ci: scope pull request checks to what the diff can affect

.github/scripts/ci-changes.sh

@@ -0,0 +1,64 @@
+#!/usr/bin/env bash
+# Prints which checks a pull request's changes can affect, as `<flag>=true|false` lines for
+# $GITHUB_OUTPUT. A flag is `false` only when every changed file is on an allowlist whose jobs
+# fully cover it. A file on no allowlist sets every flag but `desktop_live` (which keeps its own
+# skip list) to `true`; a change under .github/, an empty diff, or any failure to work out the
+# diff sets all of them.
+#
+#   full          integration and e2e: any file not allowlisted below
+#   test          the Vitest shards: content, docs app, and desktop changes, plus `full`
+#   build         the app's `next build`: content changes, plus `full`
+#   desktop_live  the live Electron suite: any file outside its own skip list
+#
+# lint always runs and is not gated.
+#
+# Usage: ci-changes.sh <base-sha>
+set -u
+
+# Posts, author profiles, and the public assets only posts reference. The test shards read them
+# (lib/content/og-image.test.ts) and `next build` prerenders every post with its MDX components,
+# which check:library-content compiles but never renders. Components under content/ are code.
+content='^apps/sim/content/([^/]+/)*[^/]+\.mdx?$|^apps/sim/content/authors/[^/]+\.json$|^apps/sim/public/(authors|blog|changelog|library)/'
+# Its own workspace tests, its type-check and docs-manifest audit in lint, and apps/sim tests
+# that read its pages (tools/netsuite/netsuite.test.ts).
+docs='^apps/docs/'
+# Nothing outside the Electron app imports it; the desktop bridge and protocol packages are
+# consumed by apps/sim and stay unlisted.
+desktop='^apps/desktop/'
+# Agent guidance, read only by the lint job's audits.
+guidance='^(\.agents|\.claude|\.cursor)/|^([^/]+/)*(AGENTS|CLAUDE)\.md$|^README\.md$'
+# The live desktop suite drives apps/desktop against the app, so it skips more than the rest.
+desktop_live_skip="$guidance"'|^(apps/docs/|apps/pii/|apps/sim/content/|apps/sim/public/(authors|blog|changelog|library)/|packages/(python-sdk|ts-sdk)/)|\.mdx?$|(^|/)LICENSE$'
+
+emit() {
+  printf 'full=%s\ntest=%s\nbuild=%s\ndesktop_live=%s\n' "$1" "$2" "$3" "$4"
+}
+run_all() {
+  emit true true true true
+  echo "Running every check: $1" >&2
+  exit 0
+}
+matches() { printf '%s\n' "$names" | grep -qE "$1"; }
+outside() { printf '%s\n' "$names" | grep -qvE "$1"; }
+
+base=${1:-}
+[ -n "$base" ] || run_all 'no base commit'
+git fetch --quiet --depth=1 origin "$base" || run_all "could not fetch $base"
+names=$(git diff --no-renames --name-only "$base" HEAD) || run_all "could not diff against $base"
+[ -n "$names" ] || run_all 'no changed files listed'
+if matches '^\.github/'; then run_all 'a CI file changed'; fi
+
+desktop_live=false
+if outside "$desktop_live_skip"; then desktop_live=true; fi
+if outside "$content|$docs|$desktop|$guidance"; then
+  emit true true true "$desktop_live"
+  unlisted=$(printf '%s\n' "$names" | grep -vE "$content|$docs|$desktop|$guidance" | head -n 1)
+  echo "Running every check (desktop_live=$desktop_live): $unlisted is not on an allowlist" >&2
+  exit 0
+fi
+
+test=false build=false
+if matches "$content|$docs|$desktop"; then test=true; fi
+if matches "$content"; then build=true; fi
+emit false "$test" "$build" "$desktop_live"
+echo "Scoped checks: test=$test build=$build desktop_live=$desktop_live" >&2

.github/scripts/ci-gate.sh

@@ -0,0 +1,39 @@
+#!/usr/bin/env bash
+# The `ci` job's verdict: the single required check, so it must never pass a run whose checks did
+# not all pass. Reads the job's `needs` context (`toJSON(needs)`) from $NEEDS and exits non-zero,
+# naming each offending job, unless every job succeeded or was skipped as `changes` intended.
+#
+# A skip passes only for a gated job whose `changes` flag is exactly `false`:
+#
+#   integration, e2e  full
+#   test              test
+#   build             build
+#   desktop-live      desktop_live
+#
+# Every other skip fails: an ungated job (lint, `changes` itself), a gated job whose flag is
+# `true`, missing, or anything else (a failed or skipped `changes` leaves no outputs). A failure
+# or a cancellation always fails.
+#
+# Usage: NEEDS='<toJSON(needs)>' ci-gate.sh
+set -euo pipefail
+
+needs=${NEEDS:-}
+if ! jq -e 'type == "object"' <<< "$needs" > /dev/null 2>&1; then
+  echo "::error::The needs context is not a JSON object: '$needs'"
+  exit 1
+fi
+
+failed="$(jq -r '
+  {"integration": "full", "e2e": "full", "test": "test", "build": "build", "desktop-live": "desktop_live"} as $gate
+  | (.changes.outputs // {}) as $flags
+  | to_entries[]
+  | select(.value.result != "success")
+  | select(.value.result != "skipped" or $gate[.key] == null or $flags[$gate[.key]] != "false")
+  | "\(.key): \(.value.result)"
+' <<< "$needs")"
+if [ -n "$failed" ]; then
+  echo "::error::Checks did not pass:"
+  echo "$failed"
+  exit 1
+fi
+echo "All checks passed."

.github/workflows/checks.yml

@@ -8,6 +8,37 @@ permissions:
   contents: read
 
 jobs:
+  # Which checks a pull request's changes can affect (.github/scripts/ci-changes.sh). A job is
+  # skipped only when the script prints `false` for its flag; an unset flag, and every push or
+  # dispatch (the deploy chain's checks), runs it. Every gated job needs this one, so its failure
+  # skips them and fails `ci` below.
+  changes:
+    name: changes
+    runs-on: &runner-2vcpu ${{ (vars.CI_PROVIDER == '' || vars.CI_PROVIDER == 'blacksmith') && 'blacksmith-2vcpu-ubuntu-2404' || 'ubuntu-latest' }}
+    timeout-minutes: 5
+    outputs:
+      full: ${{ github.event_name != 'pull_request' || steps.classify.outputs.full != 'false' }}
+      test: ${{ github.event_name != 'pull_request' || steps.classify.outputs.test != 'false' }}
+      build: ${{ github.event_name != 'pull_request' || steps.classify.outputs.build != 'false' }}
+      desktop_live: ${{ github.event_name != 'pull_request' || steps.classify.outputs.desktop_live != 'false' }}
+    steps:
+      - name: Checkout code (mirror)
+        if: github.event_name == 'pull_request' && !github.event.pull_request.head.repo.fork
+        uses: useblacksmith/checkout@25227e61ff9dafe400e22fa487b673eac4e4409a # v1.8.1
+        with:
+          fetch-depth: 2
+      - name: Checkout code
+        if: github.event_name == 'pull_request' && github.event.pull_request.head.repo.fork
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v6
+        with:
+          fetch-depth: 2
+      - name: Classify the pull request's changes
+        id: classify
+        if: github.event_name == 'pull_request'
+        env:
+          BASE: ${{ github.event.pull_request.base.sha }}
+        run: bash .github/scripts/ci-changes.sh "$BASE" >> "$GITHUB_OUTPUT"
+
   integration:
     # Runs the real-infrastructure test layer: every `*.integration.ts` in packages/db and
     # apps/sim, discovered by glob (`vitest run --mode integration`), against the database each
@@ -19,6 +50,8 @@ jobs:
     # shard runs its files one at a time against its own database. Files run serially, so a shard
     # barely uses more than one core: 4 vCPU is enough.
     name: integration (${{ matrix.provision }}, ${{ matrix.shard }}/4)
+    needs: changes
+    if: needs.changes.outputs.full == 'true'
     runs-on: &runner-4vcpu ${{ (vars.CI_PROVIDER == '' || vars.CI_PROVIDER == 'blacksmith') && 'blacksmith-4vcpu-ubuntu-2404' || 'ubuntu-latest' }}
     timeout-minutes: 20
     strategy:
@@ -143,6 +176,8 @@ jobs:
   # self-hosted suites fit on 4.
   e2e:
     name: e2e (${{ matrix.group }})
+    needs: changes
+    if: needs.changes.outputs.full == 'true'
     runs-on: ${{ (vars.CI_PROVIDER == '' || vars.CI_PROVIDER == 'blacksmith') && matrix.runner || 'ubuntu-latest' }}
     timeout-minutes: ${{ matrix.group == 'mobile' && 40 || 20 }}
     strategy:
@@ -217,39 +252,12 @@ jobs:
           if-no-files-found: ignore
           retention-days: 7
 
-  # Pull requests skip the live desktop suite only when every change is clearly unrelated to the
-  # app it drives (docs, other apps, published content). Anything else, and any failure to work
-  # out the diff, runs it: a pull request that skipped it wrongly would first fail on staging.
-  desktop-changes:
-    name: desktop-changes
-    runs-on: &runner-2vcpu ${{ (vars.CI_PROVIDER == '' || vars.CI_PROVIDER == 'blacksmith') && 'blacksmith-2vcpu-ubuntu-2404' || 'ubuntu-latest' }}
-    timeout-minutes: 5
-    outputs:
-      changed: ${{ github.event_name != 'pull_request' || steps.diff.outputs.changed != 'false' }}
-    steps:
-      - name: Checkout code (mirror)
-        if: github.event_name == 'pull_request' && !github.event.pull_request.head.repo.fork
-        uses: useblacksmith/checkout@25227e61ff9dafe400e22fa487b673eac4e4409a # v1.8.1
-        with:
-          fetch-depth: 2
-      - name: Checkout code
-        if: github.event_name == 'pull_request' && github.event.pull_request.head.repo.fork
-        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v6
-        with:
-          fetch-depth: 2
-      - name: Diff against the pull request's base
-        id: diff
-        if: github.event_name == 'pull_request'
-        env:
-          BASE: ${{ github.event.pull_request.base.sha }}
-        run: bash .github/scripts/desktop-live-changes.sh "$BASE" >> "$GITHUB_OUTPUT"
-
   # Desktop tools in the real Electron app against a local app, on its own runner: the
   # Electron app, the dev app and its realtime server together outgrow the e2e runner.
   desktop-live:
     name: desktop-live
-    needs: desktop-changes
-    if: needs.desktop-changes.outputs.changed == 'true'
+    needs: changes
+    if: needs.changes.outputs.desktop_live == 'true'
     runs-on: ${{ (vars.CI_PROVIDER == '' || vars.CI_PROVIDER == 'blacksmith') && 'blacksmith-16vcpu-ubuntu-2404' || 'ubuntu-latest' }}
     timeout-minutes: 30
     services:
@@ -592,6 +600,8 @@ jobs:
   # task hash, so a shard only ever replays its own result.
   test:
     name: test (${{ matrix.shard }}/2)
+    needs: changes
+    if: needs.changes.outputs.test == 'true'
     runs-on: ${{ (vars.CI_PROVIDER == '' || vars.CI_PROVIDER == 'blacksmith') && 'blacksmith-8vcpu-ubuntu-2404' || 'ubuntu-latest' }}
     timeout-minutes: 15
     strategy:
@@ -660,6 +670,8 @@ jobs:
   # not the native Turbopack workers that dominate, so it cannot prevent this.
   build:
     name: build
+    needs: changes
+    if: needs.changes.outputs.build == 'true'
     runs-on: ${{ (vars.CI_PROVIDER == '' || vars.CI_PROVIDER == 'blacksmith') && 'blacksmith-16vcpu-ubuntu-2404' || 'linux-x64-8-core' }}
     # Build durations crossed 15 minutes as the app grew (10m02 on Jul 29 AM,
     # 14m44 after the folders/desktop/library merges, then two straight
@@ -713,25 +725,22 @@ jobs:
         run: bunx turbo run build --filter=@sim/app
 
   # One status for every check above: the single check to require on a branch ruleset, so adding,
-  # sharding or renaming a job never means editing the ruleset. Skipped is a pass (the desktop live
-  # suite skips on pull requests that cannot affect it); a failure or a cancellation is not.
+  # sharding or renaming a job never means editing the ruleset. .github/scripts/ci-gate.sh decides:
+  # a skip passes only for a gated job whose `changes` flag is exactly `false`; any other skip (a
+  # failed or skipped `changes`, a lost output, an ungated job) fails, as does a failure or a
+  # cancellation. A failed checkout fails the gate too.
   # `always()`, not `!cancelled()`: a skipped job satisfies a required check, so a gate that skips
   # on a cancelled run would report a cancelled head commit as passing.
   ci:
     name: ci
-    needs: [integration, e2e, desktop-changes, desktop-live, lint, test, build]
+    needs: [changes, integration, e2e, desktop-live, lint, test, build]
     if: ${{ always() }}
     runs-on: *runner-2vcpu
     timeout-minutes: 5
     steps:
+      - *checkout-mirror
+      - *checkout-plain
       - name: Require every check to pass
         env:
-          RESULTS: ${{ toJSON(needs) }}
-        run: |
-          failed="$(jq -r 'to_entries[] | select(.value.result != "success" and .value.result != "skipped") | "\(.key): \(.value.result)"' <<< "$RESULTS")"
-          if [ -n "$failed" ]; then
-            echo "::error::Checks did not pass:"
-            echo "$failed"
-            exit 1
-          fi
-          echo "All checks passed."
+          NEEDS: ${{ toJSON(needs) }}
+        run: bash .github/scripts/ci-gate.sh

.github/workflows/ci.yml

@@ -18,9 +18,8 @@ name: ci
 on:
   push:
     branches: [main, staging, dev]
-  # No path filter: `checks / ci` is the required status check, and a pull request whose run
-  # never starts never reports it, so it could not merge. Docs-only pull requests are rare enough
-  # that running CI on them costs nothing measurable.
+  # No path filter: a pull request whose run never starts never reports `checks / ci`. The
+  # `changes` job in checks.yml scopes the jobs to what the diff can affect instead.
   pull_request:
     branches: [main, staging, dev]