| name | dev-workflow-monthly-consolidation |
| description | Monthly maintenance routine for dev-workflow's char-budget health (G2b — "reduce what already exists", the counterpart to G2a's "stop growing" cap and A1's "reduce via structure"). Consolidates duplicate reminders in `skills/dev-workflow/SKILL.md` / `references/*.md`, re-audits intentional reinforcement-by-repetition, converts synonymous prose to cross-references, re-opens `dev-workflow-triage` issues that were deferred by the char-budget gate, dispatches `Skill(verify-skill-refs)` for governed-site enumeration drift, and checks G4 history-note sunset eligibility. Project-local routine — not for marketplace distribution. Trigger manually (`/dev-workflow-monthly-consolidation`) on a monthly cadence, or sooner when `dev-workflow-triage`'s summary reports a growing `budget-deferred (skipped, awaiting headroom)` count. |
| allowed-tools | Read, Grep, Glob, Edit, Skill(verify-skill-refs), Bash(wc -m *), Bash(git diff *), Bash(git add *), Bash(git commit *), Bash(git checkout HEAD -- *), Bash(git reset *), Bash(git rev-parse *), Bash(cp -R *), Bash(jq *), mcp__github__list_issues, mcp__github__issue_read, mcp__github__issue_write |
dev-workflow Monthly Consolidation
It is a project-local skill (lives under .claude/skills/dev-workflow-monthly-consolidation/, not registered in .claude-plugin/marketplace.json).
Purpose
skills/dev-workflow/SKILL.md and its references/*.md grow continuously (+53% over 3 weeks, observed 2026-07-02). Three countermeasures work together: G2a (the dev-workflow-triage char-budget gate) stops runaway growth, A1 reduces the total through structural extraction/compression, and G2b — this skill — periodically reduces what already exists (duplicate reminders, stale cross-references, prose that repeats the same point in multiple places). Reference: .claude/plans/dev-workflow-improvement-ideas.md § G2b / § G4.
Trigger
- Monthly, as a rough cadence (30 days — a discretionary starting point, adjust from operational experience), or
- Sooner, when
dev-workflow-triage's run summary shows the budget-deferred (skipped, awaiting headroom) count reaching 3 or more (also discretionary)
Owner
Manual invocation (/dev-workflow-monthly-consolidation). Scheduling this via /schedule (a cron-based routine) is left to the repo operator's discretion — this skill does not register a schedule itself.
No-Stall Principle
The generic regimen (sub-skill return discipline, non-fatal error handling, no standalone waiting turns) is defined in dev-workflow-triage's SKILL.md § No-Stall Principle and applies here without modification. This skill's own delta:
Zero designed user-gate points (same shape as triage-review): there is no interactive user-gate between checklist item 1 and item 7's write-back — each item records its outcome and moves to the next in the same turn; a single summary is emitted at the end.
Fatal-abort exits: none. Every failure mode this skill can hit is non-fatal — see below.
Non-fatal error classes (record and continue through to item 7's write-back): the § Step 3.7 — Release bookkeeping procedure's own failure branches (version-skew guard, invalid JSON, CHANGELOG edit error, scope leak, commit error) all record release-bookkeeping=failed (<reason>) and proceed rather than aborting the routine — the same disposition dev-workflow-triage gives them, since Step 3.7 is a once-per-run sub-procedure with no "next Finding" to retry into; item 4's mcp__github__* calls failing for a given issue (label-failed, per dev-workflow-triage's own disposition for the same operation); item 5's Skill(verify-skill-refs) dispatch failing to return a parseable verdict (record verify-skill-refs-failed and continue to item 6 — do not abort the run for a detect-only sub-skill's dispatch failure).
This skill does not use TodoWrite — its 7 checklist items run sequentially within a single turn with no sub-agent progress to track, unlike dev-workflow-triage's per-issue loop.
GitHub access path
Every SonicGarden/dev-workflow-issues operation in this skill uses the mcp__github__* tools listed in frontmatter — never gh CLI. See dev-workflow-triage's SKILL.md § Step 1 — Pre-flight's "GitHub access path" note for why (a documented Claude Code on the Web breakage). This skill supports the same non-interactive / cron-triggered execution mode, so it follows the same tool path.
Checklist
- Pre-flight:
Read skills/dev-workflow/SKILL.md and each skills/dev-workflow/references/*.md file once, and record their current char counts (wc -m) for the summary's before/after delta. Items 2, 3, and 6 below all judge this same read, so do not re-read/re-grep the file set per item.
- Build the intentional-duplication list, then consolidate everything else: from the Pre-flight read, first list every site explicitly marked as intentional reinforcement-by-repetition (e.g. "the abstract enumeration in § No-Stall Principle is intentionally duplicated here so the rule fires at the decision moment") and re-judge each — if the decision-moment justification has weakened (e.g. the callee's own return contract now makes the reminder redundant), mark it for consolidation too; otherwise keep it and record why. Then find reminder text repeated verbatim or near-verbatim elsewhere (not on that list) and consolidate each to one canonical location, replacing the others with a stable-phrase-anchor cross-reference. Building the intentional-duplication list first avoids consolidating a site that the next check would have told you to keep.
- Convert synonymous prose to cross-references: find passages in different sections that explain the same concept in different words, and reduce to one canonical explanation plus cross-references from the others (see
.claude/rules/project.rules.local.md's "Token defined-once + cross-reference" pattern).
- Re-open budget-deferred issues (independent of items 2/3 — this is a GitHub lookup with no dependency on the local-file consolidation, so it may run before or concurrently with them): call
mcp__github__list_issues (owner=SonicGarden, repo=dev-workflow-issues, label filter budget-deferred) to fetch issues deferred by the char-budget gate. For each issue this pass created headroom for, call mcp__github__issue_read then mcp__github__issue_write (read-then-write, so the label update never silently replaces the issue's other labels) to remove the budget-deferred label — this returns the issue to dev-workflow-triage's processable set on its next run. Zero matching issues is a normal outcome (no-op), not a failure.
- Governed-site enumeration check: dispatch
Skill(verify-skill-refs) against skills/dev-workflow/ and consume its Class (d) findings (governed-site enumeration gaps — the manifest already includes a subagent-model-read-sites row) rather than hand-auditing subagent_model or other governed-key enumerations here. Record any reported gap. Return-point reminder: regardless of outcome (findings reported, no gaps, or a dispatch failure per § No-Stall Principle), proceed directly to item 6 in the same turn — do not pause on the verdict.
- G4 history-note sunset check: for each Configuration bullet carrying a version-tagged history note (
Behavior change from vX.Y.Z, (History: v... )), find that flag's actual default-flip version in CHANGELOG.md — search for a flip the <flag> default entry; do not anchor on an earlier version mentioned only inside the (History: ...) parenthetical (introduction / gating changes that happened before the default actually flipped). The flip version's release date is the ## YYYY-MM-DD heading under which that CHANGELOG entry sits. Compute days elapsed since that date. A note is sunset-eligible only when both hold: (a) ≥ 4 weeks (28 days) have elapsed, and (b) the Configuration bullet's current-behavior description still reads coherently with the version references removed. Eligible notes: strip the version-tagged history clause, keeping the current-behavior description intact.
- Write back results: record what changed (or that nothing changed) — checklist items applied, char-count delta, re-opened issue count, verify-skill-refs findings, any G4 sunset applied — in
.claude/plans/dev-workflow-improvement-ideas.md § G2b.
Editing skills/dev-workflow/ (checklist items 2 / 3 / 6 — item 5 is detect-only and only records verify-skill-refs findings, it never edits)
Any edit to skills/dev-workflow/SKILL.md or references/*.md is a distributed-skill change. Follow dev-workflow-triage's SKILL.md § Step 3.7 — Release bookkeeping procedure (version-skew guard, Edit-based marketplace.json bump, jq empty, CHANGELOG prepend, bundle-copy sync via cp -R) verbatim, with two substitutions specific to this skill: the commit subject is chore(release): bump dev-workflow / dev-workflow-bundle (monthly-consolidation YYYY-MM-DD) (per .claude/rules/project.rules.local.md's bookkeeping-commit convention), and this skill has no per-Finding commit loop to separate the bookkeeping commit from — it is simply the one commit for this run's edits. git add is scoped to the exact paths this run touched (.claude-plugin/marketplace.json, CHANGELOG.md, the edited skills/dev-workflow/ files, and their bundle copies) — never git add -A / git add ., per this repo's pathspec-scoping convention for bookkeeping commits. "Verbatim" includes § Step 3.7's failure-recovery branches (git checkout HEAD -- <paths> on invalid JSON / CHANGELOG edit error / scope leak, git reset + git checkout HEAD -- <paths> on a non-zero-exit commit) — this is why Bash(git checkout HEAD -- *) / Bash(git reset *) are granted above even though this section's prose does not restate those branches' commands inline.
When no skills/dev-workflow/ edit landed this run (items 2/3/6 all found nothing to change), skip this sub-procedure entirely.
