All documentation

Known issues

Limitations and bugs we know about in the current version, with what to expect when you run into them.

Checked against the source on September 27, 2026.

These are problems we know about in the version these pages describe. Each one names what you would notice and where it comes from, so you can plan around it until it is fixed.

More than one window

  • Terminal output goes to the first window. Main sends terminal output to the first open window only, and any window loading or reloading detaches every terminal panel. Terminals in a second window can stay blank, and opening a window can stop live output in the others until you reopen the panel. (apps/desktop/src/main/pty-host.ts)
  • Replies reach every window. The core posts every reply to every window, and each window numbers its requests starting from 1. When two windows have a request with the same number in flight, one can receive the other's reply. It is rare, and reloading the window fixes the display. (apps/desktop/src/core/index.ts)

Crashes and restarts

  • A crashed core is not reported. If the core process exits unexpectedly, windows are not told. Requests wait forever until you reload the window, which starts a new core.
  • Agents can outlive a crash. Agent CLIs run in their own process groups and OpenOrc does not record their process IDs, so after a crash they are not cleaned up on the next start.
  • Stop does not clear the queue. Pressing Stop ends the turn, and the next queued message is then sent. Queued messages are kept across restarts: those still waiting are sent when OpenOrc starts again, which starts the agent. A message that was being delivered when OpenOrc stopped is not sent again on its own; it waits in the queue with Retry.

Conversations

  • Switching back to a provider skips what happened in between. In a project conversation, returning to a provider that already worked in it resumes that provider's older session, with no summary of the other provider's work. Ask it to check the files, or start a fresh session. Workspace conversations start fresh with a short summary instead.
  • Forks remember too much. Forking from an earlier point shows the conversation up to that point, but the agent's session continues from the parent's latest state, so it may know about later messages. A fork also shares the parent's folder and branch.
  • Sending mid-turn fails for OpenCode. OpenCode cannot receive messages during a turn, but pressing Enter still tries to send into the running turn and shows "OpenCode is still answering." Use the Queue button instead.
  • Schedules need the app open. Schedules only fire while OpenOrc runs, and missed firings collapse into one.
  • Long team and task histories load all at once. Ordinary conversations load a few turns at a time, but team conversations and task activity load every event, which can be slow for long histories.

Tasks and Git

  • Apply to local checkout cannot be used. The button appears only on team tasks that have their own worktree, and it is disabled there, because team changes are applied from the team conversation. Other tasks work in their conversation's worktree and do not show it. (apps/desktop/src/renderer/src/panels/ReviewPanel.tsx, packages/core/src/services/task-checkout.ts)
  • A task's base branch can be ignored. When a task starts in the conversation it was created in, and that conversation already has a worktree, the task works on that worktree's branch, not the Base branch chosen for the task. (packages/core/src/services/thread-task-execution.ts)
  • A failed setup script is not retried. If a project's setup script fails, the worktree is kept, and the next start uses it without running the script again.
  • Terminal shells outlive their conversation. Deleting a conversation does not close its terminal, and there is no limit on open shells.

Storage

  • History does not expire. Conversations, their events, attachments, and the audit log grow until you delete them, and there is no size limit. Raw provider output is the exception: its log files in logs/provider are deleted after 14 days, oldest first once they pass 512 MB, and any raw output kept in the database is deleted after 14 days. Attachments are not deleted even when their conversation is.
  • Staged task images are never cleared. Images pasted into a task document are staged in the window's IndexedDB and stay there after they are saved.
  • Redaction does not revisit old records. Text is redacted when it is written, so a secret stored before a rule was added stays until you delete the conversation.
  • Imports ignore custom CLI folders. Importing sessions reads ~/.claude and ~/.codex even if you moved them with CLAUDE_CONFIG_DIR or CODEX_HOME.

See What deleting removes for what stays after a delete.

Memory

  • Some runs are never distilled. Runs that end because you quit are skipped, extraction jobs still waiting when you quit are discarded, failed extractions are not retried, and failures are not shown in the interface. OpenCode cannot run extraction yet, so under Automatic its runs are not summarized.
  • The embedding worker does not restart. If it crashes, memory search uses full text only until OpenOrc restarts.

Slack

  • No automatic reconnect. After OpenOrc restarts, you have to press Connect again.
  • A dropped connection may not recover. OpenOrc turns off retries for Slack's Socket Mode reconnection. If the connection drops and the first reconnect fails, for example right after your computer wakes, Slack can stay disconnected until you reconnect it by hand, while the Slack settings still show it as connected. (packages/core/src/services/slack/service.ts)
  • "Also send to channel" replies are ignored. Slack marks them with a subtype OpenOrc does not accept.
  • Large approvals are long. An approval for a large edit is posted in full, in as many Slack messages as it needs.

Team preview

  • Agents share your checkout. In local mode, the default, several team members edit your files at the same time, and file claims are advisory.
  • Stopping leaves work unmerged. Stopping an execution, or reaching its time limit, does not capture or merge work in progress, and no notification is sent.
  • A restart stops running turns. Turns that were running when OpenOrc stopped are marked as needing attention and wait for you to retry them. Retrying one does not count against the retry limit, and turns that had not started run on their own.
  • Folders and refs accumulate. Assignment worktrees, merge scratch folders, and checkpoint refs are kept until the team conversation is deleted.

More limitations are listed on Agent teams.

Windows

  • OpenOrc does not probe a login shell on Windows, so it finds agents only through the OPENORC_CODEX_BIN, OPENORC_CLAUDE_BIN, and OPENORC_OPENCODE_BIN environment variables.
  • Claude's Plan and Accept edits modes are unavailable on Windows.
  • Project setup scripts run with /bin/bash, which Windows does not have.
  • Stopping an agent signals only its main process, not the processes it started.
  • Windows beta installers are unsigned.