Agent Zero Release Notes

Highlights

This release introduces a redesigned first-run onboarding experience, separates Desktop and Office into focused plugins, adds host browser routing through the A0 CLI connector, and delivers significant prompt and tool contract improvements informed by cross-provider evidence sweeps.

New Features

• Guided first-run onboarding — A new wizard walks users through Cloud vs. Local provider selection, account connection, model picking, and a ready state, replacing the previous manual setup flow.
• Ollama Cloud provider — Available as an OpenAI-compatible chat provider with hosted API base and model listing.
• Nebius Token Factory provider — Registered as an OpenAI-compatible provider, available in onboarding and model config.
• Host Browser (Bring Your Own Browser) — Route browser actions through the A0 CLI connector to a local Chrome/Chromium instance with remote-debugging support, including profile mode selection (existing profile or clean Agent profile).
• Browser runtime selection exposed to CLI — A protected connector endpoint allows the A0 CLI to switch between Docker browser and Bring Your Own Browser mode.
• Persistent browser history screenshots — Each Browser tool call saves a static JPEG to chat history for immutable transcript previews.
• ZIP download toast feedback — Immediate grouped progress toasts for backup and file-browser ZIP downloads with error replacement.
• API chat lifetime and cleanup — lifetime_hours is validated and persisted per chat; a background job loop removes expired API chats automatically.
• Connector model overrides in Web UI — Custom per-chat model overrides now render in the model switcher instead of a generic "Custom" label, with live refresh on changes.
• chats-header-controls extension hook — Plugins can inject controls into the chats list header via the established x-extension pattern.

Architecture & Plugin Changes

• Desktop split into _desktop plugin — Xpra/Xfce lifecycle, /desktop route, Desktop state, session APIs, surface registration, and the linux-desktop skill now live in _desktop.
• Office reduced to document ownership — _office focuses on document artifacts, Markdown sessions, LibreOffice-compatible file actions, and persistence. Binary document editing routes through explicit Desktop requests.
• Shared surfaces frontend service — Browser and Desktop register docked or floating live UI through a dedicated surfaces layer, preserving active viewers across canvas/modal switches.
• Office and Desktop state moved under plugin storage — Retired /usr/_office and /usr/_desktop trees migrate into /usr/plugins/<plugin> at startup.
• Serialized runtime package preparation — Office and Desktop apt operations share an in-process retry guard to handle transient apt locks gracefully.
• Desktop install progress — The Desktop frontend shows a neutral install message during runtime preparation instead of a misleading missing-dependencies error.

Tool & Prompt Improvements

• Normalized tool action contracts — Multi-action tools standardized around tool_args.action while maintaining backward compatibility with older dialects.
• Renamed skills to task-oriented names — Office skills become Writer, Calc, Impress; high-impact skills get task-oriented names; plugin-owned skills move into their owning plugin folders.
• Reduced default prompt pressure — Browser and scheduler prompts compacted into skill-backed manifests, skill catalog descriptions shortened, noisy framework knowledge pruned.
• A0 connector remote skills split by affordance — Separate text-editor and code-execution remote skills replace the combined CLI remote workflow skill; computer-use remote scoped to desktop control and skill-gated as beta.
• Browser tool ergonomics — Selector-based reference actions, coordinate click fallbacks, focused-field typing, string key chords, and smarter content helper snapshots that ignore framework event wrappers.
• Document artifact stabilization — File creation is opt-in through document_artifact, file cards appear only in final responses, and Desktop canvas sessions refresh after saved edits.

Security Fixes

• Path-containment bypass in /api/image_get resolved — Requested images are now resolved against the Agent Zero base directory with symlink-aware validation. SVG/SVGZ responses include nosniff and sandboxed CSP headers.

Bug Fixes

• Blocking history compression edge cases — Stalled automatic compression no longer spins forever; large single-line histories (85k+ characters) split by verified token budget.
• Canvas attachment for browser and documents — Browser canvas attaches to active Docker sessions with initial snapshots; Markdown documents open through the right-canvas Desktop editor.
• Bare desktop canvas treated as ready — Linux Desktop reports healthy when XFCE has no active application window but display and screenshots are available.
• Computer-use rearm handling — Stale Free Run settings no longer treated as ready; agents provide backend-agnostic rearm guidance.
• Subordinate agent document affordances — Automatic office document response affordances skipped for delegated agents so reviews keep their actual content.
• Model config provider controls — Custom API base URL resets on provider change; Supports Vision toggle moved out of Advanced Settings.

Infrastructure & Maintenance

• Browser Playwright cache moved to /a0/tmp/playwright — Startup migration cleans retired usr-based caches automatically.
• uv cache cleaned during self-update — Best-effort uv cache clean runs when the self-update manager processes an update.
• Legacy Collabora retirement — Old a0_office_collabora supervisor program stopped and removed during Docker self-update; stale runtime paths cleaned.
• Legacy runtime artifact cleanup — Stale directories removed, retired packages (including dynamically discovered collaboraoffice* splits) purged at startup/self-update.
• Self-update backups skip runtime sockets — Live sockets, device nodes, vanished files, and unreadable entries no longer trip rollback checks.
• Connector hello metadata — Remote tool metadata stored through a dedicated hello path with context-bound websocket SID acknowledgment.
• Skill ownership hardened — Connector list/delete scoped to enabled catalog; built-in plugin skills cannot be deleted; invalid skill YAML surfaced instead of silently accepted.

UI Improvements

• Simplified project LLM setup — Project creation focuses on basic details and Git cloning; model selection moved to a collapsed Advanced Settings accordion on the edit screen.
• Settings accordion redesign — Sticky accordion with active section tracking replaces two-step settings navigation.
• Action button order aligned — Primary actions appear before Cancel across plugin settings and related modals.
• Browser chat action button removed — The Browser plugin no longer injects a button under the chat input.
• Project skills settings repositioned — Moved directly after project instructions, before file structure and secrets.
• Host Browser config simplified — "Use host when ready" replaced with clear "Bring Your Own Browser" option and Chrome remote-debugging guidance.

Documentation

• Refreshed user-facing guides — README, quickstart, and docs index updated around Browser, Desktop, A0 CLI, projects, memory, skills, profiles, and model presets with optimized screenshots.
• First-run onboarding guide — Screenshot-led walkthrough linked from docs hub, quickstart, and installation flow.
• Model efficiency scorecard — Scored tool-efficiency chart with provider/model rankings, failure clusters, and applied improvements.
• Refactor guardrails — Internal docs covering modal/surface boundary, Desktop ownership, Office scope, renamed skills, and connector ownership rules.