Configurable conversation layout and voice correction that learns your vocabulary
Startup and sign-in: theme-aware loading and better window-control spacing; email verification codes support cell-based entry, paste, and automatic verification when complete. Existing account and license checks remain in place.
Adjust the browser without leaving your work: The built-in browser now offers page zoom, device viewport controls, and in-place cookie import from its toolbar menu. When a viewport is larger than the panel, scroll horizontally or vertically to reach its edges.
Navigate files, preview media, and paste long text: File views and review previews gain a segmented path bar: open directories, pick files across folders, or copy the full path. Images support zoom and fit-to-window, SVGs offer source and preview modes, and audio uses dedicated playback controls with clearer file-type icons.
See built-in skills, project rules, and subtask details: The slash menu can now show Claude Code skills actually available in the current runtime. The right-hand skills panel lists built-in skills separately from project skills as read-only entries. New /list-agents and /autocompact entries appear when supported by the runtime.
Claude Code plugins: browse marketplace catalogs from the extension center, manage sources, install or remove entries, and inspect status or errors. Review third-party plugin contents before installing.
Read channel conversations and task progress more easily: Channel conversation titles can independently show or hide the sender and group or contact name. Tabbed organization settings make title, marker, and layout options easier to find. Navigation entries can be dragged from the whole row while customizing.
Conversation layout: markers, sources, and activity states can sit left or right of the title, with three quick presets and each position still independently adjustable.
Sidebar presentation: sidebar text size is independent from message and code size; set an initial sidebar width while a manually dragged width stays remembered; navigation entries can be shown, hidden, and reordered, with hidden entries still reachable from Explore.
Conversation organization: mark conversations with an emoji, color, or reusable label, filter by label, and rename manually or accept an editable AI suggestion — manually named conversations are not silently overwritten.
Voice correction: add custom correction rules, industry templates, and manual words; the app learns terminology from clear pre-send edits without an extra model request, excluding ambiguous single-character, numeric, negation, and broad-rewrite edits.
Dictation history: separately review the original transcript, AI-corrected text, and the text you actually sent.
New session: an empty conversation can show cached official release, documentation, and blog cards between the welcome message and composer; a new Help entry opens documentation and focused provider, shortcut, and appearance settings.
Code reading: Read tool results show file name, language, and line range with theme-aware syntax colors, exact copy behavior, and bounded excerpts for very long files.
Reliability: new-session drafts survive navigation, startup failures show a recoverable error instead of a black window, daily briefing keeps a valid cached overview when a refresh fails.
Claude Agent SDK updated from 0.3.259 to 0.3.278, with compatibility checks for the changed upstream behavior.
Changed defaults: conversation, schedule, and recap tool groups require explicit enablement; existing conversations and schedules are retained. AI naming is opt-in. Browser runtime starts off on new installations and preserves an existing enabled setting.
Upgrade: startup migrates existing conversations and settings and creates a local upgrade-backups snapshot when the database needs a schema upgrade; do not delete application data as part of an ordinary update. Older apps refuse a newer database schema; rollback requires a matching pre-upgrade backup and loses changes made afterward.
Packages and limits: macOS Apple Silicon / Intel and Windows x64 installers are provided. macOS uses ad-hoc signatures without Apple notarization; Windows has no publisher certificate signature. Application-file integrity signatures are separate. Real Intel/Windows installation and old-version differential updates have not been verified on this build host.
Redesigned workspace, conversation browser history, and clearer switching
UI/UX: reorganized conversations, sidebar, settings, and previews; subtler focus states, dark popover borders, and consistent macOS icons.
Appearance: eight built-in workspace palettes, light and dark modes, accent color, font size, font weight, and reading density; coordinated terminal and editor themes.
Plugins and skills: a project-level menu for plugins, skills, and MCP. Skill details open an editable draft in the selected project.
Browser: per-conversation history with title and URL search, a shortcut beside permissions, and collapsible history in conversation information. Loading and error states are consistent; retrying the same URL reloads correctly.
Conversation switching: show the target first, then load history in batches. Preserve drafts, attachments, split views, and background streams; hide information overlays during dragging.
Agent progress: compact cards and activity rows retain the actual current step. Very large individual messages can still take time to render.
Mermaid: diagram enlargement, source editing, type and detail settings, with less flicker and layout movement.
Settings: better service-form scrolling, External Access onboarding, saved model and reasoning preferences, usage accounting, and audio-permission recovery.
Upgrade: 1.2.0 replaces the planned 1.1.24. macOS Apple Silicon / Intel and Windows x64 installers are available. Startup migrates existing data and creates a local upgrade-backups snapshot when the database needs a schema upgrade; downgrading is not guaranteed to be lossless.
Platform notes: macOS builds are not notarized; follow the installation guide. Windows retains blockmap differential updates, but a real Windows installation from an older version through that path has not been verified for this release.
WeChat private chats — connect by scanning a QR code, send tasks and files, and receive replies alongside existing Feishu, DingTalk, and Discord channels. WeChat currently supports private chats with a single bot.
Session teams — with the team Skill available, start from @Team, agree on a plan and file ownership, then work through real child sessions. The team cockpit shows member progress, stages, and estimated cost; completed work returns to the lead for review.
Built-in integrations — connect external applications through scoped API keys, isolated sessions, and Webhooks. Manage applications and projects in Onevium, with an optional self-hosted gateway for remote access.
Browser control — keep multiple tabs open, check responsive layouts with per-tab viewports, and run page actions in batches. Clearer permission prompts and an in-page action label make ongoing work easier to follow.
Review and reply annotations — attach notes to code lines or selected reply text, send feedback to the agent, and move between changes and files with floating navigation. Large-diff scrolling does less overlay work.
Connected sessions — child, team, and scheduled turns stream into the conversation while running. Session trees, cross-session messages, and subagent details make delegated work easier to follow and continue.
Voice dictation — choose a microphone and local recognition model, including SenseVoice, and copy transcripts or download and delete recordings from history. Optional polish runs after recording stops and retains the raw text if it cannot complete.
Provider and channel updates — refresh model presets and pricing for GLM, Kimi, Moonshot, and DeepSeek; improve Discord setup, channel naming, reconnect behavior, and preservation of existing user settings.
Voice dictation, managed services, and Claude Fable 5.1
On-device voice dictation — speak directly into the composer with local speech recognition on macOS, Windows, and Linux; audio stays on your device.
Optional transcript polish and dictation history — clean up a local transcript with your configured AI provider, revisit recent dictation, and keep typing whenever voice is not the right input.
Managed services for long-running work — Onevium can keep dev servers, databases, and APIs attached to a session with real process IDs, operating-system port checks, TCP readiness, and state that survives an app restart.
Service controls you can trust — the Activity Panel reflects the real process state, offers stop and restart actions, and caps retained logs so a long-running service cannot grow without bound.
Richer team context — Feishu and DingTalk agents can use recent group conversation; DingTalk also adds member and file access, while scheduled channel pushes can address real @mentions.
Claude Fable 5.1 and Claude Agent SDK 0.3.251 — refreshed model support and the latest upstream reliability work included in this release.
Clearer work in motion — tool activity and long output overlays are easier to read, split-screen routing and the chat minimap are steadier, and the built-in browser gains console, network, and typed-input improvements.
macOS microphone permission handling and lower Autonomy prompt cost round out the release, with clearer recovery when system access is not yet granted.
Rock-solid Windows runtime and self-healing startup
A complete embedded Windows runtime now carries the native AI dependencies it needs, removing a class of message-disconnect failures on packaged installations.
Self-healing startup — if the internal server exits unexpectedly, Onevium restarts it instead of leaving the desktop window on a blank screen.
Local diagnostics under ~/.onevium/logs make startup and runtime failures easier to inspect without exposing project content.
Sign-in is more resilient on restricted networks, with a faster offline fallback when the authentication service cannot be reached.
Windows updates now wait for a safe application exit, and duplicated reasoning content no longer appears during streamed replies.
Goal mode — use /goal to give Onevium a completion condition for longer work, with a live per-session indicator that shows iterations, elapsed time, and the latest check while the agent keeps going.
A more reliable built-in browser — reopening and refreshing AI-controlled pages is steadier, browser actions read clearly in chat, overlays keep a frozen page frame behind their menus, and screenshot, download, viewport, cookie, and profile flows are more dependable.
Editable SVG canvases — open generated SVG work as a visual canvas, select or box-select elements, drag groups, edit text, add shapes and routed arrows, resize the canvas, zoom the view, and save changes back to the conversation.
Split conversations feel calmer — active columns are easier to read, dividers are lighter and can be double-clicked to restore equal widths, narrow composers adapt sooner, and long conversations mount progressively to keep multi-column work responsive.
Settings stay usable over split views — settings now live in a true overlay layer, and nested dialogs, selects, dropdowns, tooltips, focus, Escape handling, and scrolling no longer fight with the panel underneath.
Faster review loops — chat file links open in a focused preview drawer, review routing is steadier across projects, and abandoned Monaco diff work no longer surfaces a red Runtime Canceled screen.
Current Claude and provider support — Claude Agent SDK 0.3.226, Anthropic SDK 0.116.0, refreshed Claude model capabilities, safer provider migrations, and preserved custom model, role, environment, and disabled-state choices.
New Usage page — Settings → Usage now shows your Claude account, live plan limits as clear progress bars, and a behavior-analysis panel, so you can see how your session and weekly quotas are trending at a glance.
Terminals that remember — the built-in terminal is now cached per project: collapse it and your history and running processes are still there when you reopen, and every session in a project shares the same terminal instead of spawning a fresh one each time.
In-chat command cards — running /usage, /context, /config, or /recap in chat now renders a clean, readable card instead of raw text output.
Richer session right-click menu — right-click any session in the sidebar to copy its path, open it in the built-in terminal, or open it in your editor of choice.
Live session list — the sidebar now updates the moment a session changes through a lightweight push channel, without re-rendering the whole list on every tick.
Smoother display — messages scrolled off-screen skip rendering entirely, and decorative effects step down automatically on GPUs without hardware compositing, keeping long sessions and split views fluid.
Polish — corrected archived-session timestamps (previously off by your timezone offset), restored several keyboard shortcuts, and fixed a split-column rename dialog that wouldn't save.
Faster diffs & streaming, reliable workflow progress, crisper Windows UI
Big diffs open instantly — reviewing a change in a large file no longer freezes the app; diff views now highlight only the lines on screen instead of the whole file.
Smoother streaming replies — render work per streamed token is roughly halved, and long conversations skip layout entirely for messages outside the viewport, so long sessions stay fluid.
Sessions open rock-solid — opening a chat lands directly on the latest message with no content pushing or jumping while history settles in.
Workflow progress you can trust — the workflow card now reconciles against durable run records, so it always reaches its true final state (even for background workflows or after a restart), shows live per-agent completion counts, and stops all motion the moment the run finishes.
Polished workflow launch — freshly spawned agents show clean loading rows that resolve into their real task labels, with a subtle staggered entrance.
Crisper icons on Windows — toolbar and navigation icons now pin their stroke width physically, fixing the fuzzy, low-quality look on 125%/150% display scaling.
Lighter on your machine — background polling pauses while the window is minimized, decorative animations switch off automatically on GPUs without hardware compositing, and several caches are now bounded.
Windows fixes — Feishu/Lark file downloads, 'Show in File Explorer', Ctrl+↑/↓ diff navigation, and background AI features (Git Bash auto-detection) now work correctly.
Workflows no longer get interrupted mid-run — a live Workflow is now correctly recognized as in-flight, so the resume/wake path waits for it to finish instead of aborting it.
Reliable resume after a Workflow — when a turn launches a Workflow and waits, the conversation always wakes back up with the Workflow's results instead of leaving you with "it ran but nothing came back".
Model-switch alerts — if Claude declines a turn and the session falls back to a different model, Onevium now tells you instead of silently continuing on another model.
Clearer tool permissions — approval cards show human-readable tool names (especially for MCP tools) instead of raw wire identifiers, so you know exactly what you're approving.
Safety-check notices — when a risky operation is blocked by a safety check, you get a clear heads-up instead of a buried error.
Latest Claude Agent SDK (0.3.195) — bumped from 0.3.170, picking up the newest models and upstream reliability fixes.
Accurate context meter — the context-window indicator now reads your model's real window size straight from the SDK, so it's correct on large-context (1M) models instead of overstating usage.
Smoother reconnect & long-running tasks — reattaching to an in-progress turn and waiting on background work are more reliable.
Memory recall v2.2 — a cross-encoder reranker (bge-reranker-base) is now the precision gate on top of the cosine + BM25 + entity candidate pool. On real Onevium memory the old cosine path ranked unrelated same-project memories ABOVE the relevant one at 0.70+ similarity; the reranker pushes those down to ≤0.10. When the model is still warming up your chat falls back to the v2.1 path automatically — no first-message wait.
In-page reattach to running turns — open any chat with a server-side turn still running and the reply continues inline, the input is locked while it streams, and the sidebar shows a per-session in-progress indicator so you can find it. No more 'background finished but the page is dead' state.
Per-provider role-model mapping — custom Anthropic-compatible providers now expose an editable role → model table (opus / sonnet / haiku) in Settings → Providers. Useful for company gateways that default to a model you don't want; raw env_overrides still win, and built-in presets are unchanged.
Long-task protocol split — the system prompt now distinguishes self-monitoring background work from foreground build / test commands, so a long build is no longer pushed into a `run_in_background` slot that will be reaped before it finishes.
Workflow card persistence — inline sub-agent cards remember their expand state across stream → persistence handoff and across navigation, eliminating the auto-reexpand jump and the visible list shift when a workflow completes.
Streaming polish — the backdrop-blur over streaming content is gone (was causing visible flicker in split chat columns), and sidebar session polls dedupe identical responses so the list no longer rerenders on each tick.
Preview teardown — the `web-preview:get-view-for-session` IPC handler now tears down idempotently, fixing the 'Failed to Start' dialog seen on some Electron lifecycle races.
Auto-continue surfaces simplified — the 'auto-armed waiting' inline row was a misfeature (users stopped typing to wait for it). Auto-armed fallback suspensions are now an invisible safety net; promised work still resumes itself, just without an interruptive UI affordance.
Claude Fable 5, latest Claude Agent SDK, smarter continuity
Claude Fable 5 support — Anthropic's new model is now selectable alongside Opus 4.8 and Sonnet 4.6. Opt-in, priced at $10 / $50 per MTok. Opus 4.8 remains the default for new chats.
Claude Agent SDK 0.3.170 — bumped from 0.3.158 across this release, picking up upstream fixes and three latent optimizations. M1 regression resolved.
Auto-continuity for promised work — when the assistant says it'll come back with results later AND has live background work running, Onevium automatically wakes the turn when that work finishes. Stale background tasks also expose a manual continue button.
Per-session reasoning level + per-model default — each chat remembers its own reasoning effort (low / medium / high). Settings → General now has a per-model default for new sessions.
Live sub-agent following — drill into any running sub-agent and watch its transcript stream in real time. Cards anchor where you triggered them, expand smoothly, name themselves clearly, and stay visible across long workflows. Cancel any sub-agent independently with F5.
Live reattach to in-progress turns — reopen a chat that's still streaming and seamlessly pick up where it left off, with no split messages or lost permission cards.
@Recap MCP — read-only MCP tool that lets agents query your activity across any time range. Pairs naturally with the daily briefing.
Smarter daily briefing — activity items ranked by importance instead of just chronology, with one-tap copy-as-narrative. Fresh installs come pre-seeded with a 17:59 daily briefing schedule.
Review panel upgrades — per-file discard with a restorable archive, jump-to-file from any diff, virtualized highlighter for large files, auto-refresh after commits.
Activity split panel — stays warm across navigations (no reload flicker), skeleton loaders during loads, quick actions per item, and project-session filtering.
Worktime in message footer — every assistant message now shows response time.
Channels reconnection rebuilt end-to-end — bots self-heal on transient network issues with bounded, jittered backoff, and never give up while the app is running. Idle bots no longer get flagged as down by a single blip.
Connection-health indicator on the sidebar — sustained connection failures surface as a red badge on the Channels nav so you know when a bot needs attention, without having to open the panel.
System-wide network hardening — outbound connection limits and request timeouts prevent any single integration from affecting the rest of the app under poor network conditions.
Ultracode: multi-agent workflows for Claude Opus 4.8
Ultracode mode for Claude Opus 4.8 — a one-tap mode that runs your task as a multi-agent workflow: the model decomposes the work, dispatches sub-agents in parallel, verifies their findings against each other, and synthesizes one result. Toggle it from the composer; the choice persists with the session and can be flipped mid-stream.
Inline workflow card — every workflow collapses into a single live card in the conversation: sub-agents grouped by type, per-agent run / done / fail status, and a workflow title. Expand any sub-agent to read its full transcript, and follow a running agent's output as it streams without scrolling to chase the latest line.
Live reattach — reload the window, drop your connection, or switch sessions mid-turn and Onevium reconnects to the in-progress reply, keeps streaming it, then hands off seamlessly to the finished message. Turns no longer get stuck after a refresh.
Workflow progress in the Activity Panel — long multi-agent runs surface their current phase and per-agent progress in the side panel, so a large job stays legible at a glance.
Mid-stream effort and Ultracode switching — change reasoning effort, or turn Ultracode on or off, without ending the current turn.
Turn continuity hardened — a self-notifying background task (a workflow, an agent, or a background process) no longer double-wakes a parked turn, removing a spurious error and duplicate work. A turn is also no longer idle-aborted while a workflow or background agent is still running.
Confirm-on-quit and macOS reopen fix — closing the window now asks before quitting an active session, and reopening from the dock no longer hangs.
Upgraded the Claude Agent SDK to 0.3.156; project context menus now render solid over the animated sidebar, with monospace ligatures in code.
Claude Opus 4.8 on day one, Memory v2 window-extraction, chat error boundaries
Claude Opus 4.8 wired in on launch day — model picker, effort tiers (low / medium / high / xhigh / max), and usage-cost catalog all include claude-opus-4-8 from the moment you update. A database migration reseeds the model list for users who connected Anthropic before 4.8 shipped, so it appears in Settings → Providers → Manage Models without re-registering.
Memory v2 — window-extraction redesign. The extractor now sees the last N user+assistant turns instead of a single-turn snapshot (N = frequency: high=1, mid=3, low=5), and existing session memories are passed alongside the prompt so the model stops re-extracting facts it already captured. Dedup collapsed from a two-stage election + merge into a single reconcile pass that decides NONE / UPDATE / MERGE / ADD per candidate. Chinese-language recall is materially better.
Chat error boundaries shield bad messages — a malformed message can no longer crash the whole list. Individual bubbles fall back to an error placeholder while the rest of the conversation keeps rendering.
Memory recall popover redesigned — hidden scrollbar, fade gradients at the edges, click-to-expand long memories, and project-aware origin labels (so you can tell which project a global memory came from). Scrolling through long recall lists is much smoother.
Auto-continue row centered + image grid cell overflow fixed — the auto-continue affordance now sits centred under the last bubble instead of left-aligned, and image grids no longer push their cell content past the bubble edge on narrow widths.
CompactingPill animation softened — the /compact in-flight indicator now uses a calmer amplitude so it doesn't visually compete with surrounding bubbles.
Per-project memory extraction blocklist — exclude individual projects from automatic memory extraction without disabling extraction globally. Existing memories from excluded projects still recall normally; only new writes are suppressed.
Daily Briefing reliability, faster git diff, and tighter production hardening
Daily Briefing reads cleanly — the "For you" insight no longer leaks raw JSON braces or wrapper text. Content is generated as plain prose with a defence-in-depth fallback that strips JSON-shaped escapes if any slip through.
Faster git diff in the Inspector — the Review tab opens noticeably faster on large repos. Diff hunks render in chunks, the underlying API does less synchronous work, and switching between files no longer freezes the panel.
Background work keeps your machine awake — Onevium now holds a `prevent-app-suspension` power assertion while a session is running, a channel is processing a message, or a scheduled task is in flight, and releases it the moment the work ends. Sessions, channels, and schedules no longer get killed when the laptop idles to sleep.
Smoother chat motion — freshly-sent user messages slide up into place with a short, calm easing; existing history loads instantly without any entrance animation. Split columns refresh their titles and message state reliably when you swap sessions between columns, and the top progress bar no longer gets stuck in split mode.
Tighter production builds — DevTools is now disabled in packaged releases. Dev mode is unaffected.
Trial period for new email signups now reads from the admin panel — new Onevium accounts get a 7-day trial by default; the duration can be changed in real time from the admin panel without redeploying the auth worker.
Memory v2 That Actually Remembers, plus a Cross-Project Morning Briefing
Memory v2 — three-layer context injection ships across every chat. A new User Rules editor (Settings → Autonomy) lets you hand-write Markdown that attaches to every project. Pinned memories and a 24h-cached project digest sit above a per-message hybrid local search (vector + full-text, top-5) so the relevant facts land in the prompt before the AI starts thinking. No network calls; the whole pipeline runs locally against on-device embeddings and SQLite.
See exactly what got injected — every user message gets a small chip in its footer recording which memories were attached when the AI answered. Hover the bubble, click the chip, see the list with relevance scores. The snapshot is independent of the live store, so editing or archiving a memory later doesn't change the audit trail.
Memory hygiene is automatic — duplicates merge, contradictions get flagged, stale entries get archived rather than deleted. Background maintenance once a day; nothing is ever lost. Memory cost stays under ten cents per month for a typical user (extraction uses Haiku; retrieval is free).
Daily Briefing — a new sidebar entry that auto-generates a morning briefing across all your projects. "Pick up where you left off" hero zone with one-click resume into the right session, plus yesterday's decisions, things you learned, what shipped, what failed, and what's still waiting for your reply. A "For you" insight card surfaces patterns from your own messages (context-switch cost, repeated questions worth pinning, etc.) — constructive only, never judgey.
Independent of Memory — Daily Briefing has its own toggle, its own model selection (default Haiku), and its own daily-budget cap. Turn off Memory and the briefing still runs; you just lose two memory-derived blocks.
Claude Agent SDK upgraded to 0.3.150 — picks up the new TaskCreate / TaskUpdate / TaskGet / TaskList tools (TodoWrite is gone upstream; we keep both wired so old chats still render), `status: requesting / compacting` for clearer streaming state, and a friendlier error when the selected model isn't available from the current provider.
Three-layer defense against the upstream "Thinking…" stuck symptom — server-side iterator terminal-timeout, client snapshot staleness signal, and a DOM-side watchdog that reconciles from DB if the SSE channel goes silent for 30s+ on an active stream.
AskUserQuestion / ExitPlanMode no longer abort after 5 minutes — the cards now keep the stream alive for up to 6 hours so you can step away from a permission prompt without losing the turn.
ActivityPanel polish — Processes and Monitors sections gain a "Clear completed" button that wipes terminal-state rows for the current session without touching running services or parked monitors.
Scheduled task `suspend_turn` boot-stale grace extended from 5 minutes to 1 hour — covers lunch / meetings / brief lid-close without losing the suspension.
Pin / Unpin projects in the sidebar sessions list — pinned projects stay at the top across re-mounts.
Quieter Status Icons, Count Badges, and a Sharper Inspector
No more Settings hang on repeated navigation. Settings sections, Channels, and Scheduled Tasks now stay responsive under heavy back-and-forth switching. Sidebar running counts and the CLI import check are shared across re-mounts so re-mount bursts can't flood the request pipeline.
Inspector sharpened — `⌘F` searches inside the active file; the panel mounts on the `/chat` empty state so the file tree is always reachable; search no longer yanks scroll to top; "Open in external editor" gets a pending state with an opened-in-external affordance.
Quieter chat surface — lighter sidebar typography, revived chat skeleton, a muted plug icon for connection status, flat count badges (no more breathing pulse, gracefully truncates to "99+"), and the assistant message edge now aligns with the composer.
Steadier chat behavior — Image Agent state is scoped per split column; permission profile no longer flips back to Default on new or switched sessions; MessageMinimap triggers and PlusMenu close-on-toggle were polished.
Chrome bridge no longer loops on failure — a failed bridge start (e.g. Claude CLI missing) requires an explicit Retry now, no more silent auto-retry storms or lingering poll intervals in the main process.
Schedules — session_mode control is back in the create / edit dialogs.
Steer Reliability, Smoother Widgets, and a Refreshed Login
Steer messages land reliably in the current turn — queued Steer messages once again inject mid-stream, even during tool calls. The earlier 'queued message never lands' behavior is fixed at the root.
Smoother widget rendering — widgets stream progressively with a smoother finalize step, raw HTML stays hidden until layout settles, and a safety net guarantees the final widget always becomes visible. The action bar now sits in a fixed slot above the iframe so it never overlaps content; drawio previews recover gracefully when the sandbox blocks initial load.
Refreshed login & activation flow — animated particle wordmark, dedicated hero phase, and auto-verification of the email code. Activation feels noticeably faster and less click-heavy.
Per-vendor cost visibility — Anthropic rows reflect SDK-reported actual spend (Pro/Max subscribers no longer see inflated retail prices); third-party Anthropic-compatible proxies are recomputed against the local catalog. The Total Cost card now splits 'Anthropic · actual' from 'Others · list price' so you can see exactly where your spend comes from.
Channels MCP tool group — agents can now read and update channel metadata directly from inside chat, with built-in redaction and setup guidance.
Browser preview polish — comment popovers and pin overlays got a full pass: clearer affordances, better positioning, and a shared palette abstraction. Orphaned previews and lingering picker state are cleaned up reliably.
Inspector pane discipline — the Review tab hides on non-git projects and routes write actions to the File tab. Empty Review state offers a one-click 'Initialize Git' entry point.
Quieter scroll behavior — ambient background updates use soft scroll so you're no longer yanked back to the latest line; the Thinking panel passes wheel events through to the page underneath.
Faster startup — initial-mount fetches are de-duplicated through module-level promise caches, cutting redundant requests on boot.
Latest Claude Code engine — upgraded to Claude Code 2.1.140 (Claude Agent SDK 0.2.140) for the latest CLI parity.
Type @ to Build Anything — Inline Capability Chips for Skills, Memory, MCP, Schedules, Channels, and More
Type `@` in any chat to summon Onevium's nine built-in capabilities as colored inline chips — Skill, Memory, Agent, MCP, Plugin, Schedule, Channel, Browser, and Widget. Each chip launches a guided AI flow that produces a real artefact in your project: scaffold a skill, save context as memory, register an MCP server, schedule a recurring job, attach a DingTalk / Feishu / Discord channel, control the in-app browser, or build a custom widget. The chip looks identical in the input box, in historical messages, and in the message queue.
Future-proof capability system: adding new capabilities in upcoming releases never breaks existing chats — old conversations keep rendering correctly, and new chips fall into place without an upgrade ritual.
Tighter @ menu visual: tools and files share a calmer popover with smaller padding, lighter borders, and a softer label color. Type to filter capabilities; hit `@/` to walk the project file tree. Group titles, entry rows, and chip squares read as one cohesive tool list.
Cleaner session titles: titles auto-derived from messages with chips now show plain capability names in the sidebar, breadcrumb, and drag preview.
Smoother macOS update prompt: the in-app **Update Available** dialog now opens the GitHub release page directly, so a single click takes you to the latest download.
Channels Redesigned, a Smarter Browser, and Automations v2
Channels — redesigned multi-platform bridge: inline editing for every field, a unified creation flow with explicit credential registration, and a calmer list/detail layout that mirrors Automations. DingTalk bridges now @-mention people by their real display name, never lose a reply when an AI Card streaming call fails, and surface running state in the sidebar.
Browser — real browsing inside Onevium: persistent profile so cookies, history, and sessions survive launches. ⌘ / Ctrl + click on any Markdown link opens it inline with a clear hover badge — plain clicks still go to your default browser. Navigation shortcuts, auto-fit width, dialog clipping fix, and file:// support for managed temp HTML.
Automations v2 ships in full: a 1:1 redesign of the automations workspace with cleaner spacing, unified delete dialogs, and the legacy v1 path fully retired.
Proxy — Global mode with authenticated proxies: HTTP Basic auth on the global toggle, real outbound IP shown for verification, and the embedded terminal now routes through proxy-bridge so nothing leaks around the global setting.
Platform-aware keyboard shortcuts: shortcuts display ⌘ on macOS and Ctrl on Windows/Linux, with the platform-native modifier order across menus, tooltips, and the inspector.
Multi-column split view polish: active column focus is now distinguished by font weight, opacity, and a unified hairline rather than bright background tints. Hover hints toned down for a calmer multi-session view.
Chat & streaming stability: stable assistant message IDs across the bridge → DB transition (no more flash on stream completion); smooth stop transitions; the streaming status bar remains visible even when the process view is hidden; thinking panel unified with a compact left-rule style and tokenized link / search-match palette.
Inspector & diff polish: Write-tool clicks route to the Review tab; large files preserve their diff hunks instead of collapsing; the add-tab menu picked up a Kbd keycap and hairline shadow.
UI polish across the app: popover surfaces unified on a single border + shadow token; sidebar focus, hairline menu shadows, and badge anchor stability; tightened Steer tooltip; inactive macOS traffic lights are visible again.
Page scroll flows through the streaming reasoning area: hovering a long reasoning block no longer traps the wheel inside it. Long thoughts render in full and the chevron still collapses them, while the chat's auto-follow to the latest message stays intact.
File Inspector adapts to projects without git: opening a Write or Edit tool result in a non-git project now routes directly to the File tab with the same scroll-to-line behavior, instead of stranding on an empty Review tab. Running git init later in the session brings the Review tab back automatically.
Inline previews for images, audio, video, and PDF: the File tab now renders these media types directly instead of showing a binary placeholder.
Activity panel task list runs tighter: switches to the scroll-with-fade view after five items instead of ten, at a more balanced height for shorter sessions.
Sessions That Don't Get Lost, A Browser Inside Onevium, and a Faster Chat
Turn Continuity v2: a complete rewrite on top of the latest Claude Code SDK. Sessions resume reliably across sleep/wake, app restarts, and external interruptions — tool approvals, todos, and streaming output stay coherent end-to-end. The old 30-minute stream lifetime ceiling is gone.
Sandbox Browser inside Onevium: preview any URL on a popover surface, pin pages to revisit, annotate with the on-page element picker, and jump anywhere with the ⇧⌘P quick picker. New MCP tools let agents open and reload preview windows directly from a workflow.
Drag-to-split chat with persistent composer: pull any session into a new column to work on multiple threads side-by-side. Drafts, attachments, and input state now survive session switches, and an optional Message Navigator minimap lets you jump between turns in long runs.
Reasoning effort picker: choose how hard the model thinks per turn — exposed alongside the model selector so you can dial up exploration without leaving the composer.
Live preview reload: edits to HTML or source files refresh the in-app preview automatically. Next.js, Vite, and similar dev servers are detected on startup and offered as a one-click open.
Faster chat at any length: heavy markdown and Prism syntax highlighting are deferred for long messages, with file diffs and Git status streaming into the inspector incrementally as tools run. Smooth scrolling and snappy panels even on multi-thousand-line transcripts and untracked-heavy repos.
Cross-platform parity: Windows path handling rebuilt end-to-end — project trees, file previews, and diffs work correctly on non-home drives. Terminal copy/paste shortcuts now work on Windows and Linux, and packaged builds no longer leak the developer's machine path.
Networking and login hardening: HTTPS proxy testing uses a real CONNECT tunnel, SDK traffic routes through an always-on local bridge, and login resolves against the bundled claude binary's absolute path. Activity panel polish, single-indicator sidebar, no more sleep-wake header flicker, and sharper terminal contrast in light theme round out the release.
Inspector Pane, Agent Role Colors, and Connection Status Redesign
Inspector Pane: a brand-new side panel in every conversation with three tabs — Review for full inline diff viewing with syntax-highlighted additions and deletions plus direct git staging, File for previewing any project file with syntax highlighting and line numbers, and Widget for interacting with rendered widgets in a dedicated preview surface.
Agent role colors: spawned subagents now display distinct color-coded badges by role — planner, frontend-developer, code-reviewer, test-runner, and more — across tool call blocks and the Activity Panel. Instantly see who is doing what in your agentic workflow.
Background process tracking: enhanced real-time visibility into background processes spawned by agents. Track lifecycle, view output, and stop processes from the Activity Panel with one click.
Connection status redesign: the header connection indicator has been completely rebuilt with richer state transitions, detailed provider binding info, and clear error surfacing when connectivity drops.
Multi-Provider Model Support and System-Aware Proxy
Multi-provider model switching: plug in any Anthropic-compatible provider and switch between them mid-conversation. First-class presets for Anthropic Claude, Z.AI and BigModel (GLM), Moonshot and Kimi, and MiniMax — each with its own API key, base URL, and fully user-configurable model catalog. Stop being locked into one vendor.
Per-provider usage dashboard: click through directly to each provider's usage page to monitor quota and billing without leaving Onevium. Rate limit info surfaces inline.
Provider model sync: model catalogs sync from each provider's API in real time with availability checks. A unified model picker — no more stale hardcoded lists.
System-aware network proxy: auto-detects your OS-level proxy on macOS, Windows, and Linux. Dedicated Network Proxy settings page with HTTP, HTTPS, and SOCKS5 support, per-mode authentication, and a built-in connection tester. Credentials encrypted at rest via local secret keyring.
Tool detection: Onevium now auto-detects which developer tools — IDEs, shells, CLIs — are installed on your system and offers one-click launches.
Process control: Activity Panel processes section rebuilt with proper lifecycle tracking, cleaner stop semantics, and graceful recovery when the window closes. Background process spawn and kill is fully observable.
Schedule dialog redesign: Create Schedule form restructured for clarity — cleaner layout, better validation, smarter defaults across all recurrence patterns.
Connection status indicator: header connection pill now surfaces detailed state transitions, provider binding, and error reasons.
Full i18n coverage: every remaining hardcoded string across auth forms, activity utilities, CLI settings, provider pickers, and schedules now flows through the translation layer.
Steer the AI Mid-Thought, Built-in Terminal, and Real-Time Activity Feed
Steer — redirect your AI mid-stream: don't wait for the response to finish. Drop a new instruction while the model is still typing and watch it pivot in real time. Queue up to 10 steered messages with full recovery if the stream interrupts. Vibe coding, but you're actually in control.
Built-in Terminal: a full xterm.js terminal lives inside Onevium now — multi-tab, resize, spawn and kill processes without leaving the window. Zero context-switching, zero alt-tabbing. Your entire dev loop in one surface.
Activity Panel: a live feed of everything your AI agents are doing right now — running tasks, spawned subagents color-coded by role, background processes with one-click stop, and active monitors. Think of it as your agentic mission control.
Memory Explorer: browse, search, and manage every memory your AI has formed — filterable by project, categorized with confidence scores, server-side paginated. Full transparency into what your AI actually remembers about your codebase.
Message Minimap: a VS Code-style navigation rail on the right edge of long conversations. Hover to preview, click to jump. Never scroll-hunt for that one message again.
Cross-platform IDE launch: open projects in Cursor, VS Code, IntelliJ IDEA (including Community Edition), WebStorm, or Sublime Text — fully supported on both macOS and Windows.
Proxy and network settings: configure HTTP and SOCKS5 proxies directly in the UI with a built-in connection tester. Ship from behind any firewall.
Image lightbox overhaul: zoom, pan, and rotate inline images with smooth gestures. Widget diagrams and charts now export to PNG correctly.
Dozens of under-the-hood fixes: steer message ordering race conditions resolved, SSE stream-closed detection with auto-retry, channel bot keepalive hardening, and memory consolidation for long-running sessions.
AI Autonomy, Buddy Companion, and Platform Hardening
AI Autonomy System: memory extraction, tool assessment, and autonomous configuration with daily token budget controls — let Onevium learn and adapt on its own.
Buddy Companion: a persistent AI companion that evolves alongside your usage — gains stats, levels up rarity, and develops a unique personality over time.
In-conversation search: press Cmd+F (Ctrl+F) to instantly search through your chat history within any session.
Auto permission mode: intelligent SDK-level tool classification that automatically approves safe operations while prompting for risky ones.
Widget capture and watermark: offscreen rendering pipeline for capturing and sharing widget snapshots with branded watermarks.
Channel bot hardening: improved bot lifecycle management with running-count badges in the sidebar.
Activity and usage insights: redesigned Activity page with cost savings estimation based on official Anthropic pricing.
Window state persistence: Onevium remembers your window size, position, and always-on-top preference across restarts.
Soft-delete sessions: deleted sessions are archived instead of permanently removed, with SSE keepalive and idle timeout auto-retry.
License and auth fixes: offline license validation, AES-GCM token encryption, CORS policy corrections for desktop and web origins.
Integrity signing: Ed25519 build-time signing for tamper detection on all distributed binaries.
Windows stability: fixed startup crash (React error #310) and Chrome Bridge named pipe communication.
Browser automation: navigate pages, fill forms, scroll, click, and capture screenshots through natural language — no code required.
Scheduled workflows: set up recurring agents that run on cron schedules and pipe results back to any channel or log surface.
Channel bots: deploy Onevium workflows directly into Feishu, DingTalk, and Discord so your team can trigger and receive reports without leaving their tools.
Visible tool control: every file read, shell command, and browser interaction is shown in real time with approval modes you can lock per session.
Claude Code SDK: the entire automation engine runs on the Claude Code SDK, giving you state-of-the-art reasoning behind every workflow step.