Zactonz AI Connector for Ollama

1.2.0 — 29 September 2026#

Renamed to Zactonz AI Connector for Ollama, matching the rest of the connector family. The plugin folder, slug and text domain are unchanged, so this is an ordinary update.

Fixed

• Ollama Cloud and self-hosted API keys were interchangeable. A key entered on Settings › Connectors is now filed under the connection mode in use at the time, so switching mode can no longer send a Cloud key to a self-hosted address, or a private self-hosted token to ollama.com.
• A model listing containing an entry without a usable name raised a fatal error. Entries from both the self-hosted and Cloud endpoints are now validated before use.
• Events could be delivered after a stream was cancelled. The trailing block of an unterminated stream is no longer parsed once the caller has returned false, so the partial result holds only what the caller was told about.
• A callback failure on a replayed body, on the trailing block, or on the done event was discarded instead of reported.
• Streamed responses reported no token usage. An OpenAI-compatible stream omits the usage block unless it is asked for.
• Administrator AJAX errors returned HTTP 200, because a WP_Error code was passed where a status code was expected.
• The Ollama mark on Settings › Connectors was supplied at its own 646×854 proportions and did not fit the square the screen draws it in.
• The guard meant to stop the plugin loading twice was never set.
• Ten of the twelve shipped class files had no direct-access guard.
• The settings screen's JavaScript strings could not be translated, because script translations were never registered.

Changed

• Streaming now uses the WordPress HTTP API. The plugin no longer calls cURL anywhere in its own code: a stream is an ordinary wp_remote_post() and the body arrives through the requests-request.progress action.
• A transport without incremental delivery replays the whole response through the same parser, so streaming returns every event instead of failing when the cURL extension is absent.
• Cancellation stops delivering events without aborting the transfer, which no HTTP API hook can do. The partial result is unchanged.
• The reachability check is cached for five minutes per connection. The AI Client asks on every model lookup, and each answer previously cost a request — or the full request timeout when Ollama was not running.
• The Site Health connectivity test runs asynchronously, and reports a recommendation rather than a critical failure when Ollama is simply not running. A direct test held the Site Health screen open for the length of the request.
• Model preference travels through the order models are listed in, rather than five global filters. Those filters are shared with every other connector on the site, so whichever registered last won; listing order is per provider. Every capability's default now leads, not only the text one.
• The API key also lives in the option WordPress registers for the connector, so a key entered on Settings › Connectors and one entered on the plugin's own screen are the same credential. Per-connection keys are still kept, and an existing key is copied across on upgrade.
• The Ollama port widens the HTTP API's safe-port list only while the connector's own request is in flight, rather than for every request the site makes.

Added

• An uninstall handler that removes the settings and both stored keys when the plugin is deleted, on single sites and across a network.
• An External services section in the readme covering what Ollama Cloud receives.

1.1.0 — 3 August 2026#

• Added native Ollama single and batch embedding generation for the WordPress 7.1 / PHP AI Client 1.4 embedding API.
• Reworked discovery to advertise only capabilities reported by Ollama, with conservative fallback behaviour when model details are unavailable.
• Added separate default selectors for text, vision, image, embedding and tool-calling models.
• Added model-default, disabled, low, medium and high thinking preferences, with automatic migration from the previous enabled setting.
• Added a redacted diagnostics panel covering endpoint status, latency, model availability, version, embeddings and streaming readiness.
• Added a WordPress Site Health connectivity test and redacted debug-information section.
• Added stable provider-level SSE streaming with content, thinking, usage, tool-call aggregation, callback-error handling and cancellation with partial results.
• Improved WordPress.org connector-search metadata for admins coming from Settings › Connectors.
• Fixed Connector-screen saves so connection changes no longer discard model and thinking preferences.
• Fixed Ollama Cloud capability reporting so structured output is not advertised where it is unsupported.
• Added automated coverage for discovery, settings migration and preservation, embeddings, diagnostics, Site Health, streaming, cancellation and streamed tool calls.

1.0.0 — 22 May 2026#

• First Zactonz release with Settings › Connectors recognition, Ollama Cloud and self-hosted modes, separate Cloud and self-hosted API credentials, provider logo and live connected-status integration.
• Host URL or IP and port controls with model discovery from the configured Ollama instance.
• Default-model thinking mode and a longer configurable text request timeout for local Ollama responses.
• Text generation, structured output, function calling, vision-aware model metadata and image generation for compatible models.