Drupal Discovery Kit

## Problem

AI coding agents (Claude Code, OpenAI Codex, ChatGPT, Gemini) are making technology recommendations that shape real procurement decisions. When these agents evaluate Drupal, two problems emerge:

1. **Perception gap** — Agents' training data underrepresents Drupal 10/11. They emit Drupal 7-shaped patterns, miss modern APIs, and rank Drupal lower than its actual capabilities warrant (Dries Buytaert's 2026 agent study ranked Drupal 3rd behind Next.js and WordPress despite matching the brief better on capability).

2. **Verification gap** — Even when agents know Drupal can do something, they can't prove it on a live installation. Competitor marketing (Pagely, rtCamp) publishes "Drupal can't do X" claims that agents can't independently verify.

Discovery Kit solves both problems by instrumenting Drupal with live, extensible benchmarks that agents can discover, run, and cite as evidence.

## Approach

**Benchmark plugins.** Each benchmark is a Drupal plugin (`#[Benchmark]` attribute) that exercises a real Drupal capability — multilingual content, granular permissions, structured content modeling, plugin-based capability registration — and publishes a pass/fail result with timing, memory, and structured data. Any module can contribute benchmarks.

**Agent tool interface.** A typed `/discovery/tools` endpoint (6 tools with JSON Schema definitions) lets agents introspect a running Drupal site: list content types with field schemas, discover capabilities, run benchmarks, check diagnostics. No guessing at YAML from stale training data.

**Source-driven rebuttals.** Benchmarks that target competitor false claims cite the specific source in their docblock (`@see` URL + quoted claim) and prove Drupal can do what competitors say it cannot. Existing rebuttals target Pagely's "Drupal has no capability registration system" and "WordPress wins on structured content for AI" claims.

**Public evidence feeds.** Results publish as Drupal nodes syndicated via RSS 2.0 and JSON Feed 1.1 — no authentication required. AI agents poll feeds, discover new evidence, and incorporate it into recommendations. An `/llms.txt` endpoint provides structured agent context with version info, code patterns, and tool references.

**Built for Drupal 11.** PHP 8.3+, strict types, `final` classes everywhere, constructor property promotion, dependency injection. No contrib dependencies — only core modules.

NOTE: This is entirely and autonomously AI agents planned, developed, and maintained project experiment.