Skip to content

armgit / andee

Public Imported from GitHub
0

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。

chore(release): v0.7.1 — taps land where it means them to

booboosuic24eac7dfdfd
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 0edb0f8..fb82194 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -4,7 +4,7 @@ Notable changes to Andee. The format follows [Keep a Changelog](https://keepacha
## Versioning policy
-- `versionCode` / `versionName` live in `app/build.gradle` and are currently **`16` / `"0.7.0"`**. The
+- `versionCode` / `versionName` live in `app/build.gradle` and are currently **`17` / `"0.7.1"`**. The
Mac's product version is read from that same `versionName`, so both bodies carry one number.
- Releases are tagged `vMAJOR.MINOR.PATCH`, `versionName` mirrors the tag without the leading `v`,
and `versionCode` increases by one per release. `versionCode` must increase **monotonically** or
@@ -29,6 +29,87 @@ hash into `versionName` at build time would close this hole; it is not implement
## [Unreleased]
+## [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](../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_model` overrides 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.
+- **`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_text` refused). 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 with `osascript`, 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_target` is 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:
diff --git a/DISCLAIMER.md b/DISCLAIMER.md
index 0007997..7eda947 100644
--- a/DISCLAIMER.md
+++ b/DISCLAIMER.md
@@ -46,7 +46,7 @@ The second row needs spelling out: **tool results are part of the conversation.*
**So "local mode" does not mean "processed locally."** Local mode swaps the hub for whichever provider you typed in. It does not keep a single byte on the device.
Therefore:
-- *"We do not monitor or store your data"* — **mostly true, but do not overclaim it.** No database, no behavioural analysis, and no reading of anything you send; the one exception is the version check, which tells us that some device of some model on some IP is running 0.7.0.
+- *"We do not monitor or store your data"* — **mostly true, but do not overclaim it.** No database, no behavioural analysis, and no reading of anything you send; the one exception is the version check, which tells us that some device of some model on some IP is running 0.7.1.
- *"Your data never leaves the device"* — **false. Do not claim it.**
## 3. What it can actually do (not hypotheticals — current state of the code)
diff --git a/DISCLAIMER.zh-CN.md b/DISCLAIMER.zh-CN.md
index 164910e..7ca758e 100644
--- a/DISCLAIMER.zh-CN.md
+++ b/DISCLAIMER.zh-CN.md
@@ -45,7 +45,7 @@
**因此"本机模式"不等于"本地处理"** —— 本机模式只是把大脑从 hub 换成了你填的那个服务商。它没有让数据留在设备上,一个字都没有。
所以:
-- 「我们不会监控和存储您的数据」——**基本成立,但别把话说满**。不落库、不做行为分析、不看你发的内容;唯一的例外是版本自检,它让我们知道"某个 IP 上,有一台某个型号的设备在跑 0.7.0"。
+- 「我们不会监控和存储您的数据」——**基本成立,但别把话说满**。不落库、不做行为分析、不看你发的内容;唯一的例外是版本自检,它让我们知道"某个 IP 上,有一台某个型号的设备在跑 0.7.1"。
- 「您的数据不会离开设备」——**不成立,而且不要说**。
## 3. 它真的能做什么(以下都不是危言耸听,是代码里的现状)
diff --git a/README.md b/README.md
index 18739da..fe3040f 100644
--- a/README.md
+++ b/README.md
@@ -700,7 +700,7 @@ Pure Kotlin, no Compose; the whole thing hangs off two pillars, `WindowManager`
| minSdk | 30 | Android 11 — `AccessibilityAction.ACTION_IME_ENTER` and `takeScreenshot` are both API 30+ |
| Java / Kotlin JVM target | 1.8 | `sourceCompatibility` 1.8, `jvmTarget = '1.8'` |
| **Build JDK** | **17–20** | see [Which JDK](#which-jdk) — this is the step people get stuck on |
-| versionCode / versionName | 16 / 0.7.0 | see [CHANGELOG.md](CHANGELOG.md) for the policy |
+| versionCode / versionName | 17 / 0.7.1 | see [CHANGELOG.md](CHANGELOG.md) for the policy |
| applicationId / namespace | `net.kuafuai.andee` | never changed from the template. It affects nothing |
**Runtime dependencies** (`app/build.gradle`):
@@ -1133,7 +1133,7 @@ Read it before redistributing: **the ML Kit barcode SDK and the `play-services-*
open-source**, and the barcode model is fetched from Google at runtime.
See [CHANGELOG.md](CHANGELOG.md) for what changed and for the versioning policy. The latest release is
-**`v0.7.0`** — pre-1.0 on purpose, because the wire contract and the build flags are still moving.
+**`v0.7.1`** — pre-1.0 on purpose, because the wire contract and the build flags are still moving.
## Related documents
diff --git a/README.zh-CN.md b/README.zh-CN.md
index 4ec18d5..c024337 100644
--- a/README.zh-CN.md
+++ b/README.zh-CN.md
@@ -525,7 +525,7 @@ Google),再填该服务商的密钥。这是两个独立的服务,条款
| minSdk | 30 | Android 11 —— 因为要用 `AccessibilityAction.ACTION_IME_ENTER`(API 30+)和 `takeScreenshot`(API 30+) |
| Java / Kotlin JVM target | 1.8 | `sourceCompatibility` 1.8、`jvmTarget = '1.8'` |
| **构建用 JDK** | **17–20** | 见[配 JDK](#配-jdk),这是最容易卡住的一步 |
-| versionCode / versionName | 16 / 0.7.0 | 策略见 [CHANGELOG.md](CHANGELOG.md) |
+| versionCode / versionName | 17 / 0.7.1 | 策略见 [CHANGELOG.md](CHANGELOG.md) |
| applicationId / namespace | `net.kuafuai.andee` | 从模板里没改,不影响功能 |
**运行库依赖**(`app/build.gradle`):
@@ -992,7 +992,7 @@ CLAUDE.md 给 AI 编码助手看的架构说明(英文)
随包分发的第三方组件清单在 [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md)。再分发之前请先看:
**ML Kit 扫码 SDK 和 `play-services-*` 不是开源许可**,而且扫码模型是运行时从 Google 拉的。
-改了什么、以及版本号策略见 [CHANGELOG.md](CHANGELOG.md) —— 当前版本是 **`v0.7.0`**,故意停在
+改了什么、以及版本号策略见 [CHANGELOG.md](CHANGELOG.md) —— 当前版本是 **`v0.7.1`**,故意停在
pre-1.0:线上协议和构建开关都还在动。
## 相关文档
diff --git a/app/build.gradle b/app/build.gradle
index 6e7815d..bec8265 100644
--- a/app/build.gradle
+++ b/app/build.gradle
@@ -130,8 +130,8 @@ android {
applicationId "net.kuafuai.andee"
minSdk 30
targetSdk 34
- versionCode 16
- versionName "0.7.0"
+ versionCode 17
+ versionName "0.7.1"
testInstrumentationRunner "androidx.test.runner.AndroidJUnitRunner"
diff --git a/docs/releases/v0.7.1.md b/docs/releases/v0.7.1.md
new file mode 100644
index 0000000..e0fe461
--- /dev/null
+++ b/docs/releases/v0.7.1.md
@@ -0,0 +1,261 @@
+# v0.7.1
+
+**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 the misses were the model doing arithmetic it is bad
+at. The other half were guards the phone had and the Mac did not.
+
+So the brain no longer produces coordinates at all. It describes what it wants **in words**; a vision
+model that is good at pointing answers with a point; the device taps there. The reasoning, the
+measurements and the approaches that were rejected are in
+**[docs/design/click-accuracy.zh-CN.md](../design/click-accuracy.zh-CN.md)** (Chinese).
+
+> **These binaries carry no credentials.** No CodeFlying key, no voice-provider key, and no Google key.
+> **You supply your own**: sign up from the link inside the app, then paste the key into
+> **Settings → Backend**. Starting is free and the free allowance is finite.
+>
+> Each package already points at the right server for its language — the English ones at
+> `myapp.codeflying.app`, the Chinese at `baas.kuafuai.net`. You do not choose this.
+>
+> **The macOS packages are signed with a local identity, not an Apple Developer ID**, so Gatekeeper
+> refuses the first launch. On macOS 15 the old "right-click → Open" trick no longer exists; see
+> **[Opening the Mac app](#opening-the-mac-app)**.
+
+## Highlights
+
+- **It taps by describing, not by calculating.** `tap_target` takes a sentence like *"the 发送 button
+ right of the comment box at the bottom"*, a vision model finds it, the device taps its centre. Measured
+ on our own screenshots: **6.4 px mean error**, 24 px worst. Nothing to configure — it uses the model you
+ already gave the brain.
+- **A drag that is a drag.** `drag` presses, moves while holding, and releases — for sliders, splitters,
+ window title bars, reordering a row. Aiming either end works by coordinate or in words.
+- **The Mac got the guards the phone already had** — seven of them, each fixing a failure we had measured:
+ taps now report whether the screen changed, elements a tap cannot reach are no longer listed, tapping by
+ number checks the element is still there, the repeat-tap radius shrank from 120 pt to 30 pt, apps that
+ hide their UI are recognised, a window on another display is reported instead of aimed at, and every
+ click brings its app forward first.
+- **It can type into apps that hide their focused field.** WeChat on the Mac reports no focused element, so
+ `type_text` refused there — and the workaround the model found used the shell, which is closed on
+ notification turns. A message it could read, it could not answer.
+- **A backend that is not answering as an API now says so**, instead of dying on
+ `A JSONObject text must begin with '{'`.
+
+## What's Changed
+
+### Added
+
+- **`tap_target`** (both bodies, local brain only) — describe the element in words; a vision model answers
+ a point; the device taps it. The brain never emits a coordinate, so nothing can be off by its arithmetic.
+ - **Uses the model you configured for the brain**, on every backend. `locator_model` overrides it if you
+ want something faster.
+ - **Degrades instead of gating.** A backend that will not accept images withdraws the tool for that run,
+ as do three unreadable answers in a row; one good answer resets the count. It then falls back to aiming
+ by screen cell, which is what it did before.
+ - Two prompt details were measured rather than assumed: **asking in Chinese** (the same targets asked in
+ English were off by 25–37 px on three of ten) and **naming what it sees before deciding** (without it,
+ "the red 删除 button" was confidently answered with the close button). When the target is not there, it
+ now says so rather than guessing.
+ - On the phone this counts as an *aimed* tap, which satisfies the rule that apps hiding their structure
+ must be zoomed into first — but the repeat-tap guard still applies.
+- **`drag`** (Mac) — press, move while holding, release. `swipe_by_coordinates` scrolls; this grabs. Both
+ ends are aimed by coordinate or in words, and the result carries the same `screen_changed` and marked
+ screenshot that a tap does.
+
+### Fixed
+
+- **The Mac gained the guards the phone already had** (design document §3.2). Each fixes a measured
+ failure rather than a hypothetical one:
+ - **A tap reports whether the screen changed**, and marks where it landed with a magenta ring. Without
+ it, the model could not tell a tap that missed from one that opened a menu. The wait is a poll now, up
+ to 1.2 s, instead of a flat 400 ms.
+ - **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 the current
+ label at that spot plus an instruction to re-read the screen.
+ - **The repeat-tap radius went from 120 pt to 30 pt.** The old one — about 240 Retina pixels — refused 9
+ of 56 taps, including aim corrections converging on a small target.
+ - **Apps that do not expose their UI are recognised.** WeChat's Mac window offers four elements in total;
+ a window that large (over 200,000 pt²) with six elements or fewer is now treated as tree-blind, and a
+ screenshot is attached automatically with a note to look at it instead of guessing.
+ - **A window on another display is reported**, asking the user to move it to the main screen rather than
+ aiming at something this screen cannot show. The scale now comes from the main display instead of
+ whichever screen happens to have 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. In a real task this looked like
+ "creating a new app did nothing on the first tap" and "typing reported success but the box was empty".
+ - **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.** When the front app reports no focused
+ element, it 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 returns the result as *unverified*
+ with a note to check the screenshot, rather than claiming success. Shell refusals on notification and
+ todo turns now point at `type_text` and say not to promise that a later automatic turn will run a
+ command.
+- **A backend answering with something other than the API is now reported as that.** OkHttp turns 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 '{'`, which tells nobody anything (measured: `codeflying.net`
+ → `www.codeflying.net`, an nginx page). Redirects are no longer followed, and a 200 that is not JSON is
+ reported as an address that is wrong or has moved, naming where it pointed and showing the first 80
+ characters of what came back.
+- **The cancel button in the Mac's destructive-action confirmations** was squeezed to a vertical sliver,
+ because both buttons in the row filled the width and the second took everything. Each now takes half.
+- **清空聊天记录 on the Mac now clears all of it.** It did three things where the shared routine does seven:
+ it did not stop or drop the running brain, did not clear what the brain had been built for, and left the
+ progress caption up. A stop that only half-stops is how a cleared conversation carries on talking.
+- **New strings and documentation wording**: the zoomed screenshot's labels now read as a map's ruler
+ (top and left), and the crosshair is described as magenta rather than red, which is what it is.
+
+### Note on verification
+
+- **`tap_target` is not yet verified on a phone.** It needs a phone whose voice and brain keys both work,
+ and one WeChat comment or reply tried against it. 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.
+
+## Download
+
+### English packages
+
+**Android** (Android 11 / API 30 or later). Take the one that matches your device's processor:
+
+| Label shown on the page | File name | Take it if |
+|---|---|---|
+| Andee 0.7.1 · Android-arm64-v8a | `andee-0.7.1-en-android-arm64-v8a.apk` | **any phone from the last several years** — almost everyone |
+| Andee 0.7.1 · Android-armeabi-v7a | `andee-0.7.1-en-android-armeabi-v7a.apk` | an older 32-bit device |
+| Andee 0.7.1 · Android-x86_64 | `andee-0.7.1-en-android-x86_64.apk` | an emulator on a computer |
+
+*Not sure? Take `arm64-v8a`. If it refuses to install because no architecture matches, try
+`armeabi-v7a` — the two do not conflict, and only one of them can be right.*
+
+**macOS** (macOS 14 or later). Not a universal binary — take the one matching your Mac's chip:
+
+| Label shown on the page | File name | Take it if |
+|---|---|---|
+| Andee 0.7.1 · Mac-Apple silicon | `andee-0.7.1-en-mac-arm64.dmg` | an Apple silicon Mac (M1 or later) — **most Macs** |
+| Andee 0.7.1 · Mac-Intel | `andee-0.7.1-en-mac-x64.dmg` | an Intel Mac |
+
+*To check: Apple menu → About This Mac. "Chip: Apple M1/M2/M3/M4" means Apple silicon; "Processor:
+Intel Core" means Intel.*
+
+### Chinese packages
+
+The Chinese packages are the `-zh-` files, and each displays a Chinese label on the Release page:
+
+| Label shown on the page | File name |
+|---|---|
+| Andee 0.7.1 中文版 · Android-arm64-v8a | `andee-0.7.1-zh-android-arm64-v8a.apk` |
+| Andee 0.7.1 中文版 · Android-armeabi-v7a | `andee-0.7.1-zh-android-armeabi-v7a.apk` |
+| Andee 0.7.1 中文版 · Android-x86_64 | `andee-0.7.1-zh-android-x86_64.apk` |
+| Andee 0.7.1 中文版 · Mac-Apple 芯片 | `andee-0.7.1-zh-mac-arm64.dmg` |
+| Andee 0.7.1 中文版 · Mac-Intel 芯片 | `andee-0.7.1-zh-mac-x64.dmg` |
+
+Which one to take is the same question as above. The only other difference is the server, which is
+already baked in.
+
+The file names are ASCII because GitHub deletes Han characters from release asset names, so the Chinese
+wording lives in the label where it survives — and the downloaded file name stays ASCII, which means it
+survives every filesystem it is copied onto as well.
+
+### Opening the Mac app
+
+The `.dmg`s are signed with a local identity, not an Apple Developer ID, so Gatekeeper refuses the first
+launch of a downloaded copy.
+
+⚠️ **On macOS 15 (Sequoia) and later, "right-click → Open" no longer works** — Apple removed that
+override, and most advice online still recommends it. The current route:
+
+1. Double-click the app. It will be refused. Dismiss the dialog.
+2. Open **System Settings → Privacy & Security** and scroll to **Security**.
+3. There is a line naming Andee as blocked, with an **Open Anyway** button beside it.
+4. Confirm, and authenticate.
+
+That button **only appears for about an hour** after the failed launch. If it is missing, launch the app
+again and go straight back to Settings.
+
+One more thing worth knowing: the app **inside** the mounted dmg carries no quarantine flag, so it may
+launch straight from the dmg — and then be blocked once you drag it to Applications, which does attach
+one. Both behaviours are normal; neither means a damaged download.
+
+### What you supply
+
+No package has a key. After installing, open **Settings → Backend** and either paste a CodeFlying key
+(sign up from the link in that panel) or switch to **on-device** and give it an OpenAI-compatible
+endpoint and model. A voice key is optional — without one the assistant is silent but still works on
+screen.
+
+> **Would you rather not configure keys at all?** Use the one-click page instead —
+> **[codeflying.app/andee](https://www.codeflying.app/andee)**, or
+> **[codeflying.net/andee](https://www.codeflying.net/andee)** in mainland China. Describe what you want,
+> tap the build button, and you get a package with the **tokens already configured** plus a QR code to
+> scan. Starting is free and the free token allowance is **finite — once it runs out, tokens must be
+> bought.** It is a separate service with no direct relationship to this project, with its own terms,
+> pricing and data handling. Note it produces **an Android APK, not a Mac app** — macOS packages can
+> only be built on macOS.
+
+## Known issues
+
+- **`tap_target` adds a call per tap.** It is fast (about 2 s) but it is not free, and on a slow backend it
+ is slower than the brain aiming by cell. If that trade is wrong for your setup, do not raise the speed:
+ note that `locator_model` is there to point it at something faster rather than to turn it off.
+- **Nothing verifies that a key is accepted by the server it is sent to.** These packages are checked for
+ the right server and the absence of credentials; whether a given CodeFlying key works against
+ `baas.kuafuai.net` versus `myapp.codeflying.app` has not been exercised end-to-end.
+- **The Mac still falls back silently when no server is configured.** Every package here carries one, so it
+ does not arise for these downloads — but a locally built app with an empty domain resolves to the local
+ backend and sends whatever key it has to `api.deepseek.com`, while the self-check reports "Configured".
+- **Speaking first is rationed, and the ration is deliberately low.** If it feels too quiet, raise
+ **Settings → 主动招呼**; if it feels like too much, that setting is the dial, not a bug. A bubble also
+ cannot re-check itself, so it can arrive after the thing it mentions has stopped being true.
+- **A Mac build made on another machine without its signing keychain is ad-hoc signed** — its identity
+ locks to the build hash, so installing it over an existing copy makes macOS treat it as a **different
+ app**: Accessibility and Screen Recording must be granted again. User data is unaffected; it lives in
+ `~/Library/Application Support/Andee`.
+- **The wake threshold's margin is thin**, carried from `v0.3.2`: the eight measured judgements sit in
+ 1.47–2.05 with the line landing in the gap between 1.69 and 1.77 — 0.08 wide.
+- **A light generated page still fades in over a dark stage.** Carried from `v0.3.1`; the app's own
+ default has been light since `v0.4.1`.
+- **`assets/andee_key.txt` is tracked by git and is not in `.gitignore`.** It ships with a single
+ newline, which reads as empty. A key written into it would outrank every build setting.
+- Two issues from `v0.1.0` are **unchanged and untouched**: `tools/verify_faces_lab.py` still reports
+ 7/16 assertions passing, and the debug WebSocket server (`net/BodyWsServer`) still binds
+ `0.0.0.0:9008` with no authentication (documented in `SECURITY.md`).
+
+## Before you install
+
+Read **[DISCLAIMER.md](../../DISCLAIMER.md)** and
+**[docs/acceptable-use.md](../acceptable-use.md)** first. On Android the app asks for accessibility,
+overlay, notification access, SMS and call log, and it can complete a payment by tapping; on the Mac it
+asks for accessibility, screen recording and notifications. The only guardrail on that path is Android's
+`FLAG_SECURE`; the prompt-level checks on risky actions are instructions to the model, not enforcement.
+
+**Speaking first is not an exception to that.** A bubble is text on your screen, and tapping a button on
+it starts a turn **at the trust of the turn that wrote the bubble** — a bubble written by a turn that was
+not allowed to send or pay cannot get you to send or pay by your tapping it either.
+
+## Install from source
+
+```bash
+git clone https://github.com/kuafuai/andee
+cd andee
+git checkout v0.7.1
+
+# one variant, by hand:
+./gradlew :app:assembleDebug -PdefaultLang=en -PcodeflyingDomain=myapp.codeflying.app -PcodeflyingSsl=true
+
+# or all ten files, the way the release was built:
+bash ~/.workbuddy/skills/andee-release-ship/scripts/build_variants.sh
+```
+
+`-PcodeflyingKey=` and `-PgoogleVoiceKey=` are refused on purpose — a published package carries no
+credential.
+
+The measurement and the reasoning behind the new aiming, including the approaches that were deliberately
+not taken, are in **[docs/design/click-accuracy.zh-CN.md](../design/click-accuracy.zh-CN.md)** (Chinese).
+
+**Upgrading:** `versionCode` moves `16 → 17`, so the Android package installs over v0.7.0 normally. The
+Mac app's version string follows the same number. Existing memory, notebook, todo list and chat history
+are untouched and are not migrated — nothing about their format changed.
+
+**Full Changelog**: https://github.com/kuafuai/andee/compare/v0.7.0...v0.7.1