Merge pull request 'Release v26.04.13.2' (#30) from dev into main

Non-streaming generate_completion was the only LLM entry point that
didn't default num_ctx — stream_chat and stream_chat_with_tools both
fall back to Config.OLLAMA_NUM_CTX (16384). When a caller omitted the
argument, Ollama silently used the model's default window (~4k on
qwen3) and truncated the prompt.

That footgun was masked by fallback paths in the research pipeline:
_generate_outline's prompt carries ~12 sources × 2000 chars (~6k
tokens) of source material plus a system prompt, so the prompt got
chopped, the model never saw the sources, JSON parsing failed twice,
and run_research_pipeline dropped into the single-note "monolith"
fallback (research.py:251). The user reported chat 215 producing such
a monolith note for a multi-source research topic.

Two-layer fix:
- Default num_ctx to Config.OLLAMA_NUM_CTX inside generate_completion,
  matching the streaming entry points. Any current or future caller
  that forgets the argument stops silently losing input.
- Pin num_ctx=16384 explicitly in _generate_outline and
  _generate_executive_summary with comments pointing at the failure
  mode, so a refactor of the generate_completion default can't
  silently regress the research pipeline.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

src/fabledassistant/routes/briefing.py

@@ -330,10 +330,13 @@ async def reset_today_briefing():
     driven from the MCP when iterating on prompts. Pair with
     ``POST /api/briefing/trigger`` to immediately regenerate.
     """
-    from datetime import date as _date
     from sqlalchemy import delete as _delete
 
-    today = _date.today()
+    from fabledassistant.services.tz import user_briefing_date
+
+    # User-local briefing day (flips at 4am local), not ``date.today()`` —
+    # see services/tz.py for rationale.
+    today = await user_briefing_date(g.user.id)
     async with async_session() as session:
         result = await session.execute(
             select(Conversation).where(

src/fabledassistant/services/briefing_conversations.py

@@ -1,12 +1,13 @@
 """Create and manage briefing conversations."""
 
 import logging
-from datetime import date, datetime, timezone
+from datetime import datetime, timezone
 
 from sqlalchemy import select
 
 from fabledassistant.models import async_session
 from fabledassistant.models.conversation import Conversation, Message
+from fabledassistant.services.tz import user_briefing_date
 
 logger = logging.getLogger(__name__)
 
@@ -15,7 +16,10 @@ async def get_or_create_today_conversation(user_id: int, model: str) -> Conversa
     """
     Return today's briefing conversation, creating it if it doesn't exist.
     """
-    today = date.today()
+    # "Today" is the user-local briefing day (flips at 4am local), not
+    # ``date.today()`` — in a UTC container the latter rolls over at
+    # 19:00 NY local and makes the in-progress briefing disappear.
+    today = await user_briefing_date(user_id)
     async with async_session() as session:
         result = await session.execute(
             select(Conversation).where(

src/fabledassistant/services/briefing_scheduler.py

@@ -445,9 +445,12 @@ async def _run_profile_closeout(user_id: int, model: str) -> None:
     """
     from fabledassistant.services.user_profile import append_observations
     from fabledassistant.services.llm import generate_completion
+    from fabledassistant.services.tz import user_today
     from fabledassistant.models.conversation import Conversation, Message
 
-    yesterday = (date.today() - timedelta(days=1))
+    # User-local "yesterday" so closeout always targets the day that just
+    # ended in the user's timezone, regardless of container TZ.
+    yesterday = (await user_today(user_id)) - timedelta(days=1)
     async with async_session() as session:
         result = await session.execute(
             select(Conversation).where(

src/fabledassistant/services/generation_task.py

@@ -527,9 +527,22 @@ async def run_generation(
         should_gen_title = not conv_title or (msg_count > 0 and msg_count % 10 == 0)
 
         if should_gen_title:
-            title_messages = messages + [
+            # Feed the title model the *raw* conversation turns only — never
-                {"role": "assistant", "content": buf.content_so_far}
+            # the post-build_context ``messages`` list. ``build_context``
+            # prepends RAG snippets, RSS excerpts, URL content, and briefing
+            # article dumps INTO the user message string itself, so filtering
+            # by role="user" downstream still surfaces that noise as the
+            # "user's message". That pollution caused wildly-wrong titles
+            # (bug #109) — the small background model was staring at article
+            # excerpts instead of what the user actually typed. Pass the
+            # original history + the raw user_content + the assistant reply.
+            title_messages: list[dict] = [
+                {"role": m["role"], "content": m.get("content") or ""}
+                for m in history
+                if m.get("role") in ("user", "assistant")
             ]
+            title_messages.append({"role": "user", "content": user_content})
+            title_messages.append({"role": "assistant", "content": buf.content_so_far})
 
             async def _bg_title() -> None:
                 try:

src/fabledassistant/services/llm.py

@@ -317,9 +317,17 @@ async def generate_completion(
     num_ctx overrides the model's context window for this call only.
     """
     last_exc: Exception | None = None
-    options: dict = {"num_predict": max_tokens}
+    # Default num_ctx to Config.OLLAMA_NUM_CTX (matching stream_chat /
-    if num_ctx is not None:
+    # stream_chat_with_tools). Without this, Ollama silently uses the model's
-        options["num_ctx"] = num_ctx
+    # default window (~4k on qwen3) and truncates anything longer. That is
+    # how the research pipeline's outline step kept falling back to a single
+    # monolith note: its 12-source prompt is ~6k tokens and was being chopped
+    # before the model ever saw it. Non-streaming callers must not inherit
+    # that footgun — if you truly want the model default, pass num_ctx=0.
+    options: dict = {
+        "num_predict": max_tokens,
+        "num_ctx": num_ctx if num_ctx is not None else Config.OLLAMA_NUM_CTX,
+    }
     for attempt in range(3):
         if attempt > 0:
             delay = 3.0 * attempt

src/fabledassistant/services/research.py

@@ -63,7 +63,17 @@ async def _generate_outline(topic: str, sources: list[dict], model: str) -> list
     ]
     for attempt in range(2):
         try:
-            raw = await generate_completion(messages, model, max_tokens=400)
+            # Pin num_ctx explicitly. The prompt carries up to 12 sources at
+            # 2000 chars each (~6k tokens of source material alone) plus the
+            # system prompt — well over Ollama's default model window on
+            # qwen3. Without this, Ollama silently truncates the prompt, the
+            # model can't see most of the sources, JSON parsing fails twice,
+            # and the pipeline falls back to a single monolith note
+            # (`research.py:251`). Do not remove even if `generate_completion`
+            # appears to default this — see the comment there.
+            raw = await generate_completion(
+                messages, model, max_tokens=400, num_ctx=16384
+            )
             raw = raw.strip()
             raw = re.sub(r"^```(?:json)?\s*", "", raw)
             raw = re.sub(r"\s*```$", "", raw)
@@ -161,7 +171,13 @@ async def _generate_executive_summary(
         },
     ]
     try:
-        raw = await generate_completion(messages, model, max_tokens=600)
+        # Pin num_ctx explicitly — see `_generate_outline` comment for the
+        # rationale. This prompt carries N sections × 1500 chars of section
+        # prose, which can easily exceed the default model window. Don't
+        # trust the `generate_completion` default to stick.
+        raw = await generate_completion(
+            messages, model, max_tokens=600, num_ctx=16384
+        )
         return raw.strip()
     except Exception:
         logger.warning("Executive summary generation failed for '%s'", topic, exc_info=True)

src/fabledassistant/services/tools/calendar.py

@@ -14,15 +14,37 @@ from fabledassistant.services.events import (
 )
 from fabledassistant.services.tools._helpers import resolve_project
 from fabledassistant.services.tools._registry import tool
+from fabledassistant.services.tz import get_user_tz
+
+
+async def _parse_datetime_in_user_tz(
+    user_id: int, value: str
+) -> tuple[datetime, bool]:
+    """Parse a date/datetime string from the model into a UTC-aware datetime.
+
+    Naive inputs are interpreted in the **user's local timezone** and then
+    converted to UTC for storage. Never default to UTC for naive inputs —
+    that's how all-day events landed on the wrong day for non-UTC users.
+
+    Returns ``(utc_datetime, was_date_only)``.
+    """
+    was_date_only = "T" not in value and " " not in value
+    if was_date_only:
+        value = f"{value}T00:00:00"
+    dt = datetime.fromisoformat(value.replace("Z", "+00:00"))
+    if dt.tzinfo is None:
+        user_tz = await get_user_tz(user_id)
+        dt = dt.replace(tzinfo=user_tz)
+    return dt.astimezone(timezone.utc), was_date_only
 
 
 @tool(
     name="create_event",
-    description="Create a calendar event for the user. Use this when the user asks to schedule, add, or create a meeting, appointment, or event.",
+    description="Create a calendar event for the user. Use this when the user asks to schedule, add, or create a meeting, appointment, or event. Pass dates and datetimes in the user's local time — a bare date like '2026-09-30' is interpreted as that day in the user's configured timezone.",
     parameters={
         "title": {"type": "string", "description": "A descriptive event title"},
-        "start": {"type": "string", "description": "Start date (YYYY-MM-DD) or datetime (ISO 8601 with timezone offset)"},
+        "start": {"type": "string", "description": "Start date (YYYY-MM-DD) or datetime (YYYY-MM-DDTHH:MM) in the user's local time. Include a timezone offset only if the user explicitly mentions one."},
-        "end": {"type": "string", "description": "Optional end date or datetime"},
+        "end": {"type": "string", "description": "Optional end date or datetime in the user's local time"},
         "duration": {"type": "integer", "description": "Optional duration in minutes (default 60, ignored if end is set or all_day is true)"},
         "description": {"type": "string", "description": "Optional event description"},
         "location": {"type": "string", "description": "Optional event location"},
@@ -40,23 +62,21 @@ async def create_event_tool(*, user_id, arguments, **_ctx):
     start_str = arguments["start"]
     end_str = arguments.get("end")
     all_day = arguments.get("all_day", False)
-    if "T" not in start_str and " " not in start_str:
-        all_day = True
-        start_str = f"{start_str}T00:00:00"
     try:
-        start_dt = datetime.fromisoformat(start_str)
+        # Naive dates/datetimes are interpreted in the user's local
-        if start_dt.tzinfo is None:
+        # timezone, not UTC. Storing UTC for naive inputs caused all-day
-            start_dt = start_dt.replace(tzinfo=timezone.utc)
+        # events to land on the previous day for negative-offset users.
+        start_dt, start_was_date_only = await _parse_datetime_in_user_tz(
+            user_id, start_str
+        )
     except (ValueError, TypeError):
         return {"success": False, "error": f"Invalid start datetime: {start_str!r}"}
+    if start_was_date_only:
+        all_day = True
     end_dt = None
     if end_str:
-        if "T" not in end_str and " " not in end_str:
-            end_str = f"{end_str}T00:00:00"
         try:
-            end_dt = datetime.fromisoformat(end_str)
+            end_dt, _ = await _parse_datetime_in_user_tz(user_id, end_str)
-            if end_dt.tzinfo is None:
-                end_dt = end_dt.replace(tzinfo=timezone.utc)
         except (ValueError, TypeError):
             return {"success": False, "error": f"Invalid end datetime: {end_str!r}"}
     project_id = None
@@ -86,25 +106,30 @@ async def create_event_tool(*, user_id, arguments, **_ctx):
 
 @tool(
     name="list_events",
-    description="List calendar events in a date range. Use this when the user asks what events or meetings they have. Always use full-day UTC ranges: date_from at T00:00:00Z and date_to at T23:59:59Z for the days of interest so events stored in UTC are not missed.",
+    description="List calendar events in a date range. Use this when the user asks what events or meetings they have. Pass plain local dates (YYYY-MM-DD) — the server interprets them in the user's timezone and expands to a full local day.",
     parameters={
-        "date_from": {"type": "string", "description": "Start of range in ISO 8601 UTC format (e.g. 2025-01-15T00:00:00Z). Use T00:00:00Z for the start of the day."},
+        "date_from": {"type": "string", "description": "Start of range as a local date (YYYY-MM-DD) or local datetime. Interpreted in the user's timezone."},
-        "date_to": {"type": "string", "description": "End of range in ISO 8601 UTC format (e.g. 2025-01-22T23:59:59Z). Use T23:59:59Z for the end of the day."},
+        "date_to": {"type": "string", "description": "End of range as a local date (YYYY-MM-DD) or local datetime. A bare date is expanded to 23:59:59 local."},
     },
     required=["date_from", "date_to"],
     read_only=True,
     briefing=True,
 )
 async def list_events_tool(*, user_id, arguments, **_ctx):
+    # Bare local dates are expanded to a full local day in the user's TZ:
+    # date_from → 00:00 local, date_to → 23:59:59 local, both converted to
+    # UTC before the DB query. Previously the tool description told the
+    # model to pass UTC ranges, which missed events for non-UTC users.
     try:
-        date_from = datetime.fromisoformat(arguments["date_from"].replace("Z", "+00:00"))
+        date_from, _ = await _parse_datetime_in_user_tz(
-        date_to = datetime.fromisoformat(arguments["date_to"].replace("Z", "+00:00"))
+            user_id, arguments["date_from"]
+        )
+        date_to_str = arguments["date_to"]
+        if "T" not in date_to_str and " " not in date_to_str:
+            date_to_str = f"{date_to_str}T23:59:59"
+        date_to, _ = await _parse_datetime_in_user_tz(user_id, date_to_str)
     except (ValueError, TypeError, KeyError) as exc:
         return {"success": False, "error": f"Invalid date range: {exc}"}
-    if date_from.tzinfo is None:
-        date_from = date_from.replace(tzinfo=timezone.utc)
-    if date_to.tzinfo is None:
-        date_to = date_to.replace(tzinfo=timezone.utc)
     events = await events_list_events(user_id=user_id, date_from=date_from, date_to=date_to)
     return {
         "success": True,
@@ -177,9 +202,9 @@ async def update_event_tool(*, user_id, arguments, **_ctx):
         val = arguments.get(key)
         if val:
             try:
-                dt = datetime.fromisoformat(val.replace("Z", "+00:00"))
+                # Naive datetimes are user-local, not UTC — see
-                if dt.tzinfo is None:
+                # ``_parse_datetime_in_user_tz`` docstring.
-                    dt = dt.replace(tzinfo=timezone.utc)
+                dt, _ = await _parse_datetime_in_user_tz(user_id, val)
                 fields[dt_field] = dt
             except (ValueError, TypeError):
                 return {"success": False, "error": f"Invalid datetime for {key}: {val!r}"}

src/fabledassistant/services/tz.py

@@ -0,0 +1,47 @@
+"""User-timezone helpers.
+
+All datetimes in the DB are stored as UTC. The helpers here bridge between
+that UTC storage and the user's configured local timezone (IANA string in
+the ``user_timezone`` setting). Use these anywhere the model or UI talks
+in terms of "today", "tomorrow", or a bare calendar date — never
+``date.today()`` or ``datetime.now(timezone.utc)`` inside a per-user flow.
+"""
+from __future__ import annotations
+
+from datetime import date, datetime, timedelta
+from zoneinfo import ZoneInfo, ZoneInfoNotFoundError
+
+from fabledassistant.services.settings import get_setting
+
+# Briefing day boundary — kept in sync with the compilation slot in
+# ``briefing_scheduler.SLOTS``. The briefing day flips at this local hour
+# (not midnight) so the 00:00–04:00 local window still shows yesterday's
+# briefing until the 4am compilation generates the new one.
+BRIEFING_DAY_START_HOUR = 4
+
+
+async def get_user_tz(user_id: int) -> ZoneInfo:
+    """Return the user's IANA ``ZoneInfo``, falling back to UTC."""
+    tz_str = await get_setting(user_id, "user_timezone") or "UTC"
+    try:
+        return ZoneInfo(tz_str)
+    except (ZoneInfoNotFoundError, KeyError):
+        return ZoneInfo("UTC")
+
+
+async def user_today(user_id: int) -> date:
+    """Return today's calendar date in the user's local timezone."""
+    tz = await get_user_tz(user_id)
+    return datetime.now(tz).date()
+
+
+async def user_briefing_date(user_id: int) -> date:
+    """Return the current "briefing day" in the user's local timezone.
+
+    The briefing day flips at ``BRIEFING_DAY_START_HOUR`` (4am local),
+    aligned with the compilation slot that generates the day's briefing.
+    Between 00:00 and 04:00 local this still returns *yesterday*, so the
+    UI keeps showing the in-progress briefing until the new one is built.
+    """
+    tz = await get_user_tz(user_id)
+    return (datetime.now(tz) - timedelta(hours=BRIEFING_DAY_START_HOUR)).date()

tests/test_calendar_tool_tz.py

@@ -0,0 +1,134 @@
+"""Regression tests: calendar tool respects the user's configured timezone."""
+
+from datetime import timezone
+from unittest.mock import AsyncMock, patch
+
+import pytest
+
+
+@pytest.mark.asyncio
+async def test_create_event_all_day_bare_date_interprets_local_midnight():
+    """Bare date like '2026-09-30' for a NY user = 2026-09-30 00:00 NY,
+    stored as 2026-09-30 04:00 UTC — NOT 2026-09-30 00:00 UTC (which would
+    be 2026-09-29 19:00 NY and land on the wrong day)."""
+    from fabledassistant.services.tools.calendar import create_event_tool
+
+    captured = {}
+
+    async def fake_create_event(**kwargs):
+        captured.update(kwargs)
+        event = AsyncMock()
+        event.to_dict.return_value = {"id": 1, **{k: v for k, v in kwargs.items() if not callable(v)}}
+        return event
+
+    with patch(
+        "fabledassistant.services.tools.calendar.get_user_tz",
+        AsyncMock(return_value=__import__("zoneinfo").ZoneInfo("America/New_York")),
+    ), patch(
+        "fabledassistant.services.tools.calendar.events_create_event",
+        side_effect=fake_create_event,
+    ):
+        result = await create_event_tool(
+            user_id=1,
+            arguments={"title": "Birthday", "start": "2026-09-30"},
+        )
+
+    assert result["success"] is True
+    start_dt = captured["start_dt"]
+    assert start_dt.tzinfo is not None
+    # Converted to UTC: 00:00 NY (EDT, UTC-4) == 04:00 UTC on the same day
+    utc = start_dt.astimezone(timezone.utc)
+    assert (utc.year, utc.month, utc.day) == (2026, 9, 30)
+    assert utc.hour == 4  # EDT offset; would be 5 in EST
+    assert captured["all_day"] is True
+
+
+@pytest.mark.asyncio
+async def test_create_event_naive_datetime_is_user_local():
+    """'2026-03-15T14:30' for a NY user means 2:30pm NY, not 2:30pm UTC."""
+    from fabledassistant.services.tools.calendar import create_event_tool
+
+    captured = {}
+
+    async def fake_create_event(**kwargs):
+        captured.update(kwargs)
+        event = AsyncMock()
+        event.to_dict.return_value = {"id": 1}
+        return event
+
+    with patch(
+        "fabledassistant.services.tools.calendar.get_user_tz",
+        AsyncMock(return_value=__import__("zoneinfo").ZoneInfo("America/New_York")),
+    ), patch(
+        "fabledassistant.services.tools.calendar.events_create_event",
+        side_effect=fake_create_event,
+    ):
+        await create_event_tool(
+            user_id=1,
+            arguments={"title": "Meeting", "start": "2026-03-15T14:30"},
+        )
+
+    utc = captured["start_dt"].astimezone(timezone.utc)
+    # 14:30 NY on 2026-03-15 (after DST start) = 18:30 UTC
+    assert (utc.year, utc.month, utc.day, utc.hour, utc.minute) == (2026, 3, 15, 18, 30)
+
+
+@pytest.mark.asyncio
+async def test_create_event_explicit_offset_is_respected():
+    """Model-supplied timezone offsets must not be overridden by user TZ."""
+    from fabledassistant.services.tools.calendar import create_event_tool
+
+    captured = {}
+
+    async def fake_create_event(**kwargs):
+        captured.update(kwargs)
+        event = AsyncMock()
+        event.to_dict.return_value = {"id": 1}
+        return event
+
+    with patch(
+        "fabledassistant.services.tools.calendar.get_user_tz",
+        AsyncMock(return_value=__import__("zoneinfo").ZoneInfo("America/New_York")),
+    ), patch(
+        "fabledassistant.services.tools.calendar.events_create_event",
+        side_effect=fake_create_event,
+    ):
+        await create_event_tool(
+            user_id=1,
+            arguments={"title": "Meeting", "start": "2026-03-15T14:30+00:00"},
+        )
+
+    utc = captured["start_dt"].astimezone(timezone.utc)
+    assert (utc.hour, utc.minute) == (14, 30)
+
+
+@pytest.mark.asyncio
+async def test_list_events_bare_date_range_covers_local_day():
+    """date_from/date_to as bare dates should cover the full LOCAL day."""
+    from fabledassistant.services.tools.calendar import list_events_tool
+
+    captured = {}
+
+    async def fake_list_events(*, user_id, date_from, date_to):
+        captured["date_from"] = date_from
+        captured["date_to"] = date_to
+        return []
+
+    with patch(
+        "fabledassistant.services.tools.calendar.get_user_tz",
+        AsyncMock(return_value=__import__("zoneinfo").ZoneInfo("America/New_York")),
+    ), patch(
+        "fabledassistant.services.tools.calendar.events_list_events",
+        side_effect=fake_list_events,
+    ):
+        await list_events_tool(
+            user_id=1,
+            arguments={"date_from": "2026-09-30", "date_to": "2026-09-30"},
+        )
+
+    df = captured["date_from"].astimezone(timezone.utc)
+    dt = captured["date_to"].astimezone(timezone.utc)
+    # 2026-09-30 00:00 NY (EDT) = 04:00 UTC same day
+    assert (df.year, df.month, df.day, df.hour) == (2026, 9, 30, 4)
+    # 2026-09-30 23:59:59 NY (EDT) = 03:59:59 UTC next day
+    assert (dt.year, dt.month, dt.day, dt.hour) == (2026, 10, 1, 3)

tests/test_tz_helpers.py

@@ -0,0 +1,63 @@
+"""Tests for user-timezone helpers (services/tz.py)."""
+
+from datetime import datetime
+from unittest.mock import AsyncMock, patch
+from zoneinfo import ZoneInfo
+
+import pytest
+
+
+@pytest.mark.asyncio
+async def test_get_user_tz_returns_configured_zone():
+    from fabledassistant.services import tz as tz_mod
+
+    with patch.object(tz_mod, "get_setting", AsyncMock(return_value="America/New_York")):
+        zone = await tz_mod.get_user_tz(1)
+    assert zone == ZoneInfo("America/New_York")
+
+
+@pytest.mark.asyncio
+async def test_get_user_tz_falls_back_to_utc_on_bad_input():
+    from fabledassistant.services import tz as tz_mod
+
+    with patch.object(tz_mod, "get_setting", AsyncMock(return_value="Not/AZone")):
+        zone = await tz_mod.get_user_tz(1)
+    assert zone == ZoneInfo("UTC")
+
+
+@pytest.mark.asyncio
+async def test_user_briefing_date_before_4am_returns_yesterday():
+    """00:00–03:59 local still shows yesterday's briefing day."""
+    from fabledassistant.services import tz as tz_mod
+
+    ny = ZoneInfo("America/New_York")
+    # 02:00 NY on 2026-04-13 → briefing day = 2026-04-12
+    fake_now = datetime(2026, 4, 13, 2, 0, tzinfo=ny)
+
+    class _FakeDatetime(datetime):
+        @classmethod
+        def now(cls, tz=None):
+            return fake_now
+
+    with patch.object(tz_mod, "get_user_tz", AsyncMock(return_value=ny)), \
+         patch.object(tz_mod, "datetime", _FakeDatetime):
+        day = await tz_mod.user_briefing_date(1)
+    assert day.isoformat() == "2026-04-12"
+
+
+@pytest.mark.asyncio
+async def test_user_briefing_date_after_4am_returns_today():
+    from fabledassistant.services import tz as tz_mod
+
+    ny = ZoneInfo("America/New_York")
+    fake_now = datetime(2026, 4, 13, 4, 30, tzinfo=ny)
+
+    class _FakeDatetime(datetime):
+        @classmethod
+        def now(cls, tz=None):
+            return fake_now
+
+    with patch.object(tz_mod, "get_user_tz", AsyncMock(return_value=ny)), \
+         patch.object(tz_mod, "datetime", _FakeDatetime):
+        day = await tz_mod.user_briefing_date(1)
+    assert day.isoformat() == "2026-04-13"