Scarica MicroClaw 0.23.1 per Windows

Fixes found by running MicroClaw Code against real models (DeepSeek
deepseek-flash and deepseek-v4-pro) on the eval cases and by hand.
Fixed
- Approvals resumed later were not understood: answering a paused
approval with 1/2 or 3 in a later turn (mccode --session <id> -p 1,
the MCP run tool, and the Telegram / Discord / Slack / web option cards)
reached the model as a bare digit, which some models did not read as an
approval. The reply is now rewritten into an explicit instruction to retry
(or not run) the paused call.
- Compaction never ran with small context windows: the verbatim tail kept
by a Code compaction (24k tokens) did not shrink with the target, so a
session on a small window (a local model, an eval case) could not be
compacted at all and ran into a context overflow. The tail is now capped at
half of the compaction target.
- DeepSeek V4 windows: deepseek-v4-pro and deepseek-flash
(deepseek-v4-flash) are counted as 1M-token models instead of 128k.
- Language-server diagnostics:
- pyright never published diagnostics because MicroClaw Code did not send
workspace/didChangeConfiguration;
- a server on PATH that cannot start (rustup's rust-analyzer proxy
without the component) was shown as a passing check; it is now reported
once with an install hint and skipped;
- a server that sends nothing in time is reported as not checked;
- when diagnostics were the only check, a fix round that made them pass
still ended -p with exit code 1.
- @image mentions followed by punctuation (@shot.png?, @shot.png。)
were not attached.
- mccode export failed without a terminal ("the interactive interface
needs a terminal").
- --worktree summary suggested git merge while the changes were still
uncommitted in the worktree; it now says to commit them first.
- Repositories without a .gitignore: review lists at most 100
untracked files (the rest counted per directory), and the per-turn change
fingerprint no longer reads every large untracked file.
Changed
- mccode eval: min_compactions in case.yaml fails a case that did not
compact; the long-session cases use windows above the fixed prompt size so
they really compact, and the compaction count no longer races the run.
Security
- hyper 1.12.0, which drops the unmaintained pin-utils
(RUSTSEC-2026-0441, published on release day).
Full Changelog: https://github.com/microclaw/microclaw/compare/v0.23.0...v0.23.1