🔒️(summary) redact meeting content from Sentry events

Sentry attaches stack frame locals to its events. When storing a
transcript in S3 failed, the full transcript held in `data` and
`transcript` was sent to Sentry.

Keep locals for debugging, but scrub variables and nested dict keys
known to hold transcripts, summaries, LLM prompts, participants'
personal data or pre-signed URLs. Share the Sentry init between the
API and the Celery worker, and never send request bodies.

CHANGELOG.md

@@ -21,6 +21,7 @@ and this project adheres to
 - 🔒️(agents) upgrade libpcre2-8-0 to fix CVE-2026-103111
 - 🔒️(frontend) upgrade pcre2 to fix CVE-2026-103111
 - 🐛(summary) disable default S3 checksums for GCS-compatible storage
+- 🔒️(summary) redact meeting content from Sentry events
 
 ## [1.33.0] - 2026-09-30
 

src/summary/summary/core/celery_worker.py

@@ -9,7 +9,6 @@ from typing import Any
 from urllib.parse import urljoin
 
 import requests
-import sentry_sdk
 from celery import Celery, signals
 from celery.utils.log import get_task_logger
 from openai.types.audio import Transcription
@@ -42,6 +41,7 @@ from summary.core.prompt import (
     PROMPT_SYSTEM_TLDR,
     PROMPT_USER_PART,
 )
+from summary.core.sentry import init_sentry
 from summary.core.shared_models import (
     SummarizeWebhookFailurePayload,
     SummarizeWebhookSuccessPayload,
@@ -75,12 +75,11 @@ celery = Celery(
 
 celery.config_from_object("summary.core.celery_config")
 
-if settings.sentry_dsn and settings.sentry_is_enabled:
 
-    @signals.celeryd_init.connect
-    def init_sentry(**_kwargs):
-        """Initialize sentry."""
-        sentry_sdk.init(dsn=settings.sentry_dsn, enable_tracing=True)
+@signals.celeryd_init.connect
+def init_celery_sentry(**_kwargs):
+    """Initialize Sentry in the Celery worker."""
+    init_sentry()
 
 
 file_service = FileService()

src/summary/summary/core/sentry.py

@@ -0,0 +1,94 @@
+"""Sentry configuration."""
+
+import sentry_sdk
+from sentry_sdk.scrubber import DEFAULT_DENYLIST, DEFAULT_PII_DENYLIST, EventScrubber
+
+from summary.core.config import get_settings
+
+# Exact names (case-insensitive) of variables and dict keys to redact.
+SENSITIVE_DATA_DENYLIST = [
+    # Raw payloads and serialized bodies
+    "data",
+    "body",
+    "payload",
+    "args",
+    "kwargs",
+    "response",
+    "res",
+    # Transcripts
+    "transcript",
+    "transcription",
+    "transcription_json",
+    "transcription_res",
+    "new_transcription",
+    "segments",
+    "word_segments",
+    "words",
+    "text",
+    "content",
+    "formatted_output",
+    # Summaries and LLM exchanges
+    "summary",
+    "raw_summary",
+    "cleaned_summary",
+    "tldr",
+    "part",
+    "parts",
+    "parts_summarized",
+    "next_steps",
+    "title",
+    "titles",
+    "action",
+    "line",
+    "lines",
+    "user_prompt",
+    "prompt_user_part",
+    "messages",
+    "json_data",  # OpenAI client internals
+    "opts",
+    "options",
+    "input_options",
+    # Participants' personal data
+    "email",
+    "user_email",
+    "assignees",
+    "participant_name",
+    "participant_names",
+    "participants_info",
+    "speaker_to_name",
+    # Signed / pre-authenticated URLs
+    "cloud_storage_url",
+    "transcription_data_url",
+    "summary_data_url",
+]
+
+
+def build_event_scrubber() -> EventScrubber:
+    """Build the scrubber redacting meeting content and personal data."""
+    return EventScrubber(
+        denylist=DEFAULT_DENYLIST + SENSITIVE_DATA_DENYLIST,
+        pii_denylist=DEFAULT_PII_DENYLIST,
+        recursive=True,
+    )
+
+
+def init_sentry() -> None:
+    """Initialize Sentry if enabled in the settings."""
+    settings = get_settings()
+
+    if not settings.sentry_is_enabled:
+        return
+
+    if not settings.sentry_dsn:
+        return
+
+    sentry_sdk.init(
+        dsn=settings.sentry_dsn,
+        enable_tracing=True,
+        # Never attach request bodies, Celery task arguments or user data.
+        send_default_pii=False,
+        # Task creation requests carry the content to summarize.
+        max_request_body_size="never",
+        include_local_variables=True,
+        event_scrubber=build_event_scrubber(),
+    )

src/summary/summary/main.py

@@ -1,18 +1,17 @@
 """Application."""
 
-import sentry_sdk
 from dockerflow.fastapi import router as dockerflow_router
 from fastapi import FastAPI
 
 from summary.api.main import api_router_v2
 from summary.core import checks  # noqa: F401 -- registers the Dockerflow checks
 from summary.core.config import get_settings
+from summary.core.sentry import init_sentry
 
 settings = get_settings()
 
 
-if settings.sentry_dsn and settings.sentry_is_enabled:
-    sentry_sdk.init(dsn=settings.sentry_dsn, enable_tracing=True)
+init_sentry()
 
 app = FastAPI(
     title=settings.app_name,

src/summary/tests/unit/test_sentry.py

@@ -0,0 +1,207 @@
+"""Tests for the Sentry configuration.
+
+Each test raises an error from code handling meeting content and inspects the
+event Sentry would send: the content must be redacted, while harmless local
+variables are kept for debugging.
+"""
+
+import json
+from collections.abc import Callable, Iterator
+from unittest.mock import Mock
+
+import httpx
+import openai
+import pytest
+import sentry_sdk
+from botocore.exceptions import ClientError
+from botocore.stub import Stubber
+from sentry_sdk.transport import Transport
+
+from summary.core import file_service
+from summary.core import sentry as sentry_module
+from summary.core.file_service import FileService
+from summary.core.llm_service import LLMException, LLMService
+from summary.core.shared_models import WhisperXResponse
+
+CANARY = "CANARY-MEETING-CONTENT"
+
+SentryEvents = Callable[[], list[str]]
+
+
+class _CapturingTransport(Transport):
+    """Keep serialized events in memory instead of sending them."""
+
+    def __init__(self):
+        super().__init__()
+        self.events: list[str] = []
+
+    def capture_envelope(self, envelope):
+        """Store each serialized event of the envelope."""
+        for item in envelope.items:
+            if item.type == "event":
+                self.events.append(item.payload.get_bytes().decode())
+
+
+@pytest.fixture
+def sentry_events() -> Iterator[SentryEvents]:
+    """Initialize Sentry as in production, with an in-memory transport."""
+    transport = _CapturingTransport()
+    sentry_sdk.init(
+        dsn="https://public@sentry.example.com/1",
+        transport=transport,
+        send_default_pii=False,
+        include_local_variables=True,
+        event_scrubber=sentry_module.build_event_scrubber(),
+        default_integrations=False,
+    )
+
+    def flush() -> list[str]:
+        sentry_sdk.flush()
+        return transport.events
+
+    yield flush
+    sentry_sdk.init()  # Disable Sentry for the following tests
+
+
+def _transcript() -> WhisperXResponse:
+    return WhisperXResponse.model_validate(
+        {
+            "segments": [
+                {
+                    "start": 0.0,
+                    "end": 1.0,
+                    "text": f"I don't know {CANARY}",
+                    "speaker": "SPEAKER_01",
+                    "words": [
+                        {
+                            "word": CANARY,
+                            "start": 0.0,
+                            "end": 1.0,
+                            "score": 0.9,
+                            "speaker": "SPEAKER_01",
+                        }
+                    ],
+                }
+            ]
+        }
+    )
+
+
+def _local_vars(event: str) -> list[dict]:
+    """Return the local variables of every frame of the event."""
+    return [
+        frame.get("vars", {})
+        for exception in json.loads(event)["exception"]["values"]
+        for frame in exception["stacktrace"]["frames"]
+    ]
+
+
+@pytest.fixture
+def s3_stubber(monkeypatch: pytest.MonkeyPatch) -> Iterator[Stubber]:
+    """Stub the S3 client built by the file service."""
+    monkeypatch.setattr(
+        file_service,
+        "settings",
+        file_service.settings.model_copy(
+            update={
+                "aws_s3_endpoint_url": "garage:9000",
+                "aws_s3_secure_access": False,
+                "aws_s3_region_name": "fr-par",
+                "aws_storage_bucket_name": "meet-media-storage",
+            }
+        ),
+    )
+    stubber = Stubber(file_service._build_s3_client())
+    stubber.activate()
+    monkeypatch.setattr(file_service, "_build_s3_client", lambda: stubber.client)
+    yield stubber
+    stubber.assert_no_pending_responses()
+
+
+def test_sentry_store_transcript_failure_redacts_transcript(
+    sentry_events: SentryEvents, s3_stubber: Stubber
+) -> None:
+    """A failed S3 upload does not send the transcript, but keeps the job id."""
+    s3_stubber.add_client_error("put_object", service_error_code="InvalidDigest")
+
+    try:
+        FileService().store_transcript(transcript=_transcript(), job_id="job-1")
+    except ClientError:
+        sentry_sdk.capture_exception()
+    else:
+        pytest.fail("store_transcript should have failed")
+
+    [event] = sentry_events()
+    assert CANARY not in event
+
+    store_transcript_vars = next(
+        frame_vars
+        for frame_vars in _local_vars(event)
+        if "transcript_path" in frame_vars
+    )
+    assert store_transcript_vars["job_id"] == "'job-1'"
+    assert store_transcript_vars["transcript_path"] == "'transcripts/job-1.json'"
+    assert store_transcript_vars["data"] == "[Filtered]"
+    assert store_transcript_vars["transcript"] == "[Filtered]"
+
+
+def test_sentry_llm_failure_redacts_prompts(sentry_events: SentryEvents) -> None:
+    """A failed LLM call does not send the prompts, even from OpenAI internals."""
+    client = openai.OpenAI(
+        api_key="test-key",
+        base_url="https://llm.example.com/v1",
+        max_retries=0,
+        http_client=httpx.Client(
+            transport=httpx.MockTransport(lambda request: httpx.Response(500))
+        ),
+    )
+    observability = Mock(is_enabled=False)
+    observability.get_openai_client.return_value = client
+
+    try:
+        LLMService(observability).call(
+            system_prompt="Summarize this meeting.",
+            user_prompt=f"Transcript: {CANARY}",
+            name="tldr",
+        )
+    except LLMException:
+        sentry_sdk.capture_exception()
+    else:
+        pytest.fail("the LLM call should have failed")
+
+    [event] = sentry_events()
+    assert CANARY not in event
+
+
+def test_init_sentry_uses_the_event_scrubber(monkeypatch: pytest.MonkeyPatch) -> None:
+    """Sentry is initialized with local variables and the content scrubber."""
+    settings = sentry_module.get_settings().model_copy(
+        update={"sentry_is_enabled": True, "sentry_dsn": "https://k@example.com/1"}
+    )
+    monkeypatch.setattr(sentry_module, "get_settings", lambda: settings)
+    init = Mock()
+    monkeypatch.setattr(sentry_module.sentry_sdk, "init", init)
+
+    sentry_module.init_sentry()
+
+    init.assert_called_once()
+    kwargs = init.call_args.kwargs
+    assert kwargs["send_default_pii"] is False
+    assert kwargs["max_request_body_size"] == "never"
+    assert kwargs["include_local_variables"] is True
+    denylist = kwargs["event_scrubber"].denylist
+    assert {"data", "transcript", "content", "summary", "password"} <= set(denylist)
+
+
+def test_init_sentry_disabled(monkeypatch: pytest.MonkeyPatch) -> None:
+    """Sentry is not initialized when disabled."""
+    settings = sentry_module.get_settings().model_copy(
+        update={"sentry_is_enabled": False, "sentry_dsn": "https://k@example.com/1"}
+    )
+    monkeypatch.setattr(sentry_module, "get_settings", lambda: settings)
+    init = Mock()
+    monkeypatch.setattr(sentry_module.sentry_sdk, "init", init)
+
+    sentry_module.init_sentry()
+
+    init.assert_not_called()