Download the PHP package prlkhv/module-graphql-ai-profiler without Composer

Prlkhv_GraphQlAiProfiler

Why is this exists

No one longer want to read a lot of profiler data and dig into performance traces to find out why is this request slow. Another fact, almost everyone use AI tools to explain their profile result. Soooo, I decided to remove this annoying step with copy-pasting profile result to your AI assistant with this tool

Server-side GraphQL profiler for Magento 2. It instruments GraphQL execution, builds an in-memory span tree (every resolver + every SQL query), and injects it into the GraphQL response under extensions.profiling — OTLP-shaped, or a compact LLM-friendly format for AI-assisted analysis.

No OTel Collector, no exporter, no MCP. The profiling payload rides back on the same response the client already receives, so you can profile any GraphQL request by adding a header — no separate tooling, no APM.

It also ships with a Claude Code skill so you can ask an AI "why is this request slow?" and get a root-cause analysis with next steps (see Using it with Claude).

• Requirements: Magento 2.4.x, PHP 8.1–8.3.
• Safe by default: inert unless explicitly enabled, gated behind dev mode and a secret header. See Security.

Installation

Via Composer (recommended)

Manual

Copy this directory to app/code/Prlkhv/GraphQlAiProfiler, then:

Setup

The module is inert until every activation condition is met. Configure it under Stores → Configuration → Advanced → Developer → GraphQL Profiler, or via CLI:

Model\Config::isActive() returns true only when all of these hold:

If any fails, every plugin passes through with near-zero overhead and nothing profiling-related appears in the response.

All settings

Usage

Add the activation header to any GraphQL request. Choose a response format with X-GraphQl-Profiler-Format.

Verbose OTLP format (default)

The response gains extensions.profiling with an OTLP-shaped resourceSpans tree. Feed it to any OTLP-aware viewer, or read it directly.

Compact "AI" format

Add X-GraphQl-Profiler-Format: ai for a compact, LLM-friendly payload: the OTLP envelope is dropped, keys are single letters, IDs are truncated to 6 hex chars, and timestamps are microsecond offsets from trace start. Add X-GraphQl-Profiler-Sql: 1 to include the (truncated) raw SQL per query.

The payload carries no legend — the key mapping is a stable contract in AI_FORMAT_MAPPING.md.

Tip: for any query containing nested quotes, put the body in a file and use curl --data @body.json. Inline -d mangles the \" escapes and produces invalid JSON.

Request headers

Using it with Claude

This module bundles a Claude Code skill (.claude/skills/graphql-profiler/) that turns the profiler into an AI debugging tool. Ask in plain language:

why is this request slow: { products(search: "bag", pageSize: 5) { items { sku price_range { minimum_price { final_price { value } } } } } }

Claude will send the request through the profiler (asking you for the secret and endpoint if it doesn't have them), replay it warm, then analyze the span tree and report:

• Verdict — where the time actually goes (e.g. resolver vs DB, N+1).
• Breakdown — total wall time, slowest spans by self-time, DB/resolver split.
• Root cause — the specific resolver/query and why it's slow.
• Next steps — concrete, Magento-native fixes ranked by impact.

Enabling the skill

The skill lives at .claude/skills/graphql-profiler/ inside this module. Claude Code discovers skills under a project's .claude/skills/ directory, so either:

• Symlink it into your project skills dir:

• or copy that folder to <project-root>/.claude/skills/graphql-profiler.

Then, in Claude Code, invoke it explicitly with /graphql-profiler or just ask a "why is this slow" question and it triggers automatically.

The skill is two files: SKILL.md (the playbook) and scripts/analyze.py (a zero-dependency Python 3 span analyzer). You can also run the analyzer standalone:

How it works

Four plugins, all no-ops when the profiler is inactive:

Spans are held in a per-request SpanCollector; timing comes from a monotonic Clock (hrtime). The DbAdapterPlugin guards against re-entrancy so config lookups it triggers don't recurse.

OTLP encoding notes

• traceId / spanId are emitted as hex (easy for a custom reader). Switch Model\Otlp\Serializer to base64 if you pipe to a real OTLP backend.
• Nanosecond timestamps are emitted as strings (they exceed JS safe-integer range). The AI format instead uses microsecond integer offsets.

⚠️ Security

Do not enable this in production. When active it exposes server internals and SQL statements (db.statement, truncated) in the GraphQL response.

• Bind values are never logged — only the statement text.
• Activation requires all of: enabled flag + allowed deploy mode + constant-time secret match.
• The secret is stored encrypted and compared with hash_equals(). Treat it like a credential — a long, random value — and rotate it if it leaks.

These layers exist to make accidental exposure hard. Treat the secret like a credential and keep mode_allowlist at developer unless you fully understand the exposure.

License

Elastic License 2.0 (ELv2). You may use this module freely, including inside commercial projects. You may not:

• offer it to third parties as a hosted or managed service, or
• resell or sublicense it as a standalone product.

For a commercial license granting those rights, contact Ivan Perelekhov <[email protected]>.