Harden agent file exports and owner reports

api/agent/core/prompt_context.py

@@ -658,7 +658,7 @@ def _get_sqlite_guidance() -> str:
         "## SQLite Data\n\n"
         "Fetch new data with its source tool and answer small results directly. Use sqlite_batch for data already in "
         "SQLite when large/truncated or needing filtering, joins, aggregation, charts, reuse, or domain logic. Model "
-        "reusable domains once as shared entity, event, and relationship tables. Separate repeating parents from children "
+        "reusable domains as keyed entities, events, and relations; query unmatched inventory rows before reporting. Normalize parent/child data "
         "(vendors/plans, accounts/events) with PRIMARY KEY/UNIQUE identity, useful indexes, and source provenance; put "
         "logic in SQL and return only needed rows to context. Populate them from all relevant __tool_results rows with one "
         "shaped INSERT ... SELECT/json_each query and an IN/tool_name filter; never filter result_id one at a time; never "
@@ -3205,7 +3205,7 @@ def _get_web_chat_formatting_guidance() -> str:
     return (
         "Web chat and peer DM formatting:\n"
         "Start with the answer/main finding. Keep simple exchanges/outreach natural. "
-        "Reports to owners/creators and multi-part findings, metrics, or recommendations use polished sections, bullets/tables or metric blocks, status labels, and tasteful visual cues, even when not called a report. "
+        "Owner/creator reports and multi-part findings need polished hierarchy: sections, bullets, compact linked tables, metric/status blocks, or tasteful cues, even without a report label. "
         "Address known recipients naturally once around actions; avoid generic delivery logs and agent-name self-intros unless asked. "
         "Use whitespace, not separators. Charts: paste create_chart result.inline; don't attach/read/rebuild."
     )
@@ -3767,7 +3767,7 @@ def _get_system_instruction(
         "Do not invent work, results, preferences, or personal experiences.\n\n"
 
         "## Output Rules\n\n"
-        "Use the lightest clear structure: fact, list, table, or sectioned report. Ground facts, numbers, units, and URLs in tool results; do not relabel or convert units unless asked. For record lists, include item/detail URLs from results, not only feed/source URLs. Present requested returned data directly; omit unrelated fields and unavailable extras, summarize overflow, and do not add follow-up offers after simple facts, prices, statuses, or quick lookups. "
+        "Use proportionate structure. A bounded named set is a coverage contract: include each item or mark it unavailable/excluded; summarize only real overflow. Ground facts, numbers, units, and URLs in tool results; never relabel/convert units unless asked. Record lists keep item/detail URLs, not just feed/source URLs. Present requested data directly; omit unrelated/unavailable fields and follow-up offers after simple facts, prices, statuses, or lookups. "
         "Charts: create only when requested/materially useful. "
         "Paste create_chart result.inline/result.inline_html in the message; do not attach/read charts or invent paths, hashes, image tags, or <img> URLs. "
         "Use create_csv for tabular exports, create_pdf for PDFs, and create_file for other text/doc formats; create_file query mode must return exactly one row and one column.\n\n"

api/agent/files/filespace_service.py

@@ -7,10 +7,12 @@
 from dataclasses import dataclass
 from typing import Any, List
 
+from botocore.exceptions import BotoCoreError, ClientError
 from celery.utils.log import get_task_logger
 from django.core.exceptions import SuspiciousFileOperation
 from django.core.files.base import ContentFile
-from django.db import IntegrityError, transaction
+from django.db import DatabaseError, IntegrityError, transaction
+from google.api_core.exceptions import GoogleAPIError
 
 from util.analytics import Analytics, AnalyticsEvent, AnalyticsSource
 
@@ -21,6 +23,8 @@
 EXPORTS_DIR_NAME = "exports"
 DOWNLOADS_DIR_NAME = "downloads"
 UNSAFE_FILESPACE_PART_CHARS_RE = re.compile(r"(?u)[^-\w.+@]")
+FILESPACE_STORAGE_UNAVAILABLE = "Filespace storage unavailable. Retry this file tool once; do not use another file tool, Python, or shell. After another failure, report the blocker or return a small result inline."
+FILESPACE_PERSISTENCE_ERRORS = (BotoCoreError, ClientError, DatabaseError, GoogleAPIError, OSError, SuspiciousFileOperation)
 
 
 @dataclass
@@ -171,21 +175,25 @@ def _save_node_content(
     *,
     delete_node_on_failure: bool,
 ) -> dict[str, Any] | None:
+    previous_content_name = getattr(node.content, "name", None)
     try:
         node.content.save(node.name, ContentFile(content_bytes), save=False)
         node.save()
         node.refresh_from_db()
         return None
-    except Exception:
+    except FILESPACE_PERSISTENCE_ERRORS:
         logger.exception("Failed to persist file to %s for agent %s", dir_name, agent_id)
         try:
-            if node.content and getattr(node.content, "name", None):
+            if node.content and getattr(node.content, "name", None) != previous_content_name:
                 node.content.delete(save=False)
-        except Exception:
+        except FILESPACE_PERSISTENCE_ERRORS:
             logger.exception("Failed to clean up file content for node %s", node.id)
         if delete_node_on_failure:
-            node.delete()
-        return {"status": "error", "message": "Failed to save the file in the filespace."}
+            try:
+                node.delete()
+            except DatabaseError:
+                logger.exception("Failed to clean up file node %s", node.id)
+        return {"status": "error", "code": "filespace_storage_unavailable", "message": FILESPACE_STORAGE_UNAVAILABLE, "retryable": True}
 
 
 def _agent_has_access(agent: "PersistentAgent", filespace_id: "uuid.UUID") -> bool:

api/agent/tools/create_csv.py

@@ -6,8 +6,6 @@
 from api.agent.tools.file_export_helpers import resolve_export_target, write_agent_export
 from .sqlite_query_runner import run_sqlite_select
 
-EXTENSION = ".csv"
-MIME_TYPE = "text/csv"
 MAX_EXPORT_ROWS = 5000
 
 
@@ -18,7 +16,7 @@ def get_create_csv_tool() -> Dict[str, Any]:
             "name": "create_csv",
             "description": (
                 "Create a CSV file and store it in the agent filespace. "
-                "Provide either raw CSV text or a SQLite SELECT query to export query results. "
+                "Provide exactly one content source: raw CSV text, or a SQLite SELECT query for data already in SQLite. "
                 "Recommended path: /exports/your-file.csv. Returns `file`, `inline`, `inline_html`, and `attach`."
             ),
             "parameters": {
@@ -58,10 +56,8 @@ def execute_create_csv(agent: PersistentAgent, params: Dict[str, Any]) -> Dict[s
     csv_text = params.get("csv_text")
     query = params.get("query")
 
-    if not csv_text and not query:
-        return {"status": "error", "message": "Provide either csv_text or query."}
-    if csv_text and query:
-        return {"status": "error", "message": "Use csv_text OR query, not both."}
+    if bool(csv_text) == bool(query):
+        return {"status": "error", "message": "Provide exactly one of csv_text or query."}
 
     path, overwrite, error = resolve_export_target(params, agent_id=agent.id)
     if error:
@@ -92,8 +88,8 @@ def execute_create_csv(agent: PersistentAgent, params: Dict[str, Any]) -> Dict[s
     return write_agent_export(
         agent=agent,
         content_bytes=csv_text_to_write.encode("utf-8"),
-        extension=EXTENSION,
-        mime_type=MIME_TYPE,
+        extension=".csv",
+        mime_type="text/csv",
         path=path,
         overwrite=overwrite,
         size_label="CSV",

api/agent/tools/tests/test_attachment_guidance.py

@@ -80,10 +80,10 @@ def test_report_message_guidance_names_visual_quality_without_eval_prompting(sel
         self.assertIn("styled tables or metric blocks", email_tool["function"]["parameters"]["properties"]["mobile_first_html"]["description"])
         self.assertIn("false when this email is the requested final delivery", email_tool["function"]["parameters"]["properties"]["will_continue_work"]["description"])
         self.assertIn("Do not use this to simulate or confirm an email/SMS delivery", chat_tool["function"]["description"])
-        self.assertIn("status labels", chat_guidance)
-        self.assertIn("owners/creators", chat_guidance)
-        self.assertIn("even when not called a report", chat_guidance)
-        self.assertIn("tables or metric blocks", chat_guidance)
+        self.assertIn("metric/status blocks", chat_guidance)
+        self.assertIn("Owner/creator reports", chat_guidance)
+        self.assertIn("even without a report label", chat_guidance)
+        self.assertIn("compact linked tables", chat_guidance)
         self.assertIn("Address known recipients naturally once", chat_guidance)
         self.assertIn("agent-name self-intros", chat_guidance)
         self.assertIn("owners/creators", chat_tool["function"]["parameters"]["properties"]["body"]["description"])

api/evals/scenarios/behavior_micro.py

@@ -1,6 +1,8 @@
+import csv
+import io
+import json
 from copy import deepcopy
 from dataclasses import dataclass, field
-import json
 
 from api.agent.comms.human_input_requests import MAX_OPTION_COUNT, dismiss_human_input_request
 from api.agent.core.processing_flags import get_human_inbound_generation
@@ -18,6 +20,7 @@
     sqlite_batch_mutates_planning_state,
 )
 from api.models import (
+    AgentFsNode,
     EvalRunTask,
     CommsAllowlistEntry,
     CommsChannel,
@@ -2228,7 +2231,7 @@ def run(self, run_id, agent_id):
 @register_scenario
 class ToolChoiceCsvDeliverableUsesCreateCsvScenario(BehaviorMicroScenario):
     slug = TOOL_CHOICE_CSV_DELIVERABLE_USES_CREATE_CSV
-    description = "A downloadable CSV request should use create_csv."
+    description = "A downloadable CSV request should execute create_csv and persist the requested rows."
     category = "files"
     tags = ("agent_behavior", "micro", "tool_choice", "files")
     tasks = [
@@ -2241,15 +2244,6 @@ def run(self, run_id, agent_id):
         self._seed_prior_processing_run(agent_id)
         self._enable_builtin_tools(agent_id, ["create_csv"])
 
-        mock_config = {
-            "create_csv": {
-                "status": "ok",
-                "file": {"path": "/exports/q1-leads.csv"},
-                "message": "CSV created.",
-            },
-            "create_file": {"status": "error", "message": "Use create_csv for CSV deliverables."},
-        }
-
         self.record_task_result(run_id, None, EvalRunTask.Status.RUNNING, task_name="inject_prompt")
         with self.wait_for_agent_idle(agent_id, timeout=120):
             inbound = self.inject_message(
@@ -2260,10 +2254,12 @@ def run(self, run_id, agent_id):
                 ),
                 trigger_processing=True,
                 eval_run_id=run_id,
-                mock_config=mock_config,
                 eval_stop_policy={
-                    "stop_on_tool_names": ["create_file"],
-                    "stop_when_all_seen": [{"tool_name": "create_csv"}],
+                    "stop_on_tool_names_after_execution": ["create_csv"],
+                    "stop_on_unexpected_relevant_tool": True,
+                    "allowed_tool_names": ["create_csv"],
+                    "ignored_tool_names": ["update_plan", "send_chat_message", "sleep_until_next_trigger"],
+                    "max_relevant_tool_calls": 3,
                 },
             )
         self.record_task_result(
@@ -2277,22 +2273,64 @@ def run(self, run_id, agent_id):
 
         self.record_task_result(run_id, None, EvalRunTask.Status.RUNNING, task_name="verify_create_csv")
         create_csv_calls = get_tool_calls_for_run(run_id, after=inbound.timestamp, tool_names={"create_csv"})
-        if create_csv_calls:
+        successful_calls = []
+        for call in create_csv_calls:
+            result = call.result
+            if isinstance(result, str):
+                try:
+                    result = json.loads(result)
+                except json.JSONDecodeError:
+                    result = {}
+            if (
+                str(call.status or "").lower() == "complete"
+                and isinstance(result, dict)
+                and result.get("status") == "ok"
+                and result.get("file") == "$[/exports/q1-leads.csv]"
+            ):
+                successful_calls.append(call)
+
+        node = AgentFsNode.objects.filter(
+            created_by_agent_id=agent_id,
+            path="/exports/q1-leads.csv",
+        ).alive().first()
+        rows = []
+        read_error = None
+        if node is None or not node.content.name:
+            read_error = "missing content"
+        else:
+            try:
+                with node.content.open("rb") as handle:
+                    rows = list(csv.reader(io.StringIO(handle.read().decode("utf-8"))))
+            except (OSError, ValueError, UnicodeDecodeError, csv.Error) as exc:
+                read_error = exc.__class__.__name__
+
+        expected_rows = [
+            ["company", "priority"],
+            ["Acme", "high"],
+            ["Globex", "medium"],
+            ["Initech", "low"],
+        ]
+        if len(successful_calls) == 1 and len(create_csv_calls) == 1 and read_error is None and rows == expected_rows:
             self.record_task_result(
                 run_id,
                 None,
                 EvalRunTask.Status.PASSED,
                 task_name="verify_create_csv",
-                observed_summary="Agent used create_csv for the CSV deliverable.",
-                artifacts={"step": create_csv_calls[0].step},
+                observed_summary="Agent executed create_csv once and persisted the requested header and three rows.",
+                artifacts={"step": successful_calls[0].step},
             )
         else:
             self.record_task_result(
                 run_id,
                 None,
                 EvalRunTask.Status.FAILED,
                 task_name="verify_create_csv",
-                observed_summary="Agent did not use create_csv for the CSV deliverable.",
+                observed_summary=(
+                    f"Expected one successful create_csv and exact persisted rows; saw "
+                    f"{len(create_csv_calls)} call(s), {len(successful_calls)} success(es), "
+                    f"read_error={read_error}, rows={rows}."
+                ),
+                artifacts={"step": create_csv_calls[0].step} if create_csv_calls else {},
             )
 
 

api/evals/scenarios/effort_calibration.py

@@ -2473,6 +2473,12 @@ class EffortSimpleCurrentCompanyReportScenario(EffortCalibrationScenario):
         ScenarioTask(name="verify_no_config_churn", assertion_type="manual"),
         ScenarioTask(name="verify_turn_budget", assertion_type="manual"),
     ]
+    required_concept_groups = (
+        ("Northstar Robotics", "Northstar"),
+        ("Atlas", "mixed-fleet", "mixed fleet"),
+        ("Series B", "$42M", "42M", "42 million"),
+        ("18 percent", "18%", "eighteen percent"),
+    )
 
     def run(self, run_id: str, agent_id: str) -> None:
         self._ready_agent(agent_id)
@@ -2682,7 +2688,7 @@ def run(self, run_id: str, agent_id: str) -> None:
             min_source_count=2,
             min_chars=650,
             max_chars=3000,
-            required_any_groups=(("Northstar Robotics", "Northstar"),),
+            required_any_groups=self.required_concept_groups,
         )
         self._record_no_question_battery(
             run_id,