fix what the pre-launch audit found: the talker constraint, the conso… #37

Workflow file for this run

	name: CI

	# Four Dart packages live here — the SDK, the two companions and the example —
	# and each has its own pubspec. A change to the core breaks the companions
	# without touching a file in them, so all four are checked on every push.

	on:
	push:
	branches: [main]
	tags: ["v*"]
	pull_request:
	workflow_dispatch:

	# A second push supersedes the first. Nothing here deploys, so cancelling a run
	# in flight costs nothing.
	concurrency:
	group: ${{ github.workflow }}-${{ github.ref }}
	cancel-in-progress: true

	jobs:
	test:
	runs-on: ubuntu-latest
	strategy:
	# Every package is reported, rather than stopping at the first red one.
	fail-fast: false
	matrix:
	package:
	- .
	- packages/code_scout_dio
	- packages/code_scout_http
	- packages/code_scout_talker
	- example
	steps:
	- uses: actions/checkout@v5

	- uses: subosito/flutter-action@v2
	with:
	# Pinned, not `channel: stable`, and the pin is the point rather than
	# the version: 3.47 turned every package job red on a commit that
	# touched three markdown files, and a toolchain release should not be
	# able to do that. Same reasoning as pinning templ in the server repo.
	#
	# Bump it deliberately, after running the matrix locally on the new
	# version. `analysis_options.yaml` carries an `analyzer: exclude:`
	# block because 3.47's `pub get` writes one if it is missing, which
	# dirties the tree, which makes `pub publish --dry-run` exit 65 on a
	# warning. Deleting that block breaks the package jobs and nothing
	# else, so it will look unrelated.
	flutter-version: 3.47.0
	channel: stable
	cache: true

	# Every package, not just this one. `flutter analyze` at the root reaches
	# down into packages/, and without their own resolution it reports every
	# companion symbol as undefined — 39 errors about code that is fine.
	- name: Resolve every package
	run: \|
	flutter pub get
	(cd packages/code_scout_dio && flutter pub get)
	(cd packages/code_scout_http && flutter pub get)
	(cd packages/code_scout_talker && flutter pub get)
	(cd example && flutter pub get)

	# analyze before test: a type error reads better as an analyzer message
	# than as a compile failure inside a test runner.
	- run: flutter analyze
	working-directory: ${{ matrix.package }}

	- run: flutter test
	working-directory: ${{ matrix.package }}

	# What ends up in the archive is not what is in the repository, and the gap
	# is easy to miss: `example/` was gitignored for months, which meant pub
	# silently shipped a package with no example and nobody noticed until someone
	# looked at the pub.dev page.
	package:
	runs-on: ubuntu-latest
	strategy:
	fail-fast: false
	matrix:
	package:
	- .
	- packages/code_scout_dio
	- packages/code_scout_http
	- packages/code_scout_talker
	steps:
	- uses: actions/checkout@v5

	- uses: subosito/flutter-action@v2
	with:
	# Pinned, not `channel: stable`, and the pin is the point rather than
	# the version: 3.47 turned every package job red on a commit that
	# touched three markdown files, and a toolchain release should not be
	# able to do that. Same reasoning as pinning templ in the server repo.
	#
	# Bump it deliberately, after running the matrix locally on the new
	# version. `analysis_options.yaml` carries an `analyzer: exclude:`
	# block because 3.47's `pub get` writes one if it is missing, which
	# dirties the tree, which makes `pub publish --dry-run` exit 65 on a
	# warning. Deleting that block breaks the package jobs and nothing
	# else, so it will look unrelated.
	flutter-version: 3.47.0
	channel: stable
	cache: true

	- run: flutter pub get
	working-directory: ${{ matrix.package }}

	# pub refuses to package a dirty tree, and `pub get` itself rewrites the
	# example's generated_plugins.cmake — so anything generated has to be
	# untracked or this fails for a reason that has nothing to do with the
	# package.
	- run: flutter pub publish --dry-run
	working-directory: ${{ matrix.package }}

	# The SDK against a real dashboard.
	#
	# This is the one job worth the setup. Everything above proves each side
	# against its own idea of the contract; this proves the two still agree. It
	# needs the server repository, so it is also the only job that can break
	# because somebody else's repository changed — which is the point.
	sdk-e2e:
	runs-on: ubuntu-latest

	services:
	postgres:
	image: postgres:16-alpine
	env:
	POSTGRES_USER: code_scout
	POSTGRES_PASSWORD: code_scout
	POSTGRES_DB: code_scout_e2e
	ports:
	- 5432:5432
	options: >-
	--health-cmd "pg_isready -U code_scout"
	--health-interval 5s
	--health-timeout 5s
	--health-retries 10

	steps:
	- uses: actions/checkout@v5

	# Into a dot directory: pub and the Flutter tool walk the tree looking
	# for pubspecs, and ignore anything starting with a dot.
	- uses: actions/checkout@v5
	with:
	repository: getcodescout/code_scout
	path: .server

	- uses: actions/setup-go@v5
	with:
	go-version-file: .server/go.mod
	cache-dependency-path: .server/go.sum

	- uses: subosito/flutter-action@v2
	with:
	# Pinned, not `channel: stable`, and the pin is the point rather than
	# the version: 3.47 turned every package job red on a commit that
	# touched three markdown files, and a toolchain release should not be
	# able to do that. Same reasoning as pinning templ in the server repo.
	#
	# Bump it deliberately, after running the matrix locally on the new
	# version. `analysis_options.yaml` carries an `analyzer: exclude:`
	# block because 3.47's `pub get` writes one if it is missing, which
	# dirties the tree, which makes `pub publish --dry-run` exit 65 on a
	# warning. Deleting that block breaks the package jobs and nothing
	# else, so it will look unrelated.
	flutter-version: 3.47.0
	channel: stable
	cache: true

	# sqflite_common_ffi opens SQLite through dart:ffi, so the runner needs
	# the library itself. Without it every test that touches the database
	# fails at load with a missing symbol.
	- name: Install SQLite
	run: sudo apt-get update && sudo apt-get install -y libsqlite3-dev

	# No /etc/code-scout.conf here, which is fine: the loader treats the file
	# as optional and every setting has a CS_* environment override.
	- name: Start the dashboard
	working-directory: .server
	env:
	CS_DB_HOST: 127.0.0.1
	CS_DB_PORT: "5432"
	CS_DB_USER: code_scout
	CS_DB_PASSWORD: code_scout
	CS_DB_NAME: code_scout_e2e
	CS_PORT: "24284"
	run: \|
	go build -o /tmp/code_scout .
	# nohup and disown: the server has to outlive the step that starts it,
	# and a shell exiting can take an ordinary background child with it.
	nohup /tmp/code_scout > /tmp/code_scout.log 2>&1 &
	disown
	for i in $(seq 1 60); do
	curl -sf http://127.0.0.1:24284/healthz >/dev/null && exit 0
	sleep 0.5
	done
	echo "::error::The dashboard never came up."
	cat /tmp/code_scout.log
	exit 1

	- run: flutter pub get

	# Checked again here so a server that died between steps fails with that
	# sentence, rather than as a confusing "sign in failed" inside the test.
	- name: Check the dashboard is still up
	run: \|
	curl -sf http://127.0.0.1:24284/healthz \|\| {
	echo "::error::The dashboard stopped before the tests ran."
	cat /tmp/code_scout.log
	exit 1
	}

	- name: Run the SDK against it
	env:
	CS_E2E_BASE: http://127.0.0.1:24284
	run: flutter test test/e2e/

	# The server's own log is the only place an ingest rejection explains
	# itself, and a failed run is exactly when you want it.
	# Session upserts are deliberately not fatal to an upload — the logs are
	# the payload and the session is context — so a failure there is only ever
	# a line in this log. Promoted to an annotation so it shows in the run
	# summary rather than needing the step expanded.
	- name: Dashboard log
	if: failure()
	run: \|
	grep -iE "session\|error" /tmp/code_scout.log \| tail -20 \|
	while IFS= read -r line; do echo "::error::$line"; done
	cat /tmp/code_scout.log

Navigation Menu

fix what the pre-launch audit found: the talker constraint, the conso… #37

Workflow file

fix what the pre-launch audit found: the talker constraint, the conso… #37

Uh oh!

Workflow file for this run