klaude-code 1.8.0__py3-none-any.whl → 2.0.0__py3-none-any.whl

klaude_code/config/assets/builtin_config.yaml

@@ -187,6 +187,18 @@ provider_list:
         input: 0.5
         output: 3.0
         cache_read: 0.05
+  - model_name: nano-banana-pro@or
+    model_params:
+      model: google/gemini-3-pro-image-preview
+      context_limit: 1048576
+      modalities:
+      - image
+      - text
+      cost:
+        input: 2
+        output: 12
+        cache_read: 0.2
+        image: 120
   - model_name: grok
     model_params:
       model: x-ai/grok-4.1-fast
@@ -238,6 +250,7 @@ provider_list:
         input: 0.5
         output: 3.0
         cache_read: 0.05
+
 - provider_name: bedrock
   protocol: bedrock
   aws_access_key: ${AWS_ACCESS_KEY_ID}
@@ -259,7 +272,6 @@ provider_list:
   base_url: https://api.deepseek.com/anthropic
   model_list:
   - model_name: deepseek
-    provider: deepseek
     model_params:
       model: deepseek-reasoner
       context_limit: 128000
@@ -290,14 +302,51 @@ provider_list:
         cache_read: 1.0
         currency: CNY
 
+- provider_name: claude-max
+  protocol: claude_oauth
+  model_list:
+  - model_name: sonnet@claude-max
+    model_params:
+      model: claude-sonnet-4-5-20250929
+      context_limit: 200000
+      cost:
+        input: 3.0
+        output: 15.0
+        cache_read: 0.3
+        cache_write: 3.75
+  - model_name: opus@claude-max
+    model_params:
+      model: claude-opus-4-5-20251101
+      context_limit: 200000
+      verbosity: high
+      thinking:
+        type: enabled
+        budget_tokens: 2048
+      cost:
+        input: 5.0
+        output: 25.0
+        cache_read: 0.5
+        cache_write: 6.25
+  - model_name: haiku@claude-max
+    model_params:
+      model: claude-haiku-4-5-20251001
+      context_limit: 200000
+      cost:
+        input: 1.0
+        output: 5.0
+        cache_read: 0.1
+        cache_write: 1.25
+
 - provider_name: codex
-  protocol: codex
+  protocol: codex_oauth
   model_list:
   - model_name: gpt-5.2-codex
-    provider: codex
     model_params:
       model: gpt-5.2-codex
       thinking:
         reasoning_effort: medium
       context_limit: 400000
       max_tokens: 128000
+
+sub_agent_models:
+  ImageGen: nano-banana-pro@or

klaude_code/config/builtin_config.py

@@ -7,7 +7,7 @@ manually configuring providers.
 
 from functools import lru_cache
 from importlib import resources
-from typing import TYPE_CHECKING
+from typing import TYPE_CHECKING, Any
 
 import yaml
 
@@ -26,13 +26,24 @@ SUPPORTED_API_KEY_ENVS = [
 
 
 @lru_cache(maxsize=1)
+def _load_builtin_yaml() -> dict[str, Any]:
+    """Load the built-in config YAML asset."""
+    assets = resources.files("klaude_code.config.assets")
+    yaml_content = (assets / "builtin_config.yaml").read_text()
+    data: dict[str, Any] = yaml.safe_load(yaml_content)
+    return data
+
+
 def get_builtin_provider_configs() -> list["ProviderConfig"]:
     """Load built-in provider configurations from YAML asset."""
     # Import here to avoid circular import
     from klaude_code.config.config import ProviderConfig
 
-    assets = resources.files("klaude_code.config.assets")
-    yaml_content = (assets / "builtin_config.yaml").read_text()
-    data = yaml.safe_load(yaml_content)
-
+    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", {})

klaude_code/config/config.py

@@ -8,7 +8,11 @@ from typing import Any, cast
 import yaml
 from pydantic import BaseModel, Field, ValidationError, model_validator
 
-from klaude_code.config.builtin_config import SUPPORTED_API_KEY_ENVS, get_builtin_provider_configs
+from klaude_code.config.builtin_config import (
+    SUPPORTED_API_KEY_ENVS,
+    get_builtin_provider_configs,
+    get_builtin_sub_agent_models,
+)
 from klaude_code.protocol import llm_param
 from klaude_code.protocol.sub_agent import iter_sub_agent_profiles
 from klaude_code.trace import log
@@ -81,14 +85,23 @@ class ProviderConfig(llm_param.LLMConfigProviderParameter):
         """
         from klaude_code.protocol.llm_param import LLMClientProtocol
 
-        if self.protocol == LLMClientProtocol.CODEX:
+        if self.protocol == LLMClientProtocol.CODEX_OAUTH:
             # Codex uses OAuth authentication, not API key
             from klaude_code.auth.codex.token_manager import CodexTokenManager
 
             token_manager = CodexTokenManager()
             state = token_manager.get_state()
-            # Consider available if logged in and token not expired
-            return state is None or state.is_expired()
+            # Consider available if logged in. Token refresh happens on-demand.
+            return state is None
+
+        if self.protocol == LLMClientProtocol.CLAUDE_OAUTH:
+            # Claude uses OAuth authentication, not API key
+            from klaude_code.auth.claude.token_manager import ClaudeTokenManager
+
+            token_manager = ClaudeTokenManager()
+            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.
@@ -182,7 +195,17 @@ class Config(BaseModel):
         for provider in self.provider_list:
             # Resolve ${ENV_VAR} syntax for api_key
             api_key = provider.get_resolved_api_key()
-            if not api_key:
+
+            # Some protocols do not use API keys for authentication.
+            if (
+                provider.protocol
+                not in {
+                    llm_param.LLMClientProtocol.CODEX_OAUTH,
+                    llm_param.LLMClientProtocol.CLAUDE_OAUTH,
+                    llm_param.LLMClientProtocol.BEDROCK,
+                }
+                and not api_key
+            ):
                 continue
             for model in provider.model_list:
                 if model.model_name == model_name:
@@ -243,7 +266,7 @@ def get_example_config() -> UserConfig:
     """Generate example config for user reference (will be commented out)."""
     return UserConfig(
         main_model="opus",
-        sub_agent_models={"explore": "haiku", "oracle": "gpt-5.2", "webagent": "sonnet", "task": "sonnet"},
+        sub_agent_models={"explore": "haiku", "webagent": "sonnet", "task": "sonnet"},
         provider_list=[
             UserProviderConfig(
                 provider_name="my-provider",
@@ -275,7 +298,8 @@ def _get_builtin_config() -> Config:
     # 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()]
-    return Config(provider_list=providers)
+    sub_agent_models = get_builtin_sub_agent_models()
+    return Config(provider_list=providers, sub_agent_models=sub_agent_models)
 
 
 def _merge_provider(builtin: ProviderConfig, user: UserProviderConfig) -> ProviderConfig:

klaude_code/config/thinking.py

@@ -91,12 +91,12 @@ def format_current_thinking(config: llm_param.LLMConfigParameter) -> str:
 
     protocol = config.protocol
 
-    if protocol in (llm_param.LLMClientProtocol.RESPONSES, llm_param.LLMClientProtocol.CODEX):
+    if protocol in (llm_param.LLMClientProtocol.RESPONSES, llm_param.LLMClientProtocol.CODEX_OAUTH):
         if thinking.reasoning_effort:
             return f"reasoning_effort={thinking.reasoning_effort}"
         return "not set"
 
-    if protocol == llm_param.LLMClientProtocol.ANTHROPIC:
+    if protocol in (llm_param.LLMClientProtocol.ANTHROPIC, llm_param.LLMClientProtocol.CLAUDE_OAUTH):
         if thinking.type == "disabled":
             return "off"
         if thinking.type == "enabled":
@@ -201,7 +201,7 @@ def get_thinking_picker_data(config: llm_param.LLMConfigParameter) -> ThinkingPi
     model_name = config.model
     thinking = config.thinking
 
-    if protocol in (llm_param.LLMClientProtocol.RESPONSES, llm_param.LLMClientProtocol.CODEX):
+    if protocol in (llm_param.LLMClientProtocol.RESPONSES, llm_param.LLMClientProtocol.CODEX_OAUTH):
         levels = get_levels_for_responses(model_name)
         return ThinkingPickerData(
             options=_build_effort_options(levels),
@@ -209,7 +209,7 @@ def get_thinking_picker_data(config: llm_param.LLMConfigParameter) -> ThinkingPi
             current_value=_get_current_effort_value(thinking),
         )
 
-    if protocol == llm_param.LLMClientProtocol.ANTHROPIC:
+    if protocol in (llm_param.LLMClientProtocol.ANTHROPIC, llm_param.LLMClientProtocol.CLAUDE_OAUTH):
         return ThinkingPickerData(
             options=_build_budget_options(),
             message="Select thinking level:",

klaude_code/const.py

@@ -4,7 +4,10 @@ This module consolidates all magic numbers and configuration values
 that were previously scattered across the codebase.
 """
 
+from __future__ import annotations
+
 import os
+from dataclasses import dataclass
 from pathlib import Path
 
 
@@ -20,145 +23,197 @@ def _get_int_env(name: str, default: int) -> int:
 
 
 # =============================================================================
-# Agent Configuration
+# Agent / LLM Configuration
 # =============================================================================
 
-# Timeout for waiting for the first event from LLM (seconds)
-FIRST_EVENT_TIMEOUT_S = 200.0
+MAX_FAILED_TURN_RETRIES = 10  # Maximum retry attempts for failed turns
+LLM_HTTP_TIMEOUT_TOTAL = 300.0  # HTTP timeout for LLM API requests (seconds)
+LLM_HTTP_TIMEOUT_CONNECT = 15.0  # HTTP connect timeout (seconds)
+LLM_HTTP_TIMEOUT_READ = 285.0  # HTTP read timeout (seconds)
 
-# Maximum number of retry attempts for failed turns
-MAX_FAILED_TURN_RETRIES = 10
+ANTHROPIC_BETA_INTERLEAVED_THINKING = "interleaved-thinking-2025-05-14"  # Anthropic API beta flag
+ANTHROPIC_BETA_OAUTH = "oauth-2025-04-20"  # Anthropic OAuth beta flag
+ANTHROPIC_BETA_FINE_GRAINED_TOOL_STREAMING = "fine-grained-tool-streaming-2025-05-14"  # Anthropic streaming beta
+CLAUDE_CODE_IDENTITY = "You are Claude Code, Anthropic's official CLI for Claude."  # Claude identity string
 
-# Initial delay before retrying a failed turn (seconds)
-INITIAL_RETRY_DELAY_S = 1.0
+OPENROUTER_BASE_URL = "https://openrouter.ai/api/v1"  # OpenRouter API base URL
 
-# Maximum delay between retries (seconds)
-MAX_RETRY_DELAY_S = 30.0
+CODEX_BASE_URL = "https://chatgpt.com/backend-api/codex"  # Codex API base URL
+CODEX_USER_AGENT = "codex_cli_rs/0.0.0-klaude"  # Codex user agent string
 
-# Message shown when a tool call is cancelled by the user
-CANCEL_OUTPUT = "[Request interrupted by user for tool use]"
+SUPPORTED_IMAGE_SIZES = {"1K", "2K", "4K"}  # Supported image sizes for ImageGen tool
 
-# Default maximum tokens for LLM responses
-DEFAULT_MAX_TOKENS = 32000
+THROUGHPUT_MIN_DURATION_SEC = 0.15  # Minimum duration (seconds) for throughput calculation
+INITIAL_RETRY_DELAY_S = 1.0  # Initial delay before retrying a failed turn (seconds)
+MAX_RETRY_DELAY_S = 30.0  # Maximum delay between retries (seconds)
+CANCEL_OUTPUT = "[Request interrupted by user for tool use]"  # Message shown when tool call is cancelled
+INTERRUPT_MARKER = " <system>interrupted by user</system>"  # Marker appended when assistant is interrupted
+EMPTY_TOOL_OUTPUT_MESSAGE = (
+    "<system-reminder>Tool ran without output or errors</system-reminder>"  # Tool output placeholder
+)
+DEFAULT_MAX_TOKENS = 32000  # Default maximum tokens for LLM responses
+DEFAULT_TEMPERATURE = 1.0  # Default temperature for LLM requests
+DEFAULT_ANTHROPIC_THINKING_BUDGET_TOKENS = 2048  # Default thinking budget tokens for Anthropic models
 
-# Default temperature for LLM requests
-DEFAULT_TEMPERATURE = 1.0
 
-# Default thinking budget tokens for Anthropic models
-DEFAULT_ANTHROPIC_THINKING_BUDGET_TOKENS = 2048
+# =============================================================================
+# Reminders
+# =============================================================================
 
-# Tool call count threshold for todo reminder
-TODO_REMINDER_TOOL_CALL_THRESHOLD = 10
+TODO_REMINDER_TOOL_CALL_THRESHOLD = 10  # Tool call count threshold for todo reminder
+REMINDER_COOLDOWN_TURNS = 3  # Cooldown turns between reminder triggers
+MEMORY_FILE_NAMES = ["CLAUDE.md", "AGENTS.md", "AGENT.md"]  # Memory file names to search for
 
 
 # =============================================================================
-# Tool
+# Tool - Read
 # =============================================================================
 
-# -- Read Tool --
-# Maximum characters per line before truncation
-READ_CHAR_LIMIT_PER_LINE = 2000
+READ_CHAR_LIMIT_PER_LINE = 2000  # Maximum characters per line before truncation
+READ_GLOBAL_LINE_CAP = _get_int_env("KLAUDE_READ_GLOBAL_LINE_CAP", 2000)  # Maximum lines to read from a file
+READ_MAX_CHARS = _get_int_env("KLAUDE_READ_MAX_CHARS", 50000)  # Maximum total characters to read
+READ_MAX_IMAGE_BYTES = _get_int_env("KLAUDE_READ_MAX_IMAGE_BYTES", 4 * 1024 * 1024)  # Max image size (4MB)
+IMAGE_OUTPUT_MAX_BYTES = _get_int_env("KLAUDE_IMAGE_OUTPUT_MAX_BYTES", 64 * 1024 * 1024)  # Max decoded image (64MB)
+BINARY_CHECK_SIZE = 8192  # Bytes to check for binary file detection
 
-# Maximum number of lines to read from a file
-# Can be overridden via KLAUDE_READ_GLOBAL_LINE_CAP environment variable
-READ_GLOBAL_LINE_CAP = _get_int_env("KLAUDE_READ_GLOBAL_LINE_CAP", 2000)
 
-# Maximum total characters to read (truncates beyond this limit)
-# Can be overridden via KLAUDE_READ_MAX_CHARS environment variable
-READ_MAX_CHARS = _get_int_env("KLAUDE_READ_MAX_CHARS", 50000)
+# =============================================================================
+# Tool - Bash / Shell
+# =============================================================================
 
-# Maximum image file size in bytes (4MB)
-READ_MAX_IMAGE_BYTES = 4 * 1024 * 1024
+BASH_DEFAULT_TIMEOUT_MS = 120000  # Default timeout for bash commands (milliseconds)
+BASH_TERMINATE_TIMEOUT_SEC = 1.0  # Timeout before escalating to SIGKILL (seconds)
 
-# -- Bash Tool --
-# Default timeout for bash commands (milliseconds)
-BASH_DEFAULT_TIMEOUT_MS = 120000
 
-# -- Tool Output --
-# Maximum length for tool output before truncation
-TOOL_OUTPUT_MAX_LENGTH = 40000
+# =============================================================================
+# Tool - Web
+# =============================================================================
 
-# Characters to show from the beginning of truncated output
-TOOL_OUTPUT_DISPLAY_HEAD = 10000
+WEB_FETCH_DEFAULT_TIMEOUT_SEC = 30  # Default timeout for web fetch requests (seconds)
+WEB_FETCH_USER_AGENT = "Mozilla/5.0 (compatible; KlaudeCode/1.0)"  # User-Agent header for web requests
+URL_FILENAME_MAX_LENGTH = 80  # Maximum length for extracting filename from URL
+WEB_SEARCH_DEFAULT_MAX_RESULTS = 10  # Default number of search results
+WEB_SEARCH_MAX_RESULTS_LIMIT = 20  # Maximum number of search results allowed
+MERMAID_LIVE_PREFIX = "https://mermaid.live/view#pako:"  # Mermaid.live URL prefix
 
-# Characters to show from the end of truncated output
-TOOL_OUTPUT_DISPLAY_TAIL = 10000
 
-# Directory for saving full truncated output
-TOOL_OUTPUT_TRUNCATION_DIR = "/tmp/klaude"
+# =============================================================================
+# Tool - Diff
+# =============================================================================
+
+DIFF_MAX_LINE_LENGTH_FOR_CHAR_DIFF = 2000  # Maximum line length for character-level diff
+DIFF_DEFAULT_CONTEXT_LINES = 3  # Default number of context lines in diff output
 
 
 # =============================================================================
-# UI
+# Tool - Output Truncation
 # =============================================================================
 
-# Width of line number prefix in diff display
-DIFF_PREFIX_WIDTH = 4
+TOOL_OUTPUT_MAX_LENGTH = 40000  # Maximum length for tool output before truncation
+TOOL_OUTPUT_DISPLAY_HEAD = 10000  # Characters to show from the beginning of truncated output
+TOOL_OUTPUT_DISPLAY_TAIL = 10000  # Characters to show from the end of truncated output
+TOOL_OUTPUT_TRUNCATION_DIR = "/tmp/klaude"  # Directory for saving full truncated output
 
-# Maximum lines to show in diff output
-MAX_DIFF_LINES = 1000
 
-# Maximum length for invalid tool call display
-INVALID_TOOL_CALL_MAX_LENGTH = 500
+# =============================================================================
+# UI - Display
+# =============================================================================
 
-# Maximum line length for truncated display output
-TRUNCATE_DISPLAY_MAX_LINE_LENGTH = 1000
+TAB_EXPAND_WIDTH = 8  # Tab expansion width for text rendering
+DIFF_PREFIX_WIDTH = 4  # Width of line number prefix in diff display
+MAX_DIFF_LINES = 500  # Maximum lines to show in diff output
+INVALID_TOOL_CALL_MAX_LENGTH = 500  # Maximum length for invalid tool call display
+TRUNCATE_DISPLAY_MAX_LINE_LENGTH = 1000  # Maximum line length for truncated display
+TRUNCATE_DISPLAY_MAX_LINES = 6  # Maximum lines for truncated display
+MIN_HIDDEN_LINES_FOR_INDICATOR = 5  # Minimum hidden lines before showing truncation indicator
+SUB_AGENT_RESULT_MAX_LINES = 10  # Maximum lines for sub-agent result display
+TRUNCATE_HEAD_MAX_LINES = 2  # Maximum lines for sub-agent error display
+BASH_OUTPUT_PANEL_THRESHOLD = 10  # Bash output line threshold for CodePanel display
+URL_TRUNCATE_MAX_LENGTH = 400  # Maximum length for URL truncation in display
+QUERY_DISPLAY_TRUNCATE_LENGTH = 80  # Maximum length for search query display
+NOTIFY_COMPACT_LIMIT = 160  # Maximum length for notification body text
 
-# Maximum lines for truncated display output
-TRUNCATE_DISPLAY_MAX_LINES = 8
 
-# Maximum lines for sub-agent result display
-SUB_AGENT_RESULT_MAX_LINES = 50
+# =============================================================================
+# UI - Markdown Streaming
+# =============================================================================
 
+UI_REFRESH_RATE_FPS = 10  # UI refresh rate (frames per second)
+CROP_ABOVE_LIVE_REFRESH_PER_SECOND = 4.0  # CropAboveLive default refresh rate
+MARKDOWN_STREAM_LIVE_REPAINT_ENABLED = True  # Enable live area for streaming markdown
+STREAM_MAX_HEIGHT_SHRINK_RESET_LINES = 20  # Reset stream height ceiling after this shrinkage
+MARKDOWN_LEFT_MARGIN = 2  # Left margin (columns) for markdown rendering
+MARKDOWN_RIGHT_MARGIN = 2  # Right margin (columns) for markdown rendering
 
-# UI refresh rate (frames per second) for debounced content streaming
-UI_REFRESH_RATE_FPS = 10
 
-# Enable live area for streaming markdown (shows incomplete blocks being typed)
-# When False, only completed markdown blocks are displayed (more stable, less flicker)
-MARKDOWN_STREAM_LIVE_REPAINT_ENABLED = False
+# =============================================================================
+# UI - Spinner / Status
+# =============================================================================
 
-# Number of lines to keep visible at bottom of markdown streaming window
-MARKDOWN_STREAM_LIVE_WINDOW = 6
+STATUS_HINT_TEXT = " (esc to interrupt)"  # Status hint text shown after spinner
+STATUS_DEFAULT_TEXT = "Thinking …"  # Default spinner status text
+SPINNER_BREATH_PERIOD_SECONDS: float = 2.0  # Spinner breathing animation period (seconds)
+STATUS_SHIMMER_PADDING = 10  # Horizontal padding for shimmer band position
+STATUS_SHIMMER_BAND_HALF_WIDTH = 5.0  # Half-width of shimmer band in characters
+STATUS_SHIMMER_ALPHA_SCALE = 0.7  # Scale factor for shimmer intensity
 
-# Left margin (columns) to reserve when rendering markdown
-MARKDOWN_LEFT_MARGIN = 2
 
-# Right margin (columns) to reserve when rendering markdown
-MARKDOWN_RIGHT_MARGIN = 2
+# =============================================================================
+# UI - Completion System
+# =============================================================================
 
-# Status hint text shown after spinner status
-STATUS_HINT_TEXT = " (esc to interrupt)"
+COMPLETER_DEBOUNCE_SEC = 0.25  # Debounce time for file path completion (seconds)
+COMPLETER_CACHE_TTL_SEC = 60.0  # Cache TTL for completion results (seconds)
+COMPLETER_CMD_TIMEOUT_SEC = 3.0  # Timeout for completion subprocess commands (seconds)
 
-# Default spinner status text when idle/thinking
-STATUS_DEFAULT_TEXT = "Thinking …"
 
-# Status shimmer animation
-# Horizontal padding used when computing shimmer band position
-STATUS_SHIMMER_PADDING = 10
-# Half-width of the shimmer band in characters
-STATUS_SHIMMER_BAND_HALF_WIDTH = 5.0
-# Scale factor applied to shimmer intensity when blending colors
-STATUS_SHIMMER_ALPHA_SCALE = 0.7
+# =============================================================================
+# Debug / Logging
+# =============================================================================
 
-# Spinner breathing and shimmer animation period
-# Duration in seconds for one full breathe-in + breathe-out cycle (breathing)
-# and one full shimmer sweep across the text (shimmer)
-SPINNER_BREATH_PERIOD_SECONDS: float = 2.0
+DEFAULT_DEBUG_LOG_DIR = Path.home() / ".klaude" / "logs"  # Default debug log directory
+DEFAULT_DEBUG_LOG_FILE = DEFAULT_DEBUG_LOG_DIR / "debug.log"  # Default debug log file path
+LOG_MAX_BYTES = 10 * 1024 * 1024  # Maximum log file size before rotation (10MB)
+LOG_BACKUP_COUNT = 3  # Number of backup log files to keep
 
 
 # =============================================================================
-# Debug / Logging
+# Project Paths
 # =============================================================================
 
-# Default debug log directory (user cache)
-DEFAULT_DEBUG_LOG_DIR = Path.home() / ".klaude" / "logs"
 
-# Default debug log file path (symlink to latest session)
-DEFAULT_DEBUG_LOG_FILE = DEFAULT_DEBUG_LOG_DIR / "debug.log"
+def project_key_from_cwd() -> str:
+    """Derive the project key from the current working directory."""
+    return str(Path.cwd()).strip("/").replace("/", "-")
+
+
+@dataclass(frozen=True)
+class ProjectPaths:
+    """Path utilities for project-scoped storage."""
+
+    project_key: str
+
+    @property
+    def base_dir(self) -> Path:
+        return Path.home() / ".klaude" / "projects" / self.project_key
+
+    @property
+    def sessions_dir(self) -> Path:
+        return self.base_dir / "sessions"
+
+    @property
+    def exports_dir(self) -> Path:
+        return self.base_dir / "exports"
+
+    def session_dir(self, session_id: str) -> Path:
+        return self.sessions_dir / session_id
+
+    def images_dir(self, session_id: str) -> Path:
+        """Return the directory for storing session-scoped image artifacts."""
+        return self.session_dir(session_id) / "images"
 
-# Maximum log file size before rotation (10MB)
-LOG_MAX_BYTES = 10 * 1024 * 1024
+    def events_file(self, session_id: str) -> Path:
+        return self.session_dir(session_id) / "events.jsonl"
 
-# Number of backup log files to keep
-LOG_BACKUP_COUNT = 3
+    def meta_file(self, session_id: str) -> Path:
+        return self.session_dir(session_id) / "meta.json"

klaude_code/core/agent.py

@@ -9,8 +9,8 @@ from klaude_code.core.reminders import Reminder, load_agent_reminders
 from klaude_code.core.task import SessionContext, TaskExecutionContext, TaskExecutor
 from klaude_code.core.tool import build_todo_context, get_registry, load_agent_tools
 from klaude_code.llm import LLMClientABC
-from klaude_code.protocol import events, llm_param, model, tools
-from klaude_code.protocol.model import UserInputPayload
+from klaude_code.protocol import events, llm_param, tools
+from klaude_code.protocol.message import UserInputPayload
 from klaude_code.session import Session
 from klaude_code.trace import DebugType, log_debug
 
@@ -82,21 +82,12 @@ class Agent:
             self.session.model_name = profile.llm_client.model_name
 
     def cancel(self) -> Iterable[events.Event]:
-        """Handle agent cancellation and persist an interrupt marker and tool cancellations.
-
-        - Appends an `InterruptItem` into the session history so interruptions are reflected
-          in persisted conversation logs.
-        - For any tool calls that are pending or in-progress in the current task, delegate to
-          the active TaskExecutor to append synthetic ToolResultItem entries with error status
-          to indicate cancellation.
-        """
+        """Handle agent cancellation and tool cancellations."""
         # First, cancel any running task so it stops emitting events.
         if self._current_task is not None:
             yield from self._current_task.cancel()
             self._current_task = None
 
-        # Record an interrupt marker in the session history
-        self.session.append_history([model.InterruptItem()])
         log_debug(
             f"Session {self.session.id} interrupted",
             style="yellow",