deepwiki.com/badge.svg now sits behind a Vercel bot checkpoint, so
GitHub's image proxy gets an HTML challenge page and renders a broken
image.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Lead with the desktop downloads per platform and present the curl
installer as the headless (Linux server) option.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* feat(voice): add local Parakeet v3 capture and transcription backend
* feat(ui): add opt-in dictation settings and composer controls
* build(macos): package microphone permissions and voice attribution
* docs: describe desktop dictation behavior and runtime limits
* fix(ui): clarify dictation states and cancellation controls
* fix(ui): keep composer model controls right aligned
* feat(ui): give Voice a dedicated settings page
* fix(voice): support all native microphone sample formats
* fix(voice): retain model cleanup after interrupted downloads
* fix(ui): preserve dictation while pasted references resolve
* fix(ui): cancel dictation with Escape across composer controls
* fix(ui): refresh dictation state after empty undo and redo
* fix(ui): preserve dictated words on blank recognition results
* fix(voice): clarify retries while native inference finishes
* test(ui): verify repeated Send preserves unfinished dictation
* style(ui): format settings modules
* feat(voice): add microphone selection and streamline Voice settings
* feat(ui): visualize live dictation with a waveform and recording clock
* fix(ui): preserve waveform paint opacity
* feat(ui): morph composer controls into dictation track
* test(voice): exercise real inference on every desktop release target
* fix(voice): keep device choices fresh and cancellation accessible
* ci(voice): verify Linux release packages on x86 and ARM
* ci(voice): keep model and package probes out of permanent CI
* fix(voice): detect missing ALSA before activating Linux installs
* fix(voice): capture audio while Parakeet loads
Run capture independently of serialized model loading and inference. Retain stopped audio until the model is ready, close capture before reporting load failure, and hold admission through native work and stream teardown.
Cover loading, cancellation, device startup, rapid Stop, admission, and recording limits with deterministic coordinator tests.
* test(ui): verify and measure nonblocking dictation startup
Cover selected-draft preservation, delayed final results, single Send delivery, and late results after loading failures or finalization timeouts.
Add transcript-free debug timing from dictation activation to the Listening state and clarify asynchronous capture cancellation.
* docs(voice): explain concurrent capture and startup timing
Document audio retention during loading, cancellation and model-removal limits, race coverage, and the native startup timing procedure.
* fix(ui): stop Shell saves from reverting settings chosen elsewhere
The Shell republished its whole cached UiSettings on every save
(navigation, pane resize) and only refreshed a hand-kept list of fields
first. Anything missing from that list was silently reverted: leaving
Settings → Voice turned dictation off, and the microphone choice,
compact transcript, fit-content code fences, and diff split/wrap toggles
had the same bug.
Replace the list with a three-way merge against the store snapshot the
Shell last synced: fields the Shell changed win, every other field keeps
the store's value. The merge destructures UiSettings exhaustively, so a
new field fails to compile until it is merged too.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* feat(voice): hold to talk, default Cmd+D, shortcut on the Voice page
Dictation is now hold to talk. Holding the microphone (pointer, or
Enter/Space while it is focused) or the shortcut records; releasing
transcribes. A pointer release outside the button still counts. For the
shortcut, letting go of its key or any of its modifiers ends the hold,
because macOS does not deliver key-up for Command chords. A release
within 300 ms explains hold to talk instead of transcribing a blip, and a
release that only answered the permission prompt records nothing.
Assistive-technology Click still starts and then finishes, since it
cannot hold.
The default shortcut is Cmd/Ctrl+D (unbound elsewhere). Its rebinding row
moves from Settings → Shortcuts to Settings → Voice through a reusable
ShortcutField that shares the Shortcuts page's chip and conflict rules.
Settings → Shortcuts re-reads the stored keymap while not recording, so
its own commits cannot revert a binding changed on the Voice page.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(voice): harden capture, worker and download against edge cases
- Keep the capture flags outside the sample lock. The capture loop polled
them under the mutex every 10 ms, so the real-time callback's try_lock
could fail and drop a buffer of speech.
- Catch a native panic on the worker. A panicking job killed the only
worker thread, leaving every later session "unavailable" until restart.
- Keep loaded weights when a session fails for a microphone or recording
reason; dropping them made the next attempt reload 670 MB.
- Time out a download request whose response headers never arrive.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(ui): dictation review fixes for focus, shortcuts and Windows
- Enter/Space hold on the focused microphone no longer moves focus to the
editor, which sent its key-up (and auto-repeat spaces, or Enter-to-send)
to the draft and left the microphone live.
- A key release caused by focus leaving no longer pulls focus back, and an
unparseable stored shortcut falls back to the default when matching its
release, as apply_keymap already does.
- Another source can take over a hold whose release went missing, so the
microphone or shortcut can always end it.
- Dictation cannot start while a queued row is acquiring its edit lease,
which would replace the draft and silently drop the transcript.
- With dictation off (the default) ⌘D propagates to any other binding of
the chord instead of being swallowed in the composer.
- The Voice page's shortcut row is always shown, conflicts with a feature
page's shortcut name that page, and recording stops if the window closes
so Appshots' capture shortcut is never left suspended.
- Show Settings → Voice on Windows; the PR ships the Windows backend and
only Appshots is macOS/Linux-only.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(ui): give the composer microphone its siblings' hover
At rest the microphone snapped to a different surface colour on hover;
it now uses the paperclip's eased ink wash. While live it dims like Send,
and its focus ring and tooltip match the neighbouring composer buttons.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Wing <contact@winglee.io>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* Linux installers: add a launcher entry and icon
The curl installer extracted the release tarball but never ran the
tarball's install.sh, so no .desktop entry or icon was ever written and
Zeron did not show up in application launchers.
The tarball's own install.sh did write an entry, but copied the shipped
one verbatim: `Exec=zeron` and `TryExec=zeron` need ~/.local/bin on the
desktop session's PATH, which is often missing (bare Wayland/fuzzel/rofi
setups), and TryExec then hides the entry entirely. Both installers now
render the entry with absolute paths through ~/.zeron/app/current (stable
across in-app updates), reference the icon by path because the only
artwork is 1024x1024 (not a size the hicolor theme indexes), honor
XDG_DATA_HOME, quote Exec per the Desktop Entry spec, write atomically,
and refresh the desktop/icon caches only when those tools exist.
StartupNotify is added to the template; StartupWMClass=zeron already
matches the app_id the UI sets.
scripts/test-linux-desktop-entry.sh runs both installers offline against
a fake release under a throwaway HOME.
* Linux installers: keep the download cleanup, don't create an icon cache
install_desktop_entry reused `tmp` for its staging file. The curl
installer runs under sh, which has no `local`, so this overwrote the
download dir its EXIT trap removes and left every downloaded tarball
behind in $TMPDIR. Use `entry_tmp` in both copies.
gtk-update-icon-cache -f created ~/.local/share/icons/hicolor/
icon-theme.cache when none existed. Nothing else maintains a user-level
cache, and a stale one hides icons other apps later install there; the
entry references the icon by path anyway. Only refresh an existing one.
The test now gives each run its own TMPDIR and requires it empty, fails
if an icon cache appears, and checks the two copies of the function are
identical. A small workflow runs it on changes to the installers.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Wing <contact@winglee.io>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Update detection was late and fragile: the only checker lived in the engine
on a 6h monotonic timer (a laptop's sleep stretched it by the whole night),
the sidebar compared the release against whichever engine the UI attached to
(a daemon of another version hid or faked updates), nothing downloaded until
clicked, and several install paths could not update at all.
- zeron-update: hourly wall-clock schedule with failure backoff (1m→30m),
re-read on window activation; desktop (report-only) checker role; staged
payloads are unpacked beside their final name and must answer
`--version` with the promised release before they can replace anything;
up-front blockers (App Translocation, disk image, unwritable install dir)
instead of a failed download; Linux managed installs update from the app;
a service daemon restarts into a newer installed binary once idle.
- UI: app-level controller with its own checker against this binary; a found
release downloads in the background, the strip offers restart, and a
staged update installs on quit. "Check for Updates…" under About in the
macOS app menu, "Check for updates" in the account menu elsewhere, with a
live result dialog.
- Linux: the tarball's install.sh uses the self-updating ~/.zeron/app layout.
- Windows: per-user Inno Setup installer (LocalAppData, Start menu, zeron://,
uninstall entry kept current after in-app updates), built in release and
CI, with a silent install/uninstall smoke test.
- Fix: closing the window on X11 panicked in the geometry save, skipping the
engine drain and any quit hooks.
Co-authored-by: wing-anara <wing@anara.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* feat: add antigravity agy harness with grouped reasoning models
drive google's agy cli over its own stream-json print mode: one process
per session with prompts written to stdin, step updates normalized into
text, tool and context-usage events, and the terminal result frame as
the turn end. steers queue for the next turn, interrupts signal the
process, and resumes reopen the conversation with --conversation.
agy lists every effort as its own model id, so variants are grouped into
one model with a low/medium/high ladder and the picked level is folded
back into the variant id at run time.
headless agy cannot prompt for permission, so sessions run in
accept-edits mode unless auto-approve is on, and denied actions surface
as an error explaining how to allow them.
wires the harness through the proto id, engine registry, account slugs,
desktop pickers and settings, and the ios catalog and brand mark.
* fix: always skip agy permissions and report context windows
zeron has no approval ui and headless agy auto-denies anything it would
prompt for, so runs always pass --dangerously-skip-permissions, matching
the claude and codex harnesses. denied actions can now only come from
agy's own deny rules, and the error says so instead of pointing at a
missing auto-approve toggle.
agy reports token counts but no context window, so context usage now
carries the published window for the gemini, claude and gpt-oss model
families, leaving it unreported when agy picks its own default model.
* fix: count cached agy tokens and fit the context window tooltip
once prompt caching kicks in agy moves the cached prefix out of
input_tokens into cache_read_tokens, so the context indicator dropped
and climbed between steps. context usage now reports their sum, which
grows steadily across a run.
the context window tooltip was fixed at 260px regardless of its text, so
it now sizes to its content up to that width and groups token counts by
thousands.
* feat: list antigravity skills as slash commands
zeron showed no slash commands for antigravity. commands now come from
agy's own `-p=/skills` json listing, which answers without a model turn,
covering built-in and global skills. names are deduped in listing order
and descriptions trimmed to their first sentence for the picker, and a
non-empty listing is cached like the model catalog.
agy's cli commands such as /usage and /model are refused under
--input-format stream-json, so they are not offered. the listing runs
from home because the command list is not scoped to a session, so
workspace skills are not listed but still expand when typed.
* feat: drive antigravity over its acp server with sign-in on enable
replace the agy stream-json driver with google's antigravity acp server
(agy_acp_server 1.1.1) on the shared acp harness. the pinned release
archive for the platform downloads and unpacks once into the managed
adapters dir, reusing the npm installer's atomic marker and lock.
antigravity is now off by default. turning it on in settings runs the
server's google sign-in from home, outside any project, so preview
discovery never consumes the one-shot oauth redirect; the app opens the
single sign-in tab from the polled url and the toggle flips only once
sign-in succeeds. turning it off signs out through acp logout, and a run
without a login points to settings instead of opening a browser.
effort variants group by their display name, since a signed-in catalog
pairs gemini-3.1-pro-low with gemini-pro-agent for high, and each level
sends the id the server advertises. skills from the global gemini skill
folders are listed as slash commands next to the server's own, with
logout hidden. the context indicator is hidden until a window is
reported, and its tooltip no longer clips wrapped text.
* fix: address antigravity review feedback
* feat: add phased sign-in feedback for antigravity setup
* fix: read antigravity auth settings the way the server does
resolves the three auth-settings findings from the pr review: expand a
home-relative GEMINI_HOME like the server's os.path.expanduser, parse
settings.json as hjson since the server reads it with hjson.load, and
canonicalize the vertex-ai alias to the advertised agent-platform id.
* test: pin antigravity's live and fallback model catalogues
covers the review's catalogue recommendation: a newer flash the signed-in
account advertises reaches the picker as its own grouped row, while the
offline fallback stays on the models the pinned 1.1.1 archive bundles.
* feat: persist window geometry and restore its display
* fix: persist screen-sized macOS window geometry
* fix: align Antigravity auth settings paths with its server
Resolve Cargo.toml (keep main's newer Zui/gpui-base pins), crates/ui/Cargo.toml
(keep both the windows-render-fixture and appshots-fixture examples), and
Cargo.lock (main's lockfile plus this branch's windows-sys/embed-resource
dependency entries).
Resolves conflicts with the opencode native driver (#202): the
SubagentObserver enum keeps only Devin + Grok, and devin_spec drops
the retired http_sidecar field. Refreshes the fallback model rows to
the ids CLI 3000.6.14 advertises (swe-1-7-medium is the server default).
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Translate README.md to README.zh-CN.md and cross-link both files
with a language switcher line.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Registers Cognition's Devin CLI as a harness through the shared
AcpHarness spec: native ACP server, turn-boundary steering, no effort
ladder (effort rides the advertised model ids), and the `bypass` mode
value added to the generic no-prompts preference list. Brand mark,
settings row, picker icon, iOS catalog entries, and docs included.
Live-verified against devin CLI 3000.4.25: model discovery (177 wire
models) and a real prompt turn settled cleanly.
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
- binary/CLI `comet` -> `zeron`; crates comet-* -> zeron-*; COMET_* -> ZERON_*
- data dir ~/.comet-native -> ~/.zeron, with a one-shot rename migration
- services: comet-native.service -> zeron.service; launchd sh.zeron.comet ->
sh.zeron.app; macOS bundle id dev.comet.native -> sh.zeron.app
- release artifacts comet-<ver>-* -> zeron-<ver>-*; drop the Comet.app
tarball-path compat hack in package-macos.sh
- iOS: Comet -> Zeron targets/dirs, bundle id sh.zeron.Zeron, zeron:// scheme
- clients default to https://edge.zeron.sh; installer + landing on zeron.sh
- kept on purpose: repo name, worker names (comet-native-edge/comet-landing;
DO storage is bound to the script name), R2 bucket names, and legacy
comet.zeron.sh routes so 0.1.x clients keep working
zeron-<ver>-* artifacts don't exist until the next release, so curl-install
and landing downloads 404 between this deploy and the 0.2.0 tag.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Wire cursor-agent as a Comet harness with ask_question / create_plan /
update_todos extension methods and the parameterized model picker.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Add Hermes and Pi agents over ACP
Two new AcpAgentSpec entries on the shared harness: Hermes Agent (Nous
Research, native `hermes acp` server) and the pi coding agent (pi.dev,
community pi-acp adapter pinned 0.0.33 with npx fallback). Registry slots
mirror the specs (turn-boundary steering; pi's minimal->max thinking
ladder, no effort surface on Hermes yet), brand marks land in the picker
rail/tabs/sidebar/settings, and COMET_HARNESS gains grok/hermes/pi arms
(grok was a latent gap). Wire coverage rides the existing table-driven
fake-ACP tests plus descriptor and login-shell-resolution mirrors.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* ACP as source of truth for models; separate Traits dropdown
models() now runs a short-lived probe (the discover_commands pattern) and
reads the session/new response's first-class SessionModelState, falling
back to the model config option's choices — paseo's preference order. The
static catalogs survive as enrichment for matched ids and as the answer
when an agent advertises nothing (per-model effort ladders don't exist on
the ACP wire; discovered models inherit the probe session's thought_level
ladder unless a curated entry supplies one). Successful discovery caches
per harness instance; failures retry on the next picker open.
The traits picker un-merges from the harness/model popover into its own
composer chip + dropdown (t3code TraitsPicker arrangement): traits are
properties of the selected model, so they follow the pick instead of
sharing the browse surface. The chip hides entirely when a model has
neither a ladder nor options. Media re-captured: discovered Hermes list
(provider-tagged entries), the traits dropdown, both runs.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* Traits dropdown: t3code menu rows with Default badges
Replace the segmented-chip grid with t3code's TraitsPicker arrangement:
headed sections of menu rows (reasoning ladder + each model option),
hairline separators between sections, a Default badge on the section's
default choice, and the trailing check on the selected row — all reusing
the existing popover menu primitives. Menu stays open for multi-adjust.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* Drop selected checkmarks; right-align the Default badge
Selected menu rows across the app (models, branches, spaces, traits,
account switcher) now carry the active wash alone — the trailing check
duplicated what the highlight already said. The traits Default badge
moves to the row end.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* Ghost Default badge: bare muted text, no border or fill
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: wing-anara <wing@anara.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
The terminal viewport was an experiment and is no longer maintained.
comet_proto::{view,motion} stay in proto: pure, viewport-independent
derivations with their own test suite remain the right structure.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Everything else (layout, build, TUI, smoke tests, packaging) lives in
ARCHITECTURE.md and docs/; the front page is for someone deciding whether
to install it.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* comet-tui: a detachable Ratatui viewport over the same engine
A second frontend beside the gpui app, on the same typed RPC. The one
architectural difference: it never embeds an engine. A desktop window and
the work it drives share a lifetime; a terminal does not — it closes, the
SSH session drops, the lid shuts. So `comet-tui` attaches to a daemon and
starts a detached one (setsid, stdio to the daemon log) if none is up.
Quitting is therefore *detaching*: runs continue, docs keep syncing, the
DeviceRoom stays joined, and reattaching is instant.
Information architecture is comet-native's, read from the desktop shell
rather than from the reference screenshots of the original Electron app:
two sidebar sections (Spaces, then a flat global attention-ordered
Sessions list) and the selected space's own sessions as the tab strip,
which is also the header. Visual language follows herdr — no boxes, one
divider, rules that tee into it, label-left/affordance-right, a lot of
air. Colours are comet's exact oklch, pinned by a test against the
desktop theme.
New sessions work as the desktop does: `n` opens a draft, nothing is
created, and the first send materializes worktree -> createChat ->
QueueCommand as one ordered command (each step depends on the last, so a
failure leaves only a notice). Checkout is comet's two modes resolving
into three outcomes, shared via `comet_proto::view::checkout_plan`.
Performance, measured rather than asserted: the loop has no tick and
sleeps until input, an engine frame, or a deadline it armed; transcript
layout is cached per message and invalidated by content fingerprint, so a
streaming token re-wraps one message rather than the conversation; the
visible window is found by binary search over prefix sums and rendered by
reference. Idle cost is 0.00s CPU over 5s, checked by the pty smoke test.
Shared with the gpui viewport rather than duplicated, so the two surfaces
cannot drift: sort orders, staleness gating and the boot gate
(`proto::view`), tool-group summaries and chip labels, the loader curves
and the status-dot hues (`proto::motion`). comet-ui now delegates to all
of them.
Tests: 138 in the crate (unit, render-against-TestBackend, and click
resolution against real coordinates), plus scripts/tui-smoke.py, which
drives the real binary through a pty and verifies the engine survives the
viewport exiting.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* Desktop app serves its embedded engine, so comet-tui just attaches
The headed app only bound the IPC port when it *wasn't* running the engine —
the embedded case kept it to itself. Attaching a terminal viewport therefore
meant quitting the app, starting `comet headless`, and starting the app again
in that order.
Now the embed path serves the same RPC service on the port (the deferred one,
so an early viewport gets AuthRpc and gates itself instead of hanging). Binding
is best-effort: losing it costs peers, not this window. Shutdown stops
accepting before draining, so nobody connects into closing stores.
Fell out of testing it: connect_ws had no timeout, so any process holding 27654
that never completed a WebSocket handshake wedged boot forever. Bounded at 5s,
and the app now embeds rather than failing when the thing on the port isn't an
engine.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* README: the tui run command, and that a running app is a valid attach target
`cargo run -p comet-tui` picks the library crate and fails with an error that
does not hint at the fix.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* Ship cometui: bundled with comet, plus a standalone installer
The terminal viewport now reaches users two ways, both honoring the same
attach-or-spawn model: attach to a running engine, else spawn `comet headless`.
- Bundled: package-linux.sh and package-macos.sh build `--bin comet-tui`
alongside `comet` and drop `cometui` into the tarball / .app. The main
install.sh links it onto PATH next to `comet`. Because they ship together,
cometui always finds the engine binary as a sibling when it needs to spawn
one (resolve_comet_bin checks there first) — so "spawn its own" always works.
- Standalone: a new `curl … /install-tui.sh | sh` for a machine that already
runs comet and just wants the CLI. cometui has no gpui dependency, so unlike
the desktop installer this one works on Linux AND macOS. New `tui` release
job builds it for both, independent of the flaky macOS gpui build.
Edge serves /install-tui.sh (route + worker handler + wrangler carve-out).
cometui is the on-PATH command name; the bin target stays comet-tui.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* Ship the TUI as `comet tui`, not a separate cometui command
Replaces the two-installer cometui approach: the viewport is now a subcommand
of the one binary. `comet tui` attaches to a running engine or spawns one —
and because it lives in the `comet` binary, the engine it spawns is always
right there. No second binary to bundle, no standalone installer, no name that
collides with the desktop app.
The run logic (arg parsing, env resolution, file tracing, current-thread
runtime) moves into `comet_tui::cli` so the shipped `comet tui` and the
`comet-tui` dev binary share one entry point and can't drift. `comet`'s main
skips its stdout tracing for the tui path — a log line mid-frame would corrupt
the alternate screen.
Reverts the cometui packaging/installer/edge-route/release-job added in the
previous commit; the `comet` package already carries the viewport now.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
---------
Co-authored-by: wing-anara <wing@anara.com>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
The core of comet login/logout/status and comet daemon landed (swept into
45d160b): sign-in persists session.json and exits, headless off-TTY fails
fast with "run comet login first", and login/logout hold the data-dir lock
so they never rotate the single-use refresh token under a running engine.
This finishes the surface: comet daemon manages the SAME
comet-native.service the curl|sh installer writes (not a competing unit),
installed binaries point ExecStart at the app/current symlink so upgrades
aren't pinned to one version, both units get restart limits that actually
trip on the signed-out fail-fast exit (on-failure, 5x5s inside a 60s
window), and the installer + README + PARITY now document the
login-then-start flow.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- Edge Worker live on edge.comet.zeron.sh (custom domain) with real WorkOS
client id; comet.zeron.sh/install.sh + /releases/* routed here so the old
curl-install URL now installs comet-native (landing page untouched).
- /install.sh serves the native installer (edge/src/install.sh, Text rule):
~/.comet-native/app + ~/.local/bin/comet + comet-native systemd user unit;
/releases/* streams artifacts from the comet-native-releases R2 bucket.
- Zero-config headless: production edge URL + WorkOS client id baked into the
binary (COMET_EDGE_URL / COMET_WORKOS_CLIENT_ID override; a COMET_EDGE_TOKEN
dev bearer or empty client id keeps dev-mode auth for local/smoke).
- Fixed headed sign-in loopback port (27641, COMET_CALLBACK_PORT override) so
the redirect URI is exactly registrable with WorkOS.
- CI: deploy.yml autodeploys the edge Worker + Durable Objects on main pushes
touching edge/; release.yml builds Linux x86_64/aarch64 tarballs +
a macOS dmg (scripts/package-macos.sh) on v* tags, publishes a GitHub
Release, and uploads to R2 + bumps latest.txt.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Streaming: FollowMode::Tail + per-splice scroll_to_end hard-snapped growth
(one 56px jump/commit). Replaced with an owned pin + per-frame StickSpring
(mugen/use-stick-to-bottom constants: damping .7 stiffness .05 mass 1.25,
growth-EMA feed-forward, catch-up, teleport guard); splice-glue anchors the
offset so layout holds and the spring glides. Frame evidence: 2,13,13,10,6,
4,2,2,1,0 px/frame ramp vs a single 56px step. Doc->render path verified
O(changed) already.
Traffic lights: intended macOS inset was dead (.pl overwritten by .px).
TitlebarOptions (transparent, lights at 14,15, title None); animated 200ms
spacer 88px<->12px on fullscreen (per window-controls.tsx); all titlebar
strips are drag regions (start_window_move) w/ double-click zoom.
Pickers: double-highlight fixed (selected wash vs keyboard cursor wash),
search-input inset, w-72 widths, section separators, traits max-height,
skeleton sizes — all per harness-model-picker.tsx. Model description line
deferred (proto lacks the field).
Ports: comet-native defaults moved to 27654 (IPC) / 27640 (edge dev) so it
never conflicts with the original comet (26634/26638/26640/26654) running
on the same machine. 192 UI tests green.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Engine emits proto AuthState shape; LocalDevice RPC served; accounts param
aliases. scripts/e2e-smoke.sh + rpc examples (e2e_driver, rpc_probe): edge
under wrangler dev, two headless engines, chat created on A, run queued from
B through the doc command queue, host-executed on A, transcript + status
synced back to B — ALL STEPS PASS LIVE. dist/ packaging (Linux .desktop +
icon + package-linux.sh, macOS Info.plist template + docs), release profile
(thin LTO, strip). Clippy/fmt clean; docs/PARITY.md honest feature ledger;
README + ARCHITECTURE status refresh. Workspace: 311 tests green.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>