Changelog
What changed in each release, newest first.
New
- The sidebar can be collapsed to an icon rail with the new button at its bottom. The choice is remembered in this browser.
Changed
- With two workspaces open, opening another one replaces a workspace that has no running agent. If both have running agents, nothing is closed and a dialog tells you to wait or stop its agents.
- A sub-agent can run Claude Sonnet 5.5 as well as Opus 5.5. The director picks the model per task; if it does not, low or medium effort runs on Sonnet 5.5 and high or xhigh on Opus 5.5. Verifiers still run Opus 5.5. The Team tab shows each sub-agent's model next to its effort.
Changes since v2026.11.09-1.
New
- Every agent has a board: its backlog, on the new Board tab. Add, edit, move and delete cards in four columns. A card shows its number and title; click it to read the whole card and act on it. The agent can add cards too. A card handed to a sub-agent shows that task and moves by itself, to Done once the branch is merged.
- A proposal is a board card that waits for you. Approve it and it moves to To do; reject it and it moves to Dropped. You can give a reason either way. You can also tell the agent in chat to approve or reject one; the card then says the decision came through the agent. No agent decides on its own, starts work on a waiting proposal, rewords it, or brings back one you rejected. Waiting proposals are counted on the Board tab and listed on Home.
- Planner: switch it on per agent on the Board tab. The agent writes a short product brief, which you can edit, and the planner searches the web every 7 days (1 to 90, your choice) for ideas that would improve the product. It files at most 3 proposals per run and never repeats one. Scout now starts a run at once. While a run is going, the agent's card shows a "scouting" badge and the panel shows the proposals filed so far.
- A planner run is sealed, because it reads web pages nobody checked. It sees only the brief, your card titles and your past decisions. It can search, fetch pages and file proposals; it has no files, no shell, none of your other MCP servers, and a spending limit per run. It works in an empty folder that is removed when the run ends.
- Click a workspace group's name in the sidebar to fold or unfold it. A folded group shows how many workspaces it hides, and a dot if one of them is open. Each browser remembers which groups are folded.
- Once a sub-agent's branch is merged, Donezo removes its worktree and branch. It checks at startup, after each task, and every 10 minutes. Unmerged work, worktrees with uncommitted changes and running tasks are never touched. The Team tab marks those tasks merged · cleaned up.
Changed
- A message for an agent whose terminal is open is now typed into that terminal once its turn ends and you stop typing, instead of closing the terminal. The card shows how many messages are waiting.
- Opening a terminal while a background run is going waits and opens by itself when the run ends. There is no more reload-after-an-automatic-run step.
- The model picker offers Claude Fable 5.2 in place of Fable 5.1. It needs claude CLI 2.1.321 or later. The
fablealias already follows the newest Fable. - Sub-agents and verifiers always run Claude Opus 5.5. The director picks an effort per task (low to xhigh), and each failed verification raises that task's effort one level. A playbook that often needs a step-up starts one level higher. The Team tab shows the effort of each task.
Removed
- The per-role model pickers on the Team tab. Sub-agents no longer run on Antigravity.
Fixed
- An agent's terminal and a background run can no longer write the same conversation at the same time. Before, this could happen when the agent's card was off screen.
- A background run no longer hangs when the agent leaves a dev server or other background process running. Donezo stops what is left 10 seconds after the agent's turn ends and logs what it stopped, and the run keeps the agent's real report.
- Cancelling a run or pressing kill also stops the processes the agent started, not only the agent.
Changes since v2026.11.09.
New
- A Knowledge tab draws the agents, their memories and decisions, and who asked or reported to whom, as a graph you can drag around. It also shows what each agent looked up with
search_knowledge, and what it searched for and could not find. See Agents.
Changed
- Agents are told to try
search_knowledgebefore asking another agent for information. - Claude agents compact their context at about 300k tokens instead of near 1M, so each call re-reads less. Change it with
CLAUDE_CODE_AUTO_COMPACT_WINDOW; see Settings.
Fixed
- A Claude agent's terminal now runs at the card's model and effort, like its background runs. Switching between the two no longer throws away the prompt cache.
Changes since v2026.11.07.
New
- Context meters: each Claude card shows how full its context is and its prompt cache hit rate.
- Plan usage: the top bar shows your 5-hour and weekly Claude limits, amber from 80%.
search_knowledge: an agent can search the memory and decisions of the other agents in its workspace without messaging them.
Changes since v2026.08.12-1.
New
- The editor: a VS Code style editor card with Vim modes on every agent, with nothing to install. It downloads Neovim once.
- Editor keys: Normal, Insert and Visual modes, plus Ctrl S, Ctrl Z and the browser clipboard. F1 lists them, and
Turn off Vim modesgives VS Code keys. - Source Control lists every git repository in a project, with branch, changes and a diff for each file.
ORCH_NVIM_BINuses your own Neovim 0.12+;ORCH_NVIM_MIRRORdownloads from your own server.- Director and sub-agents: one agent splits the work, others build it in their own git worktrees.
- Independent verifier: work counts as done only after a fresh agent observes its Done means.
- Four sub-agent playbooks: feature, bugfix, refactor, investigate.
- Team tab: a pipeline board, the director's decision log, who may message whom, models per role.
- Reflect: turns failed runs into a playbook change you approve.
- Antigravity (Gemini) CLI support, with models read from
agy models. - Home inbox: approvals and failed sub-agents across all workspaces.
- Command palette: Ctrl K or ⌘ K to jump anywhere or run an action.
- Groups above workspaces; rename or delete a workspace in one step.
- Settings › Performance: what is using your CPU and memory, in plain words.
- Settings › Removed agents: delete leftover runs and transcripts for good.
- Settings › Media tools: shows which media commands agents can use on this machine.
- Tool servers register for every installed CLI, without a token on the command line.
- Folder card: browse a project's files on the canvas.
- Pick which past CLI session a card resumes, and delete old ones.
- Rename an agent; its old name keeps working.
- Shift + scroll zooms the canvas, even over a terminal.
- A row of chips above the canvas jumps to any card.
Changed
- Agent Orchestrator is now Donezo. Existing installs keep working: your data, your
.envand any service you already set up are unchanged. - New dashboard design: left sidebar, new fonts, Settings as a full page.
- Activity is one timeline of messages and runs, with live transcripts.
- Agent details moved into a tabbed panel: Overview, Work, Settings.
- New agent is a three-step dialog.
- Every card is a terminal card.
- Safer defaults: the server listens on your machine only, and CORS is off.
- Requests from other websites are refused.
- Claude model lists updated to the current models.
- Agents are told to signal only after their own turn is finished.
- Team tab loads 10 cards and decisions at a time as you scroll.
Removed
- Codex CLI support.
- VS Code cards. Use the folder card instead.
- The API key setting: this is a local app, reachable from your own machine by default.
- The prebuilt Linux download. Linux runs from source.
- Unity tooling and the
asset_fetchtool server. - The Compact context and Signal buttons in the agent panel. Messages go through the box at the bottom of the panel.
Fixed
- A report no longer triggers a report back.
- An agent moved to another workspace sees its new colleagues:
list_agentsnow takes only the agent's name. - Create agent is blocked while a spawn runs, so a second click cannot create a duplicate.
- A slash command in a message reaches the CLI as a command, not as text.
- The dashboard no longer slows down after hours of work.
- Ctrl-C stops the server cleanly instead of hanging.
- Selecting text in a card picks the right line at any zoom.
- A card whose transcript was deleted opens again instead of dying.
- Changing a card's model to another CLI is no longer refused.
- Icons from Nerd Fonts no longer show as empty boxes in terminals.
CLAUDE_CONFIG_DIRis respected when finding transcripts.- The dashboard can no longer serve a stale version to some workspaces.
- Windows: CLI transcripts are matched to a folder by its real path.
- Reflect and other read-only runs no longer fail on their own helper file, also in a subfolder of a repository.
- Failed Home items with no worktree have Dismiss, and an item stays as Rechecking sub-agent during Verify again.
- Double-clicking a resize handle now resets the size of the card, or of both cards in a group.