Batch page selection
Paste titles, load from a category, or find pages that transclude a template (optional prefix filter). Validation shows editable pages and skips problems early.
Features
Built for production wikis: server-side LLM calls, permissions, rate limits, and audit logging.
Six AI operations
Wikilinks, spellcheck, formatting, writing style, templates (with remote reference fetch), and custom instructions.
Server-side batch
The browser starts a batch and polls for progress. Cancel mid-run to stop remaining pages. Configurable concurrency per poll.
Diff & article preview
Lazy-loaded wikitext diffs plus a Preview article button that opens the rendered proposed page in a new tab before save.
Approve before save
Approve per page or in bulk. Saves use your edit summary, draft tokens, and the aibatcheditor change tag.
Review safeguards
Proposal warnings, diff-reviewed confirmations, post-save revision links, rate limits, and Monolog audit channel.
Compatibility
Requires MediaWiki 1.43+ (Vue and Codex in core). Tested on 1.43 LTS through 1.45+.
| Environment | Supported? | Notes |
|---|---|---|
| Self-hosted (Docker, VPS, local) | Yes | Admin installs extension and configures LLM |
| Private / corporate wikis | Yes | Same requirements |
| Wiki farms (Miraheze, ShoutWiki) | Maybe | Custom extensions + outbound HTTPS required |
| Wikimedia wikis (Wikipedia, Commons) | No | Third-party; not in WMF deployment |
Processes only existing, non-redirect wikitext pages the user can edit. PHP 8.1+, JavaScript, and outbound HTTPS to the LLM API are required.
AI capabilities and limits
The extension edits wikitext — it is not a web search or research agent.
Can do
- Edit one page's wikitext per LLM call
- Follow batch-wide and per-page instructions
- Clone templates from allowed remote wikis (e.g. es.wikipedia.org)
- Preview prompts before drafting (optional debug mode)
- Surgical-edit prompts (v3): operation scope + profile intensity, no invented facts
Cannot do
- Search the web or call external APIs (weather, news, etc.)
- Read other pages on your local wiki as context
- Share context between pages in a batch
- Edit non-wikitext content or redirects
Prompts follow a surgical-edit contract: smallest necessary change, copy everything else unchanged, return input as-is when already satisfied. Provide external facts in your instructions when needed.
Workflow
- Validate pages Title list, category, or template (+ optional prefix). Oversized or non-editable pages are flagged.
- Choose operation & profile Pick the AI task and intensity (conservative / balanced / aggressive). Custom hides the profile dropdown and uses balanced intensity. Add batch-wide instructions separate from the wiki edit summary.
-
Preview prompt
Optional: inspect system/user messages before drafting (
$wgAIBatchEditorPromptPreview). - Draft Server-side batch with status polling. Unchanged pages show as omitted.
- Review diffs Lazy-loaded diffs or Preview article in a new tab; per-page overrides, re-draft, or retry all failed pages at once.
- Approve & save Confirmations for unreviewed diffs and risky proposals. Saves require draft tokens.
Safety and review gates
| Feature | Behavior |
|---|---|
| Privacy notice | Shown when the LLM is configured |
| Rate-limit notice | Draft disabled when batch exceeds remaining hourly quota |
| Lazy diffs | Diffs load on demand, not automatically |
| Proposal warnings | Flags major deletions, near-empty output, large growth, removed headings |
| Diff-reviewed gate | Confirms approve/save if diffs were not previewed |
| Draft tokens | HMAC tokens bind proposals to user, revision, and content hash |
| Post-save links | Revision and history links after successful save |
Installation
Requires MediaWiki 1.43+ and a Grok-compatible chat completions endpoint (xAI by default).
git clone https://github.com/urimacias/AIBatchEditor/ extensions/AIBatchEditorAdd to LocalSettings.php:
wfLoadExtension( 'AIBatchEditor' );
$wgAIBatchEditorApiUrl = 'https://api.x.ai/v1/chat/completions';
// XAI_API_KEY in $IP/.env — loaded automatically; do not commit the keySensible defaults are built in (model grok-4.3, 420 s LLM timeout, batch cap 100, 500 requests/hour, concurrency 1). Set $wgAIBatchEditorConcurrency = 2 for parallel LLM calls; set $wgAIBatchEditorReasoningEffort = 'low' for faster grok-4.5. Batch progress uses CACHE_DB automatically. Store the API key in $IP/.env or Docker .env, never in git.
Users need the aibatchedit right (granted to sysop by default).
Shared hosting / cPanel
Batch progress is stored in CACHE_DB (mw_objectcache) by the extension. Optional APCu for general wiki speed:
$wgMainCacheType = CACHE_ACCEL;
$wgSessionCacheType = CACHE_DB;Configuration
| Setting | Default | Purpose |
|---|---|---|
$wgAIBatchEditorApiUrl | '' | LLM endpoint (server-side only) |
$wgAIBatchEditorApiKey | '' | API key / bearer token |
$wgAIBatchEditorModel | grok-4.3 | Model identifier |
$wgAIBatchEditorMaxBatch | 100 | Max pages per batch |
$wgAIBatchEditorMaxPageSize | 2097152 | Max wikitext bytes per page, 2 MiB (0 = no limit) |
$wgAIBatchEditorMaxInstructionsLength | 8192 | Max AI instruction bytes |
$wgAIBatchEditorRequestTimeout | 420 | LLM HTTP timeout (seconds) |
$wgAIBatchEditorTemperature | 0.1 | LLM temperature; lower = stricter instruction following |
$wgAIBatchEditorPromptPreview | false | Debug flag: expose prompts in UI/API (enable only for troubleshooting) |
$wgAIBatchEditorRateLimitPerHour | 500 | AI requests per user per hour |
$wgAIBatchEditorPollIntervalMs | 2500 | Client interval between batch status polls |
$wgAIBatchEditorConcurrency | 1 | Pages per advance (LLM in parallel when > 1) |
$wgAIBatchEditorReasoningEffort | '' | xAI effort: low / medium / high (empty = omit) |
$wgAIBatchEditorStubMode | false | Deterministic AI stub for automated browser tests only |
$wgAIBatchEditorTemplateSourceWiki | https://es.wikipedia.org | Default remote wiki for template references |
$wgAIBatchEditorTemplateSourceAllowHosts | see extension.json | Allowed hosts for template source overrides |
$wgAIBatchEditorOperationProfiles | see extension.json | Per-operation profile intensity (UI help + LLM prompt) |
$wgAIBatchEditorSystemPrompt | [] | Wiki context after ROLE (WIKI CONTEXT); prefer identity; op checklists need “UNICAMENTE con la operación equivalente” |
$wgAIBatchEditorSystemPromptAppend | [] | Wiki-wide extra system-prompt bullets at end (append-only; operation task takes precedence) |
API modules
| Module | Mode | Purpose |
|---|---|---|
aibatcheditorlist | read | Validate and list pages; rate-limit status |
aibatcheditorpreview | read | Build LLM prompts without calling the AI |
aibatcheditorbatchstart | read | Start server-side batch; returns batchId |
aibatcheditorbatchadvance | read | Process next batch chunk (LLM work) |
aibatcheditorbatchstatus | read | Read batch progress (fast; no LLM) |
aibatcheditorbatchcancel | read | Cancel a running batch; clears pending pages |
aibatcheditordiff | read | Render preview diff |
aibatcheditorarticlepreview | read | Stage proposed wikitext; return article preview URL |
aibatcheditorrefreshdrafttokens | write | Refresh draftToken values before save |
aibatcheditorsave | write | Save approved edits (requires draftToken per edit) |
The UI uses batchadvance for LLM work and lightweight batchstatus polls for progress. Wikitext and draftToken come from advance responses only. Save recovers stale tokens server-side; aibatcheditorrefreshdrafttokens is available for API integrations.
Development
127 PHPUnit tests (62 unit + 65 integration) + Playwright E2E. Run from the MediaWiki root with dev dependencies installed.
composer install --dev
./extensions/AIBatchEditor/tests/run-phpunit.sh
# E2E (Node.js + sysop credentials + stub mode)
export MW_E2E_USER=Admin
export MW_E2E_PASSWORD='your-password'
export AIBATCHEDITOR_E2E_STUB=1
./extensions/AIBatchEditor/tests/run-e2e.shSee tests/QA-CHECKLIST.md in the repository for manual release verification.