본문으로 건너뛰기

0.2.22

번역 준비 중

이 쪽은 아직 한국어로 옮기지 않았습니다. 아래는 영어 원문입니다.

Behavior changes​

  • A shared run graph that a peer edited no longer runs on the owner's machine until the owner approves that version (or saves it again); the editor run, the exec-request catalog and clone all refuse it, and the share control warns before granting edit.
  • the reviewer role can be set to a Claude subscription over ACP, falling back to the default reviewer when unavailable.
  • when a run catches up with main and only the release-note file conflicted, the run now reports it as resolved deterministically (no LLM) instead of «LLM-resolved», and the post-sync gate treats it like a clean merge.
  • the graph editor can share a graph with peers and set view or edit permission per peer.
  • the graph canvas uses one grouped node palette for workflows and harness graphs.
  • the knowledge node can use an installed knowledge pack.
  • the PWA has a plugin wizard tab.
  • Telegram opens with live runs before the /now summary.
  • /board shows the same operations board on Telegram and in the TUI.
  • control status prints a read-only seven-row board.
  • config classify reports which config keys are user, operator or internal settings, without changing them.
  • an opt-in claude-pty step can join the LLM fallback chain behind a maturity gate.
  • web push rings only for irreversible decisions or ones due within 24 hours; other decisions wait on the board.
  • ops health flags graph loops that have kept failing for over 24 hours.
  • mermaid diagrams in the PWA follow a default theme that can be changed in settings.
  • knowledge packs pass a data-leak check before they are bundled for the market.
  • supervisor answers are linked to their decision by sessionId and need an answer body.
  • a finished run's duration counts only the nodes on that run.
  • queue and immediate harness launches stamp the state directory they came from on the child.
  • the away summary takes its completed-node total from the run's graph snapshot.
  • decisions --json waits for stdout to drain, so piped output is no longer cut at 64 KB.
  • an outbound report that cannot be delivered becomes one briefing item instead of being lost.
  • the TUI no longer prints the same goal-loop completion answer twice.
  • harness policy gates also run the public export/import check.
  • the plugin validator accepts approval recipes that the graph runner already executes.
  • the loop map says the token lacks access on 401/403 instead of showing a loading or daemon message.
  • a graph's schedule trigger now starts the graph on time through the workflow scheduler.
  • graphs gain a prompt node that fills $ values into a prompt template; connecting it to your model provider in everyday runs comes in a later release.
  • the official plugin market adds image-first-shorts, a free, non-commercial plugin that makes vertical character shorts image-first (first frames, image-to-video, matting, beat-timed sound, effects, final cut) with three human approval steps; adapted with permission from Gongnyang AI Lab (credit in the plugin's SOURCE.md).
  • Claude 5.x models are recognised as reasoning-capable, so reasoning and interleaved thinking are enabled for them on the Anthropic provider.

Plus 135 internal changes.