🌐 Remote Control & Connectors

Remote Link renamed to Remote Control — The tunnel provider system was refactored, with Cloudflare Tunnel, Microsoft Dev Tunnels, Serveo, and Tailscale available from the Remote Control selector.
Connector handshakes include Agent Zero version — CLI clients can now detect version mismatches with the connected Agent Zero server.
Installed plugin toggles exposed safely — A protected endpoint lists installed plugins and supports toggling while refusing changes to protected plugins such as the connector.
Chat tab metadata exposed to connector clients — Terminal clients can receive chat number, chat name, and project metadata from the connector API.

📄 Document Query

Standalone _document_query plugin — document_query now lives in its own plugin with parser strategy support, MIME-type priority routing, and externalized configuration.
Five parser implementations — PDF, HTML, text, image, and unstructured parsers are registered through the plugin's parser strategy system.
LiteParse preferred with safer fallbacks — LiteParse is the preferred parser, with legacy fallback behavior, subprocess isolation, bounded concurrency, and adaptive OCR heuristics for large text-rich PDFs.
Progressive skill loading — Long document-query instructions moved into a skill, leaving a compact tool prompt stub.
Expanded settings panel — Parser, retrieval, fetch, LiteParse/OCR, fallback controls, and a generated thumbnail are now exposed in the plugin UI.
Large PDF processing bounded — LiteParse OCR can be disabled above a configurable page threshold, and large documents use adaptive index chunk sizing.

Durable screenshot artifacts — Browser, Desktop, computer-use, and vision screenshots are now materialized into chat-scoped artifacts so historical image references survive temporary screenshot pruning.
Background computer-use contract expanded — Native window listing, indexed window state, and element_action dispatch help agents prefer structural targeting over screenshot-only workflows.
Computer-use capability metadata preserved — Contract version and capability metadata are stored and returned so remote computer-use prompts can prefer structured dispatch.
Internal Xpra desktop control hardened — Structured-first workflows, batched desktopctl sequences, and normalized resize requests reduce distorted X11 display behavior.
Xpra Canvas sizing fixed — Backend-normalized display dimensions prevent stretched Desktop views.

Coding-agent verification discipline improved — Prompt guidance now emphasizes reading specs and tests first, making narrow edits, verifying exact artifacts, cleaning temporary work, and recovering from tool failures.
Verification habits moved into core prompt behavior — Coding discipline now applies across Agent Zero instances by default.
Code execution timeouts raised — Builds, installs, servers, tests, and training jobs get longer timeout groups, with long-running work steered toward output polling.
Remote exec timeouts aligned — Connected CLI sessions use timeout groups that match code execution behavior.

Auto-memory extraction quality improved — Passive fragments, action history, temporary artifacts, and local runtime paths are filtered before consolidation.
Durable memory guidance tightened — Memory prompts now favor stable preferences, durable project facts, and reusable solutions.
Staleness and forget guidance improved — Timestamps act as soft recency signals for mutable conflicts, and memory_forget guidance is clearer for stale entries.

Unified OAuth account management — Provider-registry-backed status summaries, redesigned settings rows, and account-backed model slots simplify connected account setup.
Provider-aware onboarding — Onboarding now supports provider-specific setup fields, device-code and browser callback flows, and model loading by provider.
Extensible OAuth providers framework — Account-backed provider support can now cover xAI, Google Cloud, and GitHub Copilot accounts.

Skills selector simplified — Visible/Pinned mode switching was replaced by a continuous view where pinned skills appear first and hide/show uses an inline eye control.
Office, Desktop, and Editor plugins can be toggled — Their manifests no longer force always_enabled.
Project instructions can include top-level AGENTS.md — A UI checkbox controls this behavior while preserving backwards-compatible defaults.
Self-update cleans desktop runtime state — Desktop SSH agent state and legacy Desktop agent sockets are cleaned best-effort without blocking updates or removing GnuPG key material.
Model config snapshot sync moved into the store — UI-only field initialization and dirty tracking are cleaner.

Per-subagent model configs fixed — Subagent model configuration now applies correctly.
document_query prompt spelling fixed — Corrected optmimize to optimize and cleaned legacy test references.
Builtin ACP plugin reverted — The Agent Client Protocol stdio plugin was added and then reverted during this release cycle.