emb_cli 0.3.6

emb_cli: ^0.3.6 copied to clipboard

Two cache-key corrections. Both are about a key that does not name everything the thing it keys was built with, so two different trees can share one entry.

• fix(cross): sysrootKey folds in cpu_flags once an augment stages into the sysroot. Omitting them is right while a sysroot is only an extracted image -- cpu-only variants of one board then share a single extraction -- and stops being right the moment an augment installs into it, because augments are compiled with the target's cpu flags. rpi5 (a76), rpi4 (a72) and rpi-zero-2w (a53) share one raspios image and so one key: prepare any of them and the other two reused a sysroot whose augments were tuned for whichever went first, with nothing detecting it. The wrong instructions surface as an illegal-instruction fault on the board, a long way from the build that chose them. host: true augments do not count -- they install to the workspace's host tools, never the sysroot -- and sysrootBaseKey is untouched, so the expensive half, the image download and extraction, stays shared as before.

• fix(cross): augmentOverlayKey folds in the sysroot and, where it matters, the host architecture. It hashed the target triple, cpu flags and augment set, which lets two overlays holding different binaries share a key: a bookworm and a trixie variant of one board match on all three and are kept apart today only by their differing augment lists. A host: true augment produces build-machine binaries, so an overlay built on x86_64 is unusable from an aarch64 host; that input is folded in only when the staged set contains such an augment, so a target-only overlay stays shared across hosts. Nothing consumes this key yet, so it is a correction ahead of use rather than a behavior change.

• fix(cross): --publish no longer skips a publish that leaves the shared cache empty. The fast path decided there was nothing to do from the image alone, but the image carries no toolchain — its dockerignore is * and it copies nothing — so the sysroot base and toolchain reach a consuming build through the shared cache instead. Skipping on the image alone left that cache empty permanently: the push that fills it runs only after a resolve populates the local store, and the resolve is exactly what the skip avoids, so the image kept looking up to date and the skip kept firing. The fast path now requires both halves, and a transport error counts as not-ready — an unreadable registry is a reason to do the work, not to assume it has been done.

• fix(cache): the shared-cache push works at all. Every push had failed since the transport was written, with absolute file path detected from oras push: the layer is a temp file this process has just written, so its path is absolute by construction. The failure was invisible because the push is best-effort — it warned, the build carried on, everything stayed green, and the cache the transport exists to fill stayed empty.

• fix(cache): the pushed layer can be pulled again. The path handed to oras push becomes the artifact's title, and oras pull uses it to decide where to write, so an absolute one asks every consumer to write outside its working directory and is refused. The first attempt at the previous fix disabled the push-side validation, which moved the failure downstream and made it worse: the artifacts reached the registry and could not be pulled at all. The layer is now pushed under its bare filename, with oras run from its own directory, so the title is relative and the artifact pulls anywhere — no flag on either side.

Together these are one bug in three parts: the shared cache had never been written to. Measured on ivi-homescreen, every consuming build re-resolved its toolchain from scratch — 91.8s on rpi4, 108.8s on rpi5, 97.3s on rpi-zero-2w, per job, roughly forty minutes of runner time per run.

• fix(cross): --publish no longer skips a publish that would move a mutable tag. The skip-on-exists fast path probed only the content-addressed tag and returned before applying the others, so whatever --tag <name> pointed at when it was last really pushed, it kept pointing at — the content tag still existed, so the skip kept firing, so the push that would move the name never ran. Anything consuming the image by that name got an older one indefinitely. The skip now compares digests and fires only when every requested tag already resolves to the content tag; a digest that cannot be read counts as a reason to republish rather than as a match. The first publish after a tag has drifted does a full resolve — that is the run that repairs it.

• chore: require packagekit_dart 0.5.0 and winget_dart 0.4.0, which migrate their native build hooks to the hooks 2.x native-assets API (and drop the unused native_toolchain_c). No behavior change for emb — both host-provisioner backends build the same native bridge and still degrade cleanly when their toolchain is absent. emb already required Dart 3.10, which hooks 2.x needs.

• feat: build the Flutter engine for riscv64 and armv7hf, in both glibc and musl. The single source-build recipe now produces a green, C-ABI-clean engine for all seven arch/libc combinations in scope — x86_64, arm64, riscv64, and armv7hf, each in glibc and musl — every one validated with a from-scratch build. armv7hf selects hard float (its gnueabihf sysroot is hard-float, but gn defaults arm to softfp); riscv64 applies the gn-riscv and swiftshader-llvm-16 patches (its Reactor JIT needs llvm-16); and musl targets add a compiler-rt builtins fallback and turn off the glibc-only mallinfo/execinfo assumptions in swiftshader's vendored LLVM.
• fix: correct the 0.3.1 note that an x86_64-musl target is unsupported. When the build uses a custom target toolchain, the engine's gn already builds host tools with the glibc host toolchain, so an x86_64-musl target builds like any other — it only needs the same complete musl sysroot (libc++, libunwind, gcc) the other musl targets do.

• feat: build the Flutter engine for musl targets. The source build now handles a musl libc target — routing the in-tree libc++ to its native musl path, fixing the flatbuffers and swiftshader glibc assumptions, dropping the glibc execinfo backtrace, and applying its patch series with plain patch so it reaches nested engine DEPS repos. Validated with a from-scratch arm64-musl build whose libflutter_engine.so loads and runs a Dart app headless in an Alpine arm64 container, within 2.3% of the published glibc engine for the same commit. musl targets must be embedded architectures (arm64/riscv/armv7): an x86_64-musl target collides with the x86_64 host toolchain.
• fix: the engine ABI gate no longer false-fails a clean musl engine. Rule 6 treated libc's __cxa_atexit / __cxa_finalize / __cxa_thread_atexit_impl as a libstdc++/libc++abi dependency, but those are the C runtime's static-destructor registration, provided by libc itself. The earlier glibc fix filtered version-tagged symbols, but musl carries no symbol version, so a clean musl engine tripped the rule. A genuine undefined __cxa_throw still fires.

• feat: build the Flutter engine from source when no prebuilt is published. When emb engine finds no published engine SDK for a (commit, arch, mode), it can now build one on demand as an opt-in augment, producing a drop-in artifact that aot/bundle consume unchanged. A fail-closed ABI gate (readelf/nm) verifies the built libflutter_engine.so is a self-contained, C-ABI-clean drop-in before it is staged. The single tool/engine/build-engine.sh recipe — a fetch/build split mirroring Yocto's do_fetch/do_compile, with --offline/--offline-strict enforcement — is shared by emb, meta-flutter/flutter-engine CI, and a Yocto recipe, so the orchestration is not triplicated.
• feat: route Linux/x86_64 aot/build/bundle into a glibc container on a host that cannot run them natively (non-Linux or non-x86_64), re-entering emb with --exec-native behind an EMB_IN_CONTAINER guard so the inner run does not recurse.
• feat: emb doctor --offline-probe --engine-commit <sha> certifies that a fetched engine source closure can build with the network denied — the local analog of a Yocto do_compile sandbox — with the verdict carried in the --json envelope for CI.
• feat: add an Alpine apk host-dependency backend and a musl -dev sysroot provider. Alpine uses apk and OpenRC, not PackageKit/systemd, so host provisioning drives apk directly — no D-Bus, no systemd, no native bridge — and the musl cross sysroot is built from apk packages, the musl analog of the Debian dpkg path.
• feat: require packagekit_dart 0.4.3, whose build hook now skips the native bridge cleanly when cmake/ninja/libsystemd are absent. Without it, installing emb aborted in that hook on Alpine, minimal images, and any non-systemd host — even though emb never uses the PackageKit backend there — which is what the engine builder/runtime container images rely on.

• fix: the update notice is no longer printed under --json. It is written to stdout, so it landed after the envelope and made the output unparseable for exactly the callers who asked for machine-readable output.
• fix: only a newer published version is offered as an update. The check compared for inequality, so any unreleased build — every developer on main, and every release branch before its publish — was told to "update" to the older published version.
• feat: emb doctor reports the board library — resolved source rung, install path, board count, and any version skew, in both the text and --json output. A reachable emb with no board library is the same false green the authorization probe exists for: nothing looks wrong until a manifest uses extends:.
• fix: commands that take no positional arguments now reject stray ones instead of silently discarding them. emb sync boards ran the repository sync and reported "No source repositories found to sync", with no hint that boards had been dropped; it now names the stray argument and suggests emb boards sync.
• fix: extends: now works on an installed emb. dart install produces a standalone binary that carries no package data files, so boards/ never reached it and every board name failed with Known boards: none. The board library is installed to <data-home>/emb/boards and resolved from there.
• fix: an empty board registry no longer reports unknown board "<name>", which read as a typo and sent people to audit a correct manifest. It now says the library was not found and lists every path tried.
• feat: emb boards list (what is loaded, from which rung, with any version skew) and emb boards sync (install the library for a pub.flutter-io.cn install, which has no checkout to copy from).
• feat: add --[no-]interactive to deps, setup, and cross, plus the EMB_NON_INTERACTIVE environment variable, and honor NONINTERACTIVE=1 (Homebrew's convention) as a secondary. DEBIAN_FRONTEND is deliberately not consulted: it suppresses the debconf package-configuration prompts rather than authorization, and images set it globally for that unrelated reason. Interactive is the unconditional default; nothing about the environment changes it, and CI-style variables are deliberately not consulted. Requires packagekit_dart 0.4.0, which sends the polkit interactive hint — without it no emb deps could install a package on a stock Fedora/openSUSE desktop unless the caller was root or covered by a permissive polkit rule.
• feat: decouple --yes from interactivity. --yes skips emb's own confirmation prompt only; --no-interactive implies --yes because nobody is there to answer, but the converse does not hold.
• feat: brew sets NONINTERACTIVE=1 when non-interactive. WinGet accepts the flag for interface parity but has no equivalent to apply it to.
• feat: classify install failures. ProvisionResult.kind carries a typed ProvisionFailure (notAuthorized, unresolved, daemonUnavailable, other) derived from the backend's error codes, so the command layer no longer string-matches error text that varies by backend.
• feat: on an authorization failure, print remediation instead of a raw exception. The advice branches on the interactivity mode, because the daemon reports the same error whether it could not prompt or prompted and was refused.
• feat: emb doctor reports whether you are authorized to install host packages. Its other checks only exercise read-only operations, which need no authorization, so a reachable backend was previously reported green even when emb deps could not install anything. The probe never prompts.
• fix: defer the install spinner until the backend reports progress. An authorization prompt can appear first, and an animating spinner drew over it and made the password prompt unreadable.
• docs: document authorization, including why polkit cannot prompt under WSL and the narrowly scoped polkit rule that works there, plus what CI needs.
• fix: the documented polkit rule now matches both wheel (Fedora/RHEL/Arch) and sudo (Debian/Ubuntu). A wheel-only rule silently does nothing on Debian, which is hard to diagnose because the rule looks installed.
• test: CI job exercising authorization against a real PackageKit daemon — refusal with remediation before the rule is installed, unattended success after, and EMB_NON_INTERACTIVE equivalence.

• fix: locate the PackageKit native bridge (libpackagekit_nc.so) when emb is installed from pub.flutter-io.cn via dart install emb_cli. The code asset ships in the dart install app-bundle's lib/ next to the executable; the resolver now finds it there (relative to Platform.resolvedExecutable). Previously the Linux backend reported "packagekit is not available" unless PK_NC_LIB or LD_LIBRARY_PATH was set by hand.

Port of meta-flutter/workspace-automation to a Dart CLI.

• feat: emb doctor — host detection + package-backend availability.
• feat: emb deps — coalesce host deps across manifests, filter, install in one PackageKit transaction (with WhatProvides resolution).
• feat: emb sync — concurrent git clone/update into app/.
• feat: emb flutter — install the Flutter SDK (auto-resolves the engine commit).
• feat: emb engine — fetch prebuilt engine artifacts from meta-flutter/flutter-engine releases (auto fetch-else-build).
• feat: emb aot — create_aot.py port producing libapp.so (profile/release).
• feat: emb bundle — assemble an ivi-homescreen bundle (debug JIT / AOT), with implicit engine fetch and cross-compile (x86_64 → arm64/riscv64).
• feat: emb build — manifest-driven build (build: block) over an arch × mode matrix.
• feat: emb cross — resolve a manifest cross: block into a toolchain + sysroot profile: cross-compile the native embedder for arm64/riscv64 with rootless sysroot extraction and Debian .deb handling.
• feat: emb fetch — fetch a cross target's toolchain + sysroot closure into the store: the online acquisition step for an offline build.
• feat: emb matrix — render a CI build matrix from manifest cross: blocks.
• feat: emb cache — inspect and reclaim the shared artifact cache.
• feat: emb update — self-update the CLI.
• feat: emb setup — one-shot provision; emits setup_env.sh.
• feat: emb env — write setup_env.sh.
• feat: install with dart install (Dart 3.10+), or with zero preinstalled toolchain via bootstrap.sh/bootstrap.ps1, which fetch a pinned Dart SDK and put emb on PATH in one step.
• feat: carry native code assets through cross-compiled bundles — stage them onto the loader path, build the Linux asset bundle, and accept native_assets.json for the kernel compile.
• feat: depend on the published packagekit_dart ^0.3.2 (hosted, not a path dependency).
• feat: auto-resolve the PackageKit native bridge (libpackagekit_nc.so) — from the package's own build or, for hosted installs, the package:hooks build-hook output under .dart_tool/ — so the Linux backend works without PK_NC_LIB.
• docs: full per-command reference (every option, default, and value set) in the README; add example/.

• feat: initial commit 🎉