The stable release of v6. Everything in the 6.0.0-beta release ships, plus the fixes from beta testing on Windows and on a real project (a landing page built with every v6 command). Upgrading from 6.0.0-beta needs no migration.
Fixed since 6.0.0-beta
- The Stop hook no longer calls the model or blocks the end of a turn. It was a
prompthook that asked Claude to applynaksha-memory-updateblocks and suggest a follow-up, and its judge could keep re-rejecting the same reply, holding a turn open indefinitely. It is now a deterministic Node script (scripts/memory-apply-hook.js) that does whatdocs/naksha-memory-schema.mdalways described: reads only new transcript lines, takes blocks only from the assistant's own text (not tool output or code fences), validates paths against the ProjectBrain fields, writesproject.jsonatomically and logs problems to~/.naksha/hook-errors.log. No more token cost per turn. Follow-up suggestions stay in each command's "What's Next" section. - Updating a local clone now works as documented. Claude Code runs its own cached copy of the plugin and refreshes it only on a version change, so
git pullalone was never picked up.install.shnow reinstalls (uninstall --keep-data+install) when naksha is already installed, and the README saysgit pull && bash install.sh. - Commands Claude runs itself now count.
/naksha-stats, the transcript scanner and the dashboard hook treat a Skill tool call fornaksha:<command>as a run of that command, so model-invoked commands get their own cost line, turn and org-chart staffing (they were 100% "unattributed" before). - Naksha Live shows the data it was missing.
/design-reviewverifier agents light up the owner roles of their findings; the findings panel lists confirmed and needs-validation findings and counts rejected ones (it used to show rejected findings as issues); the cost panel no longer times out on large transcript histories (/naksha-statsnow caches each transcript's parse in~/.naksha/cache/— counts and ids only, never message text — and the route answers "pending" instead of failing after 8 s). - The browser self-check probe is more accurate. It disables CSS transitions while measuring (no more mid-animation contrast failures after a theme change), stops flagging
:focus-visible-only focus rings as missing, and reports content clipped insideoverflow: hiddenboxes. /design-detectand the linter. Lints every UI file when the folder is not a git repo (it used to say "nothing to lint"); reads tokens from@import-ed and<link>-ed local stylesheets; skips themedia-queriesrule for tokens-only files; text mode no longer leavesdesign-lint-report.jsonin your project (only in CI mode or with--report).- Naksha Live wears the naksha-studio brand. The dashboard uses the site's colours, type, blueprint mark,
naksha-studiowordmark and favicon, with AA text contrast in both themes. The Votes panel no longer injects a<style>element, which the strict CSP blocked (the console error on load is gone). Project memory summarizes learnedtaste.observationsinstead of saying "No taste signals yet". - The website tells the v6 story.
site/index.htmlgains a "New in v6" section, correct stats (16,000+ lines, 6 platforms including Autohand Code CLI), the full two-step marketplace install, accurate update and privacy answers, a<main>landmark, AA contrast in both themes and an install-tab row that fits at 375 px.guard-counts.shnow also catches15k+-style line claims. - Windows. A
.gitattributeskeeps LF line endings on every checkout, so hook scripts and generated adapter regions work under Git for Windows' defaultcore.autocrlf.install.shreads the version withnode(it wrote 4.8.0 withoutpython3) and links the plugin with a directory junction on Git Bash. The quality scripts no longer needpython3. Tests that need POSIX symlinks skip on Windows with a reason.
Full v6 feature notes (Naksha Live dashboard, cost tracking, design review verification, design-detect, votes and taste): see the v6.0.0-beta release and CHANGELOG.md.
Upgrade: from a local clone, git pull && bash install.sh. No migration from 6.0.0-beta.