Understand any codebase instantly. System intelligence for codebases, built for humans and AI.
TypeScript aiai-memoryclaude-codeclicode-analysiscode-mappingcode-memorycodebase-understandingcodexdeveloper-toolsllmopenclawpersistent-memoryprogram-analysissoftware-architecturesystem-intelligencesystem-mapping
fix(ci): the Windows install test never ran on a release (#461)
KageBinaryb66e8a7b906c
diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.ymlindex 23cea901..6c136819 100644--- a/.github/workflows/ci.yml+++ b/.github/workflows/ci.yml@@ -53,6 +53,50 @@ jobs:fiecho "checked $checked tracked sources"exit $status+ # ── Workflow wiring ─────────────────────────────────────────────────+ #+ # In this job, not in actions-lint.yml beside zizmor, and that placement+ # is the point. The `main` ruleset requires exactly one status check,+ # `CI Passed`, whose `needs` this job is in; a check in any other workflow+ # reports and gates nothing. actions-lint.yml also carries+ # `paths: ['.github/workflows/**']`, so it is absent from most PRs and+ # could not be made required without hanging them.+ #+ # What it buys: release.yml calls windows-install.yml as a reusable+ # workflow, and nothing runs that call until a tag is pushed, so a+ # mistyped `uses:` path or a renamed input would surface for the first+ # time on a production release — the same blind spot as the+ # `release: published` trigger it replaces, which silently did nothing for+ # a whole release. actionlint resolves local `uses:` targets and matches+ # `with:` keys against the callee's declared inputs. It cannot see a+ # reusable-workflow permission restriction; that is an org setting no+ # linter reaches.+ #+ # shellcheck and pyflakes are off deliberately: actionlint shells out to+ # them for `run:` blocks when they are on PATH, which they are on+ # ubuntu-latest, so leaving them on makes a required check's verdict+ # depend on the runner image's shellcheck version rather than on the repo.+ #+ # Pinned by version and checksum, like every action here is pinned by SHA,+ # because this runs a downloaded binary. Dependabot's github-actions+ # ecosystem updates `uses:` refs and not `run:` downloads, so the bump is+ # manual — same as the `pipx install zizmor` next door, which pins nothing+ # at all.+ - name: Workflows are wired up (actionlint)+ env:+ ACTIONLINT_VERSION: 1.7.7+ ACTIONLINT_SHA256: 023070a287cd8cccd71515fedc843f1985bf96c436b7effaecce67290e7e0757+ run: |+ set -euo pipefail+ TARBALL="actionlint_${ACTIONLINT_VERSION}_linux_amd64.tar.gz"+ # Retried: this repo has already had a red check that was a release+ # asset download failing at the install step, reported as a workflow+ # problem it was not.+ curl -fsSL --retry 3 --retry-delay 2 --retry-all-errors -o "$TARBALL" \+ "https://github.com/rhysd/actionlint/releases/download/v${ACTIONLINT_VERSION}/${TARBALL}"+ echo "${ACTIONLINT_SHA256} ${TARBALL}" | sha256sum --check --strict -+ tar -xzf "$TARBALL" actionlint+ ./actionlint -shellcheck= -pyflakes=- uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7.0.0with:node-version: "22"diff --git a/.github/workflows/release.yml b/.github/workflows/release.ymlindex b769f2d8..deda4e39 100644--- a/.github/workflows/release.yml+++ b/.github/workflows/release.yml@@ -404,6 +404,8 @@ jobs:needs: buildif: ${{ !inputs.dry_run && github.event_name != 'pull_request' }}runs-on: ubuntu-latest+ outputs:+ version: ${{ steps.version.outputs.version }}permissions:contents: write # create the release + push the brew branchpull-requests: write # open the brew formula PR@@ -581,3 +583,21 @@ jobs:git commit -m "chore(brew): update formula for v${VERSION}" || exit 0git push --force-with-lease origin "$BRANCH"gh pr create --title "chore(brew): update formula for v${VERSION}" --body "Auto-generated: updates Homebrew formula SHA for v${VERSION}" --base main || true++ # ── Verify the installer against the release it just made ────────────────+ #+ # The artifact was verified by packaging it; this verifies the script that+ # unpacks it, which is where every Windows install bug has actually lived.+ # Called rather than triggered by `release: published`: the release is created+ # with `secrets.GITHUB_TOKEN`, and GitHub does not start new workflow runs+ # from events created with that token, so the trigger never fired.+ #+ # Not a gate on the release — the assets are already published by the time it+ # runs — but a failure here lands on the run that produced them, which is the+ # difference between finding out now and finding out from a user.+ verify-install:+ needs: release+ if: ${{ !inputs.dry_run && github.event_name != 'pull_request' }}+ uses: ./.github/workflows/windows-install.yml+ with:+ version: ${{ needs.release.outputs.version }}diff --git a/.github/workflows/windows-install.yml b/.github/workflows/windows-install.ymlindex a3b06e66..7dadded9 100644--- a/.github/workflows/windows-install.yml+++ b/.github/workflows/windows-install.yml@@ -11,6 +11,15 @@# So this runs the real script, on a real Windows runner, and asserts the shape# of what it leaves behind. Every check below is a bug that actually shipped.#+# It is *called* by the release workflow rather than triggered by+# `release: published`, and that is not a style choice. The release is created+# by `github-actions[bot]` with `secrets.GITHUB_TOKEN`, and GitHub does not start+# new workflow runs from events created with that token — so a `release:`+# trigger here never fires. It was written that way first and silently did+# nothing for a release; being called keeps the check inside the run that+# produced the artifact, where a failure is attached to the release that caused+# it.+## `IX_SKIP_BACKEND=1` is what makes it possible at all: the runner has no Linux# Docker, and `Write-Err` exits, so without the skip the run dies at the Docker# check having never unpacked the CLI — the part that keeps breaking.@@ -23,10 +32,14 @@ on:paths:- "scripts/install/install.ps1"- ".github/workflows/windows-install.yml"- # A published release is the moment the installer's inputs change, and the- # only moment a real user is about to run it against them.- release:- types: [published]+ # Called by the release workflow once the assets are up — see below for why+ # this is a call rather than a `release: published` trigger.+ workflow_call:+ inputs:+ version:+ description: "Version to install, without the leading v."+ required: true+ type: stringworkflow_dispatch:inputs:version:@@ -61,10 +74,14 @@ jobs:shell: pwshenv:IX_SKIP_BACKEND: "1"- IX_VERSION: ${{ inputs.version || (github.event_name == 'release' && github.event.release.tag_name) || '' }}+ IX_VERSION: ${{ inputs.version || '' }}run: |$env:IX_HOME = Join-Path $env:RUNNER_TEMP 'ix home'- # The tag carries a leading v; IX_VERSION does not.+ # Tolerance for the hand-typed case, not a translation of the tag: the+ # release workflow passes `${TAG#v}`, already stripped, so this is a+ # no-op on that path. The one input a human fills in is+ # workflow_dispatch's, where the release is called v0.10.0 everywhere+ # it is named and typing the v back in is the obvious mistake.if ($env:IX_VERSION) { $env:IX_VERSION = $env:IX_VERSION -replace '^v', '' }Write-Host "IX_HOME=$env:IX_HOME IX_VERSION=$($env:IX_VERSION -replace '^$','(latest stable)')"& .\scripts\install\install.ps1@@ -94,6 +111,8 @@ jobs:- name: Launcher runsshell: pwsh+ env:+ IX_VERSION: ${{ inputs.version || '' }}run: |# By absolute path, not through PATH: a shim that resolves only because# the session happens to have the right PATH is not a working install.@@ -103,6 +122,26 @@ jobs:if ($LASTEXITCODE -ne 0) { throw "ix --version exited $LASTEXITCODE`n$out" }Write-Host $out.Trim()+ # Assert *which* release we just installed, when one was named.+ #+ # The IX_VERSION expression above falls back to an empty string, which+ # install.ps1 reads as "latest stable" -- and workflow_call accepts an+ # empty string for a required input. So if needs.release.outputs.version+ # ever resolves empty (a renamed step id, a reworked release job) this+ # job installs an unrelated build and every check below passes against+ # it, which is the same silent no-op the release-published trigger was.+ #+ # No expression syntax in this comment. zizmor reads a template+ # expansion anywhere inside a run block as an injection site, comments+ # included, and it is right that it cannot tell the difference.+ if ($env:IX_VERSION) {+ $want = $env:IX_VERSION -replace '^v', ''+ if ($out -notmatch [regex]::Escape($want)) {+ throw "installed $($out.Trim()) but the release under test is $want"+ }+ Write-Host "verified against release $want"+ }+- name: Compass is present and stampedshell: pwshrun: |@@ -126,8 +165,9 @@ jobs:shell: pwshenv:IX_SKIP_BACKEND: "1"- IX_VERSION: ${{ inputs.version || (github.event_name == 'release' && github.event.release.tag_name) || '' }}+ IX_VERSION: ${{ inputs.version || '' }}run: |+ # Same hand-typed tolerance as the first install — see the note there.if ($env:IX_VERSION) { $env:IX_VERSION = $env:IX_VERSION -replace '^v', '' }& .\scripts\install\install.ps1if ($LASTEXITCODE -ne 0) { throw "second install.ps1 exited $LASTEXITCODE" }diff --git a/.github/zizmor.yml b/.github/zizmor.ymlindex 533a2f91..c89ec2c9 100644--- a/.github/zizmor.yml+++ b/.github/zizmor.yml@@ -13,10 +13,3 @@ rules:# setup-node's npm cache for build speed; the cache key is lockfile-pinned.# Accepted for the release workflow only.- release.yml- # windows-install.yml runs on `release: published`, which zizmor treats as- # a privileged context, and the finding is raised against setup-node on- # that basis alone. This one passes no `cache:` input, so setup-node- # neither restores nor saves a cache — there is nothing to poison, and the- # job publishes no artifact. It installs a already-published release and- # asserts the layout. Reported by zizmor at low confidence.- - windows-install.yml