Andee lives on your phone or Mac and uses any app like you do—no integrations, no root, any model; memory and passwords never leave your device | Andee 住进你的手机或 Mac,像你一样操作任何 App——无需集成、无需 Root、支持任意模型,记忆和密码绝不离开设备。
Changelog
Notable changes to Andee. The format follows Keep a Changelog.
Versioning policy
versionCode/versionNamelive inapp/build.gradleand are currently19/"0.8.1". The Mac's product version is read from that sameversionName, so both bodies carry one number.- Releases are tagged
vMAJOR.MINOR.PATCH,versionNamemirrors the tag without the leadingv, andversionCodeincreases by one per release.versionCodemust increase monotonically or Android refuses the upgrade; the same number must never be reused for two different builds. - MAJOR for a change that breaks the wire protocol or an existing configuration; MINOR for a new capability; PATCH for fixes.
The project starts at 0.x, and that is the accurate number rather than a modest one. Nothing is
frozen. The tool-schema contract a hub brain speaks has been rewritten repeatedly
(net/ToolSchemas.kt has been rewritten in most releases — it read "in every release so far" until
a release did not touch it, which is why the count is not written down here), the local.properties
build flags that decide the first-install language, the thinking default and whether CodeFlying is
compiled in have all been added since, and a saved scene is a shape that only appeared in 0.2.0. Each of those is something a user or a second implementation would have to rebuild against,
so 0.x is being honest that MINOR may break. The release that moves off 0 should be the one where
the wire contract is written down well enough that a second implementation could be built from it.
A version number identifies a release, not a build. versionCode now moves with each release,
so the installed app can say which release it came from — but nothing in it moves when main
does. A build made from a commit between two tags reports the older release's number, so for
anything not built at a tag, identify it by the short commit hash it was built from. (Embedding that
hash into versionName at build time would close this hole; it is not implemented yet.)
[Unreleased]
[0.8.1] - 2026-10-30
A cuter Andee, a notebook that starts with something in it, and four fixes that came out of using it. The default face was retuned against the reference illustrations; a fresh notebook now comes with one todo already in it; and the rest is the kind of thing that only shows up on a real device: a key saved from the self-check sheet that the running brain never saw, home gestures on the phone, and a voice handshake that failed silently.
Added
- A fresh notebook starts with one todo. Every evening at 21:00 (
0 21 * * *), it looks back over the day and makes a picture of "you and me" with some mood to it. Seeded once per notebook: a meta flag stops a todo the user dropped or paused from coming back, and a wipe (clear notebook, factory reset) takes the flag with it. Both bodies seed it on their start-up catch-up; the Mac also re-runs that after clearing the notebook or a factory reset. The todo assumes no particular tool and no particular language: image generation exists only where there is one (CodeFlying's market), so it says to use an image tool if you have one and otherwise draw the picture as a page withshow_html. Its text is model-facing English like every other wake-up line, and says to write and speak in the language of the recent conversation, falling back to the interface language — a notebook is seeded before anyone has said a word, so the interface language at that moment is the one thing that is reliably wrong. Notebooks holding an earlier wording are rewritten in place, same id, on the next start. - The Mac has the CodeFlying market.
MarketToolsmoves tocore(its only Android dependency was the logger) and the Mac registers it as a tool provider built from the brain's own config, somarket_search/market_callexist only on the CodeFlying backend, as on the phone — inert, no tools listed and nothing claimed, on any other backend. A key or backend change rebuilds the brain and this with it.
Changed
- The default face, retuned toward the reference illustrations. A periwinkle body with a peach rim
light, bigger and darker eyes with two glints, eyes and cheeks raised (
eyeY 0.09 → 0.07) with the mouth left low, rounder horns, and finer meshes for the eyes, mouth, horns and props. Gloss drops0.22 → 0.05and the rim light becomes fully mixed (rimTintMix 0.55 → 1.0), which is what moves it from plastic to matte. The look'snameis unchanged — this is the same character with a new face, not a new one, so nobody's saved choice moves. - A repeating todo is not closed when it comes due.
doneends a cron series for good, so the first evening's answer would have been the last. The wake-up line (WakeText.due(repeats)), thetodoresult on both bodies, and the prompt now all say so.
Fixed
- Thinking no longer carries the bigger eyes past the edge of the head. Oversized eyes made the old
drift values visible as an error: the thinking drift is now
0.03(was0.075) and the head turn it drives0.04(was0.09), with the slow sway and the drift capped together at0.10 radso the outer eye cannot roll off the sphere. Brief actions — a shake, a look around — are deliberately not capped, since they end before it matters. Checked over 36 frames atdp 420and at the foldeddp 136. - A key saved from the self-check sheet now reaches the running brain. Measured on a device: v0.8.0
reinstalled at 14:30 built its brain with an empty key, self-check reported
brain:FAILat 14:30:34, the key was saved at 14:31 (a 36-characterkft_value invoice_prefs.xml) — and the first typed sentence at 14:33 still answered "no key". The brain speaks that line from its owncfg, which is a snapshot taken at construction, and nothing re-read it. The self-check page can appear while the service is already running (itsonCreatefalls through wheneverensureOverlays()cannot vouch for the overlay, which includes accessibility being off while the brain is alive), so the path that assumed "no service, nothing to rebuild" was only true of one of its two entry points.BodyCoreServicealready made this comparison for its own sheet; it is lifted out asreconcileBackendAfterSettings(beforeHub, beforeBrain)and called from both hosts. - Factory reset keeps the CodeFlying key, on both bodies, and the confirmation now says so instead of
leaving it to be discovered. The phone cleared it along with the rest of
voice_prefs; the Mac deleted it too and only looked as if it kept it, because settings were saved again afterwards. - A cleared conversation no longer contains a line about itself.
clearConversationcalledstop(), whosestopAllwrites 「被你叫停了」 from the brain's own thread — after the wipe. The discarded brain is now shut down without a report and the history is cleared afterwards. A stopped review is also named 「回顾这段对话」 in that row rather than the first line of its English instruction. - Home gestures and overlay navigation on the phone. Navigation boundaries are covered and failed home actions are logged rather than dropped.
- Overlay opening and closing on the phone does less work per frame, as does layout during keyboard transitions.
- A refused voice handshake reads as a scrollback error. On the Mac a tap that could not connect used to fail silently; the reason now reaches the conversation.
- Volcengine voice endpoint and resource defaults show as hints in settings rather than looking like something already filled in.
- The cleartext exception for
124.71.176.202is gone from both network security configs. CodeFlying ishttps://baas.kuafuai.netnow, and an exception for a host nobody calls only widens what the app may talk to in the clear.
Documentation
- Volcengine voice needs an entitled resource, not just a key. A newly opened service defaults to
post-paid and can only be used from Volcengine's own trial page — calls over the API are refused, so a
correct key in a filled-in field still dies at the handshake. The switch is on the console's
开通管理 page. Recorded because the two failure shapes both surface as
ASR ws refused: …and neither message points at that page. Two ASR resource ids this build can use are named in the README, and version 2.0 is explicitly ruled out.
[0.8.0] - 2026-10-29
It has something of its own now, and it can act on it. Until this release every turn began with an event from outside: the user spoke, an alarm fired, a notification arrived, the review ran. With no event, nothing happened. This gives it a self to answer to — what it cares about, who it knows, what it has been doing — and a heartbeat that wakes it to look at that.
The design and its reasoning, including the parts chosen against, are in docs/design/inner-drive.zh-CN.md (Chinese).
Added
- A self it writes itself (
self_note) — a first-person account of who it is and what it has been busy with, rewritten as it goes. Not a persona skin: it is what the next wake-up reads to become itself again, since it has no continuous experience. Versioned (up to 20) and readable in 小本本 → 自己, alongside 在意, 日记, 人 and 维基 as tabs of the same book. - Things it cares about (
cares) — each with how often it means to look. The device computes how overdue each one is (ratio ≥ 1.0 means the time has come) and sorts by that; the ratio is the device's arithmetic, never the model's. This is what the heartbeat asks about, and the only thing it may pick up work from. Up to 7 active at once. - The people it knows (
person,people) — one entry per person, held in the encrypted notebook, because this is other people's information. Deliberately not called 通讯录: that word in this project means the system-contacts permission, which is a different thing. - A diary (
diary) — one page a day, written in the evening; 40–1200 characters, up to 60 pages. - A notebook it writes for itself, in plain files — plans, research and write-ups land here, indexed in the notebook but stored as files, so any editor can open them and they can be shared. This store is deliberately not encrypted (its own decision, as working documents), so third-party details belong in the people/notebook side, not here.
- A heartbeat — about one look every 20 minutes, at most 40 a day, deciding whether anything it cares about has waited long enough to be worth doing. Most beats stop at a free gate without any model call at all; when one does go through, it costs roughly 500 tokens.
- Work turns — what a heartbeat may start: a full turn with every tool a user's turn has, capped at 6 a day with at least an hour between them, so a day's allowance cannot be spent in its first two hours. It touches no screen, speaks nothing, and sends nothing on your behalf.
- A tap on the ball stops everything — the running turn, the queue, and wakes that were waiting to start, plus any bubble on screen. One row says what was cut off and what was dropped, so a stop is not silent.
- 动态 (Mac) / the timeline (phone) — every non-conversation event as one quiet line: a todo waking it, the review and what it folded, its writes to its own record, a heartbeat's thought and its work turns, bubbles, notifications it let pass, meetings, the camera, shell commands. A row with detail opens on click; 全部 / 只看对话 filters them. Several things that used to be fake assistant lines are events now, which is more honest and easier to scan.
- A WeChat icon beside GitHub and Discord that opens Andee's QR code, on both bodies.
- An out-of-date build is told once at launch and pointed at CodeFlying, on both bodies. "Behind" means a release strictly newer than this build — a build ahead of the release is current.
- Android can use the system TTS as its speaking provider, as an alternative to a cloud voice key.
- 13 new tools —
self_note,cares,diary,person,people, and eightwiki_*;wiki_showopens a written article from the timeline on the phone as well.
Changed
- The default step limit is 40, up from 25, on both bodies.
- Nothing asks whether you are there, or what time it is. The heartbeat, bubbles, the morning plan and the muse no longer wait for input, an unlocked screen, or 08:00. The brakes are the switches, the budgets, and a tap on the ball. The settings copy that still described the old hours has been corrected in all four language/platform files — it had been left describing behaviour the code no longer had.
Fixed
- A quiet afternoon is one review row, not eighteen. A stretch of conversation could be reviewed four times over (a sweep, the harvest straight after it, and both again after an app restart), and every write in each became its own timeline row. The Mac now records "this silence was already reviewed" across restarts, a harvest straight after a sweep writes only the hand-over summary, and a review is one row: "复盘了一遍,记下 N 处". A rewrite that changed nothing is no event at all.
- When its aim keeps missing, the device points the model at
tap_target. On 2026-10-29 the Mac missed a thin reply box six times with grid taps, and the model ended up asking the user to click it — while onetap_targetcall would have found it. Now a second coordinate tap that changes nothing, an e-number tap that changes nothing, and the same-spot refusal all carry a hint namingtap_target; nothing the device said before had mentioned it. - A todo paused or closed while its turn was waiting no longer runs, and a paused plan bubble stays paused through a reconcile.
- Reviews see every memory and pending todo before writing, so they update instead of duplicating, and a todo that tries to number several things is refused.
- The upgrade link is
codeflying.net/andee(Chinese) /codeflying.app/andee(otherwise); the service'sdownload_urlis no longer used. The Mac logs its launch version check. - Android: 小本本 keeps the picked tab in view after the card rebuilds, and the Mac settings section named 小本本 is renamed 复盘, which is what it is.
Notes
- One device is one Andee. The one on your phone and the one on your Mac are two individuals with their own narrative, cares and people; they can find each other over the peer channel but they do not share memory. This is not a limitation to be fixed — an individual does not live in two bodies, and it avoids cross-device sync, conflicts and doubled budgets as a class.
docs/design/inner-drive.zh-CN.mdstill says "草案,未实现" at the top. That line is out of date — the classes it describes are all present. Recorded here rather than silently edited, because the document's status line is the author's to set.
[0.7.1] - 2026-10-28
Taps land where it means them to. Our own numbers said the brain was a poor pointer: on the Mac between 21 and 27 October, taps aimed by element number worked about 80% of the time, taps aimed at a coordinate after zooming about 43%, and unaimed coordinate taps under 10%. Half of that was the model being asked to do arithmetic it is bad at, and half was guards the phone had and the Mac did not.
The answer is to stop asking the brain for coordinates at all. It now describes what it wants in words, a vision model that is actually good at pointing answers with a point, and the device taps there.
Reasoning, measurements and the options that were rejected are in docs/design/click-accuracy.zh-CN.md (Chinese).
Added
tap_target— describe the thing to tap, in words; a vision model finds it and the device taps its centre. On both bodies, local brain only. The brain never produces a coordinate, so nothing can be off by its arithmetic.- It uses the model you already configured for the brain, on every backend, so there is nothing new
to set up.
locator_modeloverrides it if you want something faster. - Measured on our own screenshots, ten targets with known answers: 6.4 px mean error, 24 px worst, about 2 s and 1400 input tokens each. The alternative — the brain counting cells itself — is what this replaces.
- It degrades instead of gating. A backend that will not take images withdraws the tool for that run, as do three unreadable answers in a row; one good answer resets the count. After that the brain falls back to aiming by cell, which is what it did before.
- Two prompt choices were measured, not guessed: asking in Chinese (the same targets in English were off by 25–37 px on three of them) and naming what it sees before deciding (without it, "the red 删除 button" was answered confidently with the close button). When the target is not there, it says so.
- It uses the model you already configured for the brain, on every backend, so there is nothing new
to set up.
drag— press, move while holding, release: a real drag on the Mac, for things that have to be grabbed rather than scrolled (a slider, a splitter, a window by its title bar, a row to reorder). Both ends are aimed by coordinate or in words.
Fixed
- The Mac gained the guards the phone already had. Each of these was a measured failure rather than a
hypothetical one:
- A tap now reports whether the screen changed, with a magenta ring marking where it landed. Without it the model could not tell a tap that missed from one that opened a menu.
- Elements a tap cannot reach are no longer offered. A closed menu item used to be listed, and tapping its number hit the card underneath. Checked against Chrome: 123 visible elements, no misjudgements.
- Tapping by element number checks the element is still there first, and refuses with a fresh list if it is not.
- The repeat-tap guard radius went from 120 pt to 30 pt — the old one refused 9 of 56 taps, including corrections that were being aimed at something else.
- Apps that do not expose their UI are recognised. WeChat's Mac window offers four elements in total; a window that large with so few elements is now treated as tree-blind, and a screenshot is attached automatically with a note to look at it.
- A window on another display is reported rather than aimed at, and the scale comes from the main display instead of whichever screen has keyboard focus.
- Every click brings its app forward first. On the Mac a click on an inactive window only activates it, so focus stayed on the page and a following paste went nowhere — which is how "create a new app" appeared to do nothing on the first tap.
- Hit testing only considers ordinary windows, so the Dock's full-screen window no longer answers for every point.
- The Mac can type into apps that hide their focused field (WeChat is one: it reports no focused
element, so
type_textrefused). Pastes with ⌘V only — deliberately no ⌘A, because with focus unknown select-all could take the chat history — refuses if Andee's own window is active, and says the result is unverified rather than claiming success. The model had been working around this withosascript, which is closed on notification turns: a message it could read, it could not answer. - A backend that answers with something other than the API now says so. OkHttp turned a POST meeting a
301 into a GET of the new address, and on a moved gateway that is an HTML page, so the turn died on
"A JSONObject text must begin with '{'" — a message that tells nobody anything (measured:
codeflying.net→www.codeflying.net, an nginx page). Redirects are no longer followed, and a non-JSON 200 is reported as an address that is wrong or has moved, naming where it pointed and showing what came back. - The cancel button in the Mac's destructive-action confirmations was squeezed to a sliver because both buttons filled the row; each now takes half. It read vertically.
- Clearing the conversation now clears all of it. The Mac's 清空聊天记录 did three things where the shared routine does seven — it did not stop or drop the running brain, did not clear what the brain was built for, and left the progress caption up. A stop that only half-stops is how a cleared conversation carries on talking.
Note on verification
tap_targetis not yet verified on a phone. The design document says so, and it bears repeating: it needs a phone whose voice and brain keys both work, and a WeChat comment or reply tried once. On the Mac it has been used in a real task twice (creating a new app in 码上飞), and landed both times — a sample of two, not proof.
[0.7.0] - 2026-10-27
It can speak first now, and it has a reason to. Until this release every turn began with the user:
todo alarms, notification triage and the nightly review were all reflexes to something that had
already happened. This adds the two things that turn a reflex into an intention — a way to open its
mouth that costs almost nothing, and a plan that gives the opening a why.
The design is written up in docs/design/proactive-nudge.zh-CN.md (Chinese). Its rule is that the cost of interrupting is asymmetric: missing one chance to speak costs the user a question, while one interruption too many gets the ball muted and silences the good ones with it. So the default is silence, speaking needs a reason, and the reason is rationed.
Added
- Bubbles (
nudge) — the lightest way to speak first. A small silent bubble beside the ball, on both bodies. Nothing is read aloud, the microphone stays off, and the turn is not blocked: it returns at once. It fades if ignored, always yields to a card, and is rationed per hour, per day and perkindat a level the user picks (off / few / normal); a kind can be muted by long-press. The bubbles yield, never displace — the reasoning is that a bubble is an attempt to be useful, not a claim on attention.- With up to 3 short buttons, a tap comes back later as its own turn. The button meaning "no thanks"
is marked
dismissby whoever sent the bubble rather than guessed from its label, and tapping it closes the bubble without a turn — a refusal needs no model to answer it. - Tapping the bubble's text (not a button) opens an ordinary conversation: the equivalent of the user saying "go on, then".
- With up to 3 short buttons, a tap comes back later as its own turn. The button meaning "no thanks"
is marked
- Plans (
plan_directions,day_plan) — what it is trying to get done, and how today goes.- Directions are the long-term layer: what it is helping with over these weeks. They may be created
only from what the user said or did, and each carries a
sourcenaming when and what — a goal the model thinks would be good for someone is not theirs to add. - Day plans are per-day items, each pointing at a direction or a todo. A timed item becomes a bubble raised by the device with no model call at all; outcomes are written back to the plan.
- Visible and editable: Settings → 它的打算. The user can pause or remove an entry; everything else is done by telling it, rather than through a form with a field for every case.
- One plan per body. The phone and the Mac each keep their own.
- Directions are the long-term layer: what it is helping with over these weeks. They may be created
only from what the user said or did, and each carries a
- PlanMind — a morning look and a spaced muse. Once a day it takes a look at the day ahead and drafts a plan; for items with loose timing it thinks again at intervals. Each is a single call with no tools. The nightly review compares the day's outcomes before drafting tomorrow, and looks over the directions weekly.
kittenlook and voice, and stronger personas across the existing looks.
Changed
- The step limit no longer loses the task. See Fixed below; this is listed because it changes what
the user sees after a long task, which is now resumable by
继续rather than silently truncated.
Fixed
继续after the step limit now picks up where the task stopped. Three things were wrong at once: the sentence the user heard never reached the history, so the model did not know it had been cut off; the trim protected only the newest turn, so a todo firing between the stop and the继续made the task "old" and it was dropped (measured 2026-10-22: 104 messages trimmed to 52, task gone); and tool calls left unanswered by the stopped turn made the next request fail with HTTP 400.- A turn stopped by the step limit or the repeat guard is now kept unfinished: its sentence goes into the history, its messages survive the trim, the harvest and the topic folds, and every user message says what was stopped until a later turn finishes it or an hour passes.
ContinueAfterStepLimitTestreplays the failure against a scripted model, and fails on the old code.
- Android: tapping the ball mid-task no longer ends listening immediately. Listening was ended on every loss of audio focus — a rule that speaking needs, applied to the side that makes no sound. Mid task the driven app is in front, and its autoplay, message tone or voice clip took the floor, closing the turn the tap had just opened. Now a duck request is ignored, a call still ends the turn, and any other taker is asked to hand the floor back — at most twice a turn — while listening carries on. Reasoned from the code, not from a capture: the Huawei test phone keeps no app logs and its voice key is invalid, so this is not yet verified on a device.
- macOS: the notification store is found on every version. Andee looked for
usernoted's database in one place only — the group container macOS 15 uses. Earlier systems keep it under the per-user Darwin directory, where the file does not exist, and that was reported as missing Full Disk Access, so granting it and restarting changed nothing. Now each known location is tried in turn (the old one asked of the system withconfstr(_CS_DARWIN_USER_DIR), not derived from$TMPDIR, which the JVM replaces); "no store anywhere" and "the store refuses us" are separate states, and only the second is a permission problem — the first says so plainly and offers no button to System Settings; and the read happens in-process throughlibsqlite3instead of forking/usr/bin/sqlite3, which had swallowed the reason a read failed. Verified on macOS 15.5 against the live store, including a row still in the WAL, and against a fake home for both failure states. The macOS 14 location is not verified on a device.
[0.6.2] - 2026-10-26
The download list now says what each file is for, and lists them in the order you would want them.
Every asset carries a label naming its platform and device — Andee 0.6.2 · Android-arm64-v8a,
Andee 0.6.2 中文版 · Mac-Apple 芯片 — and the list reads Android before Mac, most common device first,
English group before Chinese.
No behaviour changed in the app itself. This is a packaging release: the same ten builds as 0.6.1, named and ordered better.
Changed
- Asset file names gained a platform field:
andee-<version>-<lang>-<platform>-<device>.<ext>, soandee-0.6.2-en-android-arm64-v8a.apk. Three things depend on that shape, and each was decided by measuring rather than by preference:- The order a reader sees is the alphabetical order of these names. GitHub sorts the asset list by
name and the upload order is discarded, so the display order has to be encoded in the names —
en<zh,android<mac,arm64-v8a<armeabi-v7a<x86_64,arm64<x64. All four comparisons were checked before the shape was chosen. - No Han characters. GitHub deletes them from asset names, so a file named with the Chinese wording arrives with it removed and two separators left touching. So the name is ASCII and the Chinese wording goes in the label.
- The name is now stable across the three scripts (build, verify, publish) that parse it.
- The order a reader sees is the alphabetical order of these names. GitHub sorts the asset list by
name and the upload order is discarded, so the display order has to be encoded in the names —
- Every asset has a label, both languages. The Chinese ones carry the Chinese wording, which only survives in the label; the English ones carry a matching one so the two groups read as a set instead of the second half of the list looking bare.
Fixed
verify_variants.pyandpublish_variants.shfollowed the new naming. Both parse file names to decide what a package is; both were updated in the same change, because a name change that lands in only one of them produces a release that either fails verification or publishes the wrong pairing.
[0.6.1] - 2026-10-26
One checkout now produces every shipping variant, and the packages cannot contain a key. Four builds come out of it — Android and macOS, each in Chinese and English — totalling ten files, because Android splits by ABI and the Mac by architecture. A variant is two build settings, both non-secret and both given on the command line; the credentials are not settable at all.
This exists because v0.6.0 shipped a Mac build with no server configured, and the Mac silently fell back to a different backend rather than saying so. Two causes, both fixed here.
Fixed
- The Mac had no build-time default language, so a package built as "Chinese" came up English on an
English machine — the package was not the thing its name said it was.
DEFAULT_LANGis now written into the generated config and read byStrings.lang(), ranking below a language the user has picked and above the system locale, exactly as the phone does it. - A second variant build in one checkout reused the first one's
BuildConfig.-Pvalues are read during configuration, which AGP's up-to-date check forgenerateDebugBuildConfigdoes not see, so building Chinese and then English produced English packages filed under the Chinese name — the second run rewrote the config and re-packaged, but the first run's task had been skipped, leaving the APKs as whatever was on disk. The two settings are now declared inputs of that task. Found by readingDEFAULT_LANGout of the finished APK, not by the build log — which saidBUILD SUCCESSFULboth times. - Credentials can no longer be baked in.
CODEFLYING_KEYandGOOGLE_VOICE_KEYare hard-coded to the empty string instead of copied fromlocal.properties, and both build files refuse a-PcodeflyingKey=/-PgoogleVoiceKey=rather than ignoring it. Previously "the key ships empty" held only for as long as nobody uncommented one to test something.
Added
build_variants.sh— builds all four variants from one checkout, naming the Chinese packages in Chinese. It moves the runtime image aside between Mac architectures (the trap that produces a mixed bundle while reporting success) and checks all four architecture points before accepting a dmg.verify_variants.py— confirms each package is the variant its name claims, by reading the artifact: the APK's own dex, andcodeflying.propertiesfrom inside the dmg's jar. It checks that the server matches the language, that the language matches the name, that the architecture matches, and that no credential is present anywhere — includingassets/andee_key.txt, which is read before the build config and would override it.
Changed
-PcodeflyingDomain,-PdefaultLangand-PcodeflyingSsloverridelocal.propertiesfor both platforms. The variant is stated where it can be read back from the build log, and the properties file is never edited between builds — which is the step that eventually ships the wrong package.- Both shipping servers are HTTPS-only, measured rather than assumed:
baas.kuafuai.netanswers 200 on https and 404 on http,myapp.codeflying.appanswers 200 on https and does not connect at all on http. HencecodeflyingSsl=truetravels with the variant rather than being left to the file.
[0.6.0] - 2026-10-26
The assistant now has a second body. Andee runs on macOS, on the same brain as the phone: the
agent loop, the notebook, the memory, the tools and the emotion ball are one implementation in a new
core/ module, and each body supplies only the parts that are genuinely its own — the accessibility
service and the IME on Android, the accessibility API, real mouse and keyboard, and a sandboxed
terminal on the Mac. Two devices that run it can also be friends: they pair by QR code and ask each
other for help, on-body, over an encrypted link with no server in between.
This is the largest release so far by a wide margin — 53 commits, 217 files, +23,976 lines — and the version number finally moves on capability rather than being argued about: MINOR, as the policy says.
Two behavioural changes ride along that a reader should not miss. CodeFlying is now the default backend on a fresh install, replacing on-device as the starting point. And on both bodies the wake word, the settings, and the self-check now describe the same app, so "phone" and "Mac" differ only in what the body can physically do.
Added
- The Mac app (
desktop/, Compose Desktop). A floating ball above every window; the same chat, settings, self-check, notebook, scenes, expressions, artifacts, topic and vault cards as the phone, laid out as a sidebar with a glass panel on a real backdrop; voice in and out (the same four engines as Android, underdesktop/.../desktop/voice/); notifications read and triaged the way the phone does; mouse, keyboard, screen and window access through the accessibility API; screenshot and screen-region zoom with the same grid and crosshair; and a sandboxed terminal (run_command) as one of its hands, with a policy file and a trash broker rather than a blanket shell. :core, a platform-neutral module. The brain loop, the four voice engines, the notebook's logic, the tool schemas, the emotion-ball renderer and the wire types moved out ofapp/so both bodies share one implementation.BrainHostis the interface a body implements; the brain never sees an Android or a desktop type. A golden-text test pins the model-facing strings before the split, so a refactor that changes what the model reads fails loudly.- Friends (
core/peer/). Two Andees pair by QR code — one-time, two minutes, three tries — and the code carries the showing device's key fingerprint, so the scanner pins that key and no other. Every link afterwards is a signed handshake against pinned keys with forward-secret session keys; replayed, altered or reordered frames drop the link. Two kinds, chosen by whoever shows the code: my device (another of your own; its requests run as your own) and friend (somebody else's — the owner is asked first, five minutes then it counts as no answer, and the work runs overGuestHostwith no memory, no history, no vault, no personal data and no terminal: screen, ask, show a page, open an app or a link, and nothing else). Off until switched on; discovery is mDNS plus the last address that worked. device.permissions.requeston the Mac — a permission concierge that asks for what a task needs and explains the grant in System Settings when macOS will not show the prompt itself.- Long jobs handed to the clock on both bodies, so a slow task stops holding the conversation.
Changed
- CodeFlying is the default backend on a fresh install. The previous rule hid the CodeFlying tab entirely unless the build was configured for it, and started on-device. Now the tab always leads, is what a fresh install starts on, and says the same thing with or without a key: what it is, how to get one, and a tappable link. The key row is always visible with show/hide and copy, and it displays the key in effect — saving it back unchanged stores nothing, so a build's key is not pinned as an override. A build with no server configured, like the ones attached here, has nowhere to send and says so until a key is typed or another backend is picked.
- The Mac and the phone now describe themselves the same way. The READMEs, this file and
docs/permissions.mdsay the project is two bodies on one brain, and the permission tables gain a row for what a friend's request can and cannot reach. - The Mac reads CodeFlying's build settings from
local.properties, like the phone — a generated properties resource rather than aBuildConfig, because a JVM app has none. - The product version comes from the phone's
versionName, so both bodies carry one number. The bundle version macOS requires of a.dmgstays1.0.0and is not the product version.
Fixed
- The ball stops repainting a still window, and the GPU draws it again — a Mac idle-cost fix that also removed a thread leak.
- A tap-to-talk path on the Mac where the microphone opened but nothing listened.
- A long task no longer empties the history, a new request interrupts the old one instead of queueing behind it, and voice ends on silence rather than on a timeout.
- Words typed or spoken over a question card answer it instead of starting a new turn.
- A notebook save that did not happen fails instead of answering "Saved."
- A todo set from a notification stays untrusted when it fires, the same rule its creation had.
- Silent reviews no longer ask for thinking, which was spending reasoning tokens on a turn nobody reads.
[0.5.3] - 2026-10-22
You can now tell it to stop, out loud, and it will. "Exit Andee" / "Take a break" (and their Chinese equivalents) ends the current task and closes the assistant — microphone and wake word included — with no model call in the path. The rest of the delta is aiming: a large fix for apps that hand accessibility a blank root, and a documentation pass on the grid it aims with.
The version number is a deliberate departure from the policy below. The policy says MINOR for a new
capability and this delta has one, which would make it 0.6.0; it is 0.5.3 because the capability is
an escape hatch rather than a new way of working, and the two things it is most likely to be confused
with — asking how to exit, or cancelling a subscription — are deliberately not commands. Recording it
here rather than leaving a feat under a patch for a reader to wonder at.
This release carries work from a teammate. The voice-exit feature arrived as PR #14 from @lirongkuafuai, closing issue #13 — fifteen commits, reviewed and merged rather than imported wholesale.
Added
- Voice exit: say it and it stops (
asr/VoiceExitCommand.kt). Recognized on the device, without a model call — an utterance that matches is handled before any card answer or brain submission, so it works even when the model is unreachable. It cancels the running task, stops the service and closes the assistant, which also stops the microphone and the wake word; the app icon starts it again. The match is whole-utterance and user-initiated only: "how do I exit Andee", "cancel my subscription" and a quoted command all stay ordinary turns, and a phrase that arrives on an automatic follow-up window is ignored (that one has its own test). While a meeting is recording it refuses and says why — end the meeting and save the minutes first. Teardown is guarded so a late transcript, wake callback or TTS fallback cannot reopen an overlay after the exit was asked for. popup_open/popupson tap, long-press and screenshot results, read from the window list, which survives a withheld tree. The model could not tell an open menu from a closed one by eye and reopened one six times; a blind tap aimed inside a popup is now refused, with the reason.
Fixed
- Aiming on an app that withholds its tree (measured on WeChat 8.0.79 / HONOR, where accessibility
gets a blank root and every tap is aimed by eye off a screenshot downscaled to ~0.48×). The injection
itself was already exact — the misses were all aiming, and this is a large fix:
- The grid labels moved to the edges. Printed at each cell's centre, the
Hcolumn's label sat exactly on WeChat's 「··」 button and hid it. They are now rulers — letters along the top, numbers down the left — and nothing printed inside a cell belongs to the device. - A gesture on a tree-blind app must fall inside a zoom taken since the last one. Measured: zoom-aimed taps changed the screen 33/41 times, screenshot-aimed 124/211, and the misses opened profiles, followed people and deleted comments.
- A cell is now fifteen parts, not nine (
3 × 5:top-left…bottom-right). The old nine names still parse as aliases mapped to the outer rows, so anything written against the previous vocabulary aims where it used to. - The markers are magenta (
#E000FF), not red — red is every unread badge in every Chinese app, and a red crosshair beside WeChat's red dot read as part of the app. The caption also moved above the ring, because below is where the next target is. - Prompt side: close a small menu by tapping its opener again rather than with
back(the menu is unfocused, sobackleft Moments entirely), and keep list progress instead of restarting at item 1.
- The grid labels moved to the edges. Printed at each cell's centre, the
Changed
- The READMEs' description of aiming was corrected. They said it taps by "which third of it" and that the crosshair is red; both stopped being true in this release. They now say "which part of it" and magenta. Found by re-reading the sentence against the code, not by a search.
- The English README's language badge is back to ASCII. A web edit on 2026-10-21 changed the badge
label from
Chineseto the Chinese characters for it, which put Han characters back into a file that is checked for zero of them. The link target was never in question and did not move. If a Chinese label on the English page is actually wanted, that is a decision to make deliberately — the badge is the one place a reader who cannot read English has to be able to see themselves — and not something to arrive at by accident.
[0.5.2] - 2026-10-19
Two fixes from watching it work on an app that hides its element tree, and a new hero image. The
fixes are the substance: commenting on a WeChat Moments post took 64 steps and 341 seconds, and the
same task now takes 20 steps and 58 seconds, sent on the first tap. The version number is 0.5.2
rather than 0.6.0 because the one feat commit in this delta is the README's cover art, not a
capability — nothing a user can do changed by it.
Fixed
- A swipe no longer coasts past what the model aimed at.
swipe_by_coordinateslifted the finger while it was still moving, so the list flung several screens beyond the target and the model lost its place — the largest single cost in the 64-step measurement. The stroke now holds at its end point before lifting (SWIPE_HOLD_MS), andACTION_CANCELends a scroll without a fling, so the list moves about as far as the finger did and the distance becomes the control. A newflingparameter (defaultfalse) keeps the old behaviour for long jumps, and its tool text says to take a fresh look afterwards rather than assume where it stopped. - The tap right after blind typing is refused unless it came from a zoom. On an app that withholds
its tree, the reply box is a thin strip directly on top of the post below it, so a grid-cell tap aimed
at it landed on that post and opened it — twice out of two on the same Moments reply box. After a
type_texton such an app, only a tap carryingon="zoom"is accepted, with an error that says why; leaving the box clears the rule. Zoom-then-tap is how the same comment now sends on the first attempt. submit_inputno longer reports success on a box that cannot be submitted that way. WeChat's comment box declaresIME_FLAG_NO_ENTER_ACTIONand sends only from its own button; firing the IME action there answered "done" while sending nothing. It now says it cannot, and names the app's own send button.type_textandsubmit_inputalso work when the tree has no focus node but the input connection is bound.- The tap ring is drawn where the tap landed. The marker window is positioned in display pixels, but
an API 30+ overlay fits the system-bar insets by default, so its frame began below the status bar and
every ring sat that far under the real tap. It now carries the same
FLAG_LAYOUT_IN_SCREEN/fitInsetsTypes = 0/ cutout mode the ball window already had. The caption's 0-1000 numbers were always right.
Changed
- The tool text now corrects three things models kept getting wrong.
zoom_screen_region'sx/y/w/hare 0-1000 of the whole screen, never pixels of an image — a screenshot arrives downscaled (often ~576 px wide as the model sees it), and every zoom now reports the rectangle it actually showed so a model can notice it was in the wrong space.type_text's text states that a box which opened with the keyboard is already focused and should be typed into without a preliminary tap.take_screenshot's text states that the assistant's own ball is never in a capture — the device hides its overlays for the instant of capture — because a model that saw a round blue shape blamed it for covering a button and tried to route around it. - The README's first image is now a phone with the character sitting on it rather than a floating
head.
images/brand/andee-device.pngis the new hero art andtools/make_social_preview.pyplaces it at 720 px wide, bottom-right, keeping 60 px of white padding with a 40 px analytic edge ramp; the mood strip narrows to 512/88 so the phone's left corner clears the last cell, and the text-to-art gap check now measures the character's left edge instead of a hard-coded offset. All four generated images were regenerated; the round avatar is byte-identical, because it still uses the head-only asset. The character in this render has a body; the app still draws a floating sphere.
[0.5.1] - 2026-10-19
Two ways the assistant could quietly lose the ability to keep working, both fixed. One, its
nightly review died with the process and never came back; the other, Xiaomi's power policy was
reaping it in the background with no way for the user to see why. The version number is
a deliberate departure from the policy below: both commits are feat, and the policy would have
made this 0.6.0, but neither adds a capability — they restore two things that were already supposed
to happen, and a user who upgrades is not getting a new feature so much as getting the existing ones
to survive a night. Nothing in the wire contract or an existing configuration breaks.
Fixed
- The quiet-hour review survives the process being killed, and is re-armed after a reboot. The
sweep is an
AlarmManageralarm, so it outlives the process — but when it landed there was often no service left to run it, and a reboot did not re-arm it at all. Worse, the review reads the conversation from memory, which dies with the process, so even a recovered sweep found "0 messages to review" and did nothing. Two halves:Notebook.last_sweep_at, written only by the sweep itself, is the one fact that separates "already reviewed" from "died before it could"; andrebuildHistoryFromScrollbackrebuilds the dialogue from the persisted scrollback — user and assistant rows, each re-stamped with its own clock — when the in-memory history is empty. On start the service catches up: it waits out a quiet window that has not finished, skips a stretch that was already swept, and otherwise runs the review it missed. Measured on a device: a sweep after a restart listed the four prior turns and completed. - A Xiaomi-family device now has a self-check row for auto-start, which is what actually keeps it
alive there. MIUI reaps background processes on a timer (measured at one kill every 40–60 minutes
on the maintainer's tablet), and without the auto-start grant
START_STICKYcannot bring the process back — the symptom a user reports is "it closes by itself after a while", with nothing anywhere saying why. The new row reads the vendor AppOps op (OP_AUTO_START = 10008, by reflection, since it has a number but no name this app can spell) and offers a one-tap door to MIUI's auto-start manager, probing the known components in order because MIUI has renamed that activity across versions. It isWARN, notFAIL: with auto-start off the app is fully alive right now — it is the future that degrades, which is exactly whatWARNis for. The row only exists on Xiaomi, Redmi and POCO (matched on brand words, since the MIUI build property would need a hidden API and the words are free), because on every other ROM it would be a switch that does nothing.Fix.Screengained an optionalcomponentfor the vendor screen that has no action to name.
Changed
CLAUDE.md's self-check section no longer claims a count. It said "Seven things in the design are load-bearing" and had listed fifteen for some releases; the sentence now says "a number of" instead, on the same reasoning the versioning policy above records — a count written into prose goes stale, and this one already had.
[0.5.0] - 2026-10-16
A finished conversation is now something the device puts away and can read back. That is the new capability this release is numbered for — a stretch of talking that has ended stops costing context on every later request, and stops being lost with it. Two smaller changes ride along: a choice card's question and answer now land in the scrollback, and the README explains what signing is actually protecting.
Added
- Topics: a finished stretch of conversation is folded away and can be picked back up. A topic that
has ended was the most expensive thing in every later request and the least likely to be needed
again. The nightly review now marks where topics begin and end among the user's numbered turns
(
fold_topic), the device validates the range and swaps it for a one-line marker after the turn it names, and the verbatim conversation is archived so it can be read back. Three tools, alllocalOnly:fold_topic(review-only — an ordinary turn that calls it is refused, because folding is irreversible from the model's side and a request is not the moment to spend it),list_topics(id, title, when it ended, how many messages), andget_topic(the whole back-and-forth as plain text, truncated past a limit withtruncatedsaying so). The device refuses a fold with a reason — a turn number that is not in the list, a range that ends at the latest turn (nothing has started after it, so nothing says it ended), something inside still waiting on an answer, an overlap with a topic already put away, too short, or a scene active — and the reason names the condition rather than inviting a retry. At most three topics per review, and they are put away when the review ends rather than when the tool is called. In the scrollback a folded topic is a card standing where its messages stood — the summary shown by default, with Continue, Transcript and Made here. Continuing inlines the archived conversation into the next turn; the transcript opens fullscreen (and says so honestly when the text has been cleared and only the summary is left); Made here reopens the pages produced inside that topic, each with its version. The summary's fourth line is deliberately what the model left out, because a summary that reads complete is one whose details are lost —list_topics' own description tells the model never to answer a question about a folded topic from its one-liner alone, and never to say it did not happen.debug.sweep(loopback only) runs the review on demand for testing.
Fixed
- A choice card's question and answer now land in the scrollback.
ui.askspoke the question and handed the chosen option straight to the brain, so the card came and went and the conversation showed the brain acting on a choice nobody could see being made. The question is an Andee row now and the answer — the button, or the words said instead — is a row of the user's. A timeout writes nothing, because nothing was chosen.
Changed
- The README says what signing is for. The signature is the only thing stopping a fake "update"
from installing over a user's copy and inheriting the accessibility service, the secure-settings
grant and everything in the vault. That is now stated where signing is already discussed, along with
why a per-machine debug key is what produces
INSTALL_FAILED_UPDATE_INCOMPATIBLE. - The tool count in the READMEs and the DISCLAIMER is no longer a number. Both said "sixty-odd tools", which was accurate when written and had drifted to seventy-two. The sentences now describe what the tools do and say "a few dozen", because a count in prose is a count that goes stale — the same lesson the versioning policy above already learned once.
[0.4.1] - 2026-10-15
A day of fixes and one feature pair, and the version number is a deliberate departure from the
policy below. The delta holds three feat commits — page versioning, the brain being able to see
its own pages, and the default palette — and the policy says MINOR for a new capability, which would
have made this 0.5.0. It is 0.4.1 because the first two are two halves of repairing one
incomplete feature rather than a new capability: show_html could already make a page and the panel
could already list it, and what was missing was the model's knowledge of it and the device's
willingness to treat a change as a change. Nothing in the wire contract or an existing configuration
breaks either way, so neither number is wrong about compatibility — this one is about intent.
Added
- Revising a page makes a new version of it, not a second page. Changing a page used to archive
the result as a brand-new artifact, so two rounds of revision left three near-identical entries on
the Made for you shelf with nothing saying which one was current. An entry now carries which
artifact it belongs to and which revision of it, the panel shows one row per artifact with its
version number (
v3), and the older revisions stay on disk behind it: Versions opens the list, any of them can be opened, and Go back to this makes an old one current again — as a new version, so the history stays append-only and the rollback itself can be rolled back. The brain is told the same thing, so a request to change an earlier page reads the current source back and files its answer as the next version instead of rebuilding a lookalike beside it.show_htmlgained an optionalrevisefor naming the page explicitly; the device also infers it from theget_pagein the same turn, and from the panel row the user tapped. One artifact keeps at most 20 revisions. Which artifact a revision belongs to is decided on the device and never asked of the model — and the target is consumed rather than merely read, a distinction found by hand on a device, where a target left standing folded a second, unrelated page in as a later version of the first. - The brain can see the pages it has made.
show_htmldraws a page and the device keeps it — the Made for you panel lists them, so to the user they are a shelf — but nothing put that record into the model's context, so a page built yesterday did not exist to the brain that built it. A request to revise earlier work could come back as a fresh table that merely resembled the old one, or as a flat "I have not made that". Two read-only tools close it:list_pagesreturns every page with its id, title, date, size and whether the file is still there, andget_pagereads one back by id so a revision is a revision rather than a new page wearing the old name. Both are deliberately outside thescreen.prefix, so neither folds the card nor stops at the lock screen: reading a file reaches neither. Alongside them the newest few ride on the per-turn user message as a short index — the same place and the same reason as the notebook index, since a tool only answers a question the model thought to ask.available: falseon a page whose file the cap has taken is reported honestly rather than passed over: the title and date are still true, the contents are not there to read.
Changed
- The default appearance is light, not dark. A fresh install now comes up with the light palette.
A device that has touched the setting keeps whatever it chose — the rule reads only an explicit
night, so nothing here can overrule a user who picked dark, and the dark palette stays one tap away in Settings → Appearance. Written this way round on purpose: the test is against the non-default value, which is what makes changing the default later a one-line edit rather than a migration. Dark is not the older palette and is not being corrected — it is the look this app was designed on, and light was derived from it as the inversion.Theme.night's initial value moves with the default, so the frame drawn before the saved setting is read is already the right palette rather than flipping one frame in. - Every face got its props, its idle life and its finishing touches. Andee's mouth dropped twice
—
−0.185 → −0.235 → −0.285— so there is a finger of sphere between its eyes and its smile instead of the two crowding together. Sleep stopped being a flattening pose: thesqx/sqythat squashed the resting and sleeping ball into a puddle are gone, and the floppy horns, the ⌣ eyes and (while sleeping) the bubble carry it instead.
Fixed
- Changing the model, the thinking switch or the reasoning effort no longer wipes the
conversation. Closing the settings card rebuilt the local brain on any
BrainConfigdifference, and a rebuild is the loss of the whole conversation — flipping the thinking switch reset the history mid-task, and the model then said it had no record of a page it had just built. Only a change of who answers (mode, key, endpoint) rebuilds now; model, thinking and reasoning effort are request parameters and are retuned in place, taking effect on the next request.
[0.4.0] - 2026-10-14
Two new capabilities, both of them choices rather than defaults: a second voice provider (Google, selectable beside the existing one, contributed as PR #8) and a launcher icon that is the character's thinking pose. Behind them, three fixes to how the interface is drawn — the primary button, the type scale on a phone, and a tracker string that had been calling every new device a Volcengine device.
Added
- A second voice provider.
VoiceConfig.voiceProviderreadshuoshanorgoogle, the settings card grew a Voice provider picker above the two credential blocks, andAsrController/TtsControllernow build one of two engines behind a shared interface (AsrEngine,TtsEngine); the two are streaming on both sides and reduce to the same five calls, so nothing above them changed. Volcengine stays the fallback — anything that is not exactlygooglereads as Volcengine, and it is the leftmost option — because for every device that predates this row, Volcengine is what it has been running. Google's two halves are different protocols: recognition runs over the Gemini Live API (a WebSocket, so it needs the project number and a location) and synthesis is plain HTTP against the same Vertex Express key. Its defaults are a real region and two model names rather than a blank (asia-northeast1,gemini-3.5-transcribe-live-preview,gemini-3.1-flash-tts-preview), and every hint on its rows is the value a blank row resolves to — a card that advertises a model but silently calls another is lying in the one place the user goes to decide.GOOGLE_VOICE_KEYnon-empty at package time is also what makes Google the default provider. That is deliberate: a build is packaged for whoever it is packaged for, and baking a key is the most explicit statement of that intent there is. The repository's own source defaults to the empty string, so a checkout with nothing configured builds exactly what it built before Google existed. The self-check moved with it, and the move is the interesting part:voiceKeynow asks for the key of the provider actually in use, andvoiceReachprobes that provider's host. The rows name neither — a green tick earned by the Volcengine key while the device is talking to Google is the exact lie that having one checklist instead of two exists to prevent. (Contributed by @kuafu-limanxiang in #8.) - The launcher icon is now the thinking pose, on white. It was built from a sheet with a glow
painted around the whole head and composited over that glow; it is now built from the cut-out of the
character with a hand on its chin, on a plain
#ffffff. Same character either way, so nothing about the identity changes — the icon is now the pose the app is in when it is thinking, and the background is literally white because these illustrations were drawn on white and the warm light on the ball's edge is painted on the ball itself. The inferred-glow treatment is not deleted; it is one flag away (--bg glow), because it is real work that came close. The icon set also came out smaller, 405.4 → 258.3 KiB.
Changed
- The primary button is one flat fill instead of a blue-to-violet gradient. A gradient inside a
pill makes the button carry its own light source — the look of a landing page's call to action —
and it was the only colour in the whole palette that did not answer to the theme, so on a light card
it sat there like a sticker off another design. It is now
Palette.tint, a step deeper than each theme's own accent, which keeps the button the loudest thing on a card for the reason that survives going flat: it is the only saturated fill on it. Violet went with the gradient — the interface has exactly one hue and it is blue.Glass.TINT_STARTandGlass.TINT_ENDare gone. The label had to move with the fill, and that is measurable rather than taste: it wasGlass.TITLE, near-black by day and white at night, which works on a saturated ground at night and measures about 4:1 by day — under the line for body text. It isGlass.ON_TINTnow, white in both themes, because both fills were picked deep enough to carry it. Seven call sites, one token each; no other surface changes. - Type is sized against the device now, not just against the tablet. The six rungs were drawn on a
711dp tablet where a card is 500dp wide with air around it. A phone is 392dp and the same card fills
it edge to edge, so the same 26sp title read as a heading on one and a shout on the other — on the
phone it showed up as a row title clipped to "Accessibilit…" with the next card pushed off screen
entirely. Nothing about the type had grown: the box framing it lost a third of its width while
padding, radii and line spacing stayed in dp, and that mismatch is what reads as clumsy rather than
as large. Six rungs now carry a second column about three quarters of the first, chosen once from the
device's short side against Android's own 600dp line (
sw600dp) — the same line, not a second one. The tablet column is the original design and is untouched; a phone gets 20 / 15.5 / 13.5 / 12 / 10.5 / 9.5. MICRO moves least on purpose: it is already the floor of what can be read, and a floor that moves with the screen is not a floor. The system font scale is deliberately not read — a phone with text turned up still renders these at that scale, and the phone column is set with that in mind. - Every English-facing markdown file is now free of Chinese characters. The README had to be; the
other fourteen did not, and they had accumulated UI labels, language badges (the Chinese badge text
became
Chinese, with its link target unchanged), and one piece of shorthand. The rule is mechanical now rather than a habit, and it is worth knowing what it deliberately excludes: code identifiers, code comments andCLAUDE.mdare out of scope — the last one is written for AI assistants.
Fixed
- The voice self-check named a provider the device might not be using.
voiceKeyasked forapiKeyandvoiceReachprobedopenspeech.bytedance.com, unconditionally — correct while there was one provider, and a wrong answer the moment there were two. Both follow the selected provider now.
[0.3.2] - 2026-10-14
Wake word plus polish, and the first release that ships the APK with the release rather than telling
the reader to build one — see Added and docs/releases/v0.3.2.md.
Added
- The wake word now listens with the screen off, and stops the tools that cannot work locked. The
gate used to hold the microphone only while the screen was on — a rule with a real reason behind it,
since a dark tablet on a bedside table should not be listening all night, but the screen being off is
also the normal state of a phone in a pocket, so it meant the assistant could not be called from the
one place a person actually keeps their phone. The gate now asks
wakeListeningAllowed()— the screen is on, or the user left the "listen while locked" row on — and that setting defaults to on. Its cost is stated on the row that flips it rather than left for the status bar to mention: the microphone stays open, and so does the system's recording indicator. Woken with the keyguard up, Andee can hear and answer —tts.speakneeds nothing from the screen — but it must not pretend it can act. A gesture aimed at a locked device is dropped by the input pipeline and reports nothing; the lock screen's ownFLAG_SECUREmeans there were never any logs to search;screen.ui_treereturns the keyguard's empty placeholder, indistinguishable from the blank-tree case;device.launchopens an Activity behind the lock screen where nobody sees it. Every one of those would come back to the brain as "completed" and be repeated to the user as something that happened, sodispatch()refuses them outright with a sentence saying why and what to do instead.screen.is blocked by prefix so a new tool in that namespace inherits the guard; the names outside it —device.launch,open_url,scan,look,permissions.request,ui.ask,ui.alert,ui.show_html— are the same act by another road.device.battery,device.volume,tts.speakand the wholenote.namespace stay open, because they work perfectly well locked. Read live rather than mirrored into a flag, because the caller that matters fires at any point of a turn: locking the phone mid-answer is a normal thing to do, not a state transition to track. Verified on the tablet over the 9008 tool port, including the device being unlocked again to prove the guard reads the device rather than a memory of it. - The voice segmenter now starts on the voice, not on the microphone. Two symptoms reported
separately, neither with a log line to explain it — "it takes three tries to enrol, but if I wait
three seconds one is enough", and "it wakes on everything" — turned out to be one defect with two
faces.
VoiceSegmenterdecides an utterance has started when the level holds forSTART_FRAMESframes, three of them, 30 ms, and the floor is seeded from the first frame the microphone delivers. Three frames is a click, a tap on the tablet's edge, or the recorder's own first read, so the segment started on the hardware rather than on the person. A segment that starts early comes out long and dirty and, pastMAX_FRAMES, was dropped in silence — the caller's latch never counted down and the card sat on "listening…" for the rest of the window. That is the second and third try. The fix slows the decision down without throwing audio away:START_FRAMES3 → 8 (80 ms) so a transient cannot open a segment, andPREROLL_FRAMES15 → 30 (300 ms) so the onset is still in the ring when the decision lands. Delaying a decision is free; discarding the audio is not — which is why there is deliberately no warm-up that drops the first 200 ms. The obvious fix was proposed and withdrawn: a user who taps and speaks immediately would lose exactly the syllable the ring exists to protect. - Five of the character's faces now run along the bottom of both cover images
(
images/cover*.png,images/social-preview*.png), one word under each — Listen, Look, Think, Act, Speak. They are the app's own illustrations rather than drawings made for the web, and they are the gesture moods, the set the web renderer cannot draw because they use props — so this is the first place the cover shows the character instead of describing it. The row is confined to the left column by the head's own geometry, which is why the whole column moved up 60 px rather than the band being full width.
Changed
- The APK now ships with the release. Every release up to
0.3.1said "no binary, on purpose", and the reason has not changed — the release build type has nosigningConfigand the repository has no keystore, soassembleReleaseemits an unsigned APK that will not install. What changed is the trade: a reader who cannot install anything has no way to try the app at all short of provisioning an Android SDK. The attached APK is the debug build — signed with the standard Android debug certificate and installs as-is — and it is deliberately bare:CODEFLYING_ENABLEDisfalsewith bothCODEFLYING_KEYandCODEFLYING_DOMAINempty,DEFAULT_LANGisen, and the thinking default is off. All four were read back out of the generatedBuildConfigrather than assumed, and the artifacts were scanned for key material. The cost is stated rather than buried: a debug APK isdebuggableand is signed with the shared Android debug key, so it cannot install over a release-signed build — moving from this to a signed release takes an uninstall, and an uninstall takes the notebook and the vault with it. The READMEs' advice (hand out a signed release build, not the debug one) is still right for anything actually distributed, and both now point at this stopgap so the two do not read as a contradiction. - The wake-word prose moved with the default.
wake_done,check_detail_wake_onandsettings_wake_enrolledall said the wake word works "while the screen is on" — in both languages — and both READMEs said "it only listens while the screen is on, screen off hands the microphone back". The default is the thing that changed, so every sentence describing it had to move with it or it would have shipped stale.
Fixed
- The enrolment prompt appeared before the microphone existed. The "now speak" prompt was set on the tapper's
thread, before a recorder existed, so the user was invited to speak into a microphone that had not
started. It now moves after
mic.start()returns: the card reads "listening…" for the 100–300 ms the recorder takes to come up, and the prompt appears once something is listening. Nothing is taken from the user to pay for it. - The wake threshold was anchored on the worst take rather than the closest.
thresholdForwas anchored on the worst pair, so one dirty take set the door width for the whole system — 2.2755 and 2.1083 against a measured band of non-wake speech at 2.21–2.59, with sixteen of sixty judgements within 0.06 of the line.bestDistancereturns the nearest template, so the question the door width answers is "how far can a real attempt be from the nearest take" — and the evidence for that is the takes that landed close, not the one that did not.MAX_THRESHOLD2.8 → 1.8, which had been picked as "the widest thing still callable a threshold" and never measured. Across three enrolments twenty minutes apart the spread between takes fell from 0.42 to 0.12. VoiceSegmentergained a pure-KotlinStats(frames / loud / firstLoud / peak / floor / segments / droppedShort / droppedLong) with no Android dependency, keeping the DSP half of the wake pipeline JVM-testable;WakeEnrollUilogs it once per attempt andWakeMiclogs the read that ends a capture early, both under theWaketag so one filter sees the whole pipeline.
[0.3.1] - 2026-10-13
A patch release by decision, against this file's own definition. The delta carries six feat
commits and the versioning policy above says a new capability is MINOR; the maintainer chose PATCH
deliberately. It is written down here rather than left for someone to reconstruct from the history,
because a version number that disagrees with the policy is exactly the kind of thing that reads as an
oversight later.
Added
- A day palette, alongside the night one. The card had been one look — a translucent pane over a
blurred photograph, a bright hairline, pale type — and half the requests on top of it wanted the
opposite of those values. What blocked every one of them was that
Glass's colours wereconst vals: inlined into their readers at compile time, and so unable to vary at runtime at all. They are getters over a day and a nightPalettenow, every name unchanged, which is what let forty call sites stay where they are. Night is what the app shipped with and is the default; day is the inversion of it, with three values deliberately left alone because they name a position rather than a colour — the well stays recessed, the thumb stays raised, and the sheen flips end for end, because its job is a lit edge and the light does not come from below on either palette. A theme is a process-global with one listener, shaped afterAppLocalefor the same reason: almost no UI here lives in an Activity, so no system mechanism could reach a Service's overlays. The backdrop is the same photograph, blurred once and veiled in whichever direction the palette needs — black down to 0.22 at night (the original composite to the digit) and white up to 0.62 by day. One image, two veils, no asset change. Four faults presented as a single element looking wrong and are the reason this is one commit:CardUi's colours lived in a companion object and were evaluated once at class load, so the card kept whichever palette the process saw first for the life of the process; the window's own chrome is built once and never hears about a change, which is whatapplyPaletteToChromeis for; the row tints were cached; and the light card's dim greys could not survive being asked to sit on a photograph. - A "Looks & expressions" card in Settings, showing every look and every expression on a ball of its own. A look chip is saved on tap; an expression chip only previews.
- A generated page picks its own look. Every page came back the same dark slate, and the reason
was not taste — it was a HOUSE STYLE block in the
show_htmldescription, ending in "use these exact values, no others". Asking for a light page worked only because the block left itself a way out, and that is a lever the user had to pull every single time. It could not have been fixed by wiring the block to the theme: a tool description is a string and cannot branch onTheme.night, so on a device switched to day the cards, the ball and the backdrop all went light and the one surface the model composes stayed in the night palette. That fix was on the table and was turned down on purpose — a page's look is the model's business, not ours. Two things stay, and neither is a matter of taste: the<meta viewport>, becauseHtmlActivitysetsuseWideViewPortand a page without one is laid out at desktop width and scaled down, and the ~110×44 px the Done pill occupies in the top-right corner, because that pill is ours. The cost is on the record: for the small models this runs against, that description was the only lever on page quality, and nothing enforces any more that a generated page reads as part of Andee. - The local LLM endpoint and model follow the UI language. The on-device tab's two LLM rows already had
a hint that doubled as the value a blank row resolves to, and both were hardcoded to DeepSeek —
self-consistent for a Chinese card and wrong for an English one. The cheapest possible fix, changing
only the hint text, is the worst version of it: the box would advertise
https://api.openai.com/v1and then, on a blank row, callhttps://api.deepseek.com. The hint is read exactly when a user is deciding what to type, so a hint that lies about the fallback lies in the one place it will be believed. Hint and fallback now come from one pair of resolvers keyed offuiLanguage: an English card advertises and callshttps://api.openai.com/v1+gpt-4o-mini, a Chinese card stays onhttps://api.deepseek.com+deepseek-flash. CodeFlying is deliberately not switched by language, because its model belongs to the hosted backend and an English build inventinggpt-4o-minithere would name a model CodeFlying does not serve. - Errand trips from the check card are handed to the page instead of made by it, so the list
survives the trip. On the check page tapping a row's fix button sent the user to Settings and coming back
landed on the list; from the check card the same three taps ended on the desktop with a ball in the
corner, because the card was closed and the assistant folded before the destination launched — which
does clear the screen, since an overlay outranks every Activity, but throws the list away: a window
gets no
onResume, and the destination went out withFLAG_ACTIVITY_NEW_TASKinto a task of its own, so back from Settings had nowhere of ours to return to.SelfCheckUi.onErrand(wasonYieldScreen) now hands the host the Intent and lets the host say whether it can carry it.Fix.Grantgoes through the same door, so Allow on the microphone and notification rows stops ending on the launcher too. - The prop-and-pose layer is on every look now, not just Andee — stars, a heart, a shield, a sweat bead, the sliced lid, and the idle life that lets a look wear its own expressions and nap while nothing is happening. The argument for keeping it to one look was that the others had been tuned without it; that is overruled, because a look that is not alive standing next to one that is reads as the broken one. Switching it on for Imp needed a fix nothing at the call site would have hinted at: the lid's eye mesh was cut at a fixed 0.38 and Imp ships a resting eye already cut at 0.48, so an angry imp would have opened its eyes, in two moods, silently.
- Andee's horns stand higher, fade to a peach tip and hang on two springs, so a wiggle or a turn swings them and they ring out; the mood sets where they rest. While idle it now wears a short run of safe faces and naps for under a minute instead of sleeping until touched — faces that report what the device is doing, or that read as an alarm, are left out on purpose. Hands are a round palm, a thumb and two capsule fingers, and the wink carries a floating heart instead of a hand, because a V, a pointing finger and a thumbs-up were each tried and each misread at this size.
Changed
- Both READMEs lead with what it is. They opened by showing what the thing does under a headline that framed it as a phone upgrade. The H3, the hero and the lead now say what it is and what it can do second, and the "why it is not the assistant already in your phone" section moved above the examples, which are now explicitly what its hands look like. Two earlier drafts were self-defeating and both were caught by reading them: one claimed "not a tool, not a companion" and then proved it with a capability list, the other denied being an assistant that operates your phone while the whole section below was exactly that. Neither is fixable with an adjective; the negations are now scoped to a category, and the section opens with the axis that separates the two — its own schedule, notebook, judgement about when to speak, and limits, all of which happen in the time you did not ask for anything.
- The English README no longer contains a single Chinese character (
\p{Han}returns nothing): the language badge, five look names, a dozen Chinese-language links, the language markers on the CodeFlying sites and a button quoted in Chinese. The Chinese README keeps its own language.
Fixed
POST_NOTIFICATIONSwas declared in the manifest since the beginning and never requested — never checked, never asked for, so on Android 13+ everything this app posts was dropped in silence: the persistent notification that is the way back to it, and every reminderSchedulerfires. The codebase half knew:Schedulercarries a comment blaming this permission for its own failure to notify, and nothing anywhere tested the claim. The row that looked like it covered this does not — The notification-access row is the notification listener, which can be granted and green while every notification the app sends disappears. Two things this needed beyond a catalog entry:Perm.api, because on API 32 and below the permission does not exist andcheckSelfPermissionanswers denied for anything it does not know, so an ungated row would report a missing permission on every Android 12 device with no button and no way to fix it; andSOLO_IDS, because that set is a whitelist and a permission added by someone who never read it routes to handoff.- Three documents still described things the code had stopped doing. The data-flow table in both
DISCLAIMERs and bothpermissionsdocuments namedapi.deepseek.comas the default destination for the conversation, with no mention of the language dependency added in this release — so an English reader was told their data goes somewhere it does not.CLAUDE.mdstill called theshow_htmlhouse style "the one deliberate exception" to the single-palette rule, two paragraphs after the same file recorded that the house style had been removed; there is no exception now. And thePOST_NOTIFICATIONSrow in the permission table said only that a reminder would not arrive, which undersold it: refusing also kills the notification that is the route back to screen control.
Removed
- The raised mitts are gone. Andee's expressions were built around a pair of mitts raised at the
side — a fist under the chin for thinking, two fists up for celebrating, a hand on the belly for
resting — and every one of them read as a lump at the eighty dp the ball is actually drawn at, with
the parts sharing the body's own colour and material. Dropping them is the honest version of "this
does not work at this size": the ball keeps the paws it hangs off the screen edge with. No expression
collapses into CALM for it, because a mood is read through its own
Facerow and its body motion, neither of which ever came from the pose table. Two leftovers are deliberate and both say so where they live: WAVE keeps its fan of motion strokes with no hand to fan off them, and HOLDING keeps an empty row, because the signboard it waits behind is a real card in the ball's own window. - The
show_htmlhouse style, along with theLocalPrompt§7 sentence that pointed at it and then asserted "that style is what you look like". Both were describing the prescribed look being removed.
[0.3.0] - 2026-10-12
The release where the assistant stops living inside the accessibility service. Everything else here is smaller than that; the split is the reason it is a minor bump and the reason the upgrade notes matter, because a ROM that revokes accessibility now costs one capability instead of the whole app.
Added
- The body is a foreground service; the accessibility service is a shim. The body — ball, card,
microphone, ASR/TTS, brain socket, scrollback, todos — used to be constructed inside
ScreenBodyService.onServiceConnected(), and that class is anAccessibilityService. So a ROM that revokes accessibility (HyperOS, observed) took the whole assistant with it: no ball, no way to talk, no way back except knowing which switch in Settings to find. The failure mode was never "screen control is gone", it was "the app is gone".BodyCoreServicenow owns everything else and outlives the accessibility service, which is reduced to handing the screen capability over and taking it back — so a revoked service costs one capability, and the return trip is one tap on a notification that is still there.screen.*goes through a provider, so a call with no accessibility throws a readable error instead of no-op'ing or holding a controller built around a dead binding; the shot and gesture guards moved from aninitblock towireScreen(), called on every attach, because left on the previous controller they would simply be absent on the new one. The launch is a persistent notification that carries whether screen control is up and the way back to Settings — posted by the body, so it survives the state it describes. The launcher is the one context allowed to claim the microphone foreground-service type, because a microphone-typed service cannot be created from the background and every other starter is background. - The screen can be read with
MediaProjectionwhen accessibility is gone. The split stopped a revoked service from taking the whole assistant, except that the assistant went blind — every way of reading a screen went through the accessibility service, so "what is this screen?" was the one question it could not answer on exactly the devices this project exists for. Deliberately a fallback and deliberately weaker:screen.screenshotonly, becauselook_regionandzoom_aimexist to zoom a target for a tap anddispatchGesturehas no non-root substitute, so there is no tap on this path. The result says which way in it used, and carries no aim grid and no element list — handing the brain a picture it believes it can aim with would be worse than handing it fewer fields. Asked for only while it is the only option: the consent is a notification action, offered only when Andee is blind and has no capture, andACTION_ACCESSIBILITY_SETTINGSstays the first action because the projection buys reading, not acting. - Every mood has a
Poseon top of itsFacerow — a wink, eyes that look up and aside, lowered or worried lids, a squashed resting body, two puffy hands (open, fist, pointing) and pop-in props. Five gesture moods (wink, wave, protect, celebrate, rest) can be asked for by name or synonym in the model's own text, and the prompt says to keep them rare. Only a look withBallLook.gesturesreads it, so the other five characters are unchanged. There is deliberately no open mouth: cross-fading it with the thin smile drew both at once and one poked through the other. - The default ball is now the blue horned Andee — a cobalt head with curled horns and glassy
eyes, first in
BallLooks.ALL, so a fresh install and a factory reset both land on it. Its numbers were measured off the reference picture through the renderer's own camera rather than eyeballed. The launcher icon set, README covers, social previews and org avatar are regenerated from the same source art. - The CodeFlying per-user token is read from
assets/andee_key.txt, withBuildConfigas the fallback for dev builds. Production APKs are built once from shared source and the per-userkft_token is written into that asset at package time, so the base APK stays identical across users and the Gradle build cache no longer misses per user. -PdbgSuffix=builds a parallel package for device verification. The package on the tablet was signed with a key this checkout does not have, sopm install -rwas refused and the only alternative was uninstalling — which throws away chat history, notebook and vault. Off unless the property is passed, so ordinary builds are byte-identical.- Debug builds gain
BallLabActivity, which shows one mood on a plain background for screenshots.
Changed
- The version check moved to
https://api.heyandee.com, fromhttps://andee.kuafuai.net.VERSION_API_BASEis still a hardcoded constant indevice/SelfCheck.kt, and still the only address this app calls that the user did not configure — seedocs/permissions.md— so the move is one line of source plus the three places that quote it and have to stay in step: bothpermissionsdocuments and the note innetwork_security_config.xml. The old host was on the mainland, where the domain had to be ICP-filed before TLS would reach it at all; the new address resolves outside the mainland and retires that particular trap. It does not retire the other one: a domain still has to resolve and serve 443, and a fresh record is not everywhere at once. Measured while making the move,1.1.1.1still answeredNXDOMAINwhile223.5.5.5already had the address — on the wrong resolver that reads as "not found" on the version row, which is the intended failure rather than a fault to report. The service itself was checked before the switch: the new host answersGET /api/versionwith the same envelope, and the samedatafields the client parses. - The self-check page now stays up until the user closes it. It is a list, and a user with five rows works through them one at a time — out to a Settings screen and back for each — but the page used to hand the screen to the assistant as soon as it could draw one, which happens on the first of those returns, so the list was taken away after row one. On a page whose whole job is to walk somebody down a list, the user is the only one who knows when the list is finished. Closing it with the Done button puts the assistant back. Two surfaces fighting for the screen was the real bug underneath and it stays fixed by the same move: nothing unfolds a card while the page is up.
- Closing a page puts the assistant back at full size, unconditionally. The fold that makes room
for a page is ours — our card is a fullscreen opaque overlay and would cover it — so undoing it is
ours too, and it was missing from
pageClosed. The first cut restored the card only if it had been fullscreen when the page went up; the report was the other way round, and closing the self-check card drew the same complaint and got the same answer. - Both READMEs point their CodeFlying links at the
/andeeone-click landing page rather than the generic official site, so every jump lands on the Andee entry page.
Fixed
ui.show_htmlno longer needs accessibility, which it had started to.awaitUncovered()asked its question throughrequireScreen(), andawaitUncoveredruns offensureCompact()— whichui.show_htmlcalls too, because our card would cover the page it is about to draw. So with accessibility off, the one tool that draws a page answered "accessibility is off" and the assistant told the user it could not show them anything. Introduced byf364490, which turned a nullable-but-always-present field into a throwing provider and rewrote every call site mechanically; the other sixteen are insidescreen.*branches and are meant to throw. Returning early is also right on the merits: the wait exists so that a tap lands on the app rather than on the folded card, and there are no taps without accessibility.- The self-check no longer mistakes a re-creation for a launch. This page sends the user out
through its own rows, so the platform killing it while they are away and rebuilding it on return is
routine here rather than an edge case: grant the overlay, come back to a restarted Activity, and it
used to unfold the card and finish before a single row was drawn.
savedInstanceState != nullis what tells the two apart. A related report — "I granted it, came back, it still said no, then re-check was fine" — is the platform's answer about a setting changed seconds ago not being ready in the frame the user returns on: the notification-access row isOKonly once the listener is actually bound, and the read that finds it unbound is also the one that requests the rebind.settleRechecktakes two more local passes and stops, because a page that keeps re-reading itself is a page whose numbers stop meaning anything. (The retry is reasoned from the code path, not measured: writing the notification listener intoenabled_notification_listenersdoes not stick, so the state cannot be manufactured from adb.) - The body stays off the screen while the check page is up. The report was "the check said the
overlay was off, I granted it, came back, and I was on the assistant instead of the check" — the
check was never gone, the page was still the resumed Activity with the assistant's fullscreen card
painted over it. Two things put it there, both in the body's birth:
window.setCompact(false), andwindow.show(), which is subtler — this window is fullscreen even when the ball is compact and swallows the touches it does not use, so a page with a ball in the corner is a page whose buttons cannot be pressed (measured: the Done button did nothing). Telling a fresh launch apart from a return neededSelfCheck.isPageOpen, written down in the prefs file the startup check already uses. The startup self-check card is skipped while the page is up as well, because that page is the list. - A just-fixed row stops showing as red when the card resumes. The card's own rows send the user
out to a system screen and the only thing that happened on the way back was a language comparison,
so a row the user had just fixed stayed red until they found "check again" — and a stale red row reads as
"the fix did not take". An Activity gets a resume callback and an overlay does not, which is why
this is a change to one host rather than to the list; it is local-only on purpose, because the
probes hold the card open for seconds. The first resume is skipped (a full pass just ran), the
three probed rows keep their last reading (
SelfCheck.mergeLocal), andrenderrestores the scroll position — which only started to matter once redraws stopped being deliberate. - A perch slide dragged the whole card off-screen, and a voice failure repeated forever. A card held by the folded ball makes the window card-sized, so a perch slide (task start, page up) dragged the whole card into the bottom-right corner, half off-screen and untouchable; the perch entry points now take the signboard down first. Separately, the auto follow-up mic reopens after every answer, so a device with no voice set up raised the same failure card and scrollback row each time — now shown once per failure streak, reset by a working voice turn, leaving settings, or the user pressing the ball.
screen.screenshot's schema claimed something false. It opened "every screenshot arrives with an 8x12 AIM GRID drawn on it" and promised the element list, both exactly untrue on the projection path — which is the only screen tool that still works there. The runtimenotesaid so, but the schema is what the model reads before deciding to call it.- Three documents still described the pre-split architecture.
README.mdandREADME.zh-CN.mdsaid "this app is the accessibility service" and pointed a reader atScreenBodyService.ktas the file to read first — both false sincef364490, which is the whole point of the split. The permission table indocs/permissions.md/.zh-CN.mdsaid of accessibility that "every screen action errors", which stopped being true in82f08db:screen.screenshothas a fallback and the row now says so, along with what the fallback cannot do.
Removed
tools/check_launcher_icon.py,tools/cosmic_backdrop.pyand the oldmake_launcher_icon.py, replaced by a rewritten generator built ontools/andee_head.py. All three were written for the previous character and a star-field backdrop, neither of which is in use any more.tools/make_social_preview.pygenerates the social previews, both README covers and the org avatar from the same source art.
[0.2.0] - 2026-10-10
Three new capabilities and two permission fixes. A minor bump rather than a patch, because the three are new capabilities by this file's own definition — the first release where that rule has had to decide anything. The permission pair is the part worth reading: the app had been telling the model things about the system permission sheet that measuring on a device showed to be false, in one case in the direction that made the model act on a grant nobody had given.
Added
- A single tap on the ball now stops whatever is running and then listens. A tap used to mean
"listen" and nothing else, which left a gap the moment the ball was busy: to redirect it the user
had to stop it first (three taps) and then ask again (one more) — two gestures for one intent.
Tapping something busy and having it stop and listen is what one person does to another, and it is
the only reading of the tap that is not "talk over the thing you cannot hear". The stop is
stopEverything(), reused rather than reimplemented, so the single tap and the triple tap cannot drift into two different stops and the hub gets itstask.stopeither way. It is guarded ondispatcher.isTaskActive(): on an idle ball there is nothing to stop, and the call would still clear a pending "keep editing", tell the hub to stop, and flash "stopped" ahead of "listening…" — three side effects for a tap that only ever meant "listen". Triple-tap still stops, and is still the immediate stop, because a single tap has to wait out the double-tap window before it knows it is not the first of two. No gesture moved and no string changed: what a tap does depends on what is running, which is not something a line of text can say. Verified on device (HONOR AMM-AN00) — an idle tap listens without flashing "stopped". - A device with no conversation is offered ready-made scenes instead of a dead line. It used to
show "no conversation yet" and nothing under it. There is nothing to be helpful about yet, and a ball in the
corner has taught a newcomer nothing about what the thing does, so the one moment this app can be
proactive without being early went unused. The card now shows a hint line — "tell me, or pick one" —
then a bubble per preset from
ScenePresets, each labelled with the scene's name and carrying a sentence the user could have said: the name is what a person picks between, the sentence is exactly the part a newcomer cannot invent. Both are fields of oneScenePresets.Text, so a bubble cannot promise one thing and send another. Tapping a bubble adopts the scene before it speaks — this was wrong first and is the part worth naming: a scene is something the notebook holds, so a bubble that only sent its sentence asked the model about a scene nobody had saved, and it correctly answered that it had never seen one, in front of the user. It isadoptIfMissingrather thanadopton purpose: a bubble is offered whether or not the preset was adopted, andsaveScenereplaces the whole row includingprompt, so adopting twice would put the preset's text back over the user's own edited version with nothing to say it had happened. Pre-adopting the presets at install was the other candidate and is worse — a written-down preset is an ordinary scene, so deleting one would have it come back at the next install, the language would freeze to whatever the device spoke then, and it contradictsLocalPrompt§11. Emptiness is the whole safety story and it is mechanical: the moment there is a conversation the strip isGONEfor good, because a suggestion strip parked over a live chat would be the app talking about itself instead of answering. Two numbers are enforced rather than asked for byScenePresetsTest, which caps every summary in both languages and checks that every name a bubble can send is onebyNamecan look back up. Verified on device (HONOR AMM-AN00) by reading the notebook over the 9008 tool port. - Long-pressing a history row opens the log behind it. The scrollback says what was said;
everything that earned the right to say it — which tools ran, with what arguments, what came back,
what the model reasoned, and what each request cost in prompt tokens — was already in
files/brain_trace.jsonland reachable only withrun-as. The new page is built from those records and opened throughHtmlActivity, which is where folding the card, hanging the ball off the ledge and the Done pill were already solved. Long-press because the other two gestures on a row are taken — a single tap opens a page row's page, a double tap reads a message fullscreen — and the two detectors had to become one, becausesetOnTouchListenerreplaces rather than chains, so a second one would have silently killed the first. The row and its log are joined by<process start>-<gen>, andgenalone would have been wrong: it is an in-memory counter that restarts at 1 with every process while the trace file is persistent, so on the test device"gen":1accounted for 176 of 278 records, and installing the change reproduced the collision within 89 seconds. A scrollback matching ongenwould have opened a reinstall's first turn onto yesterday's first turn, with nothing on screen to say so;BrainTraceTestlocks the join down, dash included. Two things the page got wrong before it was ever shown to anyone, both found by reading records back off the device rather than by reasoning about them: theturncard labelled its body "Said", so the user's own sentence appeared in the assistant's mouth on the first card of the page (the heading is where the origin belongs, andkindis the field that knows it), and theendcard said "Turn ended" whenBrainTrace.endis only written on an abnormal finish — two of three turns measured had none at all. Records written before this commit have noturnfield, so a row from before the upgrade offers no gesture, and the same rule keeps hub-brain rows, scene-change rows and notification read-outs from offering a dead entry.
Fixed
request_permissionsanswered{"started": true}no matter what actually happened. That was a hard-codedtruemeaning only thatstartActivityhad not thrown. After a second refusal Android 11+ draws no dialog at all and the call still said true, so the brain read "it is on" and immediately calledget_location. Measured on turn1791621839492-4: the request returned at +19.257s,get_locationfired at +20.839s and answeredpermission denied1 ms later — the model gave the user 1.582s to find a dialog that was never drawn, then spent 34.6s looking for a way around it. The reply now carries what happened rather than that something was attempted:mode(solo/handoff/settings, who answers the sheet),result(granted/denied/no_dialog/waiting) andnext, one line saying what to do about that result.modecomes fromrouteOf, a pure function of the ids kept on a whitelist so a permission added to the catalog later falls on "the user was asked once more than necessary" rather than "the ball agreed on their behalf".resultresolves the distinction that pays for all of it: a refusal and a silent block both readPERMISSION_DENIEDand demand opposite responses, andshouldShowRequestPermissionRationale— readable only from an Activity — is what separates them.PermissionRequestActivityused to dropgrantResultson the floor; it now hands the verdict back through arm/await/settle, andPermissionsControllerwaitsDIALOG_GRACE_MSfor it, a grace period rather than a wait for the human. Verified on HONOR AMM-AN00 (location/solo/granted, contacts/handoff/waiting, steps/solo/waiting, each with the sheet really on screen); 37 unit tests pass, withPermissionsRouteTestpinning the routes from the id list without a device.- The permission sheet was drawn underneath the expanded card, so the user could never see it.
An expanded card is a fullscreen surface and our overlay sits above every activity, so
GrantPermissionsActivitywas drawn below it. Verified on device: the dialog really was up — it was bothResumedActivityandmCurrentFocus— while a screenshot showed nothing but the card. A question nobody can see is worse than no question, and it is exactly the statewaitingpromises.device.*deliberately never folds, on the reasoning that reading state is a conversation rather than an actuation; this branch is the exception that reasoning did not cover, because it addresses the user in a window of the system's own.device.scananddevice.vault.fillalready fold for the same reason, so this joins them — and only whenconfirm=true, because the first call is the script the brain reads back and that one belongs in the card. - Both READMEs gained a one-click build page and had its likely-reading fixed. The nav and the
CodeFlying paragraph now point at the
/andeepage — describe the app, tap the build button, get a pre-configured package with a QR code — and the "any app" feature line was reworded in both languages, because "not a list of integrations" was being read as a claim about integrations this app has rather than a claim about aim. The page itself is a separate service with no direct relationship to this project, as that paragraph already says.
Changed
- The prompt's rule about the system permission sheet was wrong, and is now the measured one. It
said a system permission sheet "does not take a tap you inject", and that turned out to be false.
With the card folded, an injected tap on "Always allow" really does grant:
READ_CONTACTSwent false → true, and one tap on "Allow this time" produced anACCESS_FINE_LOCATIONcarryingONE_TIME, a flag only that button can set. The earlier four failures had one cause between them and it was our own card eating the touch, not the sheet refusing it. So the rule changes from "you cannot press it" to "you cannot read it, so do not aim at it": the sheet exposes no accessibility elements at all, so any tap is blind, and the button at the bottom of this particular screen is "Deny and don't ask again", which nothing can undo. That is the real reason this step stays the user's, and it is a reason that survives being measured.LocalPrompt§0'stype_text-only rule is stated as the general one — a tool saying it pressed something is not the thing being done — and §1's list of genuine exceptions gains the sheet, which is worth naming because putting it on screen is the whole of what firing does.
[0.1.1] - 2026-10-09
Fixed
Four strings sent the user to a menu label that no longer exists — and one to a gesture that does not open settings at all. The settings card has been reorganised twice since these sentences were written, and the sentences stayed behind, which is the worst kind of copy bug: the user follows the instructions, finds nothing, and concludes the app is broken rather than the text.
check_detail_brain_key_missing,check_detail_brain_hub_missingandbrain_no_keyall said ⚙ → "Brain". The section header issettings_section_brain, which reads "Backend" — there is no label containing that word in the settings card at all. The Volcengine tab names use it only in prose ("the brain lives on a hub you run yourself"), never as a destination.check_detail_voice_key_missingsaid ⚙ → "Speech (Volcengine)" — a section that no longer exists. The Volcengine rows were moved into Backend → on-device, andsettings_section_voice/settings_section_voice_keyare still in both string files but referenced by nothing. The path is now the same Backend → on-device as the two brain keys, which is where the field actually is.brain_no_keyalso said "long-press the ball, open settings".onLongPressdoes not open settings; it callsdispatcher.expand(), which brings the full card back. Settings is the ⚙ button in that card's control bar. The sentence now names both steps, because the first one is genuinely required — the control bar only exists at full size.check_detail_brain_hub_missinggained the missing precondition: the hub tab is hidden unless Advanced → "Show cloud option" is on.showHubgates the tab in the picker, butbrainModedoes not consult it, so a device can sit in hub mode with the tab that fixes it invisible — turn the toggle on, pick the hub, turn it off again, and the self-check would have pointed at nothing.
Every label these strings now name was checked against its resource value rather than transcribed from the sentence being replaced — which is how the one remaining error was caught: the English label is "Show cloud option", singular, and the first draft of this fix wrote it plural.
Both READMEs promised a first-run wizard that no longer exists. "The first run is a stepped setup that walks you through both" — in both languages —
8dac568deleted the wizard (ui/FirstRunUi.kt). A missing setup guide is indistinguishable from a broken install, so this read as a bug report rather than as stale prose. Both now say what the app does instead: the self-check runs on every start, the card names the missing key, and its button opens settings.README.mdnamed the Volcengine field by its Chinese label, while the English build displays Volcengine API key (settings_voice_api_key). The Chinese README already had this right; only the English file was pointing at a word that never appears on an English screen.A picture, a video or a clip now goes into a page, instead of being handed over as a URL. The prompt said the opposite, and said it wrongly: "this device cannot turn an image into a URL, so an
imgtag will not open". That is false for anything on the network — the page is served against a fakehttps://origin, the app holdsINTERNET, and WebView loads an absolutehttps://subresource without complaint. What is true is narrower, and the old sentence flattened it into a blanket ban: a local file has no address a page can point at, andhttp://is refused. So the model, obeying its instructions, could not put a picture on screen — and what it did instead was hand the user a URL to open, on the device they were already holding. §7 now asks for the embedding directly (<img>,<video controls>,<audio controls>, absolutehttps://only), separates a media file from a site — the latter belongs in Chrome throughopen_url, not in aniframe, which most sites refuse — and says outright that passing over a bare URL is not an answer. The wrong half of the same sentence sat in theshow_htmldescription too, which is the text the model reads while composing; media is carved out there as well, because what that rule protects is a self-contained shell. Verified by compiling and by reading both strings back out of the compiled classes — the new text is present, all three old phrasings gone. Not verified on a device: none was attached, so a realhttpsimage and a real.mp4rendering in the page are still unconfirmed.The launcher icon could become a permanent dead end, because the overlay was inferred from a view object instead of asked for as a grant. First run, accessibility on, the overlay grant off: the checklist came up, the user closed it, and from then on the icon did nothing — no screen, no error, no way back in short of reinstalling. Nothing was removing the ball's window. On HONOR/MagicOS a denied
SYSTEM_ALERT_WINDOWdoes not makeaddViewthrow: it succeeds, and the system hides the window instead — present in the window list withalpha=0.0,mAppOpVisibility=false,isReadyForDisplay()=false. So the view object outlives the grant, andFloatingWindowUi.isShown()— which is onlyroot != null, a memory marker and nothing more — keeps answering true while the user can see nothing at all.ensureOverlays()opened with exactly that test, soSelfCheckActivitytook its expand-and-finish branch, calledexpand()on a card that cannot be drawn, finished itself, and left an empty screen; the checklist branch, which is the one that names this failure in red, never ran. The grant is now asked first and directly, asSettings.canDrawOverlays()— the same predicateSelfCheck.overlayuses for the overlay row, so the page and the door can no longer contradict each other (which is what the user hit: the page said "off" while the door said it was on screen). It also covers revoking a grant that was already given, whenrootstays non-null for the life of the process.showSelfCheckCard()guarded with the same marker, where its own comment had claimed it already did. Verified on the device (LIO-TL00) on both paths, because this edits the working path's entry test too — grant denied now reaches the checklist (13 activity records), grant allowed still takes the expand branch (0 records); checking only the first would have shipped "the icon always shows the checklist".
Changed
- The overlay-permission row now leads the self-check list, ahead of accessibility. Both are hard gates and the order between them was arbitrary; what settles it is that they fail in opposite directions. With the service off the app is visibly inert — nothing to run, and every row below is decoration. With the grant off it looks broken instead: the service runs perfectly well, silently cannot draw, and tapping the icon appears to do nothing at all. That second shape is the one users misread as "the app is dead", so it goes first.
- The accessibility row's off-state copy named the row's list position — *"the rest of this which the reorder above invalidated. It is now "nothing else here matters" — "nothing else here matters either": same meaning, no longer positional. Same edit in both string files.
Removed
- The first-run wizard is gone (
ui/FirstRunUi.kt, its wiring inScreenBodyService, the re-run-setup button on the self-check card, and its strings in both languages). The wizard and the startup self-check both wanted to put a card up on a first start, and when the check found a problem it covered the wizard, so a new user saw whichever lost the race. The wizard was removed rather than arbitrate between the two. The self-check's own prefs file stays, as doesPermissionRequestActivity, which the check still uses; an oldwizard_doneflag on existing installs is simply never read. The wake word is now offered only from settings and as a note row on the self-check. The two READMEs had promised this wizard — see the Fixed entry above. check_page_no_overlayand theSelfCheckUi.note()helper it was the only caller of. The note explained that the settings sheet is drawn as an overlay — a fact the page does not need to teach, because the overlay row directly below it already carries the fix, in that row's own words and with a button.fixButton()'s comment used to point at "the note above the list"; it now states the rule on its own: a button that does nothing is worse than no button.
[0.1.0] - 2026-10-09
The first tagged release. Because nothing was tagged before it, this section is the whole project rather than a delta from a previous version — it reads as a long list of Added for that reason, not because a release this size is the norm.
Added
docs/architecture.svg+docs/architecture.zh-CN.svg, and a## Architecturesection in both READMEs. The README described pieces — the ball, the vault, the dog — and never showed how a turn actually moves. Drawing it forced two corrections to the diagram before it was right, both of them worth recording. The first was factual: the draft showed "two brains" when there are three mutually exclusive backends, and it marked the hosted one as the default when an unconfigured device reads asBRAIN_LOCAL. The second was structural and worse: the draft stopped at the backend and drew nothing after it — no agent loop, no outputs, which is most of the interesting half. No class is named after the harness; it isLocalBrainitself. The picture now shows the system prompt held constant with the volatile parts (clock, notebook index, scenes, photos) deliberately inlined into the user message so the provider's prefix cache still hits, the tool table rebuilt every turn, and the non-streaming step loop with its guards. Outputs come back through six channels, none of which had been drawn: speech, text, a full-screen page, saved artifacts, long text, and the dog. Both files are hand-built SVG with hardcoded colours, because GitHub renders a README image through<img>andvar(--color-*)resolves to black there; both were rendered at 1:1 and read back to check for overflow, which caught three layout faults that reading the source alone would not have shown.- Both READMEs now ask for your use case (
## Showcase). The README described what the thing can do and never asked anyone to show what they did with it, which is how a project ends up with no idea whether the odd corners work. The section is an invitation and three things that make a post useful: the device and ROM (this app talks to Android at a level where ROMs disagree, so "worked on my phone" is not reproducible), the sentence you actually said, and a link if there is one. It goes to a Discussions Show and tell rather than the issue tracker, which follows the ruleconfig.ymlalready sets — the tracker is for reproducible defects, and a story is not one. Two details are deliberate rather than decorative. GitHub cannot play a video inside a README (iframes are stripped, so a YouTube link stays a link), and the section hands over the thumbnail form that makes it look like a player anyway, rather than letting twenty people discover that separately. And it says plainly that a screenshot of this app is a picture of your messages — the one thing in a showcase that can do real harm is posting someone's private screen content, including your own. LICENSE— Apache-2.0 with additional conditions (Part A: no multi-tenant SaaS, no white-labelling that removes the branding, no embedding it in another product without written authorisation). Note that GitHub reports this asNOASSERTION.SECURITY.md— threat model, the five known security limitations, and a private reporting channel.docs/permissions.md+docs/permissions.zh-CN.md— why the app asks for the permissions it asks for, each one tied to a concrete tool and to what happens if you refuse it. One file per language.CONTRIBUTING.md+CONTRIBUTING.zh-CN.md— code style, the rules for the three kinds of Chinese in this codebase, and the contribution licensing terms. One file per language.- A
Signed-off-by:sign-off is now required on every commit, and CI enforces it..github/workflows/dco.ymlchecks each commit on a pull request, and bothCONTRIBUTINGfiles plus the pull request template explain the one-line change (git commit -s). This is a DCO, not a CLA: the contributor keeps their copyright and certifies only that they had the right to submit the work. The reason it exists is LICENSE A.2 — the project is relicensed under terms that differ from stock Apache 2.0, and A.2 asks contributors to grant that. A term nobody has explicitly agreed to is a term a contributor can later dispute; a sign-off line is a per-commit record that they did. Deliberately not a CLA, which would need a signature process and an entity to hold it — friction that buys nothing here. CODE_OF_CONDUCT.md— Contributor Covenant v2.1.THIRD_PARTY_NOTICES.md— generated dependency license inventory: 93 components across 5 licenses. Notably flags that ML Kit and theplay-services-*stubs are not open-source.tools/license_audit.py— the generator behind the notices file..github/— bug report and feature request issue forms, a pull request template, and issue routing with blank issues disabled.README.zh-CN.md— the Chinese setup guide, split out fromREADME.md.DISCLAIMER.md+DISCLAIMER.zh-CN.md— the user-facing terms, one file per language (same convention as the two READMEs). Covers the four things Apache-2.0 does not: where your data actually goes (at the time this file was written, no telemetry in the repo, but speech goes to Volcengine and the conversation plus every tool result goes to the brain endpoint you configure), what the software is genuinely capable of (it can complete a payment by tapping; the only guardrail is Android'sFLAG_SECURE), who bears the consequences, and third-party platform terms.docs/acceptable-use.md+docs/acceptable-use.zh-CN.md— the prohibited uses, one file per language. This is the part that matters legally: naming the prohibited uses is what keeps this project from being characterised as providing assistance to them. It calls out fund movement (including the Article 287-2 aiding-information-network-crime exposure) and publishing false information separately.docs/permissions.mdgained a "data goes where" section — the honest version of "we do not collect your data", which is true of the maintainers and false of the device.- Both READMEs now carry a warning above the fold and a
## Disclaimer and acceptable usesection;SECURITY.mdstates the triage boundary (steering it is a vulnerability; doing what it was built to do is documented behaviour). - A version row in the self-check list — the first thing this app ever sends anywhere. It asks a
version service whether a newer build exists and reports the answer as one more row in the same list
the device diagnostics live in: one row, one renderer, no second settings page. The level mapping is
the deliberate part. Current → OK, newer release available → NOTE, forced update → WARN, and
everything undetermined → NOTE. An outdated app is not a broken one, and only FAIL and WARN may
raise the startup card, so a vendor that popped a card every time it shipped a patch would be
nagware — which is the exact failure
SelfCheck's level doc is written against. The forced case is the one real WARN, because the service is then saying this build is no longer supported. A statistics endpoint being unreachable is never the user's problem, so it never paints the card red. The address is a hardcoded constant, not a setting.VERSION_API_BASElives indevice/SelfCheck.kt(https://andee.kuafuai.net), with no settings row, no build flag and nolocal.propertieskey behind it, so turning the check off means editing that line and rebuilding. That is a deliberate trade rather than an oversight: a check which can be silently misconfigured is a check that silently stops working, and this one should be on for everyone or off for everyone. The price is that there is no build-time way to opt out, anddocs/permissions.mdnow says so instead of pretending otherwise — which it had to, because the first cut of this feature did put the address inlocal.propertiesand then advertised a fresh clone as reporting nothing. res/xml/network_security_config.xmldeliberately did not gain the version host, on the second pass. Android refuses cleartext HTTP to any host not named there, and what it produces —CLEARTEXT communication to X not permitted by network security policy— arrives as a version row reading "not found", a message about a policy rather than about anything the user did. So the first cut added the service's IP address to that file, which is the right move for ahttp://address. The address then becamehttps://andee.kuafuai.net, which needs no exception at all, and the entry was deleted rather than left in place: an exception for a host nobody calls only widens what this app will talk to in the clear. The CodeFlying backend's plain-HTTP address is still listed and still needs to be. The rule to carry forward is that this file tracks the addresses actually in use — adding a line is part of pointing the app at plain HTTP, and removing it is part of moving away.
Fixed
- A failed voice turn crashed the app with
StackOverflowErrorinstead of reporting the failure.AsrController's listener implementsonError(msg), and its body calledonError(msg)— written meaning the constructor's(String) -> Unitcallback that routes the message to the owner. Kotlin resolves an unqualified call to the nearest scope first, so the line called itself: one recognizer error became infinite recursion until the stack ran out (stack size 1039KB, thousands of frames ofAsrController.kt:152). Any Volcengine failure — a dead network, an expired token, a rejected handshake — took the whole process down, which is the opposite of what the callback is for. Now qualified[email protected](msg). The listener's three other callbacks are unaffected: none of them shares a name with an outer member.
Changed
The APK went from ~39 MB to ~8 MB. Two causes, both structural rather than a matter of trimming assets. First, every build produced one fat APK for all four ABIs, and ML Kit's
libbarhopper_v3.sois ~4 MB, so it was paid for four times over. Second, the debug variant ships un-shrunk dex (16.7 MB of it, against 3.2 MB after R8), andminifyEnabledwas off on release too. Fixed with an ABI split,minifyEnabled+shrinkResourceson release, and deflated native libs. The output paths changed: there is noapp-debug.apkany more, onlyapp-<abi>-debug.apkandapp-<abi>-release-unsigned.apk.assembleReleasealso runs in CI now, as a job parallel to the debug one, because R8 andlintVitalReleaseonly run on release — and that is exactly how a release build that failed on every invocation sat in the repository unnoticed.assembleReleasebuilds again. It had been failing inlintVitalReleaseonres/xml/backup_rules.xmlandres/xml/data_extraction_rules.xml, both of which excludedomain="device_root". That is a real domain — the framework has carriedDEVICE_ROOT_TREE_TOKENsince device-protected storage landed, and the Auto Backup reference definesdevice_rootas "like root but for the device-protected storage". AGP 8.3'sFullBackupContentcheck validates against a five-value list that predates it, so the error was a false positive. Both lines now carry a scopedtools:ignore; the check still covers everything else in those files, including a genuine typo in a domain.SECURITY.mdno longer claims a safeguard the software does not have. Its prompt-injection section stated that destructive or irreversible actions are "gated behindask_user/confirmflows". That is not true, and it was the one sentence in the repository that could be turned against the project:ask_useris an ordinary tool the model chooses to call,confirmexists on a single tool (request_permissions, where it prevents an unwanted permission dialog, not a payment), and nothing forces a confirmation beforetap_by_coordinates— which is the path a payment is completed along and the very thing the disclaimer is about.DISCLAIMER.mdanddocs/acceptable-use.mdhad both already said the opposite in as many words ("no guardrail whatsoever", "no technical guardrail on that path"), so the three files contradicted each other and the security page was the one that was wrong. It now states the posture plainly: the checks on risky actions are instructions to the model rather than enforcement, and the device should be assumed drivable into anything its permissions and the user's own logged-in state allow — including moving money — with no prompt appearing. The same edit strips detail that made the page read as a recipe rather than a warning: the example request frame, the per-method enumeration of what a LAN peer can reach, and the note that the listening port is a fixed constant. The port number stays — it is what the recommended mitigation (firewall it) refers to. Two smaller changes follow from the same reasoning. The scope list no longer counts "an injection caused an action the user did not ask for" as a vulnerability, because that is the limitation §3 now describes rather than a promise the project can be held to; and the reporting promises are now conditional ("where we can", "no timeframe is promised"), so a small project is not bound to a reply it may not send. A line at the top states the out-loud version: nothing in the file is a promise that any particular action is prevented.LICENSEPart A gained a fourth condition, (d): shipping the software inside another product now needs written authorisation. Conditions (a)–(c) already stopped white-labelling, multi-tenant SaaS and dropping the attribution, but nothing stopped someone taking the code and folding it into a product they sell — which is the one path that makes the rest moot. (d) closes it, and is written as narrowly as it can be: running it yourself, running it across your own organisation's devices, forking it and distributing the source so others can build it, and charging to install, configure or support it for a client who runs it themselves are all still permitted. The line drawn is between a service and a product — doing work that uses the software is fine, shipping the software as part of something else is not. The heading and the lead sentence were corrected from "three conditions" to match.README.md,DISCLAIMER.mdanddocs/acceptable-use.mdreference Part A without counting the clauses, so no copy drifted — the enumerations they do carry are illustrative, and onlyLICENSEitself states all four.Positioning corrected throughout the docs. Both READMEs and
CLAUDE.mddescribed this as "the body side of a Physical Agent system" — i.e. an assistant that a brain drives. That contradicts the project's own first instruction, inLocalPrompt.TEXT: "you are not a tablet's assistant, you are an individual, and the tablet is only what carries you." The docs now lead with the individual, and give the body / persona / notebook mechanisms that back it up in code (fourteen moods and idle fidgets,BallLook.personareaching the head of the system prompt, the privateNotebook, the(to yourself …)self-review turn, and §11's rule that the character is delivery and never a licence). No code changed — the code was already right; the prose was not.README.mdis now the English front page. The detailed Chinese provisioning guide moved toREADME.zh-CN.md, and both cross-link.gradle.propertiesno longer pinsorg.gradle.java.home. It previously hardcoded a machine-specific JDK path, so the build failed immediately for everyone else while looking like an environment problem. The real constraint (JDK 17–20; Gradle 8.4 cannot run on Java 21+, Android Studio's bundled JBR 25 is rejected) is now documented instead of worked around.tools/check_look_sync.pyderives the repository root from its own location instead of a hardcoded absolute path.tools/dump_faces_ascii.pyandtools/verify_faces_lab.pylocate Chrome byCHROME_PATH, then common install locations, thenPATH, instead of hardcoding a macOS-only path.CLAUDE.mdcorrected: the Gradle JDK note, the stale claim that the checkout isfrontend/, and the description of which README is which.Every bilingual document now follows one file per language, and each language links only to its own counterparts.
CONTRIBUTING.mdanddocs/permissions.mdwere written in Chinese under the English default filename — the opposite of the convention the other four files follow. They are now split into.md(English) and.zh-CN.md(Chinese), and every Chinese page that used to point at an English file now points at the Chinese one, so a Chinese reader is never dropped onto an English page.SECURITY.mdandCODE_OF_CONDUCT.mdhave no Chinese edition; the Chinese docs now say so explicitly instead of linking silently.Both READMEs restructured and fact-checked against the code. The 16-item feature list now sits above Quick start instead of below What it looks like, so the page leads with what the thing does; both nav bars follow. Three claims were corrected against the source: the ball has fourteen moods (
ui/ball/Mood.kt), not fifteen; the tool list no longer namesfiles(there is no file tool —locationdoes exist); and the Kotlin line count is now 43k+. The two-credential block puts the brain key first and marks it required, with the voice key second and optional — voice used to lead, which buried the one whose absence stops the brain from running at all. The "Things you can just say" table no longer repeats the WeChat example: that one is told as a single moment above the table ("you are driving when your mother messages you"), which is what the rest of the list is a variation on. And a short note near the top now says plainly that Android is the only supported platform today — the "any device with a body" line is about what the software is written for, and was reading as though other platforms already shipped.Four documents corrected where they contradicted the code — no behaviour changed.
DISCLAIMER.md/DISCLAIMER.zh-CN.mdanddocs/permissions.md/docs/permissions.zh-CN.mdall listedbrain=hubfirst in their "data goes where" table's Default column, which read as though the hub were the out-of-box mode. It is not:VoiceConfig.brainConfigreadsgetString("brain", BRAIN_LOCAL), and thebrainpicker lists the on-device option leftmost for the same reason. The row now readsbrain=local(the default) →api.deepseek.com, with hub mode as the other case. Separately,CLAUDE.md's "Existing but not wired up" section claimed there was notts.*dispatcher namespace — there is one (tts.speak,CommandDispatcher); what it lacks is aToolSchemasentry, which is the actual reason the model cannot call it, and the distinction matters because the fix for "let it choose to speak" is adding a schema, not a branch. The matching comment at the head ofCommandDispatcher("will be added later") is corrected too. The error was load-bearing in the disclaimer especially: a reader who believed hub was the default would size the data-path risk wrongly.CodeFlying is now introduced as what it is, and disclaimed. Quick start used to name codeflying.app / codeflying.net as simply "the fastest route", which read as though it were this project's own or affiliated build channel — an impression a reader could act on (and later hold this project to). It is now described in its own terms first — a general platform where you describe the app you want and it develops and publishes it — followed by one plain sentence: a separate service with no direct relationship to this project, its own terms, pricing and data handling. Same treatment in
## Using Andee; the Chinese getting-started section table carried it too, and has since been removed (see Removed). The practical guidance is unchanged (still the shortest path; still an APK you have to install; the free token allowance is still finite and then paid), because none of that was wrong — only the implied ownership was.The English README now addresses an English reader — in its examples and in the panel names it points at. Two different faults with one symptom: a reader concluding the page is not for them. The examples were the Chinese ones with the words swapped (
WeChat,Taobao,JD), and an example is precisely what a reader uses to judge whether something is addressed to them. They are nowWhatsApp,AmazonandeBay— WhatsApp rather than a platform-native messenger because this runs on Android only, and naming an iOS-only one would be a false claim rather than a foreign one. Two of the scene examples needed rewriting rather than substituting: "look after my WeChat" is a Chinese scene example carried across word for word, which is not a sentence English builds. Separately, the panel names were the Chinese labels, so a reader on an English device was being told to look for a Chinese label — a word their screen does not have. They now use the labelsvalues-en/actually ships: Scenes, Made for you, What it remembers, Voice API key. The measured passage keeps every fact. Xiaomi Pad 5, MIUI, WeChat 8.0.78 and the 219-node count are what was observed, and changing the app name there to make the prose consistent would be inventing a measurement. Only its framing moved: it opened on WeChat, which a non-Chinese reader reads as someone else's app having someone else's problem, so it now opens on the general behaviour and names WeChat as the one we happened to measure.The blanket "no telemetry at all" claim is gone, because the version row above made it false. It appeared in nine places —
DISCLAIMER.md,DISCLAIMER.zh-CN.md,docs/permissions.md,docs/permissions.zh-CN.md, both READMEs, and both README badges — and all nine were in the same voice: not "we anonymise it", not "opt out in settings", but a flat statement that the reader was invisible. A check that carries an IP, a phone model and a stable per-install ID is telemetry by any ordinary reading of the word, so keeping the sentence would have converted a product guarantee into a lie. The replacements keep the half that is still true and still checkable — no analytics SDK, no crash reporting, no event tracking, no advertising ID, all four confirmed by the same grep the old text quoted — and then state the single exception field by field, in a row of its own in the data table, next to an explicit list of what that request does not carry (conversation, screen, app list, contacts, location, crash log, advertising ID).docs/permissions.mdis the canonical copy; the rest point at it. The badge went fromtelemetry-nonetotelemetry-no content, because "no content" is the claim that survives. Worth recording for whoever edits these next: the old sentence was true of the repository on the day it was written and was always going to be false the moment the project shipped anything that phoned home. This file had already annotated that distinction in the DISCLAIMER's own entry above — "we do not collect your data", which is true of the maintainers and false of the device — and the claim that had to be corrected anyway was the neighbouring one about the maintainers.
Removed
brain/andee-system-prompt.md— an unused mirror of the cloud brain's prompt. It had no runtime role (no build step references it, no code loads it). The cloud brain mode itself is not retired; only this repo's copy of its prompt was removed. Recover it withgit show <rev>:brain/andee-system-prompt.md.dog/NRF24_Dongle/README.mdreference in the README — the dongle firmware is not in this repository and never was. The README now says so.- The Chinese getting-started section in
README.zh-CN.md, heading included. Its visible content was three rows announcing that the demo video and the screenshots do not exist yet, which tells a reader nothing they cannot see for themselves. The heading went too: once the table was gone the section held nothing but an HTML comment, and a heading over nothing renders as exactly that. Every pointer to it went as well — the nav anchor, the header comment, and the sentence above. In its place, a slot for the demo video sits directly under the hero in both language files, and the asset checklist moved down beside the contributors block, which is where the English file already kept it.
Known issues
tools/verify_faces_lab.pyreports 7/16 assertions passing against the committedreports/ball-faces-lab.html. This is pre-existing (identical before and after the portability changes above) and is not caused by the open-sourcing work. Left as-is; worth resolving before contributors are pointed at the ball-face lab.- The debug WebSocket server (
net/BodyWsServer) binds0.0.0.0:9008with no authentication. SeeSECURITY.md. Not fixed.