SophMate 1.2.0 Release
New Copilot files, web search tools, new Gemini provider, and fixes.
Added
- Copilot Files — Added a Pro-only, first-class file workspace for private AI-created and uploaded files, with encrypted storage, immutable revisions, provenance, scoped access grants, metadata and tag management, search and filtering, archive/trash lifecycle controls, authenticated previews and downloads, responsive desktop/mobile layouts, and accessible Markdown, CSV, JSON, TXT, PDF, and image rendering.
- Added Copilot file tools for listing, searching, inspecting, and reading accessible files; privately creating Markdown, TXT, JSON, CSV, and PDF artifacts; and approval-governed metadata, sharing, status, and revision-restore operations. File contents are treated as untrusted input, reads are bounded, and encrypted keyword indexing avoids plaintext search indexes.
- Added end-to-end chat integration for Copilot Files, including a searchable composer picker, combined five-file attachment limits, persisted attachment references, streamed and reload-safe artifact cards, authenticated preview/download actions, and cross-conversation file discovery instead of relying on conversation memory.
- Added an opt-in, Pro-only SophMate MCP Server for securely connecting Claude Code and other remote MCP clients to site-scoped SophMate tools, with Streamable HTTP, OAuth 2.1 and expiring personal tokens, connection-specific tool allowlists and permissions, read-only and all-tools snapshot presets, rate limits, audit/privacy controls, action-plan-governed write operations, and a separately confirmed administrator-only mode for always executing while bypassing human approval policy.
- Added current OpenAI GPT-5.6 variants and Claude Sonnet 5 to the Copilot model catalog; the current Gemini 3.5 Flash, 3.1 Pro, and 3.1 Flash-Lite family remains available.
- Added a read-only
sophmate.web.fetchAI tool for direct access to public HTTPS pages and textual APIs, with readable extraction, long-content chunking, link discovery, SSRF protection, bounded responses, and explicit untrusted-content labeling. - Added a provider-neutral
sophmate.web.searchAI tool backed by native OpenAI Responses or Anthropic server search, with ranked citations, domain/locale/recency controls, active encrypted-credential reuse, cache and rate controls, budget accounting, and a search-then-fetch workflow. - Added Google Gemini as a first-class LLM provider, with end-to-end support for provider setup, native streaming chat, structured output, tool calling, multimodal input, managed Agents, model selection, connection tests, and WordPress 7.0 Google connector credentials.
- Added a WordPress-first Site mode for installations without an active, supported WooCommerce version, including a shared capability profile, dependency-aware tools and automations, goal-based onboarding, WordPress-native Quick Actions, Site Insights, Site Brief onboarding, site-care watchers, and compatibility metadata for App Center apps.
Changed
- Made the
sophmate.web.fetchAI-facing description explicitly instruct models to call it when users ask to fetch, open, read, inspect, or summarize a supplied website URL instead of searching for that URL. - Adapted Home, Status, Copilot, Insights, Quick Actions, Playbooks, Agents, Workflows, Watchers, App Center, Theme Assistant, Marketing, Personalization, and Settings to present WordPress features first and expose commerce only when WooCommerce is ready or the user intentionally looks for it.
- Replaced missing optional-commerce error states with neutral, accessible dependency guidance and native WordPress plugin-management actions, while preserving blocked commerce configurations for automatic recovery after WooCommerce is reactivated.
- Made WP-Cron a supported Site-mode background backend and reserved Action Scheduler requirements for workloads that genuinely need it.
Fixed
- Added provider-specific compatibility for the newest models: GPT-5.6 uses
max_completion_tokenswhile legacy-compatible and local models retainmax_tokens, Claude 5 models omit unsupported sampling parameters and expose classifier refusals as content filtering, Gemini 3.5 omits legacy sampling parameters, and shut-down Gemini model IDs migrate to supported successors. Filtered responses now discard partial text and tool calls across stored, cached, and live-streamed output. - Hardened web fetch and search against recognizable credentials (including recursively encoded, nested, and parser-ambiguous query parameters), unsafe or looping redirects, multiplied request timeouts, known or heuristic binary payloads behind misleading content types, pathological HTML DOM breadth, cross-user/provider cache reuse and provider-configuration races, oversized structured or single-field responses and paused-turn replay state, omitted or failed-request usage accounting, mixed-answer refusals, incomplete provider responses, and Anthropic server-tool or paused-continuation failures returned with HTTP 200.
- Corrected Gemini thinking-token accounting, safety-block handling, malformed function-call rejection, thought-signature preservation, Google API-key error classification, JSON Schema compatibility, and resolved-endpoint connection testing.
- Prevented Site-mode pages, Copilot tool catalogs, slash commands, schedulers, hooks, direct deep links, and app runtimes from advertising or calling unavailable WooCommerce functionality.
- Corrected Site-mode readiness and attention counts so an absent optional WooCommerce integration does not make an otherwise healthy SophMate installation appear broken.