refactor(harbor): remove project log summary

Author: AbhiPrasad

docs/harbor-braintrust-plugin-design.md

@@ -16,7 +16,7 @@ Harbor final trial             → Braintrust root eval span
 Harbor verifier rewards        → Braintrust scores or metrics
 Harbor verifier labels         → Braintrust classifications
 Harbor lifecycle + ATIF        → child spans
-Harbor JobResult               → final reconciliation + job summary
+Harbor JobResult               → final reconciliation
 ```
 
 ### Key decisions
@@ -57,7 +57,7 @@ Version 1 does not need to:
 
 | Harbor | Braintrust | Notes |
 |---|---|---|
-| Job | Shared metadata + optional project-log summary | A job may create several experiments. |
+| Job | Shared metadata on experiments and eval roots | A job may create several experiments. |
 | Resolved dataset/task set | Dataset | Scope is the exact logical task selection. |
 | Task | Dataset record | Deterministic ID and canonical task input. |
 | Eval group/variant | Experiment | Partition by dataset and semantic agent config. |
@@ -307,7 +307,7 @@ Version 1 should only read labels from known Harbor adapters or explicit `classi
 
 If `reward-details.json` exists, put a bounded summary in the scorer output and optionally attach the complete JSON. Do not create criterion-level or judge-LLM spans until Rewardkit has a stable, tested mapping.
 
-Harbor's job metrics, custom `metric.py`, and pass@k remain authoritative aggregate results. Do not create a synthetic eval row for aggregates. Optionally write one project-log trace, `harbor.job.summary`, containing exact Harbor aggregates and links to partition experiments.
+Harbor's job metrics, custom `metric.py`, and pass@k remain authoritative aggregate results. Do not create a synthetic eval row or project-log trace for aggregates.
 
 ## ATIF import
 
@@ -495,7 +495,6 @@ HarborPlugin(
     max_attachment_bytes=5_000_000,
     max_total_attachment_bytes=20_000_000,
     max_content_bytes=20_000,
-    log_job_summary=True,
     log_retry_attempts=False,
     strict=False,
 )
@@ -527,7 +526,7 @@ Do not mutate private fields or depend on queue, metrics, progress, existing-tri
 2. Validate config/auth and read resolved tasks, lock, and custom metadata.
 3. Normalize metadata, build partitions, sync datasets, and initialize experiments.
 4. Register all trial hooks with one thin dispatcher.
-5. Dispatch job `READY`; optionally start a project-log job trace.
+5. Dispatch job `READY`.
 
 Keep blocking SDK and filesystem work off Harbor's event loop. Preserve per-trial effect ordering while allowing independent trials to make progress concurrently.
 

py/src/braintrust/integrations/harbor/config.py

@@ -103,7 +103,6 @@ class PluginConfig:
     max_attachment_bytes: int = 5_000_000
     max_total_attachment_bytes: int = 20_000_000
     max_content_bytes: int = 20_000
-    log_job_summary: bool = True
     log_retry_attempts: bool = False
     strict: bool = False
     redact_patterns: tuple[str, ...] = ()
@@ -119,7 +118,6 @@ def from_options(cls, **options: Any) -> "PluginConfig":
         for name in (
             "include_custom_metadata",
             "include_tracebacks",
-            "log_job_summary",
             "log_retry_attempts",
             "strict",
         ):

py/src/braintrust/integrations/harbor/plugin.py

@@ -14,7 +14,7 @@
 from pathlib import Path
 from typing import Any
 
-from braintrust.logger import Attachment, flush, init, init_dataset, init_logger
+from braintrust.logger import Attachment, flush, init, init_dataset
 from exceptiongroup import ExceptionGroup
 
 from .atif import _INSTRUMENTATION, ATIFImportResult, import_trajectory, summarize_trajectory
@@ -263,7 +263,6 @@ def __init__(
         max_attachment_bytes: Any = _UNSET,
         max_total_attachment_bytes: Any = _UNSET,
         max_content_bytes: Any = _UNSET,
-        log_job_summary: Any = _UNSET,
         log_retry_attempts: Any = _UNSET,
         strict: Any = _UNSET,
         **kwargs: Any,
@@ -291,7 +290,6 @@ def __init__(
             "max_attachment_bytes": max_attachment_bytes,
             "max_total_attachment_bytes": max_total_attachment_bytes,
             "max_content_bytes": max_content_bytes,
-            "log_job_summary": log_job_summary,
             "log_retry_attempts": log_retry_attempts,
             "strict": strict,
             **kwargs,
@@ -356,7 +354,7 @@ async def reconcile(result: Any) -> None:
         for name in set(self._trial_machines) - final_names:
             await self._dispatch(name, TrialEvent(TrialEventKind.OMIT))
         try:
-            await asyncio.to_thread(self._finalize, job_result)
+            await asyncio.to_thread(flush)
         except Exception as exc:
             failures.append(exc)
             self._errors.append(f"final flush: {exc}")
@@ -813,39 +811,6 @@ def _sync_final_result(self, result: Any) -> None:
         root.log(metadata=metadata)
         root.end(end_time=root_end)
 
-    def _finalize(self, job_result: Any) -> None:
-        if self._runtime is None:
-            return
-        if self.config.log_job_summary:
-            project_logger = init_logger(
-                project=self.config.project_name,
-                project_id=self.config.project_id,
-                set_current=False,
-            )
-            now = datetime.now().timestamp()
-            summary_start, summary_end = _timing(job_result, now, now)
-            summary = project_logger.start_span(
-                name="harbor.job.summary",
-                type="task",
-                id=f"harbor-job-summary-{self._runtime.snapshot.job_id}",
-                start_time=summary_start,
-                set_current=False,
-                input={"job_id": self._runtime.snapshot.job_id},
-                metadata={
-                    "harbor": {
-                        "job_id": self._runtime.snapshot.job_id,
-                        "experiments": [
-                            {"id": partition.experiment_id, "name": partition.name}
-                            for partition in self._runtime.partitions.values()
-                        ],
-                    }
-                },
-                internal={"instrumentation": _INSTRUMENTATION},
-            )
-            summary.log(output=job_result.stats.model_dump(mode="json", exclude_none=False))
-            summary.end(end_time=summary_end)
-        flush()
-
     def _persist_disabled_manifest(self) -> None:
         if self._snapshot is None:
             return