- Kotlin 93.3%
- Python 5%
- C++ 0.7%
- Java 0.6%
- Shell 0.3%
- Other 0.1%
| Filename | Latest commit message | Latest commit date |
|---|---|---|
|
All checks were successful
Build Android APK / apk (push) Successful in 9m21s
Reviewed-on: #23 |
||
| .forgejo/workflows | ||
| .idea | ||
| app | ||
| docs | ||
| gradle | ||
| models | ||
| scripts | ||
| .gitignore | ||
| AGENTS.md | ||
| build.gradle.kts | ||
| gradle.properties | ||
| gradlew | ||
| gradlew.bat | ||
| LICENSE | ||
| LICENSES.md | ||
| README.md | ||
| settings.gradle.kts | ||
| THIRD_PARTY.md | ||
| VALIDATION.md | ||
Wise Wolf · Holo
An offline Android assistant with a wise wolf identity.
Wise Wolf is the app; Holo is its assistant and wake-word name. Speak through a compact overlay, edit the transcript, and manage everyday tasks using local speech recognition and Android integrations. Light-brown ears, warm accents, and light/dark themes carry through the UI.
Status: working English-language prototype · Android: 12+ · Target API: 37
Features · Getting started · Wake words · Modules · Custom models · Permissions · Builds and releases · Documentation
Speech recognition, command interpretation and voice output run on the device. The APK has INTERNET permission to support optional model/catalog downloads and explicitly enabled Brave search. Only approved search queries go to Brave; model inference stays local. Apps opened by Holo may use their own network connections.
Features
| Capability | What Holo can do | Try saying |
|---|---|---|
| Timers | Create named timers, list them, read remaining time, cancel, or add a minute | “Start a timer for three minutes for tea” |
| Reminders | Save tasks with an optional day/time; list, complete, delete and snooze | “Remind me to buy milk tomorrow at 5 pm” |
| Calls | Resolve contact names or phone numbers and confirm before placing a call | “Call Mom” |
| Messages | Read notifications, confirm replies, and open new-text drafts or optionally send SMS directly | “Text Mom saying I'll be there at 7!” |
| Calendar | Read synced calendars and save timed/all-day events after confirmation | “Set a calendar event that I have a meeting at 1PM today” |
| Alarms | Open native Clock alarm drafts, repeat rules and the alarm list | “Wake me up at seven am on weekdays” |
| Music | Choose a default app, request songs, and save YouTube Music playlist links by spoken name | “Play chill playlist” |
| Media controls | Pause, skip or go back through active media sessions | “Pause Emby” / “Next track” |
| Time utilities | Read the device clock, convert durations exactly and calculate date/time offsets | “Convert three minutes to seconds” / “What time is it?” |
| Unit conversion | Convert length, mass, volume, temperature and fixed durations locally | “Convert 20 kilograms to pounds” / “What is 12 feet in meters?” |
| Conversation | Chat with the selected local model using in-memory history and no device actions | Switch to Conversation |
| Dictation | Append editable text and read it aloud without executing commands | Switch to Dictation |
| Speech models | Select recognition models and voices independently | Open Speech models and voices |
Ask “What can you help with?” for the capability list. Complete requests run after speech ends; missing or ambiguous information is routed according to the selected module's policy, using local field questions where supported, otherwise to the configured LLM or local help. Calls and outgoing messages always require confirmation. Calendar creation confirms the title, schedule and calendar before saving. New texts default to Google Messages drafts. Enable Send new texts directly by SMS in settings for confirmed carrier SMS without opening another app; this needs SMS permission and a default SMS SIM. Direct mode does not send RCS/MMS. Drafts still require sending in Messages.
Contact lookup includes saved Android Nickname fields and relationships on your own Me/profile contact. “Mom”, “Mum” and “Mother” match the same saved relationship. Set those in your Contacts app; Optional Wise Wolf aliases and per-app conversation shortcuts are available in Apps & integrations. Android Contacts matches take priority; a phone contact does not establish a Discord account or Matrix room. Calls/new texts and contact skills share this lookup; message filters/replies can use the resolved name when it matches a current notification. Multiple matching numbers still need clarification. Spoken call/text confirmations use the saved contact name and phone label, while the screen retains the exact target number. Unnamed numbers are spoken digit by digit with pauses, using 3–3–4 grouping for US numbers. Duplicate names/labels can use the last four digits to distinguish choices. Message contents are preserved exactly.
Music requests retain a named app at either end, including “YouTube Music play Queen” and “next song on YouTube Music”. Playback controls check the player's advertised capabilities. Saved-playlist and player selections can clarify without an LLM; an unnamed playlist request still needs interpretation and cannot silently resume music.
Reported limitation: YouTube Music song requests can find the requested song without starting playback. The 1.2.5 development path prefers public search-to-play transport, then a song-focused playback intent. Actual song playback still requires owner verification.
The assistant experience
- Compact overlay: opens above Home or the current app through Android's assistant button.
- Immediate capture: records while recognition loads, buffering up to 12 seconds in memory.
- Editable requests: use the keyboard button to stop capture/playback and type a correction. Focused drafts preserve fast typing and cursor position; the composer stays above the keyboard. Send clears the submitted draft immediately and keeps the request in history; dictation retains its editable document.
- Screen-awake conversation: wake activation keeps the display on during the visible conversation, including dictation and follow-ups. Dismissal releases the display flag.
- Readable transcripts: all-uppercase STT output becomes lowercase. Typed text and a custom model's mixed capitalization/punctuation are preserved; automatic punctuation is not added.
- Assistant turns: manual launch keeps the microphone off and stays open. Speak handles one request; voice clarification/confirmation can continue that cycle. Typed requests stay quiet. Summoned assistant sessions retain voice-first follow-ups and inactivity dismissal. Typing and explicit mic pause suspend capture. Wake idle defaults to 10 seconds; the initial wake capture has an independent 30-second absolute limit, then discards unfinished audio.
- Task memory: display history is separate from execution context. Completed requests do not supply instructions/results to the next task. Pending answers retain their exact chain until completion, cancellation or expiry. Timers and reminders persist independently.
- Appearance: choose Use device theme, Light, or Dark in Holo settings → Appearance. Both UIs share the brown ears and the same lightweight paired-perk listening animation.
Tap outside the overlay, press Back, or use its close button to dismiss it. Dismissal stops capture and cancels pending outgoing confirmations. Audio and raw transcripts are not saved to disk.
Modules and request processing
A Module is a task category (Timer, Reminder, Call, Messages, Calendar, Alarm, Music or Help).
A Module Function is an action within it, such as timer.start, timer.stop or timer.list.
Each module owns its hotwords, action keyword aliases, parameter schemas, execution policy,
confirmation requirements and manually written LLM skills.
Deterministic processing selects the first hotword match in the request, then the
first keyword match for that module. Module/action declaration order does not matter.
Matching ignores case and punctuation, respects whole words, and chooses the longest phrase
when matches start together. Payload matches remain context. Explicit independent tasks can form a bounded sequential plan; ambiguous payload/command boundaries require interpretation before any action. Explicit
parameter labels take precedence: timer start 5 minutes name is stop timer 4 starts one
five-minute timer named stop timer 4.
Speech requests and model tools share a validated execution interface:
FunctionCall("timer", "start", mapOf("duration" to 300, "name" to "Cooking")).
Unknown functions/parameters and invalid values cannot run. Calls/texts require confirmation
regardless of their source. Providers are discovered at startup from independently contributed
.provider descriptors; invalid/conflicting modules are reported while valid ones load.
Assistant settings → Advanced → Modules lists registered Modules and Module Functions and edits hotwords/keyword
aliases with conflict validation. Preferences are stored separately from module code.
Assistant settings → Local LLM enables optional MNN framework inference and downloads
Qwen3.5 2B as the recommended model, with experimental 0.8B as a resource fallback. No weights are bundled.
MiniCPM5 1B and 2B arrived in 1.2.3 / 10203 under Experimental models.
The frozen 1.2.5 / 10205 candidate remains at ec9f31d. The frozen 1.2.6 / 10206 candidate remains at 0f62713 with its failed gates intact.
The 1.2.7 / 10207 candidate has finished this iteration with open release gates.
MiniCPM5 2B passes both automated model matrices; 1B still fails compatibility, recall and
teardown, and Qwen has compatibility/workflow failures plus an interrupted host check.
Final general instrumentation has two failures. Builds, 240 JVM tests and optimized startup
pass; optimized runtime streaming, arm64 runtime, owner acceptance and licensing remain open.
Qwen is retained for compatibility. 1.2.8 is an efficiency/maintenance candidate; acceptance gates remain open.
MiniCPM remains experimental until automated and owner-phone acceptance; no installed
selection or preset changes automatically. See the 1.2.7 candidate record.
MiniCPM may use concise family-specific skill instructions; schemas and host checks stay shared.
Use is enabled only after checksum, native and structured tool-call checks pass. Experimental
models can still fail assistant requests; no family is selected automatically after a failure.
The host rejects invented timer durations and retains exact pending inputs across clarification.
All four models default to CPU, thinking off, 4,096 context tokens, four threads, 512 output
tokens and greedy sampling. The main Allow model thinking switch is saved per model and mirrors the Advanced setting.
Changing it cancels active model work. Other Advanced settings remain separate for each installation.
1.2.6 candidate: 224 JVM tests, 22 preparation/template checks and debug/release/lint
builds pass. Three wake-handoff failures remain after the bounded repair; owner acceptance
is pending. Neither 2B model passes the current full assistant/Conversation gates. The
frozen 1.2.6 release record and
1B audit retain exact model results and open gates.
The frozen 1.2.5 candidate record remains historical evidence.
Physical-device performance remains unverified. Local inference defaults off.
Conversation replaces the top Reminders shortcut; reminders remain in settings and the
assistant module. Chat uses the selected model’s published template and thinking setting,
with no skills or tools. It keeps at most eight complete turns in RAM across tabs, rotation
and backgrounding. New conversation, activity finish or process death clears it. A conservative
context budget trims oldest complete pairs and rejects oversized current messages. Changing
surfaces, models or thinking cancels generation while retaining completed chat turns. Typed
chat is silent; Speak grants one request and honors Speak responses. Tool-looking output
remains text. Explicit Assistant web search is part of the 1.2.7 candidate; OCR is deferred beyond 1.2.8.
Eight task modules plus shared time, units and explicit search modules provide manual skills, validated schemas and examples. Core tools are available across tasks; calendar work can use time tools without asking for the current time. Unit conversion uses host arithmetic, marks rounded answers as approximate and asks whether gallons are US or Imperial. These utilities also work without a model. MNN runs in a private native process; Android actions keep the existing host validation and confirmation. The upstream Android chat app and MNNCLI are references, not embedded applications. Runtime, model installation and remaining device checks.
Each activation starts deterministically. Missing timer/call/message fields ask locally. Unmatched/ambiguous requests can hand off to the model, which then handles the rest of that activation; override forces that handoff. Without a working model, Holo offers useful local help and examples. Loaded weights can remain warm for two idle minutes, while every activation and dismissal clears conversation/inference state. Preloading defaults off for new settings; an explicitly saved choice is preserved. Optional preloading runs only when inference is enabled and memory/battery state permit it; deterministic commands never wait for speculative loading. Background wake listening never preloads.
Compatible approvals combine into one host question. A correction invalidates the old approval and reconfirms exact actions. Plans run sequentially and stop after failure/uncertain completion. Direct SMS can continue to a call; opening a message draft defers remaining actions and ends that workflow. Deferred actions never carry into a later activation. Message/reminder bodies and quoted content remain payloads; conjunctions alone do not authorize more actions.
Advanced settings offers per-model overrides with Restore Recommended Defaults, a searchable
upstream catalog and direct Hugging Face repositories. Installation uses immutable revisions,
sizes/hashes, resumable staging, isolated native loading and atomic activation. Unsupported or
failed downloads preserve working installations. Catalog membership is not compatibility.
INTERNET permits model/catalog downloads; inference stays local and no prompts, contacts,
messages or recordings are uploaded. Model choices and files remain excluded from backup.
Native builds require Android NDK 28.2.13676358, CMake 3.31.6 and python3 scripts/prepare_mnn.py.
Supported ABIs are arm64-v8a and x86_64 on Android 12+. CPU is the implemented execution path;
physical-device performance and acceleration require separate validation.
To add a module, implement AssistantModule with its Module Functions and skills, then add
one independent descriptor under app/src/main/functionModules/. The
module authoring guide explains discovery, execution, session lifetime and
the engine-independent boundary, with a compiled Notes example
that includes parameters, confirmation and read-only retrieval.
Getting started
Build requirements
| Tool | Project configuration |
|---|---|
| Android Studio | Open this repository root and run the app module |
| Java | JDK 25 |
| Gradle / Android Gradle Plugin | Included wrapper 9.6.0 / AGP 9.4.1 |
| Android SDK | API 37, package platforms;android-37.0, Build Tools 36.0.0 |
| Python | Python 3, standard library only for model preparation |
| Device | Android 12+ phone or emulator; minimum SDK 31 |
CI uses command-line tools 23.0, build 16111833. Dependency versions are recorded in
the version catalog.
Prepare and build
python3 scripts/prepare_speech.py
python3 scripts/prepare_mnn.py
./gradlew :app:assembleDebug
Speech preparation downloads about 325 MB of pinned dependencies, verifies SHA-256 checksums, and reuses verified cached files. Speech models are bundled. MNN preparation stages pinned framework source for the native build; LLM weights are downloaded separately in the app. A missing-asset build error points back to the preparation script.
The debug APK is written to:
app/build/outputs/apk/debug/app-debug.apk
First launch
- Allow microphone access. Use Assistant → Speak for requests or Dictation for text capture and read-aloud. Recognition loads on startup; the voice loads when first needed.
- Enable timer alerts. Open Setup → Manage timer alerts for Android's Alarms & reminders access and notification permission. On Android 14+, also enable Full-screen alerts.
- Choose Holo as your phone assistant. Open Setup → Use Holo as phone assistant, then select Wise Wolf in Android's digital assistant settings. Some phones also need an assistant gesture enabled.
- Choose integrations. In Assistant settings, select your music app and grant the contacts, calendar or notification access needed for the features you use.
- Adjust voice and appearance. Speak responses applies to the app and overlay. Appearance, custom models, and optional wake-word training are in the same settings screen.
Setup shows only missing microphone, timer-alert, phone-assistant and full-screen access. Each completed step disappears; finishing the last step closes the dialog and hides Setup. Access is checked again when you return from Android settings, and Setup reappears if access is later removed. Optional integrations and ongoing preferences stay in Assistant settings; its Android app settings link opens Android's permission/notification controls.
Reminders is available at the top beside Assistant and Dictation. See the usage guide for timer selection, conversation behavior and integration details.
Assistant and wake words
The assistant button needs no background wake listener. Once selected as the phone's assistant, Holo opens the overlay and starts capture when microphone access is available.
Wake listening is off by default. To enable it:
- Open Assistant settings → Wake word and training.
- Select Holo as the phone assistant and allow microphone/notification access.
- Enable listening; use Allow background listening if you want screen-off activation.
- Close Holo/settings and try “Hey Holo” or “OK Holo” followed by a request.
The wake model uses phonemes, including explicit pronunciations for Holo, rather than expecting the recognizer to spell its name. Retrain after this pronunciation update: old sensitivity calibration is reset while phrase names are retained. Detection remains software/CPU based. Training checks all phrases together, matching live detection. Holo supports hoe-low with both O vowels stressed. Calibration tries trigger thresholds and, when necessary, keyword search strength, rejecting settings that trigger on the unrelated examples. One private recording now passes native Android replay; reliability across speakers, noise and distances still needs hardware checks.
You can configure up to four English phrases, each two to five words from the wake dictionary. Unknown custom words are reported when adding a phrase. Training uses three positive examples per phrase and two unrelated sentences to calibrate sensitivity. Wait for the listening ears before speaking; each recording allows five seconds. If calibration fails, Holo identifies and retries the affected example while keeping the others. Recordings stay in memory; saved profiles contain phrase/settings metadata. This is calibration, not speaker authentication or neural-weight retraining.
Background wake detection currently uses the microphone and CPU, a foreground service, and a partial wake lock. DSP/root-backed detection and DaVoice are research options; neither is an active backend. Idle battery use and screen-off reliability need phone testing.
Wake listening pauses while Holo or its settings are open. Stop it from settings or its persistent notification. A bounded audio handoff preserves the command immediately after the wake phrase; the detected wake prefix is removed from the request. Battery exemption does not reduce inference cost.
Wake behavior and limitations · SoundTrigger, DSP and DaVoice investigation
Custom speech models
Open Assistant settings → Speech models and voices to choose recognition and voice packs independently. The bundled English defaults remain available.
| Purpose | Supported format | Required files |
|---|---|---|
| Speech to text | sherpa-onnx streaming transducer: Zipformer, Conformer or LSTM; 16 kHz / 80-dimensional features | Matching encoder, decoder, joiner ONNX files and tokens |
| Text to speech | sherpa-onnx VITS/Piper exports | Voice ONNX, tokens, eSpeak NG data or lexicon/dictionary, and any required text rules |
Find models: use the official streaming transducer catalog for recognition or official voice samples for VITS/Piper voices. Download model packs rather than APKs.
Import: extract a downloaded .tar.bz2 and choose Import folder, or import a ZIP.
Keep all companion files and notices. Select the imported name to validate and activate it;
multi-speaker voices also expose Speaker ID. You can cancel imports and remove unused packs
through Manage imported models after switching away from a selected pack.
Phone storage is under the app's Android/data/com.blusnake.projectwisewolf/files/models/
folder, with separate stt/ and tts/ directories. Settings shows/copies full paths;
use the importer when Android restricts file-manager access. Repository staging folders are
models/stt and models/tts.
Generic ONNX, Whisper/GGUF, offline/CTC/Paraformer recognizers, raw Piper weights without sherpa companion files, and non-VITS voices cannot be selected through this importer. Other languages can support dictation/read-aloud; commands, replies and the wake model remain English.
Imports are validated in an isolated native process. Changed/missing packs fall back to bundled models, and runtime upgrades require revalidation. Imported models are not backed up and are removed on uninstall, so keep the originals. See formats, sources and troubleshooting.
Permissions
Grant access for the features you use through Setup and Assistant settings.
| Access | Used for |
|---|---|
| Microphone | Voice requests, optional wake listening and training |
| Notifications | Timer/reminder alerts and the wake-service Stop action |
| Alarms & reminders special access | Android scheduling for timers and timed reminders |
| Full-screen alerts, Android 14+ | Lock-screen timer activity; Android controls presentation |
| Default digital assistant | System-button overlay and optional background wake activation |
| Optional battery exemption | Software wake availability during Android sleep |
| Contacts and phone | Name/number lookup and confirmed calls |
| Optional SMS sending | Send carrier SMS after explicit confirmation when direct SMS is enabled |
| Optional Call Log | Read-only recent-call context skills; not needed to place calls |
| Calendar read/write access | Read synced events and save confirmed events in a writable calendar |
| Holo messages and media notification access | Visible Google Messages, WhatsApp, Discord and Element X notifications/replies; active music sessions |
| Selected file/folder access | Custom model imports; no broad storage access |
Message reading, calendar access and outgoing device actions require an unlocked phone. The implemented features do not require root or accessibility access.
Builds and releases
Edit WISEWOLF_VERSION_NAME and WISEWOLF_VERSION_CODE in
gradle.properties to version local and Forgejo builds. Increase the code
for each published update. Explicit environment variables can override these defaults.
Forgejo Actions runs on pushes to main, including merges.
It runs checks, signs the release APK using the stored Android signing secrets, and saves
the wise-wolf-release artifact containing wise-wolf-VERSION-release.apk and SHA256SUMS.
Feature-branch pushes, pull requests, tags and manual dispatch do not trigger it.
A signing preflight runs before SDK/model downloads and reports specific keystore, alias
and password problems. Missing/invalid signing secrets fail the build; publishing releases
remains manual. The detected keystore format is passed explicitly to Gradle, and the
packaged APK certificate must match the validated key.
For an Obtainium update, publish a release APK signed with the same permanent key as a Forgejo release attachment: download the signed APK from the successful workflow artifact and attach it to your release. Workflow artifacts are not the update feed. The four signing secret names and migration from an older debug install are documented in the release guide. Add this repository URL to Obtainium and select Forgejo (Codeberg) if the custom host is not detected:
https://git.blusnake.net/blusnake/project-wise-wolf
Runner setup, signing, version codes and manual releases
Current limits
- English commands: local parsers handle supported requests; open-ended web access is not implemented; optional local model answers are bounded.
- STT tuning: bundled English uses four-path beam decoding for alternate acoustic hypotheses. Synthetic regressions pass; real-voice accuracy gains and phone CPU/power cost remain unmeasured. Message/dictation words are never rewritten into command words.
- Managed timers/reminders: Wise Wolf schedules its own Android alerts. It cannot list/control another Clock app's timers. Timers survive normal task dismissal/process reclamation; Force stop, revoked permissions, app-data deletion or a powered-off phone can prevent alerts.
- Messages: reading/replies use current visible Google Messages, WhatsApp, Discord and Element X notifications, not full history. Explicit app selection takes priority over the configured default. Google Messages and WhatsApp support new-message drafts; Discord/Element X open text sharing with recipient selection in the destination app. Drafts require manual sending. No app or SMS substitution occurs.
- Calendar/music: event creation saves to a visible writable Android calendar after approval; multiple calendars prompt a choice. Account sync remains the calendar provider's job. Destination music apps decide which playback actions they support. YouTube Music playlist names use links saved in Settings → Apps & integrations; Holo does not sign into accounts or import its private library.
- Wake hardware: no implemented DSP, root auto-selection, speaker verification, Android Auto, or wake listening before first unlock. Software wake battery cost needs physical measurement.
Detailed behavior and recovery rules are in the usage guide.
Models and acknowledgments
| Component | Version/model | Source |
|---|---|---|
| Speech engine | sherpa-onnx Android AAR 1.13.8 | Repository/release |
| English recognition | Streaming Zipformer 2023-06-21, LibriSpeech + GigaSpeech; about 190 MB of weights, int8 encoder/joiner and float decoder | Model |
| English voice | Piper/VITS en_US-ljspeech-medium | Voice releases |
| Wake model | Phonetic Zipformer zh-en 3M, 2025-12-20, int8 encoder/joiner | Keyword model releases, provenance |
| ONNX Runtime | Included in the pinned speech AAR | Repository |
| eSpeak NG | Bundled phonemization/data | Repository |
| AndroidX / Jetpack Compose | Versions in the project catalog | Source |
| Template audit (development only) | Jinja2 3.1.6 / MarkupSafe 3.0.4, isolated Python reference; not packaged | Audit, notices |
| Optional web transport | OkHttp 4.12.0 / Okio, Brave Web Search service, explicitly approved queries | Source, notices |
| Local LLM framework | MNN 3.6.1, pinned source with Jinja block-capture repair, CPU/private process, internal CV for Qwen companions | Source pin |
| Downloadable AI models | Qwen3.5 2B / 0.8B MNN, pinned 4-bit manifests; weights outside APK | Publisher export, lock |
| Experimental AI models | MiniCPM5 1B / 2B MNN, pinned 4-bit exports, .mtok tokenizers and quantized embeddings; weights outside APK |
1B lock, 2B lock |
Exact model revisions, file choices and SHA-256 checksums are in scripts/prepare_speech.py. Imported packs retain their upstream licenses and are not automatically bundled or redistributed.
Wise Wolf's project-owned source, scripts, documentation and original artwork are
licensed under GNU GPL version 3 only (GPL-3.0-only); see LICENSE.
You may redistribute and modify that work under GPLv3; it comes without warranty.
Third-party components retain their original licenses and notices. The canonical
license inventory and mandatory dependency policy requires GPLv3
compatibility and a recorded name/source/license before every new integration.
Existing proprietary/unverified components remain blocked; selecting GPLv3 does not
clear the current candidate for distribution. See THIRD_PARTY.md.
The original ear SVG and its README PNG preview use the same brown/cream palette as the launcher and listening indicator. The app draws cached vector paths; the README preview adds no runtime dependency.
Documentation
| Guide | Contents |
|---|---|
| Using Holo | Timer conversations, alert reliability, wake setup and integration behavior |
| Speech models | Compatible packs, downloads, directories, manifests and validation |
| Playlist files | M3U investigation and why YouTube Music offline downloads are not a supported playlist-file source |
| Builds and releases | Forgejo runner setup, manual signing/releases and Obtainium |
| Architecture | Code ownership, audio/lifecycle boundaries, storage and extension points |
| Modules and LLM integration | Matching rules, independent providers, execution, global skills and a working module example |
| MNN integration | Native process, model installation, routing and validation limits |
| Shared LLM skill setup | One module skill definition for Qwen and MiniCPM5, with examples and backend rendering |
| MiniCPM5 1B audit | Published-template compatibility, controlled native fixtures and workflow limitations |
| 1.2.5–1.2.7 roadmap | Stabilization, tool-free Conversation and explicit Assistant search |
| Hardware wake research | SoundTrigger/DSP access, Pixel constraints and evaluated DaVoice options |
| Manual testing | Device acceptance checklist |
| Validation record | Completed checks, measured results and remaining device work |
For the standard app checks after preparing assets:
./gradlew :app:assembleDebug :app:assembleRelease :app:testDebugUnitTest :app:lintDebug :app:connectedDebugAndroidTest
Before publishing, run the release startup check against the installed optimized APK, then try dictation/read-aloud and assistant activation. Debug tests and successful APK signing do not cover release optimizer failures.
The maintenance baseline passed 147 JVM tests and 71 Android tests on the API 37 emulator. MNN checkpoint builds, native probes and task acceptance have separate results and remaining checks in VALIDATION.md. Physical-device audio, lifecycle, battery and installed-app behavior remain separate checks. Synthetic fixtures do not establish real-speaker accuracy.
Keep this page current when features, setup, models or limitations change. Record verification in VALIDATION.md; AGENTS.md defines the maintenance expectations.
Conversation controls
A pending question owns its next reply, including module words such as “Meeting” as a calendar title. Clarified titles and schedule fields are retained separately. Questions expire after two minutes; expired approvals cannot execute. Use “new request …” to discard an unanswered question and start another request, or “override …” for model interpretation. Holo settings → Advanced → Conversation controls → Cancel words configures exact cancellation replies (default no, cancel). Cancellation applies only while awaiting an answer, ignores surrounding punctuation/case, clears the activation and closes Holo. Extra words such as “No, change the message” do not count as exact cancellation. Completed actions remain completed. Every resummoning/new conversation clears the activation state.
Wake handoff trims recent audio to the detected phrase onset with a 160 ms safety margin, retaining commands spoken immediately after it. Synthetic priming and replayed samples are excluded from live sample positions. Synthetic emulator replay covers preceding failed attempts, pauses and a reduced-volume onset; this is not a physical accuracy measurement. Generic “play a song on YouTube Music” requests resume playback without searching for “a song.” Named searches and saved playlists keep their existing app-specific handoffs. Where notification/media access permits, Holo briefly observes playback state outside action locks and reports whether that app reports active playback. It cannot verify the selected song.
Settings now groups Setup & permissions, Voice & wake, Apps & integrations,
Local LLM, Appearance and Advanced. The Setup button opens the same checklist
used for revisits: microphone, notifications, exact/full-screen alerts, assistant selection,
battery settings and optional contacts, calls, calendar, SMS, call history and notification
access. Status refreshes on returning from Android settings. Optional access remains skippable
and is requested only by tapping its control. Timers, calls and message composition ask local
questions for missing fields. “The message is no” supplies the exact body no.
The LLM boundary receives per-function access/mode/authentication observations; the host
rechecks availability before preparation and confirmed execution.
Summoning and hands-on use
After summoning, Holo allows ten seconds to begin speaking. A pause after a detected wake phrase does not count as a completed command; ordinary command endpointing stays responsive. Button/gesture invocations and touches inside the assistant engage hands-on mode. Mic/send presses, composer edits and history touches keep the window open until explicit dismissal or 30 seconds of inactivity. Processing and spoken replies do not consume that idle allowance. Touching the composer while listening pauses capture and opens typing when ready. Explicit conversation cancellation and requested app handoffs retain their normal closing behavior.
Visible listening controls
Manual app launch never starts recording or closes for inactivity. Speak starts one voice request; only a pending clarification/confirmation retains the in-app voice cycle. Typed requests and response-button taps stay quiet and cannot start a microphone. Summoned sessions start listening automatically, retain voice follow-ups and close for inactivity. Typing pauses capture and the inactivity clock. Dictation retains explicit recording controls. Voice & wake offers wake idle dismissal (5–30 seconds, default 10), hands-on idle dismissal (15–120 seconds, default 30), and optional microphone-ready Sound/Vibration cues (default None). The initial wake capture also has a 30-second absolute deadline that noise, changing partial transcripts and decoder resets cannot extend. Empty, punctuation-only and wake-name-only captures never reach the model or restart recording. Time bounds limit accidental listening; they do not identify every background conversation. Processing and speech output suspend inactivity dismissal. Cancel stops queued work; completed actions remain.
Responses can offer module-owned timer and clarification buttons, including existing contact and calendar choices. Protected actions show the exact host-prepared plan with Confirm, Edit and Cancel. Edit discards approval and opens the request for correction. Buttons expire on use, editing, a newer response or dismissal; model-proposed buttons use registered functions.
Resource diagnostics and model controls
Paused wake listening uses signalled waiting and releases capture and its CPU wake lock. Screen-off wake capture retains the CPU lock it needs. MNN speculative preload skips Android power-saving mode and battery levels at or below 15%; on-demand interpretation remains available. No physical battery improvement has been measured.
Home suggestions refresh after access/settings changes, always include local clock/conversion examples, and fill the composer without executing. Local LLM shows friendly model names, download sizes, installed/selected state and distinct Download, Pause, Resume, Retry and Use controls. Interrupted transfers retain verified/partial files. Technical sources stay in Advanced.
Advanced offers explicit timing/resource export from bounded process memory. It records the last detected acoustic activity, recognition finalization, model preparation/inference, execution, first drawn response, task completion and spoken-output start, plus capture/CPU-lock durations and listening transitions. Missing boundaries remain unavailable, including speech boundaries for typed requests. No transcripts, contacts, message bodies or recordings are collected.
Music playback requests
“Play music” resumes the selected player's queue or asks its public media session to choose music with an empty search when no queue exists. A bounded media-browser connection supports players that are not already active. “Play free bird” in YouTube Music uses its explicit playback-search activity with the unchanged query, rather than opening its home screen or sending a background search that may be ignored. Successful play requests use visual feedback so a spoken acknowledgment does not immediately take audio focus from the player.
The player still owns catalog matching, sign-in, network access and account restrictions. Observable session requests require a playing acknowledgment; unacknowledged requests report failure without an automatic second request. Activity handoffs explicitly remain unverified. The unsigned-in emulator cannot establish actual signed-in YouTube Music song selection or streaming; those remain phone checks. No account scraping or private playback API is used.
1.2.7 Conversation streaming
Conversation previews actual generated answer text while decoding. Reasoning and model framing stay hidden. A pending turn enters history and speech only after successful completion; cancellation/failure leaves a labelled partial preview. Rotation preserves the active stream; backgrounding cancels it while retaining completed history. Assistant and the summoned overlay show loading, input-processing and generation phases without rendering partial tool calls or confirmations. Streaming does not reduce computation time. Native streaming passes the 2B and Qwen Conversation checks on x86_64. MiniCPM5 1B recall and immediate teardown fail; optimized streaming and physical-device gates remain open in the candidate release record.
OCR was removed from the unshipped 1.2.7 candidate and is deferred beyond 1.2.8. Ordinary summoned invocation disables screenshot and assist context. No replacement OCR engine or model is included.
Optional Assistant web search
Settings → Web search accepts a Brave Search API key and separately enables search; search starts disabled. The key is encrypted with Android Keystore AES-GCM in private backup-excluded storage and never enters model prompts, logs or diagnostics. Enter it in settings, never chat. This screen prevents screenshots of the key entry.
Use Search web beside an Assistant draft, or an explicit command such as “search web
for lunar phases.” Only that exact host-approved query can be dispatched once in the turn.
Models and retrieved snippets cannot grant permission or add private context. Search is a shared read-only search__web(query) module;
Conversation has no provider or gateway access in 1.2.7.
Up to five result cards appear with titles, excerpts and clickable HTTP(S) links. Search uses pinned OkHttp 4.12.0 with 5-second connect, 10-second read and 15-second overall limits, cancellation, no automatic retries and no redirects carrying the key. No full pages are fetched. Offline, authentication, rate-limit, empty-result and timeout outcomes are explicit. If Local LLM is enabled, an optional separate tool-free, thinking-off summary runs after the search gateway returns (128 output tokens, 20 seconds). Only citations for that result batch are accepted. Invalid/incomplete/failed summaries leave the result cards without a generated answer. Live configured search and owner-device acceptance are pending.
Licensing review: OCR and its proprietary transitives are removed. Missing Maven declarations were checked against exact source/artifact notices. The converted voice-weight grant, complete speech-native source/notices and historical test-audio provenance remain unresolved; 1.2.7 is not cleared for distribution. See LICENSES.md.
Production LLM prompts use current-task data and schema-derived contracts. Synthetic teaching conversations are omitted after a Qwen diagnostic copied an example recipient; registered examples remain schema-validation fixtures. Native teardown and model recall failures remain listed in the 1.2.7 validation record.
1.2.8 development: background efficiency
Software wake listening uses batched blocking microphone reads and reusable four-second audio history. Quiet audio continues through the detector; fresh streams recover speech onset after quiet rather than restarting every quiet second. Stop closes the owned input to interrupt capture. Screen-off listening retains its CPU wake lock and microphone cost. New installations load the LLM on demand; explicitly saved preload choices remain honored. Recognition and physical battery acceptance remain pending; see VALIDATION.md.
1.2.8 separates Wise Wolf Debug (com.blusnake.projectwisewolf.debug) from Release:
separate signing, data and installation. Forgejo manual dispatch offers a Debug checkbox
(default off); pushes to main build Release. Debug needs no distribution secrets. Local
Debug packaging uses bash scripts/build_debug.sh; artifacts include -debug or -release
and SHA256SUMS. Release signing/certificate verification remains mandatory in CI.
Typed startup initializes no recognizer. Speech loads for capture/output and releases on backgrounding after its worker finishes. LLM native loading uses integrity/profile and load-readiness checks; generated OK/timer probes are explicit development tests only. Old installation records remain readable without generated requalification. Load readiness does not mean workflow accuracy. Diagnostics collection/export is Debug only, and Debug does not automatically run accuracy tests. Modules must be compiled into the APK; runtime module importing is unsupported. No model, asset, architecture or dependency is replaced. Conversation search (explicit/automatic, opt-in, query approvals, sources/citations/follow-ups) is deferred to 1.2.9 or later. Explicit Assistant search remains supported.
The owner’s subsequent functional closeout supersedes the initial two-cycle stop. Build outputs and core regression evidence are recorded separately from physical power and licensing release gates in VALIDATION.md.
The 1.2.8 lifecycle closeout resets a reused summoned session before publishing it as visible. Finalized empty capture consumes the existing idle deadline rather than granting a second one; accepted speech still owns finalization. Repeated summons, idle dismissal and wake resource shutdown pass emulator checks. One fast private timer recording still fails ASR, including without wake cropping. Physical power/accuracy, arm64 execution and licensing release gates remain open; this is a candidate, not a completed release.
Smaller-model prompt retests are mixed: on the matched three-minute timer request, Qwen3.5 0.8B retained the correct 180-second call with 32.3% fewer input tokens. MiniCPM5 1B used 16.9% fewer tokens but changed the correct original result to an incorrect three-second call; the test rejected it before execution. These single-workflow observations do not qualify broader accuracy or change model recommendations. See VALIDATION.md and the recorded prompt comparison.