Skip to content

v0.7.0

Choose a tag to compare

@he-yufeng he-yufeng released this 24 Sep 18:00
· 19 commits to main since this release

Twenty-three commits since v0.6.0, still in the distilled spirit of the project:

  • Thinking models show their work (745ff65): deepseek-reasoner, kimi-k2-thinking and friends stream reasoning_content next to the answer. CoreCoder now surfaces it as it arrives instead of dropping it, and keeps it out of the reply and the history so context stays clean.
  • Streams retry wholesale, not just at connect (745ff65): a stream that died mid-drain used to kill the whole round; only the connect was retried. The drain is now inside the same backoff, so a dropped connection or a 5xx partway through re-issues the request. A 4xx still raises immediately.
  • Windows is a first-class shell citizen (8a44515, a18a4ac): command spawns route through Git Bash and drive-letter paths are posixified before they reach it. The Windows CI lanes are green again, and lint agrees.
  • cwd tracking got honest (b96a3b7, 2c49e74, 9c7fc18, ea46fd8): the shell watcher follows cd chains across semicolons, never inside subshells, quote-aware, with backslashes verbatim and Python 3.10 intact. Parallel bash calls share one cwd, and same-file edits serialize instead of racing.
  • Token estimates tier by content type (d74fdbf): code, prose and markup no longer price like the same text.
  • Direct OpenAI-path param adaptation (8990203, 66804de): rejected sampling params are adapted rather than fatal, and the tests track openai 3.14's httpx2 switch.
  • /undo recreates a deleted parent tree (27977ee), with a hooks and checkpoints audit behind it.
  • Reading track: essay 08 on the v0.6.0 extensibility trio is live in both languages, the hooks section grew two copy-pasteable cookbook examples, and the first-screen demo shows plan mode and hooks as they are now.

Full essay series and the code map are in the README. As always: readable end to end, hackable, and the gaps are still the point.