Conductor

The conductor (orchestrator) is a chat in the side panel. Describe what you want — it starts suitable agents, hands out tasks, reads their progress and passes work between them.

Who can conduct

Claude Code, Kimi Code or Hermes. The choice applies to new chats; existing ones keep their CLI.

What the conductor can do

• List sessions and read an agent's screen
• Use free agents first and send them tasks — it only starts new ones when none is free, the work must run in parallel, or you ask for it
• Run commands (e.g. npm run dev, in the terminal panel at the bottom) and open pages in the preview
• Hand off work between agents
• Read and maintain the task board

You confirm large start requests (many agents at once) beforehand. How many agents may run in total is set under Settings → System.

Tasks, reports and file areas

Every task from the conductor gets a number (e.g. t-3f9a) and moves through fixed stages: queued → delivered → working → checking → reported. So the conductor always knows where each agent stands instead of interpreting screens.

• Agents report back themselves. Each task ends with instructions on how to report. The agent sends its report with one command: done, blocked or question, plus a summary and the changed files. Agents in chat mode report with their last answer. If an agent doesn't report, the task counts as finished without a report and the conductor checks for itself.
• File areas. When several agents work in the same project, the conductor gives each its own files or folders. OpenCognit blocks changes outside them for that agent, and two open tasks can't share files. So even twelve agents don't get in each other's way.
• Mandatory check. When an agent reports done and the project has tests, OpenCognit runs them itself (stage checking). If they fail, the output goes back to the same agent automatically, up to twice. The conductor is only woken with the result. If other assigned agents still work in the same folder, a failure is not sent back automatically, because it may be theirs. The mandatory check follows the setting for automatic tests.
• Test guard. If an agent changed or deleted existing test files since it got the task, the mandatory check names them with the result. A green result only counts once the conductor has looked at the change to the tests, so no agent can quietly turn red into green by weakening a test. New test files don't count. Under Settings → Security → Test guard you choose Report (default), Strict or Off. With Strict, changed tests count as a failure and go back to the agent — unless the conductor explicitly allows them in the task (send_task with tests: true). After an app restart, tasks delivered before it have nothing to compare against.
• Merge queue. When several agents worked in their own worktree, the conductor merges them one after another: checkpoint, apply the changes, run the project's tests. If the tests fail afterwards, that one worktree is rolled back and the others continue. You confirm once for the whole batch; nothing is committed.
• Your machine stays usable. Test runs started by OpenCognit share the machine's slots (one run per four logical cores). The rest wait their turn.
• Woken with the results. When agents finish, OpenCognit wakes the conductor with their reports. With many agents this happens in batches: whoever is done gets the next task right away instead of waiting for the slowest.
• Keeps going at the limit. When an agent working on a task reaches its usage limit, the task doesn't count as finished. OpenCognit hands it over with the same number, file area and current state: first to another account of the same CLI, otherwise to another signed-in CLI. Accounts that are at their limit themselves are skipped. The chat shows who took over, and the conductor sees it in the report. If no replacement is free, the agent waits and continues by itself once the limit resets. You can turn this off under Settings → System → Hand tasks over at the limit.
• Survives restarts. If you restart OpenCognit, open tasks aren't lost. Waiting ones get delivered, and every agent continues exactly its own conversation.

Cards in the chat

When an agent is done, a card appears in the chat: who, how long, whether the tests are green, and every changed file with +/− lines. Per file Before/after (opens the comparison in the editor, see Editor) and Open; a small preview shows the first change. Below it you decide: Accept, Refine … or Discard. While an agent is still working, its card grows live, file by file. With the chat open, the card replaces the review card above the terminals.

Models

Choose model and thinking depth at the bottom of the chat. For Kimi, new models (e.g. K3.1 Pro) show up as soon as Kimi has reported them — OpenCognit remembers them itself, your Kimi config stays untouched.

Passing on code and text

When you paste long code into the chat, OpenCognit also saves it as a file — so the conductor can pass it to an agent. Every message can be copied or sent straight to an agent with Send to agent.

Chat or CLI

With the Chat | CLI switch, a Claude conductor runs as a real terminal in the side panel — not additionally in the grid.