Release: main has three merged PRs unpublished, and one question to settle first #16

Activity

1. cmj0121

   added

   questionFurther information is requestedFurther information is requested

   on Sep 16, 2026
2. cmj0121

   added a commit that references this issue on Sep 16, 2026

   Merge: settle the version-bump rule, and stage 2.2.0 without tagging it

   ...

   f2c86ff
3. cmj0121

   closed this as completedin #17on Sep 16, 2026
4. cmj0121 commented on Sep 16, 2026

   

   cmj0121

   on Sep 16, 2026

   OwnerAuthor

   More actions

   Reopening: this was closed by automation, not by being finished.

   PR #17's body opened with Closes #16, except its last acceptance line — GitHub's linker reads
   the first two words and ignores the qualifier, so the merge closed the issue at the same second
   it landed. The qualifier was the operative part.

   What #17 actually did:

   • the question answered, and the answer recorded in CLAUDE.md
   • the chosen plugin versions bumped in all three locations each — seven plugins, not the six
     this issue names in prose; agent-page's body changed in 0757a7b
   • the top-level version bumped, to 2.2.0
   • a matching v* tag pushed — deliberately not done; the maintainer holds this gate
   • make test and check-version-sync pass, with the marketplace version ahead of the tag,
     which check-version-sync reads as release-pending

   So one box is open and it is the one the issue's headline complaint depends on. Since #17 landed,
   three further PRs (#18, #19, #20) have merged, which makes the unreleased pile bigger rather than
   smaller — the second-order risk this issue named.

   Current state of the release window: seventeen plugins at 2.1.0, briefing and shortcut at
   2.0.0 (neither has phases, so #19 did not touch them), top level at 2.2.0, latest tag v2.1.0.
5. cmj0121

   reopened this on Sep 16, 2026
6. cmj0121 commented on Sep 16, 2026

   

   cmj0121

   on Sep 16, 2026

   OwnerAuthor

   More actions

   Still held, deliberately — the tag is not blocked on a defect

   make eval was run on dab7767 and came back 8/13, exit 1. The maintainer chose not to tag
   on that result, so v2.2.0 stays unpushed and this issue stays open.

   Worth separating, because the red is easy to misread: make eval failing is not a release
   blocker and was never a release gate. All four repo checks pass on main, and make eval sits
   outside make test by design — it needs a credential, a network and money per run, and CI cannot
   run it. Every one of the five failures is a routing gap that predates this work and that #14
   deliberately left unfixed with a stated reason. None was introduced by the seven merged PRs.

   The five are catalogued in #25, and the measurement problem behind them in #24 — the short version
   being that the blind router this repo has judged descriptions with is a forced choice among 19
   skills plus none, so its "skill A steals from skill B" verdicts overstate, where a real session
   often invokes nothing at all.

   Current release state

   plugins at 2.1.0	17
   plugins at 2.0.0	2 — briefing, shortcut; neither has phases, so #19 did not touch them
   top-level version	2.2.0
   latest tag	v2.1.0
   unreleased merges on main	ten — #9, #11, #12, #17, #18, #19, #20, #21, #22, #23

   check-version-sync reads the marketplace version as ahead of the tag and treats that as
   release-pending, so nothing fails; it simply stays unpublished.

   This issue's headline complaint has therefore grown from three unreleased PRs to ten. That is the
   cost of holding the gate, recorded rather than argued — the decision is the maintainer's, and the
   tag command when it is wanted is git push GITHUB main v2.2.0.
7. cmj0121 commented on Sep 16, 2026

   

   cmj0121

   on Sep 16, 2026

   OwnerAuthor

   More actions

   Released: v2.2.0 is tagged and published

   The gate this issue was held open for is now closed by the maintainer, not by automation.

   tag	v2.2.0 at ce0aa0f
   release	https://github.com/cmj0121/wisdom/releases/tag/v2.2.0
   merges published	15 since v2.1.0, 14 of them pull requests (#9 … #29)
   plugin versions	17 at 2.1.0; briefing and shortcut at 2.0.0, having no phases to change
   top-level version	2.2.0, now equal to the tag rather than ahead of it
   pushed to	GITHUB and origin — both carry main and the tag

   make test 4/4, scripts/check-version-sync passes with the tag matching the top-level
   version rather than lagging it.

   What this issue actually settled

   Its headline complaint — three unreleased PRs — grew to fifteen while it was open, which was
   the second-order risk it named. But the hold was the right call twice over, and both reasons
   are worth keeping:

   1. The question label was the real content. "Is a description rewrite a version bump?"
      is answered and recorded in CLAUDE.md (Settle whether a description rewrite is a version bump, and publish what main already carries #17), so the next rewrite does not re-open it: a
      bump tracks what a plugin does; catalogue copy alone moves the top-level version; and a
      plugin moves at most once per release window, or the number stops naming a released state.
   2. make eval at 8/13 was correctly read as not a release blocker. All four repo checks
      passed throughout. Every one of those five failures was a routing gap that predates this
      work, catalogued in Five committed cases where no skill fires at all #25 with the measurement problem behind them in The blind router and make eval disagree, and the disagreement has a shape #24 — and Record when a skill should fire, and what each runner measures #26 has since
      recorded the reading that explains four of them: a skill should not fire when the assistant
      can simply comply in the current turn.

   Also closed in this window

   • Skills: cut router cost, use the references/ machinery, remove ambiguity #8 — router cost, references/, ambiguity. Four items landed; two were answered by
     measuring rather than by doing, and the acceptance restated at the numbers the repo can hit.
   • Trigger evals: run at 3 samples per query, and harden the near-misses that measure nothing #13 — trigger evals at 3 samples. Both halves landed; the two remaining boxes are recorded
     declines, not backlog.

   No open issues remain.
8. cmj0121

   closed this as completedon Sep 16, 2026