Esta es una versión anterior. La versión actual es 0.31.0. MicroClaw 0.31.0 · Fecha de publicación:
Archivos de instalación 0.27.1
| Arquitectura | Tipo | Ámbito | Tamaño del archivo | Descargar | SHA256 |
|---|---|---|---|---|---|
| 64 bits | ZIP → Portable | — | — | github.com | VirusTotal |
El enlace y el SHA256 proceden del manifiesto comunitario del Administrador de paquetes de Windows. Conviene comparar el hash del archivo descargado antes de ejecutarlo.
Notas de la versión
Fixes found by testing MicroClaw Code 0.27.0 in Zed and in the terminal
with Anthropic's claude-opus-5, including image prompts.
Fixed
- Anthropic usage left out prompt-cache tokens (all products).
Anthropic's input_tokens excludes cache reads and writes, so with
prompt caching a request counted only a few input tokens. Token counts
and costs were too low, /context and the context meter showed 0%, and
mid-turn context-pressure compaction measured the wrong number. Cache
reads and writes now count as input, streamed and not, as they already
did for OpenAI-compatible providers. A final message_delta that carries
only output_tokens no longer drops the input count.
- MicroClaw Code's model pickers (Zed's model selector and the TUI
/model list) offered preset family names such as claude-sonnet,
claude-opus and claude-1m. These are not model names, and choosing
one failed with not_found_error. The pickers now offer only presets
that are real model names (ModelPreset::model_name).
- In Zed, a shell command that needed approval kept the title
"Allow bash?" after the answer. The permission request now keeps the
call's own title, the command; only an approval with no visible call is
titled "Allow ?".
- A failed run in mccode acp, such as an unknown model or a provider
error, was reported as a thought or an info notice, so the turn looked
empty. It is now part of the reply ("Error: …").
- Tool titles in mccode acp are Markdown code spans, so a path like
calc/__init__.py no longer renders as bold "init".
