Esta é a versão atual. · Lançamento:
Arquivos de instalação 0.31.0
| Arquitetura | Tipo | Escopo | Tamanho do arquivo | Baixar | SHA256 |
|---|---|---|---|---|---|
| 64 bits | ZIP → Portátil | — | — | github.com | VirusTotal |
O link e o SHA256 vêm do manifesto comunitário do Gerenciador de Pacotes do Windows. Compare o hash do arquivo baixado antes de executá-lo.
Notas da versão
MicroClaw Code 0.31: private project memory, Claude Code plugins and idle
compaction (docs/roadmap/mccode-competitive-plan-2026-11-cn.md, N5–N7).
Added
- Private project memory. Besides the shared .microclaw/memory.md
(in the repository), trusted projects now have a private memory in
~/.mccode/projects/<repo>-<hash>/memory.md for your preferences and this
machine's setup. Both are in the prompt of every turn; the agent is told
which kind of fact goes where (a stated preference never goes to the
shared file), and writing the private file does not ask for the
outside-the-project approval. Worktrees share the main checkout's private
memory. /memory opens both files in $EDITOR (shared, private,
show). mccode import claude adds Claude Code's auto memory of the same
repository, mccode import codex the items of Codex's memories.
- Plugins in the Claude Code layout. mccode plugin add <owner/repo | git URL | path> [--plugin <name>] [--project], list, update,
enable, disable and remove. A plugin's commands (/name and
/plugin:name), agents (for explore and sessions_spawn), skills,
hooks and MCP servers work as in Claude Code, with ${CLAUDE_PLUGIN_ROOT}
resolved; marketplace repositories are followed to the plugin's own
repository. Installing lists every hook and MCP command and asks first
(--yes without a terminal); updating asks again when new commands
appear. Unsupported parts (output styles, LSP servers, …) are listed as
skipped. User plugins are copied to ~/.mccode/plugins/, project plugins
to .microclaw/plugins/ and load only in trusted projects. /plugins
shows what is in use. Engine: extra_agents_dirs, extra_skills_dirs,
extra_mcp_configs; SDK: MicroClawBuilder::plugin_sources.
- Idle compaction. The terminal UI compacts a large session (at least
half of the compaction budget) that has been idle for 90% of the
provider's prompt cache lifetime, while the cache is still warm, so the
next prompt does not write the whole conversation to the cache again.
Anthropic only (the engine reports prompt_cache_ttl in
ContextUsage); tui.idle_compact: false turns it off. pre_compact
hooks see trigger auto.
- MicroClaw Code desktop app (preview, macOS). A GPUI app in
apps/microclaw-code-desktop with onboarding, projects, sessions and a
streaming transcript over the same ~/.mccode home and Agent Engine as
the terminal UI. Build it from source with scripts/bundle.sh; it is
not signed or shipped as a download yet.
Fixed
- glob no longer scans the disk for hours: its fallback for hidden,
ignored and absolute paths walks from the pattern's literal prefix and
stops after 50,000 entries (saying so), off the async workers.
- Plan mode and read-only profiles offer activate_skill, which only reads
a skill's instructions.
- mccode -p "/unknown" reports the unknown command instead of sending an
empty turn.
- The system prompt explains the <context_budget> line, which models
otherwise read as text injected into tool output.
