Skip to content

What's new in Kora Desktop

  • Kora Desktop 1.5.0. Goal Mode lets one session pursue a completion condition autonomously with bounded, provider-matched evaluation; inline /ask, /plan, /debug, and /build commands switch modes immediately, while /goal becomes a removable composer chip.
  • Working-tree Git diffs. Open a saved file’s current changes against HEAD directly from the editor or Git panel.
  • Isolated remote editor tabs. Files with the same path in separate remote sessions now keep independent tabs, changes, and save targets.
  • Safer tool conversations. Saved chat history now pairs tool calls with their real results before it reaches the active provider.

Kora Desktop 1.5.0 — Goal Mode, native toolbar, and safer execution

Section titled “Kora Desktop 1.5.0 — Goal Mode, native toolbar, and safer execution”

August 5, 2026

  • Native title-bar toolbar. The master toolbar is integrated into the desktop title bar, leaving the workspace surface available for the task itself.
  • Goal Mode. Set /goal <condition> to keep one session working until a tool-free evaluator confirms the condition, with visible progress, safe continuation, explicit stop, and terminal failure states. See Goal Mode.
  • Inline mode commands. Start the composer with /ask, /plan, /debug, or /build to switch execution mode immediately while keeping the rest of the text as your prompt.
  • Removable Goal chip. Selecting /goal in the composer turns it into a visible chip; type the condition beside it, remove the chip to cancel before sending, or submit to reconstruct the Goal command safely.
  • Context composition breakdown. Open the context usage indicator to see how the latest request is distributed across messages, system tools, skills, system prompt, MCP tools, and metadata.
  • Weighted cache visibility. The context details show the aggregate cache hit rate for completed requests instead of exposing only the latest request’s raw cache value.
  • Local-first tool routing. Shell, scripts, Git, system tools, and relative paths use the selected local folder by default; remote execution requires an explicit remote project directory.
  • Working-directory boundaries. Local and SSH command directories are checked against their allowed project roots before execution.

Kora Desktop 1.4.12 — Cleaner task rows and safer updates

Section titled “Kora Desktop 1.4.12 — Cleaner task rows and safer updates”

July 31, 2026

  • Cleaner todo rows. Task state stays clear through icons and styling without redundant status badges.
  • Malformed response protection. Chat no longer crashes when a model response contains malformed tool markup; code blocks and useful text remain visible.
  • More resilient releases. GitHub releases continue publishing with a safe fallback when version-specific public notes are unavailable.

Kora Desktop 1.4.11 — Per-agent control and safer queues

Section titled “Kora Desktop 1.4.11 — Per-agent control and safer queues”

July 31, 2026

A focused update improving subagent lifecycle management and queue safety.

  • Per-agent stop in the operational addendum. Each running subagent gets its own stop button; you can stop one agent, stop all, or clear completed/failed tasks with dedicated buttons and tooltips.
  • Partial results preserved on manual stop. When you stop a subagent, its output up to that point is delivered back to the parent — no more losing work because of an early stop.
  • Atomic send-now on queued messages. Message dispatch from the queue is now session-isolated and atomic, preventing cross-session message leaks during rapid Stop sequences.
  • Orphan tool-calls stay UI-only. Interrupted tool calls show a local error card but never leak synthetic results into the provider’s message history, keeping the AI SDK conversation intact.

Kora Desktop 1.4.9 — Operational addendum and timeout policy

Section titled “Kora Desktop 1.4.9 — Operational addendum and timeout policy”

July 30, 2026

  • Operational addendum overlay. Subagent and task management now lives in a fixed overlay above the chat with dynamic ResizeObserver-based padding, keeping the transcript clean and scrollable.
  • Active agent shimmer. While subagents are running, agent names in the addendum show a beat shimmer and stable ariaLabel — no more flickering numeric indices.
  • “N/N completed” green summary. When all todos are done, the addendum header shows a green checkmark summary instead of individual completed badges.
  • Centralized timeout policy. The Smart Token Engine now manages a clamped, persisted timeout profile (resolveTimeoutProfile) that all watchdogs consume — no more hardcoded magic numbers.
  • Provider error routing fix. 401 and 400 errors from fullStream now route to the failure handler instead of being persisted as assistant text. Verbose diagnostics preserved under [ProviderRegression].
  • Subagent lifecycle hardening. Terminal states, per-session waits with reference counting, and a stall watchdog with second-strike guard prevent false positives on active waits.

Kora Desktop 1.4.8 — Narrow sessions redesign and NVIDIA NIM fix

Section titled “Kora Desktop 1.4.8 — Narrow sessions redesign and NVIDIA NIM fix”

July 30, 2026

  • Redesigned narrow session rows. Under 320px, session actions split to a separate line — title and metadata stay uncompressed and readable.
  • SSH terminal button. Connected remote folders now show a terminal button that opens a shell in the remote root.
  • Italian-locale timestamps. User messages now display in dd/MM/yyyy HH:mm format with Italian locale; system messages remain unchanged.
  • NVIDIA NIM stopSequences fix. Introduced buildGenerationParameterOptions as a single choke point that drops falsy values (undefined, null, [], {}) before they reach the provider — fixes the "Stop sequences array cannot be empty" validation error.

Kora Desktop 1.4.6 — Kimi Code provider and LSP extension mapping

Section titled “Kora Desktop 1.4.6 — Kimi Code provider and LSP extension mapping”

July 29, 2026

  • Kimi Code provider. New OpenAI-compatible kimi-code provider with a dedicated Coding API endpoint and dynamic model discovery, plus automatic temperature normalization for Kimi-family models across providers.
  • Extension → language mapping UI. The LSP settings tab now has a CRUD interface for Monaco editor language overrides — map any file extension to any language ID.
  • Dart .dart extension added to the built-in mapping.

July 29, 2026

  • askUser crash resolved. Malformed historical arguments (null containers, missing multiSelect, invalid questions) are now validated at runtime with a local shadcn Alert fallback instead of a fatal render crash.
  • System prompt examples for askUser and askUserConfirmation aligned to the registered questions[] schema.

July 29, 2026

  • Dart/Flutter LSP support. Auto-detection via pubspec.yaml and analysis_options.yaml markers, using dart language-server --client-id ai.kora. Listed under a new Mobile section in LSP settings.

Kora Desktop 1.4.3 — Per-session skills and ByteDance ModelArk

Section titled “Kora Desktop 1.4.3 — Per-session skills and ByteDance ModelArk”

July 29, 2026

  • ByteDance ModelArk dynamic discovery. Real-time model catalog fetching with per-model toggles and live Test Connection in provider settings.
  • AI-powered commit messages. Staged diffs in the Changes panel can now generate a conventional commit message with a single click, using the active model.
  • Per-session skill isolation. Skills loaded via the skill tool are no longer shared globally — each session maintains its own active skill registry, isolated by sessionId. Active skill chips appear in the composer after send and after remount, with individual (X) and bulk (Clear all) removal.
  • Turn counter persistence fix. The turn counter now restores from the persisted value on remount, preventing STE degradation on session reopen.

Kora Desktop 1.4.2 — Inline @ file mentions

Section titled “Kora Desktop 1.4.2 — Inline @ file mentions”

July 28, 2026

  • Fuzzy @ mentions. Type @ anywhere in the composer to search files across all project folders. Tab browses directories, Enter confirms. The raw path stays in the message text; the user bubble renders a compact, dismissible chip.
  • Spaces in paths are auto-quoted with double quotes to prevent tokenization errors.
  • Hover tooltip on a mention chip shows the full path; clicking the chip restores the mention in the composer and reopens fuzzy search.
  • Isolated browsing roots via isWithinRoot — local and remote folders use most-specific prefix matching so mentions never cross root boundaries.

Section titled “Kora Desktop 1.4.1 — Multi-folder fuzzy search”

July 28, 2026

  • Multi-folder parallel search. @ mention fuzzy search now fans out across all local folders in parallel with a stale-result guard.
  • Raw <tool_call> markup from providers is filtered at the render boundary, preventing malformed XML envelopes from appearing in the chat.

Kora Desktop 1.4.0 — A chat that gets out of the way

Section titled “Kora Desktop 1.4.0 — A chat that gets out of the way”

July 27, 2026

This release reorganizes the chat surface around the message instead of around the participants. The transcript becomes the work area; the chrome fades into the background.

  • Full-width prompt rail for your messages. Your prompt fills the rail on the left with an accent-painted background, a left border in the primary token, and a single timestamp in the footer. The rail uses article semantics so screen readers treat it as one block.
  • Open canvas for assistant answers. Assistant messages render without avatar, author label, or duplicate timestamp. The toolbar uses shadcn Badge and ghost Button variants.
  • Tighter queue rows. Queued messages take less vertical space; the drag preview aligns with the message bounds.
  • Compact long-paste badge. A long paste is flagged with a Pasted – N lines badge rendered through the shadcn Button primitive, so it is keyboard-accessible.
  • Semantic markdown link colors. Inline links in the prompt rail use the text-primary token instead of raw blue or dark colors.
  • Surface-washed jump controls. The transcript jump controls follow the shadcn surface tokens for a consistent visual hierarchy.
  • Deterministic queue handoff. Natural turn completion is the single atomic boundary for advancing the queue. Manual edits and Stop no longer race with automatic dispatch.

1.3.5 — Onboarding, OTA, and Git workspace

Section titled “1.3.5 — Onboarding, OTA, and Git workspace”
  • Onboarding wizard expanded with ChatGPT, Models.dev, and Kora Cloud connection steps.
  • Deferred restart update flow — 3-choice update modal (install now, on quit, dismiss) with progress and phased install.
  • Context-aware auto selector — Smart Token Engine selects compression level based on actual context usage, not a fixed threshold.
  • Git power user workspace — dedicated view with branch management, commit history, and Monaco diff.
  • First-response timeout UX — non-blocking 6s toast at 3min (remote) / 5min (localhost) instead of hard abort.
  • Interactive deploy tool — unified SFTP/GitHub packaging with channel selection, CHANGELOG notes, and OTA stable tag.
  • Stream markdown in frame-aligned blocks — eliminates per-token re-render jank during long responses.
  • Apple credentials managed in-app via settings UI with .env persistence.