SkillPedia Skills Encyclopedia
中文 EN

技能 2026-09-10 19:40 UTC

pm-skillsv2.33.0

版本【v2.33.0】由提交人 jprisant 提交,主要更新类型为:功能特性新增与扩展。提交说明:ci: add the C1 delivery check and C2 manifest-tail freshness check (#292)。为项目引入了新的能力与模块支持。

jprisant 1cef1a9 product-on-purpose/pm-skills

Commit提交信息原文

ci: add the C1 delivery check and C2 manifest-tail freshness check (#292)

Phase 1 of the documentation currency program. Together these close the only
two of its seven observed defects that actually reached users, both of which
shipped with every validator green.

C1, scripts/check-delivery-pin.mjs. pm-skills ships through two repositories:
users install from the product-on-purpose marketplace, which lives in the
agent-plugins repo and pins members by commit SHA independently of anything
here. On 2026-09-01 v2.33.0 was tagged, released, CI-green and validator-green
while reaching nobody for about eight hours, because that registry still pinned
the v2.32.0 commit. A stale SHA pin resolves perfectly to real, working, older
code, so nothing errors and no other check has an opinion. It was found by a
field report.

The check compares both the served version and source.sha against the tag.
Comparing both matters: a registry that moved the label but not the pin serves
old code under a new number, and a version-only check waves that through.

Two design points the original scoping did not anticipate:

  It is not a pre-tag or CI check. The bundle runs before the tag exists, so
  the question is unanswerable rather than merely unasked, and validation.yml
  runs on pushes and pull requests where there is no new tag to ask about.

  It is not a same-sitting blocker. The real gap was about eight hours, so a
  gate that had to go green inside one G4 walkthrough would be routed around on
  first use, and a gate that is routinely bypassed is worse than none. It is a
  cheap idempotent question you re-ask until it answers yes. Exit 0 delivered,
  1 tagged-but-not-delivered, 2 unverifiable. Exit 2 never prints the success
  line, because silence being read as success is the original defect.

C2, evalManifestTail() in scripts/gen-derived-surfaces.mjs. The three manifest
descriptions are half generated: descriptionHeadline() owns the counts and
everything from the first version token onward is authored prose the generator
carries through verbatim, which is exactly why nothing read it and why all
three still pitched v2.32.0 two days after v2.33.0 shipped.

It lives in the generator's already-enforcing --check path, so it needs no new
wiring and adds nothing to the enforcing-validator count, leaving the program's
own four-validator growth budget untouched. Only the leading version token is
asserted. A tail may legitimately narrate an older release further along, and
flagging every token in sight is the heuristic that had check-version-references
reporting 1287 findings at a near-zero true-positive rate before it was retired.
A description with no tail makes no version claim and cannot be stale.

Seventeen fixtures, per the program's rule that no check ships without one
proving it fires. C1's include a reproduction of the actual v2.33.0 miss and
both half-failures (label moved but not pin, pin moved but not label). C2's
include a guard on the premise that the generator still carries tails through
verbatim, since C2 would be moot if that ever changed.

Documentation, and two more false claims found while writing it:

  ci-overview.md never catalogued gen-derived-surfaces at all. Added, covering
  both what it generates and the new tail assertion, plus a section explaining
  why C1 is deliberately outside CI and the pre-tag bundle.

  Its "How to Add a New Validator" section still taught the dual-shell trio,
  frozen since v2.30.0, and said to wire a validator into "BOTH the ubuntu and
  windows jobs" when validation.yml is one job with an OS matrix. Rewritten for
  single-source Node, the fixture rule, and where a post-tag check belongs.

  The runbook's G4 sub-check 4 said the GitHub Release body is not auto-created.
  release.yml publishes one on tag push via softprops/action-gh-release with
  body_path: RELEASE_NOTES.md, so the real task is replacing an already-public
  body. Fixed rather than silently renumbered, since C1 is inserted above it.

No tag is owed. Skills, commands and sub-agents are untouched.


Claude-Session: https://claude.ai/code/session_016M6N24i6ZZhwqjSF9rTuQc

Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>