Zybroz AI Agents – AI Automation

Changelog

1.2.0

Major update consolidating all development since 1.1.2 into one WordPress.org release.

Runner Console findings-table report format, Accessibility Scanner / SEO Plugin Auditor / Ollama-default reliability fixes, Database Backups restore hardening, and a wide batch of Runner Console retest fixes.

• Added: run reports now render a standardized findings table (Finding / To Do / Source, plus an AI Suggestion column for agents that write changes) instead of a flat bullet list, for every bundled agent — not just a hand-picked few.
• Added: the final report now shows an aggregate “N issues found” summary line instead of only echoing whichever single step happened to finish last, so a multi-check agent’s earlier results are no longer silently dropped from the summary; reports with findings also get a visual accent so a clean run and a run with issues no longer look identical at a glance.
• Added: Accessibility Scanner’s “Fix with Alt Text Architect” hand-off (previously Classic Runner only) is now available in the Runner Console, and runs directly against the posts already found — no re-selecting a target.
• Added: an agent with a review/approval step can only ever run on one post at a time (a server-side rule that already existed); the Console’s “Run on multiple” option is now disabled with an explanatory tooltip for these agents instead of offering a bulk run that was guaranteed to fail.
• Added: the “this agent has had a failed run before” confirmation can now be auto-confirmed per agent via an opt-in checkbox, collapsing to a persistent one-line reminder on later runs instead of a full dialog every time.
• Added: Database Backups’ storage summary now breaks the total file count down by type — e.g. “(4 DB, 3 Uploads)” — since “Keep max (count)” prunes each type independently and a flat total could read as overdue for pruning when neither type actually was.
• Added: the run-report bubble now shows total run duration, a dedicated “Copy this report” button (separate from “Copy transcript”), and a “View Diagnostic Log” link; a failed step in the transcript now renders with a distinct error glyph and color instead of looking identical to a successful one.
• Added: when a run fails with a recognizable auth/credential error (e.g. “Missing Authentication header”, “Invalid API key”), a “Fix it” button now appears right on that line and opens Settings scrolled to the relevant provider card in a new tab; detection now also covers “User not found”, “No auth credentials found”, “Incorrect API key”, and “Authentication failed”.
• Added: the Run button, and the “Run on multiple” scope/Run button when nothing matches, now show a tooltip explaining why they’re disabled; the bulk-scope item count shows “Counting…” while it loads; Escape closes the target-search dropdown and the “Run on multiple” scope popup; command-bar controls locked during an active run now show a tooltip explaining why.
• Fixed: Accessibility Scanner (and any other agent using the shared missing-alt-text scan) could report an image’s alt text as missing regardless of what was actually saved — a Media Library field that most authors never fill in (most set alt text through the block editor instead) was silently overriding the real, current value, and could also miss edits made directly on the attachment after the post was last saved. Editing or clearing alt text now correctly changes the scan result either way.
• Fixed: SEO Plugin Auditor’s “posts missing a meta description” table could show blank “(untitled)” rows with dead links — the underlying data was a reliable database query, but routing it back through the AI model for display risked exactly this kind of corruption; it no longer does.
• Fixed: setting Ollama as the Default AI Provider didn’t actually make agents run on it — most bundled agents request a tier-resolved cloud model by default, which Ollama routing didn’t recognize, so every run failed with “needs an AI provider that isn’t configured” even with Ollama running. Ollama-only setups also no longer see a misleading “no usable AI provider” banner.
• Fixed: a database restore could shift where the plugin looks for its own backup files, making backups taken after that snapshot disappear from the Database Backups list; typing “RESTORE” to confirm a restore also wrongly armed the Settings page’s “Unsaved changes” save bar.
• Fixed: run-failure/retry messages pointed users at WordPress core’s debug.log or Site Health, which never had the relevant detail — they now point at the plugin’s own Diagnostic Log (Settings > Debug Log); an uncaught exception during a run’s execution is now also written there instead of only showing in the transcript.
• Fixed: the Runner Console’s target picker stayed visible for site-wide agents that don’t take a target, showing a dead search field instead of “No target required – click Run”; a step’s findings list could also render after a later step’s line, or above its own step’s still-typing line, instead of directly under it.
• Fixed: run-mode confirmation dialogs (Console and REST) showed raw mode slugs like “semi-auto” instead of the user-facing label (Test/Smart/Autonomous); Classic Runner’s post-ID picker dropdown stayed open after clicking elsewhere and rendered behind other buttons due to a stale z-index.
• Fixed: a rejected OpenRouter key’s failure message could scroll out of view the moment focus jumped to the key field — it now also shows inline next to the field; saving a key now immediately runs the real verification check instead of always showing “Active” with no validation.
• Fixed: Database Backups “Prune Now” silently did nothing when the “Keep for (days)” / “Keep max (count)” fields were edited but not yet saved — it now reads the live field values directly.
• Fixed: pressing Enter on a numeric Post ID in the Runner Console’s target field picked the target using the raw “#123” as its label forever instead of looking up the real title — Enter now resolves it the same way the search dropdown already does, with a brief “Looking up #123…” placeholder.
• Fixed: picking a new Default AI Provider and clicking “Save Settings” could silently fail to apply with no indication anything was wrong — the Settings page now actually shows that warning.
• Fixed: the agent name in the Runner Console’s title bar rendered invisible (dark text on the dark panel); the “Run on multiple” scope popup had no outside-click dismissal; the final report could visibly appear before an earlier step’s text had finished its on-screen typing animation.
• Fixed: the post ID / title search dropdown never closed when clicking elsewhere and could render underneath other console elements; it also showed nothing while the lookup was in flight — now shows a “Searching…” placeholder.
• Fixed: the Runner Console’s “Copy transcript” / “Download transcript” export squashed the final report bubble’s label, agent name, and outcome text together with no spacing; the button always said “Copy transcript” even where the browser’s Clipboard API isn’t available (needs https or localhost) — it now shows “Download transcript” up front with a tooltip explaining why, in that case.
• Fixed: the Runner Console’s mode-selector dropdown (Test/Smart/Autonomous) rendered with a light background inside the Console’s dark theme, since it was appended outside the Console’s own themed container.
• Fixed: Site Health Reporter’s drift report computed plugins reactivated since the last snapshot but never showed the count — added a “Plugins Activated” report row and folded the count into the summary line.
• Changed: Runner Console header now reads “AI Agent for {agent name}” and “Agentic AI Console V 1.1.8” instead of the generic “{agent} — Runner” / “AI Agent Runner Console”; added a short clarifying note on the WordPress Core AI Client card explaining it uses its own WordPress-managed credentials, separate from this plugin’s own Anthropic/OpenAI/Google keys.
• Changed: the Post ID field and the title-search field are now one combined field (“Search by title or Post ID…”) instead of two separate inputs; action/Run buttons in the Runner Console now always use the plugin’s own brand color instead of tracking the site’s WordPress admin color scheme; the Runner Console’s outer area is light again, with the console panel itself staying dark and a clear white border separating the two.
• Security: a failed restore statement’s diagnostic-log entry could include real row data (post content, user emails) via the raw SQL statement and MySQL’s error text; the log entry now records only the statement type, table name, and error code.

1.1.2

Agent reliability fixes: Site Health Reporter and the Core AI Client provider.

• Fixed: the Site Health Reporter agent could abort its entire report when it couldn’t capture WordPress admin notices outside a real wp-admin page load — which is every normal run through the Runner. It now degrades gracefully (skips that section) instead of failing the whole run.
• Fixed: the Site Health Reporter’s “Version Changes” report section still showed the retired “Stealth Plugin Updates” label even though the wording had already been neutralized everywhere else.
• Fixed: saving “WordPress Core AI Client” as the Default AI Provider always failed with “not enabled and credentialed yet,” even when its connectors were properly configured under Settings → Connectors — the check referenced a WordPress core class name that was never correct.
• Hardened: a handful of $wpdb prepared-statement comments that weren’t actually suppressing the warnings they were written for (code-quality only, no functional change).

1.1.1

Small fixes and cleanup following the Runner Console UI release.

• Fixed: the Runner Console’s mode-selector menu, quick-help popover, and share menu each reimplemented the same viewport-clamping and outside-click-dismiss logic independently; consolidated into one shared implementation to prevent drift between them.
• Fixed: Reset / Run Again could leave stale queued transcript lines from the previous run draining into the new one; the queue is now purged on reset/reopen.
• Fixed: transcript bubbles in the Runner Console used a hardcoded dark background that stayed dark regardless of light/dark mode; now follows the active theme.
• Hidden: the Settings → “Runner UI” (Classic/Console) toggle — Console is now the sole Runner UI, and the Classic/Console selector has been retired.
• Note: if dashboard styling still looks stale after updating, a hard refresh clears it — some hosts/CDNs cache assets by version number.

1.1.0

Runner Console UI reaches full parity with the classic run modal, plus clearer failure diagnostics.

• The Runner Console (the newer, non-modal way to run an agent) now matches the classic run experience feature-for-feature: typewriter-style step output, itemized findings, 5-star run rating, rich share menu (branded image + social links) with copy-to-clipboard fallback, transcript copy/download/export, per-step issue flagging with a compiled report, backup-gated rollback, bulk post-type scoping with a live count preview, and light/dark theme parity.
• Fixed: the Alt Text Architect agent no longer burns an AI call and a consent prompt when every image on the page already has alt text (regression from an earlier refactor; now restored).
• Fixed: an approval-needed step now shows the actual AI-generated suggestions being approved (e.g. per-image alt-text suggestions, generated FAQs, meta description candidates) instead of a generic “review and confirm” prompt with nothing to review.
• Fixed: the “Run on multiple” post-type scope popup now closes properly (toggle + explicit cancel) instead of staying open over the command bar.
• Fixed: the mode-selector and the quick-help popover now stay within the visible window instead of occasionally rendering off-screen.
• Fixed: the Post ID field now offers the same live title-search suggestions as the title field, instead of requiring an exact ID typed blind.
• Fixed: the dashboard no longer gets auto-repainted black by some browsers’ OS-driven dark-mode detection.
• Improved: failed runs and apply steps now surface a clearer, more specific error message instead of a generic failure notice.
• Improved: the quick-help popover now covers the console’s main controls (target search, bulk scope, Run, approval gate), not just the Test/Smart mode explainer, and its icon moved next to the close button.
• Several smaller Runner Console fixes: correct target post-type list in target search, share menu now renders in the correct light/dark theme, mode-explainer default text no longer hardcoded, dark-console number-input styling, and a notice now appears when a saved AI provider key fails to decrypt.
• Housekeeping: internal version-tracking cleanup — this release supersedes unpublished internal builds that were briefly labeled 1.0.2–1.0.4 during development; no separate upgrade action is needed from 1.0.1.

1.0.1

Runner mode simplification, privacy disclosure correction, security hardening, and maintenance.

• Renamed the Runner’s run modes for clarity: Manual -> Test (runs each step for real, one at a time, so you can check the result before continuing — not a dry run), Semi-Auto -> Smart (always previews first, then a single Confirm or Cancel).
• Removed the run-history unlock requirement — Test and Smart are both available immediately on every agent, no “earn trust first” gate.
• Fixed: an agent’s “needs a specific post” requirement is now always computed from the agent’s own definition on the server, instead of trusting a button’s client-side guess — closes a bug where running an agent from the Library catalog’s installed-agent Run button could skip the required post picker and fail with “post_id is required.”
• Successfully completing a Test run now automatically marks an agent “Autonomous-ready” (the existing “mark safe” trust flag, renamed) for future automation surfaces.
• Added a way to flag specific steps in a Test run’s transcript and submit a combined report for agents installed from the Community Library.
• Fixed: removed a duplicate Run entry point in the Library catalog that opened a different, incomplete welcome screen — Automation > Agents is now the only Run path.
• Fixed: the issue-flag icon was missing from the consent-gate transcript card, and flagging a step no longer force-scrolls the transcript to the bottom.
• Fixed: a marked-safe agent now goes through the normal welcome screen with an inline “Auto Safe” badge (+ Undo) instead of a separate interstitial screen.
• Fixed: root-caused and removed a duplicate approval prompt that could ask for the same consent/warning gate twice in a single run.
• Fixed: the Alt Text Architect agent no longer burns an AI call and a consent prompt when every image on the page already has alt text — it now short-circuits straight to the report.
• Corrected the Privacy and External Services sections to accurately disclose the automatic daily Hub metrics sync and per-run outcome reporting that occur once a Hub account is connected (previously undisclosed as “automatic”; the behavior itself is unchanged and was already gated on an explicitly connected account).
• Security: hardened AI-prompt templating so a prior step’s AI-generated or externally-fetched output is framed as untrusted data rather than spliced unmarked into the surrounding prompt (prompt-injection hardening).
• Security: validated caller-supplied table names in the database backup module against the live table list before use.
• Fixed a stale uninstall cleanup list that missed several options and one Action Scheduler hook added since v1.0.0.
• Performance: skip the full database-schema existence check on every admin page load once the schema has been verified for the current version.

1.0.0

First stable release of Zybroz AI Agents for WordPress.

Turn your WordPress dashboard into an AI automation platform. Build reusable agents, run multi-step workflows, and automate SEO, content, accessibility, internal linking, and site maintenance—with tools to preview changes, create backups, and roll back supported changes when needed.

Agent creation and management

• Visual Agent Builder with live JSON preview, AI Prompt Generator, and guided Chat-to-Create wizard.
• Reusable agents with versioning, import/export, and support for JSON configurations and plain-text prompt templates.
• Quality Precheck to help validate agents before execution.

AI engines and providers

• Multi-model AI through OpenRouter, with a curated fallback chain that prioritizes available free models.
• Local Ollama support for running compatible models locally.
• Optional direct integrations with Anthropic, OpenAI, Google, and WordPress Core AI Client.
• InbuiltEngine for logic-based WordPress administration and maintenance tasks without requiring an AI API key.

Agents, workflows, and library

• 10 bundled starter agents for accessibility scanning, alt text, meta descriptions, FAQ generation, link health, media auditing, site health, SEO plugin auditing, orphan pages, and outdated content.
• Visual Workflows for connecting agents, actions, and conditions into multi-step automations.
• Community Library integration with the Zybroz Hub to browse, install, and publish agents, with features such as ratings, favorites, and update notifications.
• AI Action Library for reusable capabilities.
• WooCommerce product-description support.

Monitoring and productivity

• Run History and Metrics for tracking agent execution.
• Dashboard for accessing agent activity and tools.
• Advanced Backup Module and supported rollback functionality.
• Achievements and share-card images.

Safety, security, and privacy

• Dry-run previews and automatic post snapshots for supported write operations, with rollback support.
• Consent gates for write-capable steps, requiring approval before execution in Manual, Semi-Auto, and Auto modes.
• Encrypted API-key storage, with keys protected from browser-side exposure.
• SSRF protections for outbound requests to the Zybroz Hub, OpenRouter, and supported AI providers.

Includes a one-time automatic migration of internal database tables and identifiers to the zybaiag_ / zybroz- naming conventions. See the Upgrade Notice before upgrading.

0.9.392

Major update consolidating all development since 0.9.391 into one WordPress.org release.

• Migration: the original direct Anthropic/OpenAI/Gemini key fields (from 0.9.391 and earlier) are retired in favor of OpenRouter as the default external AI path. If this site has one of those original keys stored, a one-time “Migrate your AI provider keys” screen shows it once — read-only, copy-friendly, never transmitted anywhere — with instructions to move it to OpenRouter, then removes it from storage once you finish or skip. InbuiltEngine (non-AI) actions are unaffected and need no key.
• Security: write-capable agent steps that require human consent now actually halt in every run mode (Manual, Semi-Auto, and Auto) until a person approves — previously only Manual mode paused correctly. Each approval is bound to the exact run, step, and content shown, is time-limited and one-time-use, and every crossing is recorded in a new audit log.
• New: Workflows — build multi-step automations chaining agents, actions, and conditions, with a visual builder and run history. (0.9.391 shipped a “Coming soon” placeholder for this; it is now fully built.)
• New: Community Library — browse, install, and publish agents to/from the Zybroz Hub, including ratings, favorites, and update notifications for Hub-sourced items.
• New: a separate, newly-added “More AI Providers” panel lets you optionally configure a fresh direct key for Anthropic, OpenAI, Google, or WordPress’s own Core AI Client, as an alternative to routing through OpenRouter. This is a different, explicitly opt-in feature from the retired keys above — OpenRouter remains the default external AI path, and nothing is sent to these providers unless you configure a key for them here.
• Dozens of reliability, pagination, dark-mode, i18n, and security-hardening fixes across Settings, Library, Dashboard, Run History, and the AI Action library.

0.9.391

• New: Settings now has a collapsible “Data & Privacy” panel that restates, in plain language, what the plugin sends to AI providers and to the optional Central Library, and when — a summary of the “External Services” section of this readme.
• Change: The Workflows section now shows a “Coming soon” placeholder. Multi-step workflows are still in development; the menu entry, stored workflow data and REST routes are unchanged and nothing is lost. Run agents individually from the Agents page in the meantime.
• Housekeeping: Version bump on top of 0.9.390 with no database or option-schema changes — upgrading from 0.9.390 is a no-op migration.

0.9.390

• Added: Initial WordPress.org release — approved baseline v0.9.390