Shkarko MicroClaw 0.25.0

The MicroClaw Code interaction and configuration release. The planned 0.24
(TUI), 0.25 (settings and permissions) and 0.26 (innovations) ship together;
see docs/roadmap/mccode-ux-config-plan-2026-11-cn.md.
Added
- TUI look: theme tokens with truecolor, 256- and 16-colour fallback;
11 built-in themes, user YAML themes and OpenCode theme files; Markdown
replies with syntax highlighting; diffs with line numbers, word-level
highlights and a side-by-side layout on wide terminals; consecutive reads
and searches grouped into one "Explored" row; a side panel for the plan,
checks, changed files and running sub-agents; a per-item render cache; an
inline mode that uses the terminal's own scrollback (tui.screen: inline,
falling back to full screen when the terminal does not report the cursor).
- TUI interaction: Enter queues a message while the agent works and
Ctrl+S steers the running turn; fuzzy / and @ pop-ups and a Ctrl+P
palette; per-project prompt history with Ctrl+R search; Ctrl+G edits in
$EDITOR; ! runs a shell command yourself; large pastes become a chip;
Ctrl+V attaches a clipboard image; Ctrl+T opens the transcript with search;
Esc Esc rewinds the files to an earlier turn; model, theme (live preview)
and session pickers; mouse scrolling and click-to-expand; desktop
notifications (OSC 9/777) when the terminal is unfocused; rebindable keys
(tui.keymap, /keys).
- Per-turn receipts (files, checks, sandbox and network use, tokens, cost
when the price is known, checkpoint), hunk-level review on F2 (revert a
hunk, or turn comments into a prompt), /goal (keep going until a
condition holds and the checks pass), /btw (a side question in a copy
of the session) and /configure (describe a settings change; the model
proposes it and you confirm).
- Layered settings for MicroClaw Code: user, project, local and managed
YAML layers with a JSON schema (docs/schema/code-settings.v1.json) and
per-key provenance; mccode config show|get|set|unset|paths|doctor|init|schema|migrate|edit;
mccode import claude|codex|opencode; provider profiles that can replace
the configuration file; per-role models (plan, review, summary).
- Permission rules (allow / ask / deny for bash, read, edit,
net, mcp, agent, tool; Claude Code spellings accepted) and hooks
compatible with the Claude Code protocol (ten events; exit code 2 blocks;
permissionDecision, updatedInput, additionalContext). Both are
enforced through a new engine ToolGate (SDK MicroClawBuilder::tool_gate),
so sub-agents go through them too. ACP and MCP-server modes load the same
settings.
- Approvals and notifications in chat: with a paired Server, MicroClaw
Code sends approval cards (buttons on Telegram, Discord and Slack; numbered
options on Feishu) and "done" notifications while you are away
(notify.chat); Server endpoints /api/code/notify and
/api/code/approvals.
- SDK: MicroClawBuilder::summary_model; the ToolGate types are re-exported.
Fixed
Found by testing this release with DeepSeek (deepseek-flash,
deepseek-v4-pro):
- The shell tool ran commands in the login shell ($SHELL); zsh aborts on an
unmatched glob such as grep --include=*.rs. Commands now run in bash when
it is installed (Server and Code).
- bash_kill stopped only the shell, so a child (a dev server under
npm run dev) kept running and the process stayed "running". Background
commands now run in their own process group and are stopped as a group.
- Role models sent profile/model to the provider as the model name; an
ask rule could be answered by auto mode; -p 1 after a pause looped when
a prompt hook added context to the answer.
Changed
- Writing .microclaw/settings.local.yaml adds it to the repository's
.git/info/exclude.
- scripts/generate_docs_artifacts.mjs leaves #[serde(skip)] config fields
out of the generated defaults.
Security
- vodozemac 0.11.2 (Matrix end-to-end encryption), which restores secret
zeroization (RUSTSEC-2026-0445, published on release day).
Known gaps
Ctrl+B does not move a running command to the background (it toggles the
side panel); Esc Esc rewinds files but not the conversation; there is no
sub-agent role model; model presets carry no prices; read(...) rules do not
cover shell commands and net(...) rules cover web_fetch only.