Files
CST-agent/docs/android.md
T
WHC 60cbce3c5c v0.9.2: files the agent delivers come back as cards, and the saved key is used
A turn could write a report and say "here it is" -- and there was nothing to
open. The engine had no way to hand a finished file to the person, and the
front end had no way to show one. This release adds the whole path: a tool to
deliver, a store that keeps the copy, a card in the transcript, and open /
share through the platform. It also fixes two bugs the phone exposed while it
was being tried there.

deliver_file (engine)
- cst_attachment_t.for_user marks a copy that belongs to the person, not to
  the model; cst_attach_import_artifact() stores any bytes up to
  CST_ARTIFACT_MAX_BYTES (64 MB) with the mime taken from the name's
  extension. Attachments are unchanged (content decides the type, 16 MB).
- It is a read-policy engine tool: path resolves inside the workspace, a
  directory is refused, title (optional) is the display name.
- The wire is what keeps it out of the conversation: parts_add_attachments()
  skips for_user entries without even a "not sent" note, and a turn that only
  delivered a file no longer invents an empty user message. tests/test_artifacts.c
  pins that down end to end, including a session-file round trip.

The card (Qt front end)
- filesText() replaced imagesText(): one line per file -- kind, stored name,
  name, origin, mime, bytes, pixels -- decoded by ImageStrip.qml. A for_user
  entry is a file card (extension badge, name, type and size, Open, Share);
  a picture stays a thumbnail and opens on tap. Both shapes are read, so
  sessions written by an older version still draw.
- AgentCore::openArtifact() / shareArtifact() accept a name only when one of
  the session's own messages carries it, and hand the display name to the
  platform layer for the status line.
- Android opens through the app's FileProvider (content:// + ACTION_VIEW,
  resolveActivity() checked, read permission granted) and shares through the
  share sheet; a desktop uses QDesktopServices / the clipboard.
  CST_APP_NO_OPEN=1 makes --uitest run without opening anything.

Two bugs the device found
- The session writer had its own copy of the attachment JSON and forgot
  `for_user`: the file survived the file, but came back as an attachment --
  no card, and put back in front of the model. It now uses the one writer
  (cst_attachments_to_json()), with the round trip covered by a test.
- engineToolImages() looked the card up with findTool(), which only finds a
  card that is still running -- and the engine reports the result *before*
  the files, so the card was already complete and the files were dropped.
  A live turn therefore showed no card while a replay of the same session
  did. It now matches on the call id.
- A tap that could not work said nothing: ImageStrip.qml's errorText was
  never drawn. The card now carries the reason ("no app on the phone opens
  report.md").

Also
- The saved provider key reaches the engine on every start: AgentCore::open()
  was handed an empty key and the engine never resolves credentials itself
  (only the CLI did), so every start asked for it again. It resolves the same
  order main.c uses (saved key, provider environment, CST_API_KEY) and wipes
  the copy it made; --selftest checks it against a scratch credentials file.
- Android: notification_send reads the Java exception text into its result,
  and keeps failing on the NX789J -- the build() returns null and Qt clears
  the exception before C++ can read it, so the call has to be wrapped in Java
  next (docs/android.md says so).
- --uitest grew the delivered-file case (card, Open, Share, a forged name
  refused), the live order of result-then-files, and the startup row count.

Version 0.9.2.
2026-10-11 16:34:47 +08:00

9.6 KiB

CST-agent on Android

The engine runs on the phone as a plain bionic binary -- no Termux, no proot, no second process for the UI to talk to. It is the same C core the CLI, the TUI and the WebUI use, cross-compiled with the NDK; on the phone the Qt front end (app/) will link libcstcore.a in-process and register its own platform tools under the owner app:android.

Two scripts cover the engine side:

script what it does
scripts/build_third_party_android.sh ABI [all|mbedtls|curl] static mbedTLS + libcurl into third_party/prefix-android-ABI
scripts/build_android.sh [ABI] [--push] [--run] [--run-tests] libcstcore.a, cst-agent, the test binaries; optional push and on-device run

Toolchain

NixOS: nixpkgs carries the NDK (the Android SDK license makes it unfree, hence the flag; the packaged NDK has a patched interpreter, so its clang runs here):

export ANDROID_NDK_HOME=$(NIXPKGS_ALLOW_UNFREE=1 nix build --impure --no-link \
  --print-out-paths 'nixpkgs#androidenv.androidPkgs.ndk-bundle')/libexec/android-sdk/ndk-bundle

adb comes from nix-shell -p android-tools; the builds need cmake, ninja and curl. End to end:

nix-shell -p cmake ninja curl android-tools --run '
  JOBS=16 ./scripts/build_third_party_android.sh arm64-v8a all'
nix-shell -p cmake ninja curl android-tools --run '
  JOBS=16 ./scripts/build_android.sh arm64-v8a --push --run'

--run finishes with the two smoke checks on the device (cst-agent --version and tests/test_embed, the in-process embedding contract); --run-tests builds the whole tests/ suite, pushes it and runs every binary on the phone, printing a pass/fail line per test.

The APK (Qt front end)

scripts/build_apk_android.sh [--install] [--logs] [--clean] configures with the Qt for Android kit, builds the apk target (Gradle, no port opened) and optionally installs it. Bits it needs, all overridable through the environment:

variable what it points at
QT_ANDROID Qt for Android 6.11.2 (aqt install-qt all_os android 6.11.2 android_arm64_v8a -O ~/Qt --archives qtbase qtdeclarative qtquickcontrols2)
QT_HOST the host Qt of the same version (aqt install-qt linux desktop 6.11.2 linux_gcc_64 -O ~/Qt --archives qtbase qtdeclarative qtsvg)
ANDROID_SDK_ROOT SDK with platforms/android-<n> and build-tools/<v>
ANDROID_NDK_HOME the NDK the engine was built with
JAVA_HOME JDK 17+ for Gradle

The SDK without sdkmanager

nixpkgs's full androidsdk pulls the emulator and system images (gigabytes); the two components actually needed are cheap. The platform comes from the nix store (unfree, so the license flag is required) and build-tools from Google's own zip, laid out as the SDK sdkmanager would:

PLATFORM=$(NIXPKGS_ALLOW_UNFREE=1 NIXPKGS_ACCEPT_ANDROID_SDK_LICENSE=1 nix build \
  --impure --no-link --print-out-paths --expr \
  'builtins.elemAt (import <nixpkgs> {}).androidenv.androidPkgs.platforms 2')   # android-35
mkdir -p ~/android-sdk/{build-tools,platforms}
ln -s $PLATFORM/libexec/android-sdk/platforms/android-35 ~/android-sdk/platforms/

androiddeployqt then picks build-tools and platform-tools from there. For build-tools, dl.google.com is the only source: from here it hands out ~100 KB/s per connection, so the script downloads the zip in 16 ranged requests (≈25x faster) and fixes the exec bits Python's zipfile drops.

Host Qt on NixOS

The host Qt is built for a regular distribution: its tools (qmlimportscanner first of all) look for libgssapi_krb5.so.2, libglib-2.0.so.0, libz.so.1, libfontconfig, libX11, libEGL/libGL, libzstd, libbrotlidec, libdbus-1, libfreetype, libstdc++, none of which exist in NixOS's search paths. The tools carry RUNPATH=$ORIGIN/../lib, so symlinking those libraries from the nix store into <QT_HOST>/lib/ is enough:

nix build --no-link --print-out-paths nixpkgs#krb5.lib nixpkgs#dbus.lib \
  nixpkgs#libglvnd nixpkgs#fontconfig.lib nixpkgs#freetype nixpkgs#glib.lib \
  nixpkgs#libx11 nixpkgs#libxkbcommon nixpkgs#zlib nixpkgs#zstd \
  nixpkgs#brotli.lib nixpkgs#stdenv.cc.cc.lib      # then ln -s <lib> ~/Qt/6.11.2/gcc_64/lib/

Two more things the script does on the way:

  • GLESv2/EGL: Qt's FindGLESv2.cmake only searches CMAKE_FIND_ROOT_PATH prefixes (<sysroot>/include), but the NDK keeps the headers in <sysroot>/usr/include -- without -DGLESv2_INCLUDE_DIR/-DEGL_INCLUDE_DIR find_package(Qt6 Gui) fails with "dependency GLESv2 could not be found".
  • -L$PREFIX/lib on every link type: the APK's payload is a shared module, so CMAKE_SHARED_LINKER_FLAGS/CMAKE_MODULE_LINKER_FLAGS need the prefix too, not just the EXE flags.

Gradle

The generated project resolves plugins and dependencies from google() and mavenCentral(), plus a Gradle distribution from services.gradle.org. From here all three are slow (dl.google.com ~100 KB/s), so two local tweaks:

  • ~/.gradle/init.d/*.gradle (machine-local, not in the repo) prepends the Aliyun/Tencent mirrors for pluginManagement, dependencyResolutionManagement and every buildscript.
  • the wrapper in <QT_ANDROID>/src/3rdparty/gradle/gradle/wrapper/ points at the Tencent Gradle mirror (51 MB/s vs 100 KB/s).

scripts/gradle-android-init.gradle (installed into ~/.gradle/init.d/ by the build script) switches off release lint: lintVitalAnalyzeRelease resolves its own tool jars straight from dl.google.com, bypassing every repository the build configures, and fails the build after all the real work is done.

Signing

androiddeployqt copies the unsigned Gradle output and signs it itself (QT_ANDROID_SIGN_APK + the QT_ANDROID_KEYSTORE_* variables; a signed Gradle output makes it fail with "Could not copy unsigned APK"). The script exports those variables and creates the standard Android debug keystore (~/.android/debug.keystore, alias androiddebugkey) on first use; CST_ANDROID_KEYSTORE and CST_ANDROID_KEY_ALIAS/_PASS override it. apksigner is a bash script with a #!/bin/bash shebang, which NixOS does not have, so the script rewrites the shebang of the build-tools wrappers.

What is Android-specific

  • API level 24 (API_LEVEL=24): bionic provides threads, exec and /system/etc/security/cacerts, which is where libcurl is pointed (CURL_CA_PATH). Bump the level if a newer syscall or CA layout is needed.
  • No TUI, no WebUI: -DCST_BUILD_TUI=OFF -DCST_BUILD_WEBUI=OFF -- there is no ncurses on Android, and the phone front end talks to the engine in-process instead of over HTTP. src/webui_stub.c compiles in place of the server.
  • Sandbox: src/sandbox_exec.c only enables the seccomp/Landlock paths under defined(__linux__) && !defined(__ANDROID__), so the phone build gets the plain exec path -- the same rule the Termux build follows. The seccomp supervisor used to compile in and the probe reported "seccomp", but on the device a write outside the workspace was not denied (tests/test_sandbox.c:308), so the phone build now reports no sandbox instead of one that does not hold.
  • sashidown is built per ABI (vendor/sashidown/libsashidown-<abi>.a) with the NDK's aarch64-linux-android24-clang wrapper: CMake's compiler is plain clang and keeps --target/--sysroot in its own variables, which the vendored makefile cannot see. The host archive stays host-built.
  • Static libcurl needs -L$PREFIX/lib on the link line, because curl's own CMake config lists its mbedTLS dependencies as -l flags rather than paths.

Status

  • arm64-v8a, Android 16 (NX789J): cst-agent runs, tests/test_embed passes.
  • The whole suite runs on the device (scripts/build_android.sh arm64-v8a --run-tests): 36/36 after the diff test learned to tell toybox's diff from GNU diffutils.
  • The APK builds, installs and runs: the Qt front end links the engine in-process (no port), main.cpp puts the workspace, config and tmp under /data/user/0/xyz.whcself.cstagent/files/, and a turn does a real HTTPS request from inside the ART process. The saved provider key is resolved into the engine's config on every start (AgentCore::open(): saved key, the provider's environment variable, CST_API_KEY), so the phone answers without being asked for the key again. Rotation keeps the session; the sidebar appears in landscape.
  • Platform tools behind app:android (app/src/platform/android/AndroidTools.cpp): clipboard_get / clipboard_set under the engine's own tool names and share (ACTION_SEND + chooser) both work on the device. The file card's Open builds a content:// URI through the app's FileProvider (<package>.qtprovider, Qt's @xml/qtprovider_paths), checks resolveActivity() before starting ACTION_VIEW and grants the reader permission; Share reuses the share sheet. notification_send still fails on the NX789J: the channel and the small icon are found, Notification.Builder.build() returns null, and Qt clears the Java exception before C++ can read it ("no exception detail"). Catching it needs the call to be wrapped in Java (a small app/android/src/... helper class), which is the next step there.
  • A session's transcript replays a delivered file (deliver_file) as a card with Open and Share -- the copy lives next to the session, marked for_user so no request ever contains it (see app/README.md, "Files the agent delivers").
  • Still to come: the SAF picker for attaching a picture from another app, the back gesture, and the icon glyphs -- ThemeIcon finds no icon files on the phone and draws empty boxes where the desktop shows Breeze icons.