WriteWP changelog and release notes

Agentic chat, free featured images, Cloud Solver, eight prompt presets, and reliability fixes.

• New — Agentic chat: create blog drafts, import/rewrite any URL, list posts, and read site context (OpenAI, OpenRouter, Gemini).
• New — WordPress 6.9 Abilities API integration for MCP Adapter agents.
• New — WriteWP Cloud Solver for managed Cloudflare bypass (FlareSolverr still supported).
• New — Free AI featured images via Pollinations with optional paid providers and free fallback.
• New — Eight advanced prompt presets (e-commerce, news, how-to, reviews, listicles, comparisons, local SEO, tech/AI).
• Improved — Dynamic generation parameters; usage tracking covers every path; stronger duplicate detection.
• Fixed — Automation reliability (cron self-heal, site-timezone schedule hours, activity heartbeats) and admin UX fixes.

• Improved — All AI model lists refreshed to the latest models — OpenAI's GPT-5 family (default GPT-5.4 mini), Google Gemini 3.5 Flash (new default) plus the Gemini 3.1 image and TTS models, and current OpenRouter and OpenCode Zen catalogs — with matching pricing and cost estimates. Deprecated and retired models were removed from the pickers.
• Fixed — The automated daily "refresh model lists" background task was never actually scheduled, so model dropdowns only updated on a manual refresh. It now runs daily (and self-heals on existing sites after updating), keeping your available models current automatically.
• Improved — OpenRouter model lists now survive a temporary OpenRouter outage by falling back to the last successfully-fetched catalog instead of a small static list.
• Improved — Chat image support now correctly recognizes the new default models (Gemini 3.5 Flash and the GPT-5 family) as vision-capable.

• Improved — Importer: changing a source's category now instantly fetches that source's articles, and the per-source Fetch button respects the selected category. No need to use the top-right "Fetch Articles" button.
• Fixed — Cloudflare Bypass now captures and replays your browser's exact User-Agent alongside the cf_clearance cookie. Cloudflare binds the cookie to the User-Agent, so the importer's generic User-Agent previously caused the bypass to fail. (Note: some aggressively-protected sites may still block server-side requests via IP or TLS fingerprinting, which a cookie cannot override.)
• Improved — Cloudflare Bypass is now point-and-click — choose the host from a dropdown of your source sites, then paste the whole cookie string and the importer finds the cf_clearance value for you. Host matching also tolerates www vs non-www.

• Fixed — With any affiliate link configured, every import failed with "AI returned empty content" because the affiliate-link inserter used an invalid regex. It has been rewritten as a safe HTML tokenizer that never discards content.
• Fixed — Continuation overlap de-duplication never triggered, so continued chat and article output could repeat the previous paragraph. The overlap detector now works.
• Fixed — Continue-draft (long articles split across continuations) duplicated the entire existing draft into the post.
• Fixed — The Writer queue never displayed after enqueuing. It's now shown correctly.
• Fixed — Generated OpenAI images used the wrong orientation for 4:3 and 3:4 sizes.
• Fixed — Saving generated audio as Opus or PCM failed ("file type not permitted"). These formats are now allowed.
• Fixed — "Schedule Hours" (restrict automation to specific hours) never worked — the typed value was discarded. It is now parsed correctly (e.g. 9-17 or 9,15,20).
• Fixed — Chat attachments larger than 10 KB were sent to the AI as unreadable data. Text files are now sent as text.
• Fixed — The setup wizard's progress and skip actions never saved due to a nonce mismatch.
• Fixed — Dashboard "Success Rate" was understated when more than 50 imports existed. It's now computed over all records.
• Fixed — Streaming chat now falls back to the configured fallback provider on real provider errors, and comparison-mode answers no longer poison the normal chat cache.
• Removed — The non-functional "Enable Auto-Import" switch (the Schedule control is the single automation on/off).

• Fixed — Write-queue items could get stuck on "Processing" and time out if the queue was re-indexed while the background worker started. The worker now locates its item by ID instead of position.
• Fixed — Chat response caching was silently disabled for Ollama (local and cloud). Identical Ollama answers are now served from cache.
• Fixed — Studio video: choosing 1080P or 512P while Duration was set to 10s discarded the resolution. It now keeps the resolution and drops the duration to 6s.
• Fixed — Non-streaming chat: Gemini truncated-reply continuation passed no API key, and OpenCode Zen requests silently routed to Gemini. Both corrected.

• Fixed — AI Studio video generation defaulted to an unsupported "720P" resolution and failed. It now uses 768P.
• Fixed — MiniMax text-to-speech audio was decoded as base64 when the API returns hex, producing corrupt files. Audio now saves and plays correctly.
• Fixed — Gemini text-to-speech used non-existent model IDs that returned 404. Corrected to the current preview TTS models.
• Fixed — Gemini TTS output (headerless PCM) is now wrapped in a WAV container so it plays in the browser and Media Library.
• Fixed — Gemini image model IDs updated to current names; Imagen models now route to the correct API endpoint.
• Note — Confirmed the plugin uses MiniMax's international API (api.minimax.io) and removed the unused "MiniMax Group ID" field (Group IDs are only used by the separate China API).

• Fixed — Removed a PHP 8.0-only union return type that caused a fatal error on the supported PHP 7.4 minimum.
• Security — Imported settings now pass through the full sanitizer, including the Ollama host SSRF check.
• Security — The Ollama Cloud custom host, the setup-wizard host save/test, and the system health check now reject private and internal addresses.
• Privacy — "Delete all data on uninstall" now removes every plugin option, per-user chat history and templates, queues, and post metadata the plugin creates.
• Compatibility — Network-activated multisite now provisions the plugin's tables and options on every subsite, including newly-created subsites.
• Privacy — The admin interface no longer loads fonts from external CDNs — the bundled fonts are served locally.

• Fixed — Ollama host validation correctly allows localhost while still blocking other private/internal IP ranges (SSRF protection).
• Fixed — Source deletion now properly re-indexes the source sites array.
• Improved — Added configurable limits for chat conversations and messages to prevent unbounded storage growth.

• New — Complete transition from the legacy admin to the React-based admin UI.
• Security — Encryption-at-rest upgraded to authenticated encryption (libsodium with an AES-256-GCM fallback); API keys are never returned in plaintext.