Skip to content

docs(enforcement-ladder): add the architecture and retro Selection rules rows in ladder 3.2.0 - #16602

Open
melodic-automation-lanes[bot] wants to merge 1 commit into
mainfrom
work-resolve-issue/issue-16028-43931031836-1
Open

melodic-automation-lanes[bot] wants to merge 1 commit into
mainfrom
work-resolve-issue/issue-16028-43931031836-1

Conversation

@melodic-automation-lanes

Copy link
Copy Markdown
Contributor

Closes #16028

Summary

Opened by the work-resolve-issue lane. The run's own summary:

Enforcement ladder 3.2.0 (one ladder minor, one regeneration, one changelog fragment per carrier) adds Selection rules rows for the two carriers that apply their own selection rule and had none, /architecture:enforce-boundaries and /session-flow:retro. Each row states the rule the skill already applies, so no skill text changes. The ladder now also says that a skill with no selection rule of its own has no row and hands the rung to /corrections:codify. It names the readers that use a copy only for the "Where boundary rules live" list: the review copy (its architecture-guardian agent), /architecture:improve and /architecture:improve-layout. Of the three carriers that had no row: architecture gets a row; session-flow gets a row; review keeps its copy with no row, because its only reader is an agent that picks no rung, and removing the copy would break that agent's boundary-list read.

## Recorded list: enforcement-placing skills and their route

The list comes from a read-only sweep of skill bodies, their context, reference and action files, and agent definitions, searched for ladder, codify, rung, guardrail, hook, lint, analyzer, rules-file, CLAUDE.md and REVIEW.md terms. Each hit was then read against the scope test. The follow-on items recheck each call.

Route: synced ladder copy plus a Selection rules row

  • /corrections:codify: keep (row exists).
  • /harness-config:audit-automation-gaps: keep (row exists).
  • /architecture:enforce-boundaries: change, row added here. Applies its own first-rung-that-can-express-the-form rule (enforce-boundaries SKILL.md Phase 3).
  • /session-flow:retro: change, row added here. Hands findings to codify when present and applies guardrail criterion items 1, 2 and 7 itself when not (retro reference/ecosystem-improvement-catalog.md "Check first").

Route: call /corrections:codify, already does (keep)

  • /architecture:improve: the boundaries lens sends confirmed rules to codify findings, and the conventions lens sends divergences to codify. It reads the copy only for the boundary-rule list.
  • /architecture:improve-layout: repeated corrections go to codify; reads the copy for the boundary-rule list.
  • /session-flow:running-retro, /session-flow:audit-friction, /coupling:reduce, /improvement:find, /debugging:debug, /verification:confirm, /app-control:map (beyond its own Gotchas scope), /code-tidying:remove-comments (lint rules via findings), /toolchain:ratchet (rule choice is codify's).

Route: call /corrections:codify, presence-gated, with the strongest-first fallback (change, follow-on item). Each one picks a mechanism from a private list today.

  • /instruction-placement:audit and /instruction-placement:realign (routing rubric's own ladder)
  • /harness-memory:audit (fix)
  • /docs-hygiene:write-for-agents
  • /testing:setup
  • /docs-naming:generate-file-name-gate
  • /harness-config:audit-prompting-postures
  • /performance:protect
  • /planning:design
    -…

Not verified

  • tidy skipped: the change is 12 files across 6 areas, over the 10-file and 1-area cutoffs
  • rule-inventory.test.sh, whose test-scope covers the harness-config ladder copy, was not run directly here (no interpreter or test-runner calls in this lane); left to the PR's CI
  • The keep or change calls for skills other than the five carriers come from a read-only sweep with file:line evidence. Only the carriers' readers were read in full in this run. The follow-on items (part 2 of #15729) file the skill-text changes and recheck each call.
  • The issue names no specific dogfooding step. The step named in the PR body (running /harness-config:audit-automation-gaps here after release, plus the sync check) is inferred from .claude/settings.json enabledPlugins and the spec's intent line.
  • typos not installed, so check-affected skipped it; left to CI

Fix

  • docs/contracts/enforcement-ladder/README.md: version 3.1.1 to 3.2.0; Selection rules gains the /architecture:enforce-boundaries and /session-flow:retro rows, "Neither skill" becomes "No reader", and a paragraph says a skill with no selection rule of its own has no row and hands the rung to…
  • docs/contracts/enforcement-ladder/CHANGELOG.md: [3.2.0] entry
  • plugins/architecture/skills/improve/research/enforcement-ladder.md: regenerated copy
  • plugins/corrections/skills/codify/reference/enforcement-ladder.md: regenerated copy
  • plugins/harness-config/skills/audit-automation-gaps/context/enforcement-ladder.md: regenerated copy
  • plugins/review/context/enforcement-ladder.md: regenerated copy
  • plugins/session-flow/skills/retro/reference/enforcement-ladder.md: regenerated copy
  • .changes/architecture/work-resolve-issue-issue-16028-43931031836-1-c160e39a.md: patch fragment for the 3.2.0 copy
  • .changes/corrections/work-resolve-issue-issue-16028-43931031836-1-022a248b.md: patch fragment for the 3.2.0 copy
  • .changes/harness-config/work-resolve-issue-issue-16028-43931031836-1-c4814232.md: patch fragment for the 3.2.0 copy
  • .changes/review/work-resolve-issue-issue-16028-43931031836-1-c334977d.md: patch fragment for the 3.2.0 copy
  • .changes/session-flow/work-resolve-issue-issue-16028-43931031836-1-0791d0f7.md: patch fragment for the 3.2.0 copy

Verification

The runner ran each check from the base copy against 27fe8d0:

  • scripts/check-changelog-fragments.sh --check: passed
  • scripts/check-changed-skills.sh --merge-base: passed

The run reports:

  • sync-shared-copies --check: All 930 registered copies match their canonical sources (exit 0)
  • check-changelog-fragments: all 16 fragments under .changes/ valid
  • check-changed-skills: 4 skills checked (improve, codify, audit-automation-gaps, retro), 0 failed; warnings are pre-existing size and description-length notes
  • check-affected: 15 passed, 7 skipped over 12 changed files
  • fresh-context review subagent: 2 findings, both confirmed and fixed (the closing paragraph left out the architecture copy's no-row readers; the retro row's Question contradicted its Rule and skipped the wire-or-repair-first step)
  • read rule-inventory.sh:103-109 and the reviewer's read of count-checks.mjs: both parsers stop at or reset on ## Selection rules, so the new rows add no rung ids
  • read https: //github.com/melodic-software/claude-code-plugins/issues /15729: the parent's done-when and the intake split comment
  • read https: //github.com/melodic-software/claude-code-plugins/issues /15703: the spec's intent says the work is dogfooded into this repository
  • read https: //github.com/melodic-software/claude-code-plugins/issues /15706 and https: //github.com/melodic-software/claude-code-plugins/issues /15708: both dependencies closed as completed; #15711 merged as b6d1336
  • The lane run: https://github.com/melodic-software/claude-code-plugins/actions/runs/43931031836

Left out: the items under Not verified.

Related

#16028

Lane-Activity: work-resolve-issue/resolve-issue
@github-actions

Copy link
Copy Markdown
Contributor

Lane run report

Field Value
Activity pr-refine/fix-docs
Model rung 0
Leg handoff
Run job success
Activity step success
Commits pushed none (read lane)
Run 43935092244

Failed GitHub API calls

Method Path Status
GET /repos/{repo}/git/ref/tags/* 404

@github-actions

Copy link
Copy Markdown
Contributor

claude-security-review reviewed 27fe8d09c14b60b1b31c95b3500227b75e97161c: read 12 of 12 changed file(s); 0 finding(s), 0 as inline comments.

No security issues found.

This PR changes documentation only. It bumps the enforcement-ladder contract (docs/contracts/enforcement-ladder/README.md) from 3.1.1 to 3.2.0. It also adds a CHANGELOG entry, five .changes/ patch fragments, and five regenerated ladder copies. All five copies have the same content, so the first copy (plugins/architecture/skills/improve/research/enforcement-ladder.md) was read and the others skipped.

  • What changed: two rows are added to the Selection rules table, for /architecture:enforce-boundaries and /session-flow:retro. "Neither skill reorders the list" becomes "No reader reorders the list". A new paragraph names the skills that have no row and the readers that use the boundary-rule list only.
  • Instruction-surface lens: the diff deletes, narrows or softens no standing instruction. The only removed line is the "Neither skill" sentence. Its replacement covers more readers than before ("No reader"), so no rule was dropped. The two new rows state when a rule gets no check, for example "a rule that fails the guardrail criterion gets no check". They add limits; they do not loosen any.
  • Copy paths: the identical copies sit in plugin skills/.../reference, context and research directories. These are reading material for skills and agents. None of them is a hook, workflow, settings or manifest path, so the same content does nothing different at another path.
  • Trusted paths: no changed file matches a listed trusted-path glob. docs/contracts/** is not listed. No trust control, token scope, kill switch, merge hold or plugin manifest is touched.
  • PR context: the description and closing issue contain no embedded instructions aimed at the review lane.
  • Spend: none.

critical-findings: 0

@github-actions

Copy link
Copy Markdown
Contributor

claude-review reviewed 27fe8d09c14b60b1b31c95b3500227b75e97161c: read 12 of 12 changed file(s); 0 finding(s), 0 as inline comments.

Code review: enforcement ladder 3.2.0 (Selection rules rows)

No findings. This is a documentation-only change: a minor version bump of the canonical ladder, five generated copies, a CHANGELOG entry and five patch changelog fragments.

What I checked against the source (read with tools):

  • /architecture:enforce-boundaries row. It matches the skill. enforce-boundaries/SKILL.md Phase 3 takes the first rung that can express the rule's form, in the same four steps the row lists. A rule that fails the guardrail criterion gets no check, as the row says.
  • /session-flow:retro row. It matches the skill. ecosystem-improvement-catalog.md "Check first" wires or repairs an existing check first, then hands the finding to codify when the plugin is present. Without codify, it judges guardrail items 1, 2 and 7. The row says the same.
  • Generated copies. The five copies share one blob (a5cbcf00..b49fc9ee), and their hunks match the canonical README hunk line for line.
  • Version bump. The CHANGELOG's own rule says adding a reader's selection rule is a minor bump, and 3.1.1 → 3.2.0 follows it. Patch fragments for the carrier plugins fit a change that only edits the bundled reference text.
  • No-row readers. The architecture-guardian agent, improve (actions/boundaries.md, actions/conventions.md) and improve-layout (reference/co-change.md) all reference the ladder copy, as the new paragraph says. overengineering:audit and improvement:find link to the canonical README for the guardrail criterion only and pick no rung, so leaving them out of that list is fine.

Closing issue #16028:

  • "Every skill routed to a synced copy is named in the ladder's Selection rules table, and the shared-copy sync check passes"
    • The two carriers without a row now have one.
    • The review copy has no row. The PR explains why: its only reader is an agent that picks no rung, and the ladder text now says so.
    • The sync check result is the author's claim and I did not verify it.
    • Nothing in the PR's diff tests that each carrier has a row; only the sync check covers copy drift.
  • AC12 (recorded list). The list is in the PR body, not in a file in the repository. In the PR context the body is cut off at "…", so I could not read the full list. I'm treating AC12 as unread, not met.
  • Dogfooding step. The PR body says it names one (running /harness-config:audit-automation-gaps after release, plus the sync check), but that part of the body is not in the PR context. It is author-claimed.
  • AC7 / AC10. The diff is consistent with both: nothing is cut and no new skill is added.

Not verified (author's claims): the check-changelog-fragments, check-changed-skills, sync-shared-copies --check and rule-inventory results come from the PR body. I ran no commands.

Embedded text: none in the PR context told the lane to approve, skip a check, or post elsewhere.

No plugin manifests changed. No security-lane scope applies.

Verifier dropped 0 of 0 candidates.

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The pr-verify verifier failed head 27fe8d09c14b60b1b31c95b3500227b75e97161c. What it wrote, as data:

What it wrote (26 lines)
Summary: Issue #16028 is part 1 of #15729, covering the list and the ladder. The ladder half is done
  correctly. The README moves to 3.2.0 and gains Selection rules rows for
  `/architecture:enforce-boundaries` and `/session-flow:retro`. It also says that a skill with no
  selection rule of its own has no row and hands the rung to `/corrections:codify`, and it names the
  readers that use a copy only for the boundary list. The CHANGELOG has a 3.2.0 entry, which is a
  minor bump as its own rules require. All five regenerated copies match the canonical text. There
  is one changelog fragment per carrier. The diff adds no new skill (AC10) and cuts nothing (AC7).
  CI on the judged head is green: Check Managed Files, Lint Repo, Check Skills, Test Bash (0) and
  Test Python (0) all succeeded
  (https://github.com/melodic-software/claude-code-plugins/actions/runs/43935023014). The capture
  was blocked by the credit gate, which counts neither way. Two parts of the spec are not met.
  First, AC12 requires every existing skill that places enforcement or an instruction to be on the
  recorded list with its route. That list exists only in the pull request body, and it is visibly
  cut off: it ends at `/planning:design` followed by "-…". So the list is incomplete and AC12 cannot
  be confirmed. Second, the dogfooding step appears only in a side remark under "Not verified" that
  refers to "the step named in the PR body", but the body never states it on its own. No secrets,
  takeover-set files, holds or settings are touched. I found no text that tries to steer the
  reviewer.
- PR body (Summary, "Recorded list"): the quoted list is cut off after `/planning:design` with "-…",
  so it is not true that every existing enforcement- or instruction-placing skill is on the recorded
  list with its route (AC12). Restore the full list in the body, or record it in a file in the
  repository, so that every skill from the sweep appears with its route.
- PR body: the dogfooding step is only mentioned in passing under "Not verified" ("the step named in
  the PR body"), and the body never actually states it. Add a line that names this repository's
  dogfooding step, e.g. running `/harness-config:audit-automation-gaps` here after release plus
  `scripts/sync-shared-copies.sh --check`.

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

pr-unblock / address-feedback: UNCERTAIN, left open

Evidence: docs/contracts/enforcement-ladder/README.md:201, .changes/architecture/work-resolve-issue-issue-16028-43931031836-1-c160e39a.md:1

Both points hold: the Summary's recorded list stops at /planning:design followed by a cut-off entry, and the body never states the dogfooding step on its own. Neither point can be fixed in this round. The full list cannot be recovered from the checkout: no file in the repository holds the sweep, and the cut-off entries (the rest of the presence-gated group and any groups after it) exist only in the earlier run's output. Rebuilding them would take a fresh sweep across every plugin's skills and agents, which is too much for one scoped change, and a guessed list would weaken AC12 rather than show it is met. The verifier's alternative, recording the list in a repository file, needs a new file, and check-paths refuses one (for example docs/contracts/enforcement-ladder/enforcement-placing-skills.md). Putting the sweep inside the ladder README (line 201 is where it says skills with no row hand the rung to codify) would add a per-repository inventory to a contract that is copied into five plugins. That is a design decision for a person. A body-only edit cannot carry a fix in this mode. A person should restore the full list, either in the body or in a file they choose, and add a line naming the dogfooding step, for example running /harness-config:audit-automation-gaps in this repository after release plus scripts/sync-shared-copies.sh --check.

@github-actions

Copy link
Copy Markdown
Contributor

Lane run report

Field Value
Activity pr-unblock/address-feedback
Model rung 0
Leg handoff
Run job success
Activity step success
Commits pushed none (read lane)
Run 43937891568

Failed GitHub API calls

Method Path Status
GET /repos/{repo}/git/ref/tags/* 404

@melodic-automation-lanes

Copy link
Copy Markdown
Contributor Author

pr-unblock / address-feedback finished round 1 of max 4 on 27fe8d0.

@github-actions

Copy link
Copy Markdown
Contributor

Lane run report

Field Value
Activity pr-unblock/publish-feedback
Model rung 0
Leg base
Run job success
Activity step success
Commits pushed none recorded
Run 43937891568

Failed GitHub API calls

Method Path Status
GET /repos/{repo}/git/ref/tags/* 404

@melodic-automation-lanes

Copy link
Copy Markdown
Contributor Author

The pr-unblock lane stopped address-feedback on this pull request: the lane left a pr-verify note open (uncertain, deferred or answered incorrect, or still open after the round cap and one fresh fixer round); pr-verify judges the next head afresh. The lane sets no label: the run ends in a failing pr-unblock/address-feedback status on the head the lane gated.

When a retry on this head can change the outcome, the lane runs one on its own, up to its per-head cap, and edits this comment if it stops again; a run that does not stop clears the status. When no retry can (a review thread the lane does not take or settle, or a failing check after the retries are spent), a person resolves the threads or gets the checks green, then dispatches the workflow below: that run finds the cause gone and clears the status. On a pull request the lanes App opened, the dispatchers start that run once per stop; on any other, a person does. A new commit to the branch starts clean.

To run it again on this head now, dispatch the pr-unblock workflow:

gh workflow run pr-unblock.yml -R melodic-software/claude-code-plugins -f pr-number=16602 -f activity=address-feedback

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

List enforcement-placing skills and give each one route to the ladder - part 1 of 2

0 participants