Scarica MicroClaw 0.30.0 per Windows

MicroClaw Code closes four everyday gaps found by comparing it with Claude
Code, Codex CLI and OpenCode
(docs/roadmap/mccode-competitive-plan-2026-11-cn.md, N1–N4).
Added
- Ctrl+B moves a running shell command to the background. The agent's
bash call returns at once with the process id and the agent carries
on; it reads the output later with bash_output and stops it with
bash_kill. /jobs (terminal and Zed) lists background commands and
sub-agent runs, /jobs <id> shows the latest output and /jobs kill <id>
stops one, also while a turn runs. Foreground commands in a workspace now
run through the process registry for this (stdout and stderr arrive
interleaved, as in a terminal); a cancelled run still stops its command.
SDK: MicroClaw::background_running_commands, background_jobs,
background_job_output, kill_background_job.
- /rewind also rewinds the conversation. Going back to before a turn
restores its files and cuts the conversation back to where it stood when
the turn started, so the agent no longer remembers work that was undone.
The timeline (Esc Esc) takes Enter for both, f for files only, c for
the conversation only and e to just edit the prompt; /rewind <id> files|chat|both and Zed's /rewind N [files|chat] do the same. After a
files-only rewind, /undo or a conversation-only rewind, the next prompt
tells the agent that the files and the conversation may disagree. The
whole conversation is archived before it is cut, and a conversation that
was compacted since the turn is left as it is. SDK:
MicroClaw::session_mark and rewind_session.
- Sub-agents get their own model, and project agents can run as
sub-agents. model.roles.subagent (engine aux_models.subagent, SDK
MicroClawBuilder::subagent_model) now sets the model of sessions_spawn
runs and explore. sessions_spawn takes agent: <name> to run a task
as a project agent from .microclaw/agents/<name>.md with its own
instructions and model. A new tools: front-matter field (tool names
or read, write, edit, shell, search, web) narrows the tools
such an agent may use, for explore as well; it can never widen them.
- The approval reviewer can approve. permissions.reviewer: approve
lets the reviewer model approve, in ask mode, the high-risk calls it
rates SAFE (shell commands, writes outside the project, web fetches, MCP
tools); everything else still asks you, destructive commands and ask
rules always do, and a failed review asks too. The turn's receipt lists
what the reviewer approved and why. advise is the former
--review-commands behaviour (the flag still selects it), and the
reviewer now runs on model.roles.review.
Changed
- The side panel moves from Ctrl+B to Ctrl+] (Ctrl+5 in terminals without
the kitty keyboard protocol); rebind app.sidebar / app.background in
tui.keymap to change either.
- The approval reviewer's prompt says that a SAFE verdict may let a call run
without asking a person, so it keeps SAFE for clearly harmless calls.