Squad

Changelog

[1.9.1] — 2026-07-31

• Reduced the default provider timeout to 20 seconds and the connection timeout to 5 seconds.
• Capped provider calls from web workers at 20 seconds so PHP can return before the web server’s FastCGI timeout; bounded CLI workers may still request a longer timeout explicitly.

[1.9.0] — 2026-07-24

• Added provider-independent webSearch and webSearchMaxResults request options for ask() and stream().
• Added OpenRouter web-plugin support for every routed model, native Anthropic web-search tools, OpenAI/xAI Responses API search and native Google Search grounding.
• Normalized web citations into a provider-independent sources result array.
• Included web-search settings in response cache keys and rejected unsupported direct adapters explicitly instead of silently returning an unsearched answer.

[1.8.0] — 2026-07-23

• Added provider-independent stream() support for Anthropic and all OpenAI-compatible providers, including OpenRouter.
• Streaming callbacks receive text deltas immediately while the returned result retains the complete content, usage, provider, and model metadata.
• Streaming requests bypass the response cache so callers never receive a fake post-completion typing effect.

[1.7.0] — 2026-07-13

• Added native AdminThemeUikit/Konkat settings UI support based on pw-design-system patterns.
• Reworked provider key management with theme tokens, responsive layouts, provider search and status filters, accessible disclosure controls, key visibility toggles, and clearer save/test states.
• Updated Test Chat and cache management panels to use native UIkit controls and theme-aware feedback states.

[1.6.1] — 2026-07-12

• Vision inputs now enforce 12,000 px per-side and 32 MP decoded-dimension caps in addition to byte limits.

[1.6.0] — 2026-07-12

• Refreshed the OpenRouter fallback catalog from the public OpenRouter model API, using openrouter/auto as the resilient default plus current provider aliases and named models.
• Added OpenRouter’s OpenAI-compatible embeddings endpoint and a fallback embedding-model catalog, allowing Atlas to use an existing OpenRouter key.
• Added a provider-independent vision() API for bounded local PNG/JPEG/WebP/GIF inputs, with OpenAI-compatible and Anthropic multimodal request formats.

All notable changes to the Squad module (formerly AiWire) are documented in this file.

Format follows Keep a Changelog. Versions follow Semantic Versioning.

---

[1.5.1] — 2026-06-28

Fixed

• Removed deprecated curl_close() calls (no-op since PHP 8.0; emitted a deprecation notice on PHP 8.5).

---

[1.5.0] — 2026-06-25

The “AiWire → Squad” release: a rename plus a major capability expansion.

Changed

• Renamed the module from AiWire to Squad — class, files, key table, admin UI, log channel. Existing installs migrate automatically on install: the key table is renamed (aiwire_keys → squad_keys) and settings are carried over, and the encryption KDF context is preserved so stored keys keep working.
• Encrypted key storage — provider keys moved out of the module config JSON into a dedicated squad_keys table, encrypted with libsodium secretbox (secret derived from a config.php salt). A database dump no longer exposes keys. env:NAME references still supported.
• Adaptive-model aware requests — temperature/sampling params are now omitted automatically for models that reject them (Claude Opus 4.7/4.8, Fable/Mythos), preventing 400 errors.
• Model catalogue refreshed — Anthropic set updated to current IDs (added Claude Opus 4.8 as default and Claude Fable 5; bare aliases). OpenAI/Google/xAI/OpenRouter left current.

Added

• embed() — provider-independent embeddings for a string or a batch (OpenAI, Google, Qwen, Zhipu); getDefaultEmbedProvider().
• image() — text-to-image generation (xAI Grok Imagine, OpenAI gpt-image-1 / DALL·E 3); getDefaultImageProvider().
• run() — multi-step tool-use / agent loop supporting both OpenAI and Anthropic tool-calling formats.
• Direct Chinese providers — DeepSeek, Qwen (Alibaba DashScope), Kimi (Moonshot), GLM (Zhipu), MiniMax, Yi (01.AI), Doubao (ByteDance/Volcengine), Ernie (Baidu Qianfan), Hunyuan (Tencent), all via the OpenAI-compatible path (14 providers total).
• Anthropic prompt caching for large system prompts (automatic for prompts ≥ ~3 KB).

---

[1.4.0] — 2026-05-15

Added

• Added dynamic model refresh support for OpenAI and OpenRouter from the admin provider key UI.
• Added providerModels config storage for refreshed model lists, separate from static defaults and models.json.
• Added per-key custom_model support so users can enter private, preview, proxy, or newly released model IDs manually.
• Added public helpers getProviderModels() and refreshProviderModels().

Changed

• Model dropdowns now prefer refreshed provider models, then models.json, then built-in defaults.
• Runtime model resolution now prefers per-request model, then per-key custom_model, then selected key model, then provider default.

---

[1.3.0] — 2026-05-15

Added

• Added models.json as an editable provider model catalog so model IDs, labels, and defaults can be updated without changing module PHP code.
• Added support for API key references like env:OPENAI_API_KEY to avoid storing real secrets in ProcessWire module config.

Fixed

• Fixed a PHP parse error in the xAI provider definition that prevented the module from loading.
• Added CSRF validation to AiWire admin AJAX actions.
• Escaped JSON embedded in admin HTML and JavaScript to avoid broken markup and reduce XSS risk.
• Included model in cached provider instances so the same key can be used with different selected models safely.
• Restored a page’s previous output formatting state after saveTo().

---

[1.2.0] — 2026-04-23

Changed

• Anthropic models: added Claude Opus 4.7, Claude Sonnet 4.6; default changed to claude-sonnet-4-6-20260217; removed deprecated Claude Sonnet 4.5
• OpenAI models: added GPT-5.4 family (gpt-5.4, gpt-5.4-mini, gpt-5.4-nano); default changed to gpt-5.4; removed deprecated gpt-5-mini, gpt-5-nano
• Google models: added Gemini 3.1 Pro Preview, Gemini 3 Flash, Gemini 3.1 Flash Lite, Gemini 2.5 Flash; default changed to gemini-3-flash; removed deprecated gemini-flash-latest, gemini-flash-lite-latest, gemini-3-pro-preview
• xAI models: added Grok 4.20, Grok Code Fast 1; removed deprecated Grok 3 Mini
• OpenRouter models: added Amazon Nova Micro/Lite, ByteDance Seed 1.6, Xiaomi MiMo V2 Flash, Zhipu AI GLM 5; sorted by company A-Z; updated Anthropic/Google/OpenAI refs to latest; total 19 models from 13 companies

Improved

• Documentation split into README.md (overview) and DOCUMENTATION.md (full reference)
• All 25 examples now include Problem description, ProcessWire setup table, code, and Result output
• Added Table of Contents with anchor links in both README and DOCUMENTATION
• Added Result Format section explaining return arrays for all methods

---

[1.1.0] — 2026-02-19

Added

• generate() method — multi-block AI content generation with per-block settings (provider, model, temperature, systemPrompt, cache per block)
• Global options with per-block overrides: generate($page, [['field' => '...', 'prompt' => '...', 'options' => [...]]], $globalOptions)
• Each block checks field first (skip if content exists), calls AI only when needed
• Returns array keyed by field name with source: 'ai'|'field'|'error'
• 25 real-world usage examples based on lqrs.com (spirits/wine catalog)

---

[1.0.0] — 2026-02-11

Added

Core API

• chat() — simple text response, returns string
• ask() — full response with success, content, usage, raw, cached
• askWithFallback() — automatic fallback across keys and providers
• askMultiple() — same prompt to multiple providers for comparison
• askAndSave() — ask AI and save to page field (single or batch)
• saveTo() / loadFrom() — manual field storage

Providers

• 5 providers: Anthropic (Claude), OpenAI (GPT), Google (Gemini), xAI (Grok), OpenRouter (400+ models)
• Unified API across all providers via OpenAI-compatible Chat Completions endpoint
• Multiple API keys per provider with enable/disable toggle
• Default key selector per provider
• getProvider() — direct provider instance access
• getProvidersStatus() — status of all providers and keys

Cache

• File-based cache system (AiWireCache)
• TTL support: 'D' (day), 'W' (week), 'M' (month), 'Y' (year), custom like '2W', '3M'
• Page-scoped cache keys via pageId option
• clearCache($page) — clear cache for specific page
• clearAllCache() — clear entire cache
• cacheStats() — files count, total size
• Auto-cleanup of expired cache on LazyCron::everyHour

Field Storage

• Save AI content to any ProcessWire Textarea/Text field
• Skip generation if field already has content (unless overwrite: true)
• Quiet save mode (no PW hooks triggered)
• Batch mode: multiple fields from one prompt or field-to-prompt mapping

Admin Interface

• AJAX-powered admin UI with per-provider key management
• Connection test button for each key (one-click verify)
• Test Chat panel with parameter controls (provider, model, temperature, maxTokens, timeout)
• Real-time provider status display
• Cache management UI with stats and clear buttons

Options

• provider — select provider per call
• model — override model per call
• systemPrompt — system instructions
• maxTokens — response length limit
• temperature — creativity control (0.0–2.0)
• history — conversation history for multi-turn chat
• keyIndex — use specific key by index
• fallbackProviders — fallback chain
• cache — TTL-based caching
• pageId — page context for cache scoping
• timeout — request timeout
• overwrite — force regeneration
• quiet — save without triggering hooks

Logging

• Standard logging via ProcessWire wire('log')
• Debug logging (enable in module config)
• Logs: API calls, errors, cache hits/misses, field saves

---

*← Back to README.md

DOCUMENTATION.md*

This site is open source. Improve this page.