wafer-core 0.1.37__py3-none-any.whl → 0.1.39__py3-none-any.whl

wafer_core/rollouts/evaluation.py

@@ -17,6 +17,7 @@ from typing import Any
 
 import trio
 
+from ._logging.logging_config import EvalLoggingContext, setup_eval_logging
 from .agents import run_agent
 from .dtypes import (
     Actor,
@@ -29,16 +30,19 @@ from .dtypes import (
     Score,
     StopReason,
     StreamChunk,
+    TextDelta,
     TextEnd,
+    ThinkingDelta,
     ToolExecutionEnd,
     Trajectory,
 )
-from .events import EventEmitter, emit_event
-from .progress import MultiProgress
 from .training.types import Sample, Status
 
 logger = logging.getLogger(__name__)
 
+# Logger for structured eval events — handlers configured by setup_eval_logging()
+_event_logger = logging.getLogger("wafer.eval.events")
+
 
 # ── Runtime Context ───────────────────────────────────────────────────────────
 
@@ -56,7 +60,6 @@ class EvalRuntime:
     config: EvalConfig
     api_limiter: trio.CapacityLimiter | None = None
     tool_limiter: trio.CapacityLimiter | None = None
-    progress: MultiProgress | None = None
 
 
 # JSON-like recursive type for sanitize_api_keys
@@ -187,17 +190,12 @@ async def _evaluate_batch(
                            Used for incremental report writing.
     """
     config = runtime.config
-    progress = runtime.progress
     results: list[Sample] = []
     # Lock for thread-safe results access during concurrent execution
     results_lock = trio.Lock()
 
     async def run_one(sample_id: str, sample_data: dict[str, Any]) -> Sample:
         """Evaluate a single sample."""
-        task_name = sample_data.get("name", sample_id)
-        if progress:
-            progress.add_task(sample_id, name=task_name)
-
         # Get environment: prefer direct environment, fall back to factory
         if config.environment is not None:
             env = config.environment
@@ -212,17 +210,6 @@ async def _evaluate_batch(
             environment=env,
         )
 
-        # Mark task complete
-        if progress:
-            reward = result.score.reward if result.score else 0.0
-            success = result.metadata.get("status") == "success"
-            if success:
-                message = f"reward={reward:.2f}"
-            else:
-                error = result.metadata.get("error", "failed")
-                message = error[:30] if len(error) > 30 else error
-            progress.complete_task(sample_id, success=success, message=message)
-
         return result
 
     if config.max_concurrent == 1:
@@ -477,11 +464,8 @@ def _strip_ansi(text: str) -> str:
 
 
 def _truncate(text: str, max_len: int = 2000) -> str:
-    """Truncate long text and strip ANSI codes."""
-    text = _strip_ansi(text)
-    if len(text) <= max_len:
-        return text
-    return text[:max_len] + f"\n\n... (truncated, {len(text)} chars total)"
+    """Strip ANSI codes from text."""
+    return _strip_ansi(text)
 
 
 def _format_message_content(content: Any) -> str:
@@ -642,7 +626,9 @@ class EvalReport:
     timestamp: str = field(default_factory=lambda: datetime.now().isoformat())
     git_info: dict[str, Any] = field(default_factory=_get_git_info)
     config_path: str | None = None  # Path to config file relative to repo root
-    metadata: dict[str, Any] | None = None  # Custom metadata (waferbench_category, github_runner, etc.)
+    metadata: dict[str, Any] | None = (
+        None  # Custom metadata (waferbench_category, github_runner, etc.)
+    )
 
     async def save(self, output_dir: Path) -> None:
         """Save evaluation results to directory."""
@@ -974,7 +960,6 @@ async def evaluate_sample(
     """
     # Unpack runtime for convenience
     config = runtime.config
-    progress = runtime.progress
 
     # Prepare initial messages from sample
     initial_messages = config.prepare_messages(sample_data)
@@ -1004,73 +989,101 @@ async def evaluate_sample(
     async def on_chunk_with_sample_id(event: object) -> None:
         nonlocal last_status, current_turn
 
-        # Update MultiProgress on various events for granular status
         status = _get_progress_status_for_event(event)
-        turn = _get_turn_from_event(event)
-
-        if progress is not None:
-            if status is not None or turn is not None:
-                progress.update_task(
-                    sample_id,
-                    turn=turn if turn is not None else None,
-                    status=status if status is not None else None,
-                )
 
-        # Emit to file for TUI - only on status changes to avoid flooding
+        # Emit to JSONL files via logging — overview (INFO+) and per-sample (all levels)
         if isinstance(event, StreamChunk):
             if event.type == "turn_start":
                 turn_num = event.data.get("turn", 0)
                 current_turn[sample_id] = turn_num
-                emit_event("turn", id=sample_id, turn=turn_num, status="waiting")
+                _event_logger.info(
+                    "turn",
+                    extra={
+                        "sample_id": sample_id,
+                        "turn": turn_num,
+                        "status": "waiting",
+                    },
+                )
                 last_status[sample_id] = "waiting"
             elif event.type == "modal_progress":
-                emit_event("modal_progress", id=sample_id, phase=event.data.get("phase", ""))
+                _event_logger.info(
+                    "modal_progress",
+                    extra={
+                        "sample_id": sample_id,
+                        "phase": event.data.get("phase", ""),
+                    },
+                )
 
-        # Emit status changes for LLM events (streaming, thinking, tool calls)
+        # Emit status changes (dedup to avoid flooding)
         if status is not None and status != last_status.get(sample_id):
-            emit_event("turn", id=sample_id, status=status)
+            _event_logger.info("turn", extra={"sample_id": sample_id, "status": status})
             last_status[sample_id] = status
 
-        # Wide events: emit detailed timing for performance analysis
+        # Wide events: detailed timing for performance analysis
         sample_turn = current_turn.get(sample_id, 0)
         if isinstance(event, LLMCallEnd):
-            emit_event(
+            _event_logger.info(
                 "llm_call",
-                id=sample_id,
-                turn=sample_turn,
-                duration_ms=round(event.duration_ms, 1),
-                provider=event.provider,
-                model=event.model,
-                tokens_in=event.tokens_in,
-                tokens_out=event.tokens_out,
-                status=event.status,
-                error=event.error,
+                extra={
+                    "sample_id": sample_id,
+                    "turn": sample_turn,
+                    "duration_ms": round(event.duration_ms, 1),
+                    "provider": event.provider,
+                    "model": event.model,
+                    "tokens_in": event.tokens_in,
+                    "tokens_out": event.tokens_out,
+                    "status": event.status,
+                    "error": event.error,
+                },
             )
         elif isinstance(event, ToolExecutionEnd):
-            emit_event(
+            _event_logger.info(
                 "tool_execution",
-                id=sample_id,
-                turn=sample_turn,
-                tool_name=event.tool_name,
-                duration_ms=round(event.duration_ms, 1),
-                status=event.status,
-                is_error=event.is_error,
-                result_summary=event.result_summary,
+                extra={
+                    "sample_id": sample_id,
+                    "turn": sample_turn,
+                    "tool_name": event.tool_name,
+                    "duration_ms": round(event.duration_ms, 1),
+                    "status": event.status,
+                    "is_error": event.is_error,
+                    "result_summary": event.result_summary,
+                },
             )
         elif isinstance(event, TextEnd):
-            # Emit assistant message content for observability
-            # Truncate long content to avoid bloating events file
+            # Truncate for events.jsonl (INFO), full content in per-sample (also INFO)
             content = event.content
             truncated = len(content) > 2000
             if truncated:
                 content = content[:2000] + "..."
-            emit_event(
+            _event_logger.info(
                 "assistant_message",
-                id=sample_id,
-                turn=sample_turn,
-                content=content,
-                content_length=len(event.content),
-                truncated=truncated,
+                extra={
+                    "sample_id": sample_id,
+                    "turn": sample_turn,
+                    "content": content,
+                    "content_length": len(event.content),
+                    "truncated": truncated,
+                },
+            )
+
+        # DEBUG: streaming deltas — per-sample files only (filtered out of events.jsonl)
+        elif isinstance(event, TextDelta):
+            _event_logger.debug(
+                "text_delta",
+                extra={
+                    "sample_id": sample_id,
+                    "turn": sample_turn,
+                    "text": event.delta,
+                },
+            )
+        elif isinstance(event, ThinkingDelta):
+            _event_logger.debug(
+                "thinking_delta",
+                extra={
+                    "sample_id": sample_id,
+                    "turn": sample_turn,
+                    "text": event.delta,
+                },
             )
 
         # Wrap event with sample_id and forward to base handler
@@ -1107,13 +1120,13 @@ async def evaluate_sample(
         )
     )
 
-    # Also emit to file for TUI (if emitter configured)
+    # Emit sample_start for progress display
     # TODO: Retry logic can emit multiple sample_start events for the same sample_id
     # without a corresponding sample_end, causing progress display to show 100/100
     # while a sample is still running. Either emit sample_end before retry, or
     # don't emit sample_start on retries. See: chiraag/supabase-eval-traces PR #504
     sample_name = sample_data.get("name", sample_id)
-    emit_event("sample_start", id=sample_id, name=sample_name)
+    _event_logger.info("sample_start", extra={"sample_id": sample_id, "sample_name": sample_name})
 
     # Run agent with error handling
     result = await _run_agent_with_error_handling(initial_state, run_config, sample_id)
@@ -1202,8 +1215,7 @@ async def evaluate_sample(
         )
     )
 
-    # Also emit to file for TUI (if emitter configured)
-    emit_event("sample_end", id=sample_id, score=reward)
+    _event_logger.info("sample_end", extra={"sample_id": sample_id, "score": reward})
 
     return sample
 
@@ -1296,13 +1308,26 @@ async def evaluate(
         logger.info(f"max concurrent: {config.max_concurrent}")
         logger.debug("=" * 50)
 
-    # Initialize event emitter for TUI progress (writes to events.jsonl)
-    # This is separate from MultiProgress - events go to file for external TUI
-    emitter: EventEmitter | None = None
+    # Set up eval logging: events.jsonl (overview) + samples/{id}.jsonl (per-sample)
+    eval_logging: EvalLoggingContext | None = None
     if config.output_dir:
-        emitter = EventEmitter(output_dir=config.output_dir)
-        emitter.as_context()  # Make available via get_emitter()
-        emitter.emit("eval_start", name=config.eval_name, total=len(samples_to_eval))
+        eval_logging = setup_eval_logging(config.output_dir)
+        _event_logger.info(
+            "eval_start",
+            extra={
+                "eval_name": config.eval_name,
+                "total": len(samples_to_eval),
+            },
+        )
+
+    # Progress display: pytui subprocess that reads events.jsonl
+    # Requires output_dir since it reads events.jsonl from there
+    progress_ctx = None
+    if config.show_progress and config.output_dir:
+        from .progress_app import progress_display
+
+        progress_ctx = progress_display(output_dir=config.output_dir)
+        progress_ctx.__enter__()
 
     # Evaluate samples (with concurrency control)
     results: list[Sample] = []
@@ -1329,18 +1354,6 @@ async def evaluate(
             )
             last_report_count = len(all_results)
 
-    # Initialize progress display for sample-level tracking
-    # MultiProgress shows each concurrent sample with turn-by-turn updates
-    progress: MultiProgress | None = None
-    if config.show_progress:
-        progress = MultiProgress(
-            total=len(samples_to_eval),
-            desc=config.eval_name,
-            unit="sample",
-            verbose=config.verbose,  # verbose=True shows INFO logs, False shows only WARNING+
-        )
-        progress.__enter__()
-
     # Create two-level concurrency limiters if configured
     api_limiter = (
         trio.CapacityLimiter(config.max_api_concurrent)
@@ -1358,7 +1371,6 @@ async def evaluate(
         config=config,
         api_limiter=api_limiter,
         tool_limiter=tool_limiter,
-        progress=progress,
     )
 
     # Run initial evaluation batch with incremental report callback
@@ -1386,9 +1398,10 @@ async def evaluate(
         else:
             raise
 
-    # Close progress display
-    if progress:
-        progress.__exit__(None, None, None)
+    # Close progress TUI before any print() calls
+    if progress_ctx:
+        progress_ctx.__exit__(None, None, None)
+        progress_ctx = None
 
     # Write final partial report and upload if interrupted
     if interrupted and config.output_dir:
@@ -1409,10 +1422,17 @@ async def evaluate(
             print("Upload complete.")
         except Exception as e:
             print(f"Upload failed: {e}")
-        # Close emitter and exit
-        if emitter:
-            emitter.emit("eval_end", name=config.eval_name, total=len(results), interrupted=True)
-            emitter.close()
+        # Emit eval_end and clean up logging
+        if eval_logging:
+            _event_logger.info(
+                "eval_end",
+                extra={
+                    "eval_name": config.eval_name,
+                    "total": len(results),
+                    "interrupted": True,
+                },
+            )
+            eval_logging.teardown()
         # Exit cleanly - don't re-raise to avoid big traceback
         print(f"Partial results saved to {config.output_dir}")
         print("Resume with: --resume", config.output_dir)
@@ -1435,21 +1455,16 @@ async def evaluate(
             f"Retrying {len(failed_samples)} failed samples "
             f"(attempt {retry_attempt + 1}/{config.max_sample_retries}, waiting {wait_seconds}s)"
         )
-        if progress:
-            progress.log(retry_msg)
-        else:
-            logger.info(retry_msg)
+        logger.info(retry_msg)
         await trio.sleep(wait_seconds)
 
         # Remove failed samples and retry
         failed_ids = {sid for sid, _ in failed_samples}
         results = [r for r in results if r.id not in failed_ids]
-        # Create runtime without progress for retries (no incremental reports during retry)
         retry_runtime = EvalRuntime(
             config=config,
             api_limiter=api_limiter,
             tool_limiter=tool_limiter,
-            progress=None,
         )
         retry_results = await _evaluate_batch(failed_samples, retry_runtime)
         results.extend(retry_results)
@@ -1471,10 +1486,7 @@ async def evaluate(
         retry_result_msg = (
             f"Retry {retry_attempt + 1}: {succeeded} succeeded, {still_failed} still failing"
         )
-        if progress:
-            progress.log(retry_result_msg)
-        else:
-            logger.info(retry_result_msg)
+        logger.info(retry_result_msg)
 
     # Compute summary metrics
     summary_metrics = compute_summary_metrics(results)
@@ -1517,10 +1529,21 @@ async def evaluate(
             else:
                 logger.info(f"{key}: {value}")
 
-    # Close event emitter
-    if emitter:
-        emitter.emit("eval_end", name=config.eval_name, total=len(results))
-        emitter.close()
+    # Emit eval_end and clean up logging (before progress TUI so it sees eval_end)
+    if eval_logging:
+        _event_logger.info(
+            "eval_end",
+            extra={
+                "eval_name": config.eval_name,
+                "total": len(results),
+            },
+        )
+        eval_logging.teardown()
+
+    # Close progress TUI
+    if progress_ctx:
+        progress_ctx.__exit__(None, None, None)
+        progress_ctx = None
 
     return report
 
@@ -1571,7 +1594,7 @@ def compute_summary_metrics(results: list[Sample]) -> dict[str, float]:
                 median_val = (sorted_values[n // 2 - 1] + sorted_values[n // 2]) / 2
             else:
                 median_val = sorted_values[n // 2]
-            
+
             summary[f"mean_{metric_name}"] = mean_val
             summary[f"median_{metric_name}"] = median_val
             summary[f"min_{metric_name}"] = min(values)
@@ -1800,75 +1823,22 @@ def get_api_key(provider: str = "anthropic") -> str | None:
     return None
 
 
-def run_with_progress(
-    eval_fn: Callable[[Any], Any],
-    config: Any,
-    output_dir: Path,
-    quiet_config_fn: Callable[[Any], Any] | None = None,
-    async_wrapper: Callable[[Callable, Any], Callable[[], Any]] | None = None,
-) -> dict[str, Any]:
-    """Run evaluation with progress display TUI.
-
-    Wraps an async eval function with the progress_display context manager,
-    which redirects stdout/stderr to output.log and renders a TUI.
-
-    Args:
-        eval_fn: Async evaluation function that takes config and returns results
-        config: Configuration object
-        output_dir: Directory for output files (events.jsonl, output.log)
-        quiet_config_fn: Optional function to create a quiet version of config
-                        (disables internal verbose/show_progress flags)
-        async_wrapper: Optional wrapper for async runtime compatibility (e.g., trio_asyncio).
-                      Takes (eval_fn, config) and returns an async callable for trio.run().
-
-    Returns:
-        Results dict from eval_fn
-
-    Example:
-        def my_quiet_config(config):
-            return replace(config, run=replace(config.run, verbose=False, show_progress=False))
-
-        result = run_with_progress(
-            evaluate_my_task,
-            config,
-            config.output.output_dir,
-            quiet_config_fn=my_quiet_config,
-        )
-    """
-    from .progress_display import progress_display
-
-    # Apply quiet config transformation if provided
-    run_config = quiet_config_fn(config) if quiet_config_fn else config
-
-    with progress_display(output_dir=output_dir):
-        if async_wrapper:
-            result = trio.run(async_wrapper(eval_fn, run_config))
-        else:
-            result = trio.run(eval_fn, run_config)
-        assert result is not None, "Evaluation was cancelled"
-
-    return result
-
-
 def run_eval(
     eval_fn: Callable[[Any], Any],
     config: Any,
     output_dir: Path,
-    show_progress: bool = False,
-    quiet_config_fn: Callable[[Any], Any] | None = None,
     print_summary_fn: Callable[[dict[str, Any], Path], None] | None = None,
     async_wrapper: Callable[[Callable, Any], Callable[[], Any]] | None = None,
 ) -> dict[str, Any]:
     """Standard entry point for running evaluations.
 
-    Handles the common pattern of optionally wrapping eval in progress display.
+    Runs the async eval function via trio.run(). Progress display is handled
+    internally by evaluate() based on config.show_progress.
 
     Args:
         eval_fn: Async evaluation function that takes config and returns results
         config: Configuration object
         output_dir: Directory for output files
-        show_progress: Whether to show progress TUI
-        quiet_config_fn: Optional function to create quiet config for progress mode
         print_summary_fn: Optional function to print summary after completion
         async_wrapper: Optional wrapper for async runtime compatibility (e.g., trio_asyncio).
                       Takes (eval_fn, config) and returns an async callable for trio.run().
@@ -1883,8 +1853,6 @@ def run_eval(
                 eval_fn=evaluate_my_task,
                 config=config,
                 output_dir=config.output.output_dir,
-                show_progress=config.run.show_progress,
-                quiet_config_fn=lambda c: replace(c, run=replace(c.run, verbose=False, show_progress=False)),
                 print_summary_fn=print_my_summary,
             )
 
@@ -1904,14 +1872,11 @@ def run_eval(
             async_wrapper=asyncio_compat_wrapper,
         )
     """
-    if show_progress:
-        result = run_with_progress(eval_fn, config, output_dir, quiet_config_fn, async_wrapper)
+    if async_wrapper:
+        result = trio.run(async_wrapper(eval_fn, config))
     else:
-        if async_wrapper:
-            result = trio.run(async_wrapper(eval_fn, config))
-        else:
-            result = trio.run(eval_fn, config)
-        assert result is not None, "Evaluation was cancelled"
+        result = trio.run(eval_fn, config)
+    assert result is not None, "Evaluation was cancelled"
 
     if print_summary_fn:
         print_summary_fn(result, output_dir)