docs(plugins): normalize operative cross-skill chains to explicit Skill-tool phrasing - #3070
Merged
Merged
Conversation
Contributor
|
Claude finished @kyle-sexton's task in 1m 8s —— View job Security review (
|
|
Last security-reviewed head: |
Contributor
|
Claude finished @kyle-sexton's task in 3m 20s —— View job CI code review (
|
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 63888c1236
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
|
Claude has reviewed this PR 3 times. The lane skips further automatic reviews after 5; deleting this comment resets the count. |
kyle-sexton
pushed a commit
that referenced
this pull request
Aug 21, 2026
…efresh the measurement (#3002) Two review findings from PR #3070. codebase-health:audit — the sweep rewrote the "Route remediation to the dedicated lanes" preamble into a Skill-tool invocation, but the --fix paragraph four lines below requires "an explicit user-directed suggestion" and says "Do NOT auto-invoke either skill — the user drives both". A read-only audit could have launched the source-editing implementation lane. Reverted to the recommendation form; the CHANGELOG now records why that route is a mention rather than an operative chain. The rubric's committed measurement went stale against its own tree: commit 6fcb34f added four "via the Skill tool" sites after the paragraph was last written, and this commit's revert removes one. Re-measured on the final merged tree: 1,635 lines carry the token, 189 are operative, 88.4% false positives. Both copies updated — the rubric and the skill-quality CHANGELOG. The two grep commands beside the figures reproduce them exactly.
Contributor
|
Claude finished @kyle-sexton's task in 4m 13s —— View job CI code review (
|
…ll-tool phrasing
Sweep the fleet's operative cross-skill hand-offs onto the invocation-mode
rubric's explicit phrasing ("invoke `/plugin:skill` via the Skill tool"), and
retire the scoping note that limited the doctrine to new and edited text.
Enumeration was mechanical (every non-frontmatter body line in
plugins/*/skills/** and plugins/*/agents/** carrying a cross-skill
/plugin:skill token), then hand-classified. Mentions were left alone:
identity/ownership statements, "Related skills" and capability-catalog tables,
text that explicitly recommends to the human rather than invoking, emitted
templates, eval fixtures, plugin README/CHANGELOG, and any chain whose target is
`disable-model-invocation: true` (unreachable from a skill under the rubric's
invocation-reach invariant).
Enforcement decision, recorded in the rubric: no skill-quality:check criterion.
A static scan cannot separate an operative chain from a mention, and that
separation is the whole rule — 1,129 body lines fleet-wide carry a cross-skill
token against roughly 120 operative sites. `playbooks:skill-authoring` gains an
authoring-time pointer instead.
Per-plugin patch bumps and CHANGELOG entries for all 34 touched plugins.
Closes #3002
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QbfCrj3X9FfGL7VRZYmrn4
…gate requires The cross-skill phrasing sweep touched morning-brief's SKILL.md, and check-changed-skills runs check-skill with --require-evals, which FAILs any touched SKILL.md that ships no evals/evals.json. morning-brief was the one swept skill without a suite; its siblings (known-issues, observability) already have one. Seven cases pinning the shipped contract: run-the-script-and-print-verbatim, the read-only refusal, repo derived from `gh repo view` rather than hardcoded, per-section degradation when no telemetry issue or queue labels exist, the merge-readiness authority routing to /source-control:babysit-prs, local telemetry routing to /claude-ops:observability, and the post-merge timestamp discriminator behind the Stranded findings section. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01QbfCrj3X9FfGL7VRZYmrn4
… 20 misses, 2 false claims
Two independent fresh-context audits of the landed sweep found defects in
three classes. This commit fixes all of them, plus one pre-existing runtime
bug found in passing.
False positives reverted (mentions the sweep wrongly rewrote):
- `toolchain:lint`'s comparison-table positioning line, which the sweep turned
into a self-invocation (`/toolchain:lint` IS that skill).
- `planning:brainstorm`'s route-onward step, which the sweep converted into an
invocation against its own body's "**Does not decide** — the skill
recommends".
- The three private-surface banners under `docs-hygiene:extract-ssot/actions/`
— self-references and access contracts, not chains.
- Clause 1 of `verification:confirm`'s four-clause capability catalog; the
other three were left, so the paragraph no longer read as one list.
- `planning:wayfind`'s blanket rewrite over a three-way route whose other arms
are a PLUGIN name and "small enough to just do → say so"; the phrasing now
attaches to `/planning:interview` alone.
Semantics restored: `implementation:implement-dispatch` cited another skill's
NAMED STEPS in five places ("run its 'Step 1: Prerequisite Check' preflight");
the sweep turned each into a whole-skill invocation. That is a different
action, and a re-entry hazard — `/implementation:implement` Step 0 chains back
into this skill. The citation form is back, and the Prerequisites paragraph
now names the loop it would otherwise create.
Operative sites the sweep missed, now phrased (20): `planning:plan`'s
test-strategy `/tdd:principles` directive (its twin in `plan-template.md` was
rewritten); `verification:confirm`'s PRIMARY live-app delegation, left bare
while its own fallback was rewritten; `work-items:work-loop` cycle step 4;
`work-items:attend-queue`'s `[escalated]` row; `work-items:scan-todos`;
`work-items:ship`'s container-discovery dead end; `code-tidying:tidy`'s
overflow filing; `education:teach`'s tiers 2 and 3 and its adjacent-sources
line, plus `context/lessons.md`; `disk-hygiene:clean`'s worktree hand-off;
`implement-dispatch`'s `/toolchain:check` table row; both `event-storming`
glossary-graduation delegations; `discipline:do-your-research`'s escalation to
its deep sibling (the exact reciprocal of a rewritten line);
`repo-fleet-hygiene:setup`; `songwriting:diagnose`;
`discipline:pick-for-the-problem`; `session-flow:retro`'s skill-candidate
hand-off; `session-flow:workflow`'s phase-boundary save-point;
`implementation:implement`'s two remaining handoff-table rows; and
`source-control:pull-request`'s post-merge retrospective step.
Two false claims removed. The rubric's "1,129 body lines" measurement was not
reproducible under any reading; it and the CHANGELOG copy now carry 1,635 /
186 (88.6% false positives) alongside the exact two `grep` commands that
produce them, so any reader can regenerate the pair. And
`docs/upstream/aihero-course.md` still said the doctrine was "scoped to
new/edited text", contradicting the rubric's "Scope: unconditional" in the
same PR; the lane record survives with a sentence retiring the scoping note.
The rubric also gains the answer to the open classification question: a
`/plugin:name` token is not automatically a skill. `/codex:review` and
`/code-review:code-review` are slash COMMANDS in their upstream plugins, not
Skill-tool-reachable skills, so they are correct keeps; `/caveman:compress`
is a skill and takes the phrasing.
Pre-existing bug fixed in passing: `morning-brief.sh` called
`fetch_repo_label_names` 116 lines before defining it, so that call died with
`command not found` and the fixture short-circuit never fired. It reproduces
unchanged at origin/main. Both definitions now precede their first caller;
`morning-brief.test.sh` goes from exit 1 (2 failing cases) to all 83 passing.
`repo-fleet-hygiene` was never touched by the sweep and gains its first bump
and entry (0.23.7); the other 17 plugins amend the entries they already
carry.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QbfCrj3X9FfGL7VRZYmrn4
A third audit pass found three operative cross-skill chains still carrying bare `/name` prose, which falsified the rubric's own new "Scope: unconditional ... no grandfathered set" claim: - `planning:design-handoff` :32 and :56 — two of the three FAIL routes back to `/planning:design`, four lines from the one the sweep did rewrite. - `coupling:reduce` :130 — the filing arm of the same sentence whose design-exploration arm was rewritten; the hand-off table's preamble does not reach it because it is prose outside the table. - `work-items:work` :260 — the post-green PR hand-off, bare while the same paragraph's `/work-items:track add` was phrased. Each target is `disable-model-invocation: false`, so the invocation-reach carve-out does not apply, and none carried a keep rationale. The three plugin CHANGELOG entries are amended to name the added sites rather than gaining new version bumps.
Contributor
|
Claude finished @kyle-sexton's task in 55s —— View job Security review (
|
…#3002) The #3002 sweep rewrote audit's "Route remediation" preamble into an operative Skill-tool invocation. That contradicts the same section's --fix rule: present a user-directed suggestion and do not auto-invoke /implementation:implement or /verification:confirm. The preamble is a mention (recommendation to the human), so restore the original wording and record that keep in the 0.8.1 changelog entry. Co-authored-by: Kyle Sexton <kyle-sexton@users.noreply.github.com>
…#3002) Re-ran the two committed grep commands over plugins/*/skills plugins/*/agents on the rebased tree. Denominator is 1,635 (was 1,634); numerator is 189 (was 186). Derived false-positive rate is 88.4% (was 88.6%). The same pair is mirrored in skill-quality's 0.17.3 notes. Co-authored-by: Kyle Sexton <kyle-sexton@users.noreply.github.com>
… mention Expand the 0.8.1 keep rationale so the next phrasing sweep does not re-flag the user-driven --fix gate as an operative chain. Co-authored-by: Kyle Sexton <kyle-sexton@users.noreply.github.com>
cursor
Bot
force-pushed
the
claude/pocock-steering-course-00zkvd
branch
from
August 21, 2026 07:49
3f29d22 to
762ec44
Compare
Contributor
|
Claude finished @cursor[bot]'s task in 1m 48s —— View job CI code review (
|
kyle-sexton
added a commit
that referenced
this pull request
Aug 21, 2026
…ight plugins (#3069) No linked issue ## Summary A fresh-context verification pass ran over #3065 **after** it merged and found defects in what that PR shipped: **12 in the eight plugins it touched, 12 in the provenance record, and 1 wrong count inside a released changelog entry.** Two of the twelve are behavioural — a completion gate that reverses its own skill's first instruction, and a cite into another plugin's private files. The rest are unresolvable pointers, a schema field with no writer, an eval that cannot fail for the reason it claims to test, and claims in `docs/upstream/cursor-pstack.md` that do not survive checking against the pinned upstream. Two more adversarial verifiers then ran over **this PR's own first commit**, one of them reaching the live upstream. They found eight further defects in work this PR introduced, plus one from a Codex review. All nine are fixed and described below. Every plugin whose shipped text changed takes a **patch bump with its own `### Fixed` entry**. #3065's release entries describe what #3065 shipped and cannot be retrofitted to describe corrections made after it merged. ## Fix ### The two behavioural defects in #3065 **`docs-hygiene:write-for-humans` graded against the wrong standard (0.18.2).** The skill's first instruction is to resolve the consuming project's own style guide before applying anything bundled — that re-posture *is* the port, and it is what `PLUGIN-PHILOSOPHY.md:198-202`'s two-lane rule demands. But 0.18.0 shipped the seven-question self-check unconditionally, and four of those seven restate a bundled layer (1 is Diátaxis, 2 and 3 are ASD-STE100, 5 is Global English's ambiguity set). A run that followed the skill to its end would resolve a project's declared Microsoft guide, write against it, then grade the result against the bundled standards and **rewrite it to conform**. That is a lane-1 hardcode wearing lane-2 clothing, and it falsifies two of the skill's six evals. The gate is **split, not switched off wholesale** — see below for why the first attempt was wrong. **`review:quality-gate downstream` cited into another plugin's private files (0.26.2).** Its "say plainly what is unverified" step path-cited the `playbooks` plugin's `fable-5` skill inside that skill's own `context/` directory, which the encapsulation contract makes private. Now `/playbooks:fable-5 verification` — the chapter argument is load-bearing, see below. ### The rest of #3065's defects | Plugin | Bump | Defect | |---|---|---| | `architecture` | 0.6.2 | `graft-record:` was added to the candidate schema and **no step wrote it** — the Handoff step named `status` and `agreed-shape` and stopped. A field nothing fills is always empty, so the left-behind half of a graft survived nowhere. Handoff now fills it; the research file's "travels into `agreed-shape`" named the wrong field and is replaced. | | `discovery` | 0.16.2 | Two routes back to the code-shape exclusion the skill is built around: "test names" listed as source-control evidence (a test name is a symbol, not someone writing down a reason), and a gotcha calling a named constant plus a convention "a hypothesis at best", colliding with `Speculative`'s own *Competing hypotheses* output section. Separately a **body-vs-eval gap on the neighbouring rule** — the version-control-behaviour ceiling said "never `Direct`" while the eval graded "never `Direct` or `Supported`". Plus an eval whose target behaviour was **indistinguishable from correct behaviour**: it told the model not to check the tracker but never stipulated one existed, and the category is presence-gated, so a bare checkout's correct output was graded a failure. Plus a link-target normalization to the fleet's `${CLAUDE_PLUGIN_ROOT}` convention. | | `implementation` | 0.15.2 | "the same grounding rule the loop lanes apply" — an unresolvable pointer for a consumer who installs `implementation` alone. Now names `work-items:work-loop` and `source-control:babysit-loop`. | | `prototype` | 0.9.2 | `explore-directions` step 6 restated the shared capture discipline in its **pre-0.9.0 form** — "record which variant won and why" — directly above the bullets that delete the losing variants irreversibly. | | `session-flow` | 0.32.2 | Both retro skills "handed" an accepted skill candidate to `/playbooks:skill-authoring`, which takes no arguments and performs no actions. Handing it a candidate resolves to nothing, so the destination 0.32.0 set out to give did not exist. Both now read it for the doctrine and draft against it. | | `testing` | 0.7.2 | "Prefer no new test to a bad one" is the upstream cursor/plugins `tdd` cost branch (`cursor/plugins@60c641e4` `pstack/skills/tdd/SKILL.md`). Cited inline as `(upstream cursor/plugins tdd)`. `/tdd:principles` contains neither the phrase nor the six triggers. | ### One declared in-place correction to a released body `scripts/check-changelog-parity.sh` sanctions corrections inside an already-released version section when the correcting PR names each edit in its body **and** in the new release entry. Exactly one qualifies, and it is the only removed changelog line in the whole diff: architecture 0.6.0 said "the five-part schema was pinned in three places"; commit `228a2b19` changed the count in exactly two spots. Corrected in place; heading untouched. Nothing else was edited in place. Three blocks initially drafted as appends into released `### Added` sections were lifted back out and re-homed under the new patch entries, because attributing post-release work to a released version misleads anyone reading the changelog to learn what that version contains. ### Concurrent-bump resolutions — #3067, #3068, #3070 Main bumped these plugins three separate times while this branch was open, and each time it claimed a number this branch had already taken. - **#3067** shipped `review` 0.25.1 → this branch moved to 0.25.2. - **#3068** shipped `review` 0.26.0 → this branch moved to 0.26.1. - **#3070** shipped a **fleet-wide** `### Changed` sweep that claimed **the same patch number on all eight plugins this PR bumps** → every entry here moved to `.2`. The third one is worth recording, because it was invisible to the usual check. `git merge-tree` reported **zero conflict markers**: each side had edited a different region of each file, so a clean-looking merge would have produced eight changelogs in which two different releases share one heading. That is precisely the corruption `--check-preserved` exists to catch. The real merge surfaced nine conflicts once the bodies collided. In every case main's heading and body are untouched and this branch's entry moves up; no release is folded, relabelled, or dropped. **One of those nine was semantic rather than numeric.** #3070 respelled `retro/context/session.md`'s route as *"Hand it off by invoking `/playbooks:skill-authoring` via the Skill tool"* — which reintroduces the exact claim this PR fixes, since that skill takes no arguments and performs no actions. Both changes are kept: the invocation still names the Skill tool (it is invoked in order to be *read*), but it reads the doctrine and drafts against it rather than handing anything over. The sibling `running-retro` is deliberately **not** aligned to match — #3070's own entry says it left that list alone because it sits under "Offer routing — never auto-apply", so the asymmetry is intentional and is now recorded in the 0.32.2 entry. ### The provenance record — 12 corrections `docs/upstream/cursor-pstack.md` carried twelve claims that did not survive checking against `cursor/plugins@60c641e4`: 1. A judge instruction reported as an unconditional `must`; upstream says "**Prefer** a different model family from the parent's" and `must` appears zero times in that file. 2. "upstream spawns candidates under `isolation: worktree`" — the word never appears upstream. Its mechanism is a per-candidate output path, which is why this fleet's three recorded `isolation:` rejections were not in conflict with it. 3. The `arena` omission's grounds now rest on the Rule of Three alone. 4. The `technical-writing` row credited our third always-rule to upstream's above-the-layers trio; it comes from upstream's separate word-list and anti-jargon paragraphs. 5–7. Three non-verbatim quotes: `debugging:debug` phase 5, upstream's `tdd` description, and the above-the-layers rule text. 8. `ai-slop`'s catalog path — `reference/catalog.md` → `skills/audit/reference/catalog.md`. 9. A bare `context/spec.md` — the real file is `review/skills/quality-gate/context/spec.md`, and `review/context/` holds no `spec.md`, so the short form pointed nowhere. 10. "four always-listed descriptions" → six, now named individually. 11. "`bro` is seven lines, two of them body" → one. 12. "Five of its six operating principles" → upstream states **two** separate six-item lists, and across both, every item but one ("no shortcut by code-reading") restates historiographical source criticism and IC analytic tradecraft. An independent verifier re-checked 1, 2, 11 and 12 against the live upstream files and confirmed each verbatim. ### Convention record `docs/conventions/upstream-drift` 1.6.0 adds the adopter row for `write-for-humans`' source records. **Three carry a publication-event trigger; Google's is a page-content divergence**, because that guide is a continuously-edited site with no edition to pin. The contract admits either shape, and the row says which record uses which. ## Defects in this PR's own diff, found and fixed Two fresh-context verifiers ran over commit `b1bf09c6` with the rationale withheld; a Codex review ran alongside. 1. **The new `write-for-humans` gate contradicted itself** — "the seven below do not apply" under a declared guide, then "the three rules above apply either way", when questions 4, 6 and 7 **are** those three rules. Now split per question: four stand down, three always apply. 2. **An undescribed compression pass rode along in that same file** — nine hunks the changelog never mentioned, three of them lossy (a bullet lost "of a published standard"; two gotchas lost their actionable half; one lost the "only" whose placement the skill's own ambiguity rule governs). The worked example lost "(and ambiguity)" from a fix that *is* an ambiguity fix, and misquoted its own Before block. Reverted to main, with only the two described fixes re-applied. 3. **`interface-design.md` named two destinations for one record.** Replaced rather than annotated. 4. **`explore-directions` step 6 was still short a clause** — it dropped "what the discarded parts held that the graft deliberately left behind" while the changelog claimed the widened form. 5. **The bare `/playbooks:fable-5` cite (Codex).** That skill's `argument-hint` and Arguments paragraph make a bare invocation arm its **entire operating doctrine as standing session instructions for the rest of the run**. Now `/playbooks:fable-5 verification`. 6. **The `discovery` umbrella overcounted** — "three routes back to the code-shape exclusion" when the third is the version-control-behaviour ceiling, which the same skill says explicitly is *not* code shape. 7. **A link-target normalization was undescribed.** The change is right and stays; the changelog now names it. 8. **The `review` entry spelled the private path the fix removed.** Described rather than spelled. 9. **The `upstream-drift` adopter row generalized from three records to four.** Corrected. **One Codex finding was accepted after a primary-source check reversed an earlier decline.** Its claim that `/tdd:principles` contains neither the phrase nor the six triggers is correct. That skill's nearest sentence is Khorikov's *"It's better to not write a test at all than to write a bad test"* in `testable-architecture-khorikov.md` — the 2×2 / Humble Object chapter, which this port used as grounds to *reject* upstream's five-item bad-test definition as already owned. The phrase and the six triggers come from upstream `tdd` (*"Prefer no new test over a bad test"*, plus the identical trigger list), which `cursor-pstack.md:34` already recorded as **Taken**. The original cite therefore contradicted this PR's own provenance document. Now `(upstream cursor/plugins tdd)`. ## Verification Every gate re-run locally at the branch tip against the current base (`69b3584`), all green: - `check-changelog-parity` — all four modes. `--check-bump`: each of the eight version changes has its own new entry and is strictly greater than the base's. `--check-preserved`: 9 changed changelogs, **322 headings compared**, none dropped — including every heading #3067, #3068 and #3070 added. `--check-order`: newest-first, no duplicates. - `check-changed-skills` — 9 skills checked, 0 failed. - `validate-plugins` (manifests + catalog, `--strict`), `validate-plugin-contracts` (48 setup skills, 2823 files). - `check-skill-portability`, `check-skill-count-claims`, `check-skill-leaf-names`, `check-cross-plugin-source-drift`, `check-plugin-manifest-presence`, `check-orphaned-fixtures`, `check-silent-skips`, `check-fleet-audit-doc-grammar`, `check-lane-coverage`, `check-contract-clause-coverage`, `check-contract-slice-prune`, `check-stale-base-overlap`. - `markdownlint-cli2` over **every tracked markdown file** — 0 issues. - Eval sets: all `evals.json` validate against the bundled schema, 0 failures. `check-evals-quality` returns **PASS with advisory warnings fleet-wide** (the WARN tier never fails the step); **the two suites this PR touches carry 0 warnings**. An earlier revision said "PASS, 0 warnings" without that distinction — a verifier caught the overstatement. - **No executable file is touched anywhere in this PR** — the diff is markdown, eight `plugin.json` version bumps, and one `evals.json` — so the plugin test suite cannot be affected by it. ## Related - #3065 — the merged PR these corrections apply to. - #3067, #3068, #3070 — the three concurrent bumps this branch renumbered around; all of their headings survive. - `docs/upstream/cursor-pstack.md` — the single source of truth for everything derived from `cursor/plugins@60c641e4`. - `docs/conventions/upstream-drift` — the four-part stamp contract the new adopter row conforms to. - `docs/PLUGIN-PHILOSOPHY.md:198-202` — the two-lane convention posture the `write-for-humans` self-check violated. - `scripts/check-changelog-parity.sh` — the released-entry body-edit clause the one in-place correction is declared under, and the concurrent-bump guidance every resolution here follows. --------- Co-authored-by: Claude <noreply@anthropic.com> Co-authored-by: Kyle Sexton <kyle-sexton@users.noreply.github.com> Co-authored-by: Cursor Agent <cursoragent@cursor.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Closes #3002
Summary
Lane 6 of the AI Hero course vetting adopted the cross-skill invocation-phrasing doctrine
(
docs/conventions/invocation-mode/README.md, "Cross-skill invocation phrasing"), but could onlyscope it to new and edited skill text because the fleet's existing practice was mixed. This is the
sweep that retires that scoping.
Operative cross-skill hand-offs across
plugins/*/skills/**andplugins/*/agents/**now name themechanism explicitly — "invoke
/plugin:skillvia the Skill tool" — and the rubric's scope becomesunconditional. Wording only: every presence gate, fallback, threshold and step order is preserved.
The two offenders the issue named as calibration cases (
knowledge:map-corpus,knowledge:docpage-digest) are among them.Enumeration was mechanical, then hand-classified. Mentions were deliberately left alone — identity
and ownership statements, "Related skills" lists and capability catalogs, verdict annotations, text
that explicitly recommends to the human rather than invoking, emitted templates and eval fixtures,
and any chain whose target is
disable-model-invocation: true(unreachable from a skill under therubric's own invocation-reach invariant). Each keep class is recorded in the touched plugin's
CHANGELOG.
Fix
Retire the scoping.
docs/conventions/invocation-mode/README.md— "this binds NEW and EDITEDskill text" becomes "Scope: unconditional." Adds the operative-vs-mention test the sweep ran
on, so the next author has the rule rather than the precedent.
docs/upstream/aihero-course.mdgains the matching retirement sentence, with its historical wording left intact so the lane record
survives.
Rewrite the operative chains. Notable: the knowledge pipeline's queue and interview hand-offs;
implementation's dispatch cadence;planning's design→plan gate, stress-test dispatch andresearch iteration;
verification:confirm's delegation set;docs-hygiene:extract-ssot's renamesweeps. Multi-row composition tables name the mechanism once in a preamble rather than in every
cell.
A
/plugin:nametoken is not automatically a skill. New rubric subsection, because the sweephit it:
/codex:reviewand/code-review:code-revieware slash commands from upstreammarketplaces, not Skill-tool-reachable skills, so text directing them stays as written. Verified
against the upstream sources rather than inferred —
openai/codex-plugin-ccshipsplugins/codex/commands/review.md(itselfdisable-model-invocation: true), andanthropics/claude-plugins-officialshipsplugins/code-review/commands/code-review.mdwith noskills/directory at all./caveman:compressis the opposite case — a real skill — so it takesthe phrasing.
Enforcement decision: declined, and recorded with a regenerable measurement. No
skill-quality:checkcriterion. skill-quality/fleet: enforce the invocation-mode rubric — explicit disable-model-invocation on 17 skills, check criterion, skill-authoring pointer #2968's check 24 is the explicit-key criterion, a different axis.A static scan cannot separate an operative chain from a mention, and that separation is the whole
rule. The rubric now commits the two shell commands that produce its figures beside them, so any
reader can regenerate the pair rather than trusting it.
playbooks:skill-authoringgains anauthoring-time pointer, which is where the rule is decidable.
claude-ops:morning-briefgains its first eval suite. The sweep touched itsSKILL.md, andcheck-changed-skills.shruns with--require-evals. Seven cases pin the shipped contract(run-the-script-and-print-verbatim, the read-only refusal, repo resolution, per-section
degradation, the two authority route-outs, the post-merge timestamp discriminator).
Pre-existing bug fixed in passing.
plugins/claude-ops/skills/morning-brief/scripts/morning-brief.shcalled
fetch_repo_label_namesat line 221 but defined it at line 337, so that call died withcommand not foundand two of the skill's own tests had been failing. This reproduces unchangedon
mainand is not caused by this diff — it is fixed here because it lives in a shipped script,in a plugin this PR already bumps, in the very skill this PR adds evals for.
morning-brief.test.shgoes from exit 1 (2 failing cases) to exit 0 (all 83 passing).
Verification
Every gate below was confirmed green on the base commit before this work landed, so any failure
would be attributable to this diff. All re-run on the final merged tree:
scripts/validate-plugins.shgenerate-catalog.mjs --check/generate-cheatsheet.mjs --checkcheck-changelog-parity.sh—--check,--check-order,--check-bump,--check-preservedcheck-changed-skills.sh origin/mainmarkdownlint-cli2over 139 changed.md0 issuestyposskill-quality/scripts/check-skill.test.shmorning-brief.test.shcheck-jsonschema+check-evals-quality.shon the newevals.jsonPASS (0 warnings)Three independent fresh-context audits were run against the landed tree, each with the author's
rationale withheld, and their findings fixed:
instructed to invoke itself), 4 sites where a named-step citation had been converted into a
whole-skill invocation creating a re-entry loop, and 12 missed operative sites.
git diff --numstatcheck across every changed file; all 7 eval cases traced by hand to shippedbehavior; found 6 further misses and the stale
aihero-course.mdmirror.mainwhere they should be,reproduced the committed measurement to the digit, and found 3 residual bare chains
(
planning:design-handoff×2,coupling:reduce,work-items:work) which are now fixed.Two audit findings were disputed with evidence and deliberately not applied:
testing:diagnose:42andreview:quality-gate's downstream route both sit under column headersliterally reading
Suggest/ "not executed here", so rewriting them would have manufactured freshfalse positives of exactly the class audit 1 was catching. The third audit independently agreed with
both calls.
Related
Refs #2968 (check 24 — the explicit-key criterion this decision coordinates with, different axis)
Refs #2994 (AI Hero course journey roadmap)
Generated by Claude Code