Backchannel v0.5.0
Setup stops assuming which provider you use, the call surfaces stop losing
things, and the post-call record stops showing you raw identifiers. Under it
all, the app you install is now built from the exact dependency tree that was
tested.
Highlights
- First-run setup asks instead of assuming. A fresh install no longer seeds
a cloud model behind your back. Every enabled agent starts unselected and
says so, models are grouped by Google, OpenAI, and local, and one
role-appropriate recommendation is marked in each provider you actually have
available. Connecting only an OpenAI key no longer tells you Google is
required, and a keyless local setup can find the paths that work with no
cloud account at all. - Strategic signals persist. Signals raised during a call are kept and
stay readable both live and after the call, instead of scrolling away and
being gone. Briefing context built from them is deduplicated, so the same
observation stops arriving several times wearing different words. - Documents feed the live conversation. Uploaded document summaries are
stored once and reused, so asking about a document mid-call draws on the
summary rather than re-reading the file. Under Privacy First the excerpt
path keeps that working without sending the document anywhere. - The briefing names people, not identifiers. Owners and attribution now
resolve to speaker names at read time, including in exported HTML, so a
finished call reads as a record of a conversation rather than a table of
UUIDs. The raw owner chip is gone and the rationale sits directly under the
summary it explains. - Insights say which model produced them. Enhance Insights shows the model
you asked for alongside the one that actually ran, so a fallback is visible
rather than silent. Insight type labels are legible words on badges and
section headers. The Excel export is now a single enriched file instead of a
choice between two partial ones. - Ask questions without leaving the call. A new in-call ask bar answers from
the conversation as it happens, so you can check what was said earlier without
stopping the recording or waiting for the post-call briefing. - The live insight surface was redesigned. Insights arriving during a call are
laid out to be read at a glance while you are still talking, so the list can be
scanned mid-conversation rather than studied afterwards.
Reliability
Activity chips distinguish blocked work from failed work and report a current
failure count rather than a running total. Batch readiness stops reporting a
blank or unrecognized model id as ready. Oversized briefing context items are
retained rather than dropped on the floor. The live in-call ask bar no longer
starves the transcript it depends on.
Release integrity
Sixteen known vulnerabilities were cleared from the web-facing dependency
stack, and the documentation site's dependency chain was cleared separately.
More consequentially, the four ways this project builds its frontend now agree.
Continuous integration, the Windows, macOS, and Linux desktop bundles, and the
Docker image all install the exact dependency tree recorded in the repository,
and the security audit now measures that same tree. Previously the tests ran
against one set of resolved dependencies while the bundles shipped another, so
a release could differ from anything that had been tested. Windows executables
also now carry proper product name and version metadata in their file
properties.
Upgrading
No action required. Sessions, agents, documents, and endpoint configuration
carry forward. Existing installations keep the models they have selected -- the
unselected first-run state applies only to genuinely fresh databases. Desktop
installs on v0.4.0 update in place through the in-app updater.
Included
Structural debt was paid down against a refreshed baseline, a repository test
gate and dependency and static-analysis workflows were added, and the public
screenshots were refreshed for the current interface.