infer-conventions

# Infer Conventions

Learn how this application writes Laravel, then record what you learn as durable, path-scoped rules other agents will read. You are documenting reality, not improving it.

## Ground Rules (read before you start)

- Consistency first. The codebase's majority style is the convention. Never judge it, never propose a "better" pattern, never record what the code should do. If the app validates inline everywhere, that is the rule, even if Form Requests would be nicer.
- Skip what an active tool produces, keep what a tool would fight. Inspect the project's Pint and Rector configuration first; a Rector transformation is tooling-owned only when its package and relevant rule or set are installed and enabled. Active tools may rewrite code toward one canonical form: `$casts` to `casts()`, `$fillable` to attributes, magic accessors to the `Attribute` class, pipe-string rules to arrays, `$signature` to `#[Signature]`, named migrations to anonymous, and many more. When the app already sits at an active tool's target form, the tool owns it, so record nothing. But when the app deliberately holds a form an active tool would refactor away, such as legacy `getXxxAttribute()` accessors the `Attribute` class would replace, no tool can reproduce that choice and an agent defaults the other way. That against-the-grain hold is exactly what to record.
- Record decisions, not defaults. A consistent pattern earns a rule only when it reflects a choice: the app took one valid option where the framework or common practice offered others, or the pattern would surprise a competent agent. Framework defaults steer nothing, so skip them: anonymous migrations, `$signature` commands, `ShouldQueue` jobs, `casts()` on Laravel 11+, named routes, Rule objects in `app/Rules`, and `Mail::fake()` or `Bus::fake()` to isolate framework services. A real fork is not enough on its own. Weigh the side the app took, and record only the side an agent would not reach for by itself: inline closures everywhere, legacy accessors, a bespoke query layer. Watch for the false fork too. "No Mockery" next to facade fakes is not a choice against Mockery, because they double different things. The test for every candidate: without this rule, would the next agent plausibly write it differently? Only "yes" earns a rule.
- Architecture choices are the gold. Record presence and deliberate absence. The structural pattern the app commits to is the highest-signal convention and the one no tool can decide: Action classes and how they are invoked (`handle` / `execute` / `__invoke`), service objects, dedicated query objects exposing `builder()`, DTOs (spatie/laravel-data vs readonly classes), Form Request validation vs inline, an events and listeners spine vs direct calls, and domain or module folders. Also record a consistent non-pattern, such as "query Eloquent directly in controllers, no repository layer", so the next agent matches the app's altitude instead of over-engineering.
- Never duplicate `.ai/rules`. Read `.ai/rules/index.md` and the area files before the sweep. A dimension already covered there is marked done and skipped.
- Evidence or silence. A convention needs at least 3 consistent examples and no meaningful rival to become a candidate. Every Step 1 verdict applies this bar.
- The recorded rule states the convention, nothing else. One or two imperative lines: this project does X, so do X here. Keep detection evidence out. No counts, ratios, current usage, file lists, or example paths, because that is proof for the confirm step, not part of the rule. One short syntax fragment at most, and point to `search-docs` for API details.

## Process

Each step ends on a checkable completion criterion. Do not advance until it holds.

Fan out when you can. The sweep is embarrassingly parallel. If your environment can spawn subagents (a Task, dispatch, or equivalent tool), do Step 0 yourself, then hand each checklist group (A to J) and the architecture map to its own subagent. Each subagent runs the greps, reads a few representative files, and returns structured verdicts (dimension, verdict, evidence, proposed glob / title / note). You aggregate, dedupe, then run Steps 3 to 5. It is far faster on a real app. No subagents available? Run the steps in sequence, with the same bar and the same output.

### Step 0: Orient

Read `composer.json` (installed packages tell you which checklist groups apply), the `pint.json` / PHPStan / Rector config, `.ai/rules/index.md` if present, and most important, map the `app/` tree. List every directory under `app/` (and any `Modules/`, `src/`, `packages/`, or domain root). Every folder beyond Laravel's default skeleton (`Http`, `Models`, `Providers`, `Console`, `Exceptions`) is a structural pattern the app committed to and a high-value rule waiting to be written: `Actions`, `Services`, `Data` or DTOs, `Queries`, `Repositories`, `ViewModels`, `Pipelines`, `Support`, `Enums`, `Contracts`, `Observers`, or `Domain` and module roots. Note each one. You will confirm how it is used in Step 2.

This app ships a frontend stack, so the frontend checklist group applies. Sweep it.

Done when: you have the applicable checklist groups, the dimensions already recorded in `.ai/rules`, and a list of every non-default `app/` directory mapped to the pattern it represents.

### Step 1: Predefined sweep

Open `references/checklist.md` and work every applicable dimension using its search hints. Give each exactly one verdict:

- Pattern. Clears the bar, rival under ~20% of sites, and reflects a real choice (passes the decisions-not-defaults test). A recording candidate. Cite 2 to 3 example files.
- Conflict. Both styles present in meaningful numbers. Report the split with counts and example files. Never record a preferred winner while the code remains mixed, even in yolo, because that would describe an aspiration rather than reality. Record only if the user identifies a stable path or context boundary that explains both styles; othe