Repository navigation
Cut CI from 44 minutes: build once, make the ghostty cache reachable, skip docs-only runs - #37
Merged
Merged
Conversation
Every pull request runs the whole job: build libghostty-vt, three package test suites, an app build, the notices check, and a full UI pass on an iPhone and an iPad. The last of those alone is 31 minutes. A change to a Markdown file cannot affect any of it. It is gated per step rather than with `paths-ignore`, which is the obvious way and is wrong here. `ios` is a required status check on main and `enforce_admins` is on, so a workflow that never runs leaves that check pending forever — and a documentation-only pull request would become unmergeable by anybody, including the repository owner. A job that runs and skips its steps still reports `ios`, so branch protection is satisfied. What counts as unable to reach the build is deliberately narrow: Markdown anywhere, and `LICENSE` at the root. The app's sources are `App/`, its only extra bundled resource is the shortcut, and the notice generator reads `Package.resolved`, `project.yml`, `scripts/build_ghostty_vt.sh` and `scripts/third_party_licenses` — none of them Markdown. `LICENSE` at the root is not the file the app bundles; that is `App/Resources/ThirdPartyNotices.txt`, which is not Markdown and so still builds. A pull request touching one Markdown file and one source file builds, and so does an unexpectedly empty file list. The condition tests `grep -v`'s output rather than its exit status. On the machine this was written on, `/usr/bin/grep` is ugrep, where `grep -qv` returns 1 on input that `grep -v` reports with 0 — the `-q` form would have skipped a pull request that changed a README *and* a Swift file. The runner's own grep does not behave that way, so this would have been invisible in CI and wrong only for whoever ran it locally. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…eachable Measured on #34's run — 44 minutes, of which the two test destinations were 31. Splitting that 31 by log timestamps says half of it never ran a test: iPhone prepare simulator 3m31s, build and install ~5m, tests 5m03s iPad prepare simulator 3m31s, build and install 3m45s, tests 9m53s The rebuild before the iPad's first test is the part worth explaining. A destination change gives Xcode a different build description, so `generic/platform=iOS Simulator` for `make build`, then the iPhone, then the iPad each built from scratch into the same derived data — three builds of the same sources, two of which produced nothing the previous one had not. So the app and its test bundles are built once, for both destinations, with `build-for-testing`, and each idiom runs `test-without-building` against it. `make build`'s generic destination goes with it: the two concrete simulators cover the same compilation, and what they do not cover is a simulator architecture no runner has. The iPad step now runs even when the iPhone step failed. The build is already paid for by then, and one run that reports both idioms is one round trip rather than two, which is the thing that actually costs time when a pull request is waiting. Separately, every pull request was spending about four minutes installing Zig and compiling libghostty-vt at a pin that has not moved in weeks. Not a key problem — a scoping one. `ci.yml` only runs on `pull_request`, so the only caches it ever wrote were under `refs/pull/N/merge`, which no sibling pull request can read: five entries existed, under five different pull requests, all under the same key, none of them reachable by anyone. cache-ghostty.yml writes one from a push to main, where every pull request opened against main can restore it. Untouched on purpose: the two `scripts/prepare_simulator.sh` calls are 3m31s each and still sequential. Preparing them concurrently means two processes writing `com.apple.iphonesimulator DevicePreferences` at once, and that preference is the one keeping the hardware keyboard away from the test that #32 added. Seven minutes is not worth reintroducing that. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The iPhone and iPad passes were sequential and were half the run on their own. They share nothing except the build, so they are now a matrix of two jobs, with `fail-fast: false` so one idiom failing still reports the other — one round trip rather than two, which is what actually costs time when a pull request is waiting. Four jobs now: scope the docs-only decision, on Linux, seconds packages the three package suites, which never needed the app build build-for-testing once, the notices check, and the bundles test iPhone and iPad in parallel, test-without-building `packages` no longer waits behind an app build it does not use, so a package failure surfaces in about six minutes rather than ten. The build is `generic/platform=iOS Simulator` again, which is right for this shape: `build-for-testing` produces one .xctestrun both idioms consume, and the test jobs never build, so no destination gives Xcode a reason to rebuild. Both runners use the same working directory, so the .xctestrun's paths resolve on the machine that downloads it. `ios` is kept as a job whose only purpose is to carry the verdict. It is the required status check on main and `enforce_admins` is on, so exactly one check has to report under that name however the work is arranged — and it has to report on a documentation-only pull request too, where every other job is skipped. A required check that never reports leaves the pull request unmergeable by anybody, including an administrator. Simulator preparation starts in the background before the artifact download rather than after it, because it is mostly waiting for a boot. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`generic/platform=iOS Simulator` cannot build AppTests: it reaches into the app with `@testable import iOSSH`, and the generic destination builds for more than one architecture, which leaves the test host's module unresolvable — "unable to resolve module dependency: 'iOSSH'". The simulator picked here is neither booted nor the device the tests run on. Every simulator on a runner is the same architecture, so the one .xctestrun this produces still runs on both idioms. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Each `run:` step is a separate process, so backgrounding the preparation in one step and waiting for it in the next hands `wait` a pid that is not its child: "wait: pid 1666 is not a child of this shell", exit 127, on both idioms. The overlap it was buying was a minute or two of boot against the artifact download. Not worth carrying a process across steps for, so it runs in order. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Merged
m96-chan
added a commit
that referenced
this pull request
Sep 15, 2026
… not there (#42) main has been red on every push since #38. The `scope` job reads the changed files from `repos/OWNER/REPO/pulls/NUMBER/files`, and on a push event `github.event.pull_request.number` is empty, so the request goes to `/pulls//files` and comes back 404. `scope` fails, every job that needs it is skipped, and the `ios` check reports the failure. #37 added `scope` for pull requests and #38 added the push trigger. Each was green on its own; the combination was never run before it was on main. A push has nothing to skip anyway. What reached main is worth building, and that run is the only thing that reports on it now that the required check cannot. Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Every pull request runs the whole job — libghostty-vt, three package suites, an app build, the notices check, and a full UI pass on an iPhone and an iPad. #34 took 44 minutes; the UI pass alone was 31 of them. A change to a Markdown file cannot affect any of it.
Why this is gated per step and not with
paths-ignorepaths-ignoreis the obvious implementation and it would have quietly broken the repository.mainrequires theiosstatus check, andenforce_adminsis on. When a workflow is skipped by path filtering the check is never reported at all, so it stays pending forever — and a documentation-only pull request becomes unmergeable by anybody, including you. A job that runs and skips its steps still reportsios, so branch protection stays satisfied and the job costs a few seconds.What counts as unable to reach the build
Deliberately narrow: Markdown anywhere, and
LICENSEat the repository root.The app's sources are
App/, its only extra bundled resource is the shortcut, and the notice generator readsPackage.resolved,project.yml,scripts/build_ghostty_vt.shandscripts/third_party_licenses— none of them Markdown. The rootLICENSEis not the file the app bundles; that isApp/Resources/ThirdPartyNotices.txt, which is not Markdown and therefore still builds.Checked against the cases that matter:
README.md docs/APP_STORE.mdLICENSE README.mddocs/DEVELOPMENT.jp.mdREADME.md App/ConnectionModel.swiftApp/Resources/ThirdPartyNotices.txtscripts/prepare_simulator.shOne thing worth knowing
The condition tests
grep -v's output, not its exit status. On the machine this was written on/usr/bin/grepis ugrep, wheregrep -qvreturns 1 on input that plaingrep -vreports with 0. The-qform — which is what I wrote first — skipped a pull request that changed a README and a Swift file. The runner's own grep does not behave that way, so it would have been green here and wrong only for whoever ran it locally.Verification
This pull request is not documentation, so it runs the full job — which is the evidence that the gate does not break a code change. The skip path is exercised the first time a docs-only pull request runs it; #33 is exactly that shape and is a good first subject once this lands.
🤖 Generated with Claude Code