent_models()
-    return Config(provider_list=providers, sub_agent_models=sub_agent_models)
+    return get_builtin_config()
 
 
 def _merge_model(builtin: ModelConfig, user: ModelConfig) -> ModelConfig:
@@ -485,7 +495,7 @@ def _merge_provider(builtin: ProviderConfig, user: UserProviderConfig) -> Provid
         if value is not None:
             merged_data[key] = value
 
-    merged_data["model_list"] = list(merged_models.values())
+    merged_data["model_list"] = [m.model_dump() for m in merged_models.values()]
     return ProviderConfig.model_validate(merged_data)
 
 
@@ -530,11 +540,14 @@ def _merge_configs(user_config: UserConfig | None, builtin_config: Config) -> Co
     # Merge sub_agent_models
     merged_sub_agent_models = {**builtin_config.sub_agent_models, **user_config.sub_agent_models}
 
+    # Re-validate providers to ensure compatibility (tests may monkeypatch the class)
+    revalidated_providers = [ProviderConfig.model_validate(p.model_dump()) for p in merged_providers.values()]
     merged = Config(
         main_model=user_config.main_model or builtin_config.main_model,
+        compact_model=user_config.compact_model or builtin_config.compact_model,
         sub_agent_models=merged_sub_agent_models,
         theme=user_config.theme or builtin_config.theme,
-        provider_list=list(merged_providers.values()),
+        provider_list=revalidated_providers,
     )
     # Keep reference to user config for saving
     merged.set_user_config(user_config)
@@ -555,7 +568,7 @@ def _load_user_config() -> UserConfig | None:
     try:
         return UserConfig.model_validate(config_dict)
     except ValidationError as e:
-        log(f"Invalid config file: {config_path}", style="red bold")
+        log(f"Invalid config file: {config_path}", style="red")
         log(str(e), style="red")
         raise ValueError(f"Invalid config file: {config_path}") from e
 
@@ -591,6 +604,7 @@ def create_example_config() -> bool:
 def _load_config_uncached() -> Config:
     """Load and merge builtin + user config. Always returns a valid Config."""
     builtin_config = _get_builtin_config()
+
     user_config = _load_user_config()
 
     return _merge_configs(user_config, builtin_config)

klaude_code/const.py

@@ -7,6 +7,8 @@ that were previously scattered across the codebase.
 from __future__ import annotations
 
 import os
+import sys
+import tempfile
 from dataclasses import dataclass
 from pathlib import Path
 
@@ -22,6 +24,13 @@ def _get_int_env(name: str, default: int) -> int:
         return default
 
 
+def get_system_temp() -> str:
+    """Return system-level temp directory: /tmp on Unix, system temp on Windows."""
+    if sys.platform == "win32":
+        return tempfile.gettempdir()
+    return "/tmp"
+
+
 # =============================================================================
 # Agent / LLM Configuration
 # =============================================================================
@@ -115,7 +124,7 @@ TOOL_OUTPUT_DISPLAY_TAIL = 10000  # Characters to show from the end of truncated
 TOOL_OUTPUT_MAX_LINES = 2000  # Maximum lines for tool output before truncation
 TOOL_OUTPUT_DISPLAY_HEAD_LINES = 1000  # Lines to show from the beginning of truncated output
 TOOL_OUTPUT_DISPLAY_TAIL_LINES = 1000  # Lines to show from the end of truncated output
-TOOL_OUTPUT_TRUNCATION_DIR = "/tmp"  # Directory for saving full truncated output
+TOOL_OUTPUT_TRUNCATION_DIR = get_system_temp()  # Directory for saving full truncated output
 
 
 # =============================================================================
@@ -161,6 +170,7 @@ STATUS_HINT_TEXT = " (esc to interrupt)"  # Status hint text shown after spinner
 STATUS_WAITING_TEXT = "Loading …"
 STATUS_THINKING_TEXT = "Thinking …"
 STATUS_COMPOSING_TEXT = "Composing"
+STATUS_COMPACTING_TEXT = "Compacting"
 
 # Backwards-compatible alias for the default spinner status text.
 STATUS_DEFAULT_TEXT = STATUS_WAITING_TEXT

klaude_code/core/agent.py

@@ -18,9 +18,11 @@ class Agent:
         self,
         session: Session,
         profile: AgentProfile,
+        compact_llm_client: LLMClientABC | None = None,
     ):
         self.session: Session = session
         self.profile: AgentProfile = profile
+        self.compact_llm_client: LLMClientABC | None = compact_llm_client
         self._current_task: TaskExecutor | None = None
         if not self.session.model_name:
             self.session.model_name = profile.llm_client.model_name
@@ -43,18 +45,20 @@ class Agent:
     ) -> AsyncGenerator[events.Event]:
         session_ctx = SessionContext(
             session_id=self.session.id,
-            get_conversation_history=lambda: self.session.conversation_history,
+            get_conversation_history=self.session.get_llm_history,
             append_history=self.session.append_history,
             file_tracker=self.session.file_tracker,
             todo_context=build_todo_context(self.session),
             run_subtask=run_subtask,
         )
         context = TaskExecutionContext(
+            session=self.session,
             session_ctx=session_ctx,
             profile=self.profile,
             tool_registry=get_registry(),
             process_reminder=self._process_reminder,
             sub_agent_state=self.session.sub_agent_state,
+            compact_llm_client=self.compact_llm_client,
         )
 
         task = TaskExecutor(context)

klaude_code/core/agent_profile.py

@@ -12,6 +12,7 @@ from typing import TYPE_CHECKING, Any, Protocol
 if TYPE_CHECKING:
     from klaude_code.config.config import Config
 
+from klaude_code.auth.codex.exceptions import CodexUnsupportedModelError
 from klaude_code.config.sub_agent_model_helper import SubAgentModelHelper
 from klaude_code.core.reminders import (
     at_file_reader_reminder,
@@ -53,15 +54,15 @@ COMMAND_DESCRIPTIONS: dict[str, str] = {
 }
 
 
-# Mapping from logical prompt keys to resource file paths under the core/prompt directory.
-PROMPT_FILES: dict[str, str] = {
-    "main_codex": "prompts/prompt-codex.md",
-    "main_gpt_5_1_codex_max": "prompts/prompt-codex-gpt-5-1-codex-max.md",
-    "main_gpt_5_2_codex": "prompts/prompt-codex-gpt-5-2-codex.md",
-    "main": "prompts/prompt-claude-code.md",
-    "main_gemini": "prompts/prompt-gemini.md",  # https://ai.google.dev/gemini-api/docs/prompting-strategies?hl=zh-cn#agentic-si-template
+# Prompts for codex_oauth protocol - must be used exactly as-is without any additions.
+CODEX_OAUTH_PROMPTS: dict[str, str] = {
+    "gpt-5.2-codex": "prompts/prompt-codex-gpt-5-2-codex.md",
+    "gpt-5.2": "prompts/prompt-codex-gpt-5-2.md",
 }
 
+# Prompt for antigravity protocol - used exactly as-is without any additions.
+ANTIGRAVITY_PROMPT_PATH = "prompts/prompt-antigravity.md"
+
 
 STRUCTURED_OUTPUT_PROMPT_FOR_SUB_AGENT = """\
 
@@ -93,31 +94,20 @@ def _load_prompt_by_path(prompt_path: str) -> str:
     return files(__package__).joinpath(prompt_path).read_text(encoding="utf-8").strip()
 
 
-def _load_base_prompt(file_key: str) -> str:
-    """Load and cache the base prompt content from file."""
-
-    try:
-        prompt_path = PROMPT_FILES[file_key]
-    except KeyError as exc:
-        raise ValueError(f"Unknown prompt key: {file_key}") from exc
-
-    return _load_prompt_by_path(prompt_path)
-
-
-def _get_file_key(model_name: str, protocol: llm_param.LLMClientProtocol) -> str:
-    """Determine which prompt file to use based on model."""
+def _load_prompt_by_model(model_name: str) -> str:
+    """Load base prompt content based on model name."""
 
     match model_name:
         case name if "gpt-5.2-codex" in name:
-            return "main_gpt_5_2_codex"
-        case name if "gpt-5.1-codex-max" in name:
-            return "main_gpt_5_1_codex_max"
+            return _load_prompt_by_path("prompts/prompt-codex-gpt-5-2-codex.md")
+        case name if "gpt-5.2" in name:
+            return _load_prompt_by_path("prompts/prompt-codex-gpt-5-2.md")
         case name if "gpt-5" in name:
-            return "main_codex"
+            return _load_prompt_by_path("prompts/prompt-codex.md")
         case name if "gemini" in name:
-            return "main_gemini"
+            return _load_prompt_by_path("prompts/prompt-gemini.md")
         case _:
-            return "main"
+            return _load_prompt_by_path("prompts/prompt-claude-code.md")
 
 
 def _build_env_info(model_name: str) -> str:
@@ -175,16 +165,22 @@ def load_system_prompt(
 ) -> str:
     """Get system prompt content for the given model and sub-agent type."""
 
+    # For codex_oauth protocol, use exact prompts without any additions.
+    if protocol == llm_param.LLMClientProtocol.CODEX_OAUTH:
+        for model_key, prompt_path in CODEX_OAUTH_PROMPTS.items():
+            if model_key in model_name:
+                return _load_prompt_by_path(prompt_path)
+        raise CodexUnsupportedModelError(f"codex_oauth protocol does not support model: {model_name}")
+
+    # For antigravity protocol, use exact prompt without any additions.
+    if protocol == llm_param.LLMClientProtocol.ANTIGRAVITY:
+        return _load_prompt_by_path(ANTIGRAVITY_PROMPT_PATH)
+
     if sub_agent_type is not None:
         profile = get_sub_agent_profile(sub_agent_type)
         base_prompt = _load_prompt_by_path(profile.prompt_file)
     else:
-        file_key = _get_file_key(model_name, protocol)
-        base_prompt = _load_base_prompt(file_key)
-
-    if protocol == llm_param.LLMClientProtocol.CODEX_OAUTH:
-        # Do not append environment info or skills info for Codex protocol.
-        return base_prompt
+        base_prompt = _load_prompt_by_model(model_name)
 
     skills_prompt = ""
     sub_agent_prompt = ""