Skip to content

v0.17.0

Choose a tag to compare

@github-actions github-actions released this 27 Oct 09:55
· 229 commits to main since this release

Added

  • /usage shows what is left of your subscriptions. Until now the first
    you heard of a plan's limit was a refused turn. /usage opens a page with a
    section for every subscription you are signed in to, the one in use first:
    ChatGPT's 5-hour and weekly windows, its credits and the limit resets you
    can redeem, GitHub Copilot's monthly allowance (AI credits or premium
    requests, chat, code completions, each counted in its own unit, and a Free
    seat named Free), and ClinePass's 5-hour, weekly and monthly windows. Each
    limit is a bar of the share used, green until it turns amber at 75% and red
    at 90%, with the counts behind it and when it resets, both as a countdown
    and as a local time (resets in 1d 23h · Tue 2:36 PM). A spent limit says
    limit reached in red on its own row, and a paused ClinePass window says
    what to switch to. The highlighted subscription's own usage page is linked
    under the list, and c copies it. A Console sign-in has no plan, so its
    section says it is billed per token, without asking the provider anything.
    The page opens mid-turn, says how fresh its numbers are, re-reads every
    minute while it is open and at once with r, keeps the last numbers if a
    refresh fails and shows them at once when you open it again, and fits a
    short terminal by folding the other subscriptions to one line each.
    API-key providers bill per token and are not listed, and with no
    subscription signed in /usage says so in a toast (docs/usage.md).
  • An interrupted turn now shows what it cost. Pressing Esc mid-turn used
    to leave only the red Conversation interrupted line, and the status
    line's running time and token count vanished with it. The summary a
    finished turn ends on now appears under the notice:
    Worked for 8s · ≈24.3k tokens. The ≈ means part of the count is the
    app's own estimate: a provider reports a request's usage only when the
    request finishes, so the one you interrupted never does. With nothing
    estimated, as when you stop a running command after the model's request
    finished, the number is the provider's own and has no ≈. A ! command
    and an interrupted /compact still end without one (docs/interrupt.md).

Changed

  • Sessions start in auto mode. The permission mode used to start at
    manual, which stopped the turn for every file change and every command.
    A new session now starts in auto: file changes run without asking, and a
    command your allowlist doesn't cover goes to the auto mode classifier,
    which runs it, blocks it, or asks you when it can't decide. Each command it
    reviews is one request to your provider. The footer's right edge says
    auto, and Shift+Tab still cycles the modes; from auto, one press is
    master and two are manual. A project where you picked edit, auto or
    master keeps it. Earlier versions never saved manual, though, so a
    project left in manual starts in auto now; switch it back with
    Shift+Tab and it stays. To keep asking everywhere, launch with
    ALTER_ZERO_PERMISSION_MODE=manual. It takes any mode name, sets the mode
    for that launch, and is never saved (docs/permissions.md).
  • A long command's header stops at two rows. A Bash(…) header shows up
    to 160 characters of the command, and in a terminal narrower than about
    ninety columns they spilled onto a third row that held only the end of the
    cut, often part of a word (scripts/parallel_se…)). The header now ends on
    its second row with …). Wider terminals show what they did before, and
    Ctrl+O still shows the whole command. Every tool's header follows the same
    rule, so a long file path or MCP call is cut the same way (docs/tools.md
    Long headers).

Fixed

  • A thought cut short by Esc keeps its duration. Pressing Esc while the
    model was thinking settled the thought as Thought for 0s, however long it
    had been thinking. It now says how long the model actually thought.
  • A connection lost in the middle of a reply no longer ends the turn. When
    the network dropped while a reply was streaming — Wi-Fi changing, a VPN
    reconnecting, a laptop waking from sleep — the turn died on request failed: request or response body error: error reading a body from connection: Connection timed out (os error 110), or a reset, however far
    into a task it was. Now the reply starts over: what had streamed of it is
    taken back off the screen, a toast says the connection was lost, the status
    line shows retrying 1/3, and the request is sent again, waiting for the
    network first if it is still gone. The reply then streams once, whole.
    Subagents start over the same way, and hand back only the reply that got
    through. A reply that keeps dropping still stops after the attempts the
    /settings Error retry row allows (docs/offline.md).

Full changelog: v0.16.0...v0.17.0