Skip to content

Cody v0.52.0

Choose a tag to compare

@github-actions github-actions released this 07 Oct 00:09
· 33 commits to main since this release

Built for omp 18.7, plus per-device trust and scheduled sending.

New with omp 18.7

  • Side questions: type /btw , or use "Ask a side question" while
    the agent is working. The answer streams into its own panel and never
    enters the chat; follow-ups, Cancel, Copy and history are kept per chat.
  • Goals are now the engine's own: /goal [--budget 200k] shows a
    Goal panel with tokens used against the budget, time, Pause, Resume and
    Drop. "Keep working automatically" lets the agent continue toward the goal
    by itself (saved for every chat).
  • Stop gives queued messages back: anything you sent during a run that the
    agent had not read yet returns to the message box instead of running
    after you pressed Stop.
  • Fork from here on any reply, and Duplicate chat in a chat's menu and in
    the command palette.
  • Slow mode in the model menu (low-priority lane on Claude subscriptions,
    flex tier on OpenAI), and a plain line in the usage popup when an account
    is over its limit and on a slower or wrap-up allowance.
  • A chat whose saved model no longer exists opens with a notice and a model
    picker instead of failing to start; the message you sent waits and is
    delivered once you pick. omp 18.7 refuses to reopen such chats on its own.
  • models.yml editor: model kind (image, speech, embedding, judge, ...), the
    runner APIs that serve them, and per-provider/model stored Responses
    chaining.
  • Fixed for 18.7: new chats showed "(no messages)" in the sidebar and never
    got an automatic name, and a pop-up repeated the engine's internal tool
    list every time a chat started. Only MCP tool changes raise a pop-up now.
  • New engine settings labelled: terminal-only display options are chipped,
    and the subagent completion-estimate setting notes that it never runs in
    Cody. Verified against omp 18.7.0.

Devices and scheduling

  • One question per device, in the chat: Allow or Deny, with "Remember this
    device". It replaces the shell-access switch, per-operation approvals and
    typed override codes. Forget it from the device card or Settings.
  • Scheduled send: the Send button's arrow sends now, when the chat's model
    quota resets, or at a time you pick. Agents can schedule messages too.

Smaller image

  • The container no longer ships the 730 MB web build cache it never used,
    so the image is about 0.7 GB smaller on disk.

Full Changelog: v0.51.0...v0.52.0