prautoblogger/ARCHITECTURE.md

PRAutoBlogger — Architecture

Cross-app context: decisions that affect multiple plugins (Cloudflare AI Gateway routing, OpenRouter account sharing, the interface pattern, image-generation stack, social distributor choice) are recorded in Peptide Repo CTO/docs/engineering/decisions/. The incident runbook for cross-app failure modes is at Peptide Repo CTO/docs/engineering/INCIDENT-RUNBOOK.md. Engineer PRAutoBlogger should read both before making decisions that cross plugin boundaries.

PRAutoBlogger is a WordPress plugin that monitors social media (starting with Reddit) for recurring questions, complaints, and comparisons in a configured niche, uses LLM agents (via OpenRouter) to generate high-quality blog articles from those insights, runs them through a chief editor agent for quality review, and auto-publishes them on a configurable daily schedule. All collected data and generation metrics are stored for a self-improvement feedback loop.

---

Data Flow

┌─────────────────────────────────────────────────────────────────────────┐
│                    WP-Cron / Action Scheduler                          │
│                    (Daily trigger, configurable)                        │
└────────────┬────────────────────────────────────────────────────────────┘
             │
             ▼
┌─────────────────────────┐
│  1. Source Collector     │  Pulls data from enabled sources:
│  (Reddit + LLM Research)│  - Reddit RSS/Atom (primary) / .json (fallback)
│                         │  - LLM Deep Research (reasoning model)
│                         │  Stores raw data in `ab_source_data` table
└────────────┬────────────┘
             │
             ▼
┌─────────────────────────┐
│  2. Content Analyzer    │  LLM call (cheap model) to detect:
│  (Analysis Agent)       │  - Recurring questions
│                         │  - Complaints / pain points
│                         │  - Product comparisons
│                         │  Stores analysis in `ab_analysis_results` table
│                         │  Outputs ranked article ideas
└────────────┬────────────┘
             │
             ▼
┌─────────────────────────┐
│  3. Idea Scorer &       │  Deduplicates against existing posts
│     Deduplicator        │  Scores ideas by relevance, frequency, freshness
│                         │  Picks top N ideas (N = daily article target)
└────────────┬────────────┘
             │
             ▼
┌─────────────────────────┐
│  4. Writer Agent        │  Configurable pipeline:
│  (Content Generator)    │  Single-pass: one LLM call → complete article
│                         │  Multi-step: outline → draft → polish
│                         │  Uses quality model via OpenRouter
└────────────┬────────────┘
             │
             ▼
┌─────────────────────────┐
│  5. Chief Editor Agent  │  LLM-powered editorial review:
│                         │  - Checks quality, accuracy, tone
│                         │  - Checks SEO (title, headings, keyword density)
│                         │  - Can request rewrites or approve
│                         │  - Approved → publish; Rejected → flag for human
└────────────┬────────────┘
             │
             ▼
┌─────────────────────────┐
│  6. Publisher           │  Creates/updates WordPress post
│                         │  Sets categories, tags, featured image
│                         │  Stores generation metadata in post_meta
│                         │  Logs cost/tokens in `prab_generation_log` table
└────────────┬────────────┘
             │
             ▼
┌─────────────────────────┐
│  6b. Image Pipeline     │  Generates two A/B images in parallel:
│  (optional, if enabled) │  - Image A: article-driven (featured)
│                         │  - Image B: source-driven (post meta)
│                         │  Provider derived from the saved Image Model
│                         │  via Image_Model_Registry::provider_for().
│                         │  NSFW-blocked slots (CF error code 3030) are
│                         │  retried once with a rule-based fallback
│                         │  prompt via Image_NSFW_Retry; a second block
│                         │  logs WARNING and publishes without that image.
│                         │  v0.17.0: results then pass through the local
│                         │  deterministic Image_Composer (corner mark +
│                         │  branded OG/square variants for Image A; mark
│                         │  only for Image B) with a capability ladder
│                         │  (imagick → GD resize-only → pass-through).
│                         │  Sideloads to media library, logs costs.
└────────────┬────────────┘
             │
             ▼
┌─────────────────────────┐
│  7. Metrics Collector   │  Tracks post-publish performance:
│  (Separate cron job)    │  - WordPress native (views, comments)
│                         │  - GA4 (pageviews, bounce rate, time on page)
│                         │  - Composite "content score"
│                         │  Feeds back into Analysis step for self-improvement
└─────────────────────────┘

---

File Tree

prautoblogger/
├── prautoblogger.php                  # Plugin bootstrap — minimal, only hook registration
├── uninstall.php                      # Clean removal of ALL plugin data
├── readme.txt                         # WordPress.org standard readme
├── composer.json                      # Autoloading (PSR-4) and dependencies
├── ARCHITECTURE.md                    # This file
├── CONVENTIONS.md                     # Naming patterns, extension guides
├── CHANGELOG.md                       # Semantic versioning changelog
│
├── assets/
│   ├── brand/                         # Vendored brand PNGs rasterized from peptide-repo-brand SVGs (composer overlays)
│   │   ├── logo-mark-small-56.png     # Corner mark, 1x
│   │   ├── logo-mark-small-112.png    # Corner mark / OG band logo, 2x
│   │   ├── logo-horizontal-reverse-128.png # Square footer lockup, 1x
│   │   └── logo-horizontal-reverse-256.png # Square footer lockup, 2x
│   ├── fonts/                         # Bundled OFL faces for deterministic text rendering
│   │   ├── Poppins-Bold.ttf           # Square-card caption face
│   │   ├── Poppins-SemiBold.ttf       # OG band caption face
│   │   └── OFL.txt                    # SIL Open Font License
│   ├── css/
│   │   ├── admin.css                  # Admin page styles (wp-admin conventions)
│   │   └── posts-widget.css           # Frontend posts widget styles (uses theme CSS vars)
│   └── js/
│       ├── admin.js                   # Admin page interactivity (vanilla JS / Alpine.js)
│       └── posts-widget.js            # React component for frontend post cards (wp.element)
│
├── includes/
│   ├── class-prautoblogger.php        # Main orchestrator — registers all hooks, delegates execution
│   ├── class-pipeline-schema-installer.php # v0.18.0 substrate tables (prompts, run_sources, run_decisions, runs, run_stages)
│   ├── class-migrate-prompt-seed-v0180.php # One-shot v0.18.0 prompt-registry seed migration
│   ├── class-executor.php             # Cron handlers, generation AJAX (start + status), model registry
│   ├── class-ajax-handlers.php        # Non-generation AJAX: images, models, test connections
│   ├── class-generation-lock.php      # DB-level atomic mutex for single-writer generation
│   ├── class-activator.php            # Activation: create DB tables, set defaults, schedule cron
│   ├── class-deactivator.php          # Deactivation: clear cron, cleanup transients
│   ├── class-autoloader.php           # PSR-4-style autoloader for plugin classes
│   │
│   ├── admin/
│   │   ├── class-admin-page.php       # Main settings page (tabbed SaaS-style UI)
│   │   ├── class-settings-fields.php  # Declarative settings: sections + core fields (API/Models/Content/Sources)
│   │   ├── class-settings-fields-extended.php # Operational fields: schedule, publishing, display, analytics, images
│   │   ├── class-image-model-registry.php # Static list of image generation models for the model picker
│   │   ├── class-admin-notices.php    # Onboarding notices, error alerts, budget warnings
│   │   ├── class-dashboard-widget.php # WP Dashboard widget showing generation status
│   │   ├── class-post-metabox.php     # Metabox on posts showing generation metadata
│   │   ├── class-metrics-page.php     # Admin page for cost dashboard and content scores
│   │   ├── class-ideas-browser.php    # Browse all analysis results (article ideas)
│   │   ├── class-review-queue.php     # Approve/edit/reject queue for generated drafts
│   │   └── class-log-viewer.php       # Activity log viewer with level filtering
│   │
│   ├── core/
│   │   ├── class-scheduler.php        # WP-Cron / Action Scheduler job management
│   │   ├── class-source-collector.php # Orchestrates data collection from all sources
│   │   ├── class-content-analyzer.php # LLM-powered analysis of collected social data
│   │   ├── class-analysis-prompts.php # System/user prompt builders + performance context for analyzer
│   │   ├── class-idea-scorer.php      # Ranks and deduplicates article ideas (delegates to Semantic_Dedup)
│   │   ├── class-semantic-dedup.php   # Embedding-based cosine similarity dedup with keyword fallback
│   │   ├── class-content-generator.php# Writer agent — manages the generation pipeline
│   │   ├── class-content-prompts.php  # Builds all LLM prompts for content generation (system, stage, linking rules)
│   │   ├── class-peptide-linker.php   # Deterministic post-processor: hyperlinks peptide mentions to /peptides/ pages
│   │   ├── class-chief-editor.php     # Editor agent — LLM-powered editorial review
│   │   ├── class-publisher.php        # Creates WordPress posts from approved content
│   │   ├── class-post-assembler.php   # Post-creation helpers: taxonomy, log linking, images, sanitization
│   │   ├── class-image-pipeline.php   # Orchestrates A/B image generation (parallel via batch)
│   │   ├── class-image-prompt-builder.php # Generates visual prompts from article/source data
│   │   ├── class-image-template-filler.php # Fills the editorial style template's {{ topic_summary }} slot (v0.16.0)
│   │   ├── interface-image-composer.php    # Contract: provider bytes → featured-first variant list (v0.17.0)
│   │   ├── class-image-composer.php        # Composer orchestrator: cached capability probe + degradation ladder
│   │   ├── class-image-composer-imagick.php # Imagick renderer: corner mark, OG band, square card (WP-free)
│   │   ├── class-image-composer-canvas.php # Imagick primitives: smoke test, cover-crop, logo, deterministic PNG (WP-free)
│   │   ├── class-image-composer-editor.php # Resize-only rung via wp_get_image_editor (GD path)
│   │   ├── class-image-composer-layout.php # Geometry defaults + caption clamp (pure, WP-free)
│   │   ├── class-image-attacher.php        # Sideload + cost log + variant meta-linking + caption prepend
│   │   ├── class-image-media-sideloader.php # Imports images into WordPress media library
│   │   ├── class-cost-tracker.php     # Logs all API costs, enforces budget limits
│   │   ├── class-stage-display-map.php # Stage vocabulary: labels + default roles + prompt keys (old + new + image stages)
│   │   ├── class-prompt-registry.php  # Versioned prompt registry, read side (render, pins, self-healing fallback)
│   │   ├── class-prompt-registry-writer.php # Registry write side (create version, activate, seed) — versions immutable
│   │   ├── class-prompt-defaults.php  # Canonical v1 bodies: content.* + analysis.* (seed + fallback single source)
│   │   ├── class-prompt-defaults-editorial.php # Canonical v1 bodies: editor.* / research.system / image.* (composer seam)
│   │   ├── class-run-context.php      # Per-process active run id (set by Cost_Tracker::set_run_id)
│   │   ├── class-run-state.php        # runs-table ledger + lifecycle row (ceiling/reserved/settled, pins, status)
│   │   ├── class-cost-governor.php    # Per-run reserve-before-call enforcement (atomic conditional UPDATE)
│   │   ├── class-cost-ceiling-exception.php # Thrown on ceiling breach (run already halted)
│   │   ├── class-run-stage-state.php  # Per-run per-stage state machine (idempotent resume, output snapshots)
│   │   ├── class-run-reaper.php       # Stuck-run sweep + audit-payload retention (rides the #19 cron)
│   │   ├── class-audit-writer.php     # run_sources / run_decisions insert layer
│   │   ├── class-pipeline-status.php  # Status-transient + summary helpers (extracted from runner/worker)
│   │   ├── class-logger.php           # Structured logging singleton (error/warning/info/debug)
│   │   ├── class-article-worker.php   # Single-article generation (content + edit + publish)
│   │   ├── class-pipeline-runner.php  # Orchestrates pipeline; chains per-article cron jobs
│   │   ├── class-ga4-client.php       # Google Analytics 4 API client (OAuth + Data API)
│   │   └── class-metrics-collector.php# Collects post performance data (WP + GA4)
│   │
│   ├── providers/
│   │   ├── interface-llm-provider.php    # Contract for any LLM provider
│   │   ├── class-open-router-provider.php # OpenRouter API implementation (chat completions)
│   │   ├── class-open-router-completion-guard.php # Empty-completion guard: warn on finish_reason=length, one reasoning-off retry, audited failure (v0.18.1)
│   │   ├── class-open-router-embedding-provider.php # Text embeddings for semantic dedup
│   │   ├── class-open-router-pricing.php  # Model pricing lookup and cost estimation
│   │   ├── interface-source-provider.php # Contract for any social media source
│   │   ├── class-reddit-json-client.php  # Reddit HTTP client — RSS (primary) + .json (fallback)
│   │   ├── class-llm-research-provider.php # LLM deep research source (reasoning models)
│   │   ├── class-reddit-provider.php     # Reddit data collection orchestrator (RSS primary)
│   │   ├── interface-image-provider.php  # Contract for any image generation provider (incl. batch)
│   │   ├── class-open-router-image-provider.php  # OpenRouter image gen (single + batch dispatch)
│   │   ├── class-open-router-image-batch.php     # Parallel curl_multi execution for batch image gen
│   │   ├── class-open-router-image-support.php   # API key, response parsing, retry/backoff helpers
│   │   ├── class-open-router-image-pricing.php   # Model resolution + per-image cost estimation
│   │   ├── class-open-router-config.php          # API base URL (direct vs AI Gateway)
│   │   ├── class-open-router-request-builder.php # Request body assembly + reasoning token cap/headroom (v0.18.1) + headers + Hostinger cURL auth workaround
│   │   ├── class-runware-image-provider.php     # Runware FLUX.1 (default v0.9.0+)
│   │   ├── class-runware-image-pricing.php      # FLUX schnell/dev cost table + resolver
│   │   ├── class-runware-image-support.php      # Key, response parsing, retry, dimension snap
│   │   ├── class-runware-image-batch.php        # True parallel curl_multi batch dispatcher
│   │   └── (new providers go here — see CONVENTIONS.md)
│   │
│   ├── services/
│   │   ├── interface-model-registry.php           # Contract for provider-specific model registries
│   │   ├── class-open-router-model-registry.php    # Fetches, caches, queries the OpenRouter model list
│   │   ├── class-open-router-model-normalizer.php  # Maps raw OpenRouter data → standardized record shape
│   │   └── (Phase 3: class-runware-model-registry.php)
│   │
│   ├── frontend/
│   │   ├── class-article-typography.php # Inline CSS for font, size, table borders on generated posts
│   │   └── class-posts-widget.php     # [prautoblogger_posts] shortcode + REST endpoint
│   │
│   └── models/
│       ├── class-source-data.php      # Value object: raw social media post/comment
│       ├── class-analysis-result.php  # Value object: analyzed topic with scores
│       ├── class-article-idea.php     # Value object: scored article idea
│       ├── class-content-request.php  # Value object: generation request with params
│       ├── class-editorial-review.php # Value object: editor verdict, scores, revised content
│       ├── class-generation-log.php   # Value object: log entry for a generation run
│       └── class-content-score.php    # Value object: composite performance score
│
├── templates/
│   └── admin/
│       ├── settings-page.php          # Settings page template (tabbed sidebar layout)
│       ├── metrics-page.php           # Metrics/cost dashboard template
│       ├── review-queue.php           # Review queue table template
│       ├── log-viewer.php             # Activity log viewer template
│       └── metabox-generation-info.php# Post metabox template
│
├── languages/
│   └── prautoblogger.pot              # i18n translation template
│
└── tests/
    ├── unit/                          # PHPUnit tests (mocked dependencies)
    └── integration/                   # WordPress integration tests

---

Frontend Widget Data Flow

┌──────────────────────────────┐
│  WordPress Page/Post         │  Contains [prautoblogger_posts] shortcode
│  (e.g. Home page)            │
└────────────┬─────────────────┘
             │  Shortcode renders:
             │  1. <div id="prab-posts-root"></div>
             │  2. Enqueues posts-widget.js + posts-widget.css
             │  3. Passes config via wp_localize_script
             ▼
┌──────────────────────────────┐
│  React Component (wp.element)│  Mounts into #prab-posts-root
│  posts-widget.js             │  Shows loading skeleton on mount
└────────────┬─────────────────┘
             │  Fetches asynchronously
             ▼
┌──────────────────────────────┐
│  REST API Endpoint           │  GET /wp-json/prautoblogger/v1/posts
│  class-posts-widget.php      │  Queries published posts with
│                              │  _prautoblogger_generated = '1' meta
│                              │  Returns slim JSON (id, title, excerpt,
│                              │  url, date, category, image, word_count)
│                              │  5-minute Cache-Control header
└──────────────────────────────┘

Shortcode Attributes

Attribute	Default	Description
count	6	Posts to display (max 12)
category	(all)	Filter by category slug
title	"Latest Research & Insights"	Widget heading
subtitle	"Evidence-based articles..."	Widget subheading

CSS Integration

The widget CSS (posts-widget.css) uses the Peptide Starter theme's CSS custom properties (--color-*, --spacing-*, --text-*, --radius-*, --transition-*) with hardcoded fallbacks. This means:

• On sites using Peptide Starter: widget automatically adapts to light/dark mode via data-theme.
• On other themes: fallback values render a sensible dark-themed design.

---

Background Generation Flow (v0.3.0)

The "Generate Now" button runs the pipeline in a background WP-Cron process to avoid Hostinger's 120-second LiteSpeed connection timeout. The frontend polls for status.

┌──────────────────────────────┐
│  Admin clicks "Generate Now" │  admin.js sends AJAX to on_ajax_generate_now()
└────────────┬─────────────────┘
             │  Handler returns immediately (< 1 second):
             │  1. Writes "running" transient with started timestamp
             │  2. Schedules one-shot cron: prautoblogger_manual_generation
             │  3. Fires non-blocking loopback to wp-cron.php
             │  4. Returns JSON success to browser
             ▼
┌──────────────────────────────┐
│  admin.js starts polling     │  setInterval every 5 seconds
│  on_ajax_generation_status() │  Shows stage text + spinner
└────────────┬─────────────────┘
             │
             ▼ (separate PHP process)
┌──────────────────────────────┐
│  on_manual_generation()      │  WP-Cron fires this in a new request
│  (class-executor.php)        │  1. ignore_user_abort(true) — survives HTTP kill
│                              │  2. set_time_limit(300)
│                              │  3. Acquires PRAutoBlogger_Generation_Lock
│                              │  4. Runs PRAutoBlogger_Pipeline_Runner::run()
│                              │     (pipeline broadcasts stage updates to transient)
│                              │  5. Writes final result to transient
│                              │  6. Releases lock
└────────────┬─────────────────┘
             │
             ▼
┌──────────────────────────────┐
│  Poller detects completion   │  Reads transient: status = 'complete' or 'error'
│  Shows result message        │  Resets button to idle state
└──────────────────────────────┘

Fallback detection: If Hostinger kills the PHP process before it updates the transient (despite ignore_user_abort), the status endpoint detects the orphan after 180 seconds by checking for recently-created posts with _prautoblogger_run_id meta. It also releases any stale generation lock. After 600 seconds (STATUS_TTL), it gives up unconditionally and reports a timeout.

---

Database Schema

Custom Tables

All tables use $wpdb->prefix + prautoblogger_ prefix.

ab_source_data — Raw social media data

Column	Type	Description
id	BIGINT UNSIGNED	Auto-increment PK
source_type	VARCHAR(50)	'reddit' (extensible for future providers)
source_id	VARCHAR(255)	Platform-specific unique ID (post/comment ID)
subreddit	VARCHAR(255)	Subreddit or channel name (nullable)
title	TEXT	Post/video title
content	LONGTEXT	Post body / comment text / transcript
author	VARCHAR(255)	Original author username
score	INT	Upvotes/likes/engagement metric
comment_count	INT	Number of comments/replies
permalink	VARCHAR(500)	URL to original content
collected_at	DATETIME	When we collected this data
metadata_json	LONGTEXT	Platform-specific extra data (JSON)

• INDEX: source_type, collected_at
• UNIQUE INDEX: source_type, source_id

ab_analysis_results — Analyzed topics and patterns

Column	Type	Description
id	BIGINT UNSIGNED	Auto-increment PK
analysis_type	VARCHAR(50)	'question', 'complaint', 'comparison'
topic	VARCHAR(500)	Detected topic/theme
summary	TEXT	LLM-generated summary of the pattern
frequency	INT	How many source posts mention this
relevance_score	FLOAT	LLM-assigned relevance score (0-1)
source_ids_json	LONGTEXT	JSON array of ab_source_data.id references
analyzed_at	DATETIME	When analysis was performed
metadata_json	LONGTEXT	Extra analysis data (JSON)

• INDEX: analysis_type, relevance_score
• INDEX: analyzed_at

prab_generation_log — API cost and generation tracking

Column	Type	Description
id	BIGINT UNSIGNED	Auto-increment PK
post_id	BIGINT UNSIGNED	WordPress post ID (nullable, set after publish)
run_id	VARCHAR(36)	UUID linking all entries from one pipeline run
stage	VARCHAR(50)	See "Stage vocabulary" below. Canonical edit-stage name is 'polish' (this doc previously said 'edit'; the code always wrote 'polish')
provider	VARCHAR(50)	'openrouter'
model	VARCHAR(100)	Model identifier used
prompt_tokens	INT	Input tokens consumed
completion_tokens	INT	Output tokens generated
estimated_cost	DECIMAL(10,6)	Estimated USD cost
request_json	LONGTEXT	Full request payload (for debugging)
response_status	VARCHAR(20)	'success', 'error', 'timeout'
error_message	TEXT	Error details if failed (nullable)
agent_role	VARCHAR(50)	v0.18.0, nullable. Role that made the call ('writer', 'editor', 'analyst', 'researcher', 'illustrator', …). Historical rows stay NULL
prompt_version	VARCHAR(20)	v0.18.0, nullable. Pinned prompt-registry version the call rendered with. Historical rows stay NULL
created_at	DATETIME	When the API call was made

• INDEX: post_id
• INDEX: run_id
• INDEX: created_at
• INDEX: stage

Stage vocabulary. stage is a VARCHAR, not a SQL enum; the vocabulary lives in PRAutoBlogger_Stage_Display_Map (PHP), which renders historical and new values coherently and maps each stage to its default agent role + primary prompt-registry key:

• Historical values (exist on prod): analysis, outline, draft, polish, review, llm_research, image_a, image_b, image_prompt_rewrite, opik_eval_judge. Note: single-pass generation logs its one call as draft (no outline row — that absence is how single-pass and multi-step runs are told apart in cost reporting).
• Pipeline v2 values (new runs may write): research, curate, draft, editorial, seo, publish.
• Unknown values render via a humanizing fallback — the audit view never renders blank.

prab_event_log — Structured application logging

Column	Type	Description
id	BIGINT UNSIGNED	Auto-increment PK
level	VARCHAR(10)	'error', 'warning', 'info', 'debug'
context	VARCHAR(100)	Origin: 'pipeline', 'reddit', 'ga4', etc.
message	TEXT	Human-readable log message
meta_json	LONGTEXT	Optional structured data (nullable)
created_at	DATETIME	When the event occurred

• INDEX: level, created_at
• INDEX: created_at

ab_content_scores — Post performance metrics

Column	Type	Description
id	BIGINT UNSIGNED	Auto-increment PK
post_id	BIGINT UNSIGNED	WordPress post ID
pageviews	INT	Total pageviews (WP native or GA4)
avg_time_on_page	FLOAT	Average seconds on page (GA4)
bounce_rate	FLOAT	Bounce rate percentage (GA4)
comment_count	INT	WordPress comment count
composite_score	FLOAT	LLM-computed content quality score (0-100)
score_factors_json	LONGTEXT	JSON breakdown of what contributed to score
measured_at	DATETIME	When metrics were collected

• INDEX: post_id
• INDEX: measured_at
• INDEX: composite_score

Pipeline v2 substrate tables (v0.18.0, db 1.2.0)

Real table names use the standard prefix, e.g. wp_prautoblogger_prompts (the plan-of-record's prab_* names are shorthand). Created by PRAutoBlogger_Pipeline_Schema_Installer; all are dropped on uninstall.

prautoblogger_prompts — versioned prompt registry

Column	Type	Description
id	BIGINT UNSIGNED	Auto-increment PK
prompt_key	VARCHAR(64)	Registry key, e.g. 'content.single_pass' (key is reserved SQL)
version	INT UNSIGNED	Monotonic per key; rows are immutable — edits create new versions
body	LONGTEXT	Prompt template; {{ token }} placeholders filled at render time
model	VARCHAR(100)	Optional model hint (informational in Phase 1)
params_json	LONGTEXT	Optional params snapshot (temperature, max_tokens, …)
author	VARCHAR(100)	Who created the version ('seed:v0.18.0', wp user login, agent)
created_at	DATETIME	Version creation time
active	TINYINT(1)	Exactly one active row per key

• UNIQUE: prompt_key, version · INDEX: prompt_key, active

prautoblogger_run_sources — per-run source audit

Column	Type	Description
id	BIGINT UNSIGNED	Auto-increment PK
run_id	VARCHAR(36)	Pipeline run UUID
agent_role	VARCHAR(50)	Role that surfaced the source
source_url	VARCHAR(500)	Source URL (nullable)
doi	VARCHAR(255)	DOI when known (nullable)
kept	TINYINT(1)	Keep/discard decision
reason	TEXT	Why kept or discarded (nullable)
quality_score	FLOAT	Source quality weighting (nullable)
created_at	DATETIME	Row creation time

• INDEX: run_id — consolidates the old source_ids_json + _prautoblogger_research_sources scatter for new runs; historical runs are not backfilled.

prautoblogger_run_decisions — per-stage decision audit

Column	Type	Description
id	BIGINT UNSIGNED	Auto-increment PK
run_id	VARCHAR(36)	Pipeline run UUID
stage	VARCHAR(50)	Stage that decided (see stage vocabulary)
verdict	VARCHAR(50)	e.g. 'approved', 'revised', 'rejected', 'halted'
rationale	TEXT	Decision rationale (nullable)
citation_score	FLOAT	Nullable until the Phase-2 editorial loop computes it
created_at	DATETIME	Row creation time

• INDEX: run_id

prautoblogger_runs — run ledger + lifecycle

Column	Type	Description
run_id	VARCHAR(36)	PK — pipeline run UUID
status	VARCHAR(20)	pending / running / done / failed / halted
ceiling_usd	DECIMAL(10,6)	Per-run cost ceiling snapshotted at run start
reserved_usd	DECIMAL(10,6)	Outstanding reserve-before-call holds
settled_usd	DECIMAL(10,6)	Actual settled spend
overage_usd	DECIMAL(10,6)	Amount the breaching reservation exceeded the ceiling by
pinned_prompts_json	LONGTEXT	Map of prompt_key → version frozen at run start
started_at	DATETIME	Run start
updated_at	DATETIME	Last ledger/state write
finished_at	DATETIME	Set on done/failed/halted (nullable)

• INDEX: status — the Cost_Governor's reserve is a single conditional UPDATE against this row (same atomic discipline as the #10 generation lock), so concurrent curl_multi writers cannot slip past the ceiling between check and call.

prautoblogger_run_stages — per-run per-stage state machine

Column	Type	Description
id	BIGINT UNSIGNED	Auto-increment PK
run_id	VARCHAR(36)	Pipeline run UUID
stage	VARCHAR(50)	Stage name (see stage vocabulary)
agent_role	VARCHAR(50)	Fan-out dimension. Populated in Phase 1 from Stage_Display_Map::default_agent_role() (e.g. 'writer', 'editor', 'publisher'). Phase-2 quorum passes multiple roles for the same stage.
item_key	VARCHAR(64)	Scopes article-level stages within multi-article runs ('' for run-level)
status	VARCHAR(20)	pending / running / done / failed / halted
attempt	SMALLINT UNSIGNED	Re-entry counter
cost_usd	DECIMAL(10,6)	Cost attributed to the stage
meta_json	LONGTEXT	Stage output snapshot — lets resume reuse a done stage instead of re-charging it (payload pruned on the retention cron)
started_at	DATETIME	First entry
updated_at	DATETIME	Last transition
finished_at	DATETIME	Set on done/failed (nullable)

• UNIQUE: run_id, stage, agent_role, item_key (the idempotency key) · INDEX: status, updated_at

WordPress Options (wp_options)

All prefixed with prautoblogger_:

Option Key	Description
prautoblogger_openrouter_api_key	Encrypted OpenRouter API key
prautoblogger_ai_gateway_base_url	Optional Cloudflare AI Gateway URL (proxies OpenRouter); empty = direct
prautoblogger_ai_gateway_cache_ttl	Seconds Cloudflare may cache identical LLM responses (0 = off)
prautoblogger_ga4_property_id	Google Analytics 4 property ID
prautoblogger_ga4_credentials_json	Encrypted GA4 service account credentials
prautoblogger_analysis_model	OpenRouter model for analysis (default: cheap)
prautoblogger_writing_model	OpenRouter model for writing (default: quality)
prautoblogger_editor_model	OpenRouter model for chief editor (default: quality)
prautoblogger_reasoning_max_tokens	v0.18.1 — hard cap on thinking tokens per call when reasoning is enabled; sent as OpenRouter reasoning.max_tokens (replaces effort when active) and the request max_tokens is raised by the cap so reasoning can never consume the visible-content budget; 0 = legacy uncapped effort mode. Default PRAUTOBLOGGER_DEFAULT_REASONING_MAX_TOKENS (2048)
prautoblogger_daily_article_target	Number of articles per day (1-10, default: 1)
prautoblogger_writing_pipeline	'single_pass' or 'multi_step' (default: multi_step)
prautoblogger_niche_description	Text description of the site's niche
prautoblogger_target_subreddits	JSON array of subreddits to monitor
prautoblogger_monthly_budget_usd	Monthly API spend limit in USD
prautoblogger_per_run_cost_ceiling_usd	v0.18.0 — per-run hard cost ceiling (USD); reserve-before-call enforced; snapshotted onto the runs row at run start; 0 = disabled. Default PRAUTOBLOGGER_DEFAULT_RUN_CEILING_USD ($0.50)
prautoblogger_request_json_retention_days	v0.18.0 — days before heavy audit payloads (generation_log.request_json, run_stages.meta_json) are NULLed by the daily cleanup cron; 0 = keep forever. Default PRAUTOBLOGGER_DEFAULT_REQUEST_JSON_RETENTION_DAYS (14)
prautoblogger_tone	Content tone (informational, conversational, etc.)
prautoblogger_min_word_count	Minimum article word count (default: 800)
prautoblogger_max_word_count	Maximum article word count (default: 2000)
prautoblogger_topic_exclusions	JSON array of topics to never write about
prautoblogger_enabled_sources	JSON array of active source types
prautoblogger_auto_publish	Toggle: auto-publish approved posts (default: '0')
prautoblogger_default_author	WP user ID for generated post authorship
prautoblogger_default_category	WP category ID fallback for generated posts
prautoblogger_log_level	Logging threshold: error/warning/info/debug
prautoblogger_db_version	Schema version for migrations
prautoblogger_schedule_time	Daily generation time (HH:MM, default: '03:00')
prautoblogger_image_model	Image model slug from Image_Model_Registry::get_models(); provider is derived from this on save
prautoblogger_image_provider	Derived from the chosen model on save (v0.8.0+); not editable in admin UI
prautoblogger_image_prompt_instructions	System prompt given to the image rewriter LLM (v0.8.0+); falls back to Image_Prompt_Builder::REWRITER_SYSTEM_PROMPT when empty. v0.16.0: the default now instructs the LLM to emit a 1-2 sentence editorial topic/mechanism summary as the SCENE (was a comic gag)
prautoblogger_image_nsfw_retry	Toggle: retry NSFW-blocked image slots with a rule-based fallback prompt (default: '1')
prautoblogger_migrated_image_provider_v080	One-shot migration flag — auto-heals mismatched provider/model pairs on first admin_init after v0.8.0 upgrade
prautoblogger_article_font_family	Font family key: 'default', 'inter', 'georgia', 'merriweather', 'lora', 'open_sans', 'roboto', 'system'
prautoblogger_article_font_size	Body font size in px (0 = theme default). Recommended: 16–18.
prautoblogger_table_borders	Toggle: add borders/padding/striping to tables (default: '1')
prautoblogger_image_style_template	Full image prompt template (v0.16.0+); contains exactly one {{ topic_summary }} token filled per-article with the rewriter scene. Default = PRAUTOBLOGGER_DEFAULT_IMAGE_STYLE_TEMPLATE (editorial scientific illustration). Replaces the Style Suffix. Validated on save (single-token, brief A5)
prautoblogger_image_style_suffix	DEPRECATED (v0.16.0). Former comic style suffix appended to every image prompt; no longer read by the builder. Mirrored to prautoblogger_image_style_suffix_deprecated for one cycle and not deleted yet
prautoblogger_openrouter_model_registry	Normalized OpenRouter model list (JSON array, daily refresh, serves as durable cache)
prautoblogger_openrouter_model_registry_fetched_at	Unix timestamp of last successful model registry refresh
prautoblogger_image_compose_enabled	Toggle: run the deterministic image composer (default '1'; auto-degrades, never blocks publishing)
prautoblogger_image_compose_variants	Comma list of composed variants for Image A (default 'og,square'; whitelisted at point of use)
prautoblogger_image_featured_mark_enabled	Toggle: subtle corner mark on featured images (default '1')
prautoblogger_image_compose_capability	Cached capability probe {fingerprint, capability}; fingerprint = PHP version + imagick/gd presence, so it auto-invalidates on host changes

Post Meta

Stored on every PRAutoBlogger-generated post:

Meta Key	Description
_prautoblogger_generated	Boolean flag — '1' if generated by plugin
_prautoblogger_analysis_id	FK to ab_analysis_results.id
_prautoblogger_source_ids	JSON array of source data IDs used
_prautoblogger_model_used	Model that generated the content
_prautoblogger_pipeline_mode	'single_pass' or 'multi_step'
_prautoblogger_total_cost	Total USD cost for this post
_prautoblogger_total_tokens	Total tokens consumed for this post
_prautoblogger_editor_verdict	'approved', 'revised', 'rejected'
_prautoblogger_editor_notes	Chief editor's review notes
_prautoblogger_generated_at	ISO 8601 timestamp of generation
_prautoblogger_research_sources	JSON array of source URLs used
_prautoblogger_idea_hash	v0.18.0 — stable hash of the idea (title
_prautoblogger_og_image_id	Attachment ID of the composed 1200×630 OG variant (v0.17.0; the rebuilt SEO stage emits og:image from this — key name frozen)
_prautoblogger_square_image_id	Attachment ID of the composed 1080×1080 square variant (v0.17.0; stored now, no consumer yet)

Composed-variant attachment meta (v0.17.0): every pipeline attachment gets _prautoblogger_image_role (featured/og/square); OG/square variants also get _prautoblogger_base_attachment_id pointing at the featured attachment they derive from. All keys keep the _prautoblogger_ prefix, so the uninstall prefix-sweep purges them.

---

External API Integrations

Service	Purpose	Auth	Rate Limit	Code
OpenRouter	All LLM calls (analysis, writing, editing)	API key (encrypted in wp_options)	Per-model	providers/class-open-router-provider.php, providers/class-open-router-pricing.php
Cloudflare AI Gateway (optional)	Transparent proxy in front of OpenRouter — adds response caching, cost logging, rate limiting, provider fallback	Same OpenRouter key; gateway URL in prautoblogger_ai_gateway_base_url	Gateway-side quotas	Same provider file; activated when gateway URL option is non-empty
Reddit RSS	Primary Reddit data source — Atom feeds for subreddit hot posts	None (unauthenticated)	No known rate limit; reliable from datacenter IPs	providers/class-reddit-json-client.php
Reddit .json	Fallback for posts + only source for comments	None (unauthenticated)	~10 req/min (datacenter IPs often blocked)	providers/class-reddit-json-client.php
Google Analytics 4	Post performance metrics	OAuth2 service account	Standard GA4 limits	core/class-ga4-client.php, core/class-metrics-collector.php
Runware (FLUX.1 via runware.ai)	Default image backend (v0.9.0+): schnell ~$0.0006/image, dev ~$0.02/image. True parallel generation via curl_multi. v0.15.0: live model catalog sync from /v1/models endpoint (task-based API); normalized to Image_Model_Registry shape; cached with 24h TTL; on-demand refresh via AJAX.	API key (encrypted in wp_options)	Account-level quotas	providers/class-runware-image-provider.php, providers/class-runware-image-pricing.php, providers/class-runware-image-support.php, providers/class-runware-image-batch.php, providers/class-runware-model-catalog.php

---

Opik LLM Observability (Optional)

Status: Wave 1 (v0.12.0) — tracing skeleton. Wave 2 (future) — eval harness with frozen-dataset regression testing.

Opik is a cloud-hosted LLM observability platform (Comet) providing:

• Per-call tracing: Every OpenRouter call generates a span with model, tokens, latency metadata.
• Trace grouping: One trace per article generation bundles all spans (draft, QA, image prompt).
• Prompt versioning: Foundation for Wave 2 — compare new prompt revisions against a frozen reference dataset using LLM-as-judge scoring on rubric axes (factual accuracy, style match, hallucination risk).

Architecture:

• Feature flag: prautoblogger_opik_enabled (WP option, default off). When false, zero network traffic to Opik.
• Credentials: API key and workspace from PHP constants PRAUTOBLOGGER_OPIK_API_KEY, PRAUTOBLOGGER_OPIK_WORKSPACE (defined in wp-config.php). Never stored in DB.
• Span lifecycle:
  1. Opik_Trace_Context::init_trace() generates UUID at article-worker start.
  2. Each LLM call wraps with start_span() → LLM request → end_span() with tokens + response.
  3. finalize_trace() + Opik_Span_Queue::enqueue() accumulates trace + spans for async dispatch.
  4. prautoblogger_opik_dispatch cron action batches spans/traces (max 100 per POST) to /v1/private/spans/batch, /v1/private/traces/batch.
• Retry logic: Exponential backoff (2s, 4s, 8s) on 5xx errors; permanent abort on 4xx; dropped after 3 attempts per item.
• Queue: WP options-backed; max 1000 items; 12-hour TTL; persists across page reloads for truly async dispatch.

Call sites instrumented (Wave 1):

• Draft generation (single-pass or multi-step outline/draft/polish).
• Editorial review (chief editor LLM call).
• Image prompt generation (deferred; file at 314 LOC, over limit).

Cost & Performance:

• Opik free tier: 25k spans/month. Estimated current usage ~4.2k/month (20 posts/day × 7 calls × 30 days). Headroom to 5× growth.
• Impact on article generation: async dispatch, zero blocking latency overhead.
• Wave 2 (eval harness): adds LLM-as-judge calls (~$14/month at Haiku pricing, assuming 2 eval runs/week).

Code structure:

includes/services/opik/
  class-opik-client.php            — REST client (auth, retry, batch POSTs)
  class-opik-trace-context.php     — per-request trace state singleton
  class-opik-span-queue.php        — WP options-backed async queue
  class-opik-dispatcher.php        — cron handler, batch dispatch

includes/admin/
  class-opik-settings.php          — admin UI (toggle, project name, status)

---

Key Decisions

#1: OpenRouter over direct provider APIs

Gives access to many models through one API, simplifies provider management, and lets the user switch models without code changes. Trade-off: slight latency overhead and dependency on OpenRouter's availability.

#2: Reddit as first social source (RSS primary, .json fallback)

Reddit has rich discussion data ideal for identifying recurring questions and pain points. Reddit RSS/Atom feeds are the primary data source — they work reliably from datacenter IPs (Hostinger) where .json endpoints return 403. The .json endpoints serve as a fallback for posts and are the only source for comment fetching (RSS doesn't include comments). Trade-off: RSS lacks engagement metrics (score, comment count), but post titles and content are sufficient for topic analysis.

#3: Chief editor agent instead of human review queue

The user wants full automation — a second LLM pass reviews quality, SEO, and accuracy before publishing. Posts that fail editorial review are flagged for human intervention rather than published. Trade-off: doubles the LLM cost of the "review" step, but catches quality issues.

#4: Configurable pipeline (single-pass vs. multi-step)

Some users want cheap/fast, others want high quality. Making this configurable avoids forcing one approach. Trade-off: more code paths to maintain.

#5: Composite content score using LLM

Rather than just tracking raw metrics, we periodically have the LLM evaluate what made high-performing posts succeed and low-performing posts fail. This feeds back into the analysis and generation prompts. Trade-off: additional API cost for scoring, but enables the self-improvement loop.

#6: Source provider interface for future platforms

Reddit is the only implemented source, but the interface pattern means adding a new source (YouTube, TikTok, etc.) is one class implementation plus a settings checkbox. We removed the old stub provider files (TikTok, Instagram, YouTube) because dead code confuses AI agents and humans alike — the interface is the contract, not empty classes. Trade-off: slight over-engineering for day one, but pays off immediately at source #2.

#7: Custom tables for high-volume data

Source data, analysis results, and generation logs are high-write, time-series data. WordPress post_meta and options are wrong for this. Custom tables with proper indexes. Trade-off: more complex activation/uninstall, but correct.

#8: All API keys encrypted at rest

Using wp_salt() as encryption key with OpenSSL. Not bulletproof (salt is in wp-config.php) but significantly better than plaintext in wp_options. Trade-off: adds complexity to option get/set, but necessary for security.

#9: Structured logger instead of raw error_log()

All application logging flows through PRAutoBlogger_Logger singleton, which writes to the prab_event_log table and forwards errors/warnings to PHP's error_log(). This gives users an in-admin Activity Log page with filtering by level, search, and pagination — much more accessible than server logs. Trade-off: one extra DB write per log entry, but the table is prunable and indexed.

#10: Database-level atomic mutex for generation lock

Uses INSERT IGNORE on wp_options (which has a UNIQUE index on option_name) instead of transient-based locking. This eliminates the TOCTOU race condition where two concurrent cron runs could both read the transient as "not locked" before either sets it. Expired locks older than 1 hour are cleaned up first to prevent permanent deadlock.

#11: wp.element for frontend React (no build step)

The posts widget uses WordPress-bundled React (wp.element) with raw createElement calls instead of JSX. This avoids requiring a Node.js build pipeline for what is a simple card grid. Trade-off: more verbose component code, but zero tooling dependencies for the plugin consumer.

#12: REST API for frontend data fetching

The widget fetches posts asynchronously from a dedicated REST endpoint rather than rendering server-side. This keeps initial page load fast and allows the endpoint to be cached independently (5-minute Cache-Control). The permission callback uses a filter (prautoblogger_rest_posts_public) so sites can restrict access if needed.

#13: Run-ID based log linking

Each pipeline execution generates a UUID (run_id) that tags every prab_generation_log entry. When a post is published, link_generation_logs() uses UPDATE WHERE run_id = X to associate entries with the post_id. This is more reliable than the previous timestamp-window approach, especially in batch runs where multiple posts are generated in quick succession. v0.8.1+: run_id is also persisted on each post as _prautoblogger_run_id post_meta, giving the orphan-research reaper a direct path from run to siblings without walking the gen_log table (see #19).

#19: Daily orphan-research-row reaper (v0.8.1)

When a pipeline dies between post creation and the final Post_Assembler::amortize_research_costs() call (Hostinger exec kill, OOM, LiteSpeed cut, fatal), the shared llm_research cost row stays in prab_generation_log with post_id = NULL and sibling articles understate their true cost in the breakdown popover. A new daily WP-Cron event prautoblogger_reap_orphan_research_rows (hooked to PRAutoBlogger_Research_Reaper::on_cron, 03:15 server time — 15 min after the primary generation cron) scans for orphan rows older than a 1-hour grace window, looks up sibling posts by _prautoblogger_run_id post_meta (fallback: distinct post_ids in gen_log for the same run_id), and delegates to the existing idempotent amortize_research_costs() on match. Orphans that still have no siblings after 7 days are deleted outright as sunk cost. A one-time activator migration (prautoblogger_migrated_run_id_meta_v081) back-populates _prautoblogger_run_id on existing posts from gen_log so the meta-primary path works on upgrade. Manual ops trigger: wp prautoblogger reap-research.

#14: Reddit RSS replaces PullPush.io (and earlier Reddit OAuth)

Reddit rejected our OAuth API application (April 2026). We initially switched to PullPush.io, but its index was frequently stale or unavailable. Reddit's RSS/Atom feeds (/r/{sub}/hot.rss) proved the most reliable option — they work from datacenter IPs where .json gets 403, require no auth, and have no apparent rate limit. The .json endpoints are kept as a fallback for posts and as the sole source for comment data. Each collected item's metadata includes a data_source field (reddit_rss or reddit_json) for auditability.

#18: OpenRouter model registry — daily refresh, WP option + transient cache, zero-coupling

The admin model picker (v1) needs to list OpenRouter models with pricing and capabilities. We fetch https://openrouter.ai/api/v1/models (free, unauthenticated) once daily and store the normalized payload in a WP option fronted by a 24h transient. On stale-and-fetch-fails, we serve last-good + surface a warning. The registry class (class-open-router-model-registry.php) takes all config via constructor (option name, transient name, endpoint URL) — no PRAUTOBLOGGER_* constants inside the class body. Phase 2 lifts it into a shared Composer package with zero internal edits. Phase 3 adds a parallel Cloudflare_WorkersAI_Model_Registry behind the same interface. Capability vocabulary: text→text, text+image→text, text+audio→text, text→image, text→audio, text→video, text→embedding. Cost: $0/month.

#20: Editorial illustration prompt + style-template setting (v0.16.0, May 2026)

Commit 1 of the in-plugin editorial image pipeline (brief docs/proposals/2026-05-29-image-pipeline-in-plugin-brief.md). The image style pivots from a single-panel newspaper comic to a text-free editorial scientific illustration. Two changes ship here; the deterministic PHP composer is a later commit.

1. Prompt structure. Image_Prompt_Builder::REWRITER_SYSTEM_PROMPT now asks the rewriter LLM for a concise 1-2 sentence topic/mechanism summary (one concrete centered focal subject, no text/people-as-gag/logos) as the SCENE, keeping the short CAPTION line for the unchanged HTML-caption-below-image path. The scene/caption parsing contract (Image_Scene_Parser) is unchanged.
2. Template substitution. The old trim($scene . ' ' . $style_suffix) concat is replaced by PRAutoBlogger_Image_Template_Filler::fill($scene) in all three entry points (build_article_prompt, build_source_prompt, build_fallback_prompt). The filler resolves the active template (admin override → default), strips control chars and clamps the summary length (brief A5), then substitutes the single {{ topic_summary }} token. Degradation (brief A5/A6): if the template lacks exactly one token it appends the summary and logs a warning; if the summary is empty it emits the style-only prompt — a blank or token-only prompt is never sent to the provider.
3. Setting. New textarea prautoblogger_image_style_template (default PRAUTOBLOGGER_DEFAULT_IMAGE_STYLE_TEMPLATE) replaces the Style Suffix field. On save it is sanitised with sanitize_textarea_field and validated to contain exactly one token (Image_Template_Filler::sanitize_for_save). The old prautoblogger_image_style_suffix is mirrored to ..._deprecated for one cycle and not deleted (one-time migration prautoblogger_migrated_style_template_v0160).

Provider/model are unchanged — only the prompt text changes, so production keeps emitting via Runware FLUX.1 schnell but now produces text-free editorial base images.

#21: Deterministic PHP image composer with capability ladder (v0.17.0, Jun 2026)

Commit 2 (and final scope) of the in-plugin editorial image pipeline (#20). Small diffusion models cannot render clean text (2026-04-21 A/B finding), so base images are text-free — but the one image that travels without its HTML caption is the social/OG share image. A deterministic local composer (PRAutoBlogger_Image_Composer behind PRAutoBlogger_Image_Composer_Interface) now runs between provider bytes and sideload:

1. Variants. Image A → corner-marked featured (1200×632) + branded OG (1200×630, teal band, brand mark, baked caption in Poppins SemiBold) + square card (1080×1080: full base downscaled into a 1080×569 top slice — no upscaling — cream caption panel in Poppins Bold, teal footer with the horizontal reverse lockup). Image B → corner mark only. The variant set is config-driven (prautoblogger_image_compose_variants, default og,square; square has no consumer yet — CEO wants it rendered + stored now). Variants are meta-linked via _prautoblogger_image_role, _prautoblogger_base_attachment_id, _prautoblogger_og_image_id, _prautoblogger_square_image_id. The og key is frozen: the rebuilt SEO stage (pipeline-redesign §4.8) emits og:image from it.
2. Capability ladder, never fatal. Imagick is probed once (PNG format + TTF annotate smoke test) and cached in prautoblogger_image_compose_capability, fingerprinted by PHP version + imagick/gd presence so host flips (hPanel/cagefs — it has happened) auto-invalidate the cache. Rungs: imagick → wp_get_image_editor() resize-only (unbranded, square skipped because it would upscale) → pass-through. Every render is try/caught; degradation logs one WARNING per run; the publish always proceeds with the editable HTML caption intact. Test/ops override: prautoblogger_image_compose_capability filter.
3. Determinism. Bundled OFL fonts (no system-font queries), vendored brand PNGs (no runtime SVG delegate variance), stripImage() + PNG date/time chunks excluded + fixed compression → byte-stable output per environment (not across ImageMagick versions). Composition is local CPU, $0, logged as a duration-only stage on the image_composer channel — no cost-tracker rows.
4. Layout tuning. Geometry/opacity are constants in PRAutoBlogger_Image_Composer_Layout::defaults() behind the prautoblogger_image_compose_layout filter — deliberately not settings; promote only if the CEO asks to tune them. The caption is hard-clamped (word-boundary wrap + ellipsis, multibyte-safe) even though the rewriter LLM is already constrained to <15 words.

#17: Runware as default image backend (v0.9.0, Apr 2026)

After a comic-style A/B round in Apr 2026, FLUX.1 schnell via Runware was chosen as the default image backend. Cost is ~$0.0006/image vs $0.039/image for Gemini 2.5 Flash Image via OpenRouter (≈65× cheaper), and the looser schnell fidelity reads as editorial-cartoon style — a feature, not a bug, for our single-panel comic aesthetic. FLUX.1 dev ($0.02/image, 28 steps) remains available as an opt-in for higher-fidelity runs. The Runware layer mirrors the OpenRouter split (interface + provider + support + pricing + batch), all under the 300-line cap. The batch class dispatches imageInference POSTs via curl_multi_init/exec/select — wall-clock time for the A/B pair drops to ≈ the slowest single image. A v0.9.0 one-time migration (prautoblogger_migrated_default_image_v090) flips sites still on the legacy default; explicit user selections are preserved. Cloudflare Workers AI was removed as an image provider in v0.10.0 (see ADR #16 note below).

#16: Image generation via Cloudflare Workers AI (FLUX.1) — SUPERSEDED v0.10.0

Kept here for history. Until v0.8.2 the default image backend was FLUX.1 schnell on Cloudflare Workers AI, routed via the Cloudflare AI Gateway for caching + unified telemetry. In v0.9.0 (2026-04-21) Runware FLUX.1 schnell became the default (65× cheaper, see ADR #17). In v0.10.0 (2026-04-21) the Cloudflare Workers AI provider was removed entirely — classes, settings, tests, and registry entries are gone; any CF user is migrated to runware:100@1 on upgrade via PRAutoBlogger_Migrate_Remove_Cloudflare_V0100::run(). The Cloudflare AI Gateway in front of OpenRouter (ADR #15) is unaffected — it is a distinct integration.

#15: Optional Cloudflare AI Gateway in front of OpenRouter

We already use Cloudflare for DNS/CDN on peptiderepo.com, so layering AI Gateway in front of OpenRouter is zero marginal infrastructure. It gives us response caching (meaningful for repeated classification/scoring calls), a unified cost/latency dashboard, rate limiting, and provider fallback — all of which we would otherwise have to build ourselves to satisfy the CTO cost-tracking rules. Kept as an opt-in URL setting (prautoblogger_ai_gateway_base_url) so the plugin still works unchanged out of the box and can be bypassed instantly if the gateway misbehaves. The gateway is a transparent OpenRouter-compatible proxy; no new provider class is needed, and the response parsing path (usage, choices[0].message.content) is unchanged.

---

#22: Pipeline v2 Phase 1 substrate (v0.18.0, db 1.2.0, June 2026)

Phase 1 of the CEO-approved pipeline rebuild (plan of record: convo/cpo/threads/2026-06-prautoblogger-pipeline-redesign/03-cto-revised-brief.md), built on the current pipeline with no behavior change to the Economy (single-pass) and multi-step publish paths. Four subsystems:

1. Versioned prompt registry (prautoblogger_prompts). All pipeline prompt copy (content/analysis/editor/research + the illustration rewriter prompt and the image Style Template) lives as immutable versioned rows; the hardcoded v0.16.0 texts are seeded as v1 and remain the in-code fallback, so a missing/empty table can never fatal — call sites render through PRAutoBlogger_Prompt_Registry and fall back to the same constants the seed used (byte-identical output). A run pins the active version of every key at start (runs.pinned_prompts_json) and every generation_log row is stamped with the pinned prompt_version. The image-composer PR consumes the image.* keys — it must not grow its own prompt storage. No admin UI in Phase 1; the API supports list/activate/ create-version (never mutate) for the Phase-2 Prompts screen.
2. Additive audit schema. agent_role + prompt_version columns on generation_log; run_sources + run_decisions child tables; PRAutoBlogger_Stage_Display_Map renders historical + image + v2 stage values. request_json (and stage output payloads in run_stages.meta_json) are nulled after R days — R from the prautoblogger_request_json_retention_days setting (default 14), pruned on the existing daily reaper cron.
3. Per-run cost governor (PRAutoBlogger_Cost_Governor). Reserve-before-call against the per-run ledger row using a single conditional UPDATE (the #10 atomic-write discipline); estimates come from the #18 pricing chain; curl_multi image batches reserve their summed estimate before dispatch; reservations settle to actuals after each response. Breach → run halted, overage recorded, un-dispatched work aborted, surfaced on the Review Queue. Ceiling = prautoblogger_per_run_cost_ceiling_usd setting (default $0.50, 0 = disabled). The monthly Cost_Tracker cap and the Cloudflare AI Gateway (#15) path are unchanged.
4. Run state machine + idempotency (PRAutoBlogger_Run_State). Per-run per-stage state (pending/running/done/failed/halted) keyed run_id + stage + agent_role + item_key (item_key scopes article-level stages because one run_id spans all N articles of a batch); done stages persist output and are reused on re-entry, never re-charged; post creation is keyed by _prautoblogger_run_id + _prautoblogger_idea_hash (check-before-insert) so a retried run cannot duplicate a post. PRAutoBlogger_Run_Reaper (same daily cron as #19) marks runs stuck running > 2× expected stage wall-clock as failed; such runs render as "incomplete" in audit queries. Quorum logic (multiple roles per stage) itself is Phase 2. In Phase 1, agent_role is populated on every row from Stage_Display_Map::default_agent_role() so the run timeline is self-describing and the idempotency key is fully populated.

#20: PHPUnit test infrastructure + WordPress-Core PHPCS compliance (v0.10.1)

Tests use Brain\Monkey for WordPress function mocking (no database required). The BaseTestCase singleton handles setup/teardown of Monkey stubs and provides fixture helpers for common data shapes (SourceData, ArticleIdea, GenerationLog, etc.). All WordPress functions used by the plugin are stubbed in BaseTestCase::setUp() so individual test classes only override what they need. In v0.10.1, missing stubs for post_type_exists() were added to support PeptideLinker's PR-Core guards. All 1,362 WordPress-Core coding-standards violations (short array syntax, line length, spacing) were auto-fixed via phpcbf and the CI gate was flipped to strict mode — PHPCS failures now block CI. This prevents future regressions on the code style front.

Cross-System LLM Budget Coordination

PRAutoBlogger and Peptide News both call OpenRouter and may share a single API key / billing account. Their combined spend should be considered when setting per-plugin budgets.

Plugin	Default Models	Typical Daily Spend	Budget Control
PRAutoBlogger	Gemini 2.5 Flash Lite (analysis + editing), Claude Sonnet 4 (writing)	$0.05–$0.30 depending on article count	Hard-stop monthly budget in plugin settings
Peptide News	Google Gemini 2.0 Flash (keywords + summaries)	$0.01–$0.05	No hard budget yet (planned)

Important: If your OpenRouter account has a global spending limit, set each plugin's budget to less than half the total. PRAutoBlogger will hard-stop when its budget is exhausted, but Peptide News currently has no budget enforcement — a spike in news fetches could consume shared quota.

Future improvement: A shared wp_options key (e.g., ecosystem_monthly_llm_budget) that both plugins read, with each plugin reserving its allocation on startup. This requires coordination at the ecosystem level and is tracked as a medium-term goal.