synth-ai 0.2.6.dev1__py3-none-any.whl → 0.4.3__py3-none-any.whl

synth_ai/sdk/streaming/handlers.py

@@ -0,0 +1,1997 @@
+from __future__ import annotations
+
+import contextlib
+import json
+import re
+import time
+from abc import ABC, abstractmethod
+from collections import deque
+from datetime import datetime
+from pathlib import Path
+from typing import Any, Callable
+
+import click
+
+from .types import StreamMessage, StreamType
+
+
+def _mask_sensitive_urls(text: str) -> str:
+    """Mask S3/Wasabi URLs and sensitive paths in log messages.
+    
+    Replaces full S3/Wasabi URLs with masked versions to prevent leaking
+    bucket names, paths, and infrastructure details in public SDK logs.
+    
+    Examples:
+        s3://synth-artifacts/models/... -> s3://***/***/[masked]
+        Wasabi s3://bucket/path/file.tar.gz -> Wasabi s3://***/***/[masked]
+    """
+    if not text:
+        return text
+    
+    # Pattern matches:
+    # - Optional "Wasabi " prefix
+    # - s3:// or http(s):// scheme
+    # - Any bucket/host
+    # - Any path
+    # - Common model file extensions
+    pattern = r'(Wasabi\s+)?((s3|https?)://[^\s]+\.(tar\.gz|zip|pt|pth|safetensors|ckpt|bin))'
+    
+    def replace_url(match: re.Match) -> str:
+        prefix = match.group(1) or ""  # "Wasabi " or empty
+        url = match.group(2)
+        # Extract just the filename
+        filename = url.split("/")[-1] if "/" in url else "file"
+        return f'{prefix}s3://***/***/[{filename}]'
+    
+    return re.sub(pattern, replace_url, text, flags=re.IGNORECASE)
+
+
+class StreamHandler(ABC):
+    """Base class for log handlers that consume ``StreamMessage`` objects."""
+
+    @abstractmethod
+    def handle(self, message: StreamMessage) -> None:
+        """Process a message produced by the streamer."""
+
+    def should_handle(self, message: StreamMessage) -> bool:  # pragma: no cover - trivial
+        """Predicate allowing handlers to filter messages before processing."""
+        return True
+
+    def flush(self) -> None:  # pragma: no cover - optional
+        """Flush buffered output."""
+        return None
+
+
+class CLIHandler(StreamHandler):
+    """Simple CLI output mirroring current poller behaviour."""
+
+    def __init__(
+        self,
+        *,
+        hidden_event_types: set[str] | None = None,
+        hidden_event_substrings: set[str] | None = None,
+    ) -> None:
+        self._hidden_event_types = set(hidden_event_types or set())
+        self._hidden_event_substrings = {s.lower() for s in (hidden_event_substrings or set())}
+
+    def handle(self, message: StreamMessage) -> None:
+        if not self.should_handle(message):
+            return
+
+        timestamp = datetime.now().strftime("%H:%M:%S")
+        if message.stream_type is StreamType.STATUS:
+            status = str(message.data.get("status") or message.data.get("state") or "unknown")
+            click.echo(f"[{timestamp}] status={status}")
+            return
+
+        if message.stream_type is StreamType.EVENTS:
+            event_type = message.data.get("type", "event")
+            if event_type in self._hidden_event_types:
+                return
+            level = message.data.get("level")
+            msg = message.data.get("message") or ""
+            # Evaluate substring filters against lower-cased concatenated text
+            if self._hidden_event_substrings:
+                blob = " ".join(
+                    [
+                        event_type or "",
+                        str(msg),
+                        json.dumps(message.data.get("data", "")),
+                    ]
+                ).lower()
+                if any(sub in blob for sub in self._hidden_event_substrings):
+                    return
+            prefix = f"[{timestamp}] [{message.seq}] {event_type}"
+            if level:
+                prefix += f" ({level})"
+            # Mask sensitive URLs before displaying
+            sanitized_msg = _mask_sensitive_urls(msg)
+            
+            # For error events, show full details including underlying errors
+            if level == "error" or event_type.endswith(".failed"):
+                click.echo(f"{prefix}: {sanitized_msg}")
+                # Show error details from data field if available
+                data = message.data.get("data", {})
+                if isinstance(data, dict):
+                    error_detail = data.get("detail") or data.get("error") or data.get("error_detail")
+                    if error_detail and str(error_detail) != sanitized_msg:
+                        # Show underlying error if different from main message
+                        click.echo(f"    Error details: {error_detail}")
+                    # Show traceback or stack if available
+                    traceback_info = data.get("traceback") or data.get("stack")
+                    if traceback_info:
+                        lines = str(traceback_info).split("\n")
+                        # Show last few lines of traceback (most relevant)
+                        for line in lines[-5:]:
+                            if line.strip():
+                                click.echo(f"    {line}")
+            else:
+                click.echo(f"{prefix}: {sanitized_msg}".rstrip(": "))
+
+            data = message.data.get("data") if isinstance(message.data.get("data"), dict) else {}
+            if event_type == "prompt.learning.mipro.complete" and data:
+                best_prompt = data.get("best_prompt")
+                if isinstance(best_prompt, dict):
+                    sections = best_prompt.get("sections")
+                    if isinstance(sections, list) and sections:
+                        click.echo("    --- BEST PROMPT ---")
+                        for section in sections:
+                            if not isinstance(section, dict):
+                                continue
+                            role = section.get("role", "unknown").upper()
+                            name = section.get("name")
+                            header = f"    [{role}]"
+                            if name:
+                                header += f" {name}"
+                            click.echo(header)
+                            content = section.get("content", "")
+                            if isinstance(content, str) and content:
+                                click.echo(f"        {content}")
+                        click.echo("    -------------------")
+
+            if event_type == "mipro.topk.evaluated" and data:
+                rank = data.get("rank")
+                train_score = data.get("train_score")
+                test_score = data.get("test_score")
+                instruction_text = data.get("instruction_text", "")
+                demo_indices = data.get("demo_indices", [])
+                lift_abs = data.get("lift_absolute")
+                lift_pct = data.get("lift_percent")
+                stage_payloads = data.get("stage_payloads", {})
+                details: list[str] = []
+                if rank is not None:
+                    details.append(f"Rank {rank}")
+                if isinstance(train_score, int | float):
+                    train_score_float = float(train_score)
+                    details.append(f"train={train_score_float:.3f} ({train_score_float*100:.1f}%)")
+                if isinstance(test_score, int | float):
+                    test_score_float = float(test_score)
+                    details.append(f"test={test_score_float:.3f} ({test_score_float*100:.1f}%)")
+                if isinstance(lift_abs, int | float) and isinstance(lift_pct, int | float):
+                    details.append(f"lift={lift_abs:+.3f} ({lift_pct:+.1f}%)")
+                if details:
+                    click.echo("    --- TOP-K CANDIDATE ---")
+                    click.echo(f"    {' | '.join(details)}")
+                    if isinstance(instruction_text, str) and instruction_text.strip():
+                        snippet = instruction_text.strip()
+                        click.echo(f"        Instruction: {snippet}")
+                    if isinstance(demo_indices, list) and demo_indices:
+                        click.echo(f"        Demo indices: {demo_indices}")
+                    
+                    # Display per-stage information if available
+                    if isinstance(stage_payloads, dict) and stage_payloads:
+                        click.echo("        Per-stage breakdown:")
+                        for stage_id, payload in stage_payloads.items():
+                            if isinstance(payload, dict):
+                                module_id = payload.get("module_id", stage_id)
+                                instr_ids = payload.get("instruction_indices", [])
+                                demo_ids = payload.get("demo_indices", [])
+                                click.echo(f"          [{module_id}/{stage_id}] instr_ids={instr_ids} demo_ids={demo_ids}")
+                    
+                    seed_scores = data.get("test_seed_scores")
+                    if isinstance(seed_scores, list) and seed_scores:
+                        formatted_scores = ", ".join(
+                            f"{item.get('seed')}: {item.get('score'):.2f}"
+                            for item in seed_scores
+                            if isinstance(item, dict) and isinstance(item.get("seed"), int) and isinstance(item.get("score"), int | float)
+                        )
+                        if formatted_scores:
+                            click.echo(f"        Test per-seed: {formatted_scores}")
+                    click.echo("    ----------------------")
+            return
+
+        if message.stream_type is StreamType.METRICS:
+            name = message.data.get("name")
+            value = message.data.get("value")
+            step = message.data.get("step")
+            data = message.data.get("data", {})
+            
+            # Format metric display
+            metric_str = f"[{timestamp}] [metric] {name}={value:.4f}" if isinstance(value, int | float) else f"[{timestamp}] [metric] {name}={value}"
+            if step is not None:
+                metric_str += f" (step={step})"
+            
+            # Add any additional context from data field
+            if isinstance(data, dict):
+                n = data.get("n")
+                if n is not None:
+                    metric_str += f" n={n}"
+            
+            click.echo(metric_str)
+            return
+
+        if message.stream_type is StreamType.TIMELINE:
+            phase = message.data.get("phase", "phase")
+            click.echo(f"[{timestamp}] timeline={phase}")
+
+
+class JSONHandler(StreamHandler):
+    """Emit messages as JSON lines suitable for machine parsing."""
+
+    def __init__(self, output_file: str | None = None, *, indent: int | None = None) -> None:
+        self.output_file = Path(output_file).expanduser() if output_file else None
+        self._indent = indent
+
+    def handle(self, message: StreamMessage) -> None:
+        if not self.should_handle(message):
+            return
+
+        payload: dict[str, Any] = {
+            "stream_type": message.stream_type.name,
+            "timestamp": message.timestamp,
+            "job_id": message.job_id,
+            "data": message.data,
+        }
+        if message.seq is not None:
+            payload["seq"] = message.seq
+        if message.step is not None:
+            payload["step"] = message.step
+        if message.phase is not None:
+            payload["phase"] = message.phase
+
+        line = json.dumps(payload, indent=self._indent)
+        if self.output_file:
+            with self.output_file.open("a", encoding="utf-8") as fh:
+                fh.write(line)
+                if self._indent is None:
+                    fh.write("\n")
+        else:
+            click.echo(line)
+
+    def flush(self) -> None:
+        return None
+
+
+class CallbackHandler(StreamHandler):
+    """Invoke user-provided callbacks for specific stream types."""
+
+    def __init__(
+        self,
+        *,
+        on_status: Callable[[dict[str, Any]], None] | None = None,
+        on_event: Callable[[dict[str, Any]], None] | None = None,
+        on_metric: Callable[[dict[str, Any]], None] | None = None,
+        on_timeline: Callable[[dict[str, Any]], None] | None = None,
+    ) -> None:
+        self._on_status = on_status
+        self._on_event = on_event
+        self._on_metric = on_metric
+        self._on_timeline = on_timeline
+
+    def handle(self, message: StreamMessage) -> None:
+        if not self.should_handle(message):
+            return
+
+        if message.stream_type is StreamType.STATUS and self._on_status:
+            self._on_status(message.data)
+        elif message.stream_type is StreamType.EVENTS and self._on_event:
+            self._on_event(message.data)
+        elif message.stream_type is StreamType.METRICS and self._on_metric:
+            self._on_metric(message.data)
+        elif message.stream_type is StreamType.TIMELINE and self._on_timeline:
+            self._on_timeline(message.data)
+
+
+class BufferedHandler(StreamHandler):
+    """Collect messages and emit them in batches."""
+
+    def __init__(self, *, flush_interval: float = 5.0, max_buffer_size: int = 100) -> None:
+        self.flush_interval = flush_interval
+        self.max_buffer_size = max_buffer_size
+        self._buffer: list[StreamMessage] = []
+        self._last_flush = time.time()
+
+    def handle(self, message: StreamMessage) -> None:
+        if not self.should_handle(message):
+            return
+
+        self._buffer.append(message)
+        now = time.time()
+        if len(self._buffer) >= self.max_buffer_size or now - self._last_flush >= self.flush_interval:
+            self.flush()
+
+    def flush(self) -> None:
+        if not self._buffer:
+            return
+        self.process_batch(self._buffer)
+        self._buffer.clear()
+        self._last_flush = time.time()
+
+    def process_batch(self, messages: list[StreamMessage]) -> None:  # pragma: no cover - abstract
+        """Override to define how buffered messages should be processed."""
+
+
+class IntegrationTestHandler(StreamHandler):
+    """Collect messages for integration tests or programmatic assertions."""
+
+    def __init__(self) -> None:
+        self.messages: list[StreamMessage] = []
+
+    def handle(self, message: StreamMessage) -> None:
+        self.messages.append(message)
+
+    def clear(self) -> None:
+        self.messages.clear()
+
+
+class GraphGenHandler(StreamHandler):
+    """Handler for ADAS jobs that delegate child job streams to an underlying handler.
+    
+    ADAS jobs emit events from child jobs (GEPA, MIPRO, RL, SFT, etc.). This handler
+    provides light ADAS-aware filtering and routing while keeping child job output
+    intact via a delegate handler. The delegate can be supplied directly or created
+    via a factory; by default we choose a prompt-learning handler for GEPA/MIPRO and
+    a basic CLI handler for other job types.
+    """
+
+    def __init__(
+        self,
+        *,
+        child_handler: StreamHandler | None = None,
+        child_handler_factory: Callable[[str | None], StreamHandler | None] | None = None,
+        show_trial_results: bool = True,
+        show_transformations: bool = False,
+        show_validation: bool = True,
+        filter_verbose_events: bool = True,
+        wrap_child_events: bool = True,
+    ) -> None:
+        # User-supplied delegate or factory; both are optional.
+        self.child_handler = child_handler
+        self._child_handler_factory = child_handler_factory
+
+        # Options for the default prompt-learning delegate
+        self._pl_show_trial_results = show_trial_results
+        self._pl_show_transformations = show_transformations
+        self._pl_show_validation = show_validation
+
+        self.filter_verbose_events = filter_verbose_events
+        # If False, skip ADAS-specific filtering/transformations and just pass through.
+        self.wrap_child_events = wrap_child_events
+
+        # Detected child job type (gepa/mipro/rl/sft/etc.)
+        self.child_job_type: str | None = None
+        # Track whether we created the delegate automatically (so we can swap if needed)
+        self._delegate_auto_created = False
+
+    def handle(self, message: StreamMessage) -> None:
+        if not self.should_handle(message):
+            return
+
+        if message.stream_type is StreamType.EVENTS:
+            self._detect_child_job_type(message)
+            self._maybe_reset_delegate_for_child_type()
+
+            if self.wrap_child_events and self.filter_verbose_events:
+                if self._should_filter_event(message):
+                    return
+
+            if self.wrap_child_events:
+                message = self._transform_event_message(message)
+
+        delegate = self._get_child_handler()
+        if delegate:
+            delegate.handle(message)
+
+    def _get_child_handler(self) -> StreamHandler:
+        """Return or create the delegate handler used for child job events."""
+        if self.child_handler:
+            return self.child_handler
+
+        handler: StreamHandler | None = None
+        if self._child_handler_factory:
+            handler = self._child_handler_factory(self.child_job_type)
+
+        if handler is None:
+            # Choose a sensible default based on detected child job type
+            if self._is_prompt_learning_type(self.child_job_type):
+                handler = PromptLearningHandler(
+                    show_trial_results=self._pl_show_trial_results,
+                    show_transformations=self._pl_show_transformations,
+                    show_validation=self._pl_show_validation,
+                )
+            else:
+                handler = CLIHandler()
+
+        self.child_handler = handler
+        self._delegate_auto_created = self._child_handler_factory is None and self.child_handler is not None
+        return handler
+
+    def _detect_child_job_type(self, message: StreamMessage) -> None:
+        """Infer the child job type from event types."""
+        if self.child_job_type:
+            return
+
+        event_type = str(message.data.get("type") or "").lower()
+        if not event_type:
+            return
+
+        if event_type.startswith("graph_evolve."):
+            self.child_job_type = "graph_evolve"
+        elif "mipro" in event_type:
+            self.child_job_type = "mipro"
+        elif "gepa" in event_type or event_type.startswith("prompt.learning"):
+            self.child_job_type = "prompt_learning"
+        elif event_type.startswith("rl.") or ".rl." in event_type:
+            self.child_job_type = "rl"
+        elif event_type.startswith("sft.") or ".sft." in event_type:
+            self.child_job_type = "sft"
+        else:
+            # Fall back to the first segment as a hint (e.g., "adas.child_type")
+            parts = event_type.split(".")
+            if parts:
+                self.child_job_type = parts[0]
+
+    def _maybe_reset_delegate_for_child_type(self) -> None:
+        """Swap out auto-created delegates when we later detect a different child type."""
+        if not self.child_handler or not self._delegate_auto_created:
+            return
+
+        # If the detected type does not match the current delegate choice, rebuild.
+        wants_prompt_learning = self._is_prompt_learning_type(self.child_job_type)
+        has_prompt_learning_handler = isinstance(self.child_handler, PromptLearningHandler)
+
+        if wants_prompt_learning and not has_prompt_learning_handler:
+            self.child_handler = None
+            self._delegate_auto_created = False
+        elif not wants_prompt_learning and has_prompt_learning_handler:
+            self.child_handler = None
+            self._delegate_auto_created = False
+
+    def _should_filter_event(self, message: StreamMessage) -> bool:
+        """Determine if an event should be hidden from output."""
+        event_type = message.data.get("type", "") or ""
+        event_type_lower = event_type.lower()
+
+        # Never filter graph_evolve events - they're important for GraphGen jobs
+        if event_type.startswith("graph_evolve."):
+            return False
+
+        # Only filter prompt-learning style events; leave other job types untouched.
+        if not any(key in event_type_lower for key in ("prompt.learning", "gepa", "mipro")):
+            return False
+
+        important_events = {
+            "prompt.learning.created",
+            "prompt.learning.gepa.start",
+            "prompt.learning.gepa.complete",
+            "prompt.learning.mipro.job.started",
+            "prompt.learning.mipro.optimization.exhausted",
+            "prompt.learning.trial.results",
+            "prompt.learning.progress",
+            "prompt.learning.gepa.new_best",
+            "prompt.learning.validation.summary",
+            "prompt.learning.candidate.evaluated",
+            "prompt.learning.candidate.evaluation.started",
+            # GraphGen/graph_evolve important events
+            "graph_evolve.job_started",
+            "graph_evolve.generation_started",
+            "graph_evolve.generation_completed",
+            "graph_evolve.candidate_evaluated",
+            "graph_evolve.archive_updated",
+            "graph_evolve.job_completed",
+            "graph_evolve.job_failed",
+        }
+        if event_type in important_events:
+            return False
+
+        verbose_patterns = [
+            "gepa.transformation.proposed",
+            "gepa.proposal.scored",
+            "prompt.learning.proposal.scored",
+            "mipro.tpe.update",
+            "prompt.learning.stream.connected",
+        ]
+        return any(pattern in event_type_lower for pattern in verbose_patterns)
+
+    def _transform_event_message(self, message: StreamMessage) -> StreamMessage:
+        """Transform event messages for ADAS context (currently passthrough)."""
+        return message
+
+    def flush(self) -> None:
+        # Ensure delegate flushes buffered output if needed.
+        if self.child_handler and hasattr(self.child_handler, "flush"):
+            with contextlib.suppress(Exception):
+                self.child_handler.flush()
+
+    @staticmethod
+    def _is_prompt_learning_type(job_type: str | None) -> bool:
+        """Return True if the child job type should use prompt-learning formatting."""
+        return job_type in {"gepa", "mipro", "prompt_learning", "prompt-learning", None}
+
+
+class LossCurveHandler(StreamHandler):
+    """Render a live-updating loss chart inside a fixed Rich panel."""
+
+    def __init__(
+        self,
+        *,
+        metric_name: str = "train.loss",
+        max_points: int = 200,
+        width: int = 60,
+        console: Any | None = None,
+        live: Any | None = None,
+    ) -> None:
+        try:
+            from rich.console import Console
+            from rich.live import Live
+            from rich.panel import Panel
+            from rich.text import Text
+        except ImportError as exc:  # pragma: no cover - optional dependency guard
+            raise RuntimeError(
+                "LossCurveHandler requires the 'rich' package. Install synth-ai[all] or rich>=13."
+            ) from exc
+
+        self.metric_name = metric_name
+        self.max_points = max_points
+        self.width = width
+
+        self._console_class = Console
+        self._panel_class = Panel
+        self._text_class = Text
+
+        self._console = console or Console()
+        self._live = live or Live(console=self._console, transient=False, refresh_per_second=8)
+        self._started = False
+
+        self._steps: list[int] = []
+        self._values: list[float] = []
+        self._status = "waiting"
+        self._last_event: str | None = None
+
+    def handle(self, message: StreamMessage) -> None:
+        updated = False
+
+        if message.stream_type is StreamType.STATUS:
+            status = str(message.data.get("status") or message.data.get("state") or "unknown")
+            if status != self._status:
+                self._status = status
+                updated = True
+
+        elif message.stream_type is StreamType.EVENTS:
+            event_type = message.data.get("type", "")
+            msg = message.data.get("message") or ""
+            level = message.data.get("level")
+            summary = f"{event_type}".strip()
+            if level:
+                summary += f" ({level})"
+            if msg:
+                summary += f": {msg}"
+            if summary != self._last_event:
+                self._last_event = summary
+                updated = True
+
+        elif message.stream_type is StreamType.METRICS:
+            if message.data.get("name") != self.metric_name:
+                return
+            value = message.data.get("value")
+            step = message.data.get("step")
+            if not isinstance(value, int | float) or not isinstance(step, int):
+                return
+            self._values.append(float(value))
+            self._steps.append(step)
+            if len(self._values) > self.max_points:
+                self._values = self._values[-self.max_points :]
+                self._steps = self._steps[-self.max_points :]
+            updated = True
+
+        elif message.stream_type is StreamType.TIMELINE:
+            phase = message.data.get("phase")
+            if phase:
+                self._status = str(phase)
+                updated = True
+
+        if updated:
+            self._refresh()
+
+    def flush(self) -> None:
+        if self._started:
+            with contextlib.suppress(Exception):
+                self._live.stop()
+            self._started = False
+
+    def _ensure_live(self) -> None:
+        if not self._started:
+            with contextlib.suppress(Exception):
+                self._live.start()
+            self._started = True
+
+    def _refresh(self) -> None:
+        self._ensure_live()
+        body = self._build_body()
+        title = f"{self.metric_name} | status={self._status}"
+        self._live.update(self._panel_class(body, title=title, border_style="cyan"))
+
+    def _build_body(self) -> Any:
+        if not self._values:
+            return self._text_class("Waiting for metrics…", style="yellow")
+
+        chart = self._render_sparkline()
+        last_value = self._values[-1]
+        lines = [
+            chart,
+            f"latest: {last_value:.4f} (step {self._steps[-1]})",
+        ]
+        if self._last_event:
+            lines.append(f"event: {self._last_event}")
+        return "\n".join(lines)
+
+    def _render_sparkline(self) -> str:
+        blocks = "▁▂▃▄▅▆▇█"
+        tail_len = min(self.width, len(self._values))
+        tail = self._values[-tail_len:]
+        minimum = min(tail)
+        maximum = max(tail)
+        if maximum == minimum:
+            level = blocks[0]
+            return f"{minimum:.2f} {level * tail_len} {maximum:.2f}"
+        scale = (len(blocks) - 1) / (maximum - minimum)
+        chars = "".join(blocks[int((v - minimum) * scale + 0.5)] for v in tail)
+        return f"{minimum:.2f} {chars} {maximum:.2f}"
+
+    def __del__(self) -> None:  # pragma: no cover - defensive cleanup
+        with contextlib.suppress(Exception):
+            self.flush()
+
+class RichHandler(StreamHandler):
+    """Rich powered handler with live progress and metrics table."""
+
+    def __init__(
+        self,
+        *,
+        event_log_size: int = 20,
+        console: Any | None = None,
+    ) -> None:
+        try:
+            from rich.console import Console
+            from rich.progress import BarColumn, Progress, SpinnerColumn, TextColumn
+            from rich.table import Table
+        except ImportError as exc:  # pragma: no cover - requires optional dependency
+            raise RuntimeError(
+                "RichHandler requires the 'rich' package. Install synth-ai[all] or rich>=13."
+            ) from exc
+
+        self._console_class = Console
+        self._progress_class = Progress
+        self._spinner_column = SpinnerColumn
+        self._text_column = TextColumn
+        self._bar_column = BarColumn
+        self._table_class = Table
+
+        self._console = console or Console()
+        self._progress = Progress(
+            SpinnerColumn(),
+            TextColumn("[progress.description]{task.description}"),
+            BarColumn(),
+            TextColumn("{task.completed}/{task.total}" if console else ""),
+            TextColumn("[progress.percentage]{task.percentage:>3.0f}%"),
+            transient=False,
+            console=self._console,
+        )
+        self._task_id: int | None = None
+        self._current_status = "unknown"
+        self._latest_metrics: dict[str, Any] = {}
+        self._event_log: deque[str] = deque(maxlen=event_log_size)
+        self._progress_started = False
+
+    def handle(self, message: StreamMessage) -> None:
+        if not self.should_handle(message):
+            return
+
+        if message.stream_type is StreamType.STATUS:
+            self._current_status = str(message.data.get("status") or message.data.get("state"))
+            self._ensure_progress_started()
+            if self._task_id is not None:
+                description = f"Status: {self._current_status}"
+                self._progress.update(self._task_id, description=description)  # type: ignore[arg-type]
+            self._render_summary()
+            return
+
+        if message.stream_type is StreamType.EVENTS:
+            event_type = message.data.get("type", "event")
+            summary = message.data.get("message") or ""
+            level = message.data.get("level")
+            # Mask sensitive URLs before displaying
+            sanitized_summary = _mask_sensitive_urls(summary)
+            formatted = f"[{event_type}] {sanitized_summary}".strip()
+            if level:
+                formatted = f"{formatted} ({level})"
+            self._event_log.append(formatted)
+            data = message.data.get("data") or {}
+            step = data.get("step") or data.get("current_step")
+            total_steps = data.get("total_steps") or data.get("max_steps")
+            if step and total_steps:
+                self._ensure_progress_started(total_steps)
+                if self._task_id is not None:
+                    self._progress.update(self._task_id, completed=int(step), total=int(total_steps))  # type: ignore[arg-type]
+            self._render_summary()
+            return
+
+        if message.stream_type is StreamType.METRICS:
+            name = message.data.get("name", "")
+            value = message.data.get("value")
+            if name:
+                self._latest_metrics[name] = value
+            self._render_summary()
+            return
+
+        if message.stream_type is StreamType.TIMELINE:
+            phase = message.data.get("phase", "")
+            if phase and phase.lower() not in {"training", "running"}:
+                self._event_log.append(f"[timeline] {phase}")
+            self._render_summary()
+
+    def flush(self) -> None:
+        if self._progress_started:
+            self._progress.stop()
+            self._progress_started = False
+        self._render_summary(force=True)
+
+    def _ensure_progress_started(self, total: int | float | None = None) -> None:
+        if not self._progress_started:
+            self._progress.start()
+            self._progress_started = True
+        if self._task_id is None:
+            self._task_id = self._progress.add_task(
+                f"Status: {self._current_status}", total=total or 100
+            )
+        elif total is not None and self._task_id is not None:
+            self._progress.update(self._task_id, total=total)  # type: ignore[arg-type]
+
+    def _render_summary(self, force: bool = False) -> None:
+        if force and self._progress_started:
+            self._progress.refresh()
+
+        table = self._table_class(title="Latest Metrics")
+        table.add_column("Metric")
+        table.add_column("Value")
+
+        if not self._latest_metrics:
+            table.add_row("—", "—")
+        else:
+            for name, value in sorted(self._latest_metrics.items()):
+                table.add_row(str(name), str(value))
+
+        if self._progress_started:
+            self._progress.console.print(table)
+        else:
+            self._console.print(table)
+
+        if self._event_log:
+            self._console.print("\nRecent events:")
+            for entry in list(self._event_log):
+                self._console.print(f"  • {entry}")
+
+class ContextLearningHandler(StreamHandler):
+    """CLI-friendly handler for Context Learning jobs.
+
+    Emits high-signal progress similar to other infra job handlers,
+    specialized for generation-based bash context optimization.
+    """
+
+    def __init__(self) -> None:
+        self.best_score_so_far = 0.0
+        self.current_generation = 0
+
+    def handle(self, message: StreamMessage) -> None:
+        if not self.should_handle(message):
+            return
+
+        timestamp = datetime.now().strftime("%H:%M:%S")
+
+        if message.stream_type is StreamType.STATUS:
+            status = str(message.data.get("status") or message.data.get("state") or "unknown")
+            click.echo(f"[{timestamp}] status={status}")
+            return
+
+        if message.stream_type is StreamType.METRICS:
+            name = message.data.get("name")
+            value = message.data.get("value")
+            step = message.data.get("step")
+            if isinstance(value, int | float):
+                try:
+                    val_f = float(value)
+                    if val_f > self.best_score_so_far:
+                        self.best_score_so_far = val_f
+                    if isinstance(step, int):
+                        self.current_generation = max(self.current_generation, step)
+                    click.echo(f"[{timestamp}] gen={step} best={val_f:.3f}")
+                    return
+                except Exception:
+                    pass
+            click.echo(f"[{timestamp}] metric {name}={value}")
+            return
+
+        if message.stream_type is StreamType.EVENTS:
+            event_type = str(message.data.get("type") or "")
+            msg = message.data.get("message") or ""
+            data = message.data.get("data") or {}
+
+            if event_type == "context.learning.generation.completed":
+                gen = data.get("generation") or data.get("gen") or self.current_generation
+                score = data.get("best_score") or data.get("score") or self.best_score_so_far
+                try:
+                    score_f = float(score)
+                    if score_f > self.best_score_so_far:
+                        self.best_score_so_far = score_f
+                    click.echo(f"[{timestamp}] generation {gen} best={score_f:.3f}")
+                except Exception:
+                    click.echo(f"[{timestamp}] generation {gen} completed")
+                return
+
+            if event_type.endswith(".failed"):
+                click.echo(f"[{timestamp}] {event_type}: {msg}")
+                return
+
+            if msg:
+                click.echo(f"[{timestamp}] {event_type}: {msg}")
+            else:
+                click.echo(f"[{timestamp}] {event_type}")
+
+
+class PromptLearningHandler(StreamHandler):
+    """Enhanced handler for GEPA/MIPRO prompt optimization jobs with rich formatting and metrics tracking.
+    
+    This handler processes streaming events from both GEPA (Genetic Evolutionary Prompt
+    Algorithm) and MIPRO (Meta-Instruction PROposer) optimization jobs. It provides:
+    
+    - **Real-time progress tracking**: Shows trial results, rollouts, iterations, and budget usage
+    - **Optimization curve tracking**: Maintains a history of best scores over time
+    - **GEPA-specific features**: Tracks transformations, rollouts, and validation results
+    - **MIPRO-specific features**: Tracks iterations, trials, minibatch/full evaluations, and budget
+    - **Dual output**: Writes to both console (via click.echo) and optional log file
+    
+    The handler filters verbose events (like TPE updates, proposed instructions) to keep
+    output readable while preserving important progress information. It formats output
+    consistently between GEPA and MIPRO for easier comparison.
+    
+    Example:
+        >>> handler = PromptLearningHandler(
+        ...     show_trial_results=True,
+        ...     max_tokens=1_000_000,
+        ...     log_file=Path("optimization.log")
+        ... )
+        >>> # Handler is used by JobStreamer to process events
+    """
+    
+    def __init__(
+        self,
+        *,
+        show_trial_results: bool = True,
+        show_transformations: bool = False,
+        show_validation: bool = True,
+        max_tokens: int | None = None,
+        max_time_seconds: float | None = None,
+        max_rollouts: int | None = None,
+        log_file: Path | None = None,
+    ):
+        """Initialize the prompt learning handler.
+        
+        Args:
+            show_trial_results: Whether to display individual trial scores (default: True).
+                When True, shows each trial's score and best score so far.
+            show_transformations: Whether to display transformation/proposal details
+                (default: False). When True, shows verbose transformation events.
+            show_validation: Whether to display validation summaries (default: True).
+                Shows validation results comparing candidates against baseline.
+            max_tokens: Maximum token budget for MIPRO (from TOML termination_config).
+                Used to track progress and enforce limits.
+            max_time_seconds: Maximum time budget in seconds (from TOML termination_config).
+                Used to track elapsed time and ETA.
+            max_rollouts: Maximum rollouts budget (from TOML termination_config).
+                Used to track rollout progress for both GEPA and MIPRO.
+            log_file: Optional path to log file for persistent logging. If provided,
+                all output is written to both console and file. File is opened in
+                append mode and remains open for streaming.
+        """
+        self.show_trial_results = show_trial_results
+        self.show_transformations = show_transformations
+        self.show_validation = show_validation
+        self.optimization_curve: list[tuple[int, float]] = []
+        self.trial_counter = 0
+        self.best_score_so_far = 0.0
+        
+        # MIPRO progress tracking
+        self.mipro_start_time: float | None = None
+        self.mipro_total_trials: int | None = None
+        self.mipro_completed_trials: int = 0
+        self.mipro_total_tokens: int = 0
+        self.mipro_policy_tokens: int = 0  # Rollout tokens (policy only)
+        self.mipro_max_tokens: int | None = max_tokens  # From TOML termination_config
+        self.mipro_total_cost: float = 0.0
+        self.mipro_max_cost: float | None = None
+        self.mipro_current_iteration: int = 0
+        self.mipro_num_iterations: int | None = None
+        self.mipro_trials_per_iteration: int | None = None
+        self.mipro_best_score: float = 0.0  # Track best full eval score
+        self.mipro_baseline_score: float | None = None  # Track baseline for comparison
+        self.mipro_batch_size: int | None = None  # Track minibatch size (N for minibatch scores)
+        self.mipro_rollouts_completed: int = 0  # Total rollouts completed
+        self.mipro_max_rollouts: int | None = max_rollouts  # From TOML termination_config
+        self.mipro_max_time_seconds: float | None = max_time_seconds  # From TOML termination_config
+        self._last_progress_emit_time: float | None = None  # Throttle progress updates
+        self._progress_emit_interval: float = 5.0  # Emit progress at most every 5 seconds
+        
+        # Log file for real-time streaming
+        self.log_file: Path | None = log_file
+        self._log_file_handle = None
+        if self.log_file:
+            try:
+                # Create parent directory if needed
+                self.log_file.parent.mkdir(parents=True, exist_ok=True)
+                # Open file in append mode for live streaming
+                # Note: File must remain open for streaming, so we can't use context manager
+                from datetime import datetime
+                self._log_file_handle = open(self.log_file, "a", encoding="utf-8")  # noqa: SIM115
+                # Write header
+                self._log_file_handle.write("=" * 80 + "\n")
+                self._log_file_handle.write("PROMPT LEARNING VERBOSE LOG\n")
+                self._log_file_handle.write(f"Started: {datetime.now().strftime('%Y-%m-%d %H:%M:%S')}\n")
+                self._log_file_handle.write("=" * 80 + "\n\n")
+                self._log_file_handle.flush()
+            except Exception as e:
+                # If we can't open the log file, continue without it
+                click.echo(f"⚠️  Could not open log file {log_file}: {e}", err=True)
+                self.log_file = None
+                self._log_file_handle = None
+    
+    def _write_log(self, text: str) -> None:
+        """Write text to both console and log file."""
+        click.echo(text)
+        if self._log_file_handle:
+            try:
+                self._log_file_handle.write(text + "\n")
+                self._log_file_handle.flush()
+            except Exception:
+                # If write fails, close handle and continue without logging
+                from contextlib import suppress
+                with suppress(Exception):
+                    self._log_file_handle.close()
+                self._log_file_handle = None
+    
+    def handle(self, message: StreamMessage) -> None:
+        """Handle a stream message from the prompt learning job.
+        
+        Routes messages to appropriate handlers based on stream type:
+        - STATUS: Job status updates (queued, running, completed, etc.)
+        - EVENTS: Algorithm-specific events (trials, iterations, transformations)
+        - METRICS: Performance metrics (scores, accuracies, costs)
+        - TIMELINE: Phase transitions
+        
+        Filters verbose events (TPE updates, proposed instructions) to keep output
+        readable. MIPRO and GEPA events are handled by specialized methods.
+        
+        Args:
+            message: StreamMessage containing event data from the backend
+        """
+        if not self.should_handle(message):
+            return
+        
+        timestamp = datetime.now().strftime("%H:%M:%S")
+        
+        if message.stream_type is StreamType.STATUS:
+            status = str(message.data.get("status") or message.data.get("state") or "unknown")
+            self._write_log(f"[{timestamp}] status={status}")
+            return
+        
+        if message.stream_type is StreamType.EVENTS:
+            event_type = message.data.get("type", "event")
+            level = message.data.get("level")
+            msg = message.data.get("message") or ""
+            
+            # Handle MIPRO-specific events for progress tracking (before skipping hidden events)
+            if event_type == "mipro.job.started":
+                self._handle_mipro_job_started(message.data)
+                # Continue to default display
+            
+            if event_type == "mipro.budget.update":
+                self._handle_mipro_budget_update(message.data)
+                # Continue to default display
+            
+            if event_type == "mipro.trial.complete":
+                self._handle_mipro_trial_complete(message.data)
+                # Continue to default display
+            
+            # Show more MIPRO events - only hide the most verbose ones
+            _hidden_mipro_events = {
+                # Keep only the most verbose TPE updates hidden
+                "mipro.tpe.update",  # Very frequent, low value
+            }
+            if event_type in _hidden_mipro_events:
+                return
+            
+            # Show GEPA transformation proposals - they're useful for debugging
+            # if event_type == "gepa.transformation.proposed":
+            #     return
+            
+            # Handle trial results for optimization curve tracking
+            if event_type == "prompt.learning.trial.results":
+                self._handle_trial_results(message.data)
+                # Continue to default display
+            
+            # Handle validation summary
+            if event_type == "prompt.learning.validation.summary":
+                if self.show_validation:
+                    self._handle_validation_summary(message.data)
+                # Continue to default display
+            
+            # Handle progress events
+            if event_type == "prompt.learning.progress":
+                self._handle_progress(message.data)
+                # Continue to default display
+            
+            # Handle MIPRO-specific events for progress tracking
+            if event_type == "mipro.iteration.start":
+                self._handle_mipro_iteration_start(message.data)
+                # Continue to default display
+            
+            if event_type == "mipro.iteration.complete":
+                self._handle_mipro_iteration_complete(message.data)
+                # Continue to default display
+            
+            if event_type == "mipro.fulleval.complete":
+                self._handle_mipro_fulleval_complete(message.data)
+                # Continue to default display
+            
+            if event_type == "mipro.optimization.exhausted":
+                # Graceful conclusion - show final progress
+                self._emit_mipro_progress()
+                # Continue to default display
+            
+            if event_type == "mipro.new_incumbent":
+                self._handle_mipro_new_incumbent(message.data)
+                # Continue to default display
+            
+            # Handle rollouts start event
+            if event_type == "prompt.learning.rollouts.start":
+                self._handle_rollouts_start(message.data)
+                # Continue to default display
+
+            # Handle GEPA new best event
+            if event_type == "prompt.learning.gepa.new_best":
+                self._handle_gepa_new_best(message.data)
+                # Continue to default display
+
+            # Handle phase changed event
+            if event_type == "prompt.learning.phase.changed":
+                self._handle_phase_changed(message.data)
+                # Continue to default display
+
+            # Handle stream connected event (connection lifecycle)
+            if event_type == "prompt.learning.stream.connected":
+                self._handle_stream_connected(message.data)
+                # Continue to default display
+
+            # Handle proposal scored events (transformations) - show by default
+            if event_type == "prompt.learning.proposal.scored":
+                self._handle_proposal_scored(message.data)
+                # Continue to default display
+            
+            # Show verbose transformation events by default - they're useful
+            # Only skip if explicitly disabled via show_transformations=False
+            # verbose_event_types = [
+            #     "prompt.learning.proposal.scored",
+            #     "prompt.learning.eval.summary",
+            #     "prompt.learning.validation.scored",
+            #     "prompt.learning.final.results",
+            # ]
+            # if event_type in verbose_event_types and not self.show_transformations:
+            #     return
+            
+            # Default event display - show more details
+            prefix = f"[{timestamp}] {event_type}"
+            if level:
+                prefix += f" ({level})"
+            sanitized_msg = _mask_sensitive_urls(msg)
+            
+            # Include key data fields if message is empty or short
+            if not sanitized_msg or len(sanitized_msg) < 50:
+                data = message.data.get("data", {})
+                if isinstance(data, dict):
+                    # Show useful fields
+                    useful_fields = []
+                    for key in ["score", "accuracy", "mean", "step", "iteration", "trial", "completed", "total", "version_id"]:
+                        if key in data:
+                            value = data[key]
+                            if isinstance(value, (int, float)):
+                                useful_fields.append(f"{key}={value:.4f}" if isinstance(value, float) else f"{key}={value}")
+                            else:
+                                useful_fields.append(f"{key}={value}")
+                    if useful_fields:
+                        sanitized_msg = sanitized_msg + (" " if sanitized_msg else "") + " ".join(useful_fields[:5])  # Limit to 5 fields
+            
+            self._write_log(f"{prefix}: {sanitized_msg}".rstrip(": "))
+            return
+        
+        if message.stream_type is StreamType.METRICS:
+            name = message.data.get("name")
+            value = message.data.get("value")
+            step = message.data.get("step")
+            data = message.data.get("data", {})
+            
+            metric_str = f"[{timestamp}] [metric] {name}={value:.4f}" if isinstance(value, int | float) else f"[{timestamp}] [metric] {name}={value}"
+            if step is not None:
+                metric_str += f" (step={step})"
+            
+            if isinstance(data, dict):
+                n = data.get("n")
+                if n is not None:
+                    metric_str += f" n={n}"
+            
+            self._write_log(metric_str)
+            return
+        
+        if message.stream_type is StreamType.TIMELINE:
+            phase = message.data.get("phase", "phase")
+            self._write_log(f"[{timestamp}] timeline={phase}")
+    
+    def _handle_trial_results(self, event_data: dict[str, Any]) -> None:
+        """Handle GEPA trial results events and track optimization curve.
+        
+        Processes trial completion events from GEPA optimization, tracking:
+        - Mean score for the trial
+        - Best score achieved so far
+        - Number of rollouts completed (N)
+        - Optimization curve data points
+        
+        Updates the optimization curve with (trial_number, best_score) tuples
+        for visualization. Displays trial results if show_trial_results is True.
+        
+        Args:
+            event_data: Event data dictionary containing:
+                - data.mean: Mean score for this trial
+                - data.completed: Number of rollouts completed
+                - data.total: Total rollouts planned
+        """
+        data = event_data.get("data", {})
+        if not isinstance(data, dict):
+            return
+        
+        mean_score = data.get("mean")
+        if mean_score is not None:
+            self.trial_counter += 1
+            self.best_score_so_far = max(self.best_score_so_far, float(mean_score))
+            self.optimization_curve.append((self.trial_counter, self.best_score_so_far))
+            
+            if self.show_trial_results:
+                timestamp = datetime.now().strftime("%H:%M:%S")
+                
+                # Extract N (number of rollouts)
+                completed = data.get("completed")
+                total = data.get("total")
+                
+                n_str = f" N={completed}/{total}" if completed is not None and total is not None else (f" N={completed}" if completed is not None else "")
+                
+                self._write_log(f"[{timestamp}] [Trial {self.trial_counter}] Score: {mean_score:.4f} (Best: {self.best_score_so_far:.4f}){n_str}")
+    
+    def _handle_validation_summary(self, event_data: dict[str, Any]) -> None:
+        """Handle validation summary events showing candidate performance.
+        
+        Displays validation results comparing optimized prompts against a baseline.
+        Shows baseline score, number of candidates evaluated (N), and top candidate
+        scores. Only displayed if show_validation is True.
+        
+        Args:
+            event_data: Event data dictionary containing:
+                - data.baseline: Baseline score (dict with accuracy/score or number)
+                - data.results: List of candidate results with accuracy/score fields
+        """
+        data = event_data.get("data", {})
+        if not isinstance(data, dict):
+            return
+        
+        timestamp = datetime.now().strftime("%H:%M:%S")
+        
+        # Extract baseline
+        baseline = data.get("baseline")
+        baseline_score = None
+        if isinstance(baseline, dict):
+            baseline_score = baseline.get("accuracy") or baseline.get("score")
+        elif isinstance(baseline, int | float):
+            baseline_score = baseline
+        
+        # Extract results
+        results = data.get("results", [])
+        if not isinstance(results, list):
+            results = []
+        
+        # Display validation summary
+        self._write_log(f"[{timestamp}] Validation Summary:")
+        
+        # Show baseline if available
+        if baseline_score is not None:
+            self._write_log(f"  Baseline: {baseline_score:.4f}")
+        
+        # Show N (number of candidates)
+        n_candidates = len(results)
+        if n_candidates > 0:
+            self._write_log(f"  N={n_candidates}")
+        
+        # Display validation results
+        if results:
+            for i, result in enumerate(results[:10]):  # Show top 10
+                if isinstance(result, dict):
+                    accuracy = result.get("accuracy") or result.get("score")
+                    if accuracy is not None:
+                        self._write_log(f"  Candidate {i+1}: {accuracy:.4f}")
+    
+    def _handle_progress(self, event_data: dict[str, Any]) -> None:
+        """Handle GEPA progress events with detailed rollout and transformation tracking.
+        
+        Displays comprehensive progress information including:
+        - Overall completion percentage
+        - Rollout progress (completed/total with percentage)
+        - Transformation progress (tried/planned with percentage)
+        - Token usage (used/budget in millions)
+        - Elapsed time and ETA
+        
+        Formats progress in a human-readable format similar to CLI progress bars.
+        
+        Args:
+            event_data: Event data dictionary containing:
+                - data.rollouts_completed: Number of rollouts completed
+                - data.rollouts_total: Total rollouts planned
+                - data.transformations_tried: Number of transformations tried
+                - data.transformations_planned: Total transformations planned
+                - data.rollout_tokens_used: Tokens consumed
+                - data.rollout_tokens_budget: Token budget
+                - data.elapsed_seconds: Time elapsed
+                - data.eta_seconds: Estimated time remaining
+                - data.percent_overall: Overall completion percentage
+        """
+        data = event_data.get("data", {})
+        if not isinstance(data, dict):
+            return
+        
+        timestamp = datetime.now().strftime("%H:%M:%S")
+        
+        # Extract rollout progress
+        rollouts_completed = data.get("rollouts_completed")
+        rollouts_total = data.get("rollouts_total")
+        percent_rollouts = data.get("percent_rollouts")
+        
+        # Extract transformation progress
+        transformations_tried = data.get("transformations_tried")
+        transformations_planned = data.get("transformations_planned")
+        percent_transformations = data.get("percent_transformations")
+        
+        # Extract overall progress
+        percent_overall = data.get("percent_overall")
+        
+        # Extract timing
+        elapsed_seconds = data.get("elapsed_seconds")
+        eta_seconds = data.get("eta_seconds")
+        
+        # Extract token usage
+        rollout_tokens_used = data.get("rollout_tokens_used")
+        rollout_tokens_budget = data.get("rollout_tokens_budget")
+        
+        # Build progress message
+        parts = []
+        
+        # Overall percentage
+        if percent_overall is not None:
+            parts.append(f"{int(percent_overall * 100)}% complete")
+        
+        # Rollout progress
+        if rollouts_completed is not None and rollouts_total is not None:
+            parts.append(f"rollouts={rollouts_completed}/{rollouts_total}")
+            if percent_rollouts is not None:
+                parts.append(f"({int(percent_rollouts * 100)}%)")
+        elif rollouts_completed is not None:
+            parts.append(f"rollouts={rollouts_completed}")
+        
+        # Transformation progress
+        if transformations_tried is not None and transformations_planned is not None:
+            parts.append(f"transformations={transformations_tried}/{transformations_planned}")
+            if percent_transformations is not None:
+                parts.append(f"({int(percent_transformations * 100)}%)")
+        elif transformations_tried is not None:
+            parts.append(f"transformations={transformations_tried}")
+        
+        # Token usage
+        if rollout_tokens_used is not None:
+            tokens_millions = rollout_tokens_used / 1_000_000.0
+            if rollout_tokens_budget is not None:
+                budget_millions = rollout_tokens_budget / 1_000_000.0
+                parts.append(f"tokens={tokens_millions:.2f}M/{budget_millions:.2f}M")
+            else:
+                parts.append(f"tokens={tokens_millions:.2f}M")
+        
+        # Timing
+        if elapsed_seconds is not None:
+            if elapsed_seconds >= 60:
+                elapsed_str = f"{elapsed_seconds / 60:.1f}min"
+            else:
+                elapsed_str = f"{int(elapsed_seconds)}s"
+            parts.append(f"elapsed={elapsed_str}")
+        
+        if eta_seconds is not None:
+            eta_str = f"{eta_seconds / 60:.1f}min" if eta_seconds >= 60 else f"{int(eta_seconds)}s"
+            parts.append(f"eta={eta_str}")
+        
+        # Fallback to simple step/total_steps if no detailed info
+        if not parts:
+            step = data.get("step") or data.get("current_step")
+            total_steps = data.get("total_steps") or data.get("max_steps")
+            if step is not None and total_steps is not None:
+                parts.append(f"{step}/{total_steps} ({100 * step / total_steps:.1f}%)")
+        
+        if parts:
+            progress_msg = " ".join(parts)
+            self._write_log(f"[{timestamp}] Progress: {progress_msg}")
+    
+    def _handle_rollouts_start(self, event_data: dict[str, Any]) -> None:
+        """Handle GEPA rollouts start event.
+        
+        Displays when rollouts begin, showing the number of training seeds
+        that will be evaluated. This marks the start of the main optimization
+        phase for GEPA.
+        
+        Args:
+            event_data: Event data dictionary containing:
+                - data.train_seeds: List of training seed values
+        """
+        data = event_data.get("data", {})
+        if not isinstance(data, dict):
+            return
+        
+        timestamp = datetime.now().strftime("%H:%M:%S")
+        train_seeds = data.get("train_seeds", [])
+        
+        if isinstance(train_seeds, list) and train_seeds:
+            num_seeds = len(train_seeds)
+            self._write_log(f"[{timestamp}] Starting rollouts: {num_seeds} seeds")
+        else:
+            self._write_log(f"[{timestamp}] Starting rollouts")
+
+    def _handle_gepa_new_best(self, event_data: dict[str, Any]) -> None:
+        """Handle GEPA new best candidate event.
+
+        Displays when a new best candidate is found during optimization,
+        showing the improvement over the previous best.
+
+        Args:
+            event_data: Event data dictionary containing:
+                - data.accuracy: New best accuracy score
+                - data.previous_best_score: Previous best score
+                - data.improvement: Absolute improvement
+                - data.version_id: ID of the new best candidate
+        """
+        data = event_data.get("data", {})
+        if not isinstance(data, dict):
+            return
+
+        timestamp = datetime.now().strftime("%H:%M:%S")
+        accuracy = data.get("accuracy")
+        previous = data.get("previous_best_score")
+        improvement = data.get("improvement")
+
+        if accuracy is not None:
+            msg = f"[{timestamp}] \u2728 New best: {accuracy:.4f}"
+            if previous is not None and improvement is not None:
+                msg += f" (+{improvement:.4f} from {previous:.4f})"
+            elif previous is not None:
+                msg += f" (was {previous:.4f})"
+            self._write_log(msg)
+
+    def _handle_phase_changed(self, event_data: dict[str, Any]) -> None:
+        """Handle phase transition event.
+
+        Displays when the optimization transitions between phases
+        (e.g., bootstrap -> optimization -> validation -> complete).
+
+        Args:
+            event_data: Event data dictionary containing:
+                - data.from_phase: Previous phase name
+                - data.to_phase: New phase name
+                - data.phase_summary: Optional summary of completed phase
+        """
+        data = event_data.get("data", {})
+        if not isinstance(data, dict):
+            return
+
+        timestamp = datetime.now().strftime("%H:%M:%S")
+        from_phase = data.get("from_phase") or "start"
+        to_phase = data.get("to_phase")
+
+        if to_phase:
+            self._write_log(f"[{timestamp}] Phase: {from_phase} \u2192 {to_phase}")
+
+    def _handle_stream_connected(self, event_data: dict[str, Any]) -> None:
+        """Handle SSE stream connection event.
+
+        Displays connection confirmation with cursor position for debugging.
+
+        Args:
+            event_data: Event data dictionary containing:
+                - data.cursor: Current sequence cursor position
+                - data.heartbeat_interval_seconds: Heartbeat interval
+        """
+        data = event_data.get("data", {})
+        if not isinstance(data, dict):
+            return
+
+        timestamp = datetime.now().strftime("%H:%M:%S")
+        cursor = data.get("cursor", 0)
+        self._write_log(f"[{timestamp}] Stream connected (cursor={cursor})")
+
+    def _handle_mipro_job_started(self, event_data: dict[str, Any]) -> None:
+        """Handle MIPRO job start event and extract configuration.
+        
+        Captures initial MIPRO configuration from the job start event to enable
+        progress tracking. Extracts num_iterations and num_trials_per_iteration
+        to estimate total trials and rollouts.
+        
+        Args:
+            event_data: Event data dictionary containing:
+                - data.num_iterations: Total number of optimization iterations
+                - data.num_trials_per_iteration: Trials per iteration
+        """
+        data = event_data.get("data", {})
+        if not isinstance(data, dict):
+            return
+        
+        # Extract config values to estimate max rollouts
+        num_iterations = data.get("num_iterations")
+        num_trials_per_iteration = data.get("num_trials_per_iteration")
+        
+        if num_iterations is not None:
+            self.mipro_num_iterations = num_iterations
+        if num_trials_per_iteration is not None:
+            self.mipro_trials_per_iteration = num_trials_per_iteration
+    
+    def _handle_mipro_iteration_start(self, event_data: dict[str, Any]) -> None:
+        """Handle MIPRO iteration start event and initialize progress tracking.
+        
+        Called at the start of each MIPRO iteration. On the first iteration (0),
+        initializes all progress tracking variables including:
+        - Total iterations and trials per iteration
+        - Batch size (for minibatch evaluations)
+        - Max rollouts estimate (iterations * trials * batch_size)
+        - Time and token budgets
+        
+        Sets the start time for elapsed time tracking.
+        
+        Args:
+            event_data: Event data dictionary containing:
+                - data.iteration: Current iteration number (0-indexed)
+                - data.num_iterations: Total iterations
+                - data.num_trials_per_iteration: Trials per iteration
+                - data.batch_size: Minibatch size (N for minibatch scores)
+                - data.max_trials: Maximum trials limit (optional)
+                - data.max_rollouts: Maximum rollouts limit (optional)
+                - data.max_time_seconds: Maximum time limit (optional)
+        """
+        import time
+        
+        data = event_data.get("data", {})
+        if not isinstance(data, dict):
+            return
+        
+        iteration = data.get("iteration")
+        if iteration == 0 and self.mipro_start_time is None:
+            self.mipro_start_time = time.time()
+        
+        # Extract total iterations and trials per iteration from first iteration
+        if iteration == 0:
+            self.mipro_num_iterations = data.get("num_iterations") or self.mipro_num_iterations
+            self.mipro_trials_per_iteration = data.get("num_trials_per_iteration") or self.mipro_trials_per_iteration
+            batch_size = data.get("batch_size")
+            if batch_size is not None:
+                self.mipro_batch_size = batch_size
+            
+            if self.mipro_num_iterations and self.mipro_trials_per_iteration:
+                self.mipro_total_trials = self.mipro_num_iterations * self.mipro_trials_per_iteration
+            
+            # Extract max limits if available (from events, but TOML value takes precedence)
+            # Only override if TOML value wasn't set
+            max_trials = data.get("max_trials")
+            max_rollouts_from_event = data.get("max_rollouts")
+            if self.mipro_max_rollouts is None:
+                if max_rollouts_from_event is not None:
+                    # Use event value if TOML value wasn't set
+                    self.mipro_max_rollouts = max_rollouts_from_event
+                elif max_trials is not None:
+                    # Fallback: If max_trials is set, use it as max rollouts (approximation)
+                    self.mipro_max_rollouts = max_trials
+                elif self.mipro_num_iterations and self.mipro_trials_per_iteration and self.mipro_batch_size:
+                    # Estimate max rollouts: iterations * trials_per_iteration * batch_size
+                    self.mipro_max_rollouts = self.mipro_num_iterations * self.mipro_trials_per_iteration * self.mipro_batch_size
+            
+            max_time_seconds = data.get("max_time_seconds") or data.get("max_wall_clock_seconds")
+            if max_time_seconds is not None and self.mipro_max_time_seconds is None:
+                # Use event value only if TOML value wasn't set
+                self.mipro_max_time_seconds = float(max_time_seconds)
+        
+        self.mipro_current_iteration = iteration if iteration is not None else self.mipro_current_iteration
+    
+    def _handle_mipro_iteration_complete(self, event_data: dict[str, Any]) -> None:
+        """Handle MIPRO iteration completion event.
+        
+        Updates progress tracking when an iteration completes, including:
+        - Cumulative trial count
+        - Current iteration number
+        
+        Emits a progress update showing overall progress, trials completed,
+        iterations, rollouts, tokens, and time.
+        
+        Args:
+            event_data: Event data dictionary containing:
+                - data.iteration: Completed iteration number
+                - data.cumulative: Cumulative trial count across all iterations
+        """
+        data = event_data.get("data", {})
+        if not isinstance(data, dict):
+            return
+        
+        cumulative = data.get("cumulative")
+        if cumulative is not None:
+            self.mipro_completed_trials = cumulative
+        
+        # Update current iteration
+        iteration = data.get("iteration")
+        if iteration is not None:
+            self.mipro_current_iteration = iteration
+        
+        # Emit progress update
+        self._emit_mipro_progress()
+    
+    def _handle_mipro_trial_complete(self, event_data: dict[str, Any]) -> None:
+        """Handle MIPRO trial completion event (minibatch evaluation).
+        
+        Processes minibatch trial completion events, which occur frequently during
+        MIPRO optimization. Tracks:
+        - Completed trial count
+        - Rollouts completed (from num_seeds)
+        - Minibatch scores (displayed if show_trial_results is True)
+        
+        Displays trial results in GEPA-like format: [Trial X] Score: Y (Best: Z) N=W
+        where N is the minibatch size. Emits throttled progress updates.
+        
+        Args:
+            event_data: Event data dictionary containing:
+                - data.minibatch_score: Score from minibatch evaluation
+                - data.iteration: Current iteration number
+                - data.trial: Trial number within iteration
+                - data.num_seeds: Number of seeds evaluated (minibatch size N)
+        """
+        data = event_data.get("data", {})
+        if not isinstance(data, dict):
+            return
+        
+        # Increment completed trials counter
+        self.mipro_completed_trials += 1
+        
+        # Count rollouts from trial events
+        num_seeds = data.get("num_seeds") or data.get("num_instances", 0)
+        if num_seeds:
+            self.mipro_rollouts_completed += num_seeds
+        
+        # Show trial score (minibatch) - like GEPA trial format
+        if self.show_trial_results:
+            timestamp = datetime.now().strftime("%H:%M:%S")
+            minibatch_score = data.get("minibatch_score")
+            iteration = data.get("iteration")
+            trial = data.get("trial")
+            
+            if minibatch_score is not None:
+                try:
+                    score_float = float(minibatch_score)
+                    # Calculate trial number for display
+                    if iteration is not None and trial is not None and self.mipro_trials_per_iteration:
+                        trial_num_display = (iteration * self.mipro_trials_per_iteration) + (trial + 1)
+                    else:
+                        trial_num_display = self.mipro_completed_trials
+                    
+                    n_str = f" N={num_seeds}" if num_seeds else ""
+                    best_str = f" (Best: {self.mipro_best_score:.4f})" if self.mipro_best_score > 0 else ""
+                    
+                    self._write_log(
+                        f"[{timestamp}] [Trial {trial_num_display}] Score: {score_float:.4f}{best_str}{n_str}"
+                    )
+                except (ValueError, TypeError):
+                    pass
+        
+        # Emit progress update after each trial (throttled internally)
+        self._emit_mipro_progress()
+    
+    def _handle_mipro_fulleval_complete(self, event_data: dict[str, Any]) -> None:
+        """Handle MIPRO full evaluation completion event.
+        
+        Processes full evaluation events, which occur less frequently than minibatch
+        trials. Full evaluations use the full validation set and are more expensive.
+        Only displays results if the score is "promising":
+        - Better than current best score, OR
+        - At least 5% improvement over baseline
+        
+        Tracks rollouts from full evaluations and updates best score. Displays
+        results with baseline comparison and improvement percentage.
+        
+        Args:
+            event_data: Event data dictionary containing:
+                - data.score: Full evaluation score
+                - data.iteration: Current iteration number
+                - data.trial: Trial number within iteration
+                - data.num_seeds: Number of seeds evaluated (full eval size)
+                - data.seeds: List of seed values (alternative to num_seeds)
+        """
+        data = event_data.get("data", {})
+        if not isinstance(data, dict):
+            return
+        
+        # Count rollouts from full eval
+        num_seeds = data.get("num_seeds") or data.get("seeds", 0)
+        if isinstance(num_seeds, list):
+            num_seeds = len(num_seeds)
+        if num_seeds:
+            self.mipro_rollouts_completed += num_seeds
+        
+        score = data.get("score")
+        if score is None:
+            return
+        
+        try:
+            score_float = float(score)
+        except (ValueError, TypeError):
+            return
+        
+        # Initialize baseline if not set (use first score as baseline)
+        if self.mipro_baseline_score is None:
+            self.mipro_baseline_score = score_float
+        
+        # Only show if score is promising:
+        # - Better than current best, OR
+        # - At least 5% improvement over baseline
+        is_promising = False
+        if score_float > self.mipro_best_score:
+            self.mipro_best_score = score_float
+            is_promising = True
+        elif self.mipro_baseline_score is not None:
+            improvement = score_float - self.mipro_baseline_score
+            improvement_pct = (improvement / self.mipro_baseline_score * 100) if self.mipro_baseline_score > 0 else 0
+            if improvement_pct >= 5.0:  # At least 5% improvement over baseline
+                is_promising = True
+        
+        if is_promising:
+            timestamp = datetime.now().strftime("%H:%M:%S")
+            iteration = data.get("iteration")
+            trial = data.get("trial")
+            seeds = data.get("seeds") or data.get("num_seeds", 0)
+            if isinstance(seeds, list):
+                seeds = len(seeds)
+            
+            # Format similar to GEPA trial results with N displayed
+            iter_str = f" iter={iteration}" if iteration is not None else ""
+            trial_str = f" trial={trial}" if trial is not None else ""
+            n_str = f" N={seeds}" if seeds else ""
+            
+            baseline_str = ""
+            if self.mipro_baseline_score is not None:
+                improvement = score_float - self.mipro_baseline_score
+                improvement_pct = (improvement / self.mipro_baseline_score * 100) if self.mipro_baseline_score > 0 else 0
+                baseline_str = f" (Baseline: {self.mipro_baseline_score:.4f}, +{improvement_pct:.1f}%)"
+            
+            self._write_log(
+                f"[{timestamp}] Full eval: Score={score_float:.4f} (Best: {self.mipro_best_score:.4f}){n_str}{baseline_str}{iter_str}{trial_str}"
+            )
+    
+    def _handle_mipro_new_incumbent(self, event_data: dict[str, Any]) -> None:
+        """Handle MIPRO new incumbent event (best candidate found).
+        
+        Processes events when MIPRO finds a new best candidate (incumbent).
+        Updates the optimization curve and displays the result in GEPA-like format
+        for consistency. Tracks cumulative trial count for curve visualization.
+        
+        Args:
+            event_data: Event data dictionary containing:
+                - data.minibatch_score: Minibatch score of the new incumbent
+                - data.best_score: Overall best score
+                - data.iteration: Current iteration number
+                - data.trial: Trial number within iteration
+                - data.cumulative_trials: Cumulative trial count across iterations
+                - data.num_seeds: Minibatch size (N)
+        """
+        data = event_data.get("data", {})
+        if not isinstance(data, dict):
+            return
+        
+        timestamp = datetime.now().strftime("%H:%M:%S")
+        minibatch_score = data.get("minibatch_score")
+        best_score = data.get("best_score")
+        iteration = data.get("iteration")
+        trial = data.get("trial")
+        num_seeds = data.get("num_seeds")  # N for minibatch
+        
+        if minibatch_score is None:
+            return
+        
+        try:
+            score_float = float(minibatch_score)
+        except (ValueError, TypeError):
+            return
+        
+        # Update best score if this is better
+        if best_score is not None:
+            best_float = float(best_score)
+            if best_float > self.best_score_so_far:
+                self.best_score_so_far = best_float
+        elif score_float > self.best_score_so_far:
+            self.best_score_so_far = score_float
+        
+        # Track optimization curve
+        if trial is not None:
+            # Use cumulative trial count for x-axis
+            cumulative_trials = data.get("cumulative_trials")
+            if cumulative_trials is not None:
+                trial_num = cumulative_trials
+            else:
+                # Estimate: (iteration * trials_per_iteration) + trial
+                if iteration is not None and self.mipro_trials_per_iteration:
+                    trial_num = (iteration * self.mipro_trials_per_iteration) + (trial + 1)
+                else:
+                    trial_num = self.trial_counter + 1
+            
+            self.optimization_curve.append((trial_num, self.best_score_so_far))
+            self.trial_counter = trial_num
+        
+        # Format like GEPA: [Trial X] Score: X (Best: Y) N=Z
+        trial_num_display = self.trial_counter if self.trial_counter > 0 else (trial + 1 if trial is not None else 1)
+        n_str = f" N={num_seeds}" if num_seeds is not None else ""
+        
+        click.echo(
+            f"[{timestamp}] [Trial {trial_num_display}] Score: {score_float:.4f} (Best: {self.best_score_so_far:.4f}){n_str}"
+        )
+        
+        # Emit progress update after each trial (throttled internally)
+        self._emit_mipro_progress()
+    
+    def _handle_mipro_budget_update(self, event_data: dict[str, Any]) -> None:
+        """Handle MIPRO budget update events.
+        
+        Tracks token usage and cost accumulation during optimization. Updates:
+        - Total tokens consumed (all operations)
+        - Policy tokens (rollout tokens only)
+        - Total cost in USD
+        - Max token and cost limits (if provided in event)
+        
+        Emits throttled progress updates to show budget consumption.
+        
+        Args:
+            event_data: Event data dictionary containing:
+                - data.total_tokens: Total tokens consumed
+                - data.policy_tokens: Tokens used for rollouts (policy only)
+                - data.total_cost_usd: Total cost in USD
+                - data.max_token_limit: Maximum token budget (optional)
+                - data.max_spend_usd: Maximum cost budget (optional)
+        """
+        data = event_data.get("data", {})
+        if not isinstance(data, dict):
+            return
+        
+        # Update token tracking
+        total_tokens = data.get("total_tokens")
+        if total_tokens is not None:
+            self.mipro_total_tokens = total_tokens
+        
+        # Track policy tokens separately (rollout tokens)
+        policy_tokens = data.get("policy_tokens")
+        if policy_tokens is not None:
+            self.mipro_policy_tokens = policy_tokens
+        
+        # Update cost tracking
+        total_cost = data.get("total_cost_usd")
+        if total_cost is not None:
+            self.mipro_total_cost = total_cost
+        
+        # Extract max limits if available in event data
+        max_token_limit = data.get("max_token_limit")
+        if max_token_limit is not None:
+            self.mipro_max_tokens = max_token_limit
+        
+        max_spend_usd = data.get("max_spend_usd")
+        if max_spend_usd is not None:
+            self.mipro_max_cost = max_spend_usd
+        
+        # Emit progress update periodically (throttled)
+        self._emit_mipro_progress()
+    
+    def _emit_mipro_progress(self) -> None:
+        """Emit a comprehensive progress update for MIPRO (throttled).
+        
+        Formats and displays MIPRO progress in a format similar to GEPA for consistency.
+        Shows:
+        - Overall completion percentage
+        - Trial progress (completed/total with remaining)
+        - Iteration progress (current/total)
+        - Rollout progress (completed/max)
+        - Token usage (used/budget in millions)
+        - Cost (USD)
+        - Elapsed time and ETA
+        
+        Progress updates are throttled to emit at most every 5 seconds to avoid
+        overwhelming the console. This method is called after significant events
+        (trial completion, iteration completion, budget updates).
+        
+        Note:
+            Only emits if start_time is set (job has started) and sufficient time
+            has passed since the last update.
+        """
+        import time
+        
+        if self.mipro_start_time is None:
+            return
+        
+        # Throttle progress updates - only emit every N seconds
+        now = time.time()
+        if self._last_progress_emit_time is not None:
+            time_since_last = now - self._last_progress_emit_time
+            if time_since_last < self._progress_emit_interval:
+                return  # Skip this update
+        
+        self._last_progress_emit_time = now
+        
+        timestamp = datetime.now().strftime("%H:%M:%S")
+        elapsed = now - self.mipro_start_time
+        
+        parts = []
+        
+        # Overall progress percentage
+        percent_overall = None
+        if self.mipro_total_trials and self.mipro_completed_trials is not None:
+            percent_overall = (self.mipro_completed_trials / self.mipro_total_trials) * 100
+            parts.append(f"{int(percent_overall)}% complete")
+        
+        # Trial progress (like rollouts in GEPA)
+        if self.mipro_total_trials and self.mipro_completed_trials is not None:
+            parts.append(f"trials={self.mipro_completed_trials}/{self.mipro_total_trials}")
+            # Calculate remaining trials
+            remaining_trials = self.mipro_total_trials - self.mipro_completed_trials
+            if remaining_trials > 0:
+                parts.append(f"rem={remaining_trials}")
+            # Show percentage
+            if percent_overall is not None:
+                parts.append(f"({int(percent_overall)}%)")
+        elif self.mipro_completed_trials is not None:
+            parts.append(f"trials={self.mipro_completed_trials}")
+        
+        # Iteration progress
+        if self.mipro_num_iterations and self.mipro_current_iteration is not None:
+            parts.append(f"iter={self.mipro_current_iteration + 1}/{self.mipro_num_iterations}")
+        
+        # Rollouts completed vs max (like GEPA) - always show if we have any rollouts
+        if self.mipro_rollouts_completed > 0:
+            # Always try to show max if available (from TOML, event, or estimate)
+            max_rollouts_to_show = self.mipro_max_rollouts
+            if max_rollouts_to_show is None and self.mipro_total_trials and self.mipro_batch_size:
+                # Estimate max rollouts from total trials if available
+                    max_rollouts_to_show = self.mipro_total_trials * self.mipro_batch_size
+            
+            if max_rollouts_to_show:
+                rollouts_pct = (self.mipro_rollouts_completed / max_rollouts_to_show) * 100
+                parts.append(f"rollouts={self.mipro_rollouts_completed}/{max_rollouts_to_show} ({int(rollouts_pct)}%)")
+            else:
+                parts.append(f"rollouts={self.mipro_rollouts_completed}")
+        
+        # Tokens (policy tokens only, like GEPA rollout_tokens) - always show max if available
+        if self.mipro_policy_tokens > 0:
+            rollout_tokens_millions = self.mipro_policy_tokens / 1_000_000.0
+            if self.mipro_max_tokens:
+                # Use max_tokens as budget for rollout tokens (approximation)
+                budget_millions = self.mipro_max_tokens / 1_000_000.0
+                tokens_pct = (self.mipro_policy_tokens / self.mipro_max_tokens * 100) if self.mipro_max_tokens > 0 else 0
+                parts.append(f"tokens={rollout_tokens_millions:.2f}M/{budget_millions:.2f}M ({int(tokens_pct)}%)")
+            else:
+                parts.append(f"tokens={rollout_tokens_millions:.2f}M")
+        
+        # Timing (elapsed out of max, like GEPA)
+        elapsed_seconds = int(elapsed)
+        if self.mipro_max_time_seconds:
+            elapsed_pct = (elapsed / self.mipro_max_time_seconds * 100) if self.mipro_max_time_seconds > 0 else 0
+            max_time_minutes = self.mipro_max_time_seconds / 60.0
+            if elapsed_seconds >= 60:
+                elapsed_str = f"{elapsed_seconds / 60:.1f}min/{max_time_minutes:.1f}min ({int(elapsed_pct)}%)"
+            else:
+                elapsed_str = f"{elapsed_seconds}s/{int(self.mipro_max_time_seconds)}s ({int(elapsed_pct)}%)"
+        else:
+            if elapsed_seconds >= 60:
+                elapsed_str = f"{elapsed_seconds / 60:.1f}min"
+            else:
+                elapsed_str = f"{elapsed_seconds}s"
+        parts.append(f"elapsed={elapsed_str}")
+        
+        # ETA calculation (similar to GEPA) - always show if we have progress
+        eta_seconds = None
+        if self.mipro_completed_trials is not None and self.mipro_completed_trials > 0 and elapsed > 0:
+            rate = self.mipro_completed_trials / elapsed
+            if rate > 0:
+                if self.mipro_total_trials:
+                    # Calculate ETA based on remaining trials
+                    remaining = self.mipro_total_trials - self.mipro_completed_trials
+                    if remaining > 0:
+                        eta_seconds = remaining / rate
+                else:
+                    # Estimate based on iterations if we don't have total trials
+                    if self.mipro_num_iterations and self.mipro_current_iteration is not None:
+                        remaining_iterations = self.mipro_num_iterations - (self.mipro_current_iteration + 1)
+                        if remaining_iterations > 0 and self.mipro_trials_per_iteration:
+                            # Estimate: assume same rate for remaining iterations
+                            remaining_trials_estimate = remaining_iterations * self.mipro_trials_per_iteration
+                            eta_seconds = remaining_trials_estimate / rate
+        
+        if eta_seconds is not None and eta_seconds > 0:
+            eta_str = f"{eta_seconds / 60:.1f}min" if eta_seconds >= 60 else f"{int(eta_seconds)}s"
+            parts.append(f"eta={eta_str}")
+        
+        if parts:
+            progress_msg = " ".join(parts)
+            self._write_log(f"[{timestamp}] Progress: {progress_msg}")
+    
+    def flush(self) -> None:
+        """Flush buffered output and close log file."""
+        if self._log_file_handle:
+            try:
+                from datetime import datetime
+                self._log_file_handle.write("\n" + "=" * 80 + "\n")
+                self._log_file_handle.write(f"Ended: {datetime.now().strftime('%Y-%m-%d %H:%M:%S')}\n")
+                self._log_file_handle.write("=" * 80 + "\n")
+                self._log_file_handle.flush()
+                self._log_file_handle.close()
+            except Exception:
+                pass
+            finally:
+                self._log_file_handle = None
+    
+    def _handle_proposal_scored(self, event_data: dict[str, Any]) -> None:
+        """Handle GEPA proposal scored events (transformations).
+        
+        Displays transformation/proposal scoring events from GEPA optimization.
+        Only called if show_transformations is True (default: False) to avoid
+        verbose output. Shows the score assigned to each proposed transformation.
+        
+        Args:
+            event_data: Event data dictionary containing:
+                - data.score: Score assigned to the transformation/proposal
+        """
+        # Only called if show_transformations=True
+        data = event_data.get("data", {})
+        if not isinstance(data, dict):
+            return
+        
+        timestamp = datetime.now().strftime("%H:%M:%S")
+        score = data.get("score")
+        if score is not None:
+            click.echo(f"[{timestamp}] Proposal scored: {score:.4f}")
+
+
+__all__ = [
+    "GraphGenHandler",
+    "BufferedHandler",
+    "CallbackHandler",
+    "CLIHandler",
+    "PromptLearningHandler",
+    "JSONHandler",
+    "IntegrationTestHandler",
+    "LossCurveHandler",
+    "RichHandler",
+    "StreamHandler",
+]