[Lingelo]

Summary

Kaizen 5.1.0. Three ideas from the latest Compound Engineering releases (3.29–3.33), adapted to Kaizen's gates. A refreshed comparison with Spec Kit 1.2.5 and Compound Engineering 3.33.2. And fixes for what the review of this branch found in the 5.0.0 approvals: "ask" flow. No breaking change.

Added / changed

• The review flags what nobody asked for (CE #1852). The reviewer contract's intent check now runs in both directions:

  • a public surface (endpoint, flag, config key, export) that the plan's Out of scope or units rule out;
  • a new package that the standard library or an existing helper already covers.

  Each finding must be quoted, and correctness owns the check, so small diffs get it too.

• retire_when names a check that runs without changing the repo (CE #1772). /kaizen:prune-learnings runs that check itself. If the condition is met but the workaround is still in the code, the learning stays and the report recommends removing the workaround; Kaizen never removes it on its own. Also: learnings validate refuses a retire_when that is not a single string, and learnings list --retire lists the learnings that declare one.

• watch-pr asks whether an old PR still belongs (CE #1909). The new kaizen pr overlap --base <ref> reports, from local git:

  • PR files the base modified, moved or deleted;
  • folders the base removed under a path the PR adds;
  • whether the base already carries the PR's patches.

  watch-pr then gives one verdict: Update, Rework or Close. It never closes the PR itself and never rebases.

• Docs: positioning.md refreshed; pr-watch.svg regenerated from diagrams.mjs.

Fixed (5.0.0 approvals: "ask")

• node "$K" confirm <code> now reaches Claude Code's prompt. The hook only matched the literal kaizen.mjs confirm, so the documented call always failed. The hook now sees every confirm call. It requires confirm to run on its own, with the code written out. As a result:
  • a mark left by a declined prompt can no longer be reused by another visible call;
  • confirm … && git push can no longer slip a push through.
• The git pre-push hook guards Claude's pushes in "ask" mode.
  • From Claude Code (CLAUDECODE set), it only lets through the push the review hook just put to the user: same branch and commit, within 10 min, once.
  • From the user's own terminal, it only prints a reminder.
  • The review hook now also matches /usr/bin/git push and global options before push.
• friction metrics in "ask" mode:
  • a push is no longer counted twice;
  • new fields pushed_without_review (counted in the waiver share) and unreviewed_from_terminal.
• approvals.mjs reuses pendingWaiver / pendingDeploy instead of copying their lifetimes. JSON state helpers are shared in lib.mjs, and reading the state no longer creates .kaizen/state/. The push gates read checkPush's ask field; status uses approvalMode.
• pr overlap:
  • reads raw paths (-z), so non-ASCII folders are compared correctly;
  • an empty git cherry no longer counts as "absorbed".

Test plan

• node --test plugins/kaizen/tests/*.test.mjs plugins/kaizen-ops/tests/*.test.mjs: 228 pass, with and without CLAUDECODE
• New or updated tests:
  • confirm forms (prompted, refused when the code is in a variable or the call is chained);
  • git hook from Claude Code and from a terminal;
  • push spellings;
  • friction in "ask" mode;
  • learnings validate / list --retire;
  • pr overlap (non-ASCII folder, absorbed PR, empty cherry)
• diagrams.test.mjs passes after regenerating the SVGs; I opened pr-watch.svg to check the render

🤖 Generated with Claude Code

https://claude.ai/code/session_01HajgyyYNKjohDErB6g1YMt