v2.0.0 — harness-agnostic reframe

The native skill-listing budget (Claude Code v2.1.105, Codex) made ~half of the original 'measure & trim the cost' value the platform's job. So v2.0.0 reframes the skill around what's durable and portable.

• Portable core first — the problem + curation methodology (curate by INTENT not name, conservative asymmetry, blind re-rate, deterministic-over-LLM) are harness-agnostic.
• Researched cross-harness landscape (Cursor / Codex / Copilot CLI / Gemini CLI): all adopt the Agent Skills standard; a native listing budget exists only on Claude Code + Codex — on Cursor, Copilot CLI, Gemini CLI manual curation is still the live answer. disable-model-invocation is part of the standard, so it works verbatim on Cursor + Copilot CLI.
• 'Porting to another harness' recipe added.
• Claude Code levers demoted to a labeled implementation section; the retrieval-hook / 122k / forward-sweep work folded into a compact History footnote.
• Frontmatter description tightened under the 1536 per-entry cap (no self-truncation).

v1.10.0 — disable-model-invocation dual-role + reverse-audit lesson

The disable-model-invocation flag is ALSO the correct config for a user slash-command (it stops the model auto-firing /changelog, /lfg, /setup… while keeping /name), not just a trap-hide. So a reverse audit using a name-invoked → restore lens is a false-positive machine: a full body-read audit of all 487 hidden skills flagged 17 'wrongly hidden', but on a deterministic argument-hint/allowed-tools check ~16 were correctly-configured commands. Restoring a command would let the model auto-fire it and re-bloat the catalog. Genuine restores ≈ 1. A conservative re-rated forward extension also confirmed 0 new safe traps — post-budget the hide-sweep is largely played out. Added classification-rigor rules.

v1.9.0 — native catalog budget (skillListingBudgetFraction / /doctor)

Claude Code v2.1.105 (2026-04-13) shipped a native version of this skill's core thesis: a ~1% skill-listing budget that collapses the least-used skills' descriptions to bare names, plus a per-skill description cap — both surfaced by claude doctor.

What changed in v1.9.0

• New "native catalog budget" section documenting skillListingBudgetFraction (default 0.01) and maxSkillDescriptionChars (default 1536); /doctor is now the canonical readout.
• Flagged raising the budget fraction as the anti-pattern (~111k tokens/session — in the same ballpark as the skill's earlier ~122k full-description estimate; measured on the current already-trimmed catalog, so consistent rather than an exact match).
• Corrected two now-false claims — "standalone skills inject as bare names" and "name-only is a blanket no-op" — which only held before the budget made description-dropping usage-ranked.
• README + both plugin manifests updated; version 1.8.0 → 1.9.0.

v1.7.0

The root-cause work resolved: curation works, the retrieval-hook replacement doesn't.

• Separated the two strategies (curation vs retrieval-hook-replacement) and fixed the trap/procedure classifier to triage by description intent, not hyphen count (171/886 mislabeled).
• Proved the keyword hook can't replace force-load across five gate families — a base-rate wall (genuine-trap moments ~0.1% of triggers), precision-when-firing <0.3%. Embeddings then tested and fail identically (~23% recall @ ~99% firing). Shadow hook removed.
• Curation flip executed — 404 traps disable-model-invocationd after a blind second-rater + 2-of-3 tie-break rescued 33 mislabeled procedures; claudeception mint-default flipped so new traps mint hidden.
• README rewritten to match; plugin.json + marketplace.json bumped to 1.7.0.

Full changelog: v1.5.0...v1.7.0

v1.5.0

Initial public release of context-police — measure, trim, and report the runaway Claude Code skills-catalog token tax. Verified skillOverrides (per-project) + disable-model-invocation (global) levers, an interactive HTML recap, and the recall-gated retrieval-hook root-cause plan.