component-context-generator - SKILL.md Agent Skill

name: component-context-generator description: > Generates a CONTEXT.md file for a PrestaShop shared component inside the `.ai/Component/` folder. Trigger this skill when the user asks to "generate a context for [Component]", "document the [X] component", "fill in the CONTEXT.md for [Component]", or when working inside `.ai/Component/` directories. Components live under `src/Core/{Name}/` and/or `src/Adapter/{Name}/` — they are shared infrastructure, not business domains. Examples: Grid, Form, Hook, CQRS, Translation, Router. subagent: recommended

PrestaShop Component CONTEXT.md Generator

Purpose

Generate a lean, accurate CONTEXT.md for a PrestaShop shared component (src/Core/{Name}/ + src/Adapter/{Name}/) by exploring the real codebase.

Components are not business domains — they have no CQRS structure. They provide shared infrastructure consumed by many domains.

Output format

Principle: paths not inventories. Class names are greppable on demand. What earns tokens is:

The layer table (where to look)
Non-obvious patterns (things that would surprise a competent PHP developer)

# {Component Name} Component

> **Status:** Draft — this context file is a starting point and should be refined by domain experts.

## Purpose

{1–2 sentences: what this component provides and what it does NOT do}

## Layers

| Layer | Path |
|-------|------|
| {layer name} | {path} |

## Non-obvious patterns

- {bullet: surprising architectural decision, gotcha, or non-obvious constraint}

## Canonical examples

- {file path — 1-line role}

## Skills

- [`skill-name`](../../skills/skill-name/SKILL.md) — one-line description

## Related

- [{Component}]({path}) — {why related}

## Skills is optional — only include it if a skill in .ai/skills/ targets this component. Omit the section entirely if no relevant skill exists.

Do NOT include: ## Coding standards, ## Do, ## Don't, ## Testing expectations, ## Architecture overview with verbose subsections. These inflate token cost without adding value.

Target size: aim for ~30–50 lines. Simple components fit in that range. If a component legitimately needs more (multiple parallel patterns, several layers, substantial non-obvious behaviour), prefer splitting into sub-contexts over inflating a single file — see "When to split into sub-contexts" below. A single CONTEXT.md above ~80–100 lines is a smell.

Step-by-step process

1. Confirm the target is a Component, not a Domain

Component → src/Core/{Name}/ with NO Command/, Query/, Handler/ subdirectories at root
Domain → src/Core/Domain/{Name}/ — use domain-context-generator instead

2. Explore the codebase

Use the Explore agent (thoroughness: very thorough) to map:

src/Core/{Name}/ — interfaces, abstract classes, key concrete classes
src/Adapter/{Name}/ — concrete implementations, legacy bridges
Grep for the component's main interface across src/Core/Domain/ and src/PrestaShopBundle/ — identify 2–3 representative consumers
Note any sub-patterns with a non-obvious design decision

3. Fill the Layers table

One row per architectural layer actually found. Keep paths as specific as possible (file path for single-file layers, directory for multi-file layers).

4. Write Non-obvious patterns

Only include what would surprise a competent PHP developer:

Coexisting design patterns (e.g. legacy vs modern)
Constraints that break obvious assumptions (e.g. "stopPropagation() is blocked")
Generated/cached artifacts that must be refreshed manually
Performance or ordering gotchas
Subtle API distinctions that cause bugs if missed

Skip anything derivable from reading the code for 5 minutes.

5. Write Canonical examples

Pick 2–3 files: the main interface, the most-used implementation, and one domain consumer.

6. Check for relevant skills

List the contents of .ai/skills/ and check if any skill targets this component. If one exists, include a ## Skills section before ## Related linking to it.

7. Write Related (use sparingly)

Links to other .ai/Component/ or .ai/Domain/ context files — but only when the relationship is non-obvious.

The whole point of splitting contexts into separate files is to avoid loading everything at once. Every cross-reference is a potential cascade: an AI agent reads component A, follows a link to component B, follows B's link to C... and ends up loading all contexts. This defeats the purpose of the split.

Include a link when:

The relationship is architecturally surprising (e.g. PositionUpdater lives inside Grid's source tree)
Two components coexist during a migration and the coexistence has gotchas (e.g. Twig ↔ Smarty)

Do NOT include a link when:

The relationship is obvious from imports (e.g. "Controller dispatches CQRS commands")
You're linking just to mention a hook name or a specific class — those are greppable
The link points to a domain just because that domain is a heavy consumer of the component
The link would create a bidirectional reference (A → B and B → A)

When in doubt, omit the link. An agent can always find related contexts via the index in .ai/CONTEXT.md.

When to split into sub-contexts

Most components fit comfortably in a single CONTEXT.md. A few don't — when the component has two (or more) parallel patterns that share little code but coexist under the same umbrella, split the file rather than letting it grow past ~80 lines.

Trigger conditions (all should hold)

Two or more usage patterns coexist (e.g. settings forms vs CRUD forms; legacy ObjectModel vs Doctrine repositories within the same Component).
Most readers only work on one pattern at a time — loading the other pattern's rules wastes tokens.
The patterns each have non-trivial pattern-specific rules (base classes, service folders, hooks, anti-patterns) that don't reduce to a few bullets.
A meaningful shared surface exists (decision tree, shared concerns, skills table) — otherwise consider splitting into two separate Components instead.

Recommended layout

.ai/Component/{Name}/
├── CONTEXT.md            ← lean root: purpose, decision tree, shared concerns,
│                            skills table, links to each sub-context.
│                            Pure routing hub — minimal pattern-specific detail.
├── {PATTERN_A}.md        ← pattern A: required layers, service definitions,
│                            hooks, anti-patterns, canonical example.
├── {PATTERN_B}.md        ← pattern B: same shape as PATTERN_A.
└── skills/
    ├── {pattern-a skill}/  ← body opens with: Read CONTEXT.md + PATTERN_A.md
    ├── {pattern-b skill}/  ← body opens with: Read CONTEXT.md + PATTERN_B.md
    └── {generic skill}/    ← body opens with: Read CONTEXT.md only

Naming: use the existing STRUCTURE.md / GOTCHAS.md / MULTISTORE.md convention for sub-context files — UPPERCASE single noun describing the pattern (e.g. SETTINGS.md, CRUD.md).

Rules

Root CONTEXT.md stays the entry point. Every external link from other components/domains points at CONTEXT.md, not at a sub-context — the decision tree in the root routes the reader to the right sub-context. External links should only specialise to a sub-context when their own surrounding text is unambiguously pattern-specific (saves a hop without confusing readers).
No duplication between root and sub-contexts. Shared concerns live in the root. Pattern-specific rules live in exactly one sub-context.
Skills load sub-contexts conditionally. A pattern-specific skill body opens with Read @.ai/Component/{Name}/CONTEXT.md and Read @.ai/Component/{Name}/{PATTERN}.md. A skill that serves both patterns (e.g. a unified controller-actions skill) opens with Read @.ai/Component/{Name}/CONTEXT.md and instructs the reader to load the relevant sub-context based on the branch they're working on.
Skills table in the root includes a "Pattern detail to load" column so the reader sees the load decision upfront.

Reference

See .ai/Component/Forms/ for a worked example: CONTEXT.md (decision tree + shared concerns) + SETTINGS.md + CRUD.md, with skills that route to one sub-context or the other.

When NOT to split

The component has only one pattern, just lots of layers — solve with a tighter Layers table instead.
The sub-contexts would each be under ~20 lines — the split adds ceremony without saving tokens.
The "two patterns" are actually one pattern with optional features — keep them in one file with a "Conditional layers" subsection.

Reference: lean Grid component example

# Grid Component

> **Status:** Draft — this context file is a starting point and should be refined by domain experts.

## Purpose

Infrastructure for rendering and managing back-office data tables: column definitions, filters, row/bulk actions, query builders, data factories, and drag-and-drop position reordering. Does not contain any business data — each domain provides its own `GridDefinitionFactory` and Doctrine query builder.

## Layers

| Layer | Path |
|-------|------|
| Core contracts + factory | `src/Core/Grid/` |
| Column types, row/bulk actions | `src/Core/Grid/Column/`, `src/Core/Grid/Action/` |
| Query builder base | `src/Core/Grid/Query/AbstractDoctrineQueryBuilder.php` |
| Position updater | `src/Core/Grid/Position/` |
| Adapter utilities | `src/Adapter/Grid/` |

## Non-obvious patterns

- `AbstractGridDefinitionFactory` dispatches `action{GridId}GridDefinitionModifier` hook — modules add columns/actions without touching core code
- `SearchCriteriaInterface` is stored as a Symfony request attribute per grid, not a service — each grid type has its own `{Domain}Filters` class
- 60+ concrete query builders exist (one per domain grid) — all extend `AbstractDoctrineQueryBuilder` and implement `getSearchQueryBuilder()` + `getCountQueryBuilder()`

## Canonical examples

- `src/Core/Grid/Definition/Factory/AbstractGridDefinitionFactory.php`
- `src/Core/Grid/Definition/Factory/ProductGridDefinitionFactory.php`
- `src/Core/Grid/Query/AbstractDoctrineQueryBuilder.php`

## Related

- [PositionUpdater Component](../PositionUpdater/CONTEXT.md) — drag-and-drop reordering sub-layer (lives inside Grid source tree)

Output

Write the completed CONTEXT.md to:

.ai/Component/{Name}/CONTEXT.md

If the directory does not exist, create it first. After writing, confirm the file path to the user.