Backchannel v0.6.4

Backchannel v0.6.4

The exports got serious. The insights workbook now opens on a summary a
stakeholder can read in a minute, the call briefing was rewritten from a plain
sheet into a document worth forwarding, and the Privacy tab stopped reporting
a gap that was never there.

Highlights

• The briefing export is a different document. It opens with what came out
  of the call - actions, opportunities, risks, what is still open - and the
  headline finding, before anything needs scrolling. Every insight keeps the
  line from the transcript that produced it, one click away, and the full
  transcript sits at the end the same way. It is still a single file with no
  fonts, scripts or images fetched from anywhere, so it renders the same
  offline, in an email client, and years from now.
• The insights workbook opens on a summary. Three sheets in front of the
  full grid: what was captured and in what proportion, who was in the room and
  what came from them, and what still needs attention. Then Action Items and
  Opportunities, shaped to be worked down rather than filtered. The eighteen
  column Insights sheet is unchanged - nothing was removed from it.
• Sidebar groups remember how you left them. Collapsing a group used to
  come undone on its own, because the sidebar re-opened every group whenever
  the group list refreshed. Your choices now survive that, and a reload. The
  chevron also reads as a control without hovering over it, and the Groups
  header gained one collapse-all.
• The Privacy tab was wrong about transcript refinement. It showed a red
  "not covered" whenever the Transcript Refiner was switched off, which looked
  like a hole in the shield and was not one - the refiner only ever reads
  tokenized text, on any model. That row now answers the privacy question, and
  says separately whether the stage is running.
• Meetings cost fewer tokens to analyse. Two changes, both invisible in
  use. Transcript lines handed to the agents carried a 36-character identifier
  per line; they now carry a short tag with one legend per prompt, which is
  about 40 percent off every transcript the agents read, and they read each one
  many times. Separately, every agent prompt now sends its instructions as
  instructions and its data as data, in an order a provider's prompt cache can
  actually reuse.

Notes for operators

The prompt reordering is applied at the formatting seam rather than to the
prompt files, deliberately. agent_configs.prompt holds a user-editable copy
of each template and that copy is what runs, so editing the checked-in
constants reaches a fresh install and nobody else. Normalizing at format time
reaches every install, including one whose prompt was edited into a worse
order, and a template already in the right order comes through byte-identical
so no install pays a one-time cache miss for the change.

Speaker tags are derived, never stored. Every database write still uses the
real identifier, so a resumed call rebuilds the identical mapping from the same
roster and nothing persisted depends on it. Model output is accepted as either
a tag or a full identifier, which keeps a customized prompt that still asks for
identifiers working, and no tag can reach text a person reads - the prompts
forbid it and a substitution pass enforces it before an insight is saved.

The Opportunity Specialist was running with no meeting context at all, because
it formatted its prompt directly instead of going through the shared path. It
read every conversation as a sales call. It is on the shared path now.

Two things worth knowing about the token work: the savings are measured on
prompt payload size, not re-measured against a live meeting's billed usage, and
the effect of prompt reordering on analysis quality has not been compared
before and after on a real session. Both are tracked.

Upgrading

An install on v0.5.3 or later updates itself in place and verifies the new
build's signature before applying it. An install on v0.5.2 or earlier will not
see this release from inside the app: download the bundle from the release
page, quit the old version from the tray, and run the new one.

Nothing to migrate. Exports regenerate on demand, so an existing session
produces the new workbook and the new briefing with no reprocessing.