Repository navigation
Releases
v3.0 — Devin CLI model picker
Devin CLI account & model picker
New Devin CLI entry in every agent's model menu:
- Shows the logged-in
devinaccount (fromdevin auth status), plan, and last-known daily/weekly quota with reset times. - Model catalog straight from
devin models list— Free and Low cost groups up front (SWE-1.6 Slow included), full catalog under "All models". - Apply writes
agent.modelinto~/.config/devin/config.json— the nextdevinsession uses the picked model. No BunControl/OpenCode involvement. - Account switching is intentionally not in the app — it stays a manual
devin-resetoperation.
Fixes
- Audio: WAV writer copied the channel pointer table instead of sample data (bad WAVs, Whisper hallucinations).
- Speech analyzer reserves the locale before constructing the analyzer.
- Devin chooser window no longer crashes AppKit on Apply/close.
Generated with Devin
EliaTopBar v2.9.2
Fixes local Ollama model selection end to end.
- The local-model popup now sends provider_base_url with PUT /status.
- BunControl registers the selected model in config-home/opencode.json and restarts the Rust engine when needed.
- POST /server/restart now actually kills and respawns a live child instead of reporting a fake restart.
- Registration errors surface in the app instead of leaving a broken model selection.
Requires BunControl commit 74ab20c or later.
v2.9.1 — Local model picker (Ollama)
Local models
- New Local model (custom base URL)… entry at the top of every agent's model submenu.
- The popup auto-detects a running Ollama server (
OLLAMA_HOST→ saved URL →127.0.0.1:11434→localhost:11434), lists its models and appliesollama/<model>through the existing bridge path. - Falls back to any OpenAI-compatible
/v1/modelsendpoint when Ollama is absent. - Base URL persists in
UserDefaults(localModelBaseURL).
The engine needs a matching provider block (npm: @ai-sdk/openai-compatible,
options.baseURL, and a non-empty options.apiKey — any string, the engine
silently drops providers without a secret), then a restart: providers are built
once at startup and never reloaded.
Voice / audio
Pending follow-ups to the voice system: agent picker, command router, HUD and run popup.
Install
brew update && brew reinstall --cask vakandi/tap/eliatopbarv2.8 — Audio optimization pass
v2.8 — Audio optimization pass.
- Push-to-talk: Whisper live loop now runs only when the on-device analyzer is unavailable; single-flight, 15s tail slices, exponential backoff. No more permanent Python/MLX processes while idle.
- Wake mode: idle disarm (default 30 min) releases the mic + RAM/CPU.
- Optional resident Whisper daemon with configurable unload (default 5 min); off by default.
- Agent picker: single prompt instead of overlapping error HUD + panel; near-miss candidates first; hover extends auto-hide; photo cache.
- Cancel (×) button on the listening popup aborts capture/transcription/dispatch.
- Server-offline banner + trigger gating when the bridge is unreachable.
- Idle CPU back to ~0 after a capture; PerfTests 84/84.
v2.7 — Voice control
Push-to-talk, global hotkey, wake-word fallback, agent picker, Whisper transcription, macOS popup answers.
v2.6 — live model roster, clean model menu
The model menu now shows the live roster
The list was frozen. The control plane read catalog.stencil.so, a host that no
longer resolves, so /models kept serving a stale snapshot — 86 rows including
models upstream had deprecated (Muse, test, jev-1.13, a claude-opus backup).
Nothing appeared, nothing disappeared.
- The roster is now read from
models.dev/api.json— the same catalog the
opencode TUI and the Rust engine use — and keeps only theopencode
provider's free and non-deprecated models. That is exactly the ten the TUI lists. - Models that vanish upstream are dropped, not greyed out for a cycle.
- Every row is provider
opencode, so the menu is one clean block instead of
scattering entries under their maker. - The
NEWbadge and the deprecated strikethrough are gone. The menu reads like
the TUI.
The model you pick is the model that runs
POST /trigger now carries the selected model and variant, and the control plane
reads ui_electron/model-selections.json. Before, the choice made in the menu was
silently discarded and the run used whatever the config file said.
Run popups
- A popup opens on run start, not only when output arrives.
- The livestream transcript resets at the run boundary, so a fresh panel no longer
renders the previous run's transcript, todos and subagent streams. - Starting the Rust server now reconnects the bridge, so the agent list is no
longer empty behind a "Disconnected" menu.
Install
brew update
brew reinstall --cask vakandi/tap/eliatopbarbrew update is required: without it Homebrew resolves the previous cask from
its stale local clone and silently reinstalls the old version.
v2.5 — Rust server Start/Stop actually works
Start / Stop for the Rust server now works
Clicking ▶ Start Rust Server failed on any machine where EliaTopBar was not
launched from a terminal. The log said bun not found — run ./install.sh first,
which is misleading: bun was installed and working.
start.sh gates on command -v bun. A macOS .app is launched by launchd,
which gives it /usr/bin:/bin:/usr/sbin:/sbin — no ~/.bun/bin, no
~/.cargo/bin, no /opt/homebrew/bin. The script died at its first check.
Child processes now get a PATH built from the real install dirs, which also
un-hides jq: without it start.sh silently probed port 5680 instead of the
configured one.
Also fixed
- Start could hang forever. Output was read through a pipe, but
nohup bun server.tsinherits that pipe, so EOF never arrived and the menu
stayed on "Starting…" indefinitely. Capture now goes to a temp file with a
hard deadline, so every run ends in a verdict. - Stop left the server half-alive. It only matched the engine process, so it
killedopencode serveand left the Bun bridge holding:5656while the menu
read "Stopped". It now drives the repo'sstop.shand then frees the port. - Failures showed the wrong line. The popup displayed
proxy pool=100 entries— the last line printed, not the reason. It now
extracts the script's ownERROR:line.
New Start / Stop popup
Same panel as Clean RAM, stuck to the menu bar: spinner and a live elapsed
counter while it works, then ✅ or ❌ with the reason, auto-dismissing.
Install
Unzip EliaTopBar-v2.5.zip for the .app directly, or use the DMG. Ad-hoc
signed and not notarized, so remove the quarantine flag if macOS refuses:
xattr -dr com.apple.quarantine /Applications/EliaTopBar.app
Note: brew reinstall --cask eliatopbar does not install this build — the
cask still points at an older version. Copy the app to /Applications instead.
v2.4 — session selection + real provider grouping
Fixes
Picking an old session no longer snaps back to the running one.
This was the main complaint driving this release. Two real causes: the store tracked
the live session in a field it wrote but never read, so nothing stopped a live event
from landing on top of what you had just opened; and run_log frames carried no
session_id, so the app could not tell a delta for the run you were watching from one
for the session you had picked. Live and viewed are now separate, and only a tap in the
history list can move the viewed one.
Scrolling up no longer re-downloads the whole transcript.
"Load more" used to raise the limit and re-fetch the entire tail each time, and every
row was given a fresh UUID — so each poll rebuilt the whole list and lost your scroll
position. The bridge now pages backwards with a before cursor, the engine returns
stable message ids, and older pages are prepended without refetching what is already
on screen. Verified on a 67-message session: 13 pages, zero overlap, zero duplicates.
The engine honours limit on /session/{id}/message.
It ignored the parameter and returned the whole transcript every time, with no way to
tell the client how many rows exist. It now pages and reports the session total in
x-total-messages. Requests without query parameters return exactly what they did
before, so existing callers are unaffected.
The model menu is grouped by real provider.
139 models were all labelled opencode-zen and listed in one flat run. They now group
by actual provider — OpenCode first, then alphabetical — with a separator between each.
Stars mark only the OpenCode group, which is what they were always meant to mean.
Smaller things
- Session lists are cached for 3 seconds behind an ETag; a repeat fetch returns
304. - SSE replay is capped at 500 frames and yields to the event loop instead of blocking it.
- Log JSON is decoded off the main actor — a 200-message page no longer stalls the UI.
- Opening a session in the Run popup actually loads it. The row only closed the menu.
Verification
- EliaTopBar: 33/33, including new coverage for the live/viewed split.
- Engine: 39/39, including 7 new paging tests plus a round-trip check that walking
backwards reconstructs the transcript exactly. - Bridge: 42/42.
- Paging checked against a live 67-message session:
limit=5returns 5 (was 67),
13 pages walked, 0 overlap, 0 duplicates, conditional GET returns 304. - Model menu checked in the running app: 153 items, OpenCode first, separators intact.
Not verified
The log panel's selection and paging behaviour was not clicked through by hand — the
popover is not exposed to accessibility automation. It is covered by the automated
tests and the live API checks above, but not by a human clicking it.
v2.3
Fixes
OpenCode Rust block no longer reports "Not built" when the engine is installed elsewhere.
binaryPath only ever looked at <repo>/target/release/opencode. If the engine was
installed through install.sh or npm, it lives somewhere else entirely and the block
claimed it was missing. Lookup is now: repo build → explicit override → opencoderust /
opencode-rust on PATH → ~/.local/bin/opencoderust.
Both server folders are now configurable from the menu (🔌 Server Connection Settings):
- 🦀 Rust repo — the engine checkout
- 🦀 Rust binary — explicit binary, with Auto-detect to clear it
- 📁 Subworker repo — folder holding
start.sh/stop.sh
No more hardcoded /Users/vakandi. The /data/ → home-directory rewrite in the log
viewer, run popup and livestream assumed one account, so every path was wrong elsewhere.
Paths now resolve against the configured subworker repo, and model-selections.json
resolves from ~/EliaAI with a UserDefaults override.
The main dropdown stops widening. It printed the entire repo path; it now shows the
folder name, with the full path in the tooltip.
Install
brew update && brew upgrade --cask eliatopbarNot notarized — if macOS blocks it:
xattr -dr com.apple.quarantine /Applications/EliaTopBar.appv2.2 — fix the popup-burst memory leak
Fixes the popup-burst memory leak.
Opening and closing run popups used to grow memory forever: a burst of 30
popups cost ~21 MB, and it never came back — 300 popups reached 218 MB.
Two causes, both found by eliminating one variable at a time:
-
The popup windows were never freed. A closed, invisible, empty window that
was never even shown still existed 30 seconds later. macOS classifies that
memory as reachable, not leaked, so something held a live reference. We
ruled out our data path, the update rate, the anchor panel, the drop
animation, the autorelease pools, and every window property we set. macOS
simply will not release these windows — so the fix is to stop creating them.
Popped-up windows now go into a pool and get reused, capped at 12 for the
life of the app. -
Every popup also built a brand-new interface tree that was left behind.
That was the whole remaining ~19 MB per burst. The window now keeps one
interface and swaps what it shows.
300 popups across 10 bursts: 218 MB -> 32 MB.
Because a recycled popup is the riskiest part of this change, it is asserted
rather than assumed: every stress checkpoint checks that no popup is still
showing the previous agent, and it reports zero mismatches across bursts and
30 cycles using a different agent each time.
Also corrects the stress harness itself, which was passing a zero duration that
silently made every popup cycle a no-op.