aria-code 0.167.0__py3-none-win_amd64.whl

aria_code/apps/cli/runtime_consumer.py

@@ -0,0 +1,769 @@
+"""CLI consumers for runtime/model events.
+
+The runtime core emits typed events and asks for approval decisions.  This
+module keeps terminal-specific rendering and prompts at the CLI adapter layer
+instead of letting them spread through the agent loop.
+"""
+
+from __future__ import annotations
+
+import re
+import sys
+import time
+from dataclasses import dataclass, field
+from enum import Enum
+from pathlib import Path
+from typing import Any, Callable
+
+from aria_code.runtime import (
+    AgentEventStatus,
+    AgentEventThinking,
+    AgentEventToken,
+    AgentEventToolCall,
+    AgentEventToolResult,
+    ApprovalDecision,
+)
+
+
+_REPETITION_MARKER = "*[model stopped — repetition detected]*"
+_REPETITION_NOTICE = (
+    "\n\n> 已检测到模型开始重复输出，已自动停止展开。"
+    "上方结果仍然有效；如需继续，请指定要补充的部分。"
+)
+
+
+def _redact_activity_text(value: Any, limit: int = 90) -> str:
+    """One line, secrets masked — for anything echoed to the screen or a log.
+
+    Shell commands are now shown on their step line, so this also covers
+    "Authorization: Bearer …", credentials in URLs and the usual key shapes.
+    """
+    text = re.sub(r"\s+", " ", str(value or "")).strip()
+    text = re.sub(
+        r"(?i)(api[_-]?key|token|password|passwd|secret)(\s*[=:]\s*|\s+)[^\s]+",
+        r"\1\2***",
+        text,
+    )
+    text = re.sub(r"(?i)\b(bearer|basic)\s+[A-Za-z0-9._~+/=-]+", r"\1 ***", text)
+    text = re.sub(r"://[^/\s:@]+:[^@\s]+@", "://***@", text)
+    text = re.sub(r"\b(sk|pk|rk)-[A-Za-z0-9_-]{8,}|\bgh[pousr]_[A-Za-z0-9]{12,}|\bAIza[0-9A-Za-z_-]{20,}"
+                  r"|\bxox[abposr]-[A-Za-z0-9-]{10,}", "***", text)
+    return text[:limit]
+
+
+# How long the runtime measured a tool call to take, carried with its result.
+MEASURED_ELAPSED_KEY = "_elapsed_s"
+
+
+def with_measured_elapsed(result: Any, elapsed: Any) -> Any:
+    if isinstance(result, dict) and isinstance(elapsed, (int, float)):
+        return {**result, MEASURED_ELAPSED_KEY: float(elapsed)}
+    return result
+
+
+def approval_subject(tool_name: str, params: dict[str, Any]) -> str:
+    """What an answered approval leaves behind: the file, or the command with secrets masked."""
+    if tool_name == "run_command":
+        return _redact_activity_text(params.get("command", ""), limit=80)
+    if tool_name == "computer_action":
+        return _desktop_subject(params)
+    for key in ("path", "file_path"):
+        if params.get(key):
+            return Path(str(params[key])).name
+    return tool_name.replace("_", " ")
+
+
+def _desktop_subject(params: dict[str, Any]) -> str:
+    """What the mouse or keyboard is about to do; typed text is counted, never echoed."""
+    action = str(params.get("action") or "input").lower()
+    if action == "type":
+        return f"type {len(str(params.get('text') or ''))} characters"
+    if action == "key":
+        return f"press {str(params.get('key') or '?')[:40]}"
+    if action == "drag":
+        return f"drag ({params.get('x', 0)}, {params.get('y', 0)}) to ({params.get('ex', 0)}, {params.get('ey', 0)})"
+    return f"{action.replace('_', ' ')} at ({params.get('x', 0)}, {params.get('y', 0)})"
+
+
+def _tool_activity_hint(tool: str, params: dict[str, Any]) -> str:
+    """Return one safe, useful argument for a compact activity row."""
+    if not isinstance(params, dict):
+        return ""
+    for key in ("path", "file_path", "filename", "file"):
+        if params.get(key):
+            return Path(str(params[key])).name[:90]
+    for key in ("symbol", "ticker", "query", "q", "command", "url", "task"):
+        if params.get(key):
+            return _redact_activity_text(params[key])
+    title = params.get("title")
+    return _redact_activity_text(title) if title else ""
+
+
+class TurnPhase(str, Enum):
+    """User-visible lifecycle for one assistant turn."""
+
+    CONNECTING = "connecting"
+    THINKING = "thinking"
+    STREAMING = "streaming"
+    TOOL = "tool"
+    DONE = "done"
+    ERROR = "error"
+    CANCELLED = "cancelled"
+
+
+@dataclass
+class TurnLifecycle:
+    phase: TurnPhase = TurnPhase.CONNECTING
+    started_at: float = field(default_factory=time.time)
+    first_visible_at: float | None = None
+    ended_at: float | None = None
+
+    def transition(self, phase: TurnPhase) -> None:
+        self.phase = phase
+        now = time.time()
+        if phase is TurnPhase.STREAMING and self.first_visible_at is None:
+            self.first_visible_at = now
+        if phase in {TurnPhase.DONE, TurnPhase.ERROR, TurnPhase.CANCELLED}:
+            self.ended_at = now
+
+
+class TerminalRuntimeEventConsumer:
+    """Consume runtime/provider events and render them to a terminal."""
+
+    action_view: Any = None
+
+    def __init__(
+        self,
+        *,
+        terminal: Any,
+        console: Any,
+        has_rich: bool,
+        markdown_cls: type | None,
+        live_cls: type | None,
+        strip_latex: Callable[[str], str],
+        set_robot_state: Callable[[Any], None] | None = None,
+        streaming_state: Any = None,
+        print_tool_call: Callable[[str, dict], None] | None = None,
+        print_tool_done: Callable[..., None] | None = None,
+        fallback_from: str = "local",
+        live_update_interval: float = 0.08,
+        ui_lang: str = "en",
+        on_response_start: Callable[[], None] | None = None,
+        on_phase_change: Callable[[TurnPhase], None] | None = None,
+        action_view: Any = None,
+    ) -> None:
+        self.terminal = terminal
+        self.console = console
+        self.has_rich = has_rich
+        self.markdown_cls = markdown_cls
+        self.live_cls = live_cls
+        self.strip_latex = strip_latex
+        self.set_robot_state = set_robot_state
+        self.streaming_state = streaming_state
+        self.print_tool_call = print_tool_call
+        self.print_tool_done = print_tool_done
+        self.fallback_from = fallback_from
+        self.live_update_interval = live_update_interval
+        self.ui_lang = ui_lang or "en"
+        self.on_response_start = on_response_start
+        self.on_phase_change = on_phase_change
+        # ui.render.actions.ActionView: tool events as Explored / Ran / Edit
+        # cells. None keeps the per-call ⏺ / ✓ lines.
+        self.action_view = action_view
+        self.lifecycle = TurnLifecycle()
+
+        self.response_text = ""
+        self.streamed_any = False
+        self.token_count = 0
+        self.thinking_tokens = 0
+        self.thinking_shown = False
+        self.thinking_start: float | None = None
+        self.thinking_finished = False
+        self.thinking_preview_buf: list[str] = []
+        self.thinking_full_buf: list[str] = []
+        self.tool_start_times: dict[str, list[float]] = {}
+        self.tool_params: dict[str, list[dict[str, Any]]] = {}
+        self.repetition_stopped = False
+        self.repetition_notice_printed = False
+        self.response_started = False
+
+        self.live_display = None
+        self.spinner = None
+        self.first_token_received_ref = [False]
+        self.token_start_time: float | None = None
+        self.last_live_update = 0.0
+        self.use_plain_print_ref = [False]
+        self.use_batch_render_ref = [False]
+        self.latex_buf_ref = [""]
+        self.in_latex_ref = [False]
+
+    @property
+    def first_token_received(self) -> bool:
+        return self.first_token_received_ref[0]
+
+    @first_token_received.setter
+    def first_token_received(self, value: bool) -> None:
+        self.first_token_received_ref[0] = value
+
+    @property
+    def use_plain_print(self) -> bool:
+        return self.use_plain_print_ref[0]
+
+    @use_plain_print.setter
+    def use_plain_print(self, value: bool) -> None:
+        self.use_plain_print_ref[0] = value
+
+    @property
+    def use_batch_render(self) -> bool:
+        return self.use_batch_render_ref[0]
+
+    @use_batch_render.setter
+    def use_batch_render(self, value: bool) -> None:
+        self.use_batch_render_ref[0] = value
+
+    @property
+    def latex_buf(self) -> str:
+        return self.latex_buf_ref[0]
+
+    @latex_buf.setter
+    def latex_buf(self, value: str) -> None:
+        self.latex_buf_ref[0] = value
+
+    @property
+    def in_latex(self) -> bool:
+        return self.in_latex_ref[0]
+
+    @in_latex.setter
+    def in_latex(self, value: bool) -> None:
+        self.in_latex_ref[0] = value
+
+    def start_spinner(self) -> None:
+        self.set_phase(TurnPhase.THINKING)
+        if self.has_rich and self.spinner is None and not self.first_token_received:
+            label = "思考中" if self.ui_lang.lower().startswith("zh") else "Thinking"
+            cancel = "esc 取消" if self.ui_lang.lower().startswith("zh") else "esc cancel"
+            self.spinner = self.console.status(
+                f"[dim]{label}… [/dim][dim italic]{cancel}[/dim italic]",
+                spinner="dots",
+                spinner_style="dim",
+            )
+            self.spinner.__enter__()
+
+    def set_phase(self, phase: TurnPhase) -> None:
+        if self.lifecycle.phase is phase:
+            return
+        self.lifecycle.transition(phase)
+        if self.on_phase_change is not None:
+            self.on_phase_change(phase)
+
+    def ensure_response_started(self) -> None:
+        """Render the response label only when visible answer text exists."""
+        if self.response_started:
+            return
+        self._finish_thinking()
+        self.stop_spinner()
+        self.flush_actions()
+        self.response_started = True
+        if self.on_response_start is not None:
+            self.on_response_start()
+
+    def finish(self, phase: TurnPhase = TurnPhase.DONE) -> None:
+        self._finish_thinking()
+        self.stop_spinner()
+        self.close_tool_spinner()
+        self.flush_actions()
+        self.set_phase(phase)
+
+    def print_action_lines(self, lines) -> None:
+        rich = self.has_rich and self.console is not None
+        for style, text in lines:
+            if text.startswith("⏺"):  # a new cell: one blank line above it
+                self.console.print() if rich else print()
+            if rich:
+                from rich.markup import escape
+                body = f"  {escape(text)}"
+                self.console.print(f"[{style}]{body}[/{style}]" if style else body, highlight=False)
+            else:
+                print(f"  {text}")
+
+    def flush_actions(self) -> None:
+        """Print a pending Explored cell before anything else reaches the screen."""
+        if self.action_view is not None:
+            self.print_action_lines(self.action_view.flush())
+
+    def stop_spinner(self) -> None:
+        if self.spinner is not None:
+            try:
+                self.spinner.__exit__(None, None, None)
+            except Exception:
+                pass
+            self.spinner = None
+
+    def stop_live(self, discard: bool = False) -> None:
+        self.stop_spinner()
+        if self.live_display:
+            try:
+                if discard:
+                    try:
+                        from rich.text import Text as _RichText
+
+                        self.live_display.update(_RichText(""))
+                        self.live_display.refresh()
+                    except Exception:
+                        pass
+                self.live_display.stop()
+            except Exception:
+                pass
+            self.live_display = None
+        elif self.first_token_received and not discard and not self.use_batch_render:
+            print(flush=True)
+
+    def set_batch_render_mode(self, enabled: bool = True) -> None:
+        self.use_plain_print = enabled
+        self.use_batch_render = enabled
+
+    def reset_stream_state(self) -> None:
+        self.response_text = ""
+        self.streamed_any = False
+        self.token_count = 0
+        self.first_token_received = False
+        self.token_start_time = None
+        self.latex_buf = ""
+        self.in_latex = False
+
+    def flush_latex_buf(self) -> str:
+        raw = self.latex_buf
+        self.latex_buf = ""
+        self.in_latex = False
+        return self.strip_latex(raw) if raw.strip() else raw
+
+    def _show_repetition_notice(self) -> None:
+        if self.repetition_notice_printed or self.use_batch_render:
+            return
+        self.repetition_notice_printed = True
+        if self.live_display and self.has_rich and self.markdown_cls is not None:
+            clean_text = self.response_text.split(_REPETITION_MARKER, 1)[0].rstrip()
+            self.live_display.update(self.markdown_cls(self.strip_latex(clean_text + _REPETITION_NOTICE)))
+            self.live_display.refresh()
+            return
+        print(_REPETITION_NOTICE, end="", flush=True)
+
+    def finalize_text(self, final_text: str) -> str:
+        if self.in_latex and self.latex_buf:
+            leftover = self.flush_latex_buf()
+            final_text = (final_text or "") + leftover
+            if self.use_plain_print and not self.use_batch_render:
+                print(leftover, end="", flush=True)
+        return final_text
+
+    def _finish_thinking(self) -> None:
+        if not self.thinking_shown or self.thinking_finished:
+            return
+        self.thinking_finished = True
+        self.stop_spinner()
+        elapsed_t = time.time() - self.thinking_start if self.thinking_start else 0
+        self.terminal._last_thinking = "".join(self.thinking_full_buf).strip()
+        is_zh = self.ui_lang.lower().startswith("zh")
+        t_info = f"思考 {elapsed_t:.1f}s" if is_zh else f"Thought for {elapsed_t:.1f}s"
+        if self.thinking_tokens > 0:
+            t_info += f" · {self.thinking_tokens:,} tokens"
+        expand = "展开" if is_zh else "expand"
+        ctrlo = f"  [dim]· Ctrl+O {expand}[/dim]" if self.terminal._last_thinking else ""
+        if self.has_rich:
+            sys.stdout.write("\r\033[K")
+            sys.stdout.flush()
+            self.console.print(f"  [dim]✻[/dim] [dim]{t_info}[/dim]{ctrlo}")
+            if self.terminal.config.get("thinking_preview") and self.thinking_preview_buf:
+                preview_text = "".join(self.thinking_preview_buf)[:280].strip()
+                if len("".join(self.thinking_preview_buf)) > 280:
+                    preview_text += "…"
+                self.console.print(f"  [dim italic]{preview_text}[/dim italic]")
+        else:
+            print(f"\r  ✻ {t_info}")
+
+    def on_token(self, token: str) -> None:
+        if "<|im_start|>" in token or "<|im_end|>" in token:
+            token = token.replace("<|im_start|>", "").replace("<|im_end|>", "")
+            if not token.strip():
+                return
+
+        meta_artifacts = (
+            "(注释：",
+            "（注释：",
+            "(提示：",
+            "（提示：",
+            "请使用实际注入的数据",
+            "请使用实际数据",
+            "实际注入的数据",
+            "[system]",
+            "[/system]",
+            "[INST]",
+            "[/INST]",
+        )
+        if any(a in token for a in meta_artifacts):
+            token = re.sub(
+                r"\(注[释释]：[^)）]*[)）]|（注[释释]：[^)）]*[)）]"
+                r"|\(提示：[^)）]*[)）]|（提示：[^)）]*[)）]"
+                r"|请使用实际(?:注入的)?数据[^。\n]*"
+                r"|\[/?(?:system|INST)\]",
+                "",
+                token,
+            )
+            if not token.strip():
+                return
+
+        if not self.first_token_received and not token.strip():
+            self.response_text += token
+            self.streamed_any = True
+            self.token_count += 1
+            return
+
+        if not self.first_token_received:
+            self.first_token_received = True
+            self.token_start_time = time.time()
+            self.set_phase(TurnPhase.STREAMING)
+            if self.set_robot_state is not None:
+                self.set_robot_state(self.streaming_state)
+            if not self.use_batch_render:
+                self.ensure_response_started()
+
+        if _REPETITION_MARKER in token:
+            if self.use_batch_render:
+                self.response_text += token
+                self.streamed_any = True
+                self.token_count += 1
+                self.repetition_stopped = True
+                return
+            before, _, _after = token.partition(_REPETITION_MARKER)
+            if before:
+                self.on_token(before)
+            self.response_text += _REPETITION_MARKER
+            self.streamed_any = True
+            self.token_count += 1
+            self.repetition_stopped = True
+            self._show_repetition_notice()
+            return
+
+        self._finish_thinking()
+
+        if self.use_batch_render:
+            self.response_text += token
+            self.streamed_any = True
+            self.token_count += 1
+            return
+
+        open_delims = (r"\(", r"\[", "$$")
+        close_delims = (r"\)", r"\]", "$$")
+        if not self.in_latex:
+            if any(d in token for d in open_delims):
+                self.in_latex = True
+                self.latex_buf = token
+                tail = token
+                for od, cd in zip(open_delims, close_delims):
+                    if od in tail:
+                        after = tail[tail.index(od) + len(od):]
+                        if cd in after:
+                            token = self.flush_latex_buf()
+                            break
+                else:
+                    self.response_text += self.latex_buf
+                    self.streamed_any = True
+                    self.token_count += 1
+                    return
+            else:
+                token = self.strip_latex(token)
+        else:
+            self.latex_buf += token
+            if any(d in token for d in close_delims):
+                token = self.flush_latex_buf()
+            else:
+                self.response_text += token
+                self.streamed_any = True
+                self.token_count += 1
+                return
+
+        self.response_text += token
+        self.streamed_any = True
+        self.token_count += 1
+        can_live = (
+            self.has_rich
+            and not self.use_plain_print
+            and getattr(self.console, "is_terminal", False)
+            and not getattr(self.console, "is_dumb_terminal", True)
+            and self.markdown_cls is not None
+            and self.live_cls is not None
+        )
+        if can_live:
+            now = time.time()
+            if self.live_display is not None and now - self.last_live_update < self.live_update_interval:
+                return
+            md = self.markdown_cls(self.strip_latex(self.response_text))
+            if self.live_display is None:
+                self.live_display = self.live_cls(
+                    md,
+                    console=self.console,
+                    refresh_per_second=12,
+                    vertical_overflow="visible",
+                )
+                self.live_display.start()
+                self.last_live_update = now
+            elif now - self.last_live_update >= self.live_update_interval:
+                self.live_display.update(md)
+                self.last_live_update = now
+        else:
+            print(token, end="", flush=True)
+
+    def on_thinking(self, content: str) -> None:
+        self.set_phase(TurnPhase.THINKING)
+        if self.thinking_finished:
+            # 工具调用后的新一段思考:重置段状态,否则上一段的 finished 闩锁
+            # 让本段永远无法 finalize,"思考中 …(N tokens)"活行被冻结在滚动区
+            # (即多条思考行并存的根因)。full_buf 不清,Ctrl+O 展开保留全部段落。
+            self.thinking_finished = False
+            self.thinking_shown = False
+            self.thinking_tokens = 0
+            self.thinking_preview_buf = []
+        if not self.thinking_shown:
+            self.stop_spinner()
+            self.thinking_start = time.time()
+            self.thinking_shown = True
+        self.thinking_tokens += 1
+        if self.thinking_tokens % 30 == 1:
+            elapsed = time.time() - self.thinking_start
+            label = "思考中" if self.ui_lang.lower().startswith("zh") else "Thinking"
+            # 与 finalize 行(思考 Xs · N tokens)同格式,避免括号/点分隔两种样式并存
+            sys.stdout.write(
+                f"\r  \033[2m✻\033[0m \033[2m{label}  {elapsed:.1f}s"
+                f"  ·  {self.thinking_tokens} tokens\033[0m    "
+            )
+            sys.stdout.flush()
+        if len("".join(self.thinking_preview_buf)) < 300:
+            self.thinking_preview_buf.append(content)
+        if len("".join(self.thinking_full_buf)) < 8000:
+            self.thinking_full_buf.append(content)
+
+    def on_tool_call(self, tool: str, params: dict) -> None:
+        self.set_phase(TurnPhase.TOOL)
+        self._finish_thinking()
+
+        # Cursor/Windsurf 风格的实时加载动画
+        # One spinner at a time. /team's analysts can call tools in parallel;
+        # close the previous indicator before starting the next one.
+        self.close_tool_spinner()
+        if self.has_rich and hasattr(self, 'console'):
+            hint = _tool_activity_hint(tool, params)
+            verb = "调用工具" if str(self.ui_lang).lower().startswith("zh") else "Running"
+            label = f"[bold cyan]{verb}[/bold cyan] [green]{tool}[/green]"
+            if hint:
+                label += f" [dim]({hint})[/dim]"
+            self.tool_spinner = self.console.status(label, spinner="dots12", spinner_style="cyan")
+            self.tool_spinner.__enter__()
+
+        if self.action_view is not None:
+            self.print_action_lines(self.action_view.start(tool, params if isinstance(params, dict) else {}))
+        elif self.print_tool_call is not None:
+            self.print_tool_call(tool, params if isinstance(params, dict) else {})
+
+        self.tool_start_times.setdefault(tool, []).append(time.time())
+        safe_params = dict(params) if isinstance(params, dict) else {}
+        self.tool_params.setdefault(tool, []).append(safe_params)
+
+    def close_tool_spinner(self) -> None:
+        """Stop the tool spinner, if one is running."""
+        spinner = getattr(self, "tool_spinner", None)
+        if spinner is not None:
+            try:
+                spinner.__exit__(None, None, None)
+            except Exception:
+                pass
+        self.tool_spinner = None
+
+    def on_tool_result(self, tool: str, summary: Any) -> None:
+        # 关闭工具加载动画
+        self.close_tool_spinner()
+
+        starts = self.tool_start_times.get(tool) or []
+        started_at = starts.pop(0) if starts else time.time()
+        if not starts:
+            self.tool_start_times.pop(tool, None)
+        elapsed_ms = int((time.time() - started_at) * 1000)
+        # The runtime times each call itself. Measuring from the announcement
+        # counted the person reading an approval, and results of a batch
+        # arrive together — "✓ writing file test_fx.py 13.9s" for a 9ms write.
+        if isinstance(summary, dict) and isinstance(summary.get(MEASURED_ELAPSED_KEY), (int, float)):
+            elapsed_ms = int(summary[MEASURED_ELAPSED_KEY] * 1000)
+        params_queue = self.tool_params.get(tool) or []
+        params = params_queue.pop(0) if params_queue else {}
+        if not params_queue:
+            self.tool_params.pop(tool, None)
+        ok = not (isinstance(summary, dict) and not summary.get("success", True))
+        if self.action_view is not None:
+            self.print_action_lines(self.action_view.done(tool, params, summary, elapsed_ms / 1000))
+        elif self.print_tool_done is not None:
+            # Surface the failure reason on the ✗ line — a red cross with no
+            # explanation leaves both the user and the reviewer blind to WHY
+            # (observed in the channels drill: peer_comparison failed 4× with
+            # nothing on screen).
+            detail = ""
+            if not ok and isinstance(summary, dict):
+                detail = str(summary.get("error") or "")[:90]
+            elif ok:
+                # Name what was done: two bare "✓ writing file" lines in a row
+                # did not say which file each was.
+                detail = _tool_activity_hint(tool, params)
+            self.print_tool_done(tool, elapsed_ms, success=ok, summary=detail)
+
+        ts = time.strftime("%H:%M:%S")
+        icon = "✓" if ok else "✗"
+        hint = _tool_activity_hint(tool, params)
+        duration = f"{elapsed_ms}ms" if elapsed_ms < 1000 else f"{elapsed_ms / 1000:.1f}s"
+        entry = f"{ts}  {icon} {tool}"
+        if hint:
+            entry += f"  {hint}"
+        entry += f"  · {duration}"
+        if not ok and isinstance(summary, dict):
+            error = _redact_activity_text(summary.get("error"), limit=100)
+            if error:
+                entry += f"  · {error}"
+        self.terminal._transcript_log.append(entry)
+        if len(self.terminal._transcript_log) > 100:
+            self.terminal._transcript_log = self.terminal._transcript_log[-100:]
+
+        if tool in ("TaskCreate", "TaskUpdate") and isinstance(summary, dict):
+            tid = summary.get("id") or summary.get("task_id")
+            title = summary.get("title", "")
+            status = summary.get("status", "pending")
+            if tid:
+                existing = next((t for t in self.terminal._task_list if t.get("id") == tid), None)
+                if existing:
+                    existing["status"] = status
+                    if title:
+                        existing["title"] = title
+                else:
+                    self.terminal._task_list.append({"id": tid, "title": title, "status": status})
+
+    def on_status(self, state: str, message: str) -> None:
+        if state in ("acceptance_passed", "acceptance_failed"):
+            # The gate's verdict is the one piece of evidence behind "done", so
+            # it is printed rather than left in the turn result: a user who
+            # cannot see that the check went red will read the model's summary
+            # as if it had gone green.
+            passed = state == "acceptance_passed"
+            if self.has_rich and self.console is not None:
+                colour = "green" if passed else "yellow"
+                self.console.print(f"  [{colour}]{'✓' if passed else '✗'} {message}[/{colour}]")
+            else:
+                print(f"  {'✓' if passed else '✗'} {message}")
+            return
+        if state in ("review_passed", "review_blocking", "review_error"):
+            mark, colour = {"review_passed": ("✓", "green"), "review_blocking": ("✗", "yellow"),
+                            "review_error": ("⚠", "dim")}[state]
+            if self.has_rich and self.console is not None:
+                from rich.markup import escape
+                self.console.print(f"  [{colour}]{mark} {escape(message)}[/{colour}]")
+            else:
+                print(f"  {mark} {message}")
+            return
+        if state in {"max_rounds", "budget_exhausted", "loop_guard", "checks_failed"}:
+            if self.has_rich and self.console is not None:
+                self.console.print(f"  [yellow]Task incomplete: {message}[/yellow]")
+            else:
+                print(f"  Task incomplete: {message}")
+            return
+        if state != "fallback":
+            return
+        match = re.search(r"(?:from\s+)?(\w+)\s*(?:→|->|to)\s*(\w+)", message or "", re.I)
+        if match:
+            from_provider, to_provider = match.group(1), match.group(2)
+        else:
+            from_provider, to_provider = self.fallback_from, "cloud"
+        from aria_code.ui.render.output import print_fallback_toast
+
+        print_fallback_toast(
+            from_provider,
+            to_provider,
+            message or "",
+            console=self.console,
+            has_rich=self.has_rich,
+        )
+
+    def handle_runtime_event(self, event: Any) -> None:
+        if isinstance(event, AgentEventToken):
+            self.on_token(event.text)
+        elif isinstance(event, AgentEventThinking):
+            self.on_thinking(event.content)
+        elif isinstance(event, AgentEventToolCall):
+            self.on_tool_call(event.tool, event.params)
+        elif isinstance(event, AgentEventToolResult):
+            self.on_tool_result(event.tool, with_measured_elapsed(event.result, getattr(event, "elapsed", None)))
+        elif isinstance(event, AgentEventStatus):
+            self.on_status(event.state, event.message)
+
+
+class TerminalApprovalEventConsumer:
+    """Terminal-side approval prompt consumer for runtime tool execution."""
+
+    def __init__(
+        self,
+        *,
+        terminal: Any,
+        console: Any,
+        has_rich: bool,
+        confirm_decision: Callable[..., ApprovalDecision],
+        apply_decision: Callable[[dict, ApprovalDecision], dict],
+        save_config: Callable[[dict], None],
+    ) -> None:
+        self.terminal = terminal
+        self.console = console
+        self.has_rich = has_rich
+        self.confirm_decision = confirm_decision
+        self.apply_decision = apply_decision
+        self.save_config = save_config
+
+    async def approve(
+        self,
+        tool_name: str,
+        tool_params: dict,
+        *,
+        stop_before_prompt: Callable[[], None] | None = None,
+    ) -> ApprovalDecision:
+        if stop_before_prompt is not None:
+            stop_before_prompt()
+        try:
+            approval = self.confirm_decision(
+                tool_name,
+                tool_params,
+                config_policy=self.terminal.config.get("command_policy", "safe"),
+            )
+        except KeyboardInterrupt:
+            approval = ApprovalDecision.deny("KeyboardInterrupt")
+
+        self.terminal._record_feedback(
+            "tool_accept" if approval.approved else "tool_reject",
+            tool_name,
+        )
+        if not approval.approved:
+            from aria_code.ui.render.output import print_tool_blocked
+
+            zh = str(self.terminal.config.get("ui_lang", "en")).lower().startswith("zh")
+            print_tool_blocked(tool_name, "用户取消" if zh else "declined", console=self.console,
+                               has_rich=self.has_rich)
+        return approval
+
+    def apply(self, tool_params: dict, approval: ApprovalDecision) -> dict:
+        self.apply_decision(tool_params, approval)
+        if approval.upgrade_policy:
+            tool_params.pop("_upgrade_policy", None)
+            self.terminal.config["command_policy"] = "balanced"
+            try:
+                self.save_config(self.terminal.config)
+                if self.has_rich:
+                    zh = str(self.terminal.config.get("ui_lang", "en")).lower().startswith("zh")
+                    self.console.print("  [dim]策略已升级为 balanced 并保存[/dim]" if zh else
+                                       "  [dim]Policy set to balanced and saved[/dim]")
+            except Exception:
+                pass
+        return tool_params
+
+
+__all__ = ["TerminalApprovalEventConsumer", "TerminalRuntimeEventConsumer"]