feat(cron): Clean up LLM call rows

.env.example

@@ -110,3 +110,5 @@ SMTP_PASSWORD=
 EMAILS_FROM_EMAIL=
 EMAILS_FROM_NAME=Kaapi
 FRONTEND_HOST=
+
+DELETE_ROLLING_WINDOW_HOURS=

backend/app/api/routes/cron.py

@@ -8,7 +8,9 @@
 from app.api.permissions import Permission, require_permission
 from app.core.config import settings
 from app.crud.evaluations import process_all_pending_evaluations
+from app.models.llm.response import LlmCallRedactionResult
 from app.services.job_monitoring import monitor_pending_jobs
+from app.services.llm.retention import redact_aged_llm_calls
 from app.crud.stats import StatRow, get_daily_stats
 from app.services.stats import format_sections, post_to_discord
 
@@ -23,15 +25,10 @@
         "value": settings.CRON_INTERVAL_MINUTES,
         "unit": "minute",
     },
-    # Timezone for the schedule (only affects crontab-style schedules).
     "timezone": "UTC",
-    # Grace period (minutes) before a late check-in is marked as missed.
     "checkin_margin": 2,
-    # Max runtime (minutes) before an in-progress run is marked as timed out.
     "max_runtime": 2 * settings.CRON_INTERVAL_MINUTES,
-    # Consecutive failures/missed/timeouts required to open a Sentry issue.
     "failure_issue_threshold": 2,
-    # Consecutive successful check-ins required to auto-resolve the issue.
     "recovery_threshold": 1,
 }
 
@@ -59,6 +56,16 @@
 }
 
 
+LLM_CALL_RETENTION_CRON_MONITOR_CONFIG: MonitorConfig = {
+    "schedule": {"type": "crontab", "value": "0 9 * * *"},
+    "timezone": "Asia/Kolkata",
+    "checkin_margin": 5,
+    "max_runtime": 30,
+    "failure_issue_threshold": 1,
+    "recovery_threshold": 1,
+}
+
+
 @router.get(
     "/cron/evaluations",
     include_in_schema=False,
@@ -158,6 +165,35 @@ def daily_stats_cron_job(session: SessionDep) -> dict[str, list[StatRow]]:
         raise
 
 
+@router.get(
+    "/cron/llm-call-retention",
+    include_in_schema=False,
+    dependencies=[Depends(require_permission(Permission.SUPERUSER))],
+)
+@sentry_sdk.monitor(
+    monitor_slug="llm-call-retention-cron-job",
+    monitor_config=LLM_CALL_RETENTION_CRON_MONITOR_CONFIG,
+)
+def llm_call_retention_cron_job(session: SessionDep) -> LlmCallRedactionResult:
+    logger.info("[llm_call_retention_cron_job] Cron job invoked")
+
+    try:
+        result = redact_aged_llm_calls(session=session)
+        logger.info(
+            f"[llm_call_retention_cron_job] Completed: "
+            f"rows_redacted={result.rows_redacted}, "
+            f"cutoff={result.cutoff}"
+        )
+        return result
+    except Exception as e:
+        logger.error(
+            f"[llm_call_retention_cron_job] Error executing cron job: {e}",
+            exc_info=True,
+        )
+        sentry_sdk.capture_exception(e)
+        raise
+
+
 @router.get(
     "/cron/pending-jobs",
     include_in_schema=False,

backend/app/core/config.py

@@ -2,6 +2,7 @@
 import os
 import secrets
 import warnings
+from datetime import timedelta
 from typing import Any, Literal, Self
 
 from pydantic import (
@@ -196,10 +197,13 @@ def AWS_S3_BUCKET(self) -> str:
     EVAL_FAST_STALL_THRESHOLD_MINUTES: int = 15
     PENDING_JOB_QUERY_TIMEOUT_MS: int = 1000
 
-    # AI-assisted prompt improvement settings.
-    # See docs/srd-ai-prompt-improvement.md for the full design rationale.
-    # Platform-owned Anthropic key shared by every org/project for this feature,
-    # so prompt improvement works without per-project credentials.
+    DELETE_ROLLING_WINDOW_HOURS: int = 168
+
+    @computed_field  # type: ignore[prop-decorator]
+    @property
+    def DELETE_ROLLING_WINDOW_TIMEDELTA(self) -> timedelta:
+        return timedelta(hours=self.DELETE_ROLLING_WINDOW_HOURS)
+
     ANTHROPIC_API_KEY: str = ""
     PROMPT_IMPROVEMENT_MODEL: str = "claude-opus-4-8"
 

backend/app/crud/llm.py

@@ -1,9 +1,11 @@
 import logging
 import base64
 import json
+from datetime import datetime
 from uuid import UUID
 from typing import Any, Literal
 
+from sqlalchemy import text
 from sqlmodel import Session, select
 
 from app.core.util import now
@@ -335,6 +337,38 @@ def get_llm_call_by_id(
     return session.exec(statement).first()
 
 
+REDACTED_SENTINEL = "[redacted]"
+
+REDACT_LLM_CALLS_SQL = text(
+    """
+    UPDATE llm_call
+    SET input = :redacted_sentinel,
+        content = CASE
+            WHEN jsonb_typeof(content) = 'object'
+            THEN jsonb_set(content, '{content,value}', to_jsonb(CAST(:redacted_sentinel AS text)), false)
+            ELSE content
+        END
+    WHERE updated_at <= :cutoff
+      AND (input <> :redacted_sentinel OR content -> 'content' ->> 'value' IS DISTINCT FROM :redacted_sentinel)
+    """
+)
+
+
+def redact_llm_calls(*, session: Session, cutoff: datetime) -> int:
+    result = session.connection().execute(
+        REDACT_LLM_CALLS_SQL,
+        {"cutoff": cutoff, "redacted_sentinel": REDACTED_SENTINEL},
+    )
+    session.commit()
+
+    logger.info(
+        f"[redact_llm_calls] Redacted rows | rows: {result.rowcount} | "
+        f"cutoff: {cutoff.isoformat()}"
+    )
+
+    return result.rowcount
+
+
 def get_llm_calls_by_job_id(
     session: Session, job_id: UUID, project_id: int
 ) -> list[LlmCall]:

backend/app/models/llm/response.py

@@ -19,6 +19,13 @@ class Usage(SQLModel):
     reasoning_tokens: int | None = None
 
 
+class LlmCallRedactionResult(SQLModel):
+    """Outcome of a single llm_call retention redaction run."""
+
+    rows_redacted: int
+    cutoff: datetime
+
+
 class TextOutput(SQLModel):
     type: Literal["text"] = "text"
     content: TextContent

backend/app/services/llm/retention.py

@@ -0,0 +1,27 @@
+import logging
+
+from sqlmodel import Session
+
+from app.core.config import settings
+from app.core.util import now
+from app.crud.llm import redact_llm_calls
+from app.models.llm.response import LlmCallRedactionResult
+
+logger = logging.getLogger(__name__)
+
+
+def redact_aged_llm_calls(*, session: Session) -> LlmCallRedactionResult:
+    cutoff = now() - settings.DELETE_ROLLING_WINDOW_TIMEDELTA
+
+    logger.info(
+        f"[redact_aged_llm_calls] Starting redaction | cutoff: {cutoff.isoformat()}"
+    )
+
+    rows_redacted = redact_llm_calls(session=session, cutoff=cutoff)
+
+    logger.info(
+        f"[redact_aged_llm_calls] Completed | rows_redacted: {rows_redacted} | "
+        f"cutoff: {cutoff.isoformat()}"
+    )
+
+    return LlmCallRedactionResult(rows_redacted=rows_redacted, cutoff=cutoff)

backend/app/tests/api/routes/test_cron.py

@@ -1,10 +1,12 @@
+from datetime import datetime
 from unittest.mock import AsyncMock, MagicMock, patch
 
 import pytest
 from fastapi.testclient import TestClient
 
 from app.api.routes import cron
 from app.core.config import settings
+from app.models.llm.response import LlmCallRedactionResult
 from app.tests.utils.auth import TestAuthContext
 
 
@@ -326,6 +328,61 @@ def test_daily_stats_cron_job_captures_and_reraises_on_error() -> None:
     sentry.capture_exception.assert_called_once()
 
 
+def test_llm_call_retention_cron_job_success(
+    client: TestClient,
+    superuser_api_key: TestAuthContext,
+) -> None:
+    """Returns the redaction summary produced by the retention service."""
+    result = LlmCallRedactionResult(
+        rows_redacted=4137,
+        cutoff=datetime(2026, 2, 22, 12, 0, 0),
+    )
+    with patch(
+        "app.api.routes.cron.redact_aged_llm_calls",
+        return_value=result,
+    ) as redact:
+        response = client.get(
+            f"{settings.API_V1_STR}/cron/llm-call-retention",
+            headers={"X-API-KEY": superuser_api_key.key},
+        )
+
+    assert response.status_code == 200
+    assert response.json() == {
+        "rows_redacted": 4137,
+        "cutoff": "2026-02-22T12:00:00",
+    }
+    redact.assert_called_once()
+
+
+def test_llm_call_retention_cron_job_requires_superuser(
+    client: TestClient,
+    user_api_key: TestAuthContext,
+) -> None:
+    """Non-superuser cannot access the llm_call retention cron endpoint."""
+    response = client.get(
+        f"{settings.API_V1_STR}/cron/llm-call-retention",
+        headers={"X-API-KEY": user_api_key.key},
+    )
+
+    assert response.status_code == 403
+    assert "Insufficient permissions" in response.json()["error"]
+
+
+def test_llm_call_retention_cron_job_captures_and_reraises_on_error() -> None:
+    """On failure the job reports to Sentry and re-raises."""
+    with (
+        patch(
+            "app.api.routes.cron.redact_aged_llm_calls",
+            side_effect=RuntimeError("boom"),
+        ),
+        patch("app.api.routes.cron.sentry_sdk") as sentry,
+    ):
+        with pytest.raises(RuntimeError):
+            cron.llm_call_retention_cron_job(session=MagicMock())
+
+    sentry.capture_exception.assert_called_once()
+
+
 def test_evaluation_cron_job_not_in_schema(
     client: TestClient,
 ) -> None:
@@ -340,6 +397,7 @@ def test_evaluation_cron_job_not_in_schema(
     assert f"{settings.API_V1_STR}/cron/evaluations" not in paths
     assert f"{settings.API_V1_STR}/cron/pending-jobs" not in paths
     assert f"{settings.API_V1_STR}/cron/daily-stats" not in paths
+    assert f"{settings.API_V1_STR}/cron/llm-call-retention" not in paths
 
 
 def test_cron_intervals_match_to_prevent_sentry_monitor_drift() -> None: