v2.15

Agent Zero — Release Notes

Highlights

This is a large release that promotes core tools into standalone plugins, overhauls the welcome screen, hardens channel security, improves streaming and UI responsiveness, and adds new search, migration, and documentation capabilities.

---

Tool Plugin Architecture

• Extracted bundled tools into standalone plugins: wait, notify_user, search_engine, a2a_chat, scheduler, behavior_adjustment, and the scheduler Tasks UI are now standard plugins, each with dedicated badge styling, log types, and prompt files.
• Unified memory tools: The four separate memory_load, memory_save, memory_delete, and memory_forget tools are consolidated into a single memory tool with an action argument, matching the action-dispatch pattern used by other tools.
• Per-action tool permissions: Profiles can now allow or deny individual operations across all 12 bundled multi-action tools (e.g., allow file reads but block writes, or allow listing scheduled tasks but block deletion). The Agent Editor shows grouped action controls with expandable entries.
• Centralized boolean argument coercion across all promoted and core tools, removing duplicated helpers and strict-validation error paths.
• Dedicated badge rendering for goal, document_query, skills, vision, browser, office, and computer_use_remote tool calls, each with distinct accent colors and descriptive headlines.
• Dropped agent-name prefix from all tool log headlines, warning rows, and fallback entries for cleaner chat output.
• Added catalog thumbnails for all 47 bundled plugins.

Output Handling & Model Recovery

• Surface output-limit truncation instead of misformat retries: When a turn stops at the token limit without a complete tool request, the agent retries with a dedicated truncation warning instead of the generic misformat message. Fixes #1836.
• Guard truncated native calls: Reject incomplete native-call argument JSON during output-limit recovery while allowing complete calls to proceed. Stream finish metadata is now preserved through early cancellation.
• Recover DSML tool calls: Recognize standalone DSML function_calls envelopes, repeated tool names, and outer headlines, routing them through normal dispatch instead of failed conversion.
• Keep rejected tool syntax out of model history: Store neutral {} for rejected or incomplete leaked calls so retries don't reinforce malformed output. Replayable Responses output items are removed from the history copy.
• Restore Context Doctor repair before retrying malformed calls: Missing delimiters and other repairable output can execute in the same turn again, restoring pre-PR #1920 behavior.
• Keep model prose as thoughts and retry silently: Free text and reasoning-only output is preserved in the current generation and retried without intermediate warning rows.
• Fix include placeholder expansion in tool arguments: §§include(path) is now expanded recursively at the tool execution boundary for Chat, native Responses, and parallel calls.
• Explain missing LiteLLM chat streams with an actionable error instead of raising TypeError.

WebUI Improvements

• Faster response streaming: Reduced state-push coalescing from 100 ms to 25 ms, roughly halving median visible update intervals.
• Paint startup overlay immediately without waiting for render-blocking stylesheets.
• Block duplicate sends while a send is in flight, preventing multiple /message_async calls from rapid clicks or Enter presses.
• Serialize Welcome chat creation and first send so rapid input cannot create multiple contexts.
• Make composer pause optimistic against poll snapshots — clicks in flight now suppress snapshot writes to the pause flag.
• Keep action buttons mounted during streamed updates so hover and clicks aren't killed mid-stream.
• Preserve pending setup prompts and retry them automatically after chat preset changes.
• Gate model setup on backend rejection instead of a blocking round-trip before every send.
• Allow agent profile switch while chat runs, matching chat model-preset overrides.
• Keep the switcher label mounted during background refreshes instead of flashing away.
• Render GEN process steps as plain text without KaTeX, preventing dollar-sign sequences from being mangled into math.
• Scope process step DOM IDs by log type so subordinate responses can't overwrite GEN steps.
• Render pending queue thumbnails from local blob URLs until upload confirms; drop stuck entries on failure.
• Require double tildes for markdown strikethrough, matching GitHub behavior.
• Always load the full model catalog without a query so typing doesn't filter away sibling models.
• Drop chat-loading-splash background so only the dots are visible.
• Keep show-less control from shifting the action row on hover.
• Unify translucent glass style across header and canvas surfaces with a shared token.
• Replace hardcoded accent blues with the highlight token for easier theme customization.
• Copy full execution commands while keeping compact chat previews.
• Fix canvas rail dragging when pointer capture fails.
• Fix chunky response streaming by skipping inner scroller measurements for root response bodies.

Welcome Screen

• Dismissible welcome sections with browser-local persistence and unified restoration via "Show all dismissed."
• Reorderable welcome blocks with mouse, touch drag handles, keyboard controls, and persistent order. Plugins can join dynamically.
• Plugin Hub discovery from the welcome screen with a persistent puzzle-piece button.
• Half/full-width toggle for System Resources and Quick Actions panels with responsive layouts.
• Channel plugin discovery and display on the welcome screen, showing installed channel plugins with thumbnails and settings access.
• Compact Quick Actions layout with square tiles and labels below icons at half width.
• Greeting and wave kept on one line across screen sizes.

Sidebar & Search

• Configurable sidebar chat and message search with Standard, Compact, and No search modes, Unicode-safe match emphasis, and keyboard navigation.
• Chat message search in the Connector API with bounded, Unicode-safe literal highlight and 200-character query limit.

Chat Naming & Scheduling

• Exclude greeting seed from chat naming so fresh chats aren't named after the synthetic greeting; real first user messages name the chat immediately.
• Seed subordinate agents with the greeting few-shot example so all agents receive the response tool-call JSON demonstration.
• Clear scheduler errors after successful chat continuation — failed tasks now reconcile with successful replies.
• Reduce redundant same-profile delegation — specialists handle matching work themselves instead of relaying through identical agents.

Agent Editor

• Speed up agent editor loading by deferring Advanced data until first entry into Advanced mode.
• Clarify prompt editing with readable section titles, customization status, save paths, and variable hints.
• Grouped tool action permissions with expandable controls in Easy and Advanced modes.

Notifications

• Copy action below notification toast content for capturing full title and message.
• Hide Copy on notifications with action buttons and on routine confirmation toasts.
• Plugin hooks for notification arrivals and modal header actions, enabling community plugins to retain transient errors and expose history.

Migration

• Pi import support for JSONL session versions 1–3 through the existing Migrate Agents workflow.
• Claude chat history preservation with custom titles and tool-result blocks kept in historical tools.

Shared Files

• Per-chat file browser with own folder, history, and selection by default, plus a setting to share one browsing session. Welcome panels reuse the real file browser.

Security

• Enforce same-origin isolation for browser DOM helper RPC — cross-origin windows can no longer read page content or invoke DOM actions through the helper's postMessage channel.
• Require explicit Telegram administrators for settings changes — accepted senders can no longer use /permissions to remove their own tool restrictions or toggle plugins.
• Require explicit WhatsApp administrators for settings changes, with a separate admin_numbers allowlist enforced at mutation boundaries.
• Reject unsupported Telegram configuration scopes — Telegram is now explicitly global-only, and legacy scoped configurations pause the bot until reviewed.
• Require remote terminal permission for input_remote — a custom profile can no longer bypass blocked code_execution_remote through an independently enabled input tool.

Configuration & Infrastructure

• Preserve .env configuration during startup migration — existing usr/.env is kept authoritative without overwriting conflicting legacy files.
• Fix self-update ZIP dates and reduce backup compression time (~17% faster with deflate level 1).
• Raise self-update health timeout to 600 seconds for updated and restored runtimes.
• Provider plugin discovery during onboarding with puzzle-icon links opening Plugin Hub filtered by provider.
• Codex onboarding defaults updated to gpt-6.1-terra (Main) and gpt-6-luna (Utility); final onboarding action changed to "Continue" preserving the current chat.
• Codex client version bumped to 0.166.0.
• Plugin composer controls exposed to connectors via capability-gated, session-authenticated API.
• Codex OAuth response defaults (reasoning effort, summary, verbosity) stay visible and editable across connection states.
• Test suite fully collectible — fixed 127 import errors, isolated suite pollution, resolved nested dev checkout path issues, and aligned stubs with current behavior (3,067 passing, 0 failures).

Documentation

• New guides: Local folder mounts and Launcher workspace storage, OpenCode Go setup, Codex account setup from ChatGPT, real agent migration walkthroughs, sidebar search instructions.
• Refreshed: README with plain-language introduction and complete quickstart, Quick Start around first message flow, Agent Profiles for action permissions, Model Presets, CLI onboarding preview, Plugin Hub screenshot, and Launcher onboarding.
• 12 new real screenshots across Quick Start and everyday documentation, optimized from 605 KB to 251 KB total.

Breaking Changes

• Unified memory tool: The four separate memory_load/memory_save/memory_delete/memory_forget tools are replaced by a single memory tool with an action argument. Legacy calls are preserved for compatibility, but custom integrations referencing the old tool names should migrate.
• Telegram configuration is now global-only. Existing project or profile-scoped Telegram configurations will pause the bot on startup until manually reviewed and removed.
• Browser DOM helper bumped to v2 with same-origin enforcement. Existing host browsers must restart and reconnect so old initialization scripts are discarded.
• input_remote now requires remote terminal permission in addition to its own permission. Profiles granting input_remote without code_execution_remote terminal access will be denied.
• Scheduler store must be initialized before snapshots — custom integrations accessing the task store at import time may need adjustment.
• Dev checkouts nested under /a0 will see a one-time workspace ID change; existing shadow repos are re-initialized in a new location.