klaude-code 2.10.1__py3-none-any.whl → 2.10.3__py3-none-any.whl

klaude_code/config/assets/builtin_config.yaml

@@ -102,12 +102,12 @@ provider_list:
     cost: {input: 1.75, output: 14, cache_read: 0.17}
 
   - model_name: kimi
-    model_id: moonshotai/kimi-k2-thinking
+    model_id: moonshotai/kimi-k2.5
     context_limit: 262144
-    provider_routing:
-      only:
-      - moonshotai/turbo
-    cost: {input: 0.6, output: 2.5, cache_read: 0.15}
+    # provider_routing:
+    #   only:
+    #   - moonshotai/turbo
+    cost: {input: 0.6, output: 3, cache_read: 0.1}
 
   - model_name: haiku
     model_id: anthropic/claude-haiku-4.5
@@ -272,12 +272,12 @@ provider_list:
   model_list:
 
   - model_name: kimi
-    model_id: kimi-k2-thinking
+    model_id: kimi-k2.5
     context_limit: 262144
     thinking:
       type: enabled
       budget_tokens: 8192
-    cost: {input: 4, output: 16, cache_read: 1, currency: CNY}
+    cost: {input: 4, output: 21, cache_read: 0.7, currency: CNY}
 
 
 - provider_name: cerebras

klaude_code/core/agent_profile.py

@@ -130,12 +130,6 @@ def load_system_prompt(
 ) -> str:
     """Get system prompt content for the given model and sub-agent type."""
 
-    # For codex_oauth protocol, use dynamic prompts from GitHub (no additions).
-    if protocol == llm_param.LLMClientProtocol.CODEX_OAUTH:
-        from klaude_code.llm.openai_codex.prompt_sync import get_codex_instructions
-
-        return get_codex_instructions(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)

klaude_code/llm/anthropic/input.py

@@ -107,11 +107,20 @@ def _tool_blocks_to_message(blocks: list[BetaToolResultBlockParam]) -> BetaMessa
     }
 
 
+def _model_supports_unsigned_thinking(model_name: str | None) -> bool:
+    """Check if the model supports thinking blocks without signature (e.g., kimi, deepseek)."""
+    if not model_name:
+        return False
+    model_lower = model_name.lower()
+    return "kimi" in model_lower or "deepseek" in model_lower
+
+
 def _assistant_message_to_message(msg: message.AssistantMessage, model_name: str | None) -> BetaMessageParam:
     content: list[BetaContentBlockParam] = []
     current_thinking_content: str | None = None
     native_thinking_parts, _ = split_thinking_parts(msg, model_name)
     native_thinking_ids = {id(part) for part in native_thinking_parts}
+    supports_unsigned = _model_supports_unsigned_thinking(model_name)
 
     def _degraded_thinking_block(text: str) -> BetaTextBlockParam | None:
         stripped = text.strip()
@@ -125,11 +134,18 @@ def _assistant_message_to_message(msg: message.AssistantMessage, model_name: str
             },
         )
 
-    def _flush_thinking_as_text_block() -> None:
+    def _flush_thinking() -> None:
         nonlocal current_thinking_content
         if current_thinking_content is None:
             return
-        if block := _degraded_thinking_block(current_thinking_content):
+        if supports_unsigned:
+            content.append(
+                cast(
+                    BetaContentBlockParam,
+                    {"type": "thinking", "thinking": current_thinking_content},
+                )
+            )
+        elif block := _degraded_thinking_block(current_thinking_content):
             content.append(block)
         current_thinking_content = None
 
@@ -156,9 +172,17 @@ def _assistant_message_to_message(msg: message.AssistantMessage, model_name: str
                     )
                 )
                 current_thinking_content = None
+            elif supports_unsigned:
+                content.append(
+                    cast(
+                        BetaContentBlockParam,
+                        {"type": "thinking", "thinking": current_thinking_content or ""},
+                    )
+                )
+                current_thinking_content = None
             continue
 
-        _flush_thinking_as_text_block()
+        _flush_thinking()
         if isinstance(part, message.TextPart):
             content.append(cast(BetaTextBlockParam, {"type": "text", "text": part.text}))
         elif isinstance(part, message.ToolCallPart):
@@ -182,7 +206,7 @@ def _assistant_message_to_message(msg: message.AssistantMessage, model_name: str
                 )
             )
 
-    _flush_thinking_as_text_block()
+    _flush_thinking()
 
     return {"role": "assistant", "content": content}
 

klaude_code/llm/openai_codex/client.py

@@ -146,28 +146,6 @@ class CodexClient(LLMClientABC):
             )
         except (openai.OpenAIError, httpx.HTTPError) as e:
             error_message = f"{e.__class__.__name__} {e!s}"
-
-            # Check for invalid instruction error and invalidate prompt cache
-            if _is_invalid_instruction_error(e) and param.model_id:
-                _invalidate_prompt_cache_for_model(param.model_id)
-
             return error_llm_stream(metadata_tracker, error=error_message)
 
         return ResponsesLLMStream(stream, param=param, metadata_tracker=metadata_tracker)
-
-
-def _is_invalid_instruction_error(e: Exception) -> bool:
-    """Check if the error is related to invalid instructions."""
-    error_str = str(e).lower()
-    return "invalid instruction" in error_str or "invalid_instruction" in error_str
-
-
-def _invalidate_prompt_cache_for_model(model_id: str) -> None:
-    """Invalidate the cached prompt for a model to force refresh."""
-    from klaude_code.llm.openai_codex.prompt_sync import invalidate_cache
-
-    log_debug(
-        f"Invalidating prompt cache for model {model_id} due to invalid instruction error",
-        debug_type=DebugType.GENERAL,
-    )
-    invalidate_cache(model_id)

klaude_code/llm/openrouter/client.py

@@ -17,7 +17,7 @@ from klaude_code.llm.client import LLMClientABC, LLMStreamABC
 from klaude_code.llm.input_common import apply_config_defaults
 from klaude_code.llm.openai_compatible.input import convert_tool_schema
 from klaude_code.llm.openai_compatible.stream import OpenAILLMStream
-from klaude_code.llm.openrouter.input import convert_history_to_input, is_claude_model
+from klaude_code.llm.openrouter.input import convert_history_to_input, is_claude_model, is_xai_model
 from klaude_code.llm.openrouter.reasoning import ReasoningStreamHandler
 from klaude_code.llm.registry import register
 from klaude_code.llm.usage import MetadataTracker, error_llm_stream
@@ -70,6 +70,9 @@ def build_payload(
             f"{ANTHROPIC_BETA_FINE_GRAINED_TOOL_STREAMING},{ANTHROPIC_BETA_INTERLEAVED_THINKING}"
         )
 
+    if is_xai_model(param.model_id):
+        extra_body["plugins"] = [{"id": "web", "engine": "native"}]
+
     payload: CompletionCreateParamsStreaming = {
         "model": str(param.model_id),
         "tool_choice": "auto",

klaude_code/llm/openrouter/input.py

@@ -34,6 +34,12 @@ def is_gemini_model(model_name: str | None) -> bool:
     return model_name is not None and model_name.startswith("google/gemini")
 
 
+def is_xai_model(model_name: str | None) -> bool:
+    """Return True if the model name represents an xAI model."""
+
+    return model_name is not None and model_name.startswith("x-ai/")
+
+
 def _assistant_message_to_openrouter(
     msg: message.AssistantMessage, model_name: str | None
 ) -> chat.ChatCompletionMessageParam:

klaude_code/tui/components/developer.py

@@ -6,7 +6,7 @@ from klaude_code.tui.components.common import create_grid
 from klaude_code.tui.components.rich.theme import ThemeKey
 from klaude_code.tui.components.tools import render_path
 
-REMINDER_BULLET = "⧉"
+REMINDER_BULLET = " +"
 
 
 def need_render_developer_message(e: events.DeveloperMessageEvent) -> bool:

klaude_code/tui/components/errors.py

@@ -9,8 +9,7 @@ def render_error(error_msg: Text) -> RenderableType:
     """Render error with X mark for error events."""
     grid = create_grid()
     error_msg.style = ThemeKey.ERROR
-    error_msg.overflow = "ellipsis"
-    error_msg.no_wrap = True
+    error_msg.overflow = "fold"
     grid.add_row(Text("✘", style=ThemeKey.ERROR_BOLD), error_msg)
     return grid
 
@@ -19,7 +18,6 @@ def render_tool_error(error_msg: Text) -> RenderableType:
     """Render error with indent for tool results."""
     grid = create_grid()
     error_msg.style = ThemeKey.ERROR
-    error_msg.overflow = "ellipsis"
-    error_msg.no_wrap = True
+    error_msg.overflow = "fold"
     grid.add_row(Text(" "), error_msg)
     return grid

klaude_code/tui/components/metadata.py

@@ -32,9 +32,6 @@ def _render_task_metadata_block(
 
     # Second column: provider/model description / tokens / cost / …
     content = Text()
-    if metadata.provider is not None:
-        content.append_text(Text(metadata.provider.lower().replace(" ", "-"), style=ThemeKey.METADATA))
-        content.append_text(Text("/", style=ThemeKey.METADATA))
     content.append_text(Text(metadata.model_name, style=ThemeKey.METADATA))
     if metadata.description:
         content.append_text(Text(" ", style=ThemeKey.METADATA)).append_text(
@@ -129,7 +126,7 @@ def _render_task_metadata_block(
 
     if parts:
         content.append_text(Text(" ", style=ThemeKey.METADATA))
-        content.append_text(Text(" ", style=ThemeKey.METADATA).join(parts))
+        content.append_text(Text(" ", style=ThemeKey.METADATA_DIM).join(parts))
 
     grid.add_row(mark, content)
     return grid
@@ -140,15 +137,14 @@ def render_task_metadata(e: events.TaskMetadataEvent) -> RenderableType:
     renderables: list[RenderableType] = []
 
     has_sub_agents = len(e.metadata.sub_agent_task_metadata) > 0
-    # Use an extra space for the main agent mark to align with two-character marks (├─, └─)
-    main_mark_text = "●"
+    main_mark_text = "•"
     main_mark = Text(main_mark_text, style=ThemeKey.METADATA)
 
     renderables.append(_render_task_metadata_block(e.metadata.main_agent, mark=main_mark, show_context_and_time=True))
 
     # Render each sub-agent metadata block
     for meta in e.metadata.sub_agent_task_metadata:
-        sub_mark = Text("  └", style=ThemeKey.METADATA)
+        sub_mark = Text("  •", style=ThemeKey.METADATA)
         renderables.append(_render_task_metadata_block(meta, mark=sub_mark, show_context_and_time=True))
 
     # Add total cost line when there are sub-agents
@@ -165,9 +161,8 @@ def render_task_metadata(e: events.TaskMetadataEvent) -> RenderableType:
 
         currency_symbol = "¥" if currency == "CNY" else "$"
         total_line = Text.assemble(
-            ("  └", ThemeKey.METADATA),
-            (" Σ ", ThemeKey.METADATA),
-            ("total ", ThemeKey.METADATA),
+            ("  •", ThemeKey.METADATA),
+            (" total ", ThemeKey.METADATA),
             (currency_symbol, ThemeKey.METADATA),
             (f"{total_cost:.4f}", ThemeKey.METADATA),
         )

klaude_code/tui/components/rich/markdown.py

@@ -609,7 +609,11 @@ class MarkdownStream:
 
         live_text_to_set: Text | None = None
         if not final and MARKDOWN_STREAM_LIVE_REPAINT_ENABLED and self._live_sink is not None:
-            # Only update live area after we have rendered at least one stable block
+            # Only update the live area after we have rendered at least one stable block.
+            #
+            # This keeps the bottom "live" region anchored to stable scrollback, and
+            # avoids showing a live frame that would later need to be retroactively
+            # re-rendered once stable content exists.
             if not self._stable_rendered_lines:
                 return
             # When nothing is stable yet, we still want to show incremental output.

klaude_code/tui/components/rich/theme.py

@@ -262,8 +262,8 @@ def get_theme(theme: str | None = None) -> Themes:
                 ThemeKey.ERROR_DIM.value: "dim " + palette.red,
                 ThemeKey.INTERRUPT.value: palette.red,
                 # USER_INPUT
-                ThemeKey.USER_INPUT.value: f"{palette.magenta} on {palette.user_message_background}",
-                ThemeKey.USER_INPUT_PROMPT.value: f"bold {palette.magenta} on {palette.user_message_background}",
+                ThemeKey.USER_INPUT.value: f"{palette.cyan} on {palette.user_message_background}",
+                ThemeKey.USER_INPUT_PROMPT.value: f"bold {palette.cyan} on {palette.user_message_background}",
                 ThemeKey.USER_INPUT_AT_PATTERN.value: f"{palette.purple} on {palette.user_message_background}",
                 ThemeKey.USER_INPUT_SLASH_COMMAND.value: f"bold {palette.blue} on {palette.user_message_background}",
                 ThemeKey.USER_INPUT_SKILL.value: f"bold {palette.green} on {palette.user_message_background}",

klaude_code/tui/components/user_input.py

@@ -89,7 +89,7 @@ def render_user_input(content: str) -> RenderableType:
 
     return Padding(
         Group(*renderables),
-        pad=(0, 1),
+        pad=(1, 1),
         style=ThemeKey.USER_INPUT,
         expand=False,
     )

klaude_code/tui/input/completers.py

@@ -673,34 +673,68 @@ class _AtFilesCompleter(Completer):
         all_files_lower = self._git_file_list_lower or []
         kn = keyword_norm
 
-        # Bound per-keystroke work: stop scanning once enough matches are found.
+        # Bound per-keystroke work.
+        #
+        # Important: When the keyword is common (e.g. "tools"), truncating the
+        # scan purely by number of matching *files* can accidentally hide valid
+        # directory completions that appear later in the git path order.
+        #
+        # Example: multiple */tools/ directories under different parents.
+        file_quota = max_results
+        dir_quota = max_results
+        scan_cap = max(2000, max_results * 200)
+
+        keyword_stripped = keyword_norm.strip("/")
+        keyword_basename = os.path.basename(keyword_stripped)
+        explicit_parent = "/" in keyword_stripped
+
+        def dir_matches_keyword(dir_path: str) -> bool:
+            if not keyword_basename:
+                return False
+            if explicit_parent:
+                # When user typed an explicit parent segment, match against the
+                # whole directory path (not just basename).
+                return kn in f"{dir_path}/".lower()
+            # Otherwise prioritize directories by basename match.
+            return keyword_basename in os.path.basename(dir_path).lower()
+
         matching_files: list[str] = []
+        dir_list: list[str] = []
+        dir_seen: set[str] = set()
         scan_truncated = False
+        scanned = 0
+
         for p, pl in zip(all_files, all_files_lower, strict=False):
-            if kn in pl:
-                matching_files.append(p)
-                if len(matching_files) >= max_results:
+            scanned += 1
+            if kn not in pl:
+                if scanned >= scan_cap and (matching_files or dir_list):
                     scan_truncated = True
                     break
+                continue
+
+            if len(matching_files) < file_quota:
+                matching_files.append(p)
 
-        # Also include parent directories of matching files so users can
-        # complete into a folder, similar to fd's directory results.
-        dir_candidates: set[str] = set()
-        for p in matching_files[: max_results * 3]:
+            # Collect matching parent directories, walking upwards until repo root.
+            # This allows completing into directories like "image/tools/" even
+            # when the matching file is nested deeper.
             parent = os.path.dirname(p)
             while parent and parent != ".":
-                dir_candidates.add(f"{parent}/")
+                if dir_matches_keyword(parent):
+                    cand = f"{parent}/"
+                    if cand not in dir_seen:
+                        dir_seen.add(cand)
+                        dir_list.append(cand)
+                        if len(dir_list) >= dir_quota:
+                            break
                 parent = os.path.dirname(parent)
 
-        dir_list = sorted(dir_candidates)
-        dir_truncated = False
-        if len(dir_list) > max_results:
-            dir_list = dir_list[:max_results]
-            dir_truncated = True
+            if len(matching_files) >= file_quota and len(dir_list) >= dir_quota:
+                scan_truncated = True
+                break
 
-        candidates = matching_files + dir_list
-        truncated = scan_truncated or dir_truncated
-        return candidates, truncated
+        candidates = dir_list + matching_files
+        return candidates, scan_truncated
 
     def _get_git_repo_root(self, cwd: Path) -> Path | None:
         if not self._has_cmd("git"):

klaude_code/tui/input/images.py

@@ -28,6 +28,9 @@ from klaude_code.protocol.message import ImageFilePart
 
 IMAGE_SUFFIXES = frozenset({".png", ".jpg", ".jpeg", ".gif", ".webp"})
 
+# Claude API limit is 5MB, we use 4.5MB to have some margin
+MAX_IMAGE_SIZE_BYTES = 4_500_000
+
 IMAGE_MARKER_RE = re.compile(r'\[image (?P<path>"[^"]+"|[^\]]+)\]')
 
 
@@ -159,6 +162,127 @@ def _grab_clipboard_image(dest_path: Path) -> bool:
         return _grab_clipboard_image_linux(dest_path)
 
 
+# ---------------------------------------------------------------------------
+# Image resizing for size limits
+# ---------------------------------------------------------------------------
+
+
+def _get_image_dimensions_macos(path: Path) -> tuple[int, int] | None:
+    """Get image dimensions using sips on macOS."""
+    try:
+        result = subprocess.run(
+            ["sips", "-g", "pixelWidth", "-g", "pixelHeight", str(path)],
+            capture_output=True,
+            text=True,
+        )
+        if result.returncode != 0:
+            return None
+        width = height = 0
+        for line in result.stdout.splitlines():
+            if "pixelWidth" in line:
+                width = int(line.split(":")[-1].strip())
+            elif "pixelHeight" in line:
+                height = int(line.split(":")[-1].strip())
+        if width > 0 and height > 0:
+            return (width, height)
+    except (OSError, ValueError):
+        pass
+    return None
+
+
+def _resize_image_macos(path: Path, scale: float) -> bool:
+    """Resize image using sips on macOS. Modifies file in place."""
+    dims = _get_image_dimensions_macos(path)
+    if dims is None:
+        return False
+    new_width = max(1, int(dims[0] * scale))
+    try:
+        result = subprocess.run(
+            ["sips", "--resampleWidth", str(new_width), str(path)],
+            capture_output=True,
+        )
+        return result.returncode == 0
+    except OSError:
+        return False
+
+
+def _resize_image_linux(path: Path, scale: float) -> bool:
+    """Resize image using ImageMagick convert on Linux."""
+    if not shutil.which("convert"):
+        return False
+    percent = int(scale * 100)
+    try:
+        result = subprocess.run(
+            ["convert", str(path), "-resize", f"{percent}%", str(path)],
+            capture_output=True,
+        )
+        return result.returncode == 0
+    except OSError:
+        return False
+
+
+def _resize_image_windows(path: Path, scale: float) -> bool:
+    """Resize image using PowerShell on Windows."""
+    script = f'''
+    Add-Type -AssemblyName System.Drawing
+    $img = [System.Drawing.Image]::FromFile("{path}")
+    $newWidth = [int]($img.Width * {scale})
+    $newHeight = [int]($img.Height * {scale})
+    $bmp = New-Object System.Drawing.Bitmap($newWidth, $newHeight)
+    $graphics = [System.Drawing.Graphics]::FromImage($bmp)
+    $graphics.DrawImage($img, 0, 0, $newWidth, $newHeight)
+    $img.Dispose()
+    $bmp.Save("{path}", [System.Drawing.Imaging.ImageFormat]::Png)
+    $bmp.Dispose()
+    $graphics.Dispose()
+    Write-Output "ok"
+    '''
+    try:
+        result = subprocess.run(
+            ["powershell", "-Command", script],
+            capture_output=True,
+            text=True,
+        )
+        return result.returncode == 0 and "ok" in result.stdout
+    except OSError:
+        return False
+
+
+def _resize_image(path: Path, scale: float) -> bool:
+    """Resize image by scale factor. Modifies file in place."""
+    if sys.platform == "darwin":
+        return _resize_image_macos(path, scale)
+    elif sys.platform == "win32":
+        return _resize_image_windows(path, scale)
+    else:
+        return _resize_image_linux(path, scale)
+
+
+def _ensure_image_size_limit(path: Path, max_bytes: int = MAX_IMAGE_SIZE_BYTES) -> None:
+    """Resize image if it exceeds the size limit. Modifies file in place."""
+    try:
+        current_size = path.stat().st_size
+        if current_size <= max_bytes:
+            return
+
+        # Calculate scale factor based on size ratio
+        # We use sqrt because area scales with square of linear dimensions
+        scale = (max_bytes / current_size) ** 0.5
+        # Be a bit more aggressive to ensure we get under the limit
+        scale *= 0.9
+
+        for _ in range(5):  # Max 5 resize attempts
+            if not _resize_image(path, scale):
+                return
+            new_size = path.stat().st_size
+            if new_size <= max_bytes:
+                return
+            # Still too large, reduce more
+            scale = 0.8
+    except OSError:
+        pass
+
+
 def capture_clipboard_tag() -> str | None:
     """Capture an image from clipboard and return an [image ...] marker."""
 
@@ -174,6 +298,9 @@ def capture_clipboard_tag() -> str | None:
     if not _grab_clipboard_image(path):
         return None
 
+    # Resize if image exceeds size limit
+    _ensure_image_size_limit(path)
+
     return format_image_marker(str(path))
 
 

klaude_code/tui/input/key_bindings.py

@@ -76,6 +76,60 @@ def create_key_bindings(
     term_program = os.environ.get("TERM_PROGRAM", "").lower()
     swallow_next_control_j = False
 
+    def _history_backward_cursor_to_start(buf: Buffer) -> None:
+        """Switch to previous history entry and place cursor at absolute start.
+
+        prompt_toolkit's default `Buffer.history_backward()` moves the cursor to
+        the end of the (possibly multi-line) history entry. That makes it hard
+        to keep pressing Up to continue cycling history, because subsequent Up
+        key presses start moving within the multi-line buffer.
+        """
+
+        try:
+            before = int(buf.working_index)  # type: ignore[reportUnknownMemberType]
+        except Exception:
+            before = None
+
+        buf.history_backward()
+
+        try:
+            after = int(buf.working_index)  # type: ignore[reportUnknownMemberType]
+        except Exception:
+            after = None
+
+        if before is not None and after is not None and before == after:
+            return
+
+        with contextlib.suppress(Exception):
+            buf.cursor_position = 0  # type: ignore[reportUnknownMemberType]
+
+    def _history_forward_cursor_to_end(buf: Buffer) -> None:
+        """Switch to next history entry and place cursor at absolute end.
+
+        prompt_toolkit's default `Buffer.history_forward()` moves the cursor to
+        the end of the *first* line. For our multiline REPL, it's more useful to
+        land at the end so that pressing Down keeps cycling through history.
+        """
+
+        try:
+            before = int(buf.working_index)  # type: ignore[reportUnknownMemberType]
+        except Exception:
+            before = None
+
+        buf.history_forward()
+
+        try:
+            after = int(buf.working_index)  # type: ignore[reportUnknownMemberType]
+        except Exception:
+            after = None
+
+        if before is not None and after is not None and before == after:
+            return
+
+        with contextlib.suppress(Exception):
+            text = buf.text  # type: ignore[reportUnknownMemberType]
+            buf.cursor_position = len(text)  # type: ignore[reportUnknownMemberType]
+
     def _is_bash_mode_text(text: str) -> bool:
         return text.startswith(("!", "！"))
 
@@ -151,6 +205,20 @@ def create_key_bindings(
         except Exception:
             return False
 
+    def _current_cursor_row() -> int:
+        try:
+            doc = get_app().current_buffer.document
+            return int(doc.cursor_position_row)
+        except Exception:
+            return 0
+
+    def _current_line_count() -> int:
+        try:
+            doc = get_app().current_buffer.document
+            return int(doc.line_count)
+        except Exception:
+            return 1
+
     def _move_cursor_visually_within_wrapped_line(event: KeyPressEvent, *, delta_visible_y: int) -> None:
         """Move the cursor Up/Down by one wrapped screen row, keeping column."""
 
@@ -493,6 +561,32 @@ def create_key_bindings(
     def _(event: KeyPressEvent) -> None:
         _move_cursor_visually_within_wrapped_line(event, delta_visible_y=1)
 
+    @kb.add(
+        "up",
+        filter=enabled
+        & ~has_completions
+        & ~is_searching
+        & Condition(lambda: not _can_move_cursor_visually_within_wrapped_line(delta_visible_y=-1))
+        & Condition(lambda: _current_cursor_row() == 0),
+        eager=True,
+    )
+    def _(event: KeyPressEvent) -> None:
+        """Up on first logical line: switch history and keep caret at start."""
+        _history_backward_cursor_to_start(event.current_buffer)
+
+    @kb.add(
+        "down",
+        filter=enabled
+        & ~has_completions
+        & ~is_searching
+        & Condition(lambda: not _can_move_cursor_visually_within_wrapped_line(delta_visible_y=1))
+        & Condition(lambda: _current_cursor_row() >= (_current_line_count() - 1)),
+        eager=True,
+    )
+    def _(event: KeyPressEvent) -> None:
+        """Down on last logical line: switch history and keep caret at end."""
+        _history_forward_cursor_to_end(event.current_buffer)
+
     @kb.add("c-j", filter=enabled)
     def _(event: KeyPressEvent) -> None:
         nonlocal swallow_next_control_j
@@ -617,11 +711,11 @@ def create_key_bindings(
     @kb.add("escape", "up", filter=enabled & ~has_completions)
     def _(event: KeyPressEvent) -> None:
         """Option+Up switches to previous history entry."""
-        event.current_buffer.history_backward()
+        _history_backward_cursor_to_start(event.current_buffer)
 
     @kb.add("escape", "down", filter=enabled & ~has_completions)
     def _(event: KeyPressEvent) -> None:
         """Option+Down switches to next history entry."""
-        event.current_buffer.history_forward()
+        _history_forward_cursor_to_end(event.current_buffer)
 
     return kb

klaude_code/tui/input/prompt_toolkit.py

@@ -61,7 +61,7 @@ COMPLETION_SELECTED_DARK_BG = "ansigreen"
 COMPLETION_SELECTED_LIGHT_BG = "ansigreen"
 COMPLETION_SELECTED_UNKNOWN_BG = "ansigreen"
 COMPLETION_MENU = "ansibrightblack"
-INPUT_PROMPT_STYLE = "ansimagenta bold"
+INPUT_PROMPT_STYLE = "ansicyan bold"
 INPUT_PROMPT_BASH_STYLE = "ansigreen bold"
 PLACEHOLDER_TEXT_STYLE_DARK_BG = "fg:#5a5a5a"
 PLACEHOLDER_TEXT_STYLE_LIGHT_BG = "fg:#7a7a7a"

klaude_code/tui/machine.py

@@ -736,6 +736,7 @@ class DisplayStateMachine:
                     and not e.has_structured_output
                     and s.assistant_char_count == 0
                     and e.task_result.strip()
+                    and e.task_result.strip().lower() not in {"task cancelled", "task canceled"}
                 ):
                     cmds.append(StartAssistantStream(session_id=e.session_id))
                     cmds.append(AppendAssistant(session_id=e.session_id, content=e.task_result))

klaude_code/tui/renderer.py

@@ -734,6 +734,9 @@ class TUICommandRenderer:
                         self._sub_agent_thinking_buffers[session_id] = ""
                     elif not self._thinking_stream.is_active:
                         self._thinking_stream.start(self._new_thinking_mdstream())
+                        if not self._replay_mode:
+                            self._thinking_stream.append("Thinking…  \n")
+                            self._thinking_stream.render(transform=c_thinking.normalize_thinking_content)
                 case AppendThinking(session_id=session_id, content=content):
                     if self.is_sub_agent_session(session_id):
                         if session_id in self._sub_agent_thinking_buffers:

{klaude_code-2.10.1.dist-info → klaude_code-2.10.3.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: klaude-code
-Version: 2.10.1
+Version: 2.10.3
 Summary: Minimal code agent CLI
 Requires-Dist: anthropic>=0.66.0
 Requires-Dist: chardet>=5.2.0

{klaude_code-2.10.1.dist-info → klaude_code-2.10.3.dist-info}/RECORD

@@ -30,7 +30,7 @@ klaude_code/cli/main.py,sha256=bQijkLDTYtjFUdU1t-dGCzOilo7mcotIQefleaF9or8,13072
 klaude_code/cli/self_update.py,sha256=1xdG9ifvRZQDSx6RAtSSgXmw9hZNXMLvqC2zu4bS-GY,2622
 klaude_code/config/__init__.py,sha256=Qe1BeMekBfO2-Zd30x33lB70hdM1QQZGrp4DbWSQ-II,353
 klaude_code/config/assets/__init__.py,sha256=uMUfmXT3I-gYiI-HVr1DrE60mx5cY1o8V7SYuGqOmvY,32
-klaude_code/config/assets/builtin_config.yaml,sha256=EYGgBMAVtgUKGHZcfkm4sCfSB1-fBx83q10MGVnNz8I,9121
+klaude_code/config/assets/builtin_config.yaml,sha256=5S5dEox3cr5bLWk6N3z6njIP5RLe-sHE2vAvkKPBCqo,9112
 klaude_code/config/builtin_config.py,sha256=OG5VERUHo3tSojgFXfNDV6pAHNOh3kO-xFHpvTr-cpc,1786
 klaude_code/config/config.py,sha256=otBvsUkvI-2fpZzx9fO6SPnCek7FI7kufvAnGIQqTz8,27084
 klaude_code/config/model_matcher.py,sha256=3IlLU5h3NDh_bURbCW-PV027C3irG3hyitwj1cj99Ig,6179
@@ -39,7 +39,7 @@ klaude_code/config/thinking.py,sha256=5uVM0cFUJZBBsBGGdPG-tjdiNwZ-GFeWOBBWIdSPFv
 klaude_code/const.py,sha256=VCK3HgZJZO6jcYz6U2rcHS7W-n4oyKYg9AC6eTB4HIQ,11575
 klaude_code/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 klaude_code/core/agent.py,sha256=GrIg22nfoq1c90UHyEfU_bh46vtXTCo4bLezb-3mGNo,4120
-klaude_code/core/agent_profile.py,sha256=8fEiRI2emEFVlwsCEtNaZ90UUp1A7MKhzOnzQK1IOGI,11507
+klaude_code/core/agent_profile.py,sha256=wfde6zMez5abh5O04e7BzBlbLhHwd23pn4FBijOHxQg,11231
 klaude_code/core/bash_mode.py,sha256=BUy2_uOcJ8bO89t_QCwtw6GlHt7vd1t62sN5iWk-UuQ,9250
 klaude_code/core/compaction/AGENTS.md,sha256=KZR5lxe4jVAbT5K9PxbZcHWI1UwsppbGmxIfCdHYr7Q,3684
 klaude_code/core/compaction/__init__.py,sha256=CvidYx3sX0IZAa4pifX9jrQSkg4Nib7PKrcaOHswF60,329
@@ -111,7 +111,7 @@ klaude_code/core/turn.py,sha256=TK4-ZGmRo_qa21XRklGfEMnfhFGR20wnKjnvlf31HBQ,1888
 klaude_code/llm/__init__.py,sha256=b4AsqnrMIs0a5qR_ti6rZcHwFzAReTwOW96EqozEoSo,287
 klaude_code/llm/anthropic/__init__.py,sha256=PWETvaeNAAX3ue0ww1uRUIxTJG0RpWiutkn7MlwKxBs,67
 klaude_code/llm/anthropic/client.py,sha256=RpYw4UQnhLzAsp6i-FU7cDW4deqngdAoQaTPGnCeO5U,17346
-klaude_code/llm/anthropic/input.py,sha256=ObcRuQEwdiFoI7bjBzHv5bClgIYEw23bJMgHudT92C4,9923
+klaude_code/llm/anthropic/input.py,sha256=_ng3WtRWbyhmMuFDNCnJoEi-k-tEOy9ukGbzvvgZ3vQ,10819
 klaude_code/llm/antigravity/__init__.py,sha256=TuK_k4mJpBQVBCfhRFQvVLeGtHRU8_2wXO2lRC-OB9o,71
 klaude_code/llm/antigravity/client.py,sha256=XnOBw7QPNA8NRrZIdNrOD1qLrvWLyI82PL7kIolgdic,20120
 klaude_code/llm/antigravity/input.py,sha256=WblkpS4DxnqHLb9oXGS3uDlgbzdP4GKWmC9L2HEtRwk,9463
@@ -127,8 +127,7 @@ klaude_code/llm/image.py,sha256=TJzduNtSmoJLizTUDo0EjbovqOeaXkvwnRyMfuzVVUQ,4512
 klaude_code/llm/input_common.py,sha256=htV8YHDlQ7fx2qW_D9e_BgxWlyBmZwAtNxnM15uj9gQ,7563
 klaude_code/llm/json_stable.py,sha256=FrgJbJ33YrAkzl5iPOWLolFjYFZMuT97sbOegeXN7GE,1288
 klaude_code/llm/openai_codex/__init__.py,sha256=Dx2MxrZEHQZ8gsJmM05dE4qqak4F0zjwfXd5v8pIXaE,139
-klaude_code/llm/openai_codex/client.py,sha256=bpTKtp3Qum-c6OBQAZHJ9LKSzgToHOig5M03568aX6I,6248
-klaude_code/llm/openai_codex/prompt_sync.py,sha256=u9RBuVpfRJGB5cdpI_OJkJVSL33s4pZNxbiCIm1l70Y,8767
+klaude_code/llm/openai_codex/client.py,sha256=wIOfCsB1F_b9D6ymq-nyCIXAuuBTz6-CMUxD8NiuvkQ,5403
 klaude_code/llm/openai_compatible/__init__.py,sha256=ACGpnki7k53mMcCl591aw99pm9jZOZk0ghr7atOfNps,81
 klaude_code/llm/openai_compatible/client.py,sha256=oiRmqiqNJ-clLk6pK1LTOAyx1svmNBg1sNuT8uN3l5g,4717
 klaude_code/llm/openai_compatible/input.py,sha256=UfEu2c866QuVMFTY3GEPhuKV7QcCgxBwNYqGT88E_8E,2834
@@ -137,8 +136,8 @@ klaude_code/llm/openai_responses/__init__.py,sha256=WsiyvnNiIytaYcaAqNiB8GI-5zcp
 klaude_code/llm/openai_responses/client.py,sha256=qmQNcHZUBzvjQZZQkWh2f3oUYC9Lbx_dZD7KwnXbPPI,16362
 klaude_code/llm/openai_responses/input.py,sha256=eEeSHlGYY5VYGpj2DIVm8qjH2cXo-JaODT_SjFtS6Nw,9264
 klaude_code/llm/openrouter/__init__.py,sha256=_As8lHjwj6vapQhLorZttTpukk5ZiCdhFdGT38_ASPo,69
-klaude_code/llm/openrouter/client.py,sha256=P_mDQK4_i1MLF0jK4p_bKhh15ACXgw6Ie0rUrtbfsdM,5738
-klaude_code/llm/openrouter/input.py,sha256=Z_Cf6TnMZ5KQNJ0E5IIDCKK2OWlzi8IW0S5A72BBGT0,6176
+klaude_code/llm/openrouter/client.py,sha256=pBQG3JIftEyDpWNo_rT7cDaTAmeUxTktyRRTcZNq9Hg,5858
+klaude_code/llm/openrouter/input.py,sha256=VqKPWCUb-BlFwNs1uldDD_dGPJzwskG301uULaqm958,6363
 klaude_code/llm/openrouter/reasoning.py,sha256=u7ccfnGxJ4Ws8P3X5FW91d8HXie29JjeWz0hZ1r0oFg,3320
 klaude_code/llm/partial_message.py,sha256=-sjlpV-et4ViBtBpdtihK5QBjAlwS47-mBpVbRQPP3s,142
 klaude_code/llm/registry.py,sha256=oYObtCgWQxL2eFfmvpYLqMhKvYD3JoszzpvuC08sRQQ,2177
@@ -204,35 +203,35 @@ klaude_code/tui/components/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJ
 klaude_code/tui/components/bash_syntax.py,sha256=a8JMn6rFlJXcx7-WbsxIVwqtWQoI9krLUkj0imVC20I,7632
 klaude_code/tui/components/command_output.py,sha256=SnNiWdMAK37wd6PtfSWsLW2UutNfkpTKDbFmw2wu3LA,3688
 klaude_code/tui/components/common.py,sha256=dhUYLVVOSKxg5GCoS4eyUeKZ3E8Kpt4nqft4njuvPaI,4698
-klaude_code/tui/components/developer.py,sha256=m6gcnLyLoSSw3wpwYQZ_yzaT9zgXRgNzTpjZt8Yo21o,5488
+klaude_code/tui/components/developer.py,sha256=jJPx3F6T8pBc1ubgJaAeOE5dpapBi8jfXDZA_PZdQxg,5487
 klaude_code/tui/components/diffs.py,sha256=vwllnYBxC5xGjfKU3uIkCjcupr9nrjRjvvj0tg0_MQA,3085
-klaude_code/tui/components/errors.py,sha256=fSojNfRceB6eE7cyJHfwGt5Ru0OYp63fCJ-W6-3SSYs,799
+klaude_code/tui/components/errors.py,sha256=uqxVDqZLV6-Gvk46JWPFy2dc_bXHxgg_FEwrUX71jyg,733
 klaude_code/tui/components/mermaid_viewer.py,sha256=zI1FBuX6Ionx38KqkzhOIQ9tFzd7REPbjW1iqSiNrec,3086
-klaude_code/tui/components/metadata.py,sha256=JzHMi3fr_YWQmachSRukBTYO3iHvPFAXaA3f2SRyD54,7286
+klaude_code/tui/components/metadata.py,sha256=bkN2WOHFkNSEh5Ajup_976WvcU83lR6lNG03MaH9YYg,6944
 klaude_code/tui/components/rich/__init__.py,sha256=zEZjnHR3Fnv_sFMxwIMjoJfwDoC4GRGv3lHJzAGRq_o,236
 klaude_code/tui/components/rich/cjk_wrap.py,sha256=eMqBxftUtll7zrytUb9WtJ6naYLyax0W4KJRpGwWulM,7602
 klaude_code/tui/components/rich/code_panel.py,sha256=SOdyfHBZNB4gAWIbnN_enhHB1EWxw8Hxiafx6yjwdJo,5544
 klaude_code/tui/components/rich/live.py,sha256=xiMT6dPsxM_jaazddKrV9CMJQWwpe2t9OdjffHvo1JU,2821
-klaude_code/tui/components/rich/markdown.py,sha256=vas08cxz8fEU2PlTUNrgmF-zIXtFoMxP-Jefc8FSGuE,25276
+klaude_code/tui/components/rich/markdown.py,sha256=8BXTyR0UMQ6Ut62HMeFPTB_8QLAvYDtk9SbGZ6H1hkY,25518
 klaude_code/tui/components/rich/quote.py,sha256=u6sBmGdp0ckaZLw_XgJk7iHW4zxnWikUaB3GX2tkhlM,5375
 klaude_code/tui/components/rich/status.py,sha256=hSvMwEguF2DfHH3ISR0bmDg58zAnM3CTJLcRff_rtrg,14791
-klaude_code/tui/components/rich/theme.py,sha256=cp7PrhxIgOLu_AjHBOf3wCZewzngOAZDFQQmiO0sAbY,17084
+klaude_code/tui/components/rich/theme.py,sha256=eg9EIcjp2qysQi1s_D2r7QYtclOe3CMXJjf_bwTXAGM,17078
 klaude_code/tui/components/sub_agent.py,sha256=8XTWsTi9mfbNLMD8SZ__nZQmBf81rW-NWpuOT-sFbv8,4723
 klaude_code/tui/components/thinking.py,sha256=yVzY7BbcdDwB4XKi9i2sp3cKREirvmlMiEID74b-5_0,955
 klaude_code/tui/components/tools.py,sha256=MqW-_a4gtMnCR5N8APJYtN2TfSvuUdvo2WFEGm3Qa9A,27327
-klaude_code/tui/components/user_input.py,sha256=fYvdoc3gMyv90l2sQvSs4ha0wILu71rBAcwbzQPMMVc,3659
+klaude_code/tui/components/user_input.py,sha256=u3i6y3aKvUv7at39eeT2mEi0cgOsrha91FH-lUsUetw,3659
 klaude_code/tui/components/welcome.py,sha256=Ahkhg0dsSqy17pKLOp_5UZWn9vysr68T3Y-jB40yWsA,5303
 klaude_code/tui/display.py,sha256=VbBmNjGxp6R4oLjC0FrjXkr0INWSjfXlWq1mSVFGyMU,3933
 klaude_code/tui/input/AGENTS.md,sha256=2RBLz7H0JbUJv6OBzeadLOlGUF5EBqvtwTGBf6nZuN0,1633
 klaude_code/tui/input/__init__.py,sha256=wLbjqBrvP6fmbGtbKe9Wp12yxhse0faVLOxtoWua_1E,353
-klaude_code/tui/input/completers.py,sha256=MJO1nBq0V5jDbGw_o4Ab5WLVD1ns5plJRI3cIYnGfHs,33154
+klaude_code/tui/input/completers.py,sha256=vs2Lq0_PIO1FszV968gJobeqptM5Jz4YCv-MpPqHakE,34559
 klaude_code/tui/input/drag_drop.py,sha256=oyKtrHCyUiGiMLEXpsDTnTnAKJ1_xrvVkrASOiG8O4g,3974
-klaude_code/tui/input/images.py,sha256=ft2AaOg1Figdm1t_NNoBCGdp20silYXGw-m9XKDd9GU,6996
-klaude_code/tui/input/key_bindings.py,sha256=Gpc-VhQh-h431cnetheV0HF9QGcFyy82QrHWUAYF-LA,26633
+klaude_code/tui/input/images.py,sha256=dfM_BvlQE6F8BYEr2YBNTsWKd57_Jx-2wgFtv5ZSXSk,11202
+klaude_code/tui/input/key_bindings.py,sha256=6oxWdTdklFqLDOk6ZpyKijgrnQO4B9hSqgMfT2kdjXs,30032
 klaude_code/tui/input/paste.py,sha256=kELg5jC0WdBXWHJUsEjIhZ67KCvHMbN1XzyGmevVSNM,1888
-klaude_code/tui/input/prompt_toolkit.py,sha256=aE1pVK1G96DWERc2APIXq4WmSPOdJDRHTgnyyZXp_u4,30663
-klaude_code/tui/machine.py,sha256=iRTBNx22eozmT9gyya9QbpDDcmqst07bUAa2vWMDZAA,32482
-klaude_code/tui/renderer.py,sha256=E9SSuxWlQDMA2_4ENZHpwfB54ZVpkFMJYPe_WQSzq0E,34652
+klaude_code/tui/input/prompt_toolkit.py,sha256=gohHzvAWp7Wh06-VjxTZpPojPq_0lVW-KeoTUPRsgSQ,30660
+klaude_code/tui/machine.py,sha256=STwhoBw_HwpkDHDts7CrwtUN211OYaySQ0jy8r7fa08,32579
+klaude_code/tui/renderer.py,sha256=b59_Pr17seKOcQlHDEH4CMWKNWzdpMzmUhqUDqLvQyU,34884
 klaude_code/tui/runner.py,sha256=ZADAH28Iu1DU-KDCggEyvJiM_LTbN1sjPEaQsuBNTbc,13111
 klaude_code/tui/terminal/__init__.py,sha256=GIMnsEcIAGT_vBHvTlWEdyNmAEpruyscUA6M_j3GQZU,1412
 klaude_code/tui/terminal/color.py,sha256=6SJR2RA8cqJINNoRz65w0HL3x9g46ydIvDOGWMeNnQU,7195
@@ -250,7 +249,7 @@ klaude_code/ui/debug_mode.py,sha256=ZvqbOx4c_rUerMbEZzOfcbNf9leqEDFjqJUlALtzF9Y,
 klaude_code/ui/terminal/__init__.py,sha256=5OeAzr994r8-peWsLON0iXsAvJ2pexwMp36JY7FKGDc,179
 klaude_code/ui/terminal/title.py,sha256=lCk1dKk7fIe5Fb-FRU9P4ktVEfBmT3ac3wICYmC4mGE,1229
 klaude_code/update.py,sha256=QER816AZe9u3RhRvP0Z37Jh2Ch5RLy9PREyDsI0e1dA,4480
-klaude_code-2.10.1.dist-info/WHEEL,sha256=eh7sammvW2TypMMMGKgsM83HyA_3qQ5Lgg3ynoecH3M,79
-klaude_code-2.10.1.dist-info/entry_points.txt,sha256=kkXIXedaTOtjXPr2rVjRVVXZYlFUcBHELaqmyVlWUFA,92
-klaude_code-2.10.1.dist-info/METADATA,sha256=zM4Cjq5SylCPMJsKpmsr0b0eIknIojYMQPSE3ypB_6U,10120
-klaude_code-2.10.1.dist-info/RECORD,,
+klaude_code-2.10.3.dist-info/WHEEL,sha256=eh7sammvW2TypMMMGKgsM83HyA_3qQ5Lgg3ynoecH3M,79
+klaude_code-2.10.3.dist-info/entry_points.txt,sha256=kkXIXedaTOtjXPr2rVjRVVXZYlFUcBHELaqmyVlWUFA,92
+klaude_code-2.10.3.dist-info/METADATA,sha256=p1r3t9GoNu-eBxdkTC3cKWiS7kAbXF2BXyD1ALViTm0,10120
+klaude_code-2.10.3.dist-info/RECORD,,

klaude_code/llm/openai_codex/prompt_sync.py

@@ -1,237 +0,0 @@
-"""Dynamic prompt synchronization from OpenAI Codex GitHub repository."""
-
-import json
-import time
-from functools import cache
-from importlib.resources import files
-from pathlib import Path
-from typing import Any, Literal
-
-import httpx
-
-from klaude_code.log import DebugType, log_debug
-
-GITHUB_API_RELEASES = "https://api.github.com/repos/openai/codex/releases/latest"
-GITHUB_HTML_RELEASES = "https://github.com/openai/codex/releases/latest"
-GITHUB_RAW_BASE = "https://raw.githubusercontent.com/openai/codex"
-
-CACHE_DIR = Path.home() / ".klaude" / "codex-prompts"
-CACHE_TTL_SECONDS = 24 * 60 * 60  # 24 hours
-
-type ModelFamily = Literal["gpt-5.2-codex", "codex-max", "codex", "gpt-5.2", "gpt-5.1"]
-
-PROMPT_FILES: dict[ModelFamily, str] = {
-    "gpt-5.2-codex": "gpt-5.2-codex_prompt.md",
-    "codex-max": "gpt-5.1-codex-max_prompt.md",
-    "codex": "gpt_5_codex_prompt.md",
-    "gpt-5.2": "gpt_5_2_prompt.md",
-    "gpt-5.1": "gpt_5_1_prompt.md",
-}
-
-CACHE_FILES: dict[ModelFamily, str] = {
-    "gpt-5.2-codex": "gpt-5.2-codex-instructions.md",
-    "codex-max": "codex-max-instructions.md",
-    "codex": "codex-instructions.md",
-    "gpt-5.2": "gpt-5.2-instructions.md",
-    "gpt-5.1": "gpt-5.1-instructions.md",
-}
-
-
-@cache
-def _load_bundled_prompt(prompt_path: str) -> str:
-    """Load bundled prompt from package resources."""
-    return files("klaude_code.core").joinpath(prompt_path).read_text(encoding="utf-8").strip()
-
-
-class CacheMetadata:
-    def __init__(self, etag: str | None, tag: str, last_checked: int, url: str):
-        self.etag = etag
-        self.tag = tag
-        self.last_checked = last_checked
-        self.url = url
-
-    def to_dict(self) -> dict[str, str | int | None]:
-        return {
-            "etag": self.etag,
-            "tag": self.tag,
-            "last_checked": self.last_checked,
-            "url": self.url,
-        }
-
-    @classmethod
-    def from_dict(cls, data: dict[str, object]) -> "CacheMetadata":
-        etag = data.get("etag")
-        last_checked = data.get("last_checked")
-        return cls(
-            etag=etag if isinstance(etag, str) else None,
-            tag=str(data.get("tag", "")),
-            last_checked=int(last_checked) if isinstance(last_checked, int | float) else 0,
-            url=str(data.get("url", "")),
-        )
-
-
-def get_model_family(model: str) -> ModelFamily:
-    """Determine model family from model name."""
-    if "gpt-5.2-codex" in model or "gpt 5.2 codex" in model:
-        return "gpt-5.2-codex"
-    if "codex-max" in model:
-        return "codex-max"
-    if "codex" in model or model.startswith("codex-"):
-        return "codex"
-    if "gpt-5.2" in model:
-        return "gpt-5.2"
-    return "gpt-5.1"
-
-
-def _get_latest_release_tag(client: httpx.Client) -> str:
-    """Get latest release tag from GitHub."""
-    try:
-        response = client.get(GITHUB_API_RELEASES)
-        if response.status_code == 200:
-            data: dict[str, Any] = response.json()
-            tag_name: Any = data.get("tag_name")
-            if isinstance(tag_name, str):
-                return tag_name
-    except httpx.HTTPError:
-        pass
-
-    # Fallback: follow redirect from releases/latest
-    response = client.get(GITHUB_HTML_RELEASES, follow_redirects=True)
-    if response.status_code == 200:
-        final_url = str(response.url)
-        if "/tag/" in final_url:
-            parts = final_url.split("/tag/")
-            if len(parts) > 1 and "/" not in parts[-1]:
-                return parts[-1]
-
-    raise RuntimeError("Failed to determine latest release tag from GitHub")
-
-
-def _load_cache_metadata(meta_file: Path) -> CacheMetadata | None:
-    if not meta_file.exists():
-        return None
-    try:
-        data = json.loads(meta_file.read_text())
-        return CacheMetadata.from_dict(data)
-    except (json.JSONDecodeError, ValueError):
-        return None
-
-
-def _save_cache_metadata(meta_file: Path, metadata: CacheMetadata) -> None:
-    meta_file.parent.mkdir(parents=True, exist_ok=True)
-    meta_file.write_text(json.dumps(metadata.to_dict(), indent=2))
-
-
-def get_codex_instructions(model: str = "gpt-5.1-codex", force_refresh: bool = False) -> str:
-    """Get Codex instructions for the given model.
-
-    Args:
-        model: Model name to get instructions for.
-        force_refresh: If True, bypass cache TTL and fetch fresh instructions.
-
-    Returns:
-        The Codex system prompt instructions.
-    """
-    model_family = get_model_family(model)
-    prompt_file = PROMPT_FILES[model_family]
-    cache_file = CACHE_DIR / CACHE_FILES[model_family]
-    meta_file = CACHE_DIR / f"{CACHE_FILES[model_family].replace('.md', '-meta.json')}"
-
-    # Check cache unless force refresh
-    if not force_refresh:
-        metadata = _load_cache_metadata(meta_file)
-        if metadata and cache_file.exists():
-            age = int(time.time()) - metadata.last_checked
-            if age < CACHE_TTL_SECONDS:
-                log_debug(f"Using cached {model_family} instructions (age: {age}s)", debug_type=DebugType.GENERAL)
-                return cache_file.read_text()
-
-    try:
-        with httpx.Client(timeout=30.0) as client:
-            latest_tag = _get_latest_release_tag(client)
-            instructions_url = f"{GITHUB_RAW_BASE}/{latest_tag}/codex-rs/core/{prompt_file}"
-
-            # Load existing metadata for conditional request
-            metadata = _load_cache_metadata(meta_file)
-            headers: dict[str, str] = {}
-
-            # Only use ETag if tag matches (different release = different content)
-            if metadata and metadata.tag == latest_tag and metadata.etag:
-                headers["If-None-Match"] = metadata.etag
-
-            response = client.get(instructions_url, headers=headers)
-
-            if response.status_code == 304 and cache_file.exists():
-                # Not modified, update last_checked and return cached
-                if metadata:
-                    metadata.last_checked = int(time.time())
-                    _save_cache_metadata(meta_file, metadata)
-                log_debug(f"Codex {model_family} instructions not modified", debug_type=DebugType.GENERAL)
-                return cache_file.read_text()
-
-            if response.status_code == 200:
-                instructions = response.text
-                new_etag = response.headers.get("etag")
-
-                # Save to cache
-                cache_file.parent.mkdir(parents=True, exist_ok=True)
-                cache_file.write_text(instructions)
-                _save_cache_metadata(
-                    meta_file,
-                    CacheMetadata(
-                        etag=new_etag,
-                        tag=latest_tag,
-                        last_checked=int(time.time()),
-                        url=instructions_url,
-                    ),
-                )
-
-                log_debug(f"Updated {model_family} instructions from GitHub", debug_type=DebugType.GENERAL)
-                return instructions
-
-            raise RuntimeError(f"HTTP {response.status_code}")
-
-    except Exception as e:
-        log_debug(f"Failed to fetch {model_family} instructions: {e}", debug_type=DebugType.GENERAL)
-
-        # Fallback to cached version
-        if cache_file.exists():
-            log_debug(f"Using cached {model_family} instructions (fallback)", debug_type=DebugType.GENERAL)
-            return cache_file.read_text()
-
-        # Last resort: use bundled prompt
-        bundled_path = _get_bundled_prompt_path(model_family)
-        if bundled_path:
-            log_debug(f"Using bundled {model_family} instructions (fallback)", debug_type=DebugType.GENERAL)
-            return _load_bundled_prompt(bundled_path)
-
-        raise RuntimeError(f"No Codex instructions available for {model_family}") from e
-
-
-def _get_bundled_prompt_path(model_family: ModelFamily) -> str | None:
-    """Get bundled prompt path for model family."""
-    if model_family == "gpt-5.2-codex":
-        return "prompts/prompt-codex-gpt-5-2-codex.md"
-    if model_family == "gpt-5.2":
-        return "prompts/prompt-codex-gpt-5-2.md"
-    if model_family in ("codex", "codex-max", "gpt-5.1"):
-        return "prompts/prompt-codex.md"
-    return None
-
-
-def invalidate_cache(model: str | None = None) -> None:
-    """Invalidate cached instructions to force refresh on next access.
-
-    Args:
-        model: If provided, only invalidate cache for this model's family.
-               If None, invalidate all cached instructions.
-    """
-    if model:
-        model_family = get_model_family(model)
-        meta_file = CACHE_DIR / f"{CACHE_FILES[model_family].replace('.md', '-meta.json')}"
-        if meta_file.exists():
-            meta_file.unlink()
-    else:
-        if CACHE_DIR.exists():
-            for meta_file in CACHE_DIR.glob("*-meta.json"):
-                meta_file.unlink()