v0.9.15

[bxct]

Highlights

• Ukrainian throughout the agent screens. The objective chat, Debug Console, connection banner, task terminal, agent selector, and messages sent by the server now follow the project language. This includes start failures, watchdog notices, and image-paste errors. English text is unchanged.
• Finish button for tasks an agent left unfinished. When an agent's terminal ends while its task is still in progress and the required commit is missing, the task is marked urgent and gets a comment. Its card also shows a Finish button, which relaunches the agent to commit and complete the task. The flag survives reloads, other windows, and app restarts. Idle agents are also reminded to commit and complete before they stop.
• Session limit that grows with your machine. A new default "ramp" admission mode raises the number of agent sessions one slot at a time while memory stays quiet. It drops back immediately when memory pressure appears. Queued tasks show how many sessions are running, the current limit, and when it may rise or why it is being held.
• Claude usage stays visible. The sidebar usage rows no longer go blank when the usage service is briefly unavailable or rate-limited. They show the last good reading, faded, with its time. That reading is kept across restarts. The app also checks the usage service less often to avoid being rate-limited.
• Better support for Pi models without tool calls. Custom endpoint models can be marked as not supporting tool calls, and such models show a "No tools" badge. Task and chat pickers hide these models. Objective planning still works with them, and turns that hit a tool-compatibility error retry without tools and show a clear message instead of a raw error.

Improvements

• Contrast across all themes now meets accessibility guidelines for accent text, focus rings, inline-edit outlines, tag chips, dependency lines, and selection rings.
• Completed and canceled cards and done To-Do rows stay readable. They no longer fade; a subtler background, dimmer text, and strike-through for canceled items mark them instead.
• The Start button on closed cards is now a clearly visible outlined button.
• The agent notice toast and several setup and chat elements now match light themes instead of showing a dark block.
• Saving a task from the New Task form or edit modal with a tag that does not exist yet now registers that tag automatically, with a short generated description.
• The New Task form keeps the sprint you picked, including Backlog and older sprints. Only a dependency may move it later, and the preview tells you when that happens.
• The New Task preview shows your project's task prefix and the key the task will actually get. It also checks dependencies on tasks outside the current board view.
• Task keys are now always assigned by the server, so tasks created by agents no longer get out-of-sequence numbers.
• Edit Agents shows the Claude Code and Codex model choices as cards with an inline editor, like Pi models.
• Re-checking an agent after unchecking all of them makes it the default again.
• Saved Pi model choices that can no longer run tasks now show a warning pointing to a model that can.
• The macOS notification banner no longer brings the Task App forward when you click Hide. Clicking a card does bring it forward.
• Merge conflict "Resolve with agent" now starts a fresh agent run when the previous session has already exited.
• Plans approved inside the agent's own terminal are now recognized, so the completion reminders and exit checks apply to those runs too.
• When an objective turn fails in the agent's command-line tool, the chat now shows a clear, localized error with next steps instead of a broken reply.
• Claude Code hooks now run in installed builds of the app. This includes knowledge-base auto-push, search result caching, and file-access tracking.
• Codex notification settings from a project's Codex config are carried into Task App Codex sessions without being lost.
• Agents are told to record follow-up tasks only for real defects, data risks, security issues, or steps you must take yourself. They comment on existing tasks before creating new ones.
• Lower background load: task changes are checked more cheaply, large server replies can be compressed, and idle project connections are released to save memory.
• Edits made while offline now appear in the window right away, and they replay correctly once the connection returns, including new tags. You are notified if the server rejects a queued change.
• When a reverse proxy reports the server as down, the app now treats this as offline and queues your changes instead of failing.

Fixes

• When Play All or bulk Start is refused by the server, you now see one message listing the refused tasks. Those tasks return to their previous status and agent, including starts that were refused after waiting in the queue.
• A single task start that the server refuses now returns the task to its previous status.
• Error messages are more accurate. "Network error" now appears only when the server could not be reached. Server rejections, such as a failed save or a comment you are not allowed to edit, now show their own reason.
• Opening a task you cannot access no longer starts a session or opens an empty objective tab.
• Watchdog memory figures now appear in whole MiB instead of long decimals.
• The Trello import error message now follows the current theme.
• The bulk action bar no longer hides behind expanded cards.
• The setup Agent step no longer lets you continue with no default agent. It also no longer quietly saves Claude as the default when you never picked it.
• Reading Claude usage is more reliable: there is more time for a slow start, and saved credentials are found more often on macOS.
• Reconnecting after the app starts while the server is down now retries properly.

---

Installers for macOS (Intel and Apple Silicon dmg), Windows (exe) and Linux (AppImage, deb and pacman) are attached below, with a SHA256SUMS file covering every installer.

---

This discussion was created from the release v0.9.15.