Skip to content

Releases

Partner v0.3.2

Choose a tag to compare

@github-actions github-actions released this 18 Nov 15:03

English | 简体中文 | 繁體中文

Partner v0.3.2

This release makes startup and quitting feel more responsive: the API becomes reachable before local voice models load, and quitting from the tray clears the icon immediately.

  • Startup: The API starts listening first, then warms enabled local TTS and KWS/ASR engines in parallel in the background. Settings, history and schedules stay available while models load. /api/health only means the API can accept requests; /api/voice/local and /api/voice report each engine and voice path as starting, ready, or failed. A live process whose health check times out is no longer reported as ready.
  • Settings and voice: Saving settings cancels a superseded warmup and applies the latest configuration. The input and output halves now have separate lifetimes, so changing or restarting one no longer tears down the other. Keyword pushes and health polling are cancellable, and the configured "startup wait" is enforced as a real readiness deadline. A failed playback device can still be retried by saving voice settings again.
  • Quit: Tray → Quit hides the main window and removes the tray icon before backend cleanup, without changing the saved tray preference. The API, MCP servers and engine child processes are still stopped as before.
  • Desktop pet: The same immediate tray-icon cleanup applies while the pet is running.

Notes

  • These changes are covered by automated API, host and smoke tests; a real Windows cold-start timing comparison and hands-on verification of tray exit and installer handoff are still outstanding.
  • Live2D Core and models remain user-supplied. See the voice, pet and release documentation.

The voice engines are downloaded from the engines-v1 release on first use; their URLs and SHA-256 are pinned in this build. The app code is MIT licensed; the engine runtimes and model weights keep their own licences, see THIRD-PARTY-NOTICES.md in the release assets.

Partner v0.3.1

Choose a tag to compare

@github-actions github-actions released this 17 Nov 08:26

English | 简体中文 | 繁體中文

Partner v0.3.1

This update improves pet motion tuning, session storage, voice playback and settings reliability, and strengthens the build and release process.

  • Desktop pet: Procedural motion, gaze and drag spring-back are applied before physics, allowing them to drive hair and other physics effects. Map parameters per model, adjust multipliers, add offset tracks and inspect diagnostics for saved settings. Window operation failures can be inspected and retried in settings.
  • Sessions: Reads include only committed events and support concurrent history deletion. Projection after append and pet feed backlog reads are optimized. If replacing the log during history deletion fails, the store attempts to reopen the original and otherwise explicitly faults.
  • Voice: Playback buffering is bounded by duration; interruption cancels the old reply and clears the buffer. After a device failure, save voice settings again to retry. Reference-audio uploads temporarily lock settings editing, saving and discarding.
  • Settings: Whole-settings updates use a revision to detect concurrent changes. Engine detection fills the form; changes take effect after saving.
  • Engineering and releases: Updated frontend development dependencies and split the settings panels, desktop shell and smoke checks. Tag validation, release permission isolation and pet-disabled build checks are stronger. Updates stop the API process tree first and restore supervision if the installer fails to start.

Notes

  • Pet diagnostics and previews use saved settings; model parameter mappings require manual confirmation. Sleeping does not currently slow breathing.
  • Settings are read at API startup; restart after editing the file manually. On a revision conflict, reload and reconcile unsaved edits.
  • Live2D Core and models remain user-supplied. See the pet, voice and release documentation.

The voice engines are downloaded from the engines-v1 release on first use; their URLs and SHA-256 are pinned in this build. The app code is MIT licensed; the engine runtimes and model weights keep their own licences, see THIRD-PARTY-NOTICES.md in the release assets.

Partner v0.3.0

Choose a tag to compare

@github-actions github-actions released this 16 Nov 11:02

English | 简体中文 | 繁體中文

Partner v0.3.0

A desktop pet, built on Live2D. A small transparent window that stays on top of everything, reads the assistant replies as they are finalised, scores them against a keyword dictionary, and plays a fixed set of scenes — greet, stretch, head-pat, nod, lip-sync, sleep. It does not talk to the model, does not change the system prompt, and does not ask the model for an emotion tag.

The pet

  • Where it lives — its own window: transparent, borderless, always on top, out of the taskbar, not resizable. Settings → Desktop pet holds the on/off switch, click-through, motion sound, scale and a "reset position" button.
  • What drives it — two trigger kinds only: local interaction (click on head/face/body, hover, drag, double click, idle timeout) and a keyword match in a finished reply. Fifteen scenes in total, each an action combination plus an expression, with a priority order so a double click interrupts an idle stretch instead of queueing behind it.
  • Click zones — a model with HitAreas uses them; a model without gets head/face/body proportions cut from its bounding box, adjustable with two sliders.
  • Actions fall back — an abstract action resolves through the model's mapping table first, then procedural parameter keyframes (e.g. ParamAngleY for a nod), then is skipped. Not every model carries every motion, so the fallback is what makes an arbitrary model usable.
  • Lip-sync without audio — TTS plays on the .NET side and the WebView never receives the stream, so the mouth is a 6–9 Hz envelope while the backend reports speaking. It looks like talking, it is not phoneme-accurate.
  • Emotion from text — a keyword dictionary with weights, longest-match-first, negation prefixes (不开心 is not 开心), and a try-out box in settings that scores a pasted sentence with the same code the pet runs. Four emotions: smile, angry, cry, shy. Default words are editable per emotion, up to 200 each, weight 1–10.

The Core and the models are yours

  • Cubism Core — download Cubism 5 SDK for Web R5 yourself from Live2D's site and import the zip under Settings → Desktop pet. The import is strict: the whole-file SHA-256 must be 67064a7f…df06ba0, and only Core/live2dcubismcore.min.js and Core/LICENSE.md are extracted, into %LOCALAPPDATA%\Partner\pet\core\. A mismatch is refused, so an old or edited Core cannot reach the page's script-src.
  • Models — import a folder or a zip; a pack with several model3.json files asks which one. Only the files that model3.json references are copied, paths that escape the model folder are refused, .js is refused, and the caps are 200 MB / 2000 files. Chinese file names in a zip without the UTF-8 flag are decoded as GBK, so older packs work. Model copyright belongs to each author; the app grants you nothing.
  • Where it goes — everything stays under %LOCALAPPDATA%\Partner\pet\. Nothing is uploaded.

Licence

  • Live2D code is not MIT. The app's own code still is. easy-live2d bundles the Live2D Cubism Framework, which is under the Live2D Open Software License — a separate agreement, not an MIT-compatible one. The Framework's copyright notice is preserved in the built pet bundle by a banner, because a minifier would otherwise drop it and the licence forbids removing it.
  • Models keep their authors' rights. No model ships with the app and none is mirrored.
  • Both licence texts ship with the source in third-party/ and are copied into the installer's notices/licenses/; THIRD-PARTY-NOTICES.md gains a Live2D section and names every new dependency (easy-live2d, Pixi.js 8, @pixi/sound 6 and their transitive packages).

Guards

  • .gitignore refuses *.moc3, **/live2dcubismcore*.js, ui/public/Core/, ui/public/Resources/ and *.model3.json, so a model or a Core cannot be committed by accident.
  • The release workflow runs git ls-files '*.moc3' '*live2dcubismcore*' '*.model3.json' before the frontend build and fails when the output is non-empty — the ignore file is the habit, the check is the guarantee.
  • doc/pet.md (with .en and .zh-TW) documents getting the Core, importing a model, the scene table, the default dictionary, the licence rules and the known limits.

Upgrading

  • Install v0.2.2 or later directly, or let the previous version update itself.
  • Nothing existing changes: settings, sessions, the memory database, downloaded engines and the desktop-control contracts are untouched. The pet's settings are a new section, empty until you open the page.
  • No new listening port and no new permission. The pet window's capability file grants only move, resize, read the cursor and toggle click-through.
  • If you had a model or a Core sitting under ui/public/, it is now ignored by git; the app reads them from %LOCALAPPDATA% instead, through the import page.

Limits

  • The mouth is procedural — there is no audio in the WebView, so lip-sync is an envelope, not phonemes.
  • Procedural fallbacks are approximations — a model without the expected parameters gets a shallower version of a scene, or none. A model with no Idle group, no physics and no HitAreas (the sample model is one) relies on the fallbacks for most scenes.
  • Multi-monitor and scaling smoke is partial — position restore checks that the stored point is on a visible screen and otherwise moves to the main display's bottom right; 125% / 150% scaling and a secondary screen with negative coordinates still need a manual pass.
  • Sound is opt-in — motion sound is off by default.
  • A model that is missing at start-up leaves the window blank and reports the reason to the settings page instead of raising a dialog.

Engineering

  • PetAssetStore imports both the Core and the models: the Core by whole-file hash, the models by parsing the zip central directory for raw entry names (the BCL mangles GBK names, verified) and copying only referenced files through a temporary directory and an atomic rename.
  • SettingsStore gains a monotonic Revision so the pet window can hot-reload the mapping, the dictionary and the timers without a restart; the feed endpoint reports speaking from VoiceHost.IsSpeaking, which only reads the state the voice pipeline already tracks.
  • The emotion matcher and the scene director are pure modules with an injected clock and random source, so ui/tests/pet-emotion.test.ts and ui/tests/pet-director.test.ts pin the tie-breaking rules, the negation handling, priority pre-emption and the once-a-day greeting without a running pet.
  • The API tests cover zip-slip, .., absolute paths, reparse points, .js, size limits, a GBK-named zip, multiple model3.json files, a Core hash mismatch and the ValidatePet boundaries.

The voice engines are downloaded from the engines-v1 release on first use; their URLs and SHA-256 are pinned in this build. The app code is MIT licensed; the engine runtimes and model weights keep their own licences, see THIRD-PARTY-NOTICES.md in the release assets.

Partner v0.2.2

Choose a tag to compare

@github-actions github-actions released this 12 Nov 12:13

English | 简体中文 | 繁體中文

Partner v0.2.2

A hardening release for desktop control. v0.2.1 fixed a bug in Memory Services and left the desktop as it was; this release is the other half — the parts of "let the model press buttons in my apps" that were true in the code but not proven on screen. Four things changed, and they change what an action reports, not what it can do:

  • the UI Automation client is a COM client on its own thread, with timeouts, instead of the managed one;
  • password fields never report their value, and system shortcuts ask on their own;
  • every action says whether it was delivered and whether retrying is safe, instead of leaving that to be read out of prose;
  • an observation says why it is incomplete, and a capture or comparison that spans a window moving is thrown away rather than judged.

The session, memory, settings and workspace formats are untouched. This is a drop-in replacement for v0.2.1, with one caveat: a consumer that parsed the first line of a desktop_act result sees a JSON object there now (see Upgrading).

The result of an action now says what happened

A model that only reads prose had to guess whether "点击完成" meant "the click landed, the screen changed" or "SendInput accepted it, nothing else". It also could not tell a refusal from a partial delivery, so it retried clicks that may already have happened.

Every desktop_act result now begins with one line:

{"code":"OK","delivery":"posted_unverified","retry":"unsafe"}
  • code — OK, STALE_REF, AMBIGUOUS_TARGET, PERM_DENIED, POLICY_DENIED, TIMEOUT, APP_UNRESPONSIVE, NOT_FOUND, INVALID_ARGS.
  • delivery — not_delivered, delivery_uncertain, posted_unverified, delivered_verified. "Posted" means Windows accepted the whole input batch, which is not the same as the application receiving it: UIPI silently drops input aimed at a higher-integrity window while SendInput still reports success.
  • retry — safe only when nothing was injected; everything that may have landed is unsafe.

The prose after that line keeps the detail, and "注入成功" became "已发出(Windows 已接收输入,投递结果以下面的校验为准)" so the promise matches the evidence. The same three fields are in the HTTP answer of the desktop endpoint.

A partial input batch releases what it pressed

SendInput can insert a prefix of a batch and fail the rest. That leaves a modifier held — Shift down, no Shift up — and the user's keyboard is broken for reasons nothing in the result explains. A failed batch now releases exactly the keys and buttons it pressed (never a modifier the user was already holding), then reads the key state back and says whether the release was confirmed.

What a keystroke is allowed to reach

Password fields never report their value

A password field is identified three ways: the UI Automation IsPassword flag, a page input with type=password, or an autocomplete that contains password (so current-password and new-password are both covered). Such an element is marked secret, its value is null, and it never serves as the control's display name — using an unlabelled password field's value as its name put the plaintext into the observation, which is the one place it must not appear. After a write into it, the state comparison only asks empty or filled.

System shortcuts ask on their own, and never get a task lease

Win+R opens a run dialog, Win+L locks the session, Alt+F4 closes the application, Ctrl+Shift+Esc opens Task Manager. These leave the window a human approved, so "continue working in this window for 10 more steps" must not widen into them. Any Win chord, Alt+F4, Alt+Tab, every Ctrl+Alt+* and Ctrl+Shift+Esc is now named by a guard, asks under every gear except No approval, and does not spend or receive the 10-step lease an ordinary in-window action does.

A keystroke goes to the named control, not to whatever had the caret

Typing into a form whose focus still sits in the previous field writes into that field, reports success, and leaves the model to discover the misplaced text later. A KeyPress or TypeText aimed at an elementId now calls SetFocus and confirms HasKeyboardFocus before injecting; if the caret is somewhere else the action is refused and the current focus is named. The window also has to be in front first, because focusing a control in a background window can raise that window — and then the keystroke lands in a window the user was not looking at.

An incomplete observation says why

A window whose walk stopped was previously reported with a single boolean. It now carries elementsTruncated with the reason — Depth, Produced, Visited, Time, Fault, Refused, Unresponsive, Gone — because they lead to different next steps, and the note the model gets is worded for the next step rather than for the cause the code guessed:

  • unresponsive → wait or let it recover, do not re-observe in a loop;
  • refused (permissions, UIPI) → screenshots and coordinates only;
  • budget exhausted → narrow with nameContains / roles / offset, do not treat the list as complete;
  • nothing but a webview surface → it may be a Chromium/Electron app, and that is now a possibility instead of an assertion about Electron.

A window that shares an identity with another control in the same window is marked ambiguous; acting on it by elementId returns AMBIGUOUS_TARGET and asks for coordinates or a narrower observation, rather than flipping a coin between two equally plausible controls.

The UI Automation client

The managed client is replaced by the COM API (CUIAutomation8) on a dedicated MTA thread: a 2-second connection timeout and a 5-second transaction timeout, so a provider that stops answering is a named refusal rather than a stalled turn. Waiting for a Chromium tree to settle happens on the caller, not the UIA queue.

A provider in another process is read through a CacheRequest in AutomationElementMode_Full; a provider in this process is walked without caching. The mode matters: the default None returned 63 of 129 nodes on an Edge page, while Full returned 127 of 127.

Measured on this machine, the COM client returns the same node counts as the managed one (Notepad 43/43, Explorer 7/7, a Chromium window 87/87, an Edge page 127/127) — so the reason for the change is the timeouts, the thread and the cache, not a claim that the managed client could not see. Single-window observe p50: 6.9ms (Explorer), 61.7ms (Chromium, at the 40-control cap), 327.0ms (Edge page).

Two more guards came with it:

  • A window that was captured is identified before and after (handle, process start time, rectangle). A frame taken while the window moved, resized, closed or was replaced is discarded, and a comparison across such a move answers Indeterminate — comparing two coordinate spaces would otherwise be reported as "the screen changed".
  • A hung window is probed with SendMessageTimeout(WM_NULL, 2s) first and reported as APP_UNRESPONSIVE without entering UI Automation at all.

Window enumeration also stopped listing windows that are not on the screen the model is looking at: cloaked windows (another virtual desktop, a suspended UWP app) and unowned tool windows are excluded, and a UWP app hosted in ApplicationFrameHost.exe is reported as the process inside the frame rather than as the host.

Upgrading

  • Install v0.2.2 directly, or let v0.2.1 update itself.
  • No data or format changes: settings, sessions, the memory database and downloaded engines are untouched.
  • If you have a tool, script or prompt that reads the first line of a desktop_act result, it now sees {"code","delivery","retry"}. The prose follows on the next line, and "已发出" replaced "注入成功".
  • System shortcuts (Win chords, Alt+F4, Alt+Tab, Ctrl+Alt+*, Ctrl+Shift+Esc) ask even under Approve risky actions, and approving one does not grant a lease.
  • Nothing listens on a new port and no new permission is requested; the debugging port keeps its existing rule (only browsers and Electron apps started by desktop_launch).

Limits

  • Measured on one 2560×1440 screen at 96 DPI (100%), virtual-desktop origin (0,0). Dual monitors, a secondary screen with negative coordinates and 125%/150% scaling are still untested; probes/windows/run.ps1 -Command monitors prints screen identity and DPI.
  • The "first Chromium read is incomplete, then settles" behaviour was not reproduced here. The code still re-reads at 250ms until the count stops growing, but a real cold start has not been proven.
  • The page-control path has still not been verified on real Electron applications.
  • No OCR: interfaces UI Automation cannot see are left to a vision model reading screenshots.

Engineering

  • DesktopSafetyTests (new, offline) covers the partial-batch release, the shortcut rules and the governance hook, the password and ambiguity markings, and the delivery classification. DesktopLiveSafetyTests (new, PARTNER_DESKTOP_LIVE=1) covers the same guarantees against a real window: a password field's value never reaching the walk, typing going to the named control rather than the one that had focus, a capture whose window moves being discarded, and a hung window being reported without waiting for UIA.
  • probes/windows/UiaClientProbe is a standalone probe (not product code) that prints one line per measurement so a row can be copied into the ledger in probes/windows/FINDINGS.md; probe-page.html and run.ps1 are its fixtures. The ledger records both confirmations and contradictions, including the two claims in the earlier notes that this machine did not reproduce.
  • The smoke's desktop_act checks matched the result body at its start; they...
Read more

Partner v0.2.1

Choose a tag to compare

@github-actions github-actions released this 11 Nov 01:16

English | 简体中文 | 繁體中文

Partner v0.2.1

A fix release for one bug in v0.2.0's Memory Services: a streaming request to a service returned the whole answer and then the connection died, so callers saw "terminated" instead of the end of the reply. It only happened with Summaries enabled, which is the only mode where the proxy parses the stream at all. Nothing else changed — the session, memory, settings and workspace formats are untouched, and this is a drop-in replacement for v0.2.0.

The fix

A stream no longer ends in a dropped connection when Summaries is on

When summaries are enabled, the service reads the upstream's SSE frames as it forwards them, to keep the text it will later summarize. It read choices[0] unconditionally.

A stream that ends with a usage frame — what OpenAI sends when the request carries stream_options: { include_usage: true }, and what Azure, vLLM, LiteLLM and several other OpenAI-compatible servers send by default — looks like this:

data: {"choices":[{"delta":{"content":"answer"}}]}

data: {"choices":[],"usage":{"prompt_tokens":812,"completion_tokens":34}}

data: [DONE]

That second frame has an empty choices array. Indexing [0] on it threw ArgumentOutOfRangeException inside the response copy loop, which is past the point where headers have been sent: the error could no longer become an HTTP status, so Kestrel dropped the connection and the caller's HTTP client reported the stream as terminated. The forwarded text itself had already arrived, which is why it looked like a truncation rather than a failure.

The frame is now read through a helper that answers null for an empty array, so the usage frame is simply nothing to capture — exactly like a [DONE] sentinel or a keep-alive comment. The same guard covers the non-streaming body parse, where a provider that answers with choices: [] had the same hole.

Pass-through is unchanged: the usage frame is still forwarded byte-for-byte to the caller. Only the summary capture skips it.

Upgrading

  • Install v0.2.1 directly, or let v0.2.0 update itself.
  • No data or format changes: the memory database stays at schema v3, and settings, sessions and downloaded engines are untouched.
  • If you were running a service with summaries on and streaming from it, this is the release that stops those callers seeing "terminated".

Engineering

  • ServiceProxyTests gained a case that streams a delta, then an empty-choices usage frame, then [DONE], and asserts the captured stream reaches [DONE] and the round is recorded (pendingRounds = 1). It fails against v0.2.0 with the same ArgumentOutOfRangeException a caller saw.

The voice engines are downloaded from the engines-v1 release on first use; their URLs and SHA-256 are pinned in this build. The app code is MIT licensed; the engine runtimes and model weights keep their own licences, see THIRD-PARTY-NOTICES.md in the release assets.

Partner v0.2.0

Choose a tag to compare

@github-actions github-actions released this 10 Nov 15:37

English | 简体中文 | 繁體中文

Partner v0.2.0

A feature release for connecting external agents and plugins. Until now, what Partner knew about you stayed inside its own conversation. Game AI mods, editor and browser plugins and other agent programs each talked to a model of their own and started from nothing. v0.2.0 adds Memory Services: you give such a program a local endpoint, and your Partner comes along. Anything that lets you enter an OpenAI- or Anthropic-compatible address, a key and a model name can use one. That covers an NPC dialogue mod, a writing plugin or a separate agent framework. Each request it sends gets the relevant memories added, and what it asked and got back is summarized into the main memory, so the main conversation can recall it later.

The memory store gains an origin column (schema v3, migrated automatically on first start). The session and settings formats are otherwise unchanged.

New

Memory Services

Add, edit and remove endpoints on the new Memory Services page in the sidebar. Each service listens on its own 127.0.0.1 port and forwards to a real model: the main model by default, or an upstream of its own. It changes only three things in a request:

  • the model name the caller uses is replaced with the upstream's;
  • recalled memories go in front of the caller's last user message, within a token budget that the context length and the request's own size bound;
  • an optional custom prompt goes before the caller's system prompt (empty by default).

Everything else passes through unchanged: tools, tool_choice, stream, response_format, and the upstream's tool_calls, SSE frames, usage and errors. The caller keeps its own tools and its own history; Partner's MCP servers and desktop control are not exposed through a service.

Summaries back into the main memory

After a configurable idle interval, or at the latest after a maximum wait under steady traffic, a service summarizes what callers asked and what it answered into one sentence. The sentence starts with [对外服务:<name>] and goes into the main memory. Stopping, deleting or "Summarize now" flushes too, and on shutdown every service flushes before exit. Summaries can be turned off per service; recall keeps working.

Memory origins

The Memory page labels service-written memories with their service and can filter by it. A service memory never retires a main-session memory, and deleting a conversation does not delete it. Deleting the service offers to remove the memories it wrote.

Separate credentials

Every service has its own generated token (at least 32 characters, shown once, rotatable), accepted as Authorization: Bearer or x-api-key. The caller's credential is never forwarded; the upstream key stays in the settings.

Limits

  • Same machine only. A service listens on 127.0.0.1, so the mod or plugin has to run on the computer Partner runs on.
  • Pass-through, not conversion. A service's protocol must match its upstream's. A mismatch is refused when saving, and a configuration that drifts later answers 502 without contacting the upstream.
  • External text reaches the main memory. Summaries are recalled by the main conversation like any other memory, so an untrusted caller could steer it through them. Give the token only to programs you trust, or turn summaries off.

Upgrading

  • Install v0.2.0 directly, or let v0.1.4 update itself.
  • The memory database migrates to schema v3 on first start; existing memories become main-session memories.
  • No service exists until you add one, so nothing listens on a new port after the upgrade.

The voice engines are downloaded from the engines-v1 release on first use; their URLs and SHA-256 are pinned in this build. The app code is MIT licensed; the engine runtimes and model weights keep their own licences, see THIRD-PARTY-NOTICES.md in the release assets.

Partner v0.1.4

Choose a tag to compare

@github-actions github-actions released this 09 Nov 17:13

English | 简体中文 | 繁體中文

Partner v0.1.4

An update release. Until now the in-app check walked the release pages only, because api.github.com allows 60 calls an hour per address and a shared egress IP regularly spent them; that path had no release notes, no installer size, and no per-release metadata to read. The check now asks the GitHub REST API first — one JSON document carries the tag, the notes, the installer's name, size and URLs, and the address of its checksum — and keeps the page-walking code as the fallback for exactly the case that made it necessary. The settings page also says which of the two answered, instead of letting a fallback look like the only path there ever was. Nothing in the session, memory or settings format changed; this is a drop-in replacement for v0.1.3.

New

The API is the primary check, the release pages are the fallback

UpdateService used to read releases/latest, follow its redirect to releases/tag/<tag>, and take the installer's name and SHA-256 out of SHA256SUMS.txt. That works without a quota, and it is still what runs when the API cannot answer — but it can only ever produce a version and an installer: the release pages carry no notes, no size and no asset list.

The primary path is now api.github.com/repos/Jaspersol/partner/releases/latest, whose release JSON supplies all of it in one call. Everything the API answers is checked before it is believed:

  • Only HTTPS URLs on GitHub hosts enter the candidate list, and a redirect that ends anywhere else fails the request — a poisoned release document cannot point the download at a host of its choosing.
  • Asset names may not contain a path separator, so a name can never escape the download directory.
  • A release with no usable installer is a failed primary, not an empty answer: the fallback can still name one through the manifest.

A fallback is reported, not hidden

UpdateCheck gained a source field (api or direct), and Settings → General shows a line under the result saying which path answered. A check that had to fall back used to look exactly like one that did not, which is how a rate-limited machine could believe the app had no other way to check.

Fallback behaviour

  • A failed API is not retried on every check. A spent quota (403 with x-ratelimit-remaining: 0), the secondary limit (429, or 403 with Retry-After), an unreachable host and a check timeout each pause the API for 15 minutes; an HTTP error or a malformed answer pauses it for 2, the same as the check cache. The cooldown is a timestamp, so a manual check inside it goes straight to the fallback instead of paying for a request that is known to fail.
  • A missing per-asset checksum falls back too. The API path wants <installer>.exe.sha256 next to the installer; a release that predates it — every release before this one — still updates, because the manifest path takes over. Verification stays mandatory on both paths: with no digest at all, no installer is offered.
  • Both paths down is reported as both. The message keeps the direct failure in front and appends the API one, so "GitHub is unreachable" and "the release has no checksum" do not read the same.

Download

The download now has candidates instead of one URL: the release's browser_download_url, its API asset URL, releases/download/<tag>/<name>, and finally releases/latest/download/<name>. A candidate that fails moves to the next; a checksum mismatch does not — trying another mirror cannot fix bytes that do not match the digest, so it deletes the partial file and stops.

Each attempt downloads from scratch rather than resuming a partial file, because a stream that broke halfway leaves a hash that would never match. The size cap, the MZ header check and the SHA-256 comparison are unchanged, and so is /UPDATE /P: the installer overwrites api/, the shell and the two engine scripts in place and leaves the downloaded engines alone.

Upgrading

  • Install v0.1.4 directly, or let v0.1.3 update itself — both paths lead to the same installer.
  • Every app release now has to publish three assets, and the workflow does: Partner_<version>_x64-setup.exe, <installer>.exe.sha256, and the SHA256SUMS.txt that lists it. The first two feed the API path, the last one the fallback.
  • The engines are still not in the installer: download them from Settings → Local engines; the URLs and SHA-256 stay pinned to the engines-v1 release.
  • The settings, session, memory and workspace formats are unchanged, so an existing install keeps everything.

Engineering

  • UpdateServiceTests grew from 8 fact cases to 16, plus three theories (15 rows). Between them they cover the API as primary, each way the API can be rate-limited, an offline API, a release with no .exe.sha256, both paths failing, the cooldown not being spent twice, a manifest line containing a path, a foreign asset URL in the release JSON, and the download's candidate order, checksum mismatch and non-executable payload.
  • Two frontend tests assert the source line: direct when the fallback answered, api when the primary did.
  • UpdateService takes an injected clock, so the 2-minute cache and the 15-minute cooldown are tested as values rather than by sleeping.

The voice engines are downloaded from the engines-v1 release on first use; their URLs and SHA-256 are pinned in this build. The app code is MIT licensed; the engine runtimes and model weights keep their own licences, see THIRD-PARTY-NOTICES.md in the release assets.

Partner v0.1.3

Choose a tag to compare

@github-actions github-actions released this 02 Nov 05:51

English | 简体中文 | 繁體中文

Partner v0.1.3

A fix release for the two things v0.1.2 got wrong around installing and updating. Its installer wrote the two engine server scripts as files named engines\tts and engines\kws-asr, which is where the engine folders have to go: downloading the engines failed on a fresh install, and installing over an install that already had them failed too. This build writes the scripts inside their folders, clears the old placeholder from both sides (an installer hook and the engine download itself), and hands the new installer /UPDATE /P so an update installs over the previous version and leaves the downloaded engines where they are. The MCP handshake also got its own timeout, because a slow-starting server was being read as a broken one.

Fixes

The installer wrote the engine scripts as files, so downloading and updating both failed

bundle.resources in map form treats the value as the full target path, not a directory. "../engines/tts/tts-server.py": "engines/tts/" therefore installed a file called engines/tts holding the script (byte-identical, 15,453 bytes) — exactly where the engine folder belongs.

  • Downloading the engines failed. Extraction creates <root>/tts, the path was a file, so the multi-gigabyte transfer died at the unzip with "a file already exists".
  • Installing over an install that had engines failed. The installer wanted to write a file where the download had left a folder, and NSIS could not. The maintenance page's default ("uninstall before installing") did not help.

The resource targets now name the file, and nsis/hooks.nsh deletes the stray file before the copy — only when it is a file, never a directory: engines\tts\*.* matches only a directory, and that directory is the downloaded engine. EnginePackage clears the same leftover before extracting, so a person still on the v0.1.2 installer gets the engines from Download instead of an error.

An update installs over the previous version

  • The installer is started with /UPDATE /P: it installs over the recorded directory instead of running the old uninstaller, and needs no answers on the maintenance and directory pages.
  • engines/ is not in the installer's manifest, so the runtimes and models already downloaded are left alone — nothing to download a second time.
  • The API sidecar is stopped first, because the installer overwrites api/* and Windows will not replace a running image.

The MCP handshake has its own timeout

  • "Longest wait" used to bound the initialize handshake as well as tool calls, covering two different things: the server answering, and the child process starting. A cold dotnet takes 300–400 ms to boot, which a short per-call timeout cannot absorb, and on a loaded machine the handshake lost that race — the server read as broken rather than slow.
  • The handshake is now bounded by the server's own startTimeoutSeconds (default 30 seconds); every other request keeps the per-call timeout.
  • The settings hint says so: this value does not affect start-up.

Upgrading

  • Install v0.1.3 directly. The placeholder v0.1.2 left behind is cleared by the installer, or by the engine download when that comes first; either way the layout repairs itself.
  • From v0.1.3 on, an in-app update no longer uninstalls first. The v0.1.2 shell does not know the new switch, so updating from it still shows the installer's maintenance page. The uninstaller only removes engine folders that are empty, so the runtimes already downloaded stay — it is the last time that page has to be clicked through.
  • The engines are still not in the installer: download them from Settings → Local engines; the URLs and SHA-256 are pinned to the engines-v1 release.

Engineering

  • A Rust test asserts the hand-off runs the installer in update mode (/UPDATE, /P), so the switches are a value a test reads rather than a string inside a spawn call.
  • An API test covers a tree in the old layout: the stray file is removed and extraction creates the folder.
  • The desktop "look again" test asserted an elapsed time against the default delay, which measures the OS timer rather than the loop. It now waits an explicit 200 ms and asserts half of that, so timer skew cannot close the margin while a loop that stopped waiting still fails.

The voice engines are downloaded from the engines-v1 release on first use; their URLs and SHA-256 are pinned in this build. The app code is MIT licensed; the engine runtimes and model weights keep their own licences, see THIRD-PARTY-NOTICES.md in the release assets.

Partner v0.1.2

Choose a tag to compare

@github-actions github-actions released this 31 Oct 14:54

English | 简体中文 | 繁體中文

Partner v0.1.2

Each model endpoint can now speak either the OpenAI or the Anthropic wire format. Voice gained spoken approval: a request that needs your sign-off is read aloud, and with speech input on you can answer "批准执行" or "拒绝执行" out loud without saying the wake word first. The runtime changed substantially underneath — .NET 10, an incremental session projection, daily memory.db backups, and a sidecar that restarts itself after a crash and keeps a log. Several fixes land here for problems that could wedge a whole conversation or record a half-written answer as complete.

New

Pick OpenAI or Anthropic per model endpoint

Settings → Model, and each sub-agent's own model settings, gained a "protocol" choice. anthropic sends POST /messages (the Claude Messages API); openai is the previous behaviour.

  • System prompts, tool_use / tool_result blocks, image attachments, SSE deltas, stop reasons and cached token usage all map onto the internal neutral model, so tool calls, the multi-step loop, the context budget and usage reporting are one implementation for both protocols.
  • "List models" and "test connection" follow the selected protocol instead of asking an Anthropic endpoint on an OpenAI path.
  • claude is accepted as an alias for anthropic; casing and stray spaces are ignored.

Spoken approval

With speech output on, a request that needs your approval is read aloud. With speech input on as well, you can answer it by voice.

  • The prompt reads the tool name and its source only — never the raw arguments, so paths and clipboard contents in them never reach the speaker.
  • While an approval is pending and the speaker is idle, the microphone keeps listening without the wake word, and you do not have to wait for the prompt to finish. "批准执行" approves, "拒绝执行" refuses.
  • With a wake word enabled, the wake-free listening is limited to those two commands: other speech during the wait does not enter the message queue. Without a wake word, other speech is treated as an ordinary message, as before.
  • When several requests are pending, only the oldest is answered — the same one the screen lists first — and the confirmation says how many remain.
  • An approval still expires after 120 seconds, and an unanswered request is still refused.

Short fragments are no longer read on their own

A new "minimum sentence length" setting (default 10 characters) joins a fragment shorter than that to what follows, measured from the previous ending: "嗯。好。……" is read as one sentence instead of three. English and Chinese punctuation follow the same rule, and a mixed sentence is not split by language.

Over the "break without punctuation after N characters" cap, the cut now prefers the last sentence ending, then a soft break, and only then the character count.

No polling while the window is in the tray

Hiding the window to the tray pauses the UI's polling (session, jobs, schedules, settings) and restores it the moment the window comes back.

Improvements

  • Appending to the session log no longer rebuilds the whole projection. Every append used to read the entire log back and re-deserialise it; now only the line just written is parsed and added to the cache. At 40,000 events, a cache hit drops from roughly 280 ms to under a millisecond.
  • Migrated to .NET 10. .NET 8 reaches end of support on 2026-11-10, and this app ships self-contained — a runtime patch only reaches you through a new release — so the move happens here. It also switches to Workstation GC: a 25,000-event session under the same load drops from about 2.4 GB of working set to about 250 MB.
  • memory.db is backed up daily. Backups go to a backups/ directory beside the database and the newest 7 are kept. A startup catch-up runs when the last backup is more than a day old, and the maintenance pass backs up before it physically removes soft-deleted rows.
  • The API sidecar restarts itself after a crash. Its output now goes to <local app data>/Partner/partner-api.log, capped at 10 MB with one partner-api.1.log rotation. An abnormal exit restarts it with exponential backoff from 500 ms up to 30 s; an intentional stop does not.
  • Sub-agent job logs keep only the newest 30. A job that is still running is never pruned.
  • The sentence-length settings moved to Voice output → Read-aloud. Both only decide how a reply is cut for reading, yet they sat in the input card next to settings that govern recognition.

Fixes

  • Tool-call arguments in the history were truncated into invalid JSON, failing every later turn of that session. The raw arguments string was cut at 400 characters to save room, which could leave an unterminated quote; strict OpenAI-compatible servers (llama.cpp, for one) parse tool-call arguments when replaying the history and rejected the whole request with HTTP 500 — and kept rejecting every turn after it. String values inside the JSON are shortened instead, and already-stored broken arguments are replaced with {} so the conversation can recover.
  • A connection that closed before the finish frame recorded a half answer as complete. Dropping between two SSE frames left the text received so far syntactically valid, so the loop exited normally and half a sentence was stored as a stop. A lost tool_calls frame did the same: the turn ended after the preamble with no call made. A finish_reason or [DONE] is now required, and a tool_calls finish with no complete call is rejected.
  • The local model sometimes announces an action and then stops without calling a tool. It says "我先查一下:" and ends with finish_reason=stop; max_tokens is not the cause, because the server reports stop, not length. Replaying the session reproduces it, and the same request with tool_choice=required yields the call. That short colon-terminated preamble is now recognised and retried once with the tool choice forced.
  • Saving settings failed outright on a machine with no sound card or no microphone. When an audio device opened but failed to start (no playback or no capture endpoint), the exception escaped PUT /api/settings as a 500 and nothing in the settings page could be saved — the whole document failed with it. A half that will not start now reports the reason instead ("启动输出设备失败:…", "启动输入设备失败:…") and the other half keeps running. A device that will not open was already handled this way; this release covers the start failure too.

Upgrading

  • The settings file migrates itself. input.maxSplitChars and input.minSentenceChars in an old file move under output with their stored values; when the new location already has a value, that value wins.
  • This build is self-contained .NET 10 — no separate runtime to install.
  • The voice engines are still not in the installer. Download them from Settings → Local engines after installing; the URLs and SHA-256 are pinned to the engines-v1 release.

Engineering

The release pipeline used to mark the unit tests continue-on-error, so a red test still shipped: v0.1.0 and v0.1.1 both went out with 5 failing API tests. Releases now run the same gate as CI on master — docs and CSS contracts, build, unit tests, the capability-baseline smoke run, frontend typecheck and contract tests, production build — and nothing publishes while any of it is red. The capability manifest holds 70 entries; a downgrade in any one of them blocks the release.


The voice engines are downloaded from the engines-v1 release on first use; their URLs and SHA-256 are pinned in this build. The app code is MIT licensed; the engine runtimes and model weights keep their own licences, see THIRD-PARTY-NOTICES.md in the release assets.

Partner v0.1.1

Choose a tag to compare

@github-actions github-actions released this 24 Oct 07:17

Installer for v0.1.1. The voice engines are downloaded from the engines-v1 release on first use; their URLs and SHA-256 are pinned in this build.

The app code is MIT licensed; the engine runtimes and model weights keep their own licences, see THIRD-PARTY-NOTICES.md in the release assets.