* feat(speech): spell dictionary terms in dictation, with repo terms and Whisper prompt
* test(speech): pin host repo terms for phone dictation
* fix(speech): keep symbol terms whole, accept null setup params, resolve SSH repos
* refactor(speech): key transcript words once and type phone terms at the reply schema
* refactor(speech): merge speech-types imports in the speech IPC
* feat(runtime): tell phones when the desktop refreshes a PR
* fix(runtime): name every checkout a coalesced PR read stood for
* fix(runtime): don't count a read that couldn't load checks as a checks change
transcribe-cpp is now the only on-device engine on every platform, so
the sherpa-onnx packages, their packaging (extraResources, the darwin
libonnxruntime prune, the native-variant assert and the Linux libstdc++
exemption), the sherpa recognizer in the speech process and the
per-model engine choice all go. Hotwords are written as plain terms.
* perf(github): answer phone PR lookups from the desktop's last result
* fix(github): order last PR outcomes by lookup start sequence
* fix(github): keep reused phone PR outcomes out of the refresh observers
* fix(github): reuse a fallback-found PR only for the same fallback number
* feat(speech): move the speech catalog to transcribe-cpp models
Parakeet v2/v3, Whisper tiny and SenseVoice now download their GGUF
builds, which run on transcribe-cpp everywhere, Intel Macs included (on
the CPU). A model still holding its old sherpa-onnx download shows
"Update to the faster version (<size>)"; updating replaces the old
files, and dictating before then says the model needs that update.
The five models with no GGUF build are retired and their users moved
over: the Japanese Parakeet, the Chinese-English Zipformer and the
Paraformer to SenseVoice, and the Korean and Chinese streaming
Zipformers to the new Nemotron 3.5 (28 languages). Nemotron's locale
codes (ko-KR) are matched from the ISO language setting.
* feat(speech): offer the speech model update once after the engine change
A card in the notification corner tells people whose selected on-device
model needs the new download, or whose retired model was replaced, and
starts the same download as Settings > Voice. After a replacement it
offers to delete the retired model's files.
* feat(speech): move retired Korean and Chinese models to SenseVoice and offer Whisper Large v3 Turbo
On transcribe-cpp 0.4.1 with Metal, SenseVoice beat Nemotron 3.5 on Korean
and Chinese FLEURS clips, so the retired Korean and Chinese streaming
Zipformers now move to SenseVoice. Whisper Large v3 Turbo (Q8_0, 886 MB)
joins the catalog as the most accurate choice for Japanese, Korean and
Chinese. Nemotron 3.5 stays, described by the languages it does best.
* feat(settings): add a Use GPU switch for on-device dictation
On by default (Auto: the GPU when it works, the CPU when it fails).
Turning it off keeps the transcribe-cpp speech process on the CPU, for
machines whose GPU driver freezes or restarts them. The row shows when
the selected model runs on transcribe-cpp.
* fix(speech): reload the warm speech model when Use GPU changes
A loaded recognizer kept the GPU choice it started with, so turning Use GPU
off left the next dictations on the GPU until the model unloaded.
* fix(speech): reload the speech model after a dictation when Use GPU changed during it
A recognizer still in use when Use GPU changed stayed loaded afterwards with
the old choice. It now unloads as soon as its last dictation ends, and the
warm-up loads it again with the new choice.
The pill showed only the current partial and went back to Listening after
each final, so pause-committed phrases from transcribe-cpp never appeared.
The host's transcript fold moves to src/shared and the desktop pill uses it
too: committed text in the foreground, tentative text muted.
Add an OpenAI-compatible model whose base URL, model and optional key
are set up in a Voice settings dialog and sealed together in ~/.orca,
so a key is never sent to a host it wasn't saved for. It reaches
OpenAI, Groq, Mistral and local servers, and replaces the unused
userModels setting.
* feat(speech): run GGUF models on transcribe-cpp, falling back to the CPU
The speech host child can now load a second engine. A model whose files
are GGUF runs on transcribe-cpp (Metal on Apple silicon, Vulkan on
Windows and Linux, CPU otherwise); everything else stays on sherpa-onnx,
which remains the only engine on Intel Macs.
A GPU that fails to load, errors, stalls for 25 s or crashes the process
moves the dictation to a CPU-only process (TRANSCRIBE_BACKENDS=cpu) that
replays the audio the GPU had not decoded yet, so no words are lost or
repeated. A GPU that crashed or would not load is remembered per
transcribe-cpp version and OS release, and voice.speechGpu = 'off' keeps
it on the CPU.
Packaging copies transcribe-cpp, koffi and the target's native package
into resources/node_modules outside asar, fails when a target's package
is missing, and ships the MSVC runtime beside transcribe.dll on Windows.
No catalog model is GGUF yet, so nothing runs on transcribe-cpp until
the catalog change that follows.
* fix(speech): keep the speech queue running after a failed task, and drop audio frames after a discard
* fix(speech): keep the engine's file reads off the main thread's hot paths
The engine lookup and the GPU failure marker are read once per process
instead of on every dictation start and mobile catalog request, the marker
is written asynchronously, and the child reads the hotwords file with
fs/promises. The baseline moves the existing engine presence checks from
stt-worker-paths to the shared package table and adds the once-per-process
marker reads.
The phone filters in-progress VM checkouts only when it reads its host
workspace list, so a read made before a window's hide reached the main
process kept showing the checkout until the next unrelated change. Emit
worktreesChanged for each repo whose pending checkout appears or goes
away; phones already re-read their host workspace list on that event.
* fix(vm-recipes): hide an in-progress VM checkout in every window, the phone and the jump palette
A VM create's checkout was hidden only in the sidebar of the window that
started the create. Windows now publish their pending VM checkouts to the
main process, which shares every window's set with all windows and
filters the paired phone's host workspace list and workspace context.
The jump palette and the sidebar read the shared set, and a grouped VM
project no longer draws an empty-project placeholder while its create
is pending. A window's set is dropped when it closes, crashes or reloads.
* test(project-groups): reset leaked runtime environment revisions between files
* feat(vm-recipes): show VM lifecycle states in the sidebar, with Suspend/Resume VM
* fix(vm-recipes): Suspend VM suspends an idle parent with active children; dedupe repeat suspends
- With active children, Suspend VM sleeps this workspace too, so its VM
suspends once their sessions stop.
- A repeat Suspend VM while one is pending is dropped in the coordinator
instead of trusting a possibly stale suspended record.
- After a direct suspend, re-probe the host so its header reads Suspended.
* perf(github): use the desktop's PR fallback rule on the phone and forward refreshes
* fix(github): keep a merged fallback PR while the phone has no head
* feat(runtime): tell phones when relay hosts change
* fix(runtime): keep the relay host backstop live and watch every list input
* fix(mobile): hold the relay host feed only for desktops that announce changes
* perf(runtime): watch relay host inputs only while phones read the list
A recipe VM create registers the VM's project before its environment
checks run, so the VM's primary checkout showed as a normal workspace
next to the pending row while the checks ran or were paused on a
failure. Hide that checkout in both sidebar list models until the
create adopts it (or makes its worktree). A cancelled or failed create
keeps it hidden while its VM is cleaned up; a VM that could not be
cleaned up is shown again so it can be removed.
Codex reports both approved and denied automatic approval reviews on the
guardianWarning channel. Orca rendered every one as a yellow "safety review
warning", so routine approvals looked like problems. Approved reviews now
render as a neutral "approved this action" notice on desktop and mobile;
denials and timeouts keep the warning tone.
Co-authored-by: Sean876-alt <224890246+Sean876-alt@users.noreply.github.com>
Co-authored-by: Neil <neil@stably.ai>
The runtime service and failed-start cleanup now reach orca.yaml environment
recipes and plugin VM recipes through ShellRecipeDriver's recipe-shaped steps.
Same arguments, same runner, same results: records, failure copy and
diagnostics are unchanged.
* refactor: group agent launch, status, and messaging owners
* refactor: consolidate direct owner imports
* refactor: import pane ownership directly in status listeners
One lifecycle seam (create/resume/suspend/destroy/status + capability flags and a
catalogued error type) that every environment provider implements, a shared
vitest contract any driver runs from its own test, and ShellRecipeDriver, which
wraps today's orca.yaml / plugin VM recipe runner and passes the contract.
No production callers yet.
Claude paints the caret over the first glyph of a dim composer suggestion
as an undimmed inverse cell. The typed-text read kept that glyph, so an
empty composer reported a one-character draft and the stock Try "..."
placeholder was never recognized. Treat a lone inverse caret glyph ahead
of a longer dim suggestion as untyped, while keeping a real typed last
glyph that sits before a dim argument hint (e.g. /commit [message]).
Fixes#40195
Co-authored-by: Neil <4138956+nwparker@users.noreply.github.com>
The browser-extension preload was a main-build Rollup input, so it could be
split into shared CJS chunks that a sandboxed preload cannot require. Build it
with the other guest preloads as a self-contained bundle, and share the
one-byte unminified output options between the preload build and sandboxed
bundles so the extension preload stays one-byte in V8.
Co-authored-by: Neil <neil@stably.ai>
Request the project on the shared issue node fields and map it in
mapRawIssueForWorkspace, so list/search rows carry the project the same
way issue detail already did. Detail now reuses the shared mapping.
Co-authored-by: Neil <neil@stably.ai>
The Codex config line scanner only treated exactly three quotes as the
multiline closer, so a value like '''printf 'ok'''' left the scanner
inside the string and later tables (e.g. [projects."..."]) were missed
when writing project trust. Match toml_parser: a run of 3-5 quotes
closes the string, with up to two quotes belonging to the content.
Co-authored-by: Neil <neil@stably.ai>
Claude reports deferred categories (MCP tools, system tools loaded on
demand) that sit outside the used total, but summarizeContextUsage
dropped the flag, so the popover listed them as window percentages next
to real usage and the rows summed to well over 100%.
Keep the deferred flag on each summary row and render deferred rows in
their own group under a divider, with token counts instead of a share of
the window and a note that they are not counted in the used total.
Fixes#39878
Co-authored-by: Neil <neil@stably.ai>