Repository navigation
Mysti 2.0.3 · Beta 4
Your agents, the way their own CLIs run them.
Beta 4 fixes the issues behind slow, silent or stuck chats in Beta 3, found by
driving the real editor with real settings. If you use Claude Code or Codex in
Mysti, update.
Full changelog ·
Production-readiness checklist
What changed
Agents run natively
- No more restated modes. Mysti used to append its Plan, Read Only and
edit-mode instructions to every message. Claude Code, Codex, Copilot,
Cursor and Gemini already enforce those modes themselves, so Mysti now only
adds instructions a CLI cannot enforce. On Claude at maximum effort, the
extra text made a plain "hi" think about five times longer. - Questions are answered in the same turn. When Claude or Codex asks you
something, your choice goes straight back to it and it carries on, instead
of failing the question and starting a new turn. Kimi no longer answers its
own questions with the first option. Secret Codex questions are never shown
as plain cards. - You can see thinking. Claude shows "Thinking… ~N tokens" while it
reasons; Copilot's and Cursor's reasoning is shown. - Plans work. Claude's own plan mode can explore and write its plan in
Plan and Read Only tiers, so plans reach the approval card. Codex uses its
own Plan mode. - Less repeated context. A resumed session no longer gets its own last
exchange replayed as history, and Cursor gets the conversation history it
was missing.
Never stuck
- Move on any time. If an agent's question is waiting, just send your next
message (after switching model, agent or mode, too): the question is
dismissed and your message is sent. - Skip means skip. Skipping a question or plan no longer sends a message
or triggers a reply. - Cards wait for you. Unanswered permission, question and plan cards now
wait, like Claude Code, instead of being denied after a minute.
Current models by default
- Claude defaults to Sonnet (latest): the newest Sonnet your Claude Code
knows (Sonnet 5.5). Sonnet 5.5 and Haiku 5.5 are in the model menu. - Codex defaults to GPT-5.6 Terra, Codex's own current default.
ChatGPT accounts currently reject GPT-5.6 Sol, GPT-6 Astra and GPT-5.5;
if Codex reports that a model is not supported, pick Terra or Luna. - Qwen defaults to
coder-model, the only model Qwen OAuth accepts. Cline and
OpenClaw use the model configured in those CLIs.
Install
Update from the
VS Code Marketplace
(pre-release channel), or download the VSIX from the
GitHub release
and run Extensions: Install from VSIX…. Conversations and settings carry
over from Beta 3. Settings you changed yourself keep your values; a model you
previously picked stays selected until you pick another.
Mysti remains beta. Known gaps, including providers that run only with full
access and headless plan modes that can be exited, are listed in
production readiness.