Skip to content
Agent Fleet

v0.7.0

OpenCode signs in with your opencode.ai account from the Console and lets you choose which plan its sessions run on, interrupted Claude turns pick themselves back up without an assistant in the loop, and AWS’s Agent Toolkit joins the built-in MCP integrations.

New

Section titled “New”
  • Sign in to OpenCode from the Console. Settings › Agents › OpenCode runs the device flow for you, and the running daemon picks the credential up without a restart. The API-key path is untouched and the two coexist.
  • Choose which OpenCode plan a workspace uses — free tier, Go, or Zen. The free tier needs no credentials, and no OPENCODE_API_KEY is injected into a workspace set to it, so it cannot quietly bill your Go quota. The card links to your usage page and shows which quota ran out and when it resets.
  • Agent Toolkit for AWS as a fourth built-in MCP integration, alongside PagerDuty, Grafana and CloudWatch. Read-only unless you turn on write tools, and the one built-in that is handed to coding sessions as well as to chat.
  • Interrupted turns resume by themselves. When a Claude turn dies on a retryable error the agent waits out a backoff and sends “continue” itself, up to twice, instead of leaving the session parked until someone notices. Nothing is sent while a turn is running or a question, plan or permission prompt is waiting. On, in Settings › Agents › Claude.
  • An agent can propose the next session — a ready first prompt and a name, saved and shown in the mirror where you can edit or discard it. Nothing is created until you start it through the normal launch dialog.
  • Start a session in a subfolder of a working copy, useful in a monorepo when the task only touches one directory. Only the agent’s working directory moves; the last folder is remembered per repository.
  • Pull the parent repo’s commits into a worktree. The worktree badge distinguishes “parent is ahead”, “can fast-forward” and “diverged”, and a right-click item fast-forwards. Diverged histories are refused rather than merged.
  • Keep a list of Claude models, which feeds the launch dialog, the default-model setting and list_models over MCP — Claude Code has no account-linked catalogue, so a full model ID used to be gone after you typed it.
  • Tab cycles reply suggestions the way shell completion does; Shift+Tab walks back, and a full turn returns what you typed.
  • Five more surface colours in the appearance palette, and Markdown gained a copy control on quote blocks and a wrap toggle on code blocks.

Improved

Section titled “Improved”
  • An expired Claude login now looks like a failure, and offers the fix — a red error block instead of the CLI’s “Please run /login” rendered as an ordinary reply, and a card that offers re-authentication instead of a disconnect-and-reconnect dance.
  • Usage limits are visible where they stop you. A managed Codex session stopped by a limit carries the blocked — needs attention badge, and model-specific Claude limits are detected instead of being reported as a completed answer.
  • Answering a multiple-choice question is two steps — select, then confirm. Options that carry a preview show it next to the choice.
  • Reply-suggestion learning treats full-width and half-width as the same, and the chip row shows up to 20 learned replies.
  • Pinned CLI versions: Claude Code 2.1.224, Codex 0.147.0, OpenCode 1.18.15, GitHub Copilot 1.0.78, Antigravity 1.1.11, Cursor 2026.08.04, Kiro 2.16.2, rtk 0.44.2.

Fixed

Section titled “Fixed”
  • A turn could vanish from the mirror for good. A long transcript line is written in pieces, and catching one half-written moved the reader’s cursor past a turn that had never been delivered — with a prompt stuck on “pending” underneath it.
  • A schedule could fire forever where nobody could see it. A schedule whose owning membership was disabled or removed kept firing as a no-op, invisible in the list and the run history. The first such skip now disables it and records why.
  • A worktree whose submodules were never fetched now says so and repairs itself, instead of sitting in a state git itself reports as healthy.
  • Tables written with full-width ｜ render as tables — a Japanese IME commits the pipe as U+FF5C, which used to collapse the whole table into one paragraph.
  • Reply suggestions were built from progress chatter rather than the actual answer.
  • OpenCode fixes: the model list is read from the running daemon, disconnecting clears the credential, an API-key change reaches the daemon, and the final answer reaches the mirror live instead of on the next refresh.
  • Mirror fixes: opening or switching a session lands at the bottom, the answer footer shows when the turn finished rather than when it started, and a " in an answer no longer truncates the display.
  • Launch dialog fixes: the phone soft keyboard no longer opens by itself, and a failed connection check is retried instead of leaving the launch button permanently disabled.

Upgrade notes

Section titled “Upgrade notes”

Nothing to migrate. The old OpenCode model list setting is converted to the new plan selector automatically (hide-zen → Go, everything else → Zen) — check the card if you had it set. Auto-resume for interrupted Claude turns is on by default. The AWS integration is read-only until you enable its write tools. Workspaces need a Stop → Start to pick up the newly pinned agent CLIs.