TSE Content Generator

Autonomous SEO content generation system for WordPress, built on an idempotent finite-state-machine pipeline, asynchronous job orchestration, and a fault-tolerant multi-provider AI layer.

TSE Content Generator turns a queue of keywords into fully assembled, illustrated, internally-linked WordPress articles — end to end and unattended. Every article advances through a persisted state machine, every heavy operation runs asynchronously, and every step is idempotent and resume-safe, so the system can survive crashes, timeouts and provider outages without producing duplicates or runaway API spend.

This repository is a portfolio/reference implementation showcasing backend architecture in a real WordPress plugin. API credentials are never stored in code — they are configured in the WordPress admin and resolved at runtime.

Engineering highlights

• Finite-state-machine pipeline. Each task is a row in wp_tse_tasks whose lifecycle is governed by an explicit state machine with validated transitions — no implicit status changes, no ambiguous in-between states.
• Asynchronous, idempotent orchestration. All AI calls, image generation, similarity checks and housekeeping run through Action Scheduler. Jobs are designed to be re-runnable: a retried or duplicated job converges to the same result.
• Deterministic retries. Retry paths contain no randomness. Temperature, back-off and attempt bumps are computed deterministically, so failures are reproducible and debuggable.
• Fault tolerance. The Mistral client implements multi-key failover with an explicit fatal-vs-transient error taxonomy, and the image layer falls back across providers on failure. The infrastructure layer provides reusable resilience primitives — a per-service circuit breaker (closed → open → half-open), a rate limiter, and a named mutex for critical sections shared across concurrent workers.
• Multi-provider AI abstraction. Image generation fans out across 10+ providers behind a single ImageProviderInterface, with a ProviderManager handling fallback and a QuotaTracker enforcing per-provider budgets.
• Layered, dependency-light design. A pragmatic DDD-style separation — Domain, Actions, Infra, Security, Support, Contracts — with PSR-4 autoloading, declare(strict_types=1) everywhere, and no framework beyond WordPress itself.
• Security by default. Capability checks and nonces on every privileged entry point, prepared statements for all SQL, output escaping, and authenticated at-rest encryption (libsodium) for stored secrets.

Pipeline

A task moves through six states, each handled by a dedicated, resumable step:

planning ──► chunking ──► assemble ──► image ──► similarity ──► finished
   │                                                               
   └────────────────────► failed / paused ◄────────────────────────

Feature set

• Keyword engine — research and expansion (alphabet-soup generation, SERP/suggest APIs, scoring, lemmatization), quarantine of duplicates, and atomic keyword locking to prevent two workers picking the same keyword.
• Content generation — section-wise generation against a Mistral LLM, configurable system prompt, internal linking, and SimHash similarity guarding corpus uniqueness.
• Image generation — featured + inline images, SEO-aware count, optimized prompts, and a provider fan-out (Mistral/Flux, Fal.ai, Cloudflare, Bytez, Stable Horde, Pexels, Pixabay, Unsplash, and more) with an optional Cloudflare Worker proxy.
• Operations — diagnostics page, structured component logs, statistics, auto-purge of stale tasks, and Action Scheduler health checks surfaced as admin notices.

Project structure

src/
├── Actions/          # Action Scheduler entry points & FSM step handlers
│   └── Handlers/     #   Outline · Chunking · Image · Generation
├── Admin/            # Admin pages, list tables, notices
├── Contracts/        # Interfaces (Logger, Lock)
├── Database/         # Schema, migrations, keyword storage
├── Domain/
│   ├── Content/      # Generator, PromptBuilder, InternalLinker, SimilarityChecker
│   ├── Images/       # Multi-provider image generation + Providers/
│   ├── Keywords/     # Keyword domain services
│   ├── Orchestrator/ # Pipeline orchestration
│   ├── Publishing/   # Publisher, DraftScheduler
│   ├── Services/     # MediaProcessor, AutoPurge
│   ├── Stats/        # Statistics
│   └── Tasks/        # TaskManager, Planner, TaskStatus (the FSM)
├── Features/
│   └── AutoKeywords/ # Keyword research pipeline
├── Infra/            # Action Scheduler, Mistral client, CircuitBreaker, RateLimiter, Lock, Logger
├── Security/         # Capabilities, Nonces, Crypto (at-rest secret encryption)
└── Support/          # Strings, Time, Arrays, Json utilities

cloudflare-worker/    # Optional edge proxy for image providers

Tech stack

• PHP 8.0+, strict_types, PSR-4 autoloading via Composer
• WordPress 6.2+ plugin APIs, Action Scheduler for async jobs
• Mistral AI as the LLM provider; pluggable image providers
• libsodium for at-rest secret encryption
• Static analysis with PHPStan, coding standards via PHP_CodeSniffer / WPCS

Installation (development)

git clone https://github.com/Dimmer1981/tse-content-generator.git
cd tse-content-generator
composer install

Symlink or copy the directory into wp-content/plugins/, activate it, then configure the API keys and options on the plugin's Settings page. Database tables are created automatically on activation.

Configuration & secrets. All provider credentials (Mistral, image providers) are entered in the WordPress admin and stored as options — they are never committed to the repository.

License

MIT © 2026 Dmitry Duyunov

Author

Dmitry Duyunov — Senior PHP / Backend Engineer LinkedIn · GitHub