3aa6f2f290ff9d11f3c956185aae5123495a27fc
14
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
  irisandClaude Opus 5
|
3aa6f2f290 |
Move a project's checkout to a branch or commit from the phone
The project settings sheet now lists the checkout's branches and its last fifty commits, and picking one moves the checkout on the build machine and builds what that leaves behind. It is the same act as a pull on the same single checkout, so it goes through the same machinery and reports in the same place: `after_moving` is the half the two share, which is not the git command but handing the run over under one lock, so nothing can observe the moment between the move ending and the builds it decided on starting. The two lists are read when the sheet opens rather than riding on the manifest, since both spawn git and the manifest is fetched on every open, resume and Refresh. Local reads only, so opening the sheet cannot stall on a network round trip: what the remote has and this checkout has not fetched is Pull's business, and the card already says when there is any. Listing branches was wrong twice in ways only a real checkout showed. `git branch --format` adds a `(HEAD detached at abc123)` pseudo-entry that is not a branch and cannot be checked out, so the list comes from `for-each-ref refs/heads`. And `refs/remotes/origin/HEAD` abbreviates to a bare `origin`, so filtering names that end in `/HEAD` matches nothing and a phantom branch called `origin` reached the phone; it is dropped by being a symref instead. The test written to cover the second asserted the same wrong thing and passed. Moving is refused over tracked modifications only, where a pull refuses over any dirt at all. The strict check makes this a one-way door: going back to a commit from before the .gitignore that covers this project's build output leaves that output untracked, so the tree is dirty and every move afterwards is refused -- back but never forward, from a phone, with the way out being the build machine. Nothing is lost by relaxing it, because git refuses to overwrite an untracked file itself and carries across the ones it would not, and its refusal arrives as the error the card already shows. Found by moving a checkout back and forth rather than by reading it: the first version passed its tests and trapped the checkout on the second move. Picking a commit leaves the checkout on no branch, which the card now says in those words rather than showing git's literal `HEAD` -- beside a branch icon that reads as a branch somebody named HEAD. Worth saying now that the sheet can produce the state, where before it was reachable only on the build machine. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
irisandClaude Opus 5
|
17d873ab7c |
Give each component build modes and a settings sheet
A component now declares its ways of being built in one list -- `modes: ["release", "debug"]`, the first the default -- and every command it runs is handed the mode as its last argument, so a project whose script takes `release` or `debug` names that script once. A field may instead be written per mode, which is the escape hatch for the commands that cannot take the word: cargo takes `--release` or nothing, and its profile for the unoptimised build is called `dev` while the directory it writes is called `debug`, so no single word serves as both the flag and the path. A command written per mode is not handed the word as well; it already is the answer, and a stray argument to a service binary is a process that will not start. One list rather than gathering names from whichever fields happened to mention them is what makes a mode missing from one part unsayable: every per-mode map is checked against it, so a gap is named rather than resolved to some other mode's command. Which mode to build in and whether to strip are the build machine's -- there is one checkout and one set of outputs, so a per-device mode would have two phones rebuilding over each other silently. Which of the finished builds a phone installs stays that phone's. Neither choice is part of the acceptance gate, so both have to be carried across the components list being rewritten, by acceptance and by the self entry's startup reconciliation alike; without the second a mode chosen for this server's own component would not survive the restart that applies it. A mode switch moves no commit, so `builtMode` is recorded beside `builtFrom`. Without it a component built in debug and switched to release reads as current and serves the debug build for ever -- and for an APK nothing else notices, because the "never built at all" check finds any variant under the component's directory. Each component card gains a settings sheet behind a gear at the row's right-hand end, holding the mode, which build to install, strip, and an Enrol button. The variant picker moved into it: on the card the two read as one choice, both saying debug and release, and they are not. The row's text is now bounded and truncates, so a long status can no longer push the log and settings buttons off the edge. `enroll:` is a declared command whose one line of stdout is a URL for the phone to open after installing -- generic on purpose, run per press since such a link is one-shot and carries a credential, and in the acceptance gate because it runs on the build machine. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
irisandClaude Opus 5
|
0b7164bb30 |
Say on the card when a project's own file was ignored
An unknown field discards the whole declaration rather than the key it did not recognise, so a machine that takes a project's commit before it takes a dev-updater that understands it loses that project's components, strip, staleWhen and resources at once. Nothing looks broken afterwards: the card becomes a project that declares nothing, which is the ordinary case, and the only thing saying otherwise was a line in this server's log on a machine the person holding the phone cannot read. project_config now answers a ProjectFile carrying the parse error beside the declaration, and AppEntry::declaration_state returns both halves from one read -- the manifest wants both, and the file is on that path. The card draws its own sentence about what it means and what to do, then the parser's own words, which name the file, the position and the offending field; that half is selectable, since the fix happens on the other machine. Nothing is blocked by it. What was already accepted is what runs, so an unrelated typo cannot stop a project that was working -- asserted in the test, along with the error naming both the field and the file. Seen on the emulator both ways: a project whose file gained a field from the future says so under its action row, and the whole message goes when the file parses again. Raised by the tdep-survey session after its alsoWatch commit hit exactly this against the older binary running here. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
irisandClaude Opus 5
|
c7b3af4861 |
A Rebuild on every component, and the keystore AGP actually uses
Two things, both from the same stuck download. Each component row now carries its own Rebuild, which is /build scoped to it -- the route already took ?component=, only the button was missing. Update builds only what the staleness rules call behind, and those rules cannot see a command reading undeclared files, an output changed underneath this server, or a signing key replaced since the APK was made; in all of those Update does nothing and the only force was a project-wide Rebuild that rebuilds every sibling. Same action at two scales, so it takes the project Rebuild's word and colour. hasBuild is new on the manifest component so a component with no command of its own draws none, which the project-level needsBuild could not answer. And debug_keystore_path() now resolves the preferences directory the way AGP does instead of assuming ~/.android: ANDROID_USER_HOME as-is, then ANDROID_PREFS_ROOT and ANDROID_SDK_HOME with .android appended, then $XDG_CONFIG_HOME/.android when that directory exists, then $HOME/.android. Measured by running real builds against AGP 32.3.2, including the existence test on the XDG step, which the documentation does not mention. An ordinary desktop machine with XDG_CONFIG_HOME set signs its APKs with a keystore this server never looked at, so it reported "no debug keystore" about a machine that had one and was using it -- and this VM sets no XDG_CONFIG_HOME, which is why it could not happen here. The failure now names a keystore found further down the list, since an inert one and an absent one are otherwise identical from a phone. Verified on the emulator (component Rebuild: press, that component's own progress, back to resting with no sibling moving) and against a throwaway server for both keystore branches. Reported by the tdep-survey session, which found the two keystores on the serving host. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
irisandClaude Opus 5
|
b67c70fa2f |
A component can name the files its build reads elsewhere
Freshness compared the commits under a component's own cwd, which is right about where its files are and wrong about what its build reads. tdep-survey's two clients each source one shared scripts/android-sdk-env.sh: a fix committed there changed what every build does, moved no component's subtree head, rebuilt nothing, and left every card correctly reporting "current" while answering a narrower question than the reader was asking. alsoWatch names the rest. The paths join the component's own in the same subtree-head and dirty comparisons -- git takes several pathspecs, so it stays one call each -- and it is part of the acceptance gate like any other declared field. Declared rather than inferred: which files a build reads is not knowable from here, and both wrong guesses are expensive. Raised by the tdep-survey session, which traced why its dioxus client could never become stale on the serving host. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
irisandClaude Opus 5
|
f00c3b970f |
Name the button that gets past a signature mismatch
The advice was "rebuild this app on this machine", which somebody can follow and still be stuck: nothing about a changed keystore moves the checkout, so the component is not stale and neither a pull nor Update runs the build. Rebuild is the one control that builds regardless, so the message says so. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
irisandClaude Opus 5
|
141402bcd2 |
Refuse to serve a slim APK signed with the wrong key
strip.rs re-signs with this machine's ~/.android/debug.keystore on the assumption it is the key the build was signed with, and never checked. Where it isn't -- a keystore recreated after the APK was built signs as an entirely unrelated certificate -- the phone gets an APK Android refuses, reporting "App not installed" with no cause, which reads as the download rather than the signing. Both certificates are now read with `apksigner verify --print-certs`, compared as sets of digests so a v2 source and a v3 output still match, and a mismatch is refused with both digests and the keystore path in the message. The slim copy is deleted on refusal: nothing would serve it without its stamp, but serveable_now reports the size of whatever slim file is on disk. Verified both ways against tdep-survey's app-dioxus: with a fresh keystore under a throwaway HOME the download 500s with the message and leaves nothing behind, and with the real one it serves the same 52,685,076 bytes as before, signed by the same certificate as the raw build. Raised by the tdep-survey session, which measured that a fresh debug keystore is unrelated to the existing one. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
irisandClaude Opus 5
|
0e9c7842f7 |
Find the SDK without an environment, and say what broke
A service manager hands its process a scrubbed environment -- measured in
the Gentoo guest, an OpenRC user service gets 19 variables and neither
ANDROID_HOME nor ANDROID_NDK_HOME among them -- so the tool lookups have
to stand on their own. Two ways they did not:
- $ANDROID_HOME was taken as given. This machine exports a system-wide
/opt/android-sdk whose build-tools/36.0.0 holds nothing but
package.xml, so the download failed with "failed to spawn
.../zipalign", naming the tool rather than the root that was wrong.
Each lookup now takes the first candidate that actually contains what
it needs, and names every place it looked.
- The NDK search took the newest child of ~/Android and filtered it
afterwards, which can only reject that one directory. It worked by
the luck of `Sdk` sorting before `android-ndk-r27c`; a lowercase
neighbour would have hidden an NDK that was sitting right there.
newest_child_where filters before taking the max.
And the failure had nowhere to go: ApiError::Internal answered with an
empty body, so a missing NDK reached the phone as "Server returned HTTP
500" while the message naming the path stayed in a log on a machine the
person holding the phone cannot see. It now answers with the same
{err:#} chain it logs. Every route is behind the bearer token of a
device somebody enrolled themselves, so there is no third party to
withhold it from.
Diagnosis from the tdep-survey session, which measured the OpenRC
environment and the /opt/android-sdk contents.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
||
|
iris
|
aa8e2b97a5 |
Render the escapes in an error, and drop the manifest field nobody read
Two loose ends from the per-component work. The card's error is the tail of what the build actually printed, and a compiler marks its own errors in colour -- so it arrived as `[1;31merror`, with punctuation welded onto the one line somebody is trying to read. It now goes through the same `ansiAnnotated` the log dialog uses: the colour is drawn, and the sequences with no meaning on a phone are consumed rather than printed, so a cursor movement cannot arrive looking like corruption either. One composable covers it because every failure already went through OutputText. Selection copies AnnotatedString.text, which is the message with every escape already gone, so what lands on the clipboard is what was on screen rather than what was on the wire. And `/manifest` carried a `build` object that nothing on the phone ever parsed. It was going to be how a running build survived leaving the app; that is not built, so it is gone rather than left looking finished. `RunningBuild` went with it -- it was split out of `BuildStatus` for that one reader and had become an indirection with one user, so `status()` fills one flat struct again. `/status` is byte-identical either way, since the split was flattened on the wire; checked against the running server rather than assumed. Looked at on the emulator against a failing build: bold red `error`, blue `-->`, no escape text anywhere, and long-pressing it still raises the handles and the Copy toolbar. |
||
|
iris
|
90082bd286 |
Each component builds on its own, and stops reporting when it is done
A project's components were decoupled everywhere except the one place it showed: there was a single build slot per project, and a single card state in the app keyed by project alone. So pressing Update on one client of a two-client project disabled the other client's button for the length of a build it shares nothing with, drew this one's progress bar and download percentage under the other's row, and -- had the button been pressable -- would have been a silent no-op on the server, since a second request while one was running returned without starting anything. The slot is now per component. `Inner` has no `building` flag; a component's own `ComponentRun` with an open `step` is the answer, and `claim` writes that entry synchronously under the lock the route answers from, so nothing can read a just-claimed component as idle -- which the phone would take for a build that had already finished. A failure is recorded against the component whose command it was rather than in the project's one error slot, which two components building at once cannot share. A pull stays exclusive with everything, because there is one checkout and it rewrites the files every component builds from. Releasing it and claiming what it decided to build happen under one lock: a phone polling in the gap would find a project neither pulling nor building and call the run over. The app mirrors the split -- `ProjectState` for the pull and the project-wide Rebuild, `ComponentState` keyed by component for everything one component is asked to do. Two hierarchies rather than one keyed by a pair, so a download has nowhere project-wide to be stored. A component's failure is drawn in its own row beside the Retry that acts on it, which is also where a failed service action now reports. And a finished component shows nothing at all: the elapsed times are gone from both halves of the wire, and its button simply goes back to being pressable. A bar, a count and a last line all describe something happening now, and left up they sit there looking live next to a sibling that genuinely is. Verified on the emulator against test-projects/two-clients, which exists for this: while `tablet` built, its row alone carried the bar and its button alone was disabled, `phone` stayed pressable and silent, and both returned to normal with no timing left behind. |
||
|
irisandClaude Sonnet 5
|
7eaf79370e |
Build one component instead of every one with a command
Pressing Update on one client of a multi-client project ran every
component's build to get the one that was actually asked for -- cheap
while a project had one APK, expensive the moment it had two and one of
them was slow (an ARM cross-compile, say). /prepare and /build now take
?component= to restrict a run to one named component; absent still means
the whole project, which is what Pull & Build, the project-row Rebuild,
and every single-component project keep doing. There is deliberately no
component-scoped Rebuild -- forcing one component's build without
touching the rest happens by pressing Update on it, which now runs
/prepare scoped to that component.
Verified against a live server driving a scratch two-Apk project:
building one component leaves the other's marker untouched, an unknown
?component= answers 404 naming the project and the component, and naming
none still builds both.
That verification surfaced a second, sharper bug the scoping change had
not caused but did make newly visible: component_is_stale's "never built
at all" check still asked find_apks of the whole project root, left
behind when per-component discovery (
|
||
|
irisandClaude Opus 5
|
c4e19c2e2b |
A project can produce more than one APK
tdep-survey is one checkout with a backend and two independent Android clients, so a project has to be able to declare two Apk components. It could not: APK_PATTERNS was anchored at the project root, which reached the first client and stopped, and package, strip, the variant list and the download all came from AppEntry::apk_component()'s first match -- correct only while a project produced one APK. Each component's builds are now found under its own cwd, which is what cwd already meant everywhere else: the directory the build command runs in, the subtree staleness is scoped to, a server's WorkingDirectory. A component that declares none sits at the project root, so nothing about the single-APK case changes. The alternative -- naming the file on the component -- would have made a component a file path, and an app being a project path rather than a file path is this project's central invariant. Everything derived from an APK follows it onto the component: package, previousPackage, strip, size, mtime, variants and the rename note, in a nested `apk` block that a Server simply doesn't have. Two clients install over different packages, so first-wins would have checked the installed state of one and reported it as the other's -- which looks exactly like a correct answer. For the same reason a download naming no component is refused rather than guessed; naming none still answers for a project with one, which is what lets the frozen /self/apk keep working. On the phone the per-device state is keyed by project and component, so the variant picker moved inside the component's own card, beside the build it picks, and the installed state, size and icon are each their own component's. A project building two clients shows no single icon of its own rather than borrowing the first one's. Measured against the real thing: pointed at tdep-survey, the two clients resolve to their own builds (15 MB and 153 MB), strip applies only to the one that asked for it (153 MB served as 52 MB), pressing Install on one row installed that row's package and left the other row offering Install, and a download with no component named answers 400 saying which flag to pass. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
irisandClaude Opus 5
|
db47972a25 |
A component's build progress reads under its buttons, and an unrelated history can be overridden
Two things, both about a card being able to act on what it says. The progress bar, its counts and the last line a component printed now sit below that component's buttons rather than above them. A bar reports on the press that started it, so it reads in the order it happened -- and above, it pushed the buttons down the moment a build began, moving the row somebody had just pressed out from under their finger. A pull that cannot fast-forward because the checkout shares no history with its upstream now offers a way past it. There is no fast-forward between two unrelated histories and there never will be, so the card was one that could never be pulled again, with the only remedy on the build machine -- exactly where the person holding the phone isn't. The card reports the failure as before and a dialog offers a forced pull, naming the branch and the upstream it is about to overwrite; confirming sends ?force=true, which resets onto the upstream instead of merging. Whether it *is* that failure is decided structurally, by `git merge-base` finding no common ancestor, rather than by matching what git printed: those messages are translated, and a button that appeared only on an English build machine would be worse than no button. It travels to the phone as its own field for the same reason. The dirty-tree refusal stays in front of it, so a forced pull can only ever discard something that was committed, and a merely diverged history is not offered it. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
iris
|
b0e83059a3 |
dev-updater: build an app on the machine, install it on the phone
A Rust backend that discovers Android projects under configured roots, builds one on request, and serves the APK over pinned TLS on a WireGuard interface; an Android client that lists what is buildable, watches a build, and installs the result. Enrolment carries the token and the CA, so the phone trusts exactly the machine that issued it and nothing else. `AGENTS.md` is the working guide and `README.md` the configuration reference. The shared tunnel-and-TLS code lives in `vendor/wg-app-link`, which ai-app uses too. History before this point was squashed away, and a stale `config.json` went with it: nothing had read that file since the config moved to RON outside the checkout, and what it still held was one machine's absolute paths and the names of projects on it. |