File attachments, multi-model support, and resilient job infrastructure

Features

• File attachments in agent chat — Added a unified file attach button that routes images to the vision pipeline and all other file types to the VM, with production hardening for edge cases.
• Multi-model provider support — Agents can now be configured with multiple model providers. Includes a Claude Code CLI integration using short aliases (sonnet, opus, haiku) and a checklist UI for managing allowed and default models.
• Discord/Slack: pass() tool and reprompt loop — Introduced a pass() tool for Discord and Slack integrations, along with a reprompt loop and fixes to message threading behavior.
• Prompt cache token metrics — Inference events now surface prompt cache token counts, giving operators visibility into caching efficiency.
• Cloud provider refresh UI — Added a cloud provider refresh action and an allowed-models checklist to the provider settings UI.
• Resilient job recovery and client-abandon cancellation — The job queue now recovers stalled jobs and cancels work when a client abandons a session.
• OpenUI datasource cache and coordinator API routing — OpenUI gained a datasource cache layer, scoped app/dashboard views, and proper routing through the coordinator API.

Improvements

• User profiles now include first and last name fields; applied a UI refresh alongside updated design system documentation.
• Pre-inference queue depth is now visible in the chat interface, and the job poller is capped by per-consumer PEL size to prevent runaway polling.
• Allowed and default model lists are now persisted correctly for cloud providers, with future-proof model list handling.

Fixes

• Hardened cloud inference against hangs and connection races during LLM and Redis operations.
• Fixed GPU job lifecycle to prevent stuck PEL entries and infinite retry loops on failed jobs.
• Tool loop errors now hard-fail instead of silently continuing; all ACKs now include an XDel to prevent message redelivery.
• OAuth public base URL is now correctly derived from proxy headers when the server runs behind an ingress.
• File input filter was broadened from images-only to accept all file types.

Internal

• Added a make environment target to streamline worktree and CI setup.
• Fixed stub interface drift across Discord, Slack, and WebSocket packages in the test suite.