Merge pull request 'Release v26.04.14.2 — Discuss failure mode fixes' (#33) from dev into main

Discuss flow was hallucinating unrelated content when article
extraction returned empty or RAG pulled in orphan notes that looked
more relevant than the generic seed prompt.

- seed_article_discussion raises EmptyArticleError on empty body;
  briefing and /news routes return 422 instead of staging an empty
  synthetic tool result.
- build_context skips RAG auto-injection when user_message matches
  ARTICLE_DISCUSS_SEED so the article IS the context on turn one;
  follow-up turns keep RAG on.

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

.forgejo/workflows/ci.yml

@@ -1,8 +1,12 @@
 # CI runs first; build only proceeds if all checks pass.
 #
-# Push to any branch: typecheck + lint + test
+# Push to dev:       typecheck + lint + test + build :dev + :<sha>
-# Push to dev:        gates + build :dev + :<sha>
+# Tag v* (release):  typecheck + lint + test + build :latest + :<sha> + :<version>
-# Tag v* (release):   gates + build :latest + :<sha> + :<version>
+#
+# main pushes are NOT gated here: a merge to main only happens after
+# dev has already passed CI, and the release tag is the sole trigger
+# for a production image. Re-running CI on the merge commit just burns
+# runner time without changing the outcome.
 #
 # To cut a release:
 #   Create a release via the Forgejo UI on main with a v* tag name.
@@ -11,6 +15,13 @@
 # PRs aren't triggered on purpose — this is a solo dev→main flow, so
 # gating on branch push is already enough.
 #
+# NOTE on the `if:` guards below: Forgejo Actions does not consistently
+# honor `on.push.branches` as a filter — merge commits landing on main
+# still trigger the workflow, producing redundant runs on the same SHA
+# that was already gated on dev. Every job therefore repeats the ref
+# check so main pushes trigger the workflow but every job skips
+# immediately (no runner time, no duplicate work).
+#
 # Required secrets (repo → Settings → Secrets → Actions):
 #   REGISTRY_USER   — your Forgejo username
 #   REGISTRY_TOKEN  — Forgejo PAT with write:packages scope
@@ -18,7 +29,7 @@ name: CI & Build
 
 on:
   push:
-    branches: [dev, main]
+    branches: [dev]
     tags: ["v*"]
     paths:
       - "src/**"
@@ -51,6 +62,8 @@ env:
 jobs:
   typecheck:
     name: TypeScript typecheck
+    # Skip on main merge-commit pushes — see workflow header comment.
+    if: github.ref == 'refs/heads/dev' || startsWith(github.ref, 'refs/tags/v')
     runs-on: ci-runner
     steps:
       - uses: actions/checkout@v6
@@ -77,6 +90,7 @@ jobs:
 
   lint:
     name: Python lint
+    if: github.ref == 'refs/heads/dev' || startsWith(github.ref, 'refs/tags/v')
     runs-on: ci-runner
     steps:
       - uses: actions/checkout@v6
@@ -88,6 +102,7 @@ jobs:
 
   test:
     name: Python tests
+    if: github.ref == 'refs/heads/dev' || startsWith(github.ref, 'refs/tags/v')
     runs-on: ci-runner
     steps:
       - uses: actions/checkout@v6
@@ -118,11 +133,10 @@ jobs:
     name: Build & push image
     needs: [typecheck, lint, test]
     # Build on dev branch pushes and version tag pushes only.
-    # main branch pushes run the gates for safety but do not build —
+    # Mirrors the ref guard on the gate jobs above — main merge-commit
-    # the release tag (v*) is the sole trigger for a production image.
+    # pushes skip here too, so no production image is ever built from a
-    if: |
+    # raw main push (only from the v* tag the release creates).
-      github.event_name == 'push' &&
+    if: github.ref == 'refs/heads/dev' || startsWith(github.ref, 'refs/tags/v')
-      (github.ref == 'refs/heads/dev' || startsWith(github.ref, 'refs/tags/'))
     runs-on: ci-runner
     permissions:
       contents: read

alembic/versions/0038_add_rss_item_full_content.py

@@ -0,0 +1,34 @@
+"""Add content_full and context_prepared caches to rss_items.
+
+Revision ID: 0038
+Revises: 0037
+Create Date: 2026-04-13
+"""
+
+from __future__ import annotations
+
+import sqlalchemy as sa
+from alembic import op
+
+revision = "0038"
+down_revision = "0037"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    op.add_column("rss_items", sa.Column("content_full", sa.Text(), nullable=True))
+    op.add_column(
+        "rss_items",
+        sa.Column("context_prepared", sa.Text(), nullable=True),
+    )
+    op.add_column(
+        "rss_items",
+        sa.Column("content_fetched_at", sa.DateTime(timezone=True), nullable=True),
+    )
+
+
+def downgrade() -> None:
+    op.drop_column("rss_items", "content_fetched_at")
+    op.drop_column("rss_items", "context_prepared")
+    op.drop_column("rss_items", "content_full")

alembic/versions/0039_add_rss_item_discussion_note.py

@@ -0,0 +1,38 @@
+"""Link rss_items to their discussion-summary note.
+
+Revision ID: 0039
+Revises: 0038
+Create Date: 2026-04-14
+"""
+
+from __future__ import annotations
+
+import sqlalchemy as sa
+from alembic import op
+
+revision = "0039"
+down_revision = "0038"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    op.add_column(
+        "rss_items",
+        sa.Column("discussion_note_id", sa.BigInteger(), nullable=True),
+    )
+    op.create_foreign_key(
+        "fk_rss_items_discussion_note_id",
+        "rss_items",
+        "notes",
+        ["discussion_note_id"],
+        ["id"],
+        ondelete="SET NULL",
+    )
+
+
+def downgrade() -> None:
+    op.drop_constraint(
+        "fk_rss_items_discussion_note_id", "rss_items", type_="foreignkey"
+    )
+    op.drop_column("rss_items", "discussion_note_id")

frontend/src/api/client.ts

@@ -426,7 +426,9 @@ export async function deleteRssReaction(rssItemId: number): Promise<void> {
   return apiDelete(`/api/briefing/rss-reactions/${rssItemId}`);
 }
 
-export async function openArticleInChat(itemId: number): Promise<{ conversation_id: number }> {
+export async function openArticleInChat(
+  itemId: number,
+): Promise<{ conversation_id: number; assistant_message_id?: number; status?: string }> {
   return apiPost(`/api/chat/from-article/${itemId}`, {});
 }
 

src/fabledassistant/models/rss_feed.py

@@ -1,6 +1,6 @@
 from datetime import datetime, timezone
 
-from sqlalchemy import ARRAY, DateTime, ForeignKey, Index, Integer, Text, UniqueConstraint
+from sqlalchemy import ARRAY, BigInteger, DateTime, ForeignKey, Index, Integer, Text, UniqueConstraint
 from sqlalchemy.orm import Mapped, mapped_column, relationship
 
 from fabledassistant.models import Base
@@ -46,6 +46,17 @@ class RssItem(Base):
     published_at: Mapped[datetime | None] = mapped_column(DateTime(timezone=True), nullable=True)
     # Truncated to 2000 chars to keep DB size reasonable
     content: Mapped[str] = mapped_column(Text, default="")
+    # Full trafilatura-extracted article body, populated lazily on first
+    # discuss-click / enrichment pass. Nullable — most items never get this
+    # cached. Expires naturally with the item (90-day retention).
+    content_full: Mapped[str | None] = mapped_column(Text, nullable=True)
+    # Map-reduced conversation-ready context derived from content_full. See
+    # services/article_context.py — populated on first discuss click so
+    # repeat clicks skip both the fetch and the LLM map step.
+    context_prepared: Mapped[str | None] = mapped_column(Text, nullable=True)
+    content_fetched_at: Mapped[datetime | None] = mapped_column(
+        DateTime(timezone=True), nullable=True
+    )
     fetched_at: Mapped[datetime] = mapped_column(
         DateTime(timezone=True), default=lambda: datetime.now(timezone.utc)
     )
@@ -55,6 +66,13 @@ class RssItem(Base):
     classified_at: Mapped[datetime | None] = mapped_column(
         DateTime(timezone=True), nullable=True
     )
+    # Note persisting the first-click discussion summary. Set by the article
+    # discussion pipeline once the seeded chat completes its first assistant
+    # reply; links back into RAG so re-discussing the same article lands the
+    # prior summary in context.
+    discussion_note_id: Mapped[int | None] = mapped_column(
+        BigInteger, ForeignKey("notes.id", ondelete="SET NULL"), nullable=True
+    )
 
     feed: Mapped["RssFeed"] = relationship(back_populates="items")
 

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(
@@ -529,25 +532,21 @@ async def discuss_article(item_id: int):
     if get_buffer(conv_id) is not None:
         return jsonify({"error": "Generation already in progress"}), 409
 
-    from fabledassistant.services.rss import _fetch_full_article
+    # Shared helper handles the three-layer cache (context_prepared →
-    article_content = await _fetch_full_article(item.url) or item.content or ""
+    # content_full → fresh fetch), writes the synthetic read_article tool
+    # exchange and the conversational seed user prompt into the conversation.
+    # The /news from-article route calls the same helper so behavior stays
+    # byte-identical across entry points.
+    from fabledassistant.services.article_context import (
+        EmptyArticleError,
+        seed_article_discussion,
+    )
 
-    # Store synthetic assistant message with read_article tool result
+    model = await get_setting(uid, "default_model", "") or ""
-    synthetic_tool_calls = [{
+    try:
-        "function": "read_article",
+        discuss_prompt = await seed_article_discussion(conv_id, item, model)
-        "arguments": {"url": item.url},
+    except EmptyArticleError as e:
-        "result": {
+        return jsonify({"error": str(e)}), 422
-            "success": True,
-            "type": "article_content",
-            "url": item.url,
-            "content": article_content,
-            "truncated": False,
-        },
-    }]
-    await add_message(conv_id, "assistant", "", status="complete", tool_calls=synthetic_tool_calls)
-
-    # Store user message
-    await add_message(conv_id, "user", "Please summarize and discuss this article.")
 
     # Reload conversation with fresh messages to build history
     conv = await get_conversation(uid, conv_id)
@@ -569,15 +568,13 @@ async def discuss_article(item_id: int):
         else:
             history.append(msg_dict)
 
-    model = await get_setting(uid, "default_model", "") or ""
-
     assistant_msg = await add_message(conv_id, "assistant", "", status="generating")
     buf = create_buffer(conv_id, assistant_msg.id)
 
     asyncio.create_task(run_generation(
         buf, history, model,
         uid, conv_id, conv.title or "",
-        "Please summarize and discuss this article.",
+        discuss_prompt,
     ))
 
     return jsonify({"assistant_message_id": assistant_msg.id, "status": "generating"}), 202

src/fabledassistant/routes/chat.py

@@ -510,47 +510,78 @@ async def delete_model_route():
 @chat_bp.route("/from-article/<int:item_id>", methods=["POST"])
 @login_required
 async def create_conversation_from_article(item_id: int):
-    """Create a chat conversation seeded with an RSS article's content."""
+    """Create a chat conversation seeded for article discussion and auto-run.
+
+    Mirrors the briefing ``discuss_article`` route: creates a fresh
+    conversation, stages the shared synthetic read_article exchange + seed
+    prompt, then kicks off generation so the client lands on an in-flight
+    stream. The Flutter and web chat screens reconnect to the running buffer
+    on mount.
+    """
     from sqlalchemy import select as _select
     from fabledassistant.models import async_session as _async_session
-    from fabledassistant.models.conversation import Message
     from fabledassistant.models.rss_feed import RssItem, RssFeed
+    from fabledassistant.services.article_context import (
+        EmptyArticleError,
+        seed_article_discussion,
+    )
 
     uid = get_current_user_id()
 
     async with _async_session() as session:
         result = await session.execute(
-            _select(RssItem, RssFeed.title.label("feed_title"))
+            _select(RssItem)
             .join(RssFeed, RssItem.feed_id == RssFeed.id)
             .where(RssItem.id == item_id, RssFeed.user_id == uid)
         )
-        row = result.first()
+        item = result.scalars().first()
 
-    if row is None:
+    if item is None:
         return jsonify({"error": "Article not found"}), 404
 
-    item, feed_title = row
-
     conv_title = (item.title or "Article discussion")[:80]
     conv = await create_conversation(uid, title=conv_title, conversation_type="chat")
 
-    from fabledassistant.services.rss import _fetch_full_article
+    model = await get_setting(uid, "default_model", "") or Config.OLLAMA_MODEL
-    source = feed_title or "News"
+    try:
-    content_body = (await _fetch_full_article(item.url) if item.url else None) or (item.content or "").strip()
+        discuss_prompt = await seed_article_discussion(conv.id, item, model)
-    seeded_text = f"**{source}**\n\n**{item.title}**"
+    except EmptyArticleError as e:
-    if content_body:
+        # Roll back the empty conversation so the user doesn't end up with a
-        seeded_text += f"\n\n{content_body}"
+        # phantom entry in their chat list.
-    if item.url:
+        try:
-        seeded_text += f"\n\nSource: {item.url}"
+            await delete_conversation(uid, conv.id)
+        except Exception:
+            logger.warning("Failed to clean up empty article conversation %s", conv.id)
+        return jsonify({"error": str(e)}), 422
 
-    async with _async_session() as session:
+    # Reload conversation so we see the two messages the helper just added.
-        msg = Message(
+    conv = await get_conversation(uid, conv.id)
-            conversation_id=conv.id,
+    assert conv is not None
-            role="assistant",
-            content=seeded_text,
-            msg_metadata={"rss_item_ids": [item_id]},
-        )
-        session.add(msg)
-        await session.commit()
 
-    return jsonify({"conversation_id": conv.id}), 201
+    history: list[dict] = []
+    for msg in conv.messages:
+        if msg.role == "system":
+            continue
+        msg_dict: dict = {"role": msg.role, "content": msg.content or ""}
+        if msg.tool_calls:
+            msg_dict["tool_calls"] = [
+                {"function": {"name": tc["function"], "arguments": tc["arguments"]}}
+                for tc in msg.tool_calls
+            ]
+            history.append(msg_dict)
+            for tc in msg.tool_calls:
+                history.append({"role": "tool", "content": json.dumps(tc.get("result", {}))})
+        else:
+            history.append(msg_dict)
+
+    assistant_msg = await add_message(conv.id, "assistant", "", status="generating")
+    buf = create_buffer(conv.id, assistant_msg.id)
+    asyncio.create_task(run_generation(
+        buf, history, model, uid, conv.id, conv.title or "", discuss_prompt,
+    ))
+
+    return jsonify({
+        "conversation_id": conv.id,
+        "assistant_message_id": assistant_msg.id,
+        "status": "generating",
+    }), 202

src/fabledassistant/services/article_context.py

@@ -0,0 +1,270 @@
+"""Prepare article bodies as conversation-ready context.
+
+Used by the briefing ``discuss-article`` flow and the ``/news`` discuss button.
+A raw trafilatura extraction is often too large to drop whole into a chat
+history without eating the context window, so this module runs a map-reduce
+step over oversized articles and returns a compact, structured context that
+still preserves the article's meaning across sections.
+
+Small articles pass through unchanged — map-reduce only fires when the raw
+body exceeds CHAR_BUDGET. The output is cached on ``rss_items.context_prepared``
+by the caller, so repeat discuss-clicks on the same article skip this work
+entirely.
+
+The module also owns ``seed_article_discussion``, the shared routine that
+stages a synthetic ``read_article`` tool exchange plus a conversational seed
+prompt into a conversation. Both the briefing and ``/news`` entry points call
+it so the two flows stay byte-identical — the only thing that differs between
+them is whether the conversation already existed or was freshly created.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+import re
+
+from fabledassistant.models import async_session
+from fabledassistant.models.rss_feed import RssItem
+from fabledassistant.services.chat import add_message
+from fabledassistant.services.llm import generate_completion
+
+logger = logging.getLogger(__name__)
+
+# ~12k tokens at 4 chars/token. Comfortably under OLLAMA_NUM_CTX=16384
+# with room left for system prompt, chat history, and the assistant reply.
+CHAR_BUDGET = 48_000
+
+# Chunk size for the map step on oversized articles. Overlap preserves
+# context across paragraph boundaries that happen to land mid-sentence.
+CHUNK_CHARS = 8_000
+CHUNK_OVERLAP = 400
+
+_PARA_SPLIT = re.compile(r"\n\s*\n")
+
+
+def _chunk_by_paragraph(body: str) -> list[str]:
+    """Split ``body`` into chunks of up to CHUNK_CHARS, respecting paragraphs.
+
+    Paragraphs longer than CHUNK_CHARS are split mid-paragraph as a last
+    resort. Adjacent chunks share CHUNK_OVERLAP chars of trailing text so
+    a sentence straddling the boundary stays readable on both sides.
+    """
+    paragraphs = [p.strip() for p in _PARA_SPLIT.split(body) if p.strip()]
+    chunks: list[str] = []
+    current: list[str] = []
+    current_len = 0
+    for para in paragraphs:
+        para_len = len(para)
+        if para_len > CHUNK_CHARS:
+            if current:
+                chunks.append("\n\n".join(current))
+                current, current_len = [], 0
+            for i in range(0, para_len, CHUNK_CHARS - CHUNK_OVERLAP):
+                chunks.append(para[i : i + CHUNK_CHARS])
+            continue
+        if current_len + para_len + 2 > CHUNK_CHARS and current:
+            chunks.append("\n\n".join(current))
+            tail = current[-1][-CHUNK_OVERLAP:] if current else ""
+            current = [tail, para] if tail else [para]
+            current_len = len(tail) + para_len + (2 if tail else 0)
+        else:
+            current.append(para)
+            current_len += para_len + 2
+    if current:
+        chunks.append("\n\n".join(current))
+    return chunks
+
+
+async def _summarize_chunk(title: str, chunk: str, index: int, total: int, model: str) -> str:
+    """Map-step summary of one article chunk.
+
+    Aims for ~300 words of dense, factual prose — not bullet points — so the
+    downstream chat model can quote from it naturally.
+    """
+    messages = [
+        {
+            "role": "system",
+            "content": (
+                "You are summarizing one section of a larger article so a downstream "
+                "conversation model can discuss the full article without having to read "
+                "every word.\n\n"
+                "Requirements:\n"
+                "- 250–350 words of dense factual prose\n"
+                "- Preserve specific claims, numbers, names, and quotes\n"
+                "- Do NOT editorialize or add analysis\n"
+                "- Do NOT use bullet points or headings\n"
+                "- Do NOT say 'this section' or 'this article' — write content, not meta"
+            ),
+        },
+        {
+            "role": "user",
+            "content": (
+                f"Article: {title}\n"
+                f"Section {index + 1} of {total}:\n\n{chunk}"
+            ),
+        },
+    ]
+    try:
+        # Pin num_ctx — same rationale as services/research.py:66. A large
+        # chunk plus system prompt can push well past the default window;
+        # silent truncation here would drop the tail of the chunk without
+        # any error, producing a misleading summary.
+        raw = await generate_completion(
+            messages, model, max_tokens=600, num_ctx=16384
+        )
+        return raw.strip()
+    except Exception:
+        logger.warning(
+            "Article chunk summary failed for section %d/%d of '%s'",
+            index + 1, total, title, exc_info=True,
+        )
+        # Fall back to the raw chunk truncated to ~1500 chars so the overall
+        # pipeline still delivers something rather than dropping the section.
+        return chunk[:1500]
+
+
+async def prepare_article_context(
+    title: str,
+    url: str,
+    body: str,
+    model: str,
+) -> str:
+    """Return a conversation-ready context block for ``body``.
+
+    - Small article (≤ CHAR_BUDGET): returns ``body`` unchanged.
+    - Oversized article: runs a parallel map step over paragraph-aware
+      chunks and concatenates the summaries under section headers.
+
+    The returned string is what should go into the ``read_article`` synthetic
+    tool-result in chat history. Callers are responsible for caching it to
+    ``rss_items.context_prepared``.
+    """
+    body = body or ""
+    if len(body) <= CHAR_BUDGET:
+        return body
+
+    chunks = _chunk_by_paragraph(body)
+    logger.info(
+        "Article '%s' is %d chars, map-reducing into %d chunks",
+        title, len(body), len(chunks),
+    )
+
+    summaries = await asyncio.gather(
+        *[
+            _summarize_chunk(title, chunk, i, len(chunks), model)
+            for i, chunk in enumerate(chunks)
+        ]
+    )
+
+    header = (
+        f"(This article was longer than the chat window could hold verbatim, "
+        f"so the full text was split into {len(chunks)} sections and each was "
+        "summarized below. Each section preserves specific claims, numbers, "
+        "and quotes from the original.)\n\n"
+    )
+    parts = [
+        f"## Section {i + 1}\n\n{summary}"
+        for i, summary in enumerate(summaries)
+    ]
+    return header + "\n\n".join(parts)
+
+
+# Conversational seed prompt for article discussions. Kept here so both the
+# briefing and /news entry points use the exact same wording. See
+# feedback_discuss_prompt_style memory: numbered checklists produce
+# assignment-completion responses; this conversational seed opens a dialogue.
+ARTICLE_DISCUSS_SEED = (
+    "I want to talk about this article. Start with a substantive summary "
+    "of what it's arguing and the key evidence it uses, then tell me what "
+    "stood out to you or seems worth pushing back on. I'll ask follow-ups "
+    "from there."
+)
+
+
+class EmptyArticleError(Exception):
+    """Raised when an article has no extractable body text.
+
+    Callers (the briefing and /news discuss routes) map this to a 422 so the
+    user sees a clear error instead of a hallucinated summary built from an
+    empty synthetic tool result.
+    """
+
+
+async def seed_article_discussion(
+    conv_id: int,
+    item: RssItem,
+    model: str,
+) -> str:
+    """Stage the synthetic read_article tool exchange + conversational seed.
+
+    Used by both the briefing ``discuss_article`` route and the ``/news``
+    ``from-article`` conversation creator. Handles the three-layer cache
+    (``context_prepared`` → ``content_full`` → fresh fetch) and inserts two
+    messages into ``conv_id``:
+
+    1. An assistant message with a synthetic ``read_article`` tool_call whose
+       ``result.content`` carries the prepared article context. The message
+       also carries ``msg_metadata={"rss_item_id": ...}`` so the post-generation
+       hook in ``generation_task.py`` can locate it and persist the first
+       reply as a discussion-summary Note.
+    2. A user message with the shared conversational seed prompt.
+
+    Returns the seed prompt string so callers can pass it to ``run_generation``
+    as ``user_content``.
+    """
+    # Avoid circulars: rss helper imports article_context indirectly nowhere,
+    # but keep this local for symmetry with the route-level imports it
+    # replaces.
+    from fabledassistant.services.rss import get_or_fetch_full_article
+
+    if item.context_prepared:
+        article_content = item.context_prepared
+    else:
+        raw_body = await get_or_fetch_full_article(item) or item.content or ""
+        if not raw_body.strip():
+            # Hard-fail rather than stage an empty synthetic tool result.
+            # An empty `content` field silently tells the model "the article
+            # has nothing in it" and it confabulates from RAG/history. Better
+            # to surface a clean error to the user.
+            logger.warning(
+                "Article discussion aborted: empty body for rss_item %s (%s)",
+                item.id, item.url,
+            )
+            raise EmptyArticleError(
+                "Couldn't extract any readable text from this article."
+            )
+        article_content = await prepare_article_context(
+            item.title or "", item.url, raw_body, model,
+        )
+        if not article_content.strip():
+            raise EmptyArticleError(
+                "Couldn't extract any readable text from this article."
+            )
+        async with async_session() as session:
+            fresh = await session.get(RssItem, item.id)
+            if fresh is not None:
+                fresh.context_prepared = article_content
+                await session.commit()
+
+    synthetic_tool_calls = [{
+        "function": "read_article",
+        "arguments": {"url": item.url},
+        "result": {
+            "success": True,
+            "type": "article_content",
+            "url": item.url,
+            "content": article_content,
+            "truncated": False,
+        },
+    }]
+    await add_message(
+        conv_id,
+        "assistant",
+        "",
+        status="complete",
+        tool_calls=synthetic_tool_calls,
+        msg_metadata={"rss_item_id": item.id, "article_seed": True},
+    )
+    await add_message(conv_id, "user", ARTICLE_DISCUSS_SEED)
+    return ARTICLE_DISCUSS_SEED

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/chat.py

@@ -187,6 +187,7 @@ async def add_message(
     context_note_id: int | None = None,
     status: str | None = None,
     tool_calls: list | None = None,
+    msg_metadata: dict | None = None,
 ) -> Message:
     async with async_session() as session:
         kwargs: dict = dict(
@@ -199,6 +200,8 @@ async def add_message(
             kwargs["status"] = status
         if tool_calls is not None:
             kwargs["tool_calls"] = tool_calls
+        if msg_metadata is not None:
+            kwargs["msg_metadata"] = msg_metadata
         msg = Message(**kwargs)
         session.add(msg)
         # Touch conversation updated_at

src/fabledassistant/services/generation_task.py

@@ -36,6 +36,85 @@ _TOOL_CALL_MARKER = re.compile(r"^\s*\[TOOL_CALLS\]\s*", re.IGNORECASE)
 
 DB_FLUSH_INTERVAL = 5.0  # seconds between partial DB flushes
 
+
+async def _maybe_save_article_discussion_note(
+    user_id: int, conv_id: int, reply_content: str,
+) -> None:
+    """Persist a seeded article-discussion's first reply as a Note.
+
+    Fires after ``run_generation`` completes. Looks for a synthetic
+    read_article seed message on the conversation; if found AND the linked
+    ``rss_items`` row has no ``discussion_note_id`` yet, saves ``reply_content``
+    as a Note, tags it, and writes the backlink. Subsequent discuss clicks on
+    the same article are a no-op (already linked).
+
+    Failures are logged and swallowed — the chat UI should never break because
+    Note persistence hit a snag.
+    """
+    try:
+        if not reply_content or not reply_content.strip():
+            return
+        from sqlalchemy import select as _select
+        from fabledassistant.models.conversation import Message as _Message
+        from fabledassistant.models.rss_feed import RssItem as _RssItem
+        from fabledassistant.services.notes import create_note
+
+        async with async_session() as session:
+            result = await session.execute(
+                _select(_Message)
+                .where(_Message.conversation_id == conv_id)
+                .order_by(_Message.id.asc())
+            )
+            messages = result.scalars().all()
+            seed_meta = None
+            for m in messages:
+                meta = m.msg_metadata or {}
+                if meta.get("article_seed") and meta.get("rss_item_id"):
+                    seed_meta = meta
+                    break
+            if seed_meta is None:
+                return
+            item_id = int(seed_meta["rss_item_id"])
+            item = await session.get(_RssItem, item_id)
+            if item is None or item.discussion_note_id is not None:
+                return
+            article_title = (item.title or "Untitled article").strip()
+            article_url = item.url
+            article_topics = list(item.topics or [])
+
+        note_title = f"Article: {article_title}"[:200]
+        body_parts = [f"**Source:** {article_url}"] if article_url else []
+        body_parts.append(reply_content.strip())
+        note_body = "\n\n".join(body_parts)
+        tags = ["article-summary"] + [t for t in article_topics if t]
+        note = await create_note(
+            user_id=user_id,
+            title=note_title,
+            body=note_body,
+            tags=tags,
+            entity_meta={
+                "source": "article_discussion",
+                "rss_item_id": item_id,
+                "url": article_url,
+                "conversation_id": conv_id,
+            },
+        )
+
+        async with async_session() as session:
+            fresh = await session.get(_RssItem, item_id)
+            if fresh is not None and fresh.discussion_note_id is None:
+                fresh.discussion_note_id = note.id
+                await session.commit()
+        logger.info(
+            "Saved article-discussion summary as note %d for rss_item %d (conv %d)",
+            note.id, item_id, conv_id,
+        )
+    except Exception:
+        logger.warning(
+            "Failed to persist article-discussion note for conv %d",
+            conv_id, exc_info=True,
+        )
+
 # ---------------------------------------------------------------------------
 # Thinking decision
 # ---------------------------------------------------------------------------
@@ -526,10 +605,35 @@ async def run_generation(
         msg_count = len(non_system)
         should_gen_title = not conv_title or (msg_count > 0 and msg_count % 10 == 0)
 
+        # Persist article-discussion seed conversations as a Note on their
+        # first assistant reply. This makes "Discuss" summaries part of RAG
+        # so the knowledge base stops being amnesiac about articles the user
+        # has already engaged with. The hook detects a seeded conversation by
+        # finding a synthetic read_article assistant message whose
+        # msg_metadata carries ``article_seed: True`` and whose rss_items row
+        # has no discussion_note_id yet. Fire-and-forget so the done event
+        # lands immediately.
+        asyncio.create_task(_maybe_save_article_discussion_note(
+            user_id, conv_id, buf.content_so_far,
+        ))
+
         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
@@ -716,18 +724,27 @@ async def build_context(
     orphan_only = rag_project_id is None
     effective_project_id = rag_project_id if (rag_project_id is not None and rag_project_id != -1) else None
 
-    try:
+    # Skip RAG auto-injection on the first turn of a seeded article discussion.
-        from fabledassistant.services.embeddings import semantic_search_notes
+    # The article body is already the sole context the user wants — pulling in
-        for score, note in await semantic_search_notes(
+    # unrelated orphan notes tricks the model into summarizing those instead.
-            user_id, user_message, exclude_ids=search_exclude or None, limit=8,
+    # Follow-up turns keep RAG on because by then the user's own messages drive
-            project_id=effective_project_id,
+    # the query rather than the generic seed prompt.
-            orphan_only=orphan_only,
+    from fabledassistant.services.article_context import ARTICLE_DISCUSS_SEED
-        ):
+    _skip_rag_for_article_seed = user_message.strip() == ARTICLE_DISCUSS_SEED
-            found_scored.append((score, note))
-    except Exception:
-        logger.warning("Semantic note search failed, falling back to keyword search", exc_info=True)
 
-    if not found_scored:
+    if not _skip_rag_for_article_seed:
+        try:
+            from fabledassistant.services.embeddings import semantic_search_notes
+            for score, note in await semantic_search_notes(
+                user_id, user_message, exclude_ids=search_exclude or None, limit=8,
+                project_id=effective_project_id,
+                orphan_only=orphan_only,
+            ):
+                found_scored.append((score, note))
+        except Exception:
+            logger.warning("Semantic note search failed, falling back to keyword search", exc_info=True)
+
+    if not found_scored and not _skip_rag_for_article_seed:
         keywords = _extract_keywords(user_message)
         if keywords:
             try:

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/rss.py

@@ -34,6 +34,34 @@ def _html_to_text(html: str) -> str:
         return html
 
 
+async def get_or_fetch_full_article(item: RssItem) -> str | None:
+    """Return the full article body, fetching+caching on miss.
+
+    Checks ``item.content_full`` first — populated either by the enrichment
+    pass at feed-ingest time or by a previous discuss-click. On miss, fetches
+    via ``_fetch_full_article`` and writes through. Returns ``None`` only if
+    the fetch itself fails; ``item.content_full == ""`` is still a cache hit.
+
+    Callers must pass an RssItem attached to an open session if they want
+    the write-through to persist — otherwise the fetched text is returned
+    but the cache stays empty and the next click will re-fetch.
+    """
+    if item.content_full is not None:
+        return item.content_full
+    if not item.url:
+        return None
+    text = await _fetch_full_article(item.url)
+    if text is None:
+        return None
+    async with async_session() as session:
+        fresh = await session.get(RssItem, item.id)
+        if fresh is not None:
+            fresh.content_full = text
+            fresh.content_fetched_at = datetime.now(timezone.utc)
+            await session.commit()
+    return text
+
+
 async def _fetch_full_article(url: str) -> str | None:
     """Fetch a URL and extract its main article text via trafilatura.
 
@@ -209,6 +237,11 @@ async def fetch_and_cache_feed(feed_id: int, url: str) -> int:
                         item = await session.get(RssItem, item_id)
                         if item:
                             item.content = full_text
+                            # Populate the discuss-click cache too so the
+                            # first click skips straight to the map-reduce
+                            # step without re-fetching.
+                            item.content_full = full_text
+                            item.content_fetched_at = datetime.now(timezone.utc)
                             await session.commit()
                             await upsert_rss_item_embedding(
                                 item_id, feed_user_id, item.title or "", item.content

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_article_reading.py

@@ -134,3 +134,83 @@ def test_history_builder_no_tool_calls_unchanged():
     assert len(history) == 2
     assert history[0] == {"role": "user", "content": "Hello"}
     assert history[1] == {"role": "assistant", "content": "Hi there!"}
+
+
+# ---------------------------------------------------------------------------
+# prepare_article_context tests
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.asyncio
+async def test_prepare_article_context_small_passthrough():
+    """Articles under CHAR_BUDGET pass through unchanged with zero LLM calls."""
+    from fabledassistant.services import article_context
+
+    body = "A short article.\n\nWith two paragraphs."
+    with patch(
+        "fabledassistant.services.article_context.generate_completion",
+        new_callable=AsyncMock,
+    ) as mock_gen:
+        out = await article_context.prepare_article_context(
+            "Title", "https://example.com", body, "test-model",
+        )
+
+    assert out == body
+    mock_gen.assert_not_called()
+
+
+@pytest.mark.asyncio
+async def test_prepare_article_context_large_runs_map_reduce():
+    """Articles over CHAR_BUDGET are chunked and map-reduced via the background model."""
+    from fabledassistant.services import article_context
+
+    # CHAR_BUDGET is 48_000 — build a body well over that with paragraph breaks
+    # so the chunker has natural splits to work with.
+    paragraph = ("Lorem ipsum dolor sit amet, consectetur adipiscing elit. " * 40).strip()
+    body = "\n\n".join([paragraph] * 30)  # ~70k+ chars across 30 paragraphs
+    assert len(body) > article_context.CHAR_BUDGET
+
+    with patch(
+        "fabledassistant.services.article_context.generate_completion",
+        new_callable=AsyncMock,
+        return_value="Summary of this section with specific claims preserved.",
+    ) as mock_gen:
+        out = await article_context.prepare_article_context(
+            "Long Article", "https://example.com/long", body, "test-model",
+        )
+
+    # At least one LLM call fired (the map step ran)
+    assert mock_gen.await_count >= 1
+    # Output carries the oversized-article header and section markers
+    assert "longer than the chat window" in out
+    assert "## Section 1" in out
+    # Map output is much smaller than the raw body
+    assert len(out) < len(body)
+
+
+def test_chunk_by_paragraph_respects_boundaries():
+    """Chunker splits on paragraph breaks, not mid-sentence."""
+    from fabledassistant.services.article_context import _chunk_by_paragraph, CHUNK_CHARS
+
+    paragraphs = [f"Paragraph {i}. " + ("x" * 1000) for i in range(20)]
+    body = "\n\n".join(paragraphs)
+
+    chunks = _chunk_by_paragraph(body)
+
+    # Each chunk stays under the budget
+    for chunk in chunks:
+        assert len(chunk) <= CHUNK_CHARS
+    # Total content is preserved (modulo overlap duplication, so ≥ original)
+    assert sum(len(c) for c in chunks) >= len(body) * 0.9
+
+
+def test_chunk_by_paragraph_handles_oversized_paragraph():
+    """A single paragraph larger than CHUNK_CHARS gets split mid-paragraph."""
+    from fabledassistant.services.article_context import _chunk_by_paragraph, CHUNK_CHARS
+
+    body = "x" * (CHUNK_CHARS * 3)  # one huge paragraph, no breaks
+    chunks = _chunk_by_paragraph(body)
+
+    assert len(chunks) >= 3
+    for chunk in chunks:
+        assert len(chunk) <= CHUNK_CHARS

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"