whisper.cpp's encoder always runs over a full, padded 30s mel window
(1500 encoder positions) unless audio_ctx is explicitly reduced. The
~4s live-transcription streaming windows were never setting it, so
every window paid the full 30s-equivalent encode cost on one serial
worker thread -- the actual cause of the reported 30-65s lag between
speaking and a segment appearing, not "CPU is just slow".
Scales audio_ctx proportionally to the window's real sample count
instead (audio_ctx_for_window), floored and capped to whisper.cpp's
own valid range. Verified with a controlled A/B against the real
base.en-q5_1 model: ~30% faster even under heavy concurrent CPU
contention on this machine, which is masking most of the improvement
--the fix targets the encoder's O(n^2)-ish attention cost specifically
(1500 -> ~201 positions for a 4s window), so the gain on an otherwise-
idle machine should be substantially larger.
reminder_set was accepted on the model but hard-coded to 0 on insert
and never included in the UPDATE. save_action_items's return type
changes from () to Vec<ActionItem> so the caller (which needs the row
id to key reminder scheduling) doesn't need a second round-trip for
newly-inserted items. Also adds list_pending_reminders for startup
reconciliation.
Reschedules anything Windows' own toast schedule lost track of
(uninstall/reinstall, a changed AppUserModelID) — safe to run every
startup since schedule() is idempotent (cancels any existing entry for
the same action item id first).
Chosen over an app-side polling timer specifically to honor NFR-RES-1
("no polling timers running when not recording") — Windows itself
delivers the toast at due_at, and a scheduled toast is cleared by the
OS after it fires, so there's no "already fired" bookkeeping needed in
this app either. Windows-only, matching hardware::dxgi's pattern: the
real WinRT implementation lives in a #[cfg(windows)] submodule behind
plain top-level functions.
Verified against the live Windows toast API (not just compiled): a
throwaway example scheduled a reminder, confirmed
GetScheduledToastNotifications() reported it back with the correct id,
then confirmed cancel() removed it.
Extracts export_meeting_to as a shared helper (was duplicated inline
in export_meeting) so single and bulk export share identical
fetch+render logic. bulk_export_stem sanitizes+suffixes filenames with
an id prefix — meeting titles are very often duplicates ("Untitled
meeting") and would otherwise silently overwrite each other in a
batch. A meeting that fails to export is skipped rather than aborting
the whole run; also fixes export_meeting's doc comment, stale since
T8.4 added pdf/docx.
Adds StartRecordingArgs.template_id + list_note_templates command;
renames built_in_template to summary_prompt_bias to disambiguate it
from the new NoteTemplate concept (same meeting-type identifiers,
different purpose). Also fixes refresh_notes_and_notify (the
rename/merge-speaker refresh path) to route through
Store::update_notes instead of a raw std::fs::write — it was bypassing
T8.2's FTS reindexing, a gap missed there because this call site uses
meeting_dir(meeting_id) without the & that my earlier grep matched on.
NewMeeting/Meeting gain template_id; finalize_meeting now returns it
(fetched via one cheap scalar SELECT) so callers that only have
segments/speakers in memory (not a fetched Meeting) can resolve the
template without a separate round-trip.
4 built-in templates (Standup, 1:1, Sales Call, Retro), keyed by the
same meeting-type ids commands::summary_prompt_bias already uses for
LLM-prompt biasing — one identifier space shared by both, but two
independent lookups (structuring notes.md vs. biasing a summary
prompt). Section headers are prepended as an empty scaffold above the
auto-rendered speaker dialogue.
Bullet items get a literal "• " prefix rather than a real Word
numbered-list definition — this is read-only generated content, not
something the user continues typing into, so numbering-continuity
machinery isn't worth the extra API surface.
Character-count word wrap (printpdf's builtin fonts expose no glyph
metrics) and per-boldness-run line groups (a bold speaker-name prefix
gets its own line above the dialogue) — simple choices that keep the
renderer self-contained with no bundled font asset.
Fills in the ExportFormat::Pdf/Docx arms that previously errored
"lands in Phase 8". markdown_to_blocks() covers headings, paragraphs,
bullet/task lists, and bold runs — everything MarkdownNotes::to_markdown
and the notes editor's toolbar actually produce, not general Markdown.
Shared by both new renderer submodules (pdf, docx).
tags/meeting_tags existed since migration 0001, completely unused.
Adds upsert_tag/tags_for_meeting (mirroring upsert_participant's
find-or-create pattern) and real set_tags/list_tags implementations.
Extends list_meetings with tag/participant/date filters via a new
MeetingFilter struct, replacing the old query-only signature — the
participant filter matches either a speaker mapped to that participant
or an attendee of the meeting's linked calendar event, since neither
signal alone covers "who was in this meeting." Tags are populated on
Meeting/MeetingListItem/SearchHit via a per-row lookup (N+1, but
meeting counts are small enough that this is simpler than GROUP_CONCAT
string-splitting).
The initial notes.md render after finalize/resume/recovery bypassed
the Store abstraction entirely (raw std::fs::write) — routing it
through update_notes means that first render also gets FTS-indexed and
bumps updated_at, instead of only later manual edits doing either.
meeting_fts existed since migration 0001 but nothing ever wrote to it.
Adds reindex_fts (called from finalize_meeting/update_notes so the
index can't drift from what's on disk), a one-time startup backfill
for meetings finalized before this feature existed, delete_meeting
cleanup (meeting_fts is a virtual table with no FK/CASCADE support),
and a real Store::search() replacing the hardcoded stub. Query input
is phrase-quoted per token so stray FTS5 operators in free-text search
input can't throw a MATCH syntax error.
start_recording now creates the EventSink alongside the existing frame
channel and forwards Level/DeviceChanged onto a small fire-and-forget
thread that emits "recording://level" and "recording://device".
Adds AudioLevel (rms/peak per chunk, throttled to ~20Hz) and
CaptureEvent::DeviceChanged emitted through a new EventSink alongside
the existing FrameSink. On a capture read failure, reopens against
whatever is now the default render device and continues the same WAV
file if the new format is compatible; otherwise fails loudly rather
than silently corrupting the recording.
Reports LLM endpoint/local-only status and enabled sync targets so the
UI can prove local-only handling, per the shape fixed in
docs/04-api-contracts.md. Sync targets stay empty until Phase 9 lands
SyncManager; list_sync_targets' not_implemented error is swallowed
rather than failing the whole self-check.
iamdoubz