klaude-code 2.5.1__py3-none-any.whl → 2.5.3__py3-none-any.whl

klaude_code/cli/auth_cmd.py

@@ -4,20 +4,9 @@ import datetime
 import webbrowser
 
 import typer
-from prompt_toolkit.styles import Style
 
 from klaude_code.log import log
-from klaude_code.tui.terminal.selector import SelectItem, select_one
-
-_SELECT_STYLE = Style(
-    [
-        ("instruction", "ansibrightblack"),
-        ("pointer", "ansigreen"),
-        ("highlighted", "ansigreen"),
-        ("text", "ansibrightblack"),
-        ("question", "bold"),
-    ]
-)
+from klaude_code.tui.terminal.selector import DEFAULT_PICKER_STYLE, SelectItem, select_one
 
 
 def _select_provider() -> str | None:
@@ -30,7 +19,7 @@ def _select_provider() -> str | None:
         message="Select provider to login:",
         items=items,
         pointer="→",
-        style=_SELECT_STYLE,
+        style=DEFAULT_PICKER_STYLE,
         use_search_filter=False,
     )
 

klaude_code/cli/cost_cmd.py

@@ -183,14 +183,14 @@ def render_cost_table(daily_stats: dict[str, DailyStats]) -> Table:
         box=ASCII_HORIZONAL,
     )
 
-    table.add_column("Date", style="cyan", no_wrap=True)
-    table.add_column("Model", no_wrap=True)
-    table.add_column("Input", justify="right", no_wrap=True)
-    table.add_column("Output", justify="right", no_wrap=True)
-    table.add_column("Cache", justify="right", no_wrap=True)
-    table.add_column("Total", justify="right", no_wrap=True)
-    table.add_column("USD", justify="right", no_wrap=True)
-    table.add_column("CNY", justify="right", no_wrap=True)
+    table.add_column("Date", style="cyan")
+    table.add_column("Model", overflow="ellipsis")
+    table.add_column("Input", justify="right")
+    table.add_column("Output", justify="right")
+    table.add_column("Cache", justify="right")
+    table.add_column("Total", justify="right")
+    table.add_column("USD", justify="right")
+    table.add_column("CNY", justify="right")
 
     # Sort dates
     sorted_dates = sorted(daily_stats.keys())
@@ -222,7 +222,7 @@ def render_cost_table(daily_stats: dict[str, DailyStats]) -> Table:
 
             table.add_row(
                 format_date_display(date_str) if first_row else "",
-                f"- {model_name}",
+                f"{model_name}",
                 format_tokens(stats.input_tokens),
                 format_tokens(stats.output_tokens),
                 format_tokens(stats.cached_tokens),
@@ -276,7 +276,7 @@ def render_cost_table(daily_stats: dict[str, DailyStats]) -> Table:
         usd_str, cny_str = format_cost_dual(stats.cost_usd, stats.cost_cny)
         table.add_row(
             "",
-            f"- {model_name}",
+            f"{model_name}",
             format_tokens(stats.input_tokens),
             format_tokens(stats.output_tokens),
             format_tokens(stats.cached_tokens),

klaude_code/cli/list_model.py

@@ -288,6 +288,14 @@ def _build_models_table(
             name = Text.assemble((prefix, ThemeKey.LINES), (model.model_name, "dim"))
             model_id = Text(model.model_id or "", style="dim")
             params = Text("(unavailable)", style="dim")
+        elif model.disabled:
+            name = Text.assemble(
+                (prefix, ThemeKey.LINES),
+                (model.model_name, "dim strike"),
+                (" (disabled)", "dim"),
+            )
+            model_id = Text(model.model_id or "", style="dim")
+            params = Text(" · ").join(_get_model_params_display(model))
         else:
             # Build role tags for this model
             roles: list[str] = []

klaude_code/cli/main.py

@@ -13,13 +13,46 @@ from klaude_code.session import Session
 from klaude_code.tui.command.resume_cmd import select_session_sync
 from klaude_code.ui.terminal.title import update_terminal_title
 
-ENV_HELP = """\
-Environment Variables:
-
-  KLAUDE_READ_GLOBAL_LINE_CAP  Max lines to read (default: 2000)
-
-  KLAUDE_READ_MAX_CHARS        Max total chars to read (default: 50000)
-"""
+ENV_HELP_LINES = [
+    "Environment Variables:",
+    "",
+    "Provider API keys (built-in config):",
+    "  ANTHROPIC_API_KEY             Anthropic API key",
+    "  OPENAI_API_KEY                OpenAI API key",
+    "  OPENROUTER_API_KEY            OpenRouter API key",
+    "  GOOGLE_API_KEY                Google API key (Gemini)",
+    "  DEEPSEEK_API_KEY              DeepSeek API key",
+    "  MOONSHOT_API_KEY              Moonshot API key (Kimi)",
+    "",
+    "AWS credentials (Bedrock):",
+    "  AWS_ACCESS_KEY_ID             AWS access key id",
+    "  AWS_SECRET_ACCESS_KEY         AWS secret access key",
+    "  AWS_REGION                    AWS region",
+    "",
+    "Tool limits (Read):",
+    "  KLAUDE_READ_GLOBAL_LINE_CAP    Max lines to read (default: 2000)",
+    "  KLAUDE_READ_MAX_CHARS          Max total chars to read (default: 50000)",
+    "  KLAUDE_READ_MAX_IMAGE_BYTES    Max image bytes to read (default: 4MB)",
+    "  KLAUDE_IMAGE_OUTPUT_MAX_BYTES  Max decoded image bytes (default: 64MB)",
+    "",
+    "Notifications / testing:",
+    "  KLAUDE_NOTIFY                 Set to 0/off/false/disable(d) to disable task notifications",
+    "  KLAUDE_TEST_SIGNAL            In tmux, emit `tmux wait-for -S <channel>` on task completion",
+    "  TMUX                          Auto-detected; required for KLAUDE_TEST_SIGNAL",
+    "",
+    "Editor / terminal integration:",
+    "  EDITOR                        Preferred editor for `klaude config`",
+    "  TERM                          Terminal identification (auto-detected)",
+    "  TERM_PROGRAM                  Terminal identification (auto-detected)",
+    "  WT_SESSION                    Terminal hint (auto-detected)",
+    "  VTE_VERSION                   Terminal hint (auto-detected)",
+    "  GHOSTTY_RESOURCES_DIR         Ghostty detection (auto-detected)",
+    "",
+    "Compatibility:",
+    "  ANTHROPIC_AUTH_TOKEN          Reserved by anthropic SDK; temporarily unset during client init",
+]
+
+ENV_HELP = "\n\n".join(ENV_HELP_LINES)
 
 app = typer.Typer(
     add_completion=False,
@@ -192,7 +225,7 @@ def main_callback(
                 if raw_model:
                     matches = [
                         m.selector
-                        for m in cfg.iter_model_entries()
+                        for m in cfg.iter_model_entries(only_available=True, include_disabled=False)
                         if (m.model_id or "").strip().lower() == raw_model.lower()
                     ]
                     if len(matches) == 1:

klaude_code/cli/session_cmd.py

@@ -9,9 +9,7 @@ from klaude_code.session import Session
 def _session_confirm(sessions: list[Session.SessionMetaBrief], message: str) -> bool:
     """Show session list and confirm deletion using prompt_toolkit."""
 
-    from prompt_toolkit.styles import Style
-
-    from klaude_code.tui.terminal.selector import SelectItem, select_one
+    from klaude_code.tui.terminal.selector import DEFAULT_PICKER_STYLE, SelectItem, select_one
 
     def _fmt(ts: float) -> str:
         try:
@@ -37,14 +35,7 @@ def _session_confirm(sessions: list[Session.SessionMetaBrief], message: str) ->
         message=message,
         items=items,
         pointer="→",
-        style=Style(
-            [
-                ("question", "bold"),
-                ("pointer", "ansigreen"),
-                ("highlighted", "ansigreen"),
-                ("text", ""),
-            ]
-        ),
+        style=DEFAULT_PICKER_STYLE,
         use_search_filter=False,
     )
     return bool(result)

klaude_code/config/assets/builtin_config.yaml

@@ -1,4 +1,3 @@
----
 # Built-in provider and model configurations
 # Users can start using klaude by simply setting environment variables
 # (ANTHROPIC_API_KEY, OPENAI_API_KEY, etc.) without manual configuration.
@@ -25,32 +24,58 @@ provider_list:
   protocol: responses
   api_key: ${OPENAI_API_KEY}
   model_list:
-  - model_name: gpt-5.2
+  - 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}
+  - model_name: gpt-5.2-medium
+    model_id: gpt-5.2
+    context_limit: 400000
+    verbosity: high
+    thinking:
+      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
+    verbosity: low
+    thinking:
+      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
+    verbosity: low
+    thinking:
+      reasoning_effort: none
     cost: {input: 1.75, output: 14, cache_read: 0.17}
-- provider_name: openrouter
-  protocol: openrouter
-  api_key: ${OPENROUTER_API_KEY}
-  model_list:
   - model_name: gpt-5.1-codex-max
-    model_id: openai/gpt-5.1-codex-max
+    model_id: gpt-5.1-codex-max
     max_tokens: 128000
     context_limit: 400000
     thinking:
       reasoning_effort: medium
+      reasoning_summary: detailed
     cost: {input: 1.25, output: 10, cache_read: 0.13}
-  - model_name: gpt-5.2
+- 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
     context_limit: 400000
     verbosity: high
     thinking:
       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
@@ -59,22 +84,7 @@ provider_list:
     verbosity: high
     thinking:
       reasoning_effort: medium
-    cost: {input: 1.75, output: 14, cache_read: 0.17}
-  - model_name: gpt-5.2-low
-    model_id: openai/gpt-5.2
-    max_tokens: 128000
-    context_limit: 400000
-    verbosity: low
-    thinking:
-      reasoning_effort: low
-    cost: {input: 1.75, output: 14, cache_read: 0.17}
-  - model_name: gpt-5.2-fast
-    model_id: openai/gpt-5.2
-    max_tokens: 128000
-    context_limit: 400000
-    verbosity: low
-    thinking:
-      reasoning_effort: none
+      reasoning_summary: concise
     cost: {input: 1.75, output: 14, cache_read: 0.17}
   - model_name: kimi
     model_id: moonshotai/kimi-k2-thinking
@@ -96,7 +106,6 @@ provider_list:
   - model_name: opus
     model_id: anthropic/claude-4.5-opus
     context_limit: 200000
-    verbosity: high
     thinking:
       type: enabled
       budget_tokens: 2048
@@ -166,10 +175,14 @@ provider_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
@@ -178,6 +191,13 @@ 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
+    modalities:
+    - 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}
@@ -222,7 +242,6 @@ provider_list:
   - model_name: opus
     model_id: claude-opus-4-5-20251101
     context_limit: 200000
-    verbosity: high
     thinking:
       type: enabled
       budget_tokens: 2048

klaude_code/config/config.py

@@ -332,11 +332,12 @@ class Config(BaseModel):
 
         raise ValueError(f"Unknown model: {model_name}")
 
-    def iter_model_entries(self, only_available: bool = False) -> list[ModelEntry]:
+    def iter_model_entries(self, only_available: bool = False, include_disabled: bool = True) -> list[ModelEntry]:
         """Return all model entries with their provider names.
 
         Args:
             only_available: If True, only return models from providers with valid API keys.
+            include_disabled: If False, exclude models with disabled=True.
         """
         return [
             ModelEntry(
@@ -347,25 +348,26 @@ class Config(BaseModel):
             for provider in self.provider_list
             if not only_available or not provider.is_api_key_missing()
             for model in provider.model_list
+            if include_disabled or not model.disabled
         ]
 
     def has_available_image_model(self) -> bool:
         """Check if any image generation model is available."""
-        for entry in self.iter_model_entries(only_available=True):
+        for entry in self.iter_model_entries(only_available=True, include_disabled=False):
             if entry.modalities and "image" in entry.modalities:
                 return True
         return False
 
     def get_first_available_nano_banana_model(self) -> str | None:
         """Get the first available nano-banana model, or None."""
-        for entry in self.iter_model_entries(only_available=True):
+        for entry in self.iter_model_entries(only_available=True, include_disabled=False):
             if "nano-banana" in entry.model_name:
                 return entry.model_name
         return None
 
     def get_first_available_image_model(self) -> str | None:
         """Get the first available image generation model, or None."""
-        for entry in self.iter_model_entries(only_available=True):
+        for entry in self.iter_model_entries(only_available=True, include_disabled=False):
             if entry.modalities and "image" in entry.modalities:
                 return entry.model_name
         return None
@@ -435,11 +437,26 @@ def _get_builtin_config() -> Config:
     return Config(provider_list=providers, sub_agent_models=sub_agent_models)
 
 
+def _merge_model(builtin: ModelConfig, user: ModelConfig) -> ModelConfig:
+    """Merge user model config with builtin model config.
+
+    Strategy: user values take precedence if explicitly set (not default).
+    This allows users to override specific fields (e.g., disabled=true)
+    without losing other builtin settings (e.g., model_id, max_tokens).
+    """
+    merged_data = builtin.model_dump()
+    user_data = user.model_dump(exclude_defaults=True, exclude={"model_name"})
+    for key, value in user_data.items():
+        if value is not None:
+            merged_data[key] = value
+    return ModelConfig.model_validate(merged_data)
+
+
 def _merge_provider(builtin: ProviderConfig, user: UserProviderConfig) -> ProviderConfig:
     """Merge user provider config with builtin provider config.
 
     Strategy:
-    - model_list: merge by model_name, user models override builtin models with same name
+    - model_list: merge by model_name, user model fields override builtin fields
     - Other fields (api_key, base_url, etc.): user config takes precedence if set
     """
     # Merge model_list: builtin first, then user overrides/appends
@@ -447,7 +464,12 @@ def _merge_provider(builtin: ProviderConfig, user: UserProviderConfig) -> Provid
     for m in builtin.model_list:
         merged_models[m.model_name] = m
     for m in user.model_list:
-        merged_models[m.model_name] = m
+        if m.model_name in merged_models:
+            # Merge with builtin model
+            merged_models[m.model_name] = _merge_model(merged_models[m.model_name], m)
+        else:
+            # New model from user
+            merged_models[m.model_name] = m
 
     # For other fields, use user values if explicitly set, otherwise use builtin
     # We check if user explicitly provided a value by comparing to defaults
@@ -578,7 +600,8 @@ def _load_config_cached() -> Config:
 def load_config() -> Config:
     """Load config from disk (builtin + user merged).
 
-    Always returns a valid Config. Use config.iter_model_entries(only_available=True)
+    Always returns a valid Config. Use
+    ``config.iter_model_entries(only_available=True, include_disabled=False)``
     to check if any models are actually usable.
     """
     try:

klaude_code/config/model_matcher.py

@@ -48,9 +48,9 @@ def match_model_from_config(preferred: str | None = None) -> ModelMatchResult:
     """
     config = load_config()
 
-    # Only show models from providers with valid API keys
+    # Only show models from providers with valid API keys, exclude disabled models
     models: list[ModelEntry] = sorted(
-        config.iter_model_entries(only_available=True),
+        config.iter_model_entries(only_available=True, include_disabled=False),
         key=lambda m: (m.provider.lower(), m.model_name.lower()),
     )
 
@@ -102,6 +102,7 @@ def match_model_from_config(preferred: str | None = None) -> ModelMatchResult:
             )
 
         # Normalized matching (e.g. gpt52 == gpt-5.2, gpt52 in gpt-5.2-2025-...)
+        # Only match selector/model_name exactly; model_id is checked via substring match below
         preferred_norm = _normalize_model_key(preferred)
         normalized_matches: list[ModelEntry] = []
         if preferred_norm:
@@ -110,7 +111,6 @@ def match_model_from_config(preferred: str | None = None) -> ModelMatchResult:
                 for m in models
                 if preferred_norm == _normalize_model_key(m.selector)
                 or preferred_norm == _normalize_model_key(m.model_name)
-                or preferred_norm == _normalize_model_key(m.model_id or "")
             ]
             if len(normalized_matches) == 1:
                 return ModelMatchResult(

klaude_code/config/sub_agent_model_helper.py

@@ -180,7 +180,7 @@ class SubAgentModelHelper:
         - Returns all available models
         """
         profile = get_sub_agent_profile(sub_agent_type)
-        all_models = self._config.iter_model_entries(only_available=True)
+        all_models = self._config.iter_model_entries(only_available=True, include_disabled=False)
 
         if profile.availability_requirement == AVAILABILITY_IMAGE_MODEL:
             return [m for m in all_models if m.modalities and "image" in m.modalities]

klaude_code/const.py

@@ -27,6 +27,7 @@ def _get_int_env(name: str, default: int) -> int:
 # =============================================================================
 
 MAX_FAILED_TURN_RETRIES = 10  # Maximum retry attempts for failed turns
+RETRY_PRESERVE_PARTIAL_MESSAGE = True  # Preserve partial message on stream error for retry prefill
 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)
@@ -157,7 +158,7 @@ MARKDOWN_RIGHT_MARGIN = 2  # Right margin (columns) for markdown rendering
 STATUS_HINT_TEXT = " (esc to interrupt)"  # Status hint text shown after spinner
 
 # Spinner status texts
-STATUS_WAITING_TEXT = "Connecting …"
+STATUS_WAITING_TEXT = "Loading …"
 STATUS_THINKING_TEXT = "Thinking …"
 STATUS_COMPOSING_TEXT = "Composing"
 

klaude_code/core/agent_profile.py

@@ -48,6 +48,7 @@ COMMAND_DESCRIPTIONS: dict[str, str] = {
     "fd": "simple and fast alternative to find",
     "tree": "directory listing as a tree",
     "sg": "ast-grep - AST-aware code search",
+    "jq": "command-line JSON processor",
     "jj": "jujutsu - Git-compatible version control system",
 }
 

klaude_code/core/executor.py

@@ -18,6 +18,7 @@ from klaude_code.config import load_config
 from klaude_code.config.sub_agent_model_helper import SubAgentModelHelper
 from klaude_code.core.agent import Agent
 from klaude_code.core.agent_profile import DefaultModelProfileProvider, ModelProfileProvider
+from klaude_code.core.loaded_skills import get_loaded_skill_names_by_location
 from klaude_code.core.manager import LLMClients, SubAgentManager
 from klaude_code.llm.registry import create_llm_client
 from klaude_code.log import DebugType, log_debug
@@ -136,6 +137,7 @@ class AgentRuntime:
                 session_id=session.id,
                 work_dir=str(session.work_dir),
                 llm_config=self._llm_clients.main.get_llm_config(),
+                loaded_skills=get_loaded_skill_names_by_location(),
             )
         )
 
@@ -192,6 +194,7 @@ class AgentRuntime:
                 session_id=agent.session.id,
                 work_dir=str(agent.session.work_dir),
                 llm_config=self._llm_clients.main.get_llm_config(),
+                loaded_skills=get_loaded_skill_names_by_location(),
             )
         )
 
@@ -215,6 +218,7 @@ class AgentRuntime:
                 session_id=target_session.id,
                 work_dir=str(target_session.work_dir),
                 llm_config=self._llm_clients.main.get_llm_config(),
+                loaded_skills=get_loaded_skill_names_by_location(),
             )
         )
 

klaude_code/core/loaded_skills.py

@@ -0,0 +1,36 @@
+from __future__ import annotations
+
+
+def get_loaded_skill_names_by_location() -> dict[str, list[str]]:
+    """Return loaded skill names grouped by location.
+
+    The UI should not import the skill system directly. Core can expose a
+    lightweight summary suitable for WelcomeEvent rendering.
+    """
+
+    try:
+        # Import lazily to keep startup overhead minimal and avoid unnecessary
+        # coupling at module import time.
+        from klaude_code.skill.manager import get_available_skills
+    except Exception:
+        return {}
+
+    result: dict[str, list[str]] = {"user": [], "project": [], "system": []}
+    try:
+        for name, _desc, location in get_available_skills():
+            if location == "user":
+                result["user"].append(name)
+            elif location == "project":
+                result["project"].append(name)
+            elif location == "system":
+                result["system"].append(name)
+    except Exception:
+        return {}
+
+    if not result["user"] and not result["project"] and not result["system"]:
+        return {}
+
+    result["user"].sort()
+    result["project"].sort()
+    result["system"].sort()
+    return result

klaude_code/core/tool/context.py

@@ -89,7 +89,5 @@ class ToolContext:
     def with_record_sub_agent_session_id(self, callback: Callable[[str], None] | None) -> ToolContext:
         return replace(self, record_sub_agent_session_id=callback)
 
-    def with_register_sub_agent_metadata_getter(
-        self, callback: Callable[[GetMetadataFn], None] | None
-    ) -> ToolContext:
+    def with_register_sub_agent_metadata_getter(self, callback: Callable[[GetMetadataFn], None] | None) -> ToolContext:
         return replace(self, register_sub_agent_metadata_getter=callback)

klaude_code/core/tool/file/edit_tool.py

@@ -98,7 +98,7 @@ class EditTool(ToolABC):
         if is_directory(file_path):
             return message.ToolResultMessage(
                 status="error",
-                output_text="<tool_use_error>Illegal operation on a directory. edit</tool_use_error>",
+                output_text="<tool_use_error>Illegal operation on a directory: edit</tool_use_error>",
             )
 
         if args.old_string == "":

klaude_code/core/tool/file/read_tool.py

@@ -210,7 +210,7 @@ class ReadTool(ToolABC):
         if is_directory(file_path):
             return message.ToolResultMessage(
                 status="error",
-                output_text="<tool_use_error>Illegal operation on a directory. read</tool_use_error>",
+                output_text="<tool_use_error>Illegal operation on a directory: read</tool_use_error>",
             )
         if not file_exists(file_path):
             return message.ToolResultMessage(
@@ -308,7 +308,7 @@ class ReadTool(ToolABC):
         except IsADirectoryError:
             return message.ToolResultMessage(
                 status="error",
-                output_text="<tool_use_error>Illegal operation on a directory. read</tool_use_error>",
+                output_text="<tool_use_error>Illegal operation on a directory: read</tool_use_error>",
             )
 
         if offset > max(read_result.total_lines, 0):

klaude_code/core/tool/file/write_tool.py

@@ -57,7 +57,7 @@ class WriteTool(ToolABC):
         if is_directory(file_path):
             return message.ToolResultMessage(
                 status="error",
-                output_text="<tool_use_error>Illegal operation on a directory. write</tool_use_error>",
+                output_text="<tool_use_error>Illegal operation on a directory: write</tool_use_error>",
             )
 
         file_tracker = context.file_tracker

klaude_code/core/turn.py

@@ -4,7 +4,7 @@ from collections.abc import AsyncGenerator
 from dataclasses import dataclass, field
 from typing import TYPE_CHECKING
 
-from klaude_code.const import SUPPORTED_IMAGE_SIZES
+from klaude_code.const import RETRY_PRESERVE_PARTIAL_MESSAGE, SUPPORTED_IMAGE_SIZES
 from klaude_code.core.tool import ToolABC
 from klaude_code.core.tool.context import SubAgentResumeClaims, ToolContext
 
@@ -24,6 +24,12 @@ from klaude_code.llm.client import LLMStreamABC
 from klaude_code.log import DebugType, log_debug
 from klaude_code.protocol import events, llm_param, message, model, tools
 
+# Protocols that support prefill (continuing from partial assistant message)
+_PREFILL_SUPPORTED_PROTOCOLS = frozenset({
+    "anthropic",
+    "claude_oauth",
+})
+
 
 class TurnError(Exception):
     """Raised when a turn fails and should be retried."""
@@ -176,6 +182,18 @@ class TurnExecutor:
             yield event
 
         if self._turn_result.stream_error is not None:
+            # Save accumulated content for potential prefill on retry (only for supported protocols)
+            protocol = ctx.llm_client.get_llm_config().protocol
+            supports_prefill = protocol.value in _PREFILL_SUPPORTED_PROTOCOLS
+            if (
+                RETRY_PRESERVE_PARTIAL_MESSAGE
+                and supports_prefill
+                and self._turn_result.assistant_message is not None
+                and self._turn_result.assistant_message.parts
+            ):
+                session_ctx.append_history([self._turn_result.assistant_message])
+                # Add continuation prompt to avoid Anthropic thinking block requirement
+                session_ctx.append_history([message.UserMessage(parts=[message.TextPart(text="continue")])])
             session_ctx.append_history([self._turn_result.stream_error])
             yield events.TurnEndEvent(session_id=session_ctx.session_id)
             raise TurnError(self._turn_result.stream_error.error)
@@ -339,12 +357,6 @@ class TurnExecutor:
                             )
                     case message.StreamErrorItem() as msg:
                         turn_result.stream_error = msg
-                        log_debug(
-                            "[StreamError]",
-                            msg.error,
-                            style="red",
-                            debug_type=DebugType.RESPONSE,
-                        )
                     case message.ToolCallStartDelta() as msg:
                         if thinking_active:
                             thinking_active = False