README

Opinionated personal Laravel preset. One command scaffolds the dev tooling, code conventions and CI used across new projects (on top of the Laravel + Vue starter kit, so anything the starter kit already ships is not duplicated).

Install

composer require paolobellini/laravel-preset --dev
php artisan preset:install

The packages and scripts groups add their dependencies with composer require (prefixed with ./vendor/bin/sail when Laravel Sail is installed and the command runs on the host — inside the container, as with sail artisan preset:install, plain composer is used). No version constraint is ever passed, so composer resolves the newest stable release compatible with the project — the preset never carries a stale version around. Pass --no-install to write the resolved constraints into composer.json without installing.

What it does

php artisan preset:install walks through the choices one step at a time and installs nothing until the plan is confirmed:

Step	Question	Groups
1	Where the project's commands run (only on a host with Sail)	—
2	Which optional runtime packages	`packages` — always installed
3	Development tooling, yes or no, as one block	`configs`, `scripts`
4	Which AI tooling	`ai`, `skills`, `codegraph`
5	Which automation	`github`, `lefthook`
6	The plan, then Install the preset as planned?	—

Passing a group flag (see Flags) skips the steps and installs exactly the flagged groups.

The confirmed choices are saved to preset.json at the project root — commit it. The next run finds it and asks Reuse the choices saved in preset.json? instead of walking through the steps again; a non-interactive run follows it without asking. Flagged runs neither read nor write it.

{
    "runtime": "sail",
    "packages": ["data", "query-builder"],
    "groups": ["packages", "configs", "scripts", "ai", "github"]
}

`packages` — runtime packages

nunomaduro/essentials and thecodingmachine/safe always go into require. Three more are optional and asked for one by one:

Package	Offered
`spatie/laravel-data`	always
`spatie/laravel-query-builder`	always
`spatie/laravel-typescript-transformer`	only when the project requires `inertiajs/inertia-laravel`

A non-interactive run takes every package offered. All of them go into require — the application loads them at runtime, so they must survive composer install --no-dev. One of them found in require-dev is moved to require.

Also copies config/essentials.php — the nunomaduro/essentials overrides (Unguard => true, inverse of the package default).

`configs` — lint / format / static analysis

Copies the configs not already in the starter kit:

File	Tool
`pint.json`	Laravel Pint (strict types, final classes, phpdoc-only types)
`phpstan.neon`	Larastan level 7 + the pest-plugin-phpstan and phpstan-safe-rule extensions
`rector.php`	Rector + rector-laravel sets, scoped to `app/` and `database/`
`rector-tests.php`	Rector for `tests/` — `LARAVEL_TESTING` + `PestSetList::CODING_STYLE`
`psalm.xml`	Psalm, scoped to taint analysis only (`errorLevel="8"`)
`tests/Pest.php`	created when missing (`extend(TestCase)` + `RefreshDatabase`), otherwise patched with `pest()->tia()->locally()`

`ai` — conventions

Copies the .ai/ directory and adds laravel/boost to require-dev:

.ai/guidelines/personal/* — precedence, comments, commits, controllers (action pattern), actions, caching, enums, exceptions, form-requests, frontend, models, pest-agent, php (Safe functions), policies, query-builder, resources, testing, traits, translations, typescript, workflow.
.ai/mcp/mcp.json.

query-builder.md and typescript.md are copied only when their package is part of the project — picked in the same run, or already in require. The guidelines that merely mention an optional package (controllers.md, frontend.md) say what to do without it.

`scripts` — composer quality scripts + dev deps

Merges the scripts into composer.json without clobbering existing keys, then installs the dependencies with composer require — always unconstrained, so every install picks up the current stable release.

Added to require-dev (anything already required is left untouched, use --force to re-require it at the latest version): fruitcake/laravel-debugbar, larastan/larastan, laravel/pint, laravel/pail, rector/rector, driftingly/rector-laravel, pestphp/pest and the pest-plugin-{type-coverage,mutate,rector,phpstan,evals,agent,faker} plugins, thecodingmachine/phpstan-safe-rule, vimeo/psalm. nunomaduro/collision and pestphp/pest-plugin-laravel are not added — they already ship with the starter kit.

config.allow-plugins gets pestphp/pest-plugin so the pest plugins can boot.

Composer scripts are layered so a tool is declared once:

leaf — one tool each: pint, pint:dry, rector, rector:dry, rector:test, rector:test:dry, stan, taint, test, test:type-coverage, test:coverage, test:mutate, update-shards.
groups — analyse:static (pint + stan + rector on app/), analyse (analyse:static + taint), tests (type-coverage + coverage), ci:node (npm lint/format/types).
entry points — ci (ci:php + ci:node), ci:php (analyse + tests), pre-push (rector:test:dry + test:mutate), pre-commit (analyse:static + test + ci:node, the by-hand equivalent of the hook).

Entry points compose groups and never re-list a leaf, so adding a tool means editing one group.

CI and pre-push are deliberately disjoint. CI owns the application contract — pint, stan, taint, rector over app/, coverage, type-coverage — so it is asynchronous and cannot be bypassed. pre-push owns test-suite quality: the mutation score and rector over tests/, the two slowest checks and the two whose audience is the author. Nothing runs twice.

Every script that boots the application (test:*) is prefixed with @php artisan config:clear --ansi @no_additional_args — a cached bootstrap/cache/config.php silently overrides config/ and .env, and @no_additional_args stops Composer from appending your own options to config:clear. The lint/analysis scripts do not boot the app and skip it.

composer taint → psalm --taint-analysis --no-cache. Psalm is installed only for this: PHPStan/Larastan stays the static analyser, and psalm.xml runs at the most permissive error level so the report is about tainted input reaching a sink, not about typing.
composer test:mutate → pest --parallel --mutate --covered-only --min=65, with Composer\Config::disableProcessTimeout so a long mutation run is not killed at Composer's 300s limit.

npm deps and scripts are not touched — the starter kit already provides ESLint, Prettier, TypeScript and their lint/format/types:check scripts.

`lefthook` — pre-commit hooks (opt-in)

Copies lefthook.yml, with two hooks and one job per tool — granular on purpose, since lefthook only parallelises what it sees as separate jobs and a failure in one still reports the others.

pre-commit is scoped to {staged_files} with a glob per job, so a commit touching one .vue never starts a PHP job. Pint, Prettier and ESLint fix and re-stage (stage_fixed) instead of failing, so a commit is never blocked by formatting; PHPStan analyses the staged files and Pest runs with the coverage gate (cheap locally, since TIA replays everything untouched).
pre-push runs only what CI does not: rector:test:dry and test:mutate.

tests/Pest.php is the one file the preset patches rather than copies: it holds project-specific setup, so an existing one only gets pest()->tia()->locally() appended (and nothing at all if it already configures TIA). That line is what makes the commit hook affordable — locally a run replays whatever the change did not touch, while CI still measures the full suite.

Passing paths to PHPStan and Rector narrows them to those files — an error caused elsewhere will not show up until CI. That is the trade that keeps the commit hook fast.

Paths given on the command line also replace the paths of phpstan.neon, so the stan job's glob and exclude repeat them: a staged rector.php or migration is never handed to PHPStan. Change one, change the other.

Every job — the Node ones included (vendor/bin/sail npx prettier, … eslint) — is prefixed with vendor/bin/sail when Sail is detected in the project, so a hook runs the same PHP and Node as sail composer ci, never whatever the host happens to have. Without Sail the prefix is stripped from the copied file.

Opt-in — unlike the other groups it is never selected by default: pass --lefthook, or tick it in the interactive prompt.

Lefthook is not a composer package: install the binary once per machine (brew install lefthook, or npm i -D lefthook), then run lefthook install in the project to wire up .git/hooks.

Installing the group also clears the ground for a single agent: the scaffolding boost:install wrote for the others (.amp, .codex, .cursor, .factory, .gemini, .grok, .junie, .kiro, .pi, .zed, and .github/copilot-instructions.md) is deleted, and boost.json is pinned to agents: ["claude_code"].

The pin is the part that makes it stick. Boost only falls back to detecting agents when boost.json names none, and its detection is generous — Junie counts itself present because .idea exists, or because PhpStorm is installed on the machine at all. Delete the directories without pinning and the next boost:install writes them back.

.idea and .vscode are never touched: Boost writes into them, but they hold your editor's own settings.

`github` — CI workflows and dependency updates

First removes the starter-kit lint.yml + tests.yml (superseded), then copies the workflows. Each job runs the composer scripts from the scripts group directly, so the commands and their definitions stay in this repo and cannot drift apart. Only the environment setup is shared, via the composite actions in paolobellini/bellini.one.

.github/workflows/analyse.yml — on push to main / any PR, runs composer analyse (pint, phpstan, rector, psalm taint) and composer ci:node. The Node step carries if: ${{ !cancelled() }} so a PHP failure still reports the front-end result in the same run; the job fails either way. It is also gated on dorny/paths-filter, which skips it when no front-end file changed — a Composer-only bump PR has nothing for eslint or tsc to say, and neither does a documentation-only one. The filter is written as "everything except the PHP side, the docs and the agent conventions" (** plus negations, with predicate-quantifier: some-with-excludes), so an unrecognised new file runs the checks rather than silently skipping them. Note that !**/*.md also takes Markdown out of npm run format:check; Prettier still formats it at pre-commit, where stage_fixed writes the result back before the file is ever committed.
.github/workflows/tests.yml — on push to main / any PR, runs composer tests (type coverage + coverage) against a database service container. The engine comes from the DB_CONNECTION, DB_IMAGE, DB_PORT and DB_OPTIONS repository variables, defaulting to MySQL 8.0 when they are unset (see below).
.github/workflows/security.yml — on PR targeting main, staging or dev, a Trivy filesystem scan (vuln + secret) failing on HIGH,CRITICAL, plus a second report-only pass that lists the vulnerabilities with no released fix instead of hiding them behind ignore-unfixed.
.github/trivy.yaml — points Trivy at the VEX document.
.vex/openvex.json — the triage record (see below). Created empty, with its @id derived from the project (urn:vex:vendor:project, from composer.json) so two projects never share one, and never overwritten without --force.
.github/dependabot.yml — weekly Composer, npm and GitHub Actions updates. Minor and patch bumps are grouped into one PR per ecosystem and per production/development split, so the usual flood becomes a handful of PRs; major bumps stay on their own, because those are the ones worth reading. Commit prefixes follow the repo convention (chore(deps), chore(deps-dev), chore(ci)). A cooldown holds each release back before it is proposed — 3 days for a patch, 7 for a minor, 14 for a major — so a version pulled hours after publication never reaches a PR; security updates ignore the cooldown entirely. An existing dependabot.yml is never overwritten without --force.

Triaging a vulnerability with OpenVEX

A scanner has no memory: without one, a vulnerability you assessed months ago blocks every pull request forever. The record of that assessment is a VEX document, and .vex/openvex.json is it.

Most findings never reach it — a CVE with a released fix is answered by bumping the dependency, and the finding disappears on its own. A statement is for the other case: no fix exists, or one exists and cannot be applied, and you have established that the project is not exposed.

{
  "vulnerability": { "name": "GHSA-94pj-82f3-465w" },
  "products": [ { "@id": "pkg:composer/guzzlehttp/guzzle@7.9.2" } ],
  "status": "not_affected",
  "justification": "vulnerable_code_not_in_execute_path"
}

The spec requires version to be incremented whenever the document's content changes, so bump it in the same pull request as the statement. Trivy does not check it; a VEX consumer that merges documents from several sources does.

Trivy matches on the package URL, drops the finding when the status is not_affected or fixed, and leaves it in place for affected and under_investigation — so a half-finished assessment keeps the build red. Where several statements match, the last one wins: change your mind by appending, not by editing, and the reasoning stays in the history.

Including the version pins the statement to it, and a bump brings the finding back for a fresh assessment. Dropping the version (pkg:composer/vendor/name) makes it apply to every version, forever — defensible when the justification is about your own code rather than the dependency, but it never expires, so prefer the pinned form.

OpenVEX rather than the CycloneDX VEX that Dependency-Track exports, for two reasons found in the source rather than the documentation: Trivy supports CycloneDX VEX only when scanning an SBOM, not a filesystem; and Dependency-Track's exporter omits the component list for that variant, leaving every affects[].ref pointing at the application itself rather than at the vulnerable package. OpenVEX identifies products by package URL, so it needs no companion document.

Keep each statement in its own pull request. The GitHub pull_request event builds a merge ref, so as soon as a statement lands on the base branch every open pull request against it sees it — one small merge unblocks the rest, and a security decision gets reviewed on its own rather than buried in a feature branch. A CODEOWNERS entry for /.vex/ is worth adding: it is the one place in the repository where a line makes a security alert disappear.

Dependency-Track, if you run it, keeps its own role: it holds the portfolio view across projects and sends the NEW_VULNERABILITY notification. It is not in the critical path of CI, and its triage is not what the gate reads.

Renovate instead of Dependabot

--renovate writes .github/renovate.json and deletes .github/dependabot.yml, so only one bot ever watches the repo. The config mirrors the Dependabot one — weekly Monday window, minor/patch grouped per manager and per production/ development split, majors on their own, minimumReleaseAge of 3/7/14 days by update type — and adds two things Dependabot has no equivalent for:

lockFileMaintenance refreshes the whole lockfile once a month, picking up transitive dependencies that no direct constraint mentions.
dependencyDashboard opens a single issue listing everything pending, so updates can be reviewed without a PR being opened for each.

vulnerabilityAlerts sets minimumReleaseAge back to null and schedules at any time: the soak period must not delay a security fix.

Read schedule as a permission window, not a trigger: Renovate runs on its own cadence and skips the repository when it falls outside. * 0-6 * * 1 is seven hours on Monday; * * 1-7 * 1 restricts both day of month and weekday, which Renovate reads as an AND — the first Monday of the month, with a full day to land in.

Unlike Dependabot, the file alone does nothing — Renovate is not built into GitHub. Either install the hosted Renovate GitHub App on the repository, or add --renovate-selfhosted.

`--renovate-selfhosted`

Installs .github/workflows/renovate.yml, which runs Renovate from the official action. Use it when the GitHub App cannot be installed — no admin rights on the organisation, or a policy against third-party apps with write access. It implies --renovate, so the config file is installed too.

The workflow needs a RENOVATE_TOKEN secret. Do not point it at the default GITHUB_TOKEN: pull requests opened with it do not trigger workflows, so analyse and tests would never run on an update PR and you would be merging unverified bumps. Use a personal access token, or a GitHub App token, with contents and pull-request write access.

The cron covers exactly the window the config allows, in both DST regimes — Monday 00:00–06:00 in Europe/Rome is Sunday 22:00–Monday 05:00 UTC in summer and Sunday 23:00–Monday 06:00 UTC in winter:

  schedule:
    - cron: '0 22,23 * * 0'
    - cron: '0 0-6 * * 1'

Nine runs a week rather than the fifty-six an every-three-hours cron would cost. Outside the window Renovate still boots, clones and analyses before deciding to do nothing, so a wide cron burns Actions minutes to no effect.

These two schedules have to be moved together. Widen the window in renovate.json without widening the cron and Renovate simply never runs, with nothing in the logs to say why. If that bothers you, drop schedule and lockFileMaintenance.schedule from the config and drive the timing from the cron alone — one scheduler, at the cost of losing the local-time window.

Once several projects share this setup, move the body of the file into a dedicated repository and reduce each project to {"extends": ["github>paolobellini/renovate-config"]} — the policy then lives in one place and no longer has to be re-copied by preset:install. This is safe to centralise precisely because the file describes policy, not commands: unlike the reusable CI workflows, nothing in it can fall out of step with a project's composer scripts.

Sharding a large suite

--sharded swaps tests.yml for a variant that splits the suite across a matrix of jobs instead of running it in one. It replaces the file rather than adding a second one, so the suite is never run twice.

type-coverage stays a single job — --type-coverage is static analysis, there is nothing to split and it needs no database.
tests fans out over four shards, each job running composer test -- --shard=<n>/<total>. The total comes from strategy.job-total, so widening the matrix line is the only edit needed — the denominator follows on its own.

Pest balances the shards from tests/.pest/shards.json, a file of recorded per-class timings. Without it, it chunks the test classes evenly and one shard ends up dominating the wall clock; Pest warns in the run output when the file is missing entries.

--sharded therefore also installs .github/workflows/update-shards.yml, which regenerates it every Monday (and on demand via Run workflow): it runs composer update-shards against the same database service and commits the file back if it changed. Seed it once locally with composer update-shards so the first sharded runs are balanced too.

The job needs contents: write and pushes to the default branch, so a branch protection rule that admits no exception will reject it — either allow the github-actions[bot] actor or drop the workflow and refresh the file by hand. Note also that --update-shards writes nothing when the suite fails, and that GitHub disables scheduled workflows after 60 days without repository activity.

The sharded variant drops the --coverage --min=90 gate: each job only sees its own slice, so a per-shard minimum is meaningless and the reports would have to be merged to be comparable. Shard only when the suite is big enough that this is worth losing.

Branch protection needs updating too — the required check stops being tests and becomes one entry per shard (tests (1), tests (2), …).

DB_OPTIONS cannot be shared: mysqladmin is absent from the MariaDB image and Postgres has neither, so the health-check is per engine. Set all four together — a DB_IMAGE on its own leaves the workflow connecting with the wrong driver and port.

`skills` — agent skills

Opt-in, like lefthook: offered in the prompt but never pre-selected, since it needs npx and the network. It runs the skills CLI over two sources:

jpcaparas/superpowers-laravel — laravel:queues-and-horizon, laravel:http-client-resilience, laravel:performance-select-columns.
mattpocock/skills — wait-what and teach, both user-invoked only (disable-model-invocation), so they never fire unless you type them.

They were picked to fill the gaps the .ai guidelines leave, not to restate them. Skills covering form requests, policies, caching, controllers, resources or testing are deliberately excluded: those are model-invoked, so they would fire exactly while the agent writes the code your own conventions govern, and offer a second opinion on top of them.

Two details of the CLI are baked into the command because both fail quietly:

each skill is named with its own --skill flag. Written as --skill=name the filter is ignored and every skill in the repository gets installed.
the agents are named explicitly (universal, claude-code). Left out, the CLI writes into every agent directory it detects — including the ones the ai group has just removed.

Skills land in .agents/skills/ as real files, with symlinks from .claude/skills/. Commit .agents/skills/ and skills-lock.json; the lock file records a hash per skill, so npx skills update can tell you when a source has changed upstream.

`codegraph` — the project's code graph

Opt-in. Two commands, neither asking anything:

codegraph install --yes --target claude --location local
codegraph init --yes

The first wires CodeGraph into the project, for Claude Code only: the MCP server in .mcp.json, the tool permission and prompt hook in .claude/settings.json, and a marker-fenced section in .claude/CLAUDE.md. Nothing is written to your global config, and committing those files gives the whole team the same setup. The second creates .codegraph/ and builds the graph, giving the agent a call graph to query instead of grepping for call sites.

Only the binary is installed once per machine. When it is missing, the group asks whether to run npm install --global @colbymchenry/codegraph — declining leaves the machine untouched, and non-interactive runs only print the command, never run it. The upstream npx @colbymchenry/codegraph wizard is not used: it asks questions the installer has no terminal to answer.

CodeGraph belongs to the host. Under sail artisan preset:install the group only says so and stops: run it with plain php artisan instead.

An existing .codegraph/ is not rebuilt unless --force is passed, since rebuilding an index is not free; the wiring is still refreshed.

Nothing needs adding to .gitignore: CodeGraph writes one inside .codegraph/ that ignores everything but itself, because the index is local to each machine.

Flags

php artisan preset:install --configs --ai --scripts --github   # pick groups, skips the steps
php artisan preset:install --packages                          # runtime packages only
php artisan preset:install --skills                            # agent skills only
php artisan preset:install --codegraph                         # build the code graph
php artisan preset:install --runtime=local                     # sail | local, skips the question
php artisan preset:install --force                             # overwrite existing files / deps
php artisan preset:install --no-install                        # skip the auto composer update

Without flags in a non-interactive shell, preset.json is followed when it exists; otherwise the default groups install: packages, configs, scripts, ai and github.

Runtime — where the project's commands run

The first thing the installer settles, because the rest depends on it: how composer require is launched and whether the lefthook jobs carry the vendor/bin/sail prefix.

Situation	Runtime	Asked?
No Sail (`vendor/bin/sail` or a compose file missing)	local	no
Sail configured, installer run on the host	Sail by default	yes
Installer run inside the container (`sail artisan …`)	Sail	no

--runtime=sail|local answers up front; a non-interactive run takes the default. Inside the container the runtime is still Sail — the files written for the project keep the prefix — but the installer's own commands run directly, since it is already where Sail would have sent them.

Conventions in brief

Actions pattern: thin controllers — validate (Form Request) → bind → $action->handle(...) → Resource. One final action per write, single handle().
No explanatory comments; PHPDoc only (array shapes / generics).
Commits: type(scope): message, all lowercase, subject only.
Tests: Pest, ≥90% coverage. Unit tests assert the object; feature tests assert the database. tests/Feature/{Model}/{Method}Test.php.
PHP: strict types, final classes, constructor property promotion, explicit return types, curly braces always.

License

MIT

paolobellini / laravel-preset

Package info

Security

Statistics

README

Install

What it does

`packages` — runtime packages

`configs` — lint / format / static analysis

`ai` — conventions

`scripts` — composer quality scripts + dev deps

`lefthook` — pre-commit hooks (opt-in)

`github` — CI workflows and dependency updates

Triaging a vulnerability with OpenVEX

Renovate instead of Dependabot

`--renovate-selfhosted`

Sharding a large suite

`skills` — agent skills

`codegraph` — the project's code graph

Flags

Runtime — where the project's commands run

Conventions in brief

License

paolobellini / laravel-preset

Package info

Security

Statistics

README

Install

What it does

packages — runtime packages

configs — lint / format / static analysis

ai — conventions

scripts — composer quality scripts + dev deps

lefthook — pre-commit hooks (opt-in)

github — CI workflows and dependency updates

Triaging a vulnerability with OpenVEX

Renovate instead of Dependabot

--renovate-selfhosted

Sharding a large suite

skills — agent skills

codegraph — the project's code graph

Flags

Runtime — where the project's commands run

Conventions in brief

License

`packages` — runtime packages

`configs` — lint / format / static analysis

`ai` — conventions

`scripts` — composer quality scripts + dev deps

`lefthook` — pre-commit hooks (opt-in)

`github` — CI workflows and dependency updates

`--renovate-selfhosted`

`skills` — agent skills

`codegraph` — the project's code graph