The system prompt listed phantom tools (create_task, delete_task, get_note)
that don't exist, causing the model to spiral when users asked to create
tasks under a project. Replaced the stale hardcoded string with a
dynamically-built actions list matching all registered tools, and added
conditional searxng/caldav extensions.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Content-based gating (_should_think) was introduced in 87fcaa6 to cut
TTFT on simple prompts, but it has no way to tell that short prompts
like "create a task titled X" are going to trigger a tool call — and
qwen3:14b's tool-call template is unreliable at think=False, producing
intermittent silent generations where output tokens burn but nothing
parses into content or tool_calls.
Reverting to always-on thinking restores the pre-87fcaa6 reliability
of tool emission at the cost of TTFT latency on short conversational
prompts. This also lets us delete the silent-round retry loop (which
can no longer fire) along with its bookkeeping.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Retry attempts were previously conflated with the initial call,
making prompt_tokens and headroom look cumulative and useless for
diagnosing the silent-round behavior. Move start-of-attempt captures
inside the retry loop and emit attempt_start / attempt_end lines so
each attempt's numbers stand alone.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Log num_ctx, message count, prompt/output tokens, headroom, and a
silent flag per round so we can correlate silent generations against
context pressure on the dev instance.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Qwen3's tool-call tokens sometimes fail to parse into either content
or tool_calls, burning output tokens and producing empty bubbles.
Detect the signature within a round (empty content, no tool calls,
eval_count > 0) and re-run the same round once with reasoning mode
enabled, which emits more reliable output. The post-loop fallback
remains as the final catch.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Qwen3:14b sometimes burns output tokens on tool-calling attempts whose
emission doesn't parse into any field we read — eval_count > 0 but no
thinking/content/tool_calls ever stream to the caller. Generation
completes "successfully," the user sees an empty assistant bubble, and
no error is logged. Seen in conv 220 today.
Two safety rails:
- stream_chat_with_tools now tracks whether it yielded anything; when
Ollama's done frame reports eval_count > 0 with zero yields, log a
warning including the last ~5 raw frames so the next occurrence leaves
breadcrumbs for diagnosis.
- run_generation checks the same post-condition after the tool loop
exits and, if content is empty with no tool calls but output_tokens
> 0, substitutes a visible fallback message and streams it as a chunk
so the user gets something readable instead of a blank bubble.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
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>
Both the /news discuss button and the briefing discuss button now call a
shared seed_article_discussion() helper that stages the synthetic
read_article tool exchange and the conversational seed prompt — behavior
stays byte-identical across entry points. /news also auto-starts
generation so the chat screen lands on an in-flight stream.
First assistant reply in a seeded article conversation is persisted as a
Note (tags: article-summary + article topics) and backlinked via
rss_items.discussion_note_id, so the knowledge base stops being amnesiac
about articles the user has engaged with.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The Discuss button on news cards was producing one-shot replies because
the model got the whole trafilatura blob dropped into history with a
canned "summarize and discuss this article" prompt — no length guard, no
prep, no invitation to converse. Large articles got silently truncated by
Ollama; small articles got a tepid reply.
This reworks discuss_article around a three-layer cache:
context_prepared → content_full → fresh trafilatura fetch
First click on a small article fetches once, writes through to both
caches, and passes the body straight into the synthetic read_article
tool-result. First click on a large article additionally runs a parallel
map step (services/article_context.py) that chunks the body on paragraph
boundaries, summarizes each ~8k chunk to ~300 words of dense factual
prose via the background model, and concatenates the summaries under
section headers — all pinned to num_ctx=16384 so the map step doesn't
itself fall victim to silent truncation. Repeat clicks on either path
skip straight to the chat turn.
The canned summary prompt is replaced with a conversational seed that
invites the user into an actual discussion rather than a one-shot
synopsis, matching the goal of "have a conversation about an article,
not just read it."
discuss_topic is intentionally left untouched — it's the multi-article
aggregation path and needs a separate rework. Follow-up task will decide
whether to retire it or rework it on the cached-context approach.
Closes task #106.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
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>
_generate_title was receiving the full messages list from build_context,
which prepends RAG snippets, RSS excerpts, URL content, and briefing
article dumps INTO the user-role message string. The role=="user" filter
inside _generate_title then handed that composite blob (capped at 300
chars) to gemma3:4b as "the user's message", so the background model
was titling conversations based on article excerpts instead of what the
user actually typed — producing wildly wrong titles like "Briefing
Profile Preferences & Schedule" for a plain calendar query. See #109.
Pass the raw history + user_content + assistant reply instead.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Two related bugs where the server defaulted naive datetimes to UTC instead
of the configured user timezone, causing all-day events to land on the
previous day and briefings to "disappear" at UTC midnight.
- New services/tz.py helpers: get_user_tz, user_today, user_briefing_date
(the briefing day flips at 4am local to align with the compilation slot,
so the 00:00-04:00 local window still shows yesterday's briefing until
the new one is generated).
- calendar create/list/update tools now parse naive datetimes in the
user's TZ before converting to UTC for storage, and tool descriptions
tell the model to pass plain local dates.
- briefing_conversations.get_or_create_today_conversation and the
reset-today route use user_briefing_date so the in-progress briefing
doesn't get replaced at 19:00 NY / UTC midnight.
- _run_profile_closeout targets user-local "yesterday" for consistency.
Regression tests added for the TZ helpers and the calendar tool.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Adds 38 parametrized tests for the _should_think classifier covering the
explicit-override path, empty/whitespace content, short/medium/long length
boundaries, case-insensitive keyword matching, and a chatty-message negative
set. These pin the content-based semantics so future tweaks to the keyword
list or length thresholds surface regressions immediately instead of going
unnoticed behind subtle latency changes.
Also drops the `think=True` overrides from the briefing /discuss-article
and /discuss-topic entry points. With `"discuss"` added to _THINK_KEYWORDS,
those canned prompts trip the classifier naturally, so the overrides were
redundant — keeping a uniform "classifier is authoritative" rule makes the
code easier to reason about.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The frontend hardcoded think=true on every chat send (ChatPanel full +
widget variants, KnowledgeView minichat), which defeated the _should_think
gate on the backend and made qwen3:14b spend 5-20s on chain-of-thought
reasoning for every turn — even "hi". This was the root cause of the
warm-path TTFT variance tracked in followup_ttft_variance.md: the logged
ttft_ms was really prefill + full thinking phase, bouncing with the depth
of the model's reasoning, not with cache or eviction.
All three frontend callers now pass think=false and let _should_think be
authoritative. The classifier is now a real content-based gate: explicit
think_requested=True still forces on as an override (briefing discuss
actions, future UI toggles, MCP callers), otherwise messages <80 chars
without reasoning keywords skip thinking, messages >=400 chars or
containing keywords like why/explain/analyze/debug/review/etc. get it.
Generation timing now separately records think_requested, the final
think decision, first_token_ms (first any chunk), and thinking_ms
(duration of the thinking phase). ttft_ms keeps its existing semantic
(first content token) so existing log analysis still works. The timing
log line surfaces all four fields so the old "just a big ttft number"
ambiguity is gone.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Two small hardening fixes from the mistral-nemo testing round:
1. stream_chat / stream_chat_with_tools now read the Ollama response
body and log it before raising on non-2xx. Previously all we saw
was 'HTTP 400 Bad Request' — the gemma3-no-tools failure would
have been diagnosed in one step if we'd been logging the body,
which says e.g. 'model does not support tools'.
2. backfill_project_summaries() now also targets summaries stamped
before 2026-04-12 (the gemma3:4b cutover). The remaining projects
still carrying the broken qwen2.5:3b output (token repetition,
hallucinated topics) will regenerate on next startup on the
better model.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Ollama's /api/tags returns whatever casing was used at pull time
(e.g. 'gemma3:12B' if the user ran 'ollama pull gemma3:12B'), but
/api/chat rejects mixed-case tags with a 400. The two code paths
are inconsistent, which surfaces the capitalized tag in the model
dropdown and then silently kills every chat request against it.
Lowercase on read (get_installed_models), on settings write
(update_settings_route), and on ensure_model() input so a legacy
mixed-case user setting can't trigger a spurious re-pull at
startup. The dropdown and stored settings are now always in the
form Ollama will actually accept.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
qwen2.5:3b produced broken auto-summaries (misspellings, token repetition,
hallucinated topics) — its synthesis ceiling is too low for free-form
summarization. Gemma 3 4B is stronger on summarization at similar size
and still fits comfortably alongside the main chat model in VRAM, so it
preserves the KV-cache-separation strategy that keeps chat TTFT fast.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Project auto-summaries were using the 3B background model, but the
task — synthesizing a coherent paragraph over ~10 notes — is well past
what 3B can do reliably. Evidence on dev: "doging conversation
hygiene", "MCPview). MCP).", trailing stray quotes, and hallucinated
topics ("AI regulation").
Route through the user's default chat model instead. Project summary
regeneration is rare (only when a project changes) so the KV cache
eviction cost on the main model is negligible.
Title generation, tag suggestions, and RSS classification continue to
use the background model — those tasks are within what 3B handles.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Three related fixes uncovered while benchmarking qwen3:14b against 8b:
- pick_num_ctx was only counting message content, missing the ~15K
tokens of tool schemas. num_ctx=8192 was being selected while actual
prompt_tokens hit 14K+, causing silent prompt truncation on every
tool-using request. Now includes json.dumps(tools) in the estimate.
KV cache priming in app.py and routes/settings.py also fetches tools
so the primed num_ctx matches what real chat requests will use.
- _should_think's heuristic classifier was overriding explicit
think=true requests from the frontend toggle and MCP, gating on
message length and regex patterns. Now a pass-through — the caller
is the source of truth. quick_capture hardcodes think=False since
it's a fast classification path that was relying on the old gating.
- delete_note description only mentioned "note or task", so the model
refused to call it for entries created by save_person / save_place /
create_list. Description now explicitly lists all five note_types it
handles.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Merge create_task into create_note (set status='todo' for tasks, omit
for notes), merge delete_task into delete_note, consolidate entity
tools (create/update_person → save_person, create/update_place →
save_place), rename get_note → read_note with clearer descriptions,
move calculate out of rag.py into utility.py, and extract shared
duplicate detection into check_duplicate() helper.
Updates all downstream references in generation_task.py, quick_capture.py,
ToolCallCard.vue, and WorkspaceView.vue.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Some exceptions (e.g. connection errors) produce empty str(e),
resulting in "Research failed: " with no explanation. Fall back to
the exception class name when the message is blank.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Research pipeline now produces an index note with:
- Executive summary (2-3 paragraphs synthesized from sections)
- Clickable links to each section note (/notes/{id})
- Section notes have parent_id pointing to the index
Also improves outline resilience: lowered minimum sections from 3
to 2, retries once on failure before falling back to monolithic note.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Split 2566-line tools.py into a tools/ package with @tool decorator
registration. Each tool's schema, metadata, and implementation live
together. Briefing eligibility is now a briefing=True flag instead of
a separate frozenset allowlist. Conditional inclusion (CalDAV, SearXNG)
uses requires= metadata. Public API (get_tools_for_user, execute_tool)
unchanged.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Three briefing quality fixes surfaced by reading today's 2026-04-11
compilation output:
- **Stale weather**: get_weather was returning 48h-old cache data
after a missed scheduler run. Tool now auto-refreshes any cached
location older than 6h (fetching fresh data from Open-Meteo), and
stamps each location with cache_age_hours + is_stale so the model
can hedge instead of faithfully relaying old numbers.
- **Cancelled tasks leaking into prose**: briefing loop now defaults
list_tasks calls to status=["todo","in_progress"] when the model
doesn't specify, so cancelled/done items stop showing up in the
summary. Localized to the briefing path — chat still sees full
history.
- **Overdue in-progress tasks missed by midday check-in**: tightened
the check-in prompt to explicitly require two list_tasks calls —
one for in_progress (catches items dragging past their due date)
and one filtered by due date — so long-running tasks stop getting
silently dropped.
The compilation prompt mentioned "news themes" but didn't name the
tool, and the model was never calling get_rss_items. Result: today's
briefing had zero news coverage despite the tool being wired up and
in the allowlist.
- Explicitly list the tools to call in the compilation prompt so
get_rss_items gets invoked alongside list_tasks/list_events/get_weather.
- When the model calls get_rss_items during a compilation run,
intercept and return the already-scored/filtered items (topic prefs
+ reaction-weighted) instead of the raw feed dump execute_tool
would return. Aligns the model's view of news with the sidebar's
rss_item_ids metadata.
Deletes ~760 lines of legacy briefing code: format_task, compute_task_hash,
upsert_task_snapshots, _gather_internal, _gather_weekly_review,
_llm_synthesise, and the unified prompt helpers. run_compilation and
run_slot_injection are now agentic-tool-use-loop only.
briefing_scheduler and user_profile migrated from the deleted helper to
services.llm.generate_completion (retry + keep_alive baked in).
routes/briefing.manual_trigger now persists agentic tool-call receipts
via _persist_agentic_messages (previously silently dropped them) and
adds POST /api/briefing/reset-today to wipe today's briefing messages.
BREAKING: briefing_mode setting no longer honored; no legacy fallback.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
list_notes gains exclude_paused_projects; list_tasks tool sets it when
no explicit project filter is given. Paused projects no longer leak
tasks into briefings or list_tasks results.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Applies the grounding discipline from the agentic briefing work to the
main chat system prompt. The regular chat pipeline was already agentic
(it uses stream_chat_with_tools), but its system prompt never told the
model "only assert facts from tool results" or "if a tool returns
nothing, say so honestly." That left room for the same class of
hallucinations the briefings had — calling list_events, getting an
empty array, and then confidently mentioning a meeting anyway.
Adds two new static rules to the tool guidance block in llm.build_context:
GROUNDING — when the user asks about their own data, call the relevant
tool to see what exists. Never assert from memory or assumption.
HONESTY WHEN EMPTY — if a tool returns empty results, tell the user
plainly. No fabricated example items, no invented meetings, no generic
suggestions dressed up as real data.
Both rules are in the static (KV-cache-stable) portion of the system
prompt so they cost nothing on repeated requests for the same user.
Carries the hallucination fix from the briefing work directly into
every chat turn, not just chat that happens inside a briefing thread.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
PR 1.5 + PR 2 of the agentic briefing rollout. Extends the agentic
path to cover the morning/midday/afternoon check-in slots in addition
to the 4am compilation, and persists the full tool-call sequence
into the briefing conversation so chat follow-ups see the receipts.
run_slot_injection now honors the briefing_mode setting the same way
run_compilation does: agentic mode routes through run_agentic_briefing
with a check-in system prompt variant, legacy falls back automatically
if the new path returns empty. Signature changed from str to
(str, dict) to surface the agentic message list to the caller.
_agentic_system_prompt now takes the user's local-day window and
timezone, pre-computed by run_agentic_briefing, and embeds them in
the prompt. This eliminates a whole class of "wrong day" bugs that
would otherwise happen when the model tried to translate "today" into
ISO 8601 ranges without knowing the user's timezone.
briefing_scheduler._run_slot_for_user now calls _persist_agentic_messages
after each slot, which walks the agentic message list and stores
every intermediate assistant turn (with its tool calls and results
folded into the flat storage format the existing chat loader expects)
as a real message row. The synthetic "[Midday briefing update]"
user-role messages are no longer created — final prose is written
with metadata.briefing_slot so the UI can still identify it as a
scheduled entry. The agentic user-trigger ("Generate my morning
briefing…") is deliberately skipped so it doesn't recreate the same
fake-user-message problem we're trying to remove.
briefing_conversations.post_message now accepts tool_calls, matching
the schema the Message model already supports. This lets scheduled
briefings write structured tool-use history without reaching into
the model layer.
Net effect with briefing_mode="agentic" on:
- All four slots are grounded in tool results, no more hallucinated
events or tasks
- Chat follow-ups in the briefing conversation see morning's
list_events → [] receipt (and everything else), so "what meeting?"
gets an honest "nothing on the calendar" reply grounded in data
- No more fake [Briefing update] user messages in the chat scroll
Still to come (PR 3): UI polish — tool-call status row, inline cards
for list_events/list_tasks results, visual treatment for slot messages.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
services/events.py::list_events treated any event with end_dt IS NULL
as perpetually matching, so a point-event created last week would be
returned as "happening today" whenever the briefing or list_events
tool queried today's window. That manifested as briefings claiming
past events were scheduled for the current day.
Split the null-end branch: point events now only match when start_dt
falls within [date_from, date_to]. Events with an end_dt continue to
use standard overlap logic.
This bug affected both the legacy briefing synthesis and the agentic
list_events tool path, since both call the same events service.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
First cut of the agentic briefing redesign. Morning compilation can now
route through a tool-call loop that grounds every factual claim in an
actual tool result, eliminating the hallucinated meetings, tasks, and
news items the legacy one-shot path was producing. Behind a per-user
`briefing_mode` setting (default "legacy"); falls back to the legacy
path automatically if the new path returns empty (e.g. model too weak
to drive tool calls reliably).
New: services/briefing_tools.py — explicit read-only allowlist of 10
tools (tasks, events, weather, rss, projects, notes). New tools added
to tools.py must be opted in by name. Excludes all mutating tools and
external search tools (search_images, search_web, research_topic) which
are neither useful nor safe for a scheduled background job.
New: briefing_pipeline.run_agentic_briefing — wraps the existing
stream_chat_with_tools loop with slot-specific system prompts that tell
the model to only assert facts from tool results and to be honest when
tools return nothing. Max 8 rounds, per-round exception handling,
returns the full message list so tool-call receipts can be persisted
alongside the prose in a later PR.
Sentence-count floors bumped: compilation 6–10 (was 4–8), check-ins
3–5 (was 2–3). Weekly review 5–8.
Design doc: docs/2026-04-10-agentic-briefing-design.md
Out of scope for this PR (future PRs): slot-injection migration,
persisting tool-call receipts into the conversation so chat follow-ups
see them, UI polish for tool-call status, sidecar storage for
briefings. See the design doc's migration path for details.
Enable on an account with:
UPDATE settings SET value='agentic'
WHERE user_id=<id> AND key='briefing_mode';
or insert the row if missing.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replace the hardcoded "2h" keep_alive everywhere with a helper that
returns OLLAMA_KEEP_ALIVE_MAIN (default 30m) for the interactive model
and OLLAMA_KEEP_ALIVE_BACKGROUND (default 10m) for the background
model. Lets the main model release VRAM during long idle periods
while keeping it warm enough for bursty chat use, and stops the
sporadic background model from camping VRAM it rarely needs.
Seven call sites updated to route through llm.keep_alive_for(model):
the streaming helpers, generate_completion, the two startup warmers,
the settings KV-cache primer, and the chat warmer endpoint.
Override via env vars: OLLAMA_KEEP_ALIVE_MAIN, OLLAMA_KEEP_ALIVE_BACKGROUND.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replace indigo (#6366f1) with deep violet (#7c3aed) gradient header,
violet-tinted card border and background, refined border radius and
spacing, and violet-accented footer to match the app's design language.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Exposes birthday, organization, address (person) and website, category (place)
from entity_meta in the knowledge API response; updates KnowledgeView cards
to display organization and category as visible meta chips/text.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
bvandeusen
Bryan Van Deusen