agent-zero

Agent Zero v2.14

Highlights

• Significantly faster startup and chat loading — Embedding warm-up now overlaps with chat restoration and startup watchdogs, reducing median boot-to-first-token by ~27% and chat-switch latency by running model/profile refreshes independently of snapshot rendering.
• Upgraded LiteLLM (1.104.0) and OpenAI SDK (2.54.0) with an automatic startup migration for existing Docker instances, including installed-SDK HTTP test coverage.
• Native LiteLLM routing for core providers — Nebius, Venice chat, Z.AI/Coding, and OpenRouter embeddings now use LiteLLM's built-in integrations, simplifying provider configuration while preserving existing overrides.
• Responses prompt caching fix for long conversations — Completed Responses turns now retain masked extras at their original boundaries, preventing growing conversations from discarding cacheable prompt prefixes. A live test reached 99.3% cache hit rate at 131K input tokens.

Browser & Performance

• Reduced browser startup and rendering overhead — avoids launching Chromium for empty tab listings, eliminates redundant load waits, and reuses native viewer display size. Local navigation medians improved from 474→83 ms, content reads from 82→19 ms.
• Simplified Browser quick settings to focus on preset, Host/Container, zoom, and extensions; moved extension management and Chromium settings pages into Browser config. Added annotation screenshots as an opt-in setting.
• Removed overlapping tooltips from Browser quick settings popover.

Chat & Sidebar

• Fixed disappearing chats when Sidebar Folders was disabled and another plugin imported the folders store.
• Fixed sidebar folder lag during activity and focus updates by replacing repeated full-context scans with a shared parent-child index. Median activity updates dropped from ~67 ms to 15 ms.
• Persisted sidebar Project, Status, and Last Activity filters across reloads.
• Fixed chat activity timestamps not updating when continuing a chat, by writing to the context timestamp in the shared history writer.
• Revealed children of selected subagents at every delegation depth.
• Fixed chat naming to preserve manual names under "always" mode and validate custom names before marking them manual.
• Fixed pruning of just-saved chat contexts by re-checking disk before removal.

Model & Preset Management

• Reload model presets before the preset editor opens, preventing stale drafts from deleting presets created in other tabs or via the API. Fixes #1928.
• Capped preset and agent profile selector menus to stay within the viewport.
• Detect Claude models across all providers (OpenRouter, Vertex AI, Bedrock, etc.) for thoughts-prompt adaptation, not just those with "anthropic" in the provider name.
• Restored full-response repeat checks for native tool calls, including commentary comparison.
• Added context window display controls and price currency selection with Frankfurter exchange-rate conversion and 24-hour caching.
• Enabled Venice and Agent Zero API cache routing; exposed a0_prompt_cache_mode for custom providers.

OAuth & Codex

• Improved Codex reconnection flow — handles rejected refresh tokens with actionable toasts, offers Reconnect without manual disconnect, and displays device challenge codes in a copyable box with accessible feedback.
• Pinned Codex client version to 0.160.1.
• Added success feedback when copying OAuth device codes.

Agent Execution

• Direct subordinate delegation now runs through the child's own context task, making child chats visible as Working/Paused and independently stoppable. Propagates caller cancellation while treating child-only Stop as recoverable.
• Terminal I/O no longer blocks shared event loops — uses nonblocking POSIX PTY reads and cooperative writes.
• Fixed native Windows TTY cleanup to avoid evaluating POSIX-only SIGKILL.
• Fixed SSH command truncation — uses sendall through asyncio.to_thread to deliver complete comma

Agent Zero v2.13

Highlights

This release brings a significantly improved file browser, sidebar project folders, slash commands for messaging integrations, native commentary streaming, broad performance optimizations, and numerous quality-of-life and stability fixes across the WebUI, LLM transport, plugin system, and desktop environment.

File Browser & Editor

• Redesigned file browser navigation — configurable path bar with breadcrumb and text modes, folder autocomplete with keyboard navigation, and fully clickable file rows without a loading overlay
• Consolidated create-new dropdown replaces separate New File and New Folder buttons
• Inline rename and new-folder creation in the shared footer input row, with modal rename preserved for external callers
• Preserved file tree ancestors — parent folders stay clickable during nested navigation; starting folder is now configurable and defaults to /a0
• Canvas and picker lifecycle fixes — opening a file picker no longer blanks a live canvas listing, and live browsers navigate in place for path links instead of stacking windows
• File permissions and ownership preserved when applying editor patches — executable bits and host-user ownership are no longer replaced with root-owned 0600 files
• Skip hidden Markdown previews when opening source files, reducing median open time for large files from ~22.5 s to ~0.37 s
• Text editor line edits explicitly supported — removed the legacy label and documented all three patch forms
• LibreOffice saves follow renamed documents — desktop rename now uses native Save As so Writer retains the correct save target

Sidebar & Chat Organization

• Project folders for chats and tasks — collapsible sidebar folders grouped by project, with flat-view option, creation-time sorting, drag-and-drop reordering, and whole-family project moves
• Pin to Top improvements — project folder pins including a default "No project" pin, reactive pin indicators beside chat/task titles, and scoped indicator rendering
• Deep worker rendering — subordinate workers beyond the first child level now render correctly with recursive expansion and tapered indentation
• Drag performance — median hover-handler time in a 240-chat sidebar dropped from ~10 ms to ~0.1 ms
• Clearer status filters — consistent All / Working / Idle labels; "Recent activity" shortened to "Last active"

Messaging Integrations

• Slash commands for Telegram — built-in, plugin, global, and project commands with paginated menus, compaction confirmation, transcript export, and fresh-chat session binding
• Slash commands for WhatsApp — reuses the shared context-aware command catalog with text menus, tool permissions, plugin settings, and persistent chat selection
• Shared slash command documentation added to the plugin creation skill
• Improved settings experience for Telegram, Email, and WhatsApp — project and agent-profile scoped configuration, token-native status pages, and shared settings accordions
• Telegram bot management — stable names for unnamed bots, duplicate name rejection, and automatic restart on connection changes

LLM & Model Handling

• Native commentary streaming — public commentary from native tool calls now displays as streamed thoughts in generation logs
• Commentary preserved in Responses history replay so eligible native turns retain their original message items
• Stabilized Chat Completions cache boundaries — cache breakpoints no longer shift on every tool turn; marks the last leading context and last two nonempty assistant messages
• Provider refusals surfaced — content-filter and refusal finish reasons now reach the visible error handler instead of being silently retried
• Anthropic-specific thoughts instructions — scoped prompt adjustments for Anthropic models (including via OpenRouter and Bedrock) requesting public action summaries
• **Scoped subagent model preset

Agent Zero Release Notes

Connector & Transport Hardening

• Enforced negotiated WebSocket payload ceilings with a symmetric checksummed transfer protocol for large operations, including bounded transfer state, disk spooling, and automatic cancellation on pause, reset, delete, or disconnect
• Made workspace HTTP file transfers atomic with streaming uploads into partial files, size verification, fsync, and SHA-256 download receipts
• Bounded remote connector tool payloads with explicit continuation metadata for large content, rejecting oversized inline artifacts before base64 expansion and routing bulk files to authenticated HTTP transfers
• Documented connector payload contracts including WebSocket ceilings, transfer state machine, cancellation lifecycle, and remote text limits

File Browser & Editor

• Added independent, lazily loaded directory trees to Files and Editor in canvas and modal modes with unified styling, lighter panel backgrounds, and flat toolbar controls
• Unified text and code editing in the shared Editor with ACE syntax highlighting, line numbers, native source search, and Markdown preview — removing the legacy file-editor modal
• Added File Browser preferences for sorting, view mode, initial tree visibility, and independent editing and transfer limits
• Exposed connected CLI and Launcher folders in Files with host permissions, authenticated HTTP transfers, size/SHA-256 checks, and atomic publication
• Centralized streamed file transfers and hardened archive operations with expansion and entry budgets, unsafe member rejection, and owned temporary-file cleanup
• Moved File Browser controls into a dedicated Settings category with appearance defaults, remote connections, file limits, and archive limits
• Added configurable editing (10 MiB default) and transfer (100 MiB default) limits shared across Files and Editor
• Fixed file links in chat messages to recognize file URLs and download API paths in the shared path-link renderer
• Polished Files selection toolbar, removed tree tooltips, matched folder icon tint to standard file icons, and hid file sizes in icon view

Responses API & Native Tool Calls

• Preserved native tool calls through Context Doctor, leaving canonical function calls untouched by text repair
• Stored canonical function-call content through assistant history and masking, with proper repeated-response detection
• Streamed native tool arguments through the existing live display with alias normalization and masking
• Described bundled tool arguments in Responses schemas for 15 tool implementations with non-strict definitions
• Preserved policy-filtered guidance in native Responses function descriptions with A0 envelope-to-argument conversion
• Projected native system instructions only for Responses requests, reducing Codex input from ~16.4K to ~11.3K tokens
• Preserved native Responses history within prepared prompts, retaining original IDs and encrypted reasoning
• Consolidated Responses prompt bookkeeping outside Agent, reducing agent.py additions from 61 to 24 lines
• Made llm_result required in hist_add_ai_response with LLMResult.non_llm sentinel for non-LLM turns
• Kept Responses stream reconstruction inside the parser with one authoritative output-item store
• Skipped unused Chat replay preparation and shared input conversion for Chat Completions

Context Doctor & Response Handling

• Added split-thoughts repair strategy extracting value selection into a helper with configurable settings
• Handled reasoning-only and plain-text fallback responses with dedicated retry warnings and user notices
• Tightened usable-response detection requiring at least one A0 key with a non-empty value
• Fixed raw-text fallback classification by moving raw thoughts wrapping after repair selection
• Counted thoughts_fallback warning toward the unusable-response limit
• Let Responses text reach the core response dispatcher while continuing to repair recognizable tool intent

Prompt & Profile Im

Release Notes

This release introduces several new bundled plugins, significantly improves streaming and real-time UI performance, adds vision sidecar integration with model presets, and delivers broad reliability fixes across memory, embeddings, context handling, and the interactive browser.

New Plugins & Extensions

• Context Window Usage Indicator — A bundled plugin showing a composer usage ring, responsive breakdown popover with token counts, percentages, free-space reporting, and provider-reported price/cache hit details. Includes independent mobile and desktop visibility controls.
• Context Doctor — Automatically repairs malformed tool-call JSON before dispatch, handles incomplete or thoughts-only responses, supports XML suppression, and persists repaired output in WebUI logs. Includes reasoning key-value pair handling and configurable log-detail settings.
• Vision Sidecar Integration — Strict per-preset sidecar slot with Main-first routing, one-call multi-image analysis, parallel-worker image resolution, configurable timeout/max-token controls, and inline model settings. Native vision output from parallel workers is preserved and promoted correctly.
• Result Extension Point — Extensions can now normalize or consume completed model results before default history and tool processing. Repeat detection and empty-response handling are split into dedicated, configurable loop-control extensions with prompt-backed retry limits.
• WebUI Plugin Slots — New extension slots beside the model/profile strip and at the end of interface visibility controls, allowing plugins to register responsive UI affordances.

Streaming & Real-Time Performance

• Responses now render with safe Markdown and LaTeX throughout live streaming, removing the raw-to-formatted swap at completion.
• Native response tool output (e.g., Codex answers) streams progressively through the canonical tool envelope.
• Expensive full-response parsing is coalesced behind a 128-character or 50-millisecond cadence while retaining per-chunk security hooks and strict tool validation.
• Message action buttons remain stable during streamed updates, preserving hover, focus, and extension controls.
• Native reasoning summaries are promoted as thoughts after streaming completes.
• Incomplete stream roots are rejected before invoking the dirty JSON scanner, avoiding repeated scans of partial prefixes.

WebUI & Synchronization

• Optimized real-time WebUI synchronization with collection deltas, stable sidebar row identity, bounded streamed log fields, deferred response formatting, and capped process group updates.
• Chat list state is preserved when filtered/sorted snapshots are byte-equivalent, with proper selection and tree synchronization.
• Branched chats detach provider state cleanly — inherited Responses continuation, response-ID ownership, and cached context data are rebuilt from trimmed local history.
• Context window indicator refreshes when each new generation starts, with deduplicated streamed updates.
• Step detail header actions are aligned with consistent height and mobile-friendly wrapping.
• First user turn renders optimistically; WebSocket state pushes are serialized so welcome-screen messages remain visible.
• Credential fields render as masked text inputs instead of password fields, preventing browser password-save prompts.

Embeddings & Memory

• Gateway model IDs (e.g., nvidia/llama-nemotron-embed-vl-1b-v2:free, openrouter/openai/text-embedding-3-small) are now kept intact for OpenAI-compatible embeddings — the provider prefix is always applied, matching chat wrapper behavior.
• NumPy scalar relevance scores are coerced to native floats at the source, fixing float32 is not JSON serializable crashes. Legacy NumPy scores are converted only in the memory dashboard response, keeping shared API serialization strict.
• Preloaded local SentenceTransformer models are cached by constr