metergraphrelay 0.7.2__tar.gz → 0.7.3__tar.gz

{metergraphrelay-0.7.2/src/metergraphrelay.egg-info → metergraphrelay-0.7.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: metergraphrelay
-Version: 0.7.2
+Version: 0.7.3
 Summary: Move LLM trace data between systems: pull from providers, push to metergraph.
 Author: Vasiliy Radostev
 License-Expression: MIT
@@ -309,7 +309,7 @@ but LLM spans is ever imported, but `--name` has no effect there.
 
 **Setup:** nothing, for a local Phoenix. The default base URL is
 `http://localhost:6006`. For a remote or authenticated Phoenix, set these
-in `.env` (or pass `--base-url` / `--phoenix-api-key` per-command):
+in `.env` (and pass `--base-url` only when selecting a different endpoint):
 
     PHOENIX_BASE_URL=https://phoenix.example.com
     PHOENIX_API_KEY=...

{metergraphrelay-0.7.2 → metergraphrelay-0.7.3}/README.md

@@ -286,7 +286,7 @@ but LLM spans is ever imported, but `--name` has no effect there.
 
 **Setup:** nothing, for a local Phoenix. The default base URL is
 `http://localhost:6006`. For a remote or authenticated Phoenix, set these
-in `.env` (or pass `--base-url` / `--phoenix-api-key` per-command):
+in `.env` (and pass `--base-url` only when selecting a different endpoint):
 
     PHOENIX_BASE_URL=https://phoenix.example.com
     PHOENIX_API_KEY=...

{metergraphrelay-0.7.2 → metergraphrelay-0.7.3}/pyproject.toml

@@ -1,10 +1,10 @@
 [build-system]
-requires = ["setuptools>=77"]
+requires = ["setuptools==80.9.0"]
 build-backend = "setuptools.build_meta"
 
 [project]
 name = "metergraphrelay"
-version = "0.7.2"
+version = "0.7.3"
 description = "Move LLM trace data between systems: pull from providers, push to metergraph."
 readme = "README.md"
 requires-python = ">=3.10"

metergraphrelay-0.7.3/src/metergraphrelay/__init__.py

@@ -0,0 +1 @@
+__version__ = "0.7.3"

{metergraphrelay-0.7.2 → metergraphrelay-0.7.3}/src/metergraphrelay/cli.py

@@ -12,7 +12,7 @@ from openai import OpenAI
 from .config import ConfigError, require_credentials
 from .demo import run_demo
 from .metergraph_sync import MeterGraphSyncClient, MeterGraphSyncError
-from .portkey_sync import run_portkey_sync
+from .portkey_sync import _make_content_blind, run_portkey_sync
 from .provider_sync import UNBOUNDED_COUNT, run_pull_sync
 from .providers.braintrust import (
     DEFAULT_BRAINTRUST_URL,
@@ -159,24 +159,6 @@ def build_parser() -> argparse.ArgumentParser:
         default=".env",
         help="Path to a .env file to load credentials from. (default: .env)",
     )
-    pull_langfuse_parser.add_argument(
-        "--langfuse-public-key",
-        default=None,
-        metavar="KEY",
-        help=(
-            "Langfuse public key (Basic Auth username). Overrides "
-            "$LANGFUSE_PUBLIC_KEY / .env if given; env/.env is the preferred path."
-        ),
-    )
-    pull_langfuse_parser.add_argument(
-        "--langfuse-secret-key",
-        default=None,
-        metavar="KEY",
-        help=(
-            "Langfuse secret key (Basic Auth password). Overrides "
-            "$LANGFUSE_SECRET_KEY / .env if given; env/.env is the preferred path."
-        ),
-    )
 
     pull_braintrust_parser = pull_subparsers.add_parser(
         "braintrust",
@@ -262,15 +244,6 @@ def build_parser() -> argparse.ArgumentParser:
         default=".env",
         help="Path to a .env file to load credentials from. (default: .env)",
     )
-    pull_braintrust_parser.add_argument(
-        "--braintrust-api-key",
-        default=None,
-        metavar="KEY",
-        help=(
-            "Braintrust API key (Bearer token). Overrides $BRAINTRUST_API_KEY "
-            "/ .env if given; env/.env is the preferred path."
-        ),
-    )
 
     pull_langsmith_parser = pull_subparsers.add_parser(
         "langsmith",
@@ -350,13 +323,6 @@ def build_parser() -> argparse.ArgumentParser:
         "--env-file", default=".env",
         help="Path to a .env file to load credentials from. (default: .env)",
     )
-    pull_langsmith_parser.add_argument(
-        "--langsmith-api-key", default=None, metavar="KEY",
-        help=(
-            "LangSmith API key. Overrides $LANGSMITH_API_KEY / .env if given; "
-            "env/.env is the preferred path."
-        ),
-    )
 
     pull_phoenix_parser = pull_subparsers.add_parser(
         "phoenix",
@@ -451,16 +417,6 @@ def build_parser() -> argparse.ArgumentParser:
         default=".env",
         help="Path to a .env file to load settings from. (default: .env)",
     )
-    pull_phoenix_parser.add_argument(
-        "--phoenix-api-key",
-        default=None,
-        metavar="KEY",
-        help=(
-            "Phoenix API key (Bearer token), only needed when the server has "
-            "authentication enabled. Overrides $PHOENIX_API_KEY / .env if "
-            "given. (default: none)"
-        ),
-    )
 
     sync_parser = subparsers.add_parser(
         "sync",
@@ -502,8 +458,8 @@ def build_parser() -> argparse.ArgumentParser:
             "--initial-since seeds only the first run; --max-window-seconds "
             "caps the window at 3600. One Portkey workspace per metergraph app "
             "(MVP). A 'busy' lease or a 'caught up' server exit 0 as clean "
-            "no-ops. In both modes request and response content is uploaded to "
-            "MeterGraph with no opt-out."
+            "no-ops. Scheduled imports omit request and response content unless "
+            "--include-content is explicitly passed."
         ),
         formatter_class=argparse.RawDescriptionHelpFormatter,
         help=(
@@ -565,6 +521,11 @@ def build_parser() -> argparse.ArgumentParser:
             "(both modes) and PORTKEY_API_KEY (API mode). (default: .env)"
         ),
     )
+    sync_portkey_parser.add_argument(
+        "--include-content",
+        action="store_true",
+        help="Include request and response content in the upload.",
+    )
 
     # --- sync langfuse / braintrust / phoenix: server-coordinated cron mode ---
     # One shape for the three cursor-paged providers. The server picks the
@@ -613,14 +574,6 @@ def build_parser() -> argparse.ArgumentParser:
             f"{DEFAULT_LANGFUSE_HOST})"
         ),
     )
-    sync_langfuse_parser.add_argument(
-        "--langfuse-public-key", default=None, metavar="KEY",
-        help="Langfuse public key. Overrides $LANGFUSE_PUBLIC_KEY / .env if given.",
-    )
-    sync_langfuse_parser.add_argument(
-        "--langfuse-secret-key", default=None, metavar="KEY",
-        help="Langfuse secret key. Overrides $LANGFUSE_SECRET_KEY / .env if given.",
-    )
 
     sync_braintrust_parser = sync_subparsers.add_parser(
         "braintrust",
@@ -653,10 +606,6 @@ def build_parser() -> argparse.ArgumentParser:
             "else the US data plane)"
         ),
     )
-    sync_braintrust_parser.add_argument(
-        "--braintrust-api-key", default=None, metavar="KEY",
-        help="Braintrust API key. Overrides $BRAINTRUST_API_KEY / .env if given.",
-    )
 
     sync_phoenix_parser = sync_subparsers.add_parser(
         "phoenix",
@@ -694,13 +643,6 @@ def build_parser() -> argparse.ArgumentParser:
             f"{DEFAULT_PHOENIX_URL})"
         ),
     )
-    sync_phoenix_parser.add_argument(
-        "--phoenix-api-key", default=None, metavar="KEY",
-        help=(
-            "Phoenix API key, only when the server has authentication enabled. "
-            "Overrides $PHOENIX_API_KEY / .env if given."
-        ),
-    )
 
     sync_langsmith_parser = sync_subparsers.add_parser(
         "langsmith",
@@ -742,10 +684,6 @@ def build_parser() -> argparse.ArgumentParser:
             f"$LANGSMITH_BASE_URL if set, else {DEFAULT_LANGSMITH_URL})"
         ),
     )
-    sync_langsmith_parser.add_argument(
-        "--langsmith-api-key", default=None, metavar="KEY",
-        help="LangSmith API key. Overrides $LANGSMITH_API_KEY / .env if given.",
-    )
 
     demo_parser = subparsers.add_parser(
         "demo", help="Run 1-2 demo conversations with store=True"
@@ -801,6 +739,11 @@ def _add_sync_common(parser: argparse.ArgumentParser, *, scope_help: str) -> Non
         default=".env",
         help="Path to a .env file to load credentials from. (default: .env)",
     )
+    parser.add_argument(
+        "--include-content",
+        action="store_true",
+        help="Explicitly opt this scheduled import into prompt and completion content.",
+    )
 
 
 def _config_error(exc: ConfigError) -> int:
@@ -827,8 +770,6 @@ def _resolve_langfuse_credentials(args: argparse.Namespace) -> tuple[str, str]:
     # settings (e.g. LANGFUSE_BASE_URL) may still live only in that file, not
     # the real process environment, and must still resolve.
     load_dotenv(args.env_file, override=True)
-    if args.langfuse_public_key and args.langfuse_secret_key:
-        return args.langfuse_public_key, args.langfuse_secret_key
     creds = require_credentials("langfuse", args.env_file)
     return creds["LANGFUSE_PUBLIC_KEY"], creds["LANGFUSE_SECRET_KEY"]
 
@@ -839,8 +780,6 @@ def _resolve_braintrust_credential(args: argparse.Namespace) -> str:
     # BRAINTRUST_BASE_URL) may live only in that file, not the real process
     # environment, and must still resolve.
     load_dotenv(args.env_file, override=True)
-    if args.braintrust_api_key:
-        return args.braintrust_api_key
     return require_credentials("braintrust", args.env_file)["BRAINTRUST_API_KEY"]
 
 
@@ -855,8 +794,6 @@ def _resolve_langsmith_credential(args: argparse.Namespace) -> str:
     # Load the selected --env-file unconditionally, even when the flag is
     # given, so optional settings such as LANGSMITH_ENDPOINT are read too.
     load_dotenv(args.env_file, override=True)
-    if args.langsmith_api_key:
-        return args.langsmith_api_key
     return require_credentials("langsmith", args.env_file)["LANGSMITH_API_KEY"]
 
 
@@ -911,6 +848,8 @@ def _run_sync_portkey(args: argparse.Namespace) -> int:
 
     try:
         converted, skipped = convert_portkey_export(args.export_file, tmp_path)
+        if not args.include_content:
+            _make_content_blind(tmp_path)
     except (OSError, UnicodeDecodeError) as exc:
         _cleanup_temp_file(tmp_path)
         return _os_error(exc)
@@ -1016,6 +955,7 @@ def _run_sync_portkey_api(args: argparse.Namespace) -> int:
             max_window_seconds=max_window,
             push_token=push_creds["METERGRAPH_APP_TOKEN"],
             ingest_base_url=ingest_base,
+            content_opted_in=args.include_content,
         )
     except (MeterGraphSyncError, PortkeyExportError, OSError) as exc:
         print(f"Error: {exc}", file=sys.stderr)
@@ -1115,6 +1055,7 @@ def _run_sync_pull(args: argparse.Namespace, *, source: str, source_scope: str,
             ingest_base_url=ingest_base,
             provider_errors=provider_errors,
             allow_skipped=args.allow_skipped,
+            content_opted_in=args.include_content,
         )
     except (MeterGraphSyncError, OSError, *provider_errors) as exc:
         print(f"Error: {exc}", file=sys.stderr)
@@ -1224,7 +1165,7 @@ def _run_sync_phoenix(args: argparse.Namespace) -> int:
     except ConfigError as exc:
         return _config_error(exc)
     # require_credentials() has loaded the env file; optional settings follow.
-    api_key = args.phoenix_api_key or os.environ.get("PHOENIX_API_KEY") or None
+    api_key = os.environ.get("PHOENIX_API_KEY") or None
     base_url = (
         args.base_url or os.environ.get("PHOENIX_BASE_URL") or DEFAULT_PHOENIX_URL
     )
@@ -1316,7 +1257,7 @@ def _run_pull_phoenix(args: argparse.Namespace) -> int:
     # the .env file is loaded for its optional settings rather than through
     # require_credentials, which would fail on a missing key.
     load_dotenv(args.env_file, override=True)
-    api_key = args.phoenix_api_key or os.environ.get("PHOENIX_API_KEY") or None
+    api_key = os.environ.get("PHOENIX_API_KEY") or None
     base_url = (
         args.base_url or os.environ.get("PHOENIX_BASE_URL") or DEFAULT_PHOENIX_URL
     )

metergraphrelay-0.7.3/src/metergraphrelay/finish_reason.py

@@ -0,0 +1,107 @@
+"""Extract provider completion reasons without guessing from token counts."""
+
+from __future__ import annotations
+
+from collections.abc import Mapping
+from typing import Any
+
+_MAX_REASON_LENGTH = 128
+# LangChain runs nest the reason as outputs.generations[[...]].generation_info
+# or .message.kwargs.response_metadata, six levels below the run.
+_MAX_DEPTH = 8
+_REASON_KEYS = ("stop_reason", "finish_reason", "finishReason")
+_REASON_CONTAINER_KEYS = (
+    "incomplete_details",
+    "response",
+    "output",
+    "outputs",
+    "metadata",
+    "invocation_params",
+    "generations",
+    "generation_info",
+    "message",
+    "kwargs",
+    "response_metadata",
+    "extra",
+)
+
+
+def _string_reason(value: object) -> str | None:
+    if not isinstance(value, str):
+        return None
+    reason = value.strip()
+    if not reason or len(reason) > _MAX_REASON_LENGTH:
+        return None
+    return reason
+
+
+def _reason_value(value: object) -> str | None:
+    direct = _string_reason(value)
+    if direct is not None:
+        return direct
+    if isinstance(value, (list, tuple)):
+        for item in value:
+            reason = _reason_value(item)
+            if reason is not None:
+                return reason
+    return None
+
+
+def _find_reason(value: object, *, depth: int = 0) -> str | None:
+    if depth > _MAX_DEPTH:
+        return None
+    if isinstance(value, (list, tuple)):
+        for item in value:
+            reason = _find_reason(item, depth=depth + 1)
+            if reason is not None:
+                return reason
+        return None
+    if not isinstance(value, Mapping):
+        return None
+
+    for key in _REASON_KEYS + ("finish_reasons", "finishReasons"):
+        reason = _reason_value(value.get(key))
+        if reason is not None:
+            return reason
+    incomplete_details = value.get("incomplete_details")
+    if isinstance(incomplete_details, Mapping):
+        reason = _reason_value(incomplete_details.get("reason"))
+        if reason is not None:
+            return reason
+    for key in ("choices", "candidates") + tuple(
+        key for key in _REASON_CONTAINER_KEYS if key != "incomplete_details"
+    ):
+        reason = _find_reason(value.get(key), depth=depth + 1)
+        if reason is not None:
+            return reason
+    return None
+
+
+def extract_stop_reason(*values: Any) -> str | None:
+    """Return the first known provider reason from documented response shapes."""
+    for value in values:
+        reason = _reason_value(value) or _find_reason(value)
+        if reason is not None:
+            return reason
+    return None
+
+
+def extract_phoenix_stop_reason(attributes: Mapping[str, Any]) -> str | None:
+    """Read finish reasons from Phoenix's flattened OpenTelemetry attributes."""
+    reason = extract_stop_reason(attributes)
+    if reason is not None:
+        return reason
+    for key, value in attributes.items():
+        if (
+            key.startswith("llm.output_messages.")
+            and key.endswith((".message.finish_reason", ".message.stop_reason"))
+        ) or key in {
+            "gen_ai.response.finish_reason",
+            "gen_ai.response.finish_reasons",
+            "llm.response.finish_reason",
+            "llm.response.stop_reason",
+        }:
+            reason = _reason_value(value) or _find_reason(value)
+            if reason is not None:
+                return reason
+    return None

metergraphrelay-0.7.3/src/metergraphrelay/http_limits.py

@@ -0,0 +1,36 @@
+"""Small bounds shared by provider HTTP clients."""
+
+from __future__ import annotations
+
+
+MAX_JSON_RESPONSE_BYTES = 2 * 1024 * 1024
+READ_CHUNK_SIZE = 64 * 1024
+
+
+class ResponseTooLarge(ValueError):
+    """Raised when a provider response exceeds the JSON body budget."""
+
+
+def read_bounded(response, *, max_bytes: int = MAX_JSON_RESPONSE_BYTES) -> bytes:
+    """Read a response without allowing an untrusted body to fill memory."""
+    declared = response.headers.get("Content-Length")
+    if declared:
+        try:
+            declared_bytes = int(declared)
+        except (TypeError, ValueError):
+            pass
+        else:
+            if declared_bytes > max_bytes:
+                raise ResponseTooLarge(
+                    f"response exceeds the {max_bytes}-byte JSON body limit"
+                )
+
+    # ``read(n)`` is bounded by the requested byte count for urllib response
+    # objects. Reading one byte beyond the budget detects oversized bodies
+    # without ever retaining more than the allowed response plus one byte.
+    body = response.read(max_bytes + 1)
+    if len(body) > max_bytes:
+        raise ResponseTooLarge(
+            f"response exceeds the {max_bytes}-byte JSON body limit"
+        )
+    return body

{metergraphrelay-0.7.2 → metergraphrelay-0.7.3}/src/metergraphrelay/import_identity.py

@@ -28,6 +28,7 @@ class ImportIdentityError(ValueError):
 class ImportContext:
     source: str
     source_scope: str
+    content_opted_in: bool = False
 
 
 def canonical_import_event_id(raw: Any) -> str:

{metergraphrelay-0.7.2 → metergraphrelay-0.7.3}/src/metergraphrelay/portkey_sync.py

@@ -26,6 +26,7 @@ boundaries force an unconditional renewal.
 from __future__ import annotations
 
 import os
+import json
 import sys
 import tempfile
 import time
@@ -75,6 +76,7 @@ def run_portkey_sync(
     max_poll_seconds: float = MAX_POLL_SECONDS,
     renew_interval_seconds: float = RENEW_INTERVAL_SECONDS,
     push: Callable[..., tuple[int, int]] = push_file,
+    content_opted_in: bool = False,
 ) -> SyncOutcome:
     # Acquire is the one step before a lease exists: any failure here (network error,
     # unknown status, or an "acquired" response with no lease) must NOT attempt an
@@ -105,7 +107,11 @@ def run_portkey_sync(
         )
 
     lease = acquire.lease
-    ctx = ImportContext(source=PORTKEY_SOURCE, source_scope=source_scope)
+    ctx = ImportContext(
+        source=PORTKEY_SOURCE,
+        source_scope=source_scope,
+        content_opted_in=content_opted_in,
+    )
     created_export_ids: list[str] = []  # every draft this run created, for best-effort cancel
 
     def renew() -> None:
@@ -138,6 +144,8 @@ def run_portkey_sync(
                 _, sk = convert_portkey_export(
                     raw, converted_path, import_context=ctx, on_progress=renewer.tick
                 )
+                if not content_opted_in:
+                    _make_content_blind(converted_path)
                 _discard_file(raw)  # raw fully consumed — free it before the upload
                 skipped += sk
                 renewer.force()  # normalize done — force before the row-by-row upload
@@ -236,6 +244,20 @@ def _plan_exports(pk_client, lease, created_export_ids: list[str], renewer) -> l
     return export_ids
 
 
+def _make_content_blind(path: str) -> None:
+    """Remove request and response content before a scheduled upload."""
+    temporary = f"{path}.content-blind"
+    with open(path, encoding="utf-8") as source, open(temporary, "w", encoding="utf-8") as target:
+        for line in source:
+            row = json.loads(line)
+            if isinstance(row, dict):
+                for field in ("request_json", "request_text", "response_text", "tool_calls"):
+                    row.pop(field, None)
+                row["content_opted_in"] = False
+            target.write(json.dumps(row, separators=(",", ":")) + "\n")
+    os.replace(temporary, path)
+
+
 def _poll_all(export_ids, pk_client, renewer, sleep, poll_interval, max_poll_seconds, clock) -> None:
     """Poll every export to a terminal state, renewing the lease on a time cadence.
 

{metergraphrelay-0.7.2 → metergraphrelay-0.7.3}/src/metergraphrelay/provider_sync.py

@@ -37,6 +37,7 @@ import os
 import sys
 import tempfile
 import time
+import json
 from typing import Callable, Protocol
 
 from .metergraph_sync import LeaseLostError, MeterGraphSyncError
@@ -76,6 +77,7 @@ def run_pull_sync(
     ingest_base_url: str | None,
     provider_errors: tuple[type[Exception], ...] = (),
     allow_skipped: bool = False,
+    content_opted_in: bool = False,
     work_dir: str | None = None,
     clock: Callable[[], float] = time.monotonic,
     renew_interval_seconds: float = RENEW_INTERVAL_SECONDS,
@@ -111,7 +113,11 @@ def run_pull_sync(
         )
 
     lease = acquire.lease
-    ctx = ImportContext(source=source, source_scope=source_scope)
+    ctx = ImportContext(
+        source=source,
+        source_scope=source_scope,
+        content_opted_in=content_opted_in,
+    )
 
     def renew() -> None:
         mg_client.renew(lease.lease_id)
@@ -130,6 +136,8 @@ def run_pull_sync(
                 import_context=ctx,
                 on_progress=renewer.tick,
             )
+            if not ctx.content_opted_in:
+                _make_content_blind(rows_path)
             renewer.force()  # pull done — force before the row-by-row upload
             if skipped and not allow_skipped:
                 # In sync mode there is no export file to recover a skipped row
@@ -204,6 +212,20 @@ def run_pull_sync(
         raise
 
 
+def _make_content_blind(path: str) -> None:
+    """Remove provider content before a scheduled import can upload it."""
+    temporary = f"{path}.content-blind"
+    with open(path, encoding="utf-8") as source, open(temporary, "w", encoding="utf-8") as target:
+        for line in source:
+            row = json.loads(line)
+            if isinstance(row, dict):
+                for field in ("request_json", "request_text", "response_text", "tool_calls"):
+                    row.pop(field, None)
+                row["content_opted_in"] = False
+            target.write(json.dumps(row, separators=(",", ":")) + "\n")
+    os.replace(temporary, path)
+
+
 def _already_sent(phase: str, pushed: int) -> str:
     """Say so when rows reached the server before the failure.
 

{metergraphrelay-0.7.2 → metergraphrelay-0.7.3}/src/metergraphrelay/providers/braintrust.py

@@ -9,8 +9,10 @@ import urllib.request
 from typing import Any, Callable
 
 from .. import __version__
+from ..http_limits import ResponseTooLarge, read_bounded
 from ..billing_evidence import reported_cost
 from ..capture_contract import capture_response_text, capture_row, capture_tool_calls
+from ..finish_reason import extract_stop_reason
 from ..import_identity import ImportContext, canonical_import_event_id
 from ..window import normalize_utc_designator
 
@@ -20,6 +22,8 @@ from ..window import normalize_utc_designator
 DEFAULT_BRAINTRUST_URL = "https://api.braintrust.dev"
 BTQL_PATH = "/btql"
 PAGE_LIMIT = 1000
+MAX_PAGES_PER_PULL = 10_000
+MAX_ITEMS_PER_PULL = 100_000
 # span_attributes.type is Braintrust's own span classifier ("llm", "score",
 # "function", "eval", "task", "tool", "review"). Only "llm" spans are model
 # calls; every other type is application/eval structure and is never imported.
@@ -154,7 +158,7 @@ def fetch_spans_page(
         with urllib.request.urlopen(
             request, timeout=REQUEST_TIMEOUT_SECONDS
         ) as response:
-            raw = response.read()
+            raw = read_bounded(response)
             cursor = _read_cursor(response.headers)
     except urllib.error.HTTPError as exc:
         raise BraintrustAPIError(
@@ -164,9 +168,11 @@ def fetch_spans_page(
         raise BraintrustAPIError(
             f"Braintrust API request failed: {exc.reason}"
         ) from exc
+    except ResponseTooLarge as exc:
+        raise BraintrustAPIError(str(exc)) from exc
     try:
         payload = json.loads(raw)
-    except json.JSONDecodeError as exc:
+    except (json.JSONDecodeError, RecursionError) as exc:
         raise BraintrustAPIError(
             f"Braintrust API returned invalid JSON: {exc}"
         ) from exc
@@ -484,6 +490,7 @@ def normalize_span(
     usage = map_metrics(metrics)
     request_json, request_text = _map_content(span.get("input"))
     response_text, tool_calls = _extract_output(span.get("output"))
+    stop_reason = extract_stop_reason(span)
 
     span_parents = span.get("span_parents")
     parent_span_id = (
@@ -530,6 +537,8 @@ def normalize_span(
         "span_id": span.get("span_id") or span["id"],
         "parent_span_id": parent_span_id,
     }
+    if stop_reason is not None:
+        row["stop_reason"] = stop_reason
     if import_context is not None:
         # The row id is Braintrust's stable identity for this span, so an
         # overlap re-pull deduplicates on the server instead of double counting.
@@ -563,6 +572,8 @@ def pull_braintrust(
     skipped = 0
     cursor: str | None = None
     used_cursors: set[str] = set()
+    page_count = 0
+    item_count = 0
     # The cursor is bound to the query that produced it, so every page must ask
     # for the same LIMIT — only the OFFSET clause changes between pages. The cap
     # on `count` is enforced while writing rows instead.
@@ -575,6 +586,11 @@ def pull_braintrust(
     try:
         with os.fdopen(fd, "w") as f:
             while imported < count:
+                page_count += 1
+                if page_count > MAX_PAGES_PER_PULL:
+                    raise BraintrustAPIError(
+                        f"Braintrust pull exceeded the maximum of {MAX_PAGES_PER_PULL} pages"
+                    )
                 if cursor is not None:
                     if cursor in used_cursors:
                         raise BraintrustAPIError(
@@ -597,6 +613,11 @@ def pull_braintrust(
                     on_progress()  # a page fetch is progress too
                 if not spans:
                     break
+                item_count += len(spans)
+                if item_count > MAX_ITEMS_PER_PULL:
+                    raise BraintrustAPIError(
+                        f"Braintrust pull exceeded the maximum of {MAX_ITEMS_PER_PULL} items"
+                    )
                 for span in spans:
                     if imported >= count:
                         break