[Maintenance] Listing-budget watch: report the shared aggregate and drift

[kyle-sexton]

Context

Self-maintenance pilot (Boris-talk digest menu #9, first of two report-only routines). The full-install listing surface drifts upward silently — every new skill and every description edit moves the aggregate, and the shared-budget overflow degrades auto-invocation for the least-invoked tail without any error. #2022 holds the current baseline (86,499 chars / 134 skills at 19a3f7b3) and the pending target decision; this routine keeps the number observed either way.

Proposed work

Each cycle (report-only — this routine never edits descriptions):

• Run the pooled aggregate from a current main checkout: bash <skill-quality>/scripts/check-listing-budget.sh plugins/*/skills.
• Record on this item: the aggregate, skill count, delta vs the previous cycle's recorded number, and the top-10 contributors.
• If the aggregate moved >5% since the last cycle, or a single new skill enters the top-10, say so explicitly — that is the drift signal the routine exists to catch.
• Advance the schedule row via /work-items:track recheck listing-budget-watch.

Acceptance criteria

• Cycle numbers recorded as a comment on this item before it is closed by recheck.
• Delta vs prior cycle stated (or "first cycle — baseline Listing-budget decision: full-install aggregate is 86.5k chars (~10.8x the default budget) #2022").

References

• Listing-budget decision: full-install aggregate is 86.5k chars (~10.8x the default budget) #2022 — baseline numbers + pending listing-budget target decision (questions.md ci: onboard to the CI platform (ci-workflows + standards) #4)
• plugins/skill-quality/skills/check — listing-budget action (advisory by design)

Metadata

Field	Value
Category	general
Area	unspecified
Ecosystem	unspecified

Recurring

Cadence: quarterly
Triggers: none configured

[kyle-sexton]

First cycle discharged early (2026-08-08, ahead of the 2026-11-06 due date — exercised as a pilot while the listing-budget decision round was live).

Aggregate report (check-listing-budget.sh pooled over plugins/*/skills):

Measurement point	Aggregate	Skills	Roots
#2022 baseline (main 19a3f7b3)	86,499 chars (~10.8x default 8k)	134	61
Pre-tighten derived (main d894e3b3)	88,149 chars	136	62
Post-tighten (PR #2054 branch)	85,650 chars (~10.7x)	136	62

Drift observed this cycle: +1,650 chars of organic growth between the two mains (2 new listing-eligible skills, 1 new root) — the growth law the option-3 design issue (#2034) targets. The option-2 tighten pass (#2054) removes 2,499 chars from the 10 worst entries with all trigger phrases preserved (skill-quality check-3 verified).

Decision state: operator chose option 2 (executed, #2054) plus option 3 (design filed as #2034: curated bundles + per-plugin ceiling; that issue also proposes this routine gain per-plugin subtotals, which would make the next cycle's drift attributable per plugin).

Routine stays open as the standing recurring instance; next due 2026-11-06.

[kyle-sexton]

Out-of-cycle measurement — both drift triggers have fired

Recorded from a handoff-inbox validation pass, not a discharged cycle. The schedule row is untouched — last_checked stays 2026-08-08, next_due stays 2026-11-06, and /work-items:track recheck listing-budget-watch was not run. Treat this as an interim drift reading; the quarterly cycle is still outstanding.

Aggregate at HEAD

bash plugins/skill-quality/scripts/check-listing-budget.sh plugins/*/skills, run at main 685dd381:

Shared listing-budget estimate over 143 listing-eligible skill(s) across 65 root(s):
  aggregate: 93872 chars
  budget:    8000 chars (documented default (SLASH_COMMAND_TOOL_CHAR_BUDGET fallback))
WARN: aggregate exceeds the budget by 85872 chars (~11.7x over)
      Biggest contributors (entry chars, skill, root):
          1290  audit-prompting-postures  (plugins/claude-config/skills)
          1197  babysit-loop              (plugins/source-control/skills)
          1158  check                     (plugins/skill-quality/skills)
           999  draft-goal-condition      (plugins/planning/skills)
           980  manage                    (plugins/kindle-dedrm/skills)
           960  shape                     (plugins/adhd/skills)
           957  docpage-digest            (plugins/knowledge/skills)
           956  diagnose                  (plugins/computer-use/skills)
           948  audit-instructions        (plugins/claude-config/skills)
           948  visualize                 (plugins/visualization/skills)

CAPPED = 0 — no entry reaches the 1,536-char per-entry cap. The overflow is purely aggregate.

Delta — stated against both reference points

Reference	Aggregate	Skills	Δ chars	Δ %	Δ skills
#2022 baseline, main 19a3f7b3	86,499	134	+7,373	+8.5%	+9
Last recorded cycle (2026-08-08 post-tighten, PR #2054 branch, not main)	85,650	136	+8,222	+9.6%	+7
Validation-pass measurement, main dacb8661 (3 commits back)	93,037	142	+835	+0.9%	+1

The last recorded cycle's final number was measured on a branch, so the 19a3f7b3 main-to-main comparison is the cleaner one; both are given because #2023's procedure asks for "delta vs the previous cycle's recorded number". Both cross the >5% threshold.

The third row is its own datum: three commits moved the aggregate +0.9% in four and a half hours (dacb8661 2026-08-11 16:02 UTC → 685dd381 2026-08-11 20:26 UTC) — songwriting content plus the additive claude-ops:audit-install-state skill, 835 chars. Worth noting for anyone re-using a listing figure: a repo-wide aggregate goes stale on any commit that adds a listing-eligible skill, even one that touches no other surface.

Trigger 2 — a new skill entered the top-10, at #1

claude-config/audit-prompting-postures (1,290 chars) does not exist at 19a3f7b3 and is now the single largest entry in the marketplace. computer-use/diagnose (956) also entered. Eight of HEAD's top-10 are absent from #2022's top-10 — because every one of #2022's top-10 shrank under the option-2 tightening.

Attribution — option 2 landed, and was outrun

Every one of #2022's published top-10 is smaller at HEAD:

skill	19a3f7b3	HEAD
source-control/babysit-loop	1,468	1,197
claude-config/audit-instructions	1,197	948
docs-hygiene/audit-derivability	1,161	876
session-flow/running-retro	1,116	869
session-flow/reconcile	1,072	837
adhd/clarify	1,058	766
architecture/improve	1,057	900
discipline/script-the-deterministic-work	1,054	872
planning/audit-answers	1,028	899
discipline/reuse-or-replace	1,022	836

That is −2,233 chars from those ten alone — #2054 measurably landed. (Computed from the column difference in the table above: 11,233 → 9,000. Two other figures for this quantity are in circulation — this issue's first cycle comment says 2,499, and the source ledger says 2,378. Neither reconciles with the per-skill numbers above; 2,233 is the one those numbers support.) Against it, 9 new listing-eligible skills added +7,084 chars:

new skill	chars	note
claude-config/audit-prompting-postures	1,290	now the largest entry in the marketplace
computer-use/diagnose	956	new plugin
claude-ops/audit-install-state	835	landed after dacb8661
mutation-testing/audit	773	new plugin
songwriting/metaphor	771
mutation-testing/principles	695	new plugin
evals/design	625	new plugin
evals/methodology	624	new plugin
wizard/generate	515	new plugin

Four brand-new plugins (mutation-testing, evals, computer-use, wizard). Per-plugin, growth is mutation-testing +1,468, evals +1,249, claude-config +1,041, computer-use +956, claude-ops +835, songwriting +782, wizard +515; reductions are discipline −368, docs-hygiene −295, adhd −292, session-flow −282, source-control −271, architecture −157.

This is direct empirical support for #2034's growth-law premise — "tightening lowers the constant but not the growth law", now demonstrated rather than predicted. Trimming removed 2,233 chars from #2022's top-10; new surface added 7,084 in the same window. (These are the two largest forces, not a complete decomposition of the +7,373 — the remainder is ordinary edit churn across the rest of the fleet.)

Correction to a circulating figure

An intake item reported the melodic-software listing as 109,779 chars / 176 skills. That is a ~/.claude/plugins/cache/ figure and it omits the disable-model-invocation: true exclusion — 54 skills whose descriptions the docs say are "Description not in context" and which spend none of the shared budget. It overstates the repo's real listing surface by roughly 30%. The repo's own gate is the number to quote: 93,872 / 143 at 685dd381, ~11.7× the documented 8,000-char default. (Re-running the item's own method — every SKILL.md, description only, no exclusion — gives 120,658 / 196 at dacb8661 and 121,493 / 197 at 685dd381, so 109,779/176 is not that number either; the cache held only the enabled subset of plugins.)

Tooling caveat found while measuring

The command in this routine's own procedure takes 289 s on Windows and is killed at 180 s with zero output when run in the foreground — filed as #2216. It completes only when detached. That is a plausible contributing reason this routine has one recorded cycle: an operator or agent following the procedure literally on the machine where it is driven gets a hang, not a number. Cross-validated with an independent single-pass awk (567 ms) that reproduces this tool's output exactly at both 19a3f7b3 and HEAD.

Source ledger for every figure above: .work/handoff-inbox-batch-4/ledgers/I4-listing-budget.md.

[kyle-sexton]

Listing-budget watch — cycle report (2026-08-12)

Baseline: #2022 recorded 86,499 chars / 134 skills at 19a3f7b3.

This cycle (origin/main @ da05706):

bash plugins/skill-quality/scripts/check-listing-budget.sh plugins/*/skills

Metric	Value	Δ vs #2022 baseline
Aggregate	98,382 chars	+11,883 (+13.7%)
Skills	149	+15
Roots	65	—
Budget	8,000 chars	—
Over budget	90,382 chars (~12.3×)	+4,382 chars over prior gap

Top-10 contributors (entry chars):

1. audit-prompting-postures — 1305 (claude-config)
2. audit-performance — 1200 (claude-ops)
3. babysit-loop — 1197 (source-control)
4. check — 1158 (skill-quality)
5. draft-goal-condition — 999 (planning)
6. manage — 980 (kindle-dedrm)
7. inventory — 975 (claude-ops)
8. shape — 960 (adhd)
9. docpage-digest — 956 (knowledge)
10. diagnose — 956 (computer-use)

Drift signal: aggregate moved >5% since baseline (+13.7%). No new skill entered the top-10 this cycle (rank churn only among incumbents).

Routine is report-only; no description edits from this pass.

[kyle-sexton]

Recurring maintenance watch item for listing-budget aggregate reporting. Same rationale as #2024: the watch is automated; this issue is the report surface, not a backlog item. Closing.